openssl: add generated asm code

deps/openssl/asm/x64-elf-gas/bn/x86_64-mont.s

@@ -0,0 +1,172 @@
+.text
+
+
+.globl	bn_mul_mont
+.type	bn_mul_mont,@function
+.align	16
+bn_mul_mont:
+	pushq	%rbx
+	pushq	%rbp
+	pushq	%r12
+	pushq	%r13
+	pushq	%r14
+	pushq	%r15
+
+	movl	%r9d,%r9d
+	leaq	2(%r9),%r10
+	movq	%rsp,%r11
+	negq	%r10
+	leaq	(%rsp,%r10,8),%rsp
+	andq	$-1024,%rsp
+
+	movq	%r11,8(%rsp,%r9,8)
+.Lprologue:
+	movq	%rdx,%r12
+
+	movq	(%r8),%r8
+
+	xorq	%r14,%r14
+	xorq	%r15,%r15
+
+	movq	(%r12),%rbx
+	movq	(%rsi),%rax
+	mulq	%rbx
+	movq	%rax,%r10
+	movq	%rdx,%r11
+
+	imulq	%r8,%rax
+	movq	%rax,%rbp
+
+	mulq	(%rcx)
+	addq	%r10,%rax
+	adcq	$0,%rdx
+	movq	%rdx,%r13
+
+	leaq	1(%r15),%r15
+.L1st:
+	movq	(%rsi,%r15,8),%rax
+	mulq	%rbx
+	addq	%r11,%rax
+	adcq	$0,%rdx
+	movq	%rax,%r10
+	movq	(%rcx,%r15,8),%rax
+	movq	%rdx,%r11
+
+	mulq	%rbp
+	addq	%r13,%rax
+	leaq	1(%r15),%r15
+	adcq	$0,%rdx
+	addq	%r10,%rax
+	adcq	$0,%rdx
+	movq	%rax,-16(%rsp,%r15,8)
+	cmpq	%r9,%r15
+	movq	%rdx,%r13
+	jl	.L1st
+
+	xorq	%rdx,%rdx
+	addq	%r11,%r13
+	adcq	$0,%rdx
+	movq	%r13,-8(%rsp,%r9,8)
+	movq	%rdx,(%rsp,%r9,8)
+
+	leaq	1(%r14),%r14
+.align	4
+.Louter:
+	xorq	%r15,%r15
+
+	movq	(%r12,%r14,8),%rbx
+	movq	(%rsi),%rax
+	mulq	%rbx
+	addq	(%rsp),%rax
+	adcq	$0,%rdx
+	movq	%rax,%r10
+	movq	%rdx,%r11
+
+	imulq	%r8,%rax
+	movq	%rax,%rbp
+
+	mulq	(%rcx,%r15,8)
+	addq	%r10,%rax
+	movq	8(%rsp),%r10
+	adcq	$0,%rdx
+	movq	%rdx,%r13
+
+	leaq	1(%r15),%r15
+.align	4
+.Linner:
+	movq	(%rsi,%r15,8),%rax
+	mulq	%rbx
+	addq	%r11,%rax
+	adcq	$0,%rdx
+	addq	%rax,%r10
+	movq	(%rcx,%r15,8),%rax
+	adcq	$0,%rdx
+	movq	%rdx,%r11
+
+	mulq	%rbp
+	addq	%r13,%rax
+	leaq	1(%r15),%r15
+	adcq	$0,%rdx
+	addq	%r10,%rax
+	adcq	$0,%rdx
+	movq	(%rsp,%r15,8),%r10
+	cmpq	%r9,%r15
+	movq	%rax,-16(%rsp,%r15,8)
+	movq	%rdx,%r13
+	jl	.Linner
+
+	xorq	%rdx,%rdx
+	addq	%r11,%r13
+	adcq	$0,%rdx
+	addq	%r10,%r13
+	adcq	$0,%rdx
+	movq	%r13,-8(%rsp,%r9,8)
+	movq	%rdx,(%rsp,%r9,8)
+
+	leaq	1(%r14),%r14
+	cmpq	%r9,%r14
+	jl	.Louter
+
+	leaq	(%rsp),%rsi
+	leaq	-1(%r9),%r15
+
+	movq	(%rsi),%rax
+	xorq	%r14,%r14
+	jmp	.Lsub
+.align	16
+.Lsub:	sbbq	(%rcx,%r14,8),%rax
+	movq	%rax,(%rdi,%r14,8)
+	decq	%r15
+	movq	8(%rsi,%r14,8),%rax
+	leaq	1(%r14),%r14
+	jge	.Lsub
+
+	sbbq	$0,%rax
+	andq	%rax,%rsi
+	notq	%rax
+	movq	%rdi,%rcx
+	andq	%rax,%rcx
+	leaq	-1(%r9),%r15
+	orq	%rcx,%rsi
+.align	16
+.Lcopy:
+	movq	(%rsi,%r15,8),%rax
+	movq	%rax,(%rdi,%r15,8)
+	movq	%r14,(%rsp,%r15,8)
+	decq	%r15
+	jge	.Lcopy
+
+	movq	8(%rsp,%r9,8),%rsi
+	movq	$1,%rax
+	movq	(%rsi),%r15
+	movq	8(%rsi),%r14
+	movq	16(%rsi),%r13
+	movq	24(%rsi),%r12
+	movq	32(%rsi),%rbp
+	movq	40(%rsi),%rbx
+	leaq	48(%rsi),%rsp
+.Lepilogue:
+	.byte	0xf3,0xc3
+.size	bn_mul_mont,.-bn_mul_mont
+.byte	77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105,112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.align	16

deps/openssl/asm/x64-elf-gas/md5/md5-x86_64.s

@@ -0,0 +1,671 @@
+.text
+
+.align	16
+
+.globl	md5_block_asm_data_order
+.type	md5_block_asm_data_order,@function
+md5_block_asm_data_order:
+	pushq	%rbp
+	pushq	%rbx
+	pushq	%r12
+	pushq	%r14
+	pushq	%r15
+.Lprologue:
+
+
+
+
+	movq	%rdi,%rbp
+	shlq	$6,%rdx
+	leaq	(%rsi,%rdx,1),%rdi
+	movl	0(%rbp),%eax
+	movl	4(%rbp),%ebx
+	movl	8(%rbp),%ecx
+	movl	12(%rbp),%edx
+
+
+
+
+
+
+
+	cmpq	%rdi,%rsi
+	je	.Lend
+
+
+
+.Lloop:
+	movl	%eax,%r8d
+	movl	%ebx,%r9d
+	movl	%ecx,%r14d
+	movl	%edx,%r15d
+	movl	0(%rsi),%r10d
+	movl	%edx,%r11d
+	xorl	%ecx,%r11d
+	leal	-680876936(%rax,%r10,1),%eax
+	andl	%ebx,%r11d
+	xorl	%edx,%r11d
+	movl	4(%rsi),%r10d
+	addl	%r11d,%eax
+	roll	$7,%eax
+	movl	%ecx,%r11d
+	addl	%ebx,%eax
+	xorl	%ebx,%r11d
+	leal	-389564586(%rdx,%r10,1),%edx
+	andl	%eax,%r11d
+	xorl	%ecx,%r11d
+	movl	8(%rsi),%r10d
+	addl	%r11d,%edx
+	roll	$12,%edx
+	movl	%ebx,%r11d
+	addl	%eax,%edx
+	xorl	%eax,%r11d
+	leal	606105819(%rcx,%r10,1),%ecx
+	andl	%edx,%r11d
+	xorl	%ebx,%r11d
+	movl	12(%rsi),%r10d
+	addl	%r11d,%ecx
+	roll	$17,%ecx
+	movl	%eax,%r11d
+	addl	%edx,%ecx
+	xorl	%edx,%r11d
+	leal	-1044525330(%rbx,%r10,1),%ebx
+	andl	%ecx,%r11d
+	xorl	%eax,%r11d
+	movl	16(%rsi),%r10d
+	addl	%r11d,%ebx
+	roll	$22,%ebx
+	movl	%edx,%r11d
+	addl	%ecx,%ebx
+	xorl	%ecx,%r11d
+	leal	-176418897(%rax,%r10,1),%eax
+	andl	%ebx,%r11d
+	xorl	%edx,%r11d
+	movl	20(%rsi),%r10d
+	addl	%r11d,%eax
+	roll	$7,%eax
+	movl	%ecx,%r11d
+	addl	%ebx,%eax
+	xorl	%ebx,%r11d
+	leal	1200080426(%rdx,%r10,1),%edx
+	andl	%eax,%r11d
+	xorl	%ecx,%r11d
+	movl	24(%rsi),%r10d
+	addl	%r11d,%edx
+	roll	$12,%edx
+	movl	%ebx,%r11d
+	addl	%eax,%edx
+	xorl	%eax,%r11d
+	leal	-1473231341(%rcx,%r10,1),%ecx
+	andl	%edx,%r11d
+	xorl	%ebx,%r11d
+	movl	28(%rsi),%r10d
+	addl	%r11d,%ecx
+	roll	$17,%ecx
+	movl	%eax,%r11d
+	addl	%edx,%ecx
+	xorl	%edx,%r11d
+	leal	-45705983(%rbx,%r10,1),%ebx
+	andl	%ecx,%r11d
+	xorl	%eax,%r11d
+	movl	32(%rsi),%r10d
+	addl	%r11d,%ebx
+	roll	$22,%ebx
+	movl	%edx,%r11d
+	addl	%ecx,%ebx
+	xorl	%ecx,%r11d
+	leal	1770035416(%rax,%r10,1),%eax
+	andl	%ebx,%r11d
+	xorl	%edx,%r11d
+	movl	36(%rsi),%r10d
+	addl	%r11d,%eax
+	roll	$7,%eax
+	movl	%ecx,%r11d
+	addl	%ebx,%eax
+	xorl	%ebx,%r11d
+	leal	-1958414417(%rdx,%r10,1),%edx
+	andl	%eax,%r11d
+	xorl	%ecx,%r11d
+	movl	40(%rsi),%r10d
+	addl	%r11d,%edx
+	roll	$12,%edx
+	movl	%ebx,%r11d
+	addl	%eax,%edx
+	xorl	%eax,%r11d
+	leal	-42063(%rcx,%r10,1),%ecx
+	andl	%edx,%r11d
+	xorl	%ebx,%r11d
+	movl	44(%rsi),%r10d
+	addl	%r11d,%ecx
+	roll	$17,%ecx
+	movl	%eax,%r11d
+	addl	%edx,%ecx
+	xorl	%edx,%r11d
+	leal	-1990404162(%rbx,%r10,1),%ebx
+	andl	%ecx,%r11d
+	xorl	%eax,%r11d
+	movl	48(%rsi),%r10d
+	addl	%r11d,%ebx
+	roll	$22,%ebx
+	movl	%edx,%r11d
+	addl	%ecx,%ebx
+	xorl	%ecx,%r11d
+	leal	1804603682(%rax,%r10,1),%eax
+	andl	%ebx,%r11d
+	xorl	%edx,%r11d
+	movl	52(%rsi),%r10d
+	addl	%r11d,%eax
+	roll	$7,%eax
+	movl	%ecx,%r11d
+	addl	%ebx,%eax
+	xorl	%ebx,%r11d
+	leal	-40341101(%rdx,%r10,1),%edx
+	andl	%eax,%r11d
+	xorl	%ecx,%r11d
+	movl	56(%rsi),%r10d
+	addl	%r11d,%edx
+	roll	$12,%edx
+	movl	%ebx,%r11d
+	addl	%eax,%edx
+	xorl	%eax,%r11d
+	leal	-1502002290(%rcx,%r10,1),%ecx
+	andl	%edx,%r11d
+	xorl	%ebx,%r11d
+	movl	60(%rsi),%r10d
+	addl	%r11d,%ecx
+	roll	$17,%ecx
+	movl	%eax,%r11d
+	addl	%edx,%ecx
+	xorl	%edx,%r11d
+	leal	1236535329(%rbx,%r10,1),%ebx
+	andl	%ecx,%r11d
+	xorl	%eax,%r11d
+	movl	0(%rsi),%r10d
+	addl	%r11d,%ebx
+	roll	$22,%ebx
+	movl	%edx,%r11d
+	addl	%ecx,%ebx
+	movl	4(%rsi),%r10d
+	movl	%edx,%r11d
+	movl	%edx,%r12d
+	notl	%r11d
+	leal	-165796510(%rax,%r10,1),%eax
+	andl	%ebx,%r12d
+	andl	%ecx,%r11d
+	movl	24(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ecx,%r11d
+	addl	%r12d,%eax
+	movl	%ecx,%r12d
+	roll	$5,%eax
+	addl	%ebx,%eax
+	notl	%r11d
+	leal	-1069501632(%rdx,%r10,1),%edx
+	andl	%eax,%r12d
+	andl	%ebx,%r11d
+	movl	44(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ebx,%r11d
+	addl	%r12d,%edx
+	movl	%ebx,%r12d
+	roll	$9,%edx
+	addl	%eax,%edx
+	notl	%r11d
+	leal	643717713(%rcx,%r10,1),%ecx
+	andl	%edx,%r12d
+	andl	%eax,%r11d
+	movl	0(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%eax,%r11d
+	addl	%r12d,%ecx
+	movl	%eax,%r12d
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	notl	%r11d
+	leal	-373897302(%rbx,%r10,1),%ebx
+	andl	%ecx,%r12d
+	andl	%edx,%r11d
+	movl	20(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%edx,%r11d
+	addl	%r12d,%ebx
+	movl	%edx,%r12d
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	notl	%r11d
+	leal	-701558691(%rax,%r10,1),%eax
+	andl	%ebx,%r12d
+	andl	%ecx,%r11d
+	movl	40(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ecx,%r11d
+	addl	%r12d,%eax
+	movl	%ecx,%r12d
+	roll	$5,%eax
+	addl	%ebx,%eax
+	notl	%r11d
+	leal	38016083(%rdx,%r10,1),%edx
+	andl	%eax,%r12d
+	andl	%ebx,%r11d
+	movl	60(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ebx,%r11d
+	addl	%r12d,%edx
+	movl	%ebx,%r12d
+	roll	$9,%edx
+	addl	%eax,%edx
+	notl	%r11d
+	leal	-660478335(%rcx,%r10,1),%ecx
+	andl	%edx,%r12d
+	andl	%eax,%r11d
+	movl	16(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%eax,%r11d
+	addl	%r12d,%ecx
+	movl	%eax,%r12d
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	notl	%r11d
+	leal	-405537848(%rbx,%r10,1),%ebx
+	andl	%ecx,%r12d
+	andl	%edx,%r11d
+	movl	36(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%edx,%r11d
+	addl	%r12d,%ebx
+	movl	%edx,%r12d
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	notl	%r11d
+	leal	568446438(%rax,%r10,1),%eax
+	andl	%ebx,%r12d
+	andl	%ecx,%r11d
+	movl	56(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ecx,%r11d
+	addl	%r12d,%eax
+	movl	%ecx,%r12d
+	roll	$5,%eax
+	addl	%ebx,%eax
+	notl	%r11d
+	leal	-1019803690(%rdx,%r10,1),%edx
+	andl	%eax,%r12d
+	andl	%ebx,%r11d
+	movl	12(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ebx,%r11d
+	addl	%r12d,%edx
+	movl	%ebx,%r12d
+	roll	$9,%edx
+	addl	%eax,%edx
+	notl	%r11d
+	leal	-187363961(%rcx,%r10,1),%ecx
+	andl	%edx,%r12d
+	andl	%eax,%r11d
+	movl	32(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%eax,%r11d
+	addl	%r12d,%ecx
+	movl	%eax,%r12d
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	notl	%r11d
+	leal	1163531501(%rbx,%r10,1),%ebx
+	andl	%ecx,%r12d
+	andl	%edx,%r11d
+	movl	52(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%edx,%r11d
+	addl	%r12d,%ebx
+	movl	%edx,%r12d
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	notl	%r11d
+	leal	-1444681467(%rax,%r10,1),%eax
+	andl	%ebx,%r12d
+	andl	%ecx,%r11d
+	movl	8(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ecx,%r11d
+	addl	%r12d,%eax
+	movl	%ecx,%r12d
+	roll	$5,%eax
+	addl	%ebx,%eax
+	notl	%r11d
+	leal	-51403784(%rdx,%r10,1),%edx
+	andl	%eax,%r12d
+	andl	%ebx,%r11d
+	movl	28(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ebx,%r11d
+	addl	%r12d,%edx
+	movl	%ebx,%r12d
+	roll	$9,%edx
+	addl	%eax,%edx
+	notl	%r11d
+	leal	1735328473(%rcx,%r10,1),%ecx
+	andl	%edx,%r12d
+	andl	%eax,%r11d
+	movl	48(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%eax,%r11d
+	addl	%r12d,%ecx
+	movl	%eax,%r12d
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	notl	%r11d
+	leal	-1926607734(%rbx,%r10,1),%ebx
+	andl	%ecx,%r12d
+	andl	%edx,%r11d
+	movl	0(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%edx,%r11d
+	addl	%r12d,%ebx
+	movl	%edx,%r12d
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	movl	20(%rsi),%r10d
+	movl	%ecx,%r11d
+	leal	-378558(%rax,%r10,1),%eax
+	movl	32(%rsi),%r10d
+	xorl	%edx,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%eax
+	roll	$4,%eax
+	movl	%ebx,%r11d
+	addl	%ebx,%eax
+	leal	-2022574463(%rdx,%r10,1),%edx
+	movl	44(%rsi),%r10d
+	xorl	%ecx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%edx
+	roll	$11,%edx
+	movl	%eax,%r11d
+	addl	%eax,%edx
+	leal	1839030562(%rcx,%r10,1),%ecx
+	movl	56(%rsi),%r10d
+	xorl	%ebx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ecx
+	roll	$16,%ecx
+	movl	%edx,%r11d
+	addl	%edx,%ecx
+	leal	-35309556(%rbx,%r10,1),%ebx
+	movl	4(%rsi),%r10d
+	xorl	%eax,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%ebx
+	roll	$23,%ebx
+	movl	%ecx,%r11d
+	addl	%ecx,%ebx
+	leal	-1530992060(%rax,%r10,1),%eax
+	movl	16(%rsi),%r10d
+	xorl	%edx,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%eax
+	roll	$4,%eax
+	movl	%ebx,%r11d
+	addl	%ebx,%eax
+	leal	1272893353(%rdx,%r10,1),%edx
+	movl	28(%rsi),%r10d
+	xorl	%ecx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%edx
+	roll	$11,%edx
+	movl	%eax,%r11d
+	addl	%eax,%edx
+	leal	-155497632(%rcx,%r10,1),%ecx
+	movl	40(%rsi),%r10d
+	xorl	%ebx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ecx
+	roll	$16,%ecx
+	movl	%edx,%r11d
+	addl	%edx,%ecx
+	leal	-1094730640(%rbx,%r10,1),%ebx
+	movl	52(%rsi),%r10d
+	xorl	%eax,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%ebx
+	roll	$23,%ebx
+	movl	%ecx,%r11d
+	addl	%ecx,%ebx
+	leal	681279174(%rax,%r10,1),%eax
+	movl	0(%rsi),%r10d
+	xorl	%edx,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%eax
+	roll	$4,%eax
+	movl	%ebx,%r11d
+	addl	%ebx,%eax
+	leal	-358537222(%rdx,%r10,1),%edx
+	movl	12(%rsi),%r10d
+	xorl	%ecx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%edx
+	roll	$11,%edx
+	movl	%eax,%r11d
+	addl	%eax,%edx
+	leal	-722521979(%rcx,%r10,1),%ecx
+	movl	24(%rsi),%r10d
+	xorl	%ebx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ecx
+	roll	$16,%ecx
+	movl	%edx,%r11d
+	addl	%edx,%ecx
+	leal	76029189(%rbx,%r10,1),%ebx
+	movl	36(%rsi),%r10d
+	xorl	%eax,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%ebx
+	roll	$23,%ebx
+	movl	%ecx,%r11d
+	addl	%ecx,%ebx
+	leal	-640364487(%rax,%r10,1),%eax
+	movl	48(%rsi),%r10d
+	xorl	%edx,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%eax
+	roll	$4,%eax
+	movl	%ebx,%r11d
+	addl	%ebx,%eax
+	leal	-421815835(%rdx,%r10,1),%edx
+	movl	60(%rsi),%r10d
+	xorl	%ecx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%edx
+	roll	$11,%edx
+	movl	%eax,%r11d
+	addl	%eax,%edx
+	leal	530742520(%rcx,%r10,1),%ecx
+	movl	8(%rsi),%r10d
+	xorl	%ebx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ecx
+	roll	$16,%ecx
+	movl	%edx,%r11d
+	addl	%edx,%ecx
+	leal	-995338651(%rbx,%r10,1),%ebx
+	movl	0(%rsi),%r10d
+	xorl	%eax,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%ebx
+	roll	$23,%ebx
+	movl	%ecx,%r11d
+	addl	%ecx,%ebx
+	movl	0(%rsi),%r10d
+	movl	$4294967295,%r11d
+	xorl	%edx,%r11d
+	leal	-198630844(%rax,%r10,1),%eax
+	orl	%ebx,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%eax
+	movl	28(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$6,%eax
+	xorl	%ecx,%r11d
+	addl	%ebx,%eax
+	leal	1126891415(%rdx,%r10,1),%edx
+	orl	%eax,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%edx
+	movl	56(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$10,%edx
+	xorl	%ebx,%r11d
+	addl	%eax,%edx
+	leal	-1416354905(%rcx,%r10,1),%ecx
+	orl	%edx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%ecx
+	movl	20(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$15,%ecx
+	xorl	%eax,%r11d
+	addl	%edx,%ecx
+	leal	-57434055(%rbx,%r10,1),%ebx
+	orl	%ecx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ebx
+	movl	48(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$21,%ebx
+	xorl	%edx,%r11d
+	addl	%ecx,%ebx
+	leal	1700485571(%rax,%r10,1),%eax
+	orl	%ebx,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%eax
+	movl	12(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$6,%eax
+	xorl	%ecx,%r11d
+	addl	%ebx,%eax
+	leal	-1894986606(%rdx,%r10,1),%edx
+	orl	%eax,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%edx
+	movl	40(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$10,%edx
+	xorl	%ebx,%r11d
+	addl	%eax,%edx
+	leal	-1051523(%rcx,%r10,1),%ecx
+	orl	%edx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%ecx
+	movl	4(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$15,%ecx
+	xorl	%eax,%r11d
+	addl	%edx,%ecx
+	leal	-2054922799(%rbx,%r10,1),%ebx
+	orl	%ecx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ebx
+	movl	32(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$21,%ebx
+	xorl	%edx,%r11d
+	addl	%ecx,%ebx
+	leal	1873313359(%rax,%r10,1),%eax
+	orl	%ebx,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%eax
+	movl	60(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$6,%eax
+	xorl	%ecx,%r11d
+	addl	%ebx,%eax
+	leal	-30611744(%rdx,%r10,1),%edx
+	orl	%eax,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%edx
+	movl	24(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$10,%edx
+	xorl	%ebx,%r11d
+	addl	%eax,%edx
+	leal	-1560198380(%rcx,%r10,1),%ecx
+	orl	%edx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%ecx
+	movl	52(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$15,%ecx
+	xorl	%eax,%r11d
+	addl	%edx,%ecx
+	leal	1309151649(%rbx,%r10,1),%ebx
+	orl	%ecx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ebx
+	movl	16(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$21,%ebx
+	xorl	%edx,%r11d
+	addl	%ecx,%ebx
+	leal	-145523070(%rax,%r10,1),%eax
+	orl	%ebx,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%eax
+	movl	44(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$6,%eax
+	xorl	%ecx,%r11d
+	addl	%ebx,%eax
+	leal	-1120210379(%rdx,%r10,1),%edx
+	orl	%eax,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%edx
+	movl	8(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$10,%edx
+	xorl	%ebx,%r11d
+	addl	%eax,%edx
+	leal	718787259(%rcx,%r10,1),%ecx
+	orl	%edx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%ecx
+	movl	36(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$15,%ecx
+	xorl	%eax,%r11d
+	addl	%edx,%ecx
+	leal	-343485551(%rbx,%r10,1),%ebx
+	orl	%ecx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ebx
+	movl	0(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$21,%ebx
+	xorl	%edx,%r11d
+	addl	%ecx,%ebx
+
+	addl	%r8d,%eax
+	addl	%r9d,%ebx
+	addl	%r14d,%ecx
+	addl	%r15d,%edx
+
+
+	addq	$64,%rsi
+	cmpq	%rdi,%rsi
+	jb	.Lloop
+
+
+
+.Lend:
+	movl	%eax,0(%rbp)
+	movl	%ebx,4(%rbp)
+	movl	%ecx,8(%rbp)
+	movl	%edx,12(%rbp)
+
+	movq	(%rsp),%r15
+	movq	8(%rsp),%r14
+	movq	16(%rsp),%r12
+	movq	24(%rsp),%rbx
+	movq	32(%rsp),%rbp
+	addq	$40,%rsp
+.Lepilogue:
+	.byte	0xf3,0xc3
+.size	md5_block_asm_data_order,.-md5_block_asm_data_order

deps/openssl/asm/x64-elf-gas/rc4/rc4-x86_64.s

@@ -0,0 +1,430 @@
+.text
+
+
+.globl	RC4
+.type	RC4,@function
+.align	16
+RC4:	orq	%rsi,%rsi
+	jne	.Lentry
+	.byte	0xf3,0xc3
+.Lentry:
+	pushq	%rbx
+	pushq	%r12
+	pushq	%r13
+.Lprologue:
+
+	addq	$8,%rdi
+	movl	-8(%rdi),%r8d
+	movl	-4(%rdi),%r12d
+	cmpl	$-1,256(%rdi)
+	je	.LRC4_CHAR
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	testq	$-8,%rsi
+	jz	.Lloop1
+	jmp	.Lloop8
+.align	16
+.Lloop8:
+	addb	%r9b,%r12b
+	movq	%r8,%r10
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r10b
+	movl	(%rdi,%r10,4),%r11d
+	cmpq	%r10,%r12
+	movl	%r9d,(%rdi,%r12,4)
+	cmoveq	%r9,%r11
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r9b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r11b,%r12b
+	movq	%r10,%r8
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	cmpq	%r8,%r12
+	movl	%r11d,(%rdi,%r12,4)
+	cmoveq	%r11,%r9
+	movl	%r13d,(%rdi,%r10,4)
+	addb	%r11b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r9b,%r12b
+	movq	%r8,%r10
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r10b
+	movl	(%rdi,%r10,4),%r11d
+	cmpq	%r10,%r12
+	movl	%r9d,(%rdi,%r12,4)
+	cmoveq	%r9,%r11
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r9b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r11b,%r12b
+	movq	%r10,%r8
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	cmpq	%r8,%r12
+	movl	%r11d,(%rdi,%r12,4)
+	cmoveq	%r11,%r9
+	movl	%r13d,(%rdi,%r10,4)
+	addb	%r11b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r9b,%r12b
+	movq	%r8,%r10
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r10b
+	movl	(%rdi,%r10,4),%r11d
+	cmpq	%r10,%r12
+	movl	%r9d,(%rdi,%r12,4)
+	cmoveq	%r9,%r11
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r9b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r11b,%r12b
+	movq	%r10,%r8
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	cmpq	%r8,%r12
+	movl	%r11d,(%rdi,%r12,4)
+	cmoveq	%r11,%r9
+	movl	%r13d,(%rdi,%r10,4)
+	addb	%r11b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r9b,%r12b
+	movq	%r8,%r10
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r10b
+	movl	(%rdi,%r10,4),%r11d
+	cmpq	%r10,%r12
+	movl	%r9d,(%rdi,%r12,4)
+	cmoveq	%r9,%r11
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r9b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r11b,%r12b
+	movq	%r10,%r8
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	cmpq	%r8,%r12
+	movl	%r11d,(%rdi,%r12,4)
+	cmoveq	%r11,%r9
+	movl	%r13d,(%rdi,%r10,4)
+	addb	%r11b,%r13b
+	movb	(%rdi,%r13,4),%al
+	rorq	$8,%rax
+	subq	$8,%rsi
+
+	xorq	(%rdx),%rax
+	addq	$8,%rdx
+	movq	%rax,(%rcx)
+	addq	$8,%rcx
+
+	testq	$-8,%rsi
+	jnz	.Lloop8
+	cmpq	$0,%rsi
+	jne	.Lloop1
+	jmp	.Lexit
+
+.align	16
+.Lloop1:
+	addb	%r9b,%r12b
+	movl	(%rdi,%r12,4),%r13d
+	movl	%r9d,(%rdi,%r12,4)
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r13b,%r9b
+	incb	%r8b
+	movl	(%rdi,%r9,4),%r13d
+	movl	(%rdi,%r8,4),%r9d
+	xorb	(%rdx),%r13b
+	incq	%rdx
+	movb	%r13b,(%rcx)
+	incq	%rcx
+	decq	%rsi
+	jnz	.Lloop1
+	jmp	.Lexit
+
+.align	16
+.LRC4_CHAR:
+	addb	$1,%r8b
+	movzbl	(%rdi,%r8,1),%r9d
+	testq	$-8,%rsi
+	jz	.Lcloop1
+	cmpl	$0,260(%rdi)
+	jnz	.Lcloop1
+	jmp	.Lcloop8
+.align	16
+.Lcloop8:
+	movl	(%rdx),%eax
+	movl	4(%rdx),%ebx
+	addb	%r9b,%r12b
+	leaq	1(%r8),%r10
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r10b,%r10d
+	movzbl	(%rdi,%r10,1),%r11d
+	movb	%r9b,(%rdi,%r12,1)
+	cmpq	%r10,%r12
+	movb	%r13b,(%rdi,%r8,1)
+	jne	.Lcmov0
+
+	movq	%r9,%r11
+.Lcmov0:
+	addb	%r9b,%r13b
+	xorb	(%rdi,%r13,1),%al
+	rorl	$8,%eax
+	addb	%r11b,%r12b
+	leaq	1(%r10),%r8
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r8,1),%r9d
+	movb	%r11b,(%rdi,%r12,1)
+	cmpq	%r8,%r12
+	movb	%r13b,(%rdi,%r10,1)
+	jne	.Lcmov1
+
+	movq	%r11,%r9
+.Lcmov1:
+	addb	%r11b,%r13b
+	xorb	(%rdi,%r13,1),%al
+	rorl	$8,%eax
+	addb	%r9b,%r12b
+	leaq	1(%r8),%r10
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r10b,%r10d
+	movzbl	(%rdi,%r10,1),%r11d
+	movb	%r9b,(%rdi,%r12,1)
+	cmpq	%r10,%r12
+	movb	%r13b,(%rdi,%r8,1)
+	jne	.Lcmov2
+
+	movq	%r9,%r11
+.Lcmov2:
+	addb	%r9b,%r13b
+	xorb	(%rdi,%r13,1),%al
+	rorl	$8,%eax
+	addb	%r11b,%r12b
+	leaq	1(%r10),%r8
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r8,1),%r9d
+	movb	%r11b,(%rdi,%r12,1)
+	cmpq	%r8,%r12
+	movb	%r13b,(%rdi,%r10,1)
+	jne	.Lcmov3
+
+	movq	%r11,%r9
+.Lcmov3:
+	addb	%r11b,%r13b
+	xorb	(%rdi,%r13,1),%al
+	rorl	$8,%eax
+	addb	%r9b,%r12b
+	leaq	1(%r8),%r10
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r10b,%r10d
+	movzbl	(%rdi,%r10,1),%r11d
+	movb	%r9b,(%rdi,%r12,1)
+	cmpq	%r10,%r12
+	movb	%r13b,(%rdi,%r8,1)
+	jne	.Lcmov4
+
+	movq	%r9,%r11
+.Lcmov4:
+	addb	%r9b,%r13b
+	xorb	(%rdi,%r13,1),%bl
+	rorl	$8,%ebx
+	addb	%r11b,%r12b
+	leaq	1(%r10),%r8
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r8,1),%r9d
+	movb	%r11b,(%rdi,%r12,1)
+	cmpq	%r8,%r12
+	movb	%r13b,(%rdi,%r10,1)
+	jne	.Lcmov5
+
+	movq	%r11,%r9
+.Lcmov5:
+	addb	%r11b,%r13b
+	xorb	(%rdi,%r13,1),%bl
+	rorl	$8,%ebx
+	addb	%r9b,%r12b
+	leaq	1(%r8),%r10
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r10b,%r10d
+	movzbl	(%rdi,%r10,1),%r11d
+	movb	%r9b,(%rdi,%r12,1)
+	cmpq	%r10,%r12
+	movb	%r13b,(%rdi,%r8,1)
+	jne	.Lcmov6
+
+	movq	%r9,%r11
+.Lcmov6:
+	addb	%r9b,%r13b
+	xorb	(%rdi,%r13,1),%bl
+	rorl	$8,%ebx
+	addb	%r11b,%r12b
+	leaq	1(%r10),%r8
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r8,1),%r9d
+	movb	%r11b,(%rdi,%r12,1)
+	cmpq	%r8,%r12
+	movb	%r13b,(%rdi,%r10,1)
+	jne	.Lcmov7
+
+	movq	%r11,%r9
+.Lcmov7:
+	addb	%r11b,%r13b
+	xorb	(%rdi,%r13,1),%bl
+	rorl	$8,%ebx
+	leaq	-8(%rsi),%rsi
+	movl	%eax,(%rcx)
+	leaq	8(%rdx),%rdx
+	movl	%ebx,4(%rcx)
+	leaq	8(%rcx),%rcx
+
+	testq	$-8,%rsi
+	jnz	.Lcloop8
+	cmpq	$0,%rsi
+	jne	.Lcloop1
+	jmp	.Lexit
+.align	16
+.Lcloop1:
+	addb	%r9b,%r12b
+	movzbl	(%rdi,%r12,1),%r13d
+	movb	%r9b,(%rdi,%r12,1)
+	movb	%r13b,(%rdi,%r8,1)
+	addb	%r9b,%r13b
+	addb	$1,%r8b
+	movzbl	%r13b,%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r13,1),%r13d
+	movzbl	(%rdi,%r8,1),%r9d
+	xorb	(%rdx),%r13b
+	leaq	1(%rdx),%rdx
+	movb	%r13b,(%rcx)
+	leaq	1(%rcx),%rcx
+	subq	$1,%rsi
+	jnz	.Lcloop1
+	jmp	.Lexit
+
+.align	16
+.Lexit:
+	subb	$1,%r8b
+	movl	%r8d,-8(%rdi)
+	movl	%r12d,-4(%rdi)
+
+	movq	(%rsp),%r13
+	movq	8(%rsp),%r12
+	movq	16(%rsp),%rbx
+	addq	$24,%rsp
+.Lepilogue:
+	.byte	0xf3,0xc3
+.size	RC4,.-RC4
+
+.globl	RC4_set_key
+.type	RC4_set_key,@function
+.align	16
+RC4_set_key:
+	leaq	8(%rdi),%rdi
+	leaq	(%rdx,%rsi,1),%rdx
+	negq	%rsi
+	movq	%rsi,%rcx
+	xorl	%eax,%eax
+	xorq	%r9,%r9
+	xorq	%r10,%r10
+	xorq	%r11,%r11
+
+	movl	OPENSSL_ia32cap_P(%rip),%r8d
+	btl	$20,%r8d
+	jnc	.Lw1stloop
+	btl	$30,%r8d
+	setc	%r9b
+	movl	%r9d,260(%rdi)
+	jmp	.Lc1stloop
+
+.align	16
+.Lw1stloop:
+	movl	%eax,(%rdi,%rax,4)
+	addb	$1,%al
+	jnc	.Lw1stloop
+
+	xorq	%r9,%r9
+	xorq	%r8,%r8
+.align	16
+.Lw2ndloop:
+	movl	(%rdi,%r9,4),%r10d
+	addb	(%rdx,%rsi,1),%r8b
+	addb	%r10b,%r8b
+	addq	$1,%rsi
+	movl	(%rdi,%r8,4),%r11d
+	cmovzq	%rcx,%rsi
+	movl	%r10d,(%rdi,%r8,4)
+	movl	%r11d,(%rdi,%r9,4)
+	addb	$1,%r9b
+	jnc	.Lw2ndloop
+	jmp	.Lexit_key
+
+.align	16
+.Lc1stloop:
+	movb	%al,(%rdi,%rax,1)
+	addb	$1,%al
+	jnc	.Lc1stloop
+
+	xorq	%r9,%r9
+	xorq	%r8,%r8
+.align	16
+.Lc2ndloop:
+	movb	(%rdi,%r9,1),%r10b
+	addb	(%rdx,%rsi,1),%r8b
+	addb	%r10b,%r8b
+	addq	$1,%rsi
+	movb	(%rdi,%r8,1),%r11b
+	jnz	.Lcnowrap
+	movq	%rcx,%rsi
+.Lcnowrap:
+	movb	%r10b,(%rdi,%r8,1)
+	movb	%r11b,(%rdi,%r9,1)
+	addb	$1,%r9b
+	jnc	.Lc2ndloop
+	movl	$-1,256(%rdi)
+
+.align	16
+.Lexit_key:
+	xorl	%eax,%eax
+	movl	%eax,-8(%rdi)
+	movl	%eax,-4(%rdi)
+	.byte	0xf3,0xc3
+.size	RC4_set_key,.-RC4_set_key
+
+.globl	RC4_options
+.type	RC4_options,@function
+.align	16
+RC4_options:
+	leaq	.Lopts(%rip),%rax
+	movl	OPENSSL_ia32cap_P(%rip),%edx
+	btl	$20,%edx
+	jnc	.Ldone
+	addq	$12,%rax
+	btl	$30,%edx
+	jnc	.Ldone
+	addq	$13,%rax
+.Ldone:
+	.byte	0xf3,0xc3
+.align	64
+.Lopts:
+.byte	114,99,52,40,56,120,44,105,110,116,41,0
+.byte	114,99,52,40,56,120,44,99,104,97,114,41,0
+.byte	114,99,52,40,49,120,44,99,104,97,114,41,0
+.byte	82,67,52,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.align	64
+.size	RC4_options,.-RC4_options

deps/openssl/asm/x64-elf-gas/whrlpool/wp-x86_64.s

@@ -0,0 +1,859 @@
+.text
+
+
+.globl	whirlpool_block
+.type	whirlpool_block,@function
+.align	16
+whirlpool_block:
+	pushq	%rbx
+	pushq	%rbp
+	pushq	%r12
+	pushq	%r13
+	pushq	%r14
+	pushq	%r15
+
+	movq	%rsp,%r11
+	subq	$128+40,%rsp
+	andq	$-64,%rsp
+
+	leaq	128(%rsp),%r10
+	movq	%rdi,0(%r10)
+	movq	%rsi,8(%r10)
+	movq	%rdx,16(%r10)
+	movq	%r11,32(%r10)
+.Lprologue:
+
+	movq	%r10,%rbx
+	leaq	.Ltable(%rip),%rbp
+
+	xorq	%rcx,%rcx
+	xorq	%rdx,%rdx
+	movq	0(%rdi),%r8
+	movq	8(%rdi),%r9
+	movq	16(%rdi),%r10
+	movq	24(%rdi),%r11
+	movq	32(%rdi),%r12
+	movq	40(%rdi),%r13
+	movq	48(%rdi),%r14
+	movq	56(%rdi),%r15
+.Louterloop:
+	movq	%r8,0(%rsp)
+	movq	%r9,8(%rsp)
+	movq	%r10,16(%rsp)
+	movq	%r11,24(%rsp)
+	movq	%r12,32(%rsp)
+	movq	%r13,40(%rsp)
+	movq	%r14,48(%rsp)
+	movq	%r15,56(%rsp)
+	xorq	0(%rsi),%r8
+	xorq	8(%rsi),%r9
+	xorq	16(%rsi),%r10
+	xorq	24(%rsi),%r11
+	xorq	32(%rsi),%r12
+	xorq	40(%rsi),%r13
+	xorq	48(%rsi),%r14
+	xorq	56(%rsi),%r15
+	movq	%r8,64+0(%rsp)
+	movq	%r9,64+8(%rsp)
+	movq	%r10,64+16(%rsp)
+	movq	%r11,64+24(%rsp)
+	movq	%r12,64+32(%rsp)
+	movq	%r13,64+40(%rsp)
+	movq	%r14,64+48(%rsp)
+	movq	%r15,64+56(%rsp)
+	xorq	%rsi,%rsi
+	movq	%rsi,24(%rbx)
+.align	16
+.Lround:
+	movq	4096(%rbp,%rsi,8),%r8
+	movl	0(%rsp),%eax
+	movl	4(%rsp),%ebx
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r8
+	movq	7(%rbp,%rdi,8),%r9
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	0+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	movq	6(%rbp,%rsi,8),%r10
+	movq	5(%rbp,%rdi,8),%r11
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	movq	4(%rbp,%rsi,8),%r12
+	movq	3(%rbp,%rdi,8),%r13
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	0+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	movq	2(%rbp,%rsi,8),%r14
+	movq	1(%rbp,%rdi,8),%r15
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r9
+	xorq	7(%rbp,%rdi,8),%r10
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	8+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r11
+	xorq	5(%rbp,%rdi,8),%r12
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r13
+	xorq	3(%rbp,%rdi,8),%r14
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	8+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r15
+	xorq	1(%rbp,%rdi,8),%r8
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r10
+	xorq	7(%rbp,%rdi,8),%r11
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	16+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r12
+	xorq	5(%rbp,%rdi,8),%r13
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r14
+	xorq	3(%rbp,%rdi,8),%r15
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	16+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r8
+	xorq	1(%rbp,%rdi,8),%r9
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r11
+	xorq	7(%rbp,%rdi,8),%r12
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	24+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r13
+	xorq	5(%rbp,%rdi,8),%r14
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r15
+	xorq	3(%rbp,%rdi,8),%r8
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	24+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r9
+	xorq	1(%rbp,%rdi,8),%r10
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r12
+	xorq	7(%rbp,%rdi,8),%r13
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	32+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r14
+	xorq	5(%rbp,%rdi,8),%r15
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r8
+	xorq	3(%rbp,%rdi,8),%r9
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	32+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r10
+	xorq	1(%rbp,%rdi,8),%r11
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r13
+	xorq	7(%rbp,%rdi,8),%r14
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	40+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r15
+	xorq	5(%rbp,%rdi,8),%r8
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r9
+	xorq	3(%rbp,%rdi,8),%r10
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	40+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r11
+	xorq	1(%rbp,%rdi,8),%r12
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r14
+	xorq	7(%rbp,%rdi,8),%r15
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	48+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r8
+	xorq	5(%rbp,%rdi,8),%r9
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r10
+	xorq	3(%rbp,%rdi,8),%r11
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	48+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r12
+	xorq	1(%rbp,%rdi,8),%r13
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r15
+	xorq	7(%rbp,%rdi,8),%r8
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	56+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r9
+	xorq	5(%rbp,%rdi,8),%r10
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r11
+	xorq	3(%rbp,%rdi,8),%r12
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	56+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r13
+	xorq	1(%rbp,%rdi,8),%r14
+	movq	%r8,0(%rsp)
+	movq	%r9,8(%rsp)
+	movq	%r10,16(%rsp)
+	movq	%r11,24(%rsp)
+	movq	%r12,32(%rsp)
+	movq	%r13,40(%rsp)
+	movq	%r14,48(%rsp)
+	movq	%r15,56(%rsp)
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r8
+	xorq	7(%rbp,%rdi,8),%r9
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+0+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r10
+	xorq	5(%rbp,%rdi,8),%r11
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r12
+	xorq	3(%rbp,%rdi,8),%r13
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+0+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r14
+	xorq	1(%rbp,%rdi,8),%r15
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r9
+	xorq	7(%rbp,%rdi,8),%r10
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+8+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r11
+	xorq	5(%rbp,%rdi,8),%r12
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r13
+	xorq	3(%rbp,%rdi,8),%r14
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+8+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r15
+	xorq	1(%rbp,%rdi,8),%r8
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r10
+	xorq	7(%rbp,%rdi,8),%r11
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+16+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r12
+	xorq	5(%rbp,%rdi,8),%r13
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r14
+	xorq	3(%rbp,%rdi,8),%r15
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+16+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r8
+	xorq	1(%rbp,%rdi,8),%r9
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r11
+	xorq	7(%rbp,%rdi,8),%r12
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+24+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r13
+	xorq	5(%rbp,%rdi,8),%r14
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r15
+	xorq	3(%rbp,%rdi,8),%r8
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+24+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r9
+	xorq	1(%rbp,%rdi,8),%r10
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r12
+	xorq	7(%rbp,%rdi,8),%r13
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+32+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r14
+	xorq	5(%rbp,%rdi,8),%r15
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r8
+	xorq	3(%rbp,%rdi,8),%r9
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+32+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r10
+	xorq	1(%rbp,%rdi,8),%r11
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r13
+	xorq	7(%rbp,%rdi,8),%r14
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+40+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r15
+	xorq	5(%rbp,%rdi,8),%r8
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r9
+	xorq	3(%rbp,%rdi,8),%r10
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+40+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r11
+	xorq	1(%rbp,%rdi,8),%r12
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r14
+	xorq	7(%rbp,%rdi,8),%r15
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+48+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r8
+	xorq	5(%rbp,%rdi,8),%r9
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r10
+	xorq	3(%rbp,%rdi,8),%r11
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+48+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r12
+	xorq	1(%rbp,%rdi,8),%r13
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r15
+	xorq	7(%rbp,%rdi,8),%r8
+	movb	%al,%cl
+	movb	%ah,%dl
+
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r9
+	xorq	5(%rbp,%rdi,8),%r10
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r11
+	xorq	3(%rbp,%rdi,8),%r12
+	movb	%bl,%cl
+	movb	%bh,%dl
+
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r13
+	xorq	1(%rbp,%rdi,8),%r14
+	leaq	128(%rsp),%rbx
+	movq	24(%rbx),%rsi
+	addq	$1,%rsi
+	cmpq	$10,%rsi
+	je	.Lroundsdone
+
+	movq	%rsi,24(%rbx)
+	movq	%r8,64+0(%rsp)
+	movq	%r9,64+8(%rsp)
+	movq	%r10,64+16(%rsp)
+	movq	%r11,64+24(%rsp)
+	movq	%r12,64+32(%rsp)
+	movq	%r13,64+40(%rsp)
+	movq	%r14,64+48(%rsp)
+	movq	%r15,64+56(%rsp)
+	jmp	.Lround
+.align	16
+.Lroundsdone:
+	movq	0(%rbx),%rdi
+	movq	8(%rbx),%rsi
+	movq	16(%rbx),%rax
+	xorq	0(%rsi),%r8
+	xorq	8(%rsi),%r9
+	xorq	16(%rsi),%r10
+	xorq	24(%rsi),%r11
+	xorq	32(%rsi),%r12
+	xorq	40(%rsi),%r13
+	xorq	48(%rsi),%r14
+	xorq	56(%rsi),%r15
+	xorq	0(%rdi),%r8
+	xorq	8(%rdi),%r9
+	xorq	16(%rdi),%r10
+	xorq	24(%rdi),%r11
+	xorq	32(%rdi),%r12
+	xorq	40(%rdi),%r13
+	xorq	48(%rdi),%r14
+	xorq	56(%rdi),%r15
+	movq	%r8,0(%rdi)
+	movq	%r9,8(%rdi)
+	movq	%r10,16(%rdi)
+	movq	%r11,24(%rdi)
+	movq	%r12,32(%rdi)
+	movq	%r13,40(%rdi)
+	movq	%r14,48(%rdi)
+	movq	%r15,56(%rdi)
+	leaq	64(%rsi),%rsi
+	subq	$1,%rax
+	jz	.Lalldone
+	movq	%rsi,8(%rbx)
+	movq	%rax,16(%rbx)
+	jmp	.Louterloop
+.Lalldone:
+	movq	32(%rbx),%rsi
+	movq	(%rsi),%r15
+	movq	8(%rsi),%r14
+	movq	16(%rsi),%r13
+	movq	24(%rsi),%r12
+	movq	32(%rsi),%rbp
+	movq	40(%rsi),%rbx
+	leaq	48(%rsi),%rsp
+.Lepilogue:
+	.byte	0xf3,0xc3
+.size	whirlpool_block,.-whirlpool_block
+
+.align	64
+.type	.Ltable,@object
+.Ltable:
+.byte	24,24,96,24,192,120,48,216,24,24,96,24,192,120,48,216
+.byte	35,35,140,35,5,175,70,38,35,35,140,35,5,175,70,38
+.byte	198,198,63,198,126,249,145,184,198,198,63,198,126,249,145,184
+.byte	232,232,135,232,19,111,205,251,232,232,135,232,19,111,205,251
+.byte	135,135,38,135,76,161,19,203,135,135,38,135,76,161,19,203
+.byte	184,184,218,184,169,98,109,17,184,184,218,184,169,98,109,17
+.byte	1,1,4,1,8,5,2,9,1,1,4,1,8,5,2,9
+.byte	79,79,33,79,66,110,158,13,79,79,33,79,66,110,158,13
+.byte	54,54,216,54,173,238,108,155,54,54,216,54,173,238,108,155
+.byte	166,166,162,166,89,4,81,255,166,166,162,166,89,4,81,255
+.byte	210,210,111,210,222,189,185,12,210,210,111,210,222,189,185,12
+.byte	245,245,243,245,251,6,247,14,245,245,243,245,251,6,247,14
+.byte	121,121,249,121,239,128,242,150,121,121,249,121,239,128,242,150
+.byte	111,111,161,111,95,206,222,48,111,111,161,111,95,206,222,48
+.byte	145,145,126,145,252,239,63,109,145,145,126,145,252,239,63,109
+.byte	82,82,85,82,170,7,164,248,82,82,85,82,170,7,164,248
+.byte	96,96,157,96,39,253,192,71,96,96,157,96,39,253,192,71
+.byte	188,188,202,188,137,118,101,53,188,188,202,188,137,118,101,53
+.byte	155,155,86,155,172,205,43,55,155,155,86,155,172,205,43,55
+.byte	142,142,2,142,4,140,1,138,142,142,2,142,4,140,1,138
+.byte	163,163,182,163,113,21,91,210,163,163,182,163,113,21,91,210
+.byte	12,12,48,12,96,60,24,108,12,12,48,12,96,60,24,108
+.byte	123,123,241,123,255,138,246,132,123,123,241,123,255,138,246,132
+.byte	53,53,212,53,181,225,106,128,53,53,212,53,181,225,106,128
+.byte	29,29,116,29,232,105,58,245,29,29,116,29,232,105,58,245
+.byte	224,224,167,224,83,71,221,179,224,224,167,224,83,71,221,179
+.byte	215,215,123,215,246,172,179,33,215,215,123,215,246,172,179,33
+.byte	194,194,47,194,94,237,153,156,194,194,47,194,94,237,153,156
+.byte	46,46,184,46,109,150,92,67,46,46,184,46,109,150,92,67
+.byte	75,75,49,75,98,122,150,41,75,75,49,75,98,122,150,41
+.byte	254,254,223,254,163,33,225,93,254,254,223,254,163,33,225,93
+.byte	87,87,65,87,130,22,174,213,87,87,65,87,130,22,174,213
+.byte	21,21,84,21,168,65,42,189,21,21,84,21,168,65,42,189
+.byte	119,119,193,119,159,182,238,232,119,119,193,119,159,182,238,232
+.byte	55,55,220,55,165,235,110,146,55,55,220,55,165,235,110,146
+.byte	229,229,179,229,123,86,215,158,229,229,179,229,123,86,215,158
+.byte	159,159,70,159,140,217,35,19,159,159,70,159,140,217,35,19
+.byte	240,240,231,240,211,23,253,35,240,240,231,240,211,23,253,35
+.byte	74,74,53,74,106,127,148,32,74,74,53,74,106,127,148,32
+.byte	218,218,79,218,158,149,169,68,218,218,79,218,158,149,169,68
+.byte	88,88,125,88,250,37,176,162,88,88,125,88,250,37,176,162
+.byte	201,201,3,201,6,202,143,207,201,201,3,201,6,202,143,207
+.byte	41,41,164,41,85,141,82,124,41,41,164,41,85,141,82,124
+.byte	10,10,40,10,80,34,20,90,10,10,40,10,80,34,20,90
+.byte	177,177,254,177,225,79,127,80,177,177,254,177,225,79,127,80
+.byte	160,160,186,160,105,26,93,201,160,160,186,160,105,26,93,201
+.byte	107,107,177,107,127,218,214,20,107,107,177,107,127,218,214,20
+.byte	133,133,46,133,92,171,23,217,133,133,46,133,92,171,23,217
+.byte	189,189,206,189,129,115,103,60,189,189,206,189,129,115,103,60
+.byte	93,93,105,93,210,52,186,143,93,93,105,93,210,52,186,143
+.byte	16,16,64,16,128,80,32,144,16,16,64,16,128,80,32,144
+.byte	244,244,247,244,243,3,245,7,244,244,247,244,243,3,245,7
+.byte	203,203,11,203,22,192,139,221,203,203,11,203,22,192,139,221
+.byte	62,62,248,62,237,198,124,211,62,62,248,62,237,198,124,211
+.byte	5,5,20,5,40,17,10,45,5,5,20,5,40,17,10,45
+.byte	103,103,129,103,31,230,206,120,103,103,129,103,31,230,206,120
+.byte	228,228,183,228,115,83,213,151,228,228,183,228,115,83,213,151
+.byte	39,39,156,39,37,187,78,2,39,39,156,39,37,187,78,2
+.byte	65,65,25,65,50,88,130,115,65,65,25,65,50,88,130,115
+.byte	139,139,22,139,44,157,11,167,139,139,22,139,44,157,11,167
+.byte	167,167,166,167,81,1,83,246,167,167,166,167,81,1,83,246
+.byte	125,125,233,125,207,148,250,178,125,125,233,125,207,148,250,178
+.byte	149,149,110,149,220,251,55,73,149,149,110,149,220,251,55,73
+.byte	216,216,71,216,142,159,173,86,216,216,71,216,142,159,173,86
+.byte	251,251,203,251,139,48,235,112,251,251,203,251,139,48,235,112
+.byte	238,238,159,238,35,113,193,205,238,238,159,238,35,113,193,205
+.byte	124,124,237,124,199,145,248,187,124,124,237,124,199,145,248,187
+.byte	102,102,133,102,23,227,204,113,102,102,133,102,23,227,204,113
+.byte	221,221,83,221,166,142,167,123,221,221,83,221,166,142,167,123
+.byte	23,23,92,23,184,75,46,175,23,23,92,23,184,75,46,175
+.byte	71,71,1,71,2,70,142,69,71,71,1,71,2,70,142,69
+.byte	158,158,66,158,132,220,33,26,158,158,66,158,132,220,33,26
+.byte	202,202,15,202,30,197,137,212,202,202,15,202,30,197,137,212
+.byte	45,45,180,45,117,153,90,88,45,45,180,45,117,153,90,88
+.byte	191,191,198,191,145,121,99,46,191,191,198,191,145,121,99,46
+.byte	7,7,28,7,56,27,14,63,7,7,28,7,56,27,14,63
+.byte	173,173,142,173,1,35,71,172,173,173,142,173,1,35,71,172
+.byte	90,90,117,90,234,47,180,176,90,90,117,90,234,47,180,176
+.byte	131,131,54,131,108,181,27,239,131,131,54,131,108,181,27,239
+.byte	51,51,204,51,133,255,102,182,51,51,204,51,133,255,102,182
+.byte	99,99,145,99,63,242,198,92,99,99,145,99,63,242,198,92
+.byte	2,2,8,2,16,10,4,18,2,2,8,2,16,10,4,18
+.byte	170,170,146,170,57,56,73,147,170,170,146,170,57,56,73,147
+.byte	113,113,217,113,175,168,226,222,113,113,217,113,175,168,226,222
+.byte	200,200,7,200,14,207,141,198,200,200,7,200,14,207,141,198
+.byte	25,25,100,25,200,125,50,209,25,25,100,25,200,125,50,209
+.byte	73,73,57,73,114,112,146,59,73,73,57,73,114,112,146,59
+.byte	217,217,67,217,134,154,175,95,217,217,67,217,134,154,175,95
+.byte	242,242,239,242,195,29,249,49,242,242,239,242,195,29,249,49
+.byte	227,227,171,227,75,72,219,168,227,227,171,227,75,72,219,168
+.byte	91,91,113,91,226,42,182,185,91,91,113,91,226,42,182,185
+.byte	136,136,26,136,52,146,13,188,136,136,26,136,52,146,13,188
+.byte	154,154,82,154,164,200,41,62,154,154,82,154,164,200,41,62
+.byte	38,38,152,38,45,190,76,11,38,38,152,38,45,190,76,11
+.byte	50,50,200,50,141,250,100,191,50,50,200,50,141,250,100,191
+.byte	176,176,250,176,233,74,125,89,176,176,250,176,233,74,125,89
+.byte	233,233,131,233,27,106,207,242,233,233,131,233,27,106,207,242
+.byte	15,15,60,15,120,51,30,119,15,15,60,15,120,51,30,119
+.byte	213,213,115,213,230,166,183,51,213,213,115,213,230,166,183,51
+.byte	128,128,58,128,116,186,29,244,128,128,58,128,116,186,29,244
+.byte	190,190,194,190,153,124,97,39,190,190,194,190,153,124,97,39
+.byte	205,205,19,205,38,222,135,235,205,205,19,205,38,222,135,235
+.byte	52,52,208,52,189,228,104,137,52,52,208,52,189,228,104,137
+.byte	72,72,61,72,122,117,144,50,72,72,61,72,122,117,144,50
+.byte	255,255,219,255,171,36,227,84,255,255,219,255,171,36,227,84
+.byte	122,122,245,122,247,143,244,141,122,122,245,122,247,143,244,141
+.byte	144,144,122,144,244,234,61,100,144,144,122,144,244,234,61,100
+.byte	95,95,97,95,194,62,190,157,95,95,97,95,194,62,190,157
+.byte	32,32,128,32,29,160,64,61,32,32,128,32,29,160,64,61
+.byte	104,104,189,104,103,213,208,15,104,104,189,104,103,213,208,15
+.byte	26,26,104,26,208,114,52,202,26,26,104,26,208,114,52,202
+.byte	174,174,130,174,25,44,65,183,174,174,130,174,25,44,65,183
+.byte	180,180,234,180,201,94,117,125,180,180,234,180,201,94,117,125
+.byte	84,84,77,84,154,25,168,206,84,84,77,84,154,25,168,206
+.byte	147,147,118,147,236,229,59,127,147,147,118,147,236,229,59,127
+.byte	34,34,136,34,13,170,68,47,34,34,136,34,13,170,68,47
+.byte	100,100,141,100,7,233,200,99,100,100,141,100,7,233,200,99
+.byte	241,241,227,241,219,18,255,42,241,241,227,241,219,18,255,42
+.byte	115,115,209,115,191,162,230,204,115,115,209,115,191,162,230,204
+.byte	18,18,72,18,144,90,36,130,18,18,72,18,144,90,36,130
+.byte	64,64,29,64,58,93,128,122,64,64,29,64,58,93,128,122
+.byte	8,8,32,8,64,40,16,72,8,8,32,8,64,40,16,72
+.byte	195,195,43,195,86,232,155,149,195,195,43,195,86,232,155,149
+.byte	236,236,151,236,51,123,197,223,236,236,151,236,51,123,197,223
+.byte	219,219,75,219,150,144,171,77,219,219,75,219,150,144,171,77
+.byte	161,161,190,161,97,31,95,192,161,161,190,161,97,31,95,192
+.byte	141,141,14,141,28,131,7,145,141,141,14,141,28,131,7,145
+.byte	61,61,244,61,245,201,122,200,61,61,244,61,245,201,122,200
+.byte	151,151,102,151,204,241,51,91,151,151,102,151,204,241,51,91
+.byte	0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+.byte	207,207,27,207,54,212,131,249,207,207,27,207,54,212,131,249
+.byte	43,43,172,43,69,135,86,110,43,43,172,43,69,135,86,110
+.byte	118,118,197,118,151,179,236,225,118,118,197,118,151,179,236,225
+.byte	130,130,50,130,100,176,25,230,130,130,50,130,100,176,25,230
+.byte	214,214,127,214,254,169,177,40,214,214,127,214,254,169,177,40
+.byte	27,27,108,27,216,119,54,195,27,27,108,27,216,119,54,195
+.byte	181,181,238,181,193,91,119,116,181,181,238,181,193,91,119,116
+.byte	175,175,134,175,17,41,67,190,175,175,134,175,17,41,67,190
+.byte	106,106,181,106,119,223,212,29,106,106,181,106,119,223,212,29
+.byte	80,80,93,80,186,13,160,234,80,80,93,80,186,13,160,234
+.byte	69,69,9,69,18,76,138,87,69,69,9,69,18,76,138,87
+.byte	243,243,235,243,203,24,251,56,243,243,235,243,203,24,251,56
+.byte	48,48,192,48,157,240,96,173,48,48,192,48,157,240,96,173
+.byte	239,239,155,239,43,116,195,196,239,239,155,239,43,116,195,196
+.byte	63,63,252,63,229,195,126,218,63,63,252,63,229,195,126,218
+.byte	85,85,73,85,146,28,170,199,85,85,73,85,146,28,170,199
+.byte	162,162,178,162,121,16,89,219,162,162,178,162,121,16,89,219
+.byte	234,234,143,234,3,101,201,233,234,234,143,234,3,101,201,233
+.byte	101,101,137,101,15,236,202,106,101,101,137,101,15,236,202,106
+.byte	186,186,210,186,185,104,105,3,186,186,210,186,185,104,105,3
+.byte	47,47,188,47,101,147,94,74,47,47,188,47,101,147,94,74
+.byte	192,192,39,192,78,231,157,142,192,192,39,192,78,231,157,142
+.byte	222,222,95,222,190,129,161,96,222,222,95,222,190,129,161,96
+.byte	28,28,112,28,224,108,56,252,28,28,112,28,224,108,56,252
+.byte	253,253,211,253,187,46,231,70,253,253,211,253,187,46,231,70
+.byte	77,77,41,77,82,100,154,31,77,77,41,77,82,100,154,31
+.byte	146,146,114,146,228,224,57,118,146,146,114,146,228,224,57,118
+.byte	117,117,201,117,143,188,234,250,117,117,201,117,143,188,234,250
+.byte	6,6,24,6,48,30,12,54,6,6,24,6,48,30,12,54
+.byte	138,138,18,138,36,152,9,174,138,138,18,138,36,152,9,174
+.byte	178,178,242,178,249,64,121,75,178,178,242,178,249,64,121,75
+.byte	230,230,191,230,99,89,209,133,230,230,191,230,99,89,209,133
+.byte	14,14,56,14,112,54,28,126,14,14,56,14,112,54,28,126
+.byte	31,31,124,31,248,99,62,231,31,31,124,31,248,99,62,231
+.byte	98,98,149,98,55,247,196,85,98,98,149,98,55,247,196,85
+.byte	212,212,119,212,238,163,181,58,212,212,119,212,238,163,181,58
+.byte	168,168,154,168,41,50,77,129,168,168,154,168,41,50,77,129
+.byte	150,150,98,150,196,244,49,82,150,150,98,150,196,244,49,82
+.byte	249,249,195,249,155,58,239,98,249,249,195,249,155,58,239,98
+.byte	197,197,51,197,102,246,151,163,197,197,51,197,102,246,151,163
+.byte	37,37,148,37,53,177,74,16,37,37,148,37,53,177,74,16
+.byte	89,89,121,89,242,32,178,171,89,89,121,89,242,32,178,171
+.byte	132,132,42,132,84,174,21,208,132,132,42,132,84,174,21,208
+.byte	114,114,213,114,183,167,228,197,114,114,213,114,183,167,228,197
+.byte	57,57,228,57,213,221,114,236,57,57,228,57,213,221,114,236
+.byte	76,76,45,76,90,97,152,22,76,76,45,76,90,97,152,22
+.byte	94,94,101,94,202,59,188,148,94,94,101,94,202,59,188,148
+.byte	120,120,253,120,231,133,240,159,120,120,253,120,231,133,240,159
+.byte	56,56,224,56,221,216,112,229,56,56,224,56,221,216,112,229
+.byte	140,140,10,140,20,134,5,152,140,140,10,140,20,134,5,152
+.byte	209,209,99,209,198,178,191,23,209,209,99,209,198,178,191,23
+.byte	165,165,174,165,65,11,87,228,165,165,174,165,65,11,87,228
+.byte	226,226,175,226,67,77,217,161,226,226,175,226,67,77,217,161
+.byte	97,97,153,97,47,248,194,78,97,97,153,97,47,248,194,78
+.byte	179,179,246,179,241,69,123,66,179,179,246,179,241,69,123,66
+.byte	33,33,132,33,21,165,66,52,33,33,132,33,21,165,66,52
+.byte	156,156,74,156,148,214,37,8,156,156,74,156,148,214,37,8
+.byte	30,30,120,30,240,102,60,238,30,30,120,30,240,102,60,238
+.byte	67,67,17,67,34,82,134,97,67,67,17,67,34,82,134,97
+.byte	199,199,59,199,118,252,147,177,199,199,59,199,118,252,147,177
+.byte	252,252,215,252,179,43,229,79,252,252,215,252,179,43,229,79
+.byte	4,4,16,4,32,20,8,36,4,4,16,4,32,20,8,36
+.byte	81,81,89,81,178,8,162,227,81,81,89,81,178,8,162,227
+.byte	153,153,94,153,188,199,47,37,153,153,94,153,188,199,47,37
+.byte	109,109,169,109,79,196,218,34,109,109,169,109,79,196,218,34
+.byte	13,13,52,13,104,57,26,101,13,13,52,13,104,57,26,101
+.byte	250,250,207,250,131,53,233,121,250,250,207,250,131,53,233,121
+.byte	223,223,91,223,182,132,163,105,223,223,91,223,182,132,163,105
+.byte	126,126,229,126,215,155,252,169,126,126,229,126,215,155,252,169
+.byte	36,36,144,36,61,180,72,25,36,36,144,36,61,180,72,25
+.byte	59,59,236,59,197,215,118,254,59,59,236,59,197,215,118,254
+.byte	171,171,150,171,49,61,75,154,171,171,150,171,49,61,75,154
+.byte	206,206,31,206,62,209,129,240,206,206,31,206,62,209,129,240
+.byte	17,17,68,17,136,85,34,153,17,17,68,17,136,85,34,153
+.byte	143,143,6,143,12,137,3,131,143,143,6,143,12,137,3,131
+.byte	78,78,37,78,74,107,156,4,78,78,37,78,74,107,156,4
+.byte	183,183,230,183,209,81,115,102,183,183,230,183,209,81,115,102
+.byte	235,235,139,235,11,96,203,224,235,235,139,235,11,96,203,224
+.byte	60,60,240,60,253,204,120,193,60,60,240,60,253,204,120,193
+.byte	129,129,62,129,124,191,31,253,129,129,62,129,124,191,31,253
+.byte	148,148,106,148,212,254,53,64,148,148,106,148,212,254,53,64
+.byte	247,247,251,247,235,12,243,28,247,247,251,247,235,12,243,28
+.byte	185,185,222,185,161,103,111,24,185,185,222,185,161,103,111,24
+.byte	19,19,76,19,152,95,38,139,19,19,76,19,152,95,38,139
+.byte	44,44,176,44,125,156,88,81,44,44,176,44,125,156,88,81
+.byte	211,211,107,211,214,184,187,5,211,211,107,211,214,184,187,5
+.byte	231,231,187,231,107,92,211,140,231,231,187,231,107,92,211,140
+.byte	110,110,165,110,87,203,220,57,110,110,165,110,87,203,220,57
+.byte	196,196,55,196,110,243,149,170,196,196,55,196,110,243,149,170
+.byte	3,3,12,3,24,15,6,27,3,3,12,3,24,15,6,27
+.byte	86,86,69,86,138,19,172,220,86,86,69,86,138,19,172,220
+.byte	68,68,13,68,26,73,136,94,68,68,13,68,26,73,136,94
+.byte	127,127,225,127,223,158,254,160,127,127,225,127,223,158,254,160
+.byte	169,169,158,169,33,55,79,136,169,169,158,169,33,55,79,136
+.byte	42,42,168,42,77,130,84,103,42,42,168,42,77,130,84,103
+.byte	187,187,214,187,177,109,107,10,187,187,214,187,177,109,107,10
+.byte	193,193,35,193,70,226,159,135,193,193,35,193,70,226,159,135
+.byte	83,83,81,83,162,2,166,241,83,83,81,83,162,2,166,241
+.byte	220,220,87,220,174,139,165,114,220,220,87,220,174,139,165,114
+.byte	11,11,44,11,88,39,22,83,11,11,44,11,88,39,22,83
+.byte	157,157,78,157,156,211,39,1,157,157,78,157,156,211,39,1
+.byte	108,108,173,108,71,193,216,43,108,108,173,108,71,193,216,43
+.byte	49,49,196,49,149,245,98,164,49,49,196,49,149,245,98,164
+.byte	116,116,205,116,135,185,232,243,116,116,205,116,135,185,232,243
+.byte	246,246,255,246,227,9,241,21,246,246,255,246,227,9,241,21
+.byte	70,70,5,70,10,67,140,76,70,70,5,70,10,67,140,76
+.byte	172,172,138,172,9,38,69,165,172,172,138,172,9,38,69,165
+.byte	137,137,30,137,60,151,15,181,137,137,30,137,60,151,15,181
+.byte	20,20,80,20,160,68,40,180,20,20,80,20,160,68,40,180
+.byte	225,225,163,225,91,66,223,186,225,225,163,225,91,66,223,186
+.byte	22,22,88,22,176,78,44,166,22,22,88,22,176,78,44,166
+.byte	58,58,232,58,205,210,116,247,58,58,232,58,205,210,116,247
+.byte	105,105,185,105,111,208,210,6,105,105,185,105,111,208,210,6
+.byte	9,9,36,9,72,45,18,65,9,9,36,9,72,45,18,65
+.byte	112,112,221,112,167,173,224,215,112,112,221,112,167,173,224,215
+.byte	182,182,226,182,217,84,113,111,182,182,226,182,217,84,113,111
+.byte	208,208,103,208,206,183,189,30,208,208,103,208,206,183,189,30
+.byte	237,237,147,237,59,126,199,214,237,237,147,237,59,126,199,214
+.byte	204,204,23,204,46,219,133,226,204,204,23,204,46,219,133,226
+.byte	66,66,21,66,42,87,132,104,66,66,21,66,42,87,132,104
+.byte	152,152,90,152,180,194,45,44,152,152,90,152,180,194,45,44
+.byte	164,164,170,164,73,14,85,237,164,164,170,164,73,14,85,237
+.byte	40,40,160,40,93,136,80,117,40,40,160,40,93,136,80,117
+.byte	92,92,109,92,218,49,184,134,92,92,109,92,218,49,184,134
+.byte	248,248,199,248,147,63,237,107,248,248,199,248,147,63,237,107
+.byte	134,134,34,134,68,164,17,194,134,134,34,134,68,164,17,194
+.byte	24,35,198,232,135,184,1,79
+.byte	54,166,210,245,121,111,145,82
+.byte	96,188,155,142,163,12,123,53
+.byte	29,224,215,194,46,75,254,87
+.byte	21,119,55,229,159,240,74,218
+.byte	88,201,41,10,177,160,107,133
+.byte	189,93,16,244,203,62,5,103
+.byte	228,39,65,139,167,125,149,216
+.byte	251,238,124,102,221,23,71,158
+.byte	202,45,191,7,173,90,131,51

deps/openssl/asm/x64-elf-gas/x86_64cpuid.s

@@ -0,0 +1,194 @@
+
+.section	.init
+	call	OPENSSL_cpuid_setup
+
+.text
+
+
+.globl	OPENSSL_atomic_add
+.type	OPENSSL_atomic_add,@function
+.align	16
+OPENSSL_atomic_add:
+	movl	(%rdi),%eax
+.Lspin:	leaq	(%rsi,%rax,1),%r8
+.byte	0xf0
+
+	cmpxchgl	%r8d,(%rdi)
+	jne	.Lspin
+	movl	%r8d,%eax
+.byte	0x48,0x98
+
+	.byte	0xf3,0xc3
+.size	OPENSSL_atomic_add,.-OPENSSL_atomic_add
+
+.globl	OPENSSL_rdtsc
+.type	OPENSSL_rdtsc,@function
+.align	16
+OPENSSL_rdtsc:
+	rdtsc
+	shlq	$32,%rdx
+	orq	%rdx,%rax
+	.byte	0xf3,0xc3
+.size	OPENSSL_rdtsc,.-OPENSSL_rdtsc
+
+.globl	OPENSSL_ia32_cpuid
+.type	OPENSSL_ia32_cpuid,@function
+.align	16
+OPENSSL_ia32_cpuid:
+	movq	%rbx,%r8
+
+	xorl	%eax,%eax
+	cpuid
+	movl	%eax,%r11d
+
+	xorl	%eax,%eax
+	cmpl	$1970169159,%ebx
+	setne	%al
+	movl	%eax,%r9d
+	cmpl	$1231384169,%edx
+	setne	%al
+	orl	%eax,%r9d
+	cmpl	$1818588270,%ecx
+	setne	%al
+	orl	%eax,%r9d
+	jz	.Lintel
+
+	cmpl	$1752462657,%ebx
+	setne	%al
+	movl	%eax,%r10d
+	cmpl	$1769238117,%edx
+	setne	%al
+	orl	%eax,%r10d
+	cmpl	$1145913699,%ecx
+	setne	%al
+	orl	%eax,%r10d
+	jnz	.Lintel
+
+
+	movl	$2147483648,%eax
+	cpuid
+	cmpl	$2147483656,%eax
+	jb	.Lintel
+
+	movl	$2147483656,%eax
+	cpuid
+	movzbq	%cl,%r10
+	incq	%r10
+
+	movl	$1,%eax
+	cpuid
+	btl	$28,%edx
+	jnc	.Ldone
+	shrl	$16,%ebx
+	cmpb	%r10b,%bl
+	ja	.Ldone
+	andl	$4026531839,%edx
+	jmp	.Ldone
+
+.Lintel:
+	cmpl	$4,%r11d
+	movl	$-1,%r10d
+	jb	.Lnocacheinfo
+
+	movl	$4,%eax
+	movl	$0,%ecx
+	cpuid
+	movl	%eax,%r10d
+	shrl	$14,%r10d
+	andl	$4095,%r10d
+
+.Lnocacheinfo:
+	movl	$1,%eax
+	cpuid
+	cmpl	$0,%r9d
+	jne	.Lnotintel
+	orl	$1048576,%edx
+	andb	$15,%ah
+	cmpb	$15,%ah
+	je	.Lnotintel
+	orl	$1073741824,%edx
+.Lnotintel:
+	btl	$28,%edx
+	jnc	.Ldone
+	andl	$4026531839,%edx
+	cmpl	$0,%r10d
+	je	.Ldone
+
+	orl	$268435456,%edx
+	shrl	$16,%ebx
+	cmpb	$1,%bl
+	ja	.Ldone
+	andl	$4026531839,%edx
+.Ldone:
+	shlq	$32,%rcx
+	movl	%edx,%eax
+	movq	%r8,%rbx
+	orq	%rcx,%rax
+	.byte	0xf3,0xc3
+.size	OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid
+
+.globl	OPENSSL_cleanse
+.type	OPENSSL_cleanse,@function
+.align	16
+OPENSSL_cleanse:
+	xorq	%rax,%rax
+	cmpq	$15,%rsi
+	jae	.Lot
+	cmpq	$0,%rsi
+	je	.Lret
+.Little:
+	movb	%al,(%rdi)
+	subq	$1,%rsi
+	leaq	1(%rdi),%rdi
+	jnz	.Little
+.Lret:
+	.byte	0xf3,0xc3
+.align	16
+.Lot:
+	testq	$7,%rdi
+	jz	.Laligned
+	movb	%al,(%rdi)
+	leaq	-1(%rsi),%rsi
+	leaq	1(%rdi),%rdi
+	jmp	.Lot
+.Laligned:
+	movq	%rax,(%rdi)
+	leaq	-8(%rsi),%rsi
+	testq	$-8,%rsi
+	leaq	8(%rdi),%rdi
+	jnz	.Laligned
+	cmpq	$0,%rsi
+	jne	.Little
+	.byte	0xf3,0xc3
+.size	OPENSSL_cleanse,.-OPENSSL_cleanse
+.globl	OPENSSL_wipe_cpu
+.type	OPENSSL_wipe_cpu,@function
+.align	16
+OPENSSL_wipe_cpu:
+	pxor	%xmm0,%xmm0
+	pxor	%xmm1,%xmm1
+	pxor	%xmm2,%xmm2
+	pxor	%xmm3,%xmm3
+	pxor	%xmm4,%xmm4
+	pxor	%xmm5,%xmm5
+	pxor	%xmm6,%xmm6
+	pxor	%xmm7,%xmm7
+	pxor	%xmm8,%xmm8
+	pxor	%xmm9,%xmm9
+	pxor	%xmm10,%xmm10
+	pxor	%xmm11,%xmm11
+	pxor	%xmm12,%xmm12
+	pxor	%xmm13,%xmm13
+	pxor	%xmm14,%xmm14
+	pxor	%xmm15,%xmm15
+	xorq	%rcx,%rcx
+	xorq	%rdx,%rdx
+	xorq	%rsi,%rsi
+	xorq	%rdi,%rdi
+	xorq	%r8,%r8
+	xorq	%r9,%r9
+	xorq	%r10,%r10
+	xorq	%r11,%r11
+	leaq	8(%rsp),%rax
+	.byte	0xf3,0xc3
+.size	OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu

deps/openssl/asm/x64-macosx-gas/bn/x86_64-mont.s

@@ -0,0 +1,172 @@
+.text
+
+
+.globl	_bn_mul_mont
+
+.p2align	4
+_bn_mul_mont:
+	pushq	%rbx
+	pushq	%rbp
+	pushq	%r12
+	pushq	%r13
+	pushq	%r14
+	pushq	%r15
+
+	movl	%r9d,%r9d
+	leaq	2(%r9),%r10
+	movq	%rsp,%r11
+	negq	%r10
+	leaq	(%rsp,%r10,8),%rsp
+	andq	$-1024,%rsp
+
+	movq	%r11,8(%rsp,%r9,8)
+L$prologue:
+	movq	%rdx,%r12
+
+	movq	(%r8),%r8
+
+	xorq	%r14,%r14
+	xorq	%r15,%r15
+
+	movq	(%r12),%rbx
+	movq	(%rsi),%rax
+	mulq	%rbx
+	movq	%rax,%r10
+	movq	%rdx,%r11
+
+	imulq	%r8,%rax
+	movq	%rax,%rbp
+
+	mulq	(%rcx)
+	addq	%r10,%rax
+	adcq	$0,%rdx
+	movq	%rdx,%r13
+
+	leaq	1(%r15),%r15
+L$1st:
+	movq	(%rsi,%r15,8),%rax
+	mulq	%rbx
+	addq	%r11,%rax
+	adcq	$0,%rdx
+	movq	%rax,%r10
+	movq	(%rcx,%r15,8),%rax
+	movq	%rdx,%r11
+
+	mulq	%rbp
+	addq	%r13,%rax
+	leaq	1(%r15),%r15
+	adcq	$0,%rdx
+	addq	%r10,%rax
+	adcq	$0,%rdx
+	movq	%rax,-16(%rsp,%r15,8)
+	cmpq	%r9,%r15
+	movq	%rdx,%r13
+	jl	L$1st
+
+	xorq	%rdx,%rdx
+	addq	%r11,%r13
+	adcq	$0,%rdx
+	movq	%r13,-8(%rsp,%r9,8)
+	movq	%rdx,(%rsp,%r9,8)
+
+	leaq	1(%r14),%r14
+.p2align	2
+L$outer:
+	xorq	%r15,%r15
+
+	movq	(%r12,%r14,8),%rbx
+	movq	(%rsi),%rax
+	mulq	%rbx
+	addq	(%rsp),%rax
+	adcq	$0,%rdx
+	movq	%rax,%r10
+	movq	%rdx,%r11
+
+	imulq	%r8,%rax
+	movq	%rax,%rbp
+
+	mulq	(%rcx,%r15,8)
+	addq	%r10,%rax
+	movq	8(%rsp),%r10
+	adcq	$0,%rdx
+	movq	%rdx,%r13
+
+	leaq	1(%r15),%r15
+.p2align	2
+L$inner:
+	movq	(%rsi,%r15,8),%rax
+	mulq	%rbx
+	addq	%r11,%rax
+	adcq	$0,%rdx
+	addq	%rax,%r10
+	movq	(%rcx,%r15,8),%rax
+	adcq	$0,%rdx
+	movq	%rdx,%r11
+
+	mulq	%rbp
+	addq	%r13,%rax
+	leaq	1(%r15),%r15
+	adcq	$0,%rdx
+	addq	%r10,%rax
+	adcq	$0,%rdx
+	movq	(%rsp,%r15,8),%r10
+	cmpq	%r9,%r15
+	movq	%rax,-16(%rsp,%r15,8)
+	movq	%rdx,%r13
+	jl	L$inner
+
+	xorq	%rdx,%rdx
+	addq	%r11,%r13
+	adcq	$0,%rdx
+	addq	%r10,%r13
+	adcq	$0,%rdx
+	movq	%r13,-8(%rsp,%r9,8)
+	movq	%rdx,(%rsp,%r9,8)
+
+	leaq	1(%r14),%r14
+	cmpq	%r9,%r14
+	jl	L$outer
+
+	leaq	(%rsp),%rsi
+	leaq	-1(%r9),%r15
+
+	movq	(%rsi),%rax
+	xorq	%r14,%r14
+	jmp	L$sub
+.p2align	4
+L$sub:	sbbq	(%rcx,%r14,8),%rax
+	movq	%rax,(%rdi,%r14,8)
+	decq	%r15
+	movq	8(%rsi,%r14,8),%rax
+	leaq	1(%r14),%r14
+	jge	L$sub
+
+	sbbq	$0,%rax
+	andq	%rax,%rsi
+	notq	%rax
+	movq	%rdi,%rcx
+	andq	%rax,%rcx
+	leaq	-1(%r9),%r15
+	orq	%rcx,%rsi
+.p2align	4
+L$copy:
+	movq	(%rsi,%r15,8),%rax
+	movq	%rax,(%rdi,%r15,8)
+	movq	%r14,(%rsp,%r15,8)
+	decq	%r15
+	jge	L$copy
+
+	movq	8(%rsp,%r9,8),%rsi
+	movq	$1,%rax
+	movq	(%rsi),%r15
+	movq	8(%rsi),%r14
+	movq	16(%rsi),%r13
+	movq	24(%rsi),%r12
+	movq	32(%rsi),%rbp
+	movq	40(%rsi),%rbx
+	leaq	48(%rsi),%rsp
+L$epilogue:
+	.byte	0xf3,0xc3
+
+.byte	77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105,112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.p2align	4

deps/openssl/asm/x64-macosx-gas/md5/md5-x86_64.s

@@ -0,0 +1,671 @@
+.text
+
+.p2align	4
+
+.globl	_md5_block_asm_data_order
+
+_md5_block_asm_data_order:
+	pushq	%rbp
+	pushq	%rbx
+	pushq	%r12
+	pushq	%r14
+	pushq	%r15
+L$prologue:
+
+
+
+
+	movq	%rdi,%rbp
+	shlq	$6,%rdx
+	leaq	(%rsi,%rdx,1),%rdi
+	movl	0(%rbp),%eax
+	movl	4(%rbp),%ebx
+	movl	8(%rbp),%ecx
+	movl	12(%rbp),%edx
+
+
+
+
+
+
+
+	cmpq	%rdi,%rsi
+	je	L$end
+
+
+
+L$loop:
+	movl	%eax,%r8d
+	movl	%ebx,%r9d
+	movl	%ecx,%r14d
+	movl	%edx,%r15d
+	movl	0(%rsi),%r10d
+	movl	%edx,%r11d
+	xorl	%ecx,%r11d
+	leal	-680876936(%rax,%r10,1),%eax
+	andl	%ebx,%r11d
+	xorl	%edx,%r11d
+	movl	4(%rsi),%r10d
+	addl	%r11d,%eax
+	roll	$7,%eax
+	movl	%ecx,%r11d
+	addl	%ebx,%eax
+	xorl	%ebx,%r11d
+	leal	-389564586(%rdx,%r10,1),%edx
+	andl	%eax,%r11d
+	xorl	%ecx,%r11d
+	movl	8(%rsi),%r10d
+	addl	%r11d,%edx
+	roll	$12,%edx
+	movl	%ebx,%r11d
+	addl	%eax,%edx
+	xorl	%eax,%r11d
+	leal	606105819(%rcx,%r10,1),%ecx
+	andl	%edx,%r11d
+	xorl	%ebx,%r11d
+	movl	12(%rsi),%r10d
+	addl	%r11d,%ecx
+	roll	$17,%ecx
+	movl	%eax,%r11d
+	addl	%edx,%ecx
+	xorl	%edx,%r11d
+	leal	-1044525330(%rbx,%r10,1),%ebx
+	andl	%ecx,%r11d
+	xorl	%eax,%r11d
+	movl	16(%rsi),%r10d
+	addl	%r11d,%ebx
+	roll	$22,%ebx
+	movl	%edx,%r11d
+	addl	%ecx,%ebx
+	xorl	%ecx,%r11d
+	leal	-176418897(%rax,%r10,1),%eax
+	andl	%ebx,%r11d
+	xorl	%edx,%r11d
+	movl	20(%rsi),%r10d
+	addl	%r11d,%eax
+	roll	$7,%eax
+	movl	%ecx,%r11d
+	addl	%ebx,%eax
+	xorl	%ebx,%r11d
+	leal	1200080426(%rdx,%r10,1),%edx
+	andl	%eax,%r11d
+	xorl	%ecx,%r11d
+	movl	24(%rsi),%r10d
+	addl	%r11d,%edx
+	roll	$12,%edx
+	movl	%ebx,%r11d
+	addl	%eax,%edx
+	xorl	%eax,%r11d
+	leal	-1473231341(%rcx,%r10,1),%ecx
+	andl	%edx,%r11d
+	xorl	%ebx,%r11d
+	movl	28(%rsi),%r10d
+	addl	%r11d,%ecx
+	roll	$17,%ecx
+	movl	%eax,%r11d
+	addl	%edx,%ecx
+	xorl	%edx,%r11d
+	leal	-45705983(%rbx,%r10,1),%ebx
+	andl	%ecx,%r11d
+	xorl	%eax,%r11d
+	movl	32(%rsi),%r10d
+	addl	%r11d,%ebx
+	roll	$22,%ebx
+	movl	%edx,%r11d
+	addl	%ecx,%ebx
+	xorl	%ecx,%r11d
+	leal	1770035416(%rax,%r10,1),%eax
+	andl	%ebx,%r11d
+	xorl	%edx,%r11d
+	movl	36(%rsi),%r10d
+	addl	%r11d,%eax
+	roll	$7,%eax
+	movl	%ecx,%r11d
+	addl	%ebx,%eax
+	xorl	%ebx,%r11d
+	leal	-1958414417(%rdx,%r10,1),%edx
+	andl	%eax,%r11d
+	xorl	%ecx,%r11d
+	movl	40(%rsi),%r10d
+	addl	%r11d,%edx
+	roll	$12,%edx
+	movl	%ebx,%r11d
+	addl	%eax,%edx
+	xorl	%eax,%r11d
+	leal	-42063(%rcx,%r10,1),%ecx
+	andl	%edx,%r11d
+	xorl	%ebx,%r11d
+	movl	44(%rsi),%r10d
+	addl	%r11d,%ecx
+	roll	$17,%ecx
+	movl	%eax,%r11d
+	addl	%edx,%ecx
+	xorl	%edx,%r11d
+	leal	-1990404162(%rbx,%r10,1),%ebx
+	andl	%ecx,%r11d
+	xorl	%eax,%r11d
+	movl	48(%rsi),%r10d
+	addl	%r11d,%ebx
+	roll	$22,%ebx
+	movl	%edx,%r11d
+	addl	%ecx,%ebx
+	xorl	%ecx,%r11d
+	leal	1804603682(%rax,%r10,1),%eax
+	andl	%ebx,%r11d
+	xorl	%edx,%r11d
+	movl	52(%rsi),%r10d
+	addl	%r11d,%eax
+	roll	$7,%eax
+	movl	%ecx,%r11d
+	addl	%ebx,%eax
+	xorl	%ebx,%r11d
+	leal	-40341101(%rdx,%r10,1),%edx
+	andl	%eax,%r11d
+	xorl	%ecx,%r11d
+	movl	56(%rsi),%r10d
+	addl	%r11d,%edx
+	roll	$12,%edx
+	movl	%ebx,%r11d
+	addl	%eax,%edx
+	xorl	%eax,%r11d
+	leal	-1502002290(%rcx,%r10,1),%ecx
+	andl	%edx,%r11d
+	xorl	%ebx,%r11d
+	movl	60(%rsi),%r10d
+	addl	%r11d,%ecx
+	roll	$17,%ecx
+	movl	%eax,%r11d
+	addl	%edx,%ecx
+	xorl	%edx,%r11d
+	leal	1236535329(%rbx,%r10,1),%ebx
+	andl	%ecx,%r11d
+	xorl	%eax,%r11d
+	movl	0(%rsi),%r10d
+	addl	%r11d,%ebx
+	roll	$22,%ebx
+	movl	%edx,%r11d
+	addl	%ecx,%ebx
+	movl	4(%rsi),%r10d
+	movl	%edx,%r11d
+	movl	%edx,%r12d
+	notl	%r11d
+	leal	-165796510(%rax,%r10,1),%eax
+	andl	%ebx,%r12d
+	andl	%ecx,%r11d
+	movl	24(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ecx,%r11d
+	addl	%r12d,%eax
+	movl	%ecx,%r12d
+	roll	$5,%eax
+	addl	%ebx,%eax
+	notl	%r11d
+	leal	-1069501632(%rdx,%r10,1),%edx
+	andl	%eax,%r12d
+	andl	%ebx,%r11d
+	movl	44(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ebx,%r11d
+	addl	%r12d,%edx
+	movl	%ebx,%r12d
+	roll	$9,%edx
+	addl	%eax,%edx
+	notl	%r11d
+	leal	643717713(%rcx,%r10,1),%ecx
+	andl	%edx,%r12d
+	andl	%eax,%r11d
+	movl	0(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%eax,%r11d
+	addl	%r12d,%ecx
+	movl	%eax,%r12d
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	notl	%r11d
+	leal	-373897302(%rbx,%r10,1),%ebx
+	andl	%ecx,%r12d
+	andl	%edx,%r11d
+	movl	20(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%edx,%r11d
+	addl	%r12d,%ebx
+	movl	%edx,%r12d
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	notl	%r11d
+	leal	-701558691(%rax,%r10,1),%eax
+	andl	%ebx,%r12d
+	andl	%ecx,%r11d
+	movl	40(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ecx,%r11d
+	addl	%r12d,%eax
+	movl	%ecx,%r12d
+	roll	$5,%eax
+	addl	%ebx,%eax
+	notl	%r11d
+	leal	38016083(%rdx,%r10,1),%edx
+	andl	%eax,%r12d
+	andl	%ebx,%r11d
+	movl	60(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ebx,%r11d
+	addl	%r12d,%edx
+	movl	%ebx,%r12d
+	roll	$9,%edx
+	addl	%eax,%edx
+	notl	%r11d
+	leal	-660478335(%rcx,%r10,1),%ecx
+	andl	%edx,%r12d
+	andl	%eax,%r11d
+	movl	16(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%eax,%r11d
+	addl	%r12d,%ecx
+	movl	%eax,%r12d
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	notl	%r11d
+	leal	-405537848(%rbx,%r10,1),%ebx
+	andl	%ecx,%r12d
+	andl	%edx,%r11d
+	movl	36(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%edx,%r11d
+	addl	%r12d,%ebx
+	movl	%edx,%r12d
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	notl	%r11d
+	leal	568446438(%rax,%r10,1),%eax
+	andl	%ebx,%r12d
+	andl	%ecx,%r11d
+	movl	56(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ecx,%r11d
+	addl	%r12d,%eax
+	movl	%ecx,%r12d
+	roll	$5,%eax
+	addl	%ebx,%eax
+	notl	%r11d
+	leal	-1019803690(%rdx,%r10,1),%edx
+	andl	%eax,%r12d
+	andl	%ebx,%r11d
+	movl	12(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ebx,%r11d
+	addl	%r12d,%edx
+	movl	%ebx,%r12d
+	roll	$9,%edx
+	addl	%eax,%edx
+	notl	%r11d
+	leal	-187363961(%rcx,%r10,1),%ecx
+	andl	%edx,%r12d
+	andl	%eax,%r11d
+	movl	32(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%eax,%r11d
+	addl	%r12d,%ecx
+	movl	%eax,%r12d
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	notl	%r11d
+	leal	1163531501(%rbx,%r10,1),%ebx
+	andl	%ecx,%r12d
+	andl	%edx,%r11d
+	movl	52(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%edx,%r11d
+	addl	%r12d,%ebx
+	movl	%edx,%r12d
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	notl	%r11d
+	leal	-1444681467(%rax,%r10,1),%eax
+	andl	%ebx,%r12d
+	andl	%ecx,%r11d
+	movl	8(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ecx,%r11d
+	addl	%r12d,%eax
+	movl	%ecx,%r12d
+	roll	$5,%eax
+	addl	%ebx,%eax
+	notl	%r11d
+	leal	-51403784(%rdx,%r10,1),%edx
+	andl	%eax,%r12d
+	andl	%ebx,%r11d
+	movl	28(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%ebx,%r11d
+	addl	%r12d,%edx
+	movl	%ebx,%r12d
+	roll	$9,%edx
+	addl	%eax,%edx
+	notl	%r11d
+	leal	1735328473(%rcx,%r10,1),%ecx
+	andl	%edx,%r12d
+	andl	%eax,%r11d
+	movl	48(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%eax,%r11d
+	addl	%r12d,%ecx
+	movl	%eax,%r12d
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	notl	%r11d
+	leal	-1926607734(%rbx,%r10,1),%ebx
+	andl	%ecx,%r12d
+	andl	%edx,%r11d
+	movl	0(%rsi),%r10d
+	orl	%r11d,%r12d
+	movl	%edx,%r11d
+	addl	%r12d,%ebx
+	movl	%edx,%r12d
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	movl	20(%rsi),%r10d
+	movl	%ecx,%r11d
+	leal	-378558(%rax,%r10,1),%eax
+	movl	32(%rsi),%r10d
+	xorl	%edx,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%eax
+	roll	$4,%eax
+	movl	%ebx,%r11d
+	addl	%ebx,%eax
+	leal	-2022574463(%rdx,%r10,1),%edx
+	movl	44(%rsi),%r10d
+	xorl	%ecx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%edx
+	roll	$11,%edx
+	movl	%eax,%r11d
+	addl	%eax,%edx
+	leal	1839030562(%rcx,%r10,1),%ecx
+	movl	56(%rsi),%r10d
+	xorl	%ebx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ecx
+	roll	$16,%ecx
+	movl	%edx,%r11d
+	addl	%edx,%ecx
+	leal	-35309556(%rbx,%r10,1),%ebx
+	movl	4(%rsi),%r10d
+	xorl	%eax,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%ebx
+	roll	$23,%ebx
+	movl	%ecx,%r11d
+	addl	%ecx,%ebx
+	leal	-1530992060(%rax,%r10,1),%eax
+	movl	16(%rsi),%r10d
+	xorl	%edx,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%eax
+	roll	$4,%eax
+	movl	%ebx,%r11d
+	addl	%ebx,%eax
+	leal	1272893353(%rdx,%r10,1),%edx
+	movl	28(%rsi),%r10d
+	xorl	%ecx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%edx
+	roll	$11,%edx
+	movl	%eax,%r11d
+	addl	%eax,%edx
+	leal	-155497632(%rcx,%r10,1),%ecx
+	movl	40(%rsi),%r10d
+	xorl	%ebx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ecx
+	roll	$16,%ecx
+	movl	%edx,%r11d
+	addl	%edx,%ecx
+	leal	-1094730640(%rbx,%r10,1),%ebx
+	movl	52(%rsi),%r10d
+	xorl	%eax,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%ebx
+	roll	$23,%ebx
+	movl	%ecx,%r11d
+	addl	%ecx,%ebx
+	leal	681279174(%rax,%r10,1),%eax
+	movl	0(%rsi),%r10d
+	xorl	%edx,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%eax
+	roll	$4,%eax
+	movl	%ebx,%r11d
+	addl	%ebx,%eax
+	leal	-358537222(%rdx,%r10,1),%edx
+	movl	12(%rsi),%r10d
+	xorl	%ecx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%edx
+	roll	$11,%edx
+	movl	%eax,%r11d
+	addl	%eax,%edx
+	leal	-722521979(%rcx,%r10,1),%ecx
+	movl	24(%rsi),%r10d
+	xorl	%ebx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ecx
+	roll	$16,%ecx
+	movl	%edx,%r11d
+	addl	%edx,%ecx
+	leal	76029189(%rbx,%r10,1),%ebx
+	movl	36(%rsi),%r10d
+	xorl	%eax,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%ebx
+	roll	$23,%ebx
+	movl	%ecx,%r11d
+	addl	%ecx,%ebx
+	leal	-640364487(%rax,%r10,1),%eax
+	movl	48(%rsi),%r10d
+	xorl	%edx,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%eax
+	roll	$4,%eax
+	movl	%ebx,%r11d
+	addl	%ebx,%eax
+	leal	-421815835(%rdx,%r10,1),%edx
+	movl	60(%rsi),%r10d
+	xorl	%ecx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%edx
+	roll	$11,%edx
+	movl	%eax,%r11d
+	addl	%eax,%edx
+	leal	530742520(%rcx,%r10,1),%ecx
+	movl	8(%rsi),%r10d
+	xorl	%ebx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ecx
+	roll	$16,%ecx
+	movl	%edx,%r11d
+	addl	%edx,%ecx
+	leal	-995338651(%rbx,%r10,1),%ebx
+	movl	0(%rsi),%r10d
+	xorl	%eax,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%ebx
+	roll	$23,%ebx
+	movl	%ecx,%r11d
+	addl	%ecx,%ebx
+	movl	0(%rsi),%r10d
+	movl	$4294967295,%r11d
+	xorl	%edx,%r11d
+	leal	-198630844(%rax,%r10,1),%eax
+	orl	%ebx,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%eax
+	movl	28(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$6,%eax
+	xorl	%ecx,%r11d
+	addl	%ebx,%eax
+	leal	1126891415(%rdx,%r10,1),%edx
+	orl	%eax,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%edx
+	movl	56(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$10,%edx
+	xorl	%ebx,%r11d
+	addl	%eax,%edx
+	leal	-1416354905(%rcx,%r10,1),%ecx
+	orl	%edx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%ecx
+	movl	20(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$15,%ecx
+	xorl	%eax,%r11d
+	addl	%edx,%ecx
+	leal	-57434055(%rbx,%r10,1),%ebx
+	orl	%ecx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ebx
+	movl	48(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$21,%ebx
+	xorl	%edx,%r11d
+	addl	%ecx,%ebx
+	leal	1700485571(%rax,%r10,1),%eax
+	orl	%ebx,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%eax
+	movl	12(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$6,%eax
+	xorl	%ecx,%r11d
+	addl	%ebx,%eax
+	leal	-1894986606(%rdx,%r10,1),%edx
+	orl	%eax,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%edx
+	movl	40(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$10,%edx
+	xorl	%ebx,%r11d
+	addl	%eax,%edx
+	leal	-1051523(%rcx,%r10,1),%ecx
+	orl	%edx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%ecx
+	movl	4(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$15,%ecx
+	xorl	%eax,%r11d
+	addl	%edx,%ecx
+	leal	-2054922799(%rbx,%r10,1),%ebx
+	orl	%ecx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ebx
+	movl	32(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$21,%ebx
+	xorl	%edx,%r11d
+	addl	%ecx,%ebx
+	leal	1873313359(%rax,%r10,1),%eax
+	orl	%ebx,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%eax
+	movl	60(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$6,%eax
+	xorl	%ecx,%r11d
+	addl	%ebx,%eax
+	leal	-30611744(%rdx,%r10,1),%edx
+	orl	%eax,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%edx
+	movl	24(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$10,%edx
+	xorl	%ebx,%r11d
+	addl	%eax,%edx
+	leal	-1560198380(%rcx,%r10,1),%ecx
+	orl	%edx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%ecx
+	movl	52(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$15,%ecx
+	xorl	%eax,%r11d
+	addl	%edx,%ecx
+	leal	1309151649(%rbx,%r10,1),%ebx
+	orl	%ecx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ebx
+	movl	16(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$21,%ebx
+	xorl	%edx,%r11d
+	addl	%ecx,%ebx
+	leal	-145523070(%rax,%r10,1),%eax
+	orl	%ebx,%r11d
+	xorl	%ecx,%r11d
+	addl	%r11d,%eax
+	movl	44(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$6,%eax
+	xorl	%ecx,%r11d
+	addl	%ebx,%eax
+	leal	-1120210379(%rdx,%r10,1),%edx
+	orl	%eax,%r11d
+	xorl	%ebx,%r11d
+	addl	%r11d,%edx
+	movl	8(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$10,%edx
+	xorl	%ebx,%r11d
+	addl	%eax,%edx
+	leal	718787259(%rcx,%r10,1),%ecx
+	orl	%edx,%r11d
+	xorl	%eax,%r11d
+	addl	%r11d,%ecx
+	movl	36(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$15,%ecx
+	xorl	%eax,%r11d
+	addl	%edx,%ecx
+	leal	-343485551(%rbx,%r10,1),%ebx
+	orl	%ecx,%r11d
+	xorl	%edx,%r11d
+	addl	%r11d,%ebx
+	movl	0(%rsi),%r10d
+	movl	$4294967295,%r11d
+	roll	$21,%ebx
+	xorl	%edx,%r11d
+	addl	%ecx,%ebx
+
+	addl	%r8d,%eax
+	addl	%r9d,%ebx
+	addl	%r14d,%ecx
+	addl	%r15d,%edx
+
+
+	addq	$64,%rsi
+	cmpq	%rdi,%rsi
+	jb	L$loop
+
+
+
+L$end:
+	movl	%eax,0(%rbp)
+	movl	%ebx,4(%rbp)
+	movl	%ecx,8(%rbp)
+	movl	%edx,12(%rbp)
+
+	movq	(%rsp),%r15
+	movq	8(%rsp),%r14
+	movq	16(%rsp),%r12
+	movq	24(%rsp),%rbx
+	movq	32(%rsp),%rbp
+	addq	$40,%rsp
+L$epilogue:
+	.byte	0xf3,0xc3
+

deps/openssl/asm/x64-macosx-gas/rc4/rc4-x86_64.s

@@ -0,0 +1,430 @@
+.text
+
+
+.globl	_RC4
+
+.p2align	4
+_RC4:	orq	%rsi,%rsi
+	jne	L$entry
+	.byte	0xf3,0xc3
+L$entry:
+	pushq	%rbx
+	pushq	%r12
+	pushq	%r13
+L$prologue:
+
+	addq	$8,%rdi
+	movl	-8(%rdi),%r8d
+	movl	-4(%rdi),%r12d
+	cmpl	$-1,256(%rdi)
+	je	L$RC4_CHAR
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	testq	$-8,%rsi
+	jz	L$loop1
+	jmp	L$loop8
+.p2align	4
+L$loop8:
+	addb	%r9b,%r12b
+	movq	%r8,%r10
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r10b
+	movl	(%rdi,%r10,4),%r11d
+	cmpq	%r10,%r12
+	movl	%r9d,(%rdi,%r12,4)
+	cmoveq	%r9,%r11
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r9b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r11b,%r12b
+	movq	%r10,%r8
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	cmpq	%r8,%r12
+	movl	%r11d,(%rdi,%r12,4)
+	cmoveq	%r11,%r9
+	movl	%r13d,(%rdi,%r10,4)
+	addb	%r11b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r9b,%r12b
+	movq	%r8,%r10
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r10b
+	movl	(%rdi,%r10,4),%r11d
+	cmpq	%r10,%r12
+	movl	%r9d,(%rdi,%r12,4)
+	cmoveq	%r9,%r11
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r9b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r11b,%r12b
+	movq	%r10,%r8
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	cmpq	%r8,%r12
+	movl	%r11d,(%rdi,%r12,4)
+	cmoveq	%r11,%r9
+	movl	%r13d,(%rdi,%r10,4)
+	addb	%r11b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r9b,%r12b
+	movq	%r8,%r10
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r10b
+	movl	(%rdi,%r10,4),%r11d
+	cmpq	%r10,%r12
+	movl	%r9d,(%rdi,%r12,4)
+	cmoveq	%r9,%r11
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r9b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r11b,%r12b
+	movq	%r10,%r8
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	cmpq	%r8,%r12
+	movl	%r11d,(%rdi,%r12,4)
+	cmoveq	%r11,%r9
+	movl	%r13d,(%rdi,%r10,4)
+	addb	%r11b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r9b,%r12b
+	movq	%r8,%r10
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r10b
+	movl	(%rdi,%r10,4),%r11d
+	cmpq	%r10,%r12
+	movl	%r9d,(%rdi,%r12,4)
+	cmoveq	%r9,%r11
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r9b,%r13b
+	movb	(%rdi,%r13,4),%al
+	addb	%r11b,%r12b
+	movq	%r10,%r8
+	movl	(%rdi,%r12,4),%r13d
+	rorq	$8,%rax
+	incb	%r8b
+	movl	(%rdi,%r8,4),%r9d
+	cmpq	%r8,%r12
+	movl	%r11d,(%rdi,%r12,4)
+	cmoveq	%r11,%r9
+	movl	%r13d,(%rdi,%r10,4)
+	addb	%r11b,%r13b
+	movb	(%rdi,%r13,4),%al
+	rorq	$8,%rax
+	subq	$8,%rsi
+
+	xorq	(%rdx),%rax
+	addq	$8,%rdx
+	movq	%rax,(%rcx)
+	addq	$8,%rcx
+
+	testq	$-8,%rsi
+	jnz	L$loop8
+	cmpq	$0,%rsi
+	jne	L$loop1
+	jmp	L$exit
+
+.p2align	4
+L$loop1:
+	addb	%r9b,%r12b
+	movl	(%rdi,%r12,4),%r13d
+	movl	%r9d,(%rdi,%r12,4)
+	movl	%r13d,(%rdi,%r8,4)
+	addb	%r13b,%r9b
+	incb	%r8b
+	movl	(%rdi,%r9,4),%r13d
+	movl	(%rdi,%r8,4),%r9d
+	xorb	(%rdx),%r13b
+	incq	%rdx
+	movb	%r13b,(%rcx)
+	incq	%rcx
+	decq	%rsi
+	jnz	L$loop1
+	jmp	L$exit
+
+.p2align	4
+L$RC4_CHAR:
+	addb	$1,%r8b
+	movzbl	(%rdi,%r8,1),%r9d
+	testq	$-8,%rsi
+	jz	L$cloop1
+	cmpl	$0,260(%rdi)
+	jnz	L$cloop1
+	jmp	L$cloop8
+.p2align	4
+L$cloop8:
+	movl	(%rdx),%eax
+	movl	4(%rdx),%ebx
+	addb	%r9b,%r12b
+	leaq	1(%r8),%r10
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r10b,%r10d
+	movzbl	(%rdi,%r10,1),%r11d
+	movb	%r9b,(%rdi,%r12,1)
+	cmpq	%r10,%r12
+	movb	%r13b,(%rdi,%r8,1)
+	jne	L$cmov0
+
+	movq	%r9,%r11
+L$cmov0:
+	addb	%r9b,%r13b
+	xorb	(%rdi,%r13,1),%al
+	rorl	$8,%eax
+	addb	%r11b,%r12b
+	leaq	1(%r10),%r8
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r8,1),%r9d
+	movb	%r11b,(%rdi,%r12,1)
+	cmpq	%r8,%r12
+	movb	%r13b,(%rdi,%r10,1)
+	jne	L$cmov1
+
+	movq	%r11,%r9
+L$cmov1:
+	addb	%r11b,%r13b
+	xorb	(%rdi,%r13,1),%al
+	rorl	$8,%eax
+	addb	%r9b,%r12b
+	leaq	1(%r8),%r10
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r10b,%r10d
+	movzbl	(%rdi,%r10,1),%r11d
+	movb	%r9b,(%rdi,%r12,1)
+	cmpq	%r10,%r12
+	movb	%r13b,(%rdi,%r8,1)
+	jne	L$cmov2
+
+	movq	%r9,%r11
+L$cmov2:
+	addb	%r9b,%r13b
+	xorb	(%rdi,%r13,1),%al
+	rorl	$8,%eax
+	addb	%r11b,%r12b
+	leaq	1(%r10),%r8
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r8,1),%r9d
+	movb	%r11b,(%rdi,%r12,1)
+	cmpq	%r8,%r12
+	movb	%r13b,(%rdi,%r10,1)
+	jne	L$cmov3
+
+	movq	%r11,%r9
+L$cmov3:
+	addb	%r11b,%r13b
+	xorb	(%rdi,%r13,1),%al
+	rorl	$8,%eax
+	addb	%r9b,%r12b
+	leaq	1(%r8),%r10
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r10b,%r10d
+	movzbl	(%rdi,%r10,1),%r11d
+	movb	%r9b,(%rdi,%r12,1)
+	cmpq	%r10,%r12
+	movb	%r13b,(%rdi,%r8,1)
+	jne	L$cmov4
+
+	movq	%r9,%r11
+L$cmov4:
+	addb	%r9b,%r13b
+	xorb	(%rdi,%r13,1),%bl
+	rorl	$8,%ebx
+	addb	%r11b,%r12b
+	leaq	1(%r10),%r8
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r8,1),%r9d
+	movb	%r11b,(%rdi,%r12,1)
+	cmpq	%r8,%r12
+	movb	%r13b,(%rdi,%r10,1)
+	jne	L$cmov5
+
+	movq	%r11,%r9
+L$cmov5:
+	addb	%r11b,%r13b
+	xorb	(%rdi,%r13,1),%bl
+	rorl	$8,%ebx
+	addb	%r9b,%r12b
+	leaq	1(%r8),%r10
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r10b,%r10d
+	movzbl	(%rdi,%r10,1),%r11d
+	movb	%r9b,(%rdi,%r12,1)
+	cmpq	%r10,%r12
+	movb	%r13b,(%rdi,%r8,1)
+	jne	L$cmov6
+
+	movq	%r9,%r11
+L$cmov6:
+	addb	%r9b,%r13b
+	xorb	(%rdi,%r13,1),%bl
+	rorl	$8,%ebx
+	addb	%r11b,%r12b
+	leaq	1(%r10),%r8
+	movzbl	(%rdi,%r12,1),%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r8,1),%r9d
+	movb	%r11b,(%rdi,%r12,1)
+	cmpq	%r8,%r12
+	movb	%r13b,(%rdi,%r10,1)
+	jne	L$cmov7
+
+	movq	%r11,%r9
+L$cmov7:
+	addb	%r11b,%r13b
+	xorb	(%rdi,%r13,1),%bl
+	rorl	$8,%ebx
+	leaq	-8(%rsi),%rsi
+	movl	%eax,(%rcx)
+	leaq	8(%rdx),%rdx
+	movl	%ebx,4(%rcx)
+	leaq	8(%rcx),%rcx
+
+	testq	$-8,%rsi
+	jnz	L$cloop8
+	cmpq	$0,%rsi
+	jne	L$cloop1
+	jmp	L$exit
+.p2align	4
+L$cloop1:
+	addb	%r9b,%r12b
+	movzbl	(%rdi,%r12,1),%r13d
+	movb	%r9b,(%rdi,%r12,1)
+	movb	%r13b,(%rdi,%r8,1)
+	addb	%r9b,%r13b
+	addb	$1,%r8b
+	movzbl	%r13b,%r13d
+	movzbl	%r8b,%r8d
+	movzbl	(%rdi,%r13,1),%r13d
+	movzbl	(%rdi,%r8,1),%r9d
+	xorb	(%rdx),%r13b
+	leaq	1(%rdx),%rdx
+	movb	%r13b,(%rcx)
+	leaq	1(%rcx),%rcx
+	subq	$1,%rsi
+	jnz	L$cloop1
+	jmp	L$exit
+
+.p2align	4
+L$exit:
+	subb	$1,%r8b
+	movl	%r8d,-8(%rdi)
+	movl	%r12d,-4(%rdi)
+
+	movq	(%rsp),%r13
+	movq	8(%rsp),%r12
+	movq	16(%rsp),%rbx
+	addq	$24,%rsp
+L$epilogue:
+	.byte	0xf3,0xc3
+
+
+.globl	_RC4_set_key
+
+.p2align	4
+_RC4_set_key:
+	leaq	8(%rdi),%rdi
+	leaq	(%rdx,%rsi,1),%rdx
+	negq	%rsi
+	movq	%rsi,%rcx
+	xorl	%eax,%eax
+	xorq	%r9,%r9
+	xorq	%r10,%r10
+	xorq	%r11,%r11
+
+	movl	_OPENSSL_ia32cap_P(%rip),%r8d
+	btl	$20,%r8d
+	jnc	L$w1stloop
+	btl	$30,%r8d
+	setc	%r9b
+	movl	%r9d,260(%rdi)
+	jmp	L$c1stloop
+
+.p2align	4
+L$w1stloop:
+	movl	%eax,(%rdi,%rax,4)
+	addb	$1,%al
+	jnc	L$w1stloop
+
+	xorq	%r9,%r9
+	xorq	%r8,%r8
+.p2align	4
+L$w2ndloop:
+	movl	(%rdi,%r9,4),%r10d
+	addb	(%rdx,%rsi,1),%r8b
+	addb	%r10b,%r8b
+	addq	$1,%rsi
+	movl	(%rdi,%r8,4),%r11d
+	cmovzq	%rcx,%rsi
+	movl	%r10d,(%rdi,%r8,4)
+	movl	%r11d,(%rdi,%r9,4)
+	addb	$1,%r9b
+	jnc	L$w2ndloop
+	jmp	L$exit_key
+
+.p2align	4
+L$c1stloop:
+	movb	%al,(%rdi,%rax,1)
+	addb	$1,%al
+	jnc	L$c1stloop
+
+	xorq	%r9,%r9
+	xorq	%r8,%r8
+.p2align	4
+L$c2ndloop:
+	movb	(%rdi,%r9,1),%r10b
+	addb	(%rdx,%rsi,1),%r8b
+	addb	%r10b,%r8b
+	addq	$1,%rsi
+	movb	(%rdi,%r8,1),%r11b
+	jnz	L$cnowrap
+	movq	%rcx,%rsi
+L$cnowrap:
+	movb	%r10b,(%rdi,%r8,1)
+	movb	%r11b,(%rdi,%r9,1)
+	addb	$1,%r9b
+	jnc	L$c2ndloop
+	movl	$-1,256(%rdi)
+
+.p2align	4
+L$exit_key:
+	xorl	%eax,%eax
+	movl	%eax,-8(%rdi)
+	movl	%eax,-4(%rdi)
+	.byte	0xf3,0xc3
+
+
+.globl	_RC4_options
+
+.p2align	4
+_RC4_options:
+	leaq	L$opts(%rip),%rax
+	movl	_OPENSSL_ia32cap_P(%rip),%edx
+	btl	$20,%edx
+	jnc	L$done
+	addq	$12,%rax
+	btl	$30,%edx
+	jnc	L$done
+	addq	$13,%rax
+L$done:
+	.byte	0xf3,0xc3
+.p2align	6
+L$opts:
+.byte	114,99,52,40,56,120,44,105,110,116,41,0
+.byte	114,99,52,40,56,120,44,99,104,97,114,41,0
+.byte	114,99,52,40,49,120,44,99,104,97,114,41,0
+.byte	82,67,52,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.p2align	6
+

deps/openssl/asm/x64-macosx-gas/whrlpool/wp-x86_64.s

@@ -0,0 +1,859 @@
+.text
+
+
+.globl	_whirlpool_block
+
+.p2align	4
+_whirlpool_block:
+	pushq	%rbx
+	pushq	%rbp
+	pushq	%r12
+	pushq	%r13
+	pushq	%r14
+	pushq	%r15
+
+	movq	%rsp,%r11
+	subq	$128+40,%rsp
+	andq	$-64,%rsp
+
+	leaq	128(%rsp),%r10
+	movq	%rdi,0(%r10)
+	movq	%rsi,8(%r10)
+	movq	%rdx,16(%r10)
+	movq	%r11,32(%r10)
+L$prologue:
+
+	movq	%r10,%rbx
+	leaq	L$table(%rip),%rbp
+
+	xorq	%rcx,%rcx
+	xorq	%rdx,%rdx
+	movq	0(%rdi),%r8
+	movq	8(%rdi),%r9
+	movq	16(%rdi),%r10
+	movq	24(%rdi),%r11
+	movq	32(%rdi),%r12
+	movq	40(%rdi),%r13
+	movq	48(%rdi),%r14
+	movq	56(%rdi),%r15
+L$outerloop:
+	movq	%r8,0(%rsp)
+	movq	%r9,8(%rsp)
+	movq	%r10,16(%rsp)
+	movq	%r11,24(%rsp)
+	movq	%r12,32(%rsp)
+	movq	%r13,40(%rsp)
+	movq	%r14,48(%rsp)
+	movq	%r15,56(%rsp)
+	xorq	0(%rsi),%r8
+	xorq	8(%rsi),%r9
+	xorq	16(%rsi),%r10
+	xorq	24(%rsi),%r11
+	xorq	32(%rsi),%r12
+	xorq	40(%rsi),%r13
+	xorq	48(%rsi),%r14
+	xorq	56(%rsi),%r15
+	movq	%r8,64+0(%rsp)
+	movq	%r9,64+8(%rsp)
+	movq	%r10,64+16(%rsp)
+	movq	%r11,64+24(%rsp)
+	movq	%r12,64+32(%rsp)
+	movq	%r13,64+40(%rsp)
+	movq	%r14,64+48(%rsp)
+	movq	%r15,64+56(%rsp)
+	xorq	%rsi,%rsi
+	movq	%rsi,24(%rbx)
+.p2align	4
+L$round:
+	movq	4096(%rbp,%rsi,8),%r8
+	movl	0(%rsp),%eax
+	movl	4(%rsp),%ebx
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r8
+	movq	7(%rbp,%rdi,8),%r9
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	0+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	movq	6(%rbp,%rsi,8),%r10
+	movq	5(%rbp,%rdi,8),%r11
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	movq	4(%rbp,%rsi,8),%r12
+	movq	3(%rbp,%rdi,8),%r13
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	0+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	movq	2(%rbp,%rsi,8),%r14
+	movq	1(%rbp,%rdi,8),%r15
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r9
+	xorq	7(%rbp,%rdi,8),%r10
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	8+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r11
+	xorq	5(%rbp,%rdi,8),%r12
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r13
+	xorq	3(%rbp,%rdi,8),%r14
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	8+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r15
+	xorq	1(%rbp,%rdi,8),%r8
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r10
+	xorq	7(%rbp,%rdi,8),%r11
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	16+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r12
+	xorq	5(%rbp,%rdi,8),%r13
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r14
+	xorq	3(%rbp,%rdi,8),%r15
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	16+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r8
+	xorq	1(%rbp,%rdi,8),%r9
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r11
+	xorq	7(%rbp,%rdi,8),%r12
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	24+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r13
+	xorq	5(%rbp,%rdi,8),%r14
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r15
+	xorq	3(%rbp,%rdi,8),%r8
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	24+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r9
+	xorq	1(%rbp,%rdi,8),%r10
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r12
+	xorq	7(%rbp,%rdi,8),%r13
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	32+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r14
+	xorq	5(%rbp,%rdi,8),%r15
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r8
+	xorq	3(%rbp,%rdi,8),%r9
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	32+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r10
+	xorq	1(%rbp,%rdi,8),%r11
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r13
+	xorq	7(%rbp,%rdi,8),%r14
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	40+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r15
+	xorq	5(%rbp,%rdi,8),%r8
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r9
+	xorq	3(%rbp,%rdi,8),%r10
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	40+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r11
+	xorq	1(%rbp,%rdi,8),%r12
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r14
+	xorq	7(%rbp,%rdi,8),%r15
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	48+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r8
+	xorq	5(%rbp,%rdi,8),%r9
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r10
+	xorq	3(%rbp,%rdi,8),%r11
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	48+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r12
+	xorq	1(%rbp,%rdi,8),%r13
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r15
+	xorq	7(%rbp,%rdi,8),%r8
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	56+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r9
+	xorq	5(%rbp,%rdi,8),%r10
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r11
+	xorq	3(%rbp,%rdi,8),%r12
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	56+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r13
+	xorq	1(%rbp,%rdi,8),%r14
+	movq	%r8,0(%rsp)
+	movq	%r9,8(%rsp)
+	movq	%r10,16(%rsp)
+	movq	%r11,24(%rsp)
+	movq	%r12,32(%rsp)
+	movq	%r13,40(%rsp)
+	movq	%r14,48(%rsp)
+	movq	%r15,56(%rsp)
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r8
+	xorq	7(%rbp,%rdi,8),%r9
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+0+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r10
+	xorq	5(%rbp,%rdi,8),%r11
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r12
+	xorq	3(%rbp,%rdi,8),%r13
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+0+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r14
+	xorq	1(%rbp,%rdi,8),%r15
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r9
+	xorq	7(%rbp,%rdi,8),%r10
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+8+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r11
+	xorq	5(%rbp,%rdi,8),%r12
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r13
+	xorq	3(%rbp,%rdi,8),%r14
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+8+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r15
+	xorq	1(%rbp,%rdi,8),%r8
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r10
+	xorq	7(%rbp,%rdi,8),%r11
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+16+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r12
+	xorq	5(%rbp,%rdi,8),%r13
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r14
+	xorq	3(%rbp,%rdi,8),%r15
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+16+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r8
+	xorq	1(%rbp,%rdi,8),%r9
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r11
+	xorq	7(%rbp,%rdi,8),%r12
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+24+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r13
+	xorq	5(%rbp,%rdi,8),%r14
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r15
+	xorq	3(%rbp,%rdi,8),%r8
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+24+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r9
+	xorq	1(%rbp,%rdi,8),%r10
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r12
+	xorq	7(%rbp,%rdi,8),%r13
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+32+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r14
+	xorq	5(%rbp,%rdi,8),%r15
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r8
+	xorq	3(%rbp,%rdi,8),%r9
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+32+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r10
+	xorq	1(%rbp,%rdi,8),%r11
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r13
+	xorq	7(%rbp,%rdi,8),%r14
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+40+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r15
+	xorq	5(%rbp,%rdi,8),%r8
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r9
+	xorq	3(%rbp,%rdi,8),%r10
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+40+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r11
+	xorq	1(%rbp,%rdi,8),%r12
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r14
+	xorq	7(%rbp,%rdi,8),%r15
+	movb	%al,%cl
+	movb	%ah,%dl
+	movl	64+48+8(%rsp),%eax
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r8
+	xorq	5(%rbp,%rdi,8),%r9
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r10
+	xorq	3(%rbp,%rdi,8),%r11
+	movb	%bl,%cl
+	movb	%bh,%dl
+	movl	64+48+8+4(%rsp),%ebx
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r12
+	xorq	1(%rbp,%rdi,8),%r13
+	movb	%al,%cl
+	movb	%ah,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%eax
+	xorq	0(%rbp,%rsi,8),%r15
+	xorq	7(%rbp,%rdi,8),%r8
+	movb	%al,%cl
+	movb	%ah,%dl
+
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	6(%rbp,%rsi,8),%r9
+	xorq	5(%rbp,%rdi,8),%r10
+	movb	%bl,%cl
+	movb	%bh,%dl
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	shrl	$16,%ebx
+	xorq	4(%rbp,%rsi,8),%r11
+	xorq	3(%rbp,%rdi,8),%r12
+	movb	%bl,%cl
+	movb	%bh,%dl
+
+	leaq	(%rcx,%rcx,1),%rsi
+	leaq	(%rdx,%rdx,1),%rdi
+	xorq	2(%rbp,%rsi,8),%r13
+	xorq	1(%rbp,%rdi,8),%r14
+	leaq	128(%rsp),%rbx
+	movq	24(%rbx),%rsi
+	addq	$1,%rsi
+	cmpq	$10,%rsi
+	je	L$roundsdone
+
+	movq	%rsi,24(%rbx)
+	movq	%r8,64+0(%rsp)
+	movq	%r9,64+8(%rsp)
+	movq	%r10,64+16(%rsp)
+	movq	%r11,64+24(%rsp)
+	movq	%r12,64+32(%rsp)
+	movq	%r13,64+40(%rsp)
+	movq	%r14,64+48(%rsp)
+	movq	%r15,64+56(%rsp)
+	jmp	L$round
+.p2align	4
+L$roundsdone:
+	movq	0(%rbx),%rdi
+	movq	8(%rbx),%rsi
+	movq	16(%rbx),%rax
+	xorq	0(%rsi),%r8
+	xorq	8(%rsi),%r9
+	xorq	16(%rsi),%r10
+	xorq	24(%rsi),%r11
+	xorq	32(%rsi),%r12
+	xorq	40(%rsi),%r13
+	xorq	48(%rsi),%r14
+	xorq	56(%rsi),%r15
+	xorq	0(%rdi),%r8
+	xorq	8(%rdi),%r9
+	xorq	16(%rdi),%r10
+	xorq	24(%rdi),%r11
+	xorq	32(%rdi),%r12
+	xorq	40(%rdi),%r13
+	xorq	48(%rdi),%r14
+	xorq	56(%rdi),%r15
+	movq	%r8,0(%rdi)
+	movq	%r9,8(%rdi)
+	movq	%r10,16(%rdi)
+	movq	%r11,24(%rdi)
+	movq	%r12,32(%rdi)
+	movq	%r13,40(%rdi)
+	movq	%r14,48(%rdi)
+	movq	%r15,56(%rdi)
+	leaq	64(%rsi),%rsi
+	subq	$1,%rax
+	jz	L$alldone
+	movq	%rsi,8(%rbx)
+	movq	%rax,16(%rbx)
+	jmp	L$outerloop
+L$alldone:
+	movq	32(%rbx),%rsi
+	movq	(%rsi),%r15
+	movq	8(%rsi),%r14
+	movq	16(%rsi),%r13
+	movq	24(%rsi),%r12
+	movq	32(%rsi),%rbp
+	movq	40(%rsi),%rbx
+	leaq	48(%rsi),%rsp
+L$epilogue:
+	.byte	0xf3,0xc3
+
+
+.p2align	6
+
+L$table:
+.byte	24,24,96,24,192,120,48,216,24,24,96,24,192,120,48,216
+.byte	35,35,140,35,5,175,70,38,35,35,140,35,5,175,70,38
+.byte	198,198,63,198,126,249,145,184,198,198,63,198,126,249,145,184
+.byte	232,232,135,232,19,111,205,251,232,232,135,232,19,111,205,251
+.byte	135,135,38,135,76,161,19,203,135,135,38,135,76,161,19,203
+.byte	184,184,218,184,169,98,109,17,184,184,218,184,169,98,109,17
+.byte	1,1,4,1,8,5,2,9,1,1,4,1,8,5,2,9
+.byte	79,79,33,79,66,110,158,13,79,79,33,79,66,110,158,13
+.byte	54,54,216,54,173,238,108,155,54,54,216,54,173,238,108,155
+.byte	166,166,162,166,89,4,81,255,166,166,162,166,89,4,81,255
+.byte	210,210,111,210,222,189,185,12,210,210,111,210,222,189,185,12
+.byte	245,245,243,245,251,6,247,14,245,245,243,245,251,6,247,14
+.byte	121,121,249,121,239,128,242,150,121,121,249,121,239,128,242,150
+.byte	111,111,161,111,95,206,222,48,111,111,161,111,95,206,222,48
+.byte	145,145,126,145,252,239,63,109,145,145,126,145,252,239,63,109
+.byte	82,82,85,82,170,7,164,248,82,82,85,82,170,7,164,248
+.byte	96,96,157,96,39,253,192,71,96,96,157,96,39,253,192,71
+.byte	188,188,202,188,137,118,101,53,188,188,202,188,137,118,101,53
+.byte	155,155,86,155,172,205,43,55,155,155,86,155,172,205,43,55
+.byte	142,142,2,142,4,140,1,138,142,142,2,142,4,140,1,138
+.byte	163,163,182,163,113,21,91,210,163,163,182,163,113,21,91,210
+.byte	12,12,48,12,96,60,24,108,12,12,48,12,96,60,24,108
+.byte	123,123,241,123,255,138,246,132,123,123,241,123,255,138,246,132
+.byte	53,53,212,53,181,225,106,128,53,53,212,53,181,225,106,128
+.byte	29,29,116,29,232,105,58,245,29,29,116,29,232,105,58,245
+.byte	224,224,167,224,83,71,221,179,224,224,167,224,83,71,221,179
+.byte	215,215,123,215,246,172,179,33,215,215,123,215,246,172,179,33
+.byte	194,194,47,194,94,237,153,156,194,194,47,194,94,237,153,156
+.byte	46,46,184,46,109,150,92,67,46,46,184,46,109,150,92,67
+.byte	75,75,49,75,98,122,150,41,75,75,49,75,98,122,150,41
+.byte	254,254,223,254,163,33,225,93,254,254,223,254,163,33,225,93
+.byte	87,87,65,87,130,22,174,213,87,87,65,87,130,22,174,213
+.byte	21,21,84,21,168,65,42,189,21,21,84,21,168,65,42,189
+.byte	119,119,193,119,159,182,238,232,119,119,193,119,159,182,238,232
+.byte	55,55,220,55,165,235,110,146,55,55,220,55,165,235,110,146
+.byte	229,229,179,229,123,86,215,158,229,229,179,229,123,86,215,158
+.byte	159,159,70,159,140,217,35,19,159,159,70,159,140,217,35,19
+.byte	240,240,231,240,211,23,253,35,240,240,231,240,211,23,253,35
+.byte	74,74,53,74,106,127,148,32,74,74,53,74,106,127,148,32
+.byte	218,218,79,218,158,149,169,68,218,218,79,218,158,149,169,68
+.byte	88,88,125,88,250,37,176,162,88,88,125,88,250,37,176,162
+.byte	201,201,3,201,6,202,143,207,201,201,3,201,6,202,143,207
+.byte	41,41,164,41,85,141,82,124,41,41,164,41,85,141,82,124
+.byte	10,10,40,10,80,34,20,90,10,10,40,10,80,34,20,90
+.byte	177,177,254,177,225,79,127,80,177,177,254,177,225,79,127,80
+.byte	160,160,186,160,105,26,93,201,160,160,186,160,105,26,93,201
+.byte	107,107,177,107,127,218,214,20,107,107,177,107,127,218,214,20
+.byte	133,133,46,133,92,171,23,217,133,133,46,133,92,171,23,217
+.byte	189,189,206,189,129,115,103,60,189,189,206,189,129,115,103,60
+.byte	93,93,105,93,210,52,186,143,93,93,105,93,210,52,186,143
+.byte	16,16,64,16,128,80,32,144,16,16,64,16,128,80,32,144
+.byte	244,244,247,244,243,3,245,7,244,244,247,244,243,3,245,7
+.byte	203,203,11,203,22,192,139,221,203,203,11,203,22,192,139,221
+.byte	62,62,248,62,237,198,124,211,62,62,248,62,237,198,124,211
+.byte	5,5,20,5,40,17,10,45,5,5,20,5,40,17,10,45
+.byte	103,103,129,103,31,230,206,120,103,103,129,103,31,230,206,120
+.byte	228,228,183,228,115,83,213,151,228,228,183,228,115,83,213,151
+.byte	39,39,156,39,37,187,78,2,39,39,156,39,37,187,78,2
+.byte	65,65,25,65,50,88,130,115,65,65,25,65,50,88,130,115
+.byte	139,139,22,139,44,157,11,167,139,139,22,139,44,157,11,167
+.byte	167,167,166,167,81,1,83,246,167,167,166,167,81,1,83,246
+.byte	125,125,233,125,207,148,250,178,125,125,233,125,207,148,250,178
+.byte	149,149,110,149,220,251,55,73,149,149,110,149,220,251,55,73
+.byte	216,216,71,216,142,159,173,86,216,216,71,216,142,159,173,86
+.byte	251,251,203,251,139,48,235,112,251,251,203,251,139,48,235,112
+.byte	238,238,159,238,35,113,193,205,238,238,159,238,35,113,193,205
+.byte	124,124,237,124,199,145,248,187,124,124,237,124,199,145,248,187
+.byte	102,102,133,102,23,227,204,113,102,102,133,102,23,227,204,113
+.byte	221,221,83,221,166,142,167,123,221,221,83,221,166,142,167,123
+.byte	23,23,92,23,184,75,46,175,23,23,92,23,184,75,46,175
+.byte	71,71,1,71,2,70,142,69,71,71,1,71,2,70,142,69
+.byte	158,158,66,158,132,220,33,26,158,158,66,158,132,220,33,26
+.byte	202,202,15,202,30,197,137,212,202,202,15,202,30,197,137,212
+.byte	45,45,180,45,117,153,90,88,45,45,180,45,117,153,90,88
+.byte	191,191,198,191,145,121,99,46,191,191,198,191,145,121,99,46
+.byte	7,7,28,7,56,27,14,63,7,7,28,7,56,27,14,63
+.byte	173,173,142,173,1,35,71,172,173,173,142,173,1,35,71,172
+.byte	90,90,117,90,234,47,180,176,90,90,117,90,234,47,180,176
+.byte	131,131,54,131,108,181,27,239,131,131,54,131,108,181,27,239
+.byte	51,51,204,51,133,255,102,182,51,51,204,51,133,255,102,182
+.byte	99,99,145,99,63,242,198,92,99,99,145,99,63,242,198,92
+.byte	2,2,8,2,16,10,4,18,2,2,8,2,16,10,4,18
+.byte	170,170,146,170,57,56,73,147,170,170,146,170,57,56,73,147
+.byte	113,113,217,113,175,168,226,222,113,113,217,113,175,168,226,222
+.byte	200,200,7,200,14,207,141,198,200,200,7,200,14,207,141,198
+.byte	25,25,100,25,200,125,50,209,25,25,100,25,200,125,50,209
+.byte	73,73,57,73,114,112,146,59,73,73,57,73,114,112,146,59
+.byte	217,217,67,217,134,154,175,95,217,217,67,217,134,154,175,95
+.byte	242,242,239,242,195,29,249,49,242,242,239,242,195,29,249,49
+.byte	227,227,171,227,75,72,219,168,227,227,171,227,75,72,219,168
+.byte	91,91,113,91,226,42,182,185,91,91,113,91,226,42,182,185
+.byte	136,136,26,136,52,146,13,188,136,136,26,136,52,146,13,188
+.byte	154,154,82,154,164,200,41,62,154,154,82,154,164,200,41,62
+.byte	38,38,152,38,45,190,76,11,38,38,152,38,45,190,76,11
+.byte	50,50,200,50,141,250,100,191,50,50,200,50,141,250,100,191
+.byte	176,176,250,176,233,74,125,89,176,176,250,176,233,74,125,89
+.byte	233,233,131,233,27,106,207,242,233,233,131,233,27,106,207,242
+.byte	15,15,60,15,120,51,30,119,15,15,60,15,120,51,30,119
+.byte	213,213,115,213,230,166,183,51,213,213,115,213,230,166,183,51
+.byte	128,128,58,128,116,186,29,244,128,128,58,128,116,186,29,244
+.byte	190,190,194,190,153,124,97,39,190,190,194,190,153,124,97,39
+.byte	205,205,19,205,38,222,135,235,205,205,19,205,38,222,135,235
+.byte	52,52,208,52,189,228,104,137,52,52,208,52,189,228,104,137
+.byte	72,72,61,72,122,117,144,50,72,72,61,72,122,117,144,50
+.byte	255,255,219,255,171,36,227,84,255,255,219,255,171,36,227,84
+.byte	122,122,245,122,247,143,244,141,122,122,245,122,247,143,244,141
+.byte	144,144,122,144,244,234,61,100,144,144,122,144,244,234,61,100
+.byte	95,95,97,95,194,62,190,157,95,95,97,95,194,62,190,157
+.byte	32,32,128,32,29,160,64,61,32,32,128,32,29,160,64,61
+.byte	104,104,189,104,103,213,208,15,104,104,189,104,103,213,208,15
+.byte	26,26,104,26,208,114,52,202,26,26,104,26,208,114,52,202
+.byte	174,174,130,174,25,44,65,183,174,174,130,174,25,44,65,183
+.byte	180,180,234,180,201,94,117,125,180,180,234,180,201,94,117,125
+.byte	84,84,77,84,154,25,168,206,84,84,77,84,154,25,168,206
+.byte	147,147,118,147,236,229,59,127,147,147,118,147,236,229,59,127
+.byte	34,34,136,34,13,170,68,47,34,34,136,34,13,170,68,47
+.byte	100,100,141,100,7,233,200,99,100,100,141,100,7,233,200,99
+.byte	241,241,227,241,219,18,255,42,241,241,227,241,219,18,255,42
+.byte	115,115,209,115,191,162,230,204,115,115,209,115,191,162,230,204
+.byte	18,18,72,18,144,90,36,130,18,18,72,18,144,90,36,130
+.byte	64,64,29,64,58,93,128,122,64,64,29,64,58,93,128,122
+.byte	8,8,32,8,64,40,16,72,8,8,32,8,64,40,16,72
+.byte	195,195,43,195,86,232,155,149,195,195,43,195,86,232,155,149
+.byte	236,236,151,236,51,123,197,223,236,236,151,236,51,123,197,223
+.byte	219,219,75,219,150,144,171,77,219,219,75,219,150,144,171,77
+.byte	161,161,190,161,97,31,95,192,161,161,190,161,97,31,95,192
+.byte	141,141,14,141,28,131,7,145,141,141,14,141,28,131,7,145
+.byte	61,61,244,61,245,201,122,200,61,61,244,61,245,201,122,200
+.byte	151,151,102,151,204,241,51,91,151,151,102,151,204,241,51,91
+.byte	0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+.byte	207,207,27,207,54,212,131,249,207,207,27,207,54,212,131,249
+.byte	43,43,172,43,69,135,86,110,43,43,172,43,69,135,86,110
+.byte	118,118,197,118,151,179,236,225,118,118,197,118,151,179,236,225
+.byte	130,130,50,130,100,176,25,230,130,130,50,130,100,176,25,230
+.byte	214,214,127,214,254,169,177,40,214,214,127,214,254,169,177,40
+.byte	27,27,108,27,216,119,54,195,27,27,108,27,216,119,54,195
+.byte	181,181,238,181,193,91,119,116,181,181,238,181,193,91,119,116
+.byte	175,175,134,175,17,41,67,190,175,175,134,175,17,41,67,190
+.byte	106,106,181,106,119,223,212,29,106,106,181,106,119,223,212,29
+.byte	80,80,93,80,186,13,160,234,80,80,93,80,186,13,160,234
+.byte	69,69,9,69,18,76,138,87,69,69,9,69,18,76,138,87
+.byte	243,243,235,243,203,24,251,56,243,243,235,243,203,24,251,56
+.byte	48,48,192,48,157,240,96,173,48,48,192,48,157,240,96,173
+.byte	239,239,155,239,43,116,195,196,239,239,155,239,43,116,195,196
+.byte	63,63,252,63,229,195,126,218,63,63,252,63,229,195,126,218
+.byte	85,85,73,85,146,28,170,199,85,85,73,85,146,28,170,199
+.byte	162,162,178,162,121,16,89,219,162,162,178,162,121,16,89,219
+.byte	234,234,143,234,3,101,201,233,234,234,143,234,3,101,201,233
+.byte	101,101,137,101,15,236,202,106,101,101,137,101,15,236,202,106
+.byte	186,186,210,186,185,104,105,3,186,186,210,186,185,104,105,3
+.byte	47,47,188,47,101,147,94,74,47,47,188,47,101,147,94,74
+.byte	192,192,39,192,78,231,157,142,192,192,39,192,78,231,157,142
+.byte	222,222,95,222,190,129,161,96,222,222,95,222,190,129,161,96
+.byte	28,28,112,28,224,108,56,252,28,28,112,28,224,108,56,252
+.byte	253,253,211,253,187,46,231,70,253,253,211,253,187,46,231,70
+.byte	77,77,41,77,82,100,154,31,77,77,41,77,82,100,154,31
+.byte	146,146,114,146,228,224,57,118,146,146,114,146,228,224,57,118
+.byte	117,117,201,117,143,188,234,250,117,117,201,117,143,188,234,250
+.byte	6,6,24,6,48,30,12,54,6,6,24,6,48,30,12,54
+.byte	138,138,18,138,36,152,9,174,138,138,18,138,36,152,9,174
+.byte	178,178,242,178,249,64,121,75,178,178,242,178,249,64,121,75
+.byte	230,230,191,230,99,89,209,133,230,230,191,230,99,89,209,133
+.byte	14,14,56,14,112,54,28,126,14,14,56,14,112,54,28,126
+.byte	31,31,124,31,248,99,62,231,31,31,124,31,248,99,62,231
+.byte	98,98,149,98,55,247,196,85,98,98,149,98,55,247,196,85
+.byte	212,212,119,212,238,163,181,58,212,212,119,212,238,163,181,58
+.byte	168,168,154,168,41,50,77,129,168,168,154,168,41,50,77,129
+.byte	150,150,98,150,196,244,49,82,150,150,98,150,196,244,49,82
+.byte	249,249,195,249,155,58,239,98,249,249,195,249,155,58,239,98
+.byte	197,197,51,197,102,246,151,163,197,197,51,197,102,246,151,163
+.byte	37,37,148,37,53,177,74,16,37,37,148,37,53,177,74,16
+.byte	89,89,121,89,242,32,178,171,89,89,121,89,242,32,178,171
+.byte	132,132,42,132,84,174,21,208,132,132,42,132,84,174,21,208
+.byte	114,114,213,114,183,167,228,197,114,114,213,114,183,167,228,197
+.byte	57,57,228,57,213,221,114,236,57,57,228,57,213,221,114,236
+.byte	76,76,45,76,90,97,152,22,76,76,45,76,90,97,152,22
+.byte	94,94,101,94,202,59,188,148,94,94,101,94,202,59,188,148
+.byte	120,120,253,120,231,133,240,159,120,120,253,120,231,133,240,159
+.byte	56,56,224,56,221,216,112,229,56,56,224,56,221,216,112,229
+.byte	140,140,10,140,20,134,5,152,140,140,10,140,20,134,5,152
+.byte	209,209,99,209,198,178,191,23,209,209,99,209,198,178,191,23
+.byte	165,165,174,165,65,11,87,228,165,165,174,165,65,11,87,228
+.byte	226,226,175,226,67,77,217,161,226,226,175,226,67,77,217,161
+.byte	97,97,153,97,47,248,194,78,97,97,153,97,47,248,194,78
+.byte	179,179,246,179,241,69,123,66,179,179,246,179,241,69,123,66
+.byte	33,33,132,33,21,165,66,52,33,33,132,33,21,165,66,52
+.byte	156,156,74,156,148,214,37,8,156,156,74,156,148,214,37,8
+.byte	30,30,120,30,240,102,60,238,30,30,120,30,240,102,60,238
+.byte	67,67,17,67,34,82,134,97,67,67,17,67,34,82,134,97
+.byte	199,199,59,199,118,252,147,177,199,199,59,199,118,252,147,177
+.byte	252,252,215,252,179,43,229,79,252,252,215,252,179,43,229,79
+.byte	4,4,16,4,32,20,8,36,4,4,16,4,32,20,8,36
+.byte	81,81,89,81,178,8,162,227,81,81,89,81,178,8,162,227
+.byte	153,153,94,153,188,199,47,37,153,153,94,153,188,199,47,37
+.byte	109,109,169,109,79,196,218,34,109,109,169,109,79,196,218,34
+.byte	13,13,52,13,104,57,26,101,13,13,52,13,104,57,26,101
+.byte	250,250,207,250,131,53,233,121,250,250,207,250,131,53,233,121
+.byte	223,223,91,223,182,132,163,105,223,223,91,223,182,132,163,105
+.byte	126,126,229,126,215,155,252,169,126,126,229,126,215,155,252,169
+.byte	36,36,144,36,61,180,72,25,36,36,144,36,61,180,72,25
+.byte	59,59,236,59,197,215,118,254,59,59,236,59,197,215,118,254
+.byte	171,171,150,171,49,61,75,154,171,171,150,171,49,61,75,154
+.byte	206,206,31,206,62,209,129,240,206,206,31,206,62,209,129,240
+.byte	17,17,68,17,136,85,34,153,17,17,68,17,136,85,34,153
+.byte	143,143,6,143,12,137,3,131,143,143,6,143,12,137,3,131
+.byte	78,78,37,78,74,107,156,4,78,78,37,78,74,107,156,4
+.byte	183,183,230,183,209,81,115,102,183,183,230,183,209,81,115,102
+.byte	235,235,139,235,11,96,203,224,235,235,139,235,11,96,203,224
+.byte	60,60,240,60,253,204,120,193,60,60,240,60,253,204,120,193
+.byte	129,129,62,129,124,191,31,253,129,129,62,129,124,191,31,253
+.byte	148,148,106,148,212,254,53,64,148,148,106,148,212,254,53,64
+.byte	247,247,251,247,235,12,243,28,247,247,251,247,235,12,243,28
+.byte	185,185,222,185,161,103,111,24,185,185,222,185,161,103,111,24
+.byte	19,19,76,19,152,95,38,139,19,19,76,19,152,95,38,139
+.byte	44,44,176,44,125,156,88,81,44,44,176,44,125,156,88,81
+.byte	211,211,107,211,214,184,187,5,211,211,107,211,214,184,187,5
+.byte	231,231,187,231,107,92,211,140,231,231,187,231,107,92,211,140
+.byte	110,110,165,110,87,203,220,57,110,110,165,110,87,203,220,57
+.byte	196,196,55,196,110,243,149,170,196,196,55,196,110,243,149,170
+.byte	3,3,12,3,24,15,6,27,3,3,12,3,24,15,6,27
+.byte	86,86,69,86,138,19,172,220,86,86,69,86,138,19,172,220
+.byte	68,68,13,68,26,73,136,94,68,68,13,68,26,73,136,94
+.byte	127,127,225,127,223,158,254,160,127,127,225,127,223,158,254,160
+.byte	169,169,158,169,33,55,79,136,169,169,158,169,33,55,79,136
+.byte	42,42,168,42,77,130,84,103,42,42,168,42,77,130,84,103
+.byte	187,187,214,187,177,109,107,10,187,187,214,187,177,109,107,10
+.byte	193,193,35,193,70,226,159,135,193,193,35,193,70,226,159,135
+.byte	83,83,81,83,162,2,166,241,83,83,81,83,162,2,166,241
+.byte	220,220,87,220,174,139,165,114,220,220,87,220,174,139,165,114
+.byte	11,11,44,11,88,39,22,83,11,11,44,11,88,39,22,83
+.byte	157,157,78,157,156,211,39,1,157,157,78,157,156,211,39,1
+.byte	108,108,173,108,71,193,216,43,108,108,173,108,71,193,216,43
+.byte	49,49,196,49,149,245,98,164,49,49,196,49,149,245,98,164
+.byte	116,116,205,116,135,185,232,243,116,116,205,116,135,185,232,243
+.byte	246,246,255,246,227,9,241,21,246,246,255,246,227,9,241,21
+.byte	70,70,5,70,10,67,140,76,70,70,5,70,10,67,140,76
+.byte	172,172,138,172,9,38,69,165,172,172,138,172,9,38,69,165
+.byte	137,137,30,137,60,151,15,181,137,137,30,137,60,151,15,181
+.byte	20,20,80,20,160,68,40,180,20,20,80,20,160,68,40,180
+.byte	225,225,163,225,91,66,223,186,225,225,163,225,91,66,223,186
+.byte	22,22,88,22,176,78,44,166,22,22,88,22,176,78,44,166
+.byte	58,58,232,58,205,210,116,247,58,58,232,58,205,210,116,247
+.byte	105,105,185,105,111,208,210,6,105,105,185,105,111,208,210,6
+.byte	9,9,36,9,72,45,18,65,9,9,36,9,72,45,18,65
+.byte	112,112,221,112,167,173,224,215,112,112,221,112,167,173,224,215
+.byte	182,182,226,182,217,84,113,111,182,182,226,182,217,84,113,111
+.byte	208,208,103,208,206,183,189,30,208,208,103,208,206,183,189,30
+.byte	237,237,147,237,59,126,199,214,237,237,147,237,59,126,199,214
+.byte	204,204,23,204,46,219,133,226,204,204,23,204,46,219,133,226
+.byte	66,66,21,66,42,87,132,104,66,66,21,66,42,87,132,104
+.byte	152,152,90,152,180,194,45,44,152,152,90,152,180,194,45,44
+.byte	164,164,170,164,73,14,85,237,164,164,170,164,73,14,85,237
+.byte	40,40,160,40,93,136,80,117,40,40,160,40,93,136,80,117
+.byte	92,92,109,92,218,49,184,134,92,92,109,92,218,49,184,134
+.byte	248,248,199,248,147,63,237,107,248,248,199,248,147,63,237,107
+.byte	134,134,34,134,68,164,17,194,134,134,34,134,68,164,17,194
+.byte	24,35,198,232,135,184,1,79
+.byte	54,166,210,245,121,111,145,82
+.byte	96,188,155,142,163,12,123,53
+.byte	29,224,215,194,46,75,254,87
+.byte	21,119,55,229,159,240,74,218
+.byte	88,201,41,10,177,160,107,133
+.byte	189,93,16,244,203,62,5,103
+.byte	228,39,65,139,167,125,149,216
+.byte	251,238,124,102,221,23,71,158
+.byte	202,45,191,7,173,90,131,51

deps/openssl/asm/x64-macosx-gas/x86_64cpuid.s

@@ -0,0 +1,195 @@
+
+.mod_init_func
+	.p2align	3
+	.quad	_OPENSSL_cpuid_setup
+
+.text
+
+
+.globl	_OPENSSL_atomic_add
+
+.p2align	4
+_OPENSSL_atomic_add:
+	movl	(%rdi),%eax
+L$spin:	leaq	(%rsi,%rax,1),%r8
+.byte	0xf0
+
+	cmpxchgl	%r8d,(%rdi)
+	jne	L$spin
+	movl	%r8d,%eax
+.byte	0x48,0x98
+
+	.byte	0xf3,0xc3
+
+
+.globl	_OPENSSL_rdtsc
+
+.p2align	4
+_OPENSSL_rdtsc:
+	rdtsc
+	shlq	$32,%rdx
+	orq	%rdx,%rax
+	.byte	0xf3,0xc3
+
+
+.globl	_OPENSSL_ia32_cpuid
+
+.p2align	4
+_OPENSSL_ia32_cpuid:
+	movq	%rbx,%r8
+
+	xorl	%eax,%eax
+	cpuid
+	movl	%eax,%r11d
+
+	xorl	%eax,%eax
+	cmpl	$1970169159,%ebx
+	setne	%al
+	movl	%eax,%r9d
+	cmpl	$1231384169,%edx
+	setne	%al
+	orl	%eax,%r9d
+	cmpl	$1818588270,%ecx
+	setne	%al
+	orl	%eax,%r9d
+	jz	L$intel
+
+	cmpl	$1752462657,%ebx
+	setne	%al
+	movl	%eax,%r10d
+	cmpl	$1769238117,%edx
+	setne	%al
+	orl	%eax,%r10d
+	cmpl	$1145913699,%ecx
+	setne	%al
+	orl	%eax,%r10d
+	jnz	L$intel
+
+
+	movl	$2147483648,%eax
+	cpuid
+	cmpl	$2147483656,%eax
+	jb	L$intel
+
+	movl	$2147483656,%eax
+	cpuid
+	movzbq	%cl,%r10
+	incq	%r10
+
+	movl	$1,%eax
+	cpuid
+	btl	$28,%edx
+	jnc	L$done
+	shrl	$16,%ebx
+	cmpb	%r10b,%bl
+	ja	L$done
+	andl	$4026531839,%edx
+	jmp	L$done
+
+L$intel:
+	cmpl	$4,%r11d
+	movl	$-1,%r10d
+	jb	L$nocacheinfo
+
+	movl	$4,%eax
+	movl	$0,%ecx
+	cpuid
+	movl	%eax,%r10d
+	shrl	$14,%r10d
+	andl	$4095,%r10d
+
+L$nocacheinfo:
+	movl	$1,%eax
+	cpuid
+	cmpl	$0,%r9d
+	jne	L$notintel
+	orl	$1048576,%edx
+	andb	$15,%ah
+	cmpb	$15,%ah
+	je	L$notintel
+	orl	$1073741824,%edx
+L$notintel:
+	btl	$28,%edx
+	jnc	L$done
+	andl	$4026531839,%edx
+	cmpl	$0,%r10d
+	je	L$done
+
+	orl	$268435456,%edx
+	shrl	$16,%ebx
+	cmpb	$1,%bl
+	ja	L$done
+	andl	$4026531839,%edx
+L$done:
+	shlq	$32,%rcx
+	movl	%edx,%eax
+	movq	%r8,%rbx
+	orq	%rcx,%rax
+	.byte	0xf3,0xc3
+
+
+.globl	_OPENSSL_cleanse
+
+.p2align	4
+_OPENSSL_cleanse:
+	xorq	%rax,%rax
+	cmpq	$15,%rsi
+	jae	L$ot
+	cmpq	$0,%rsi
+	je	L$ret
+L$ittle:
+	movb	%al,(%rdi)
+	subq	$1,%rsi
+	leaq	1(%rdi),%rdi
+	jnz	L$ittle
+L$ret:
+	.byte	0xf3,0xc3
+.p2align	4
+L$ot:
+	testq	$7,%rdi
+	jz	L$aligned
+	movb	%al,(%rdi)
+	leaq	-1(%rsi),%rsi
+	leaq	1(%rdi),%rdi
+	jmp	L$ot
+L$aligned:
+	movq	%rax,(%rdi)
+	leaq	-8(%rsi),%rsi
+	testq	$-8,%rsi
+	leaq	8(%rdi),%rdi
+	jnz	L$aligned
+	cmpq	$0,%rsi
+	jne	L$ittle
+	.byte	0xf3,0xc3
+
+.globl	_OPENSSL_wipe_cpu
+
+.p2align	4
+_OPENSSL_wipe_cpu:
+	pxor	%xmm0,%xmm0
+	pxor	%xmm1,%xmm1
+	pxor	%xmm2,%xmm2
+	pxor	%xmm3,%xmm3
+	pxor	%xmm4,%xmm4
+	pxor	%xmm5,%xmm5
+	pxor	%xmm6,%xmm6
+	pxor	%xmm7,%xmm7
+	pxor	%xmm8,%xmm8
+	pxor	%xmm9,%xmm9
+	pxor	%xmm10,%xmm10
+	pxor	%xmm11,%xmm11
+	pxor	%xmm12,%xmm12
+	pxor	%xmm13,%xmm13
+	pxor	%xmm14,%xmm14
+	pxor	%xmm15,%xmm15
+	xorq	%rcx,%rcx
+	xorq	%rdx,%rdx
+	xorq	%rsi,%rsi
+	xorq	%rdi,%rdi
+	xorq	%r8,%r8
+	xorq	%r9,%r9
+	xorq	%r10,%r10
+	xorq	%r11,%r11
+	leaq	8(%rsp),%rax
+	.byte	0xf3,0xc3
+

deps/openssl/asm/x64-win32-masm/bn/x86_64-mont.asm

@@ -0,0 +1,293 @@
+OPTION	DOTNAME
+.text$	SEGMENT ALIGN(64) 'CODE'
+
+PUBLIC	bn_mul_mont
+
+ALIGN	16
+bn_mul_mont	PROC PUBLIC
+	mov	QWORD PTR[8+rsp],rdi	;WIN64 prologue
+	mov	QWORD PTR[16+rsp],rsi
+	mov	rax,rsp
+$L$SEH_begin_bn_mul_mont::
+	mov	rdi,rcx
+	mov	rsi,rdx
+	mov	rdx,r8
+	mov	rcx,r9
+	mov	r8,QWORD PTR[40+rsp]
+	mov	r9,QWORD PTR[48+rsp]
+
+
+	push	rbx
+	push	rbp
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+
+	mov	r9d,r9d
+	lea	r10,QWORD PTR[2+r9]
+	mov	r11,rsp
+	neg	r10
+	lea	rsp,QWORD PTR[r10*8+rsp]
+	and	rsp,-1024
+
+	mov	QWORD PTR[8+r9*8+rsp],r11
+$L$prologue::
+	mov	r12,rdx
+
+	mov	r8,QWORD PTR[r8]
+
+	xor	r14,r14
+	xor	r15,r15
+
+	mov	rbx,QWORD PTR[r12]
+	mov	rax,QWORD PTR[rsi]
+	mul	rbx
+	mov	r10,rax
+	mov	r11,rdx
+
+	imul	rax,r8
+	mov	rbp,rax
+
+	mul	QWORD PTR[rcx]
+	add	rax,r10
+	adc	rdx,0
+	mov	r13,rdx
+
+	lea	r15,QWORD PTR[1+r15]
+$L$1st::
+	mov	rax,QWORD PTR[r15*8+rsi]
+	mul	rbx
+	add	rax,r11
+	adc	rdx,0
+	mov	r10,rax
+	mov	rax,QWORD PTR[r15*8+rcx]
+	mov	r11,rdx
+
+	mul	rbp
+	add	rax,r13
+	lea	r15,QWORD PTR[1+r15]
+	adc	rdx,0
+	add	rax,r10
+	adc	rdx,0
+	mov	QWORD PTR[((-16))+r15*8+rsp],rax
+	cmp	r15,r9
+	mov	r13,rdx
+	jl	$L$1st
+
+	xor	rdx,rdx
+	add	r13,r11
+	adc	rdx,0
+	mov	QWORD PTR[((-8))+r9*8+rsp],r13
+	mov	QWORD PTR[r9*8+rsp],rdx
+
+	lea	r14,QWORD PTR[1+r14]
+ALIGN	4
+$L$outer::
+	xor	r15,r15
+
+	mov	rbx,QWORD PTR[r14*8+r12]
+	mov	rax,QWORD PTR[rsi]
+	mul	rbx
+	add	rax,QWORD PTR[rsp]
+	adc	rdx,0
+	mov	r10,rax
+	mov	r11,rdx
+
+	imul	rax,r8
+	mov	rbp,rax
+
+	mul	QWORD PTR[r15*8+rcx]
+	add	rax,r10
+	mov	r10,QWORD PTR[8+rsp]
+	adc	rdx,0
+	mov	r13,rdx
+
+	lea	r15,QWORD PTR[1+r15]
+ALIGN	4
+$L$inner::
+	mov	rax,QWORD PTR[r15*8+rsi]
+	mul	rbx
+	add	rax,r11
+	adc	rdx,0
+	add	r10,rax
+	mov	rax,QWORD PTR[r15*8+rcx]
+	adc	rdx,0
+	mov	r11,rdx
+
+	mul	rbp
+	add	rax,r13
+	lea	r15,QWORD PTR[1+r15]
+	adc	rdx,0
+	add	rax,r10
+	adc	rdx,0
+	mov	r10,QWORD PTR[r15*8+rsp]
+	cmp	r15,r9
+	mov	QWORD PTR[((-16))+r15*8+rsp],rax
+	mov	r13,rdx
+	jl	$L$inner
+
+	xor	rdx,rdx
+	add	r13,r11
+	adc	rdx,0
+	add	r13,r10
+	adc	rdx,0
+	mov	QWORD PTR[((-8))+r9*8+rsp],r13
+	mov	QWORD PTR[r9*8+rsp],rdx
+
+	lea	r14,QWORD PTR[1+r14]
+	cmp	r14,r9
+	jl	$L$outer
+
+	lea	rsi,QWORD PTR[rsp]
+	lea	r15,QWORD PTR[((-1))+r9]
+
+	mov	rax,QWORD PTR[rsi]
+	xor	r14,r14
+	jmp	$L$sub
+ALIGN	16
+$L$sub::	sbb	rax,QWORD PTR[r14*8+rcx]
+	mov	QWORD PTR[r14*8+rdi],rax
+	dec	r15
+	mov	rax,QWORD PTR[8+r14*8+rsi]
+	lea	r14,QWORD PTR[1+r14]
+	jge	$L$sub
+
+	sbb	rax,0
+	and	rsi,rax
+	not	rax
+	mov	rcx,rdi
+	and	rcx,rax
+	lea	r15,QWORD PTR[((-1))+r9]
+	or	rsi,rcx
+ALIGN	16
+$L$copy::
+	mov	rax,QWORD PTR[r15*8+rsi]
+	mov	QWORD PTR[r15*8+rdi],rax
+	mov	QWORD PTR[r15*8+rsp],r14
+	dec	r15
+	jge	$L$copy
+
+	mov	rsi,QWORD PTR[8+r9*8+rsp]
+	mov	rax,1
+	mov	r15,QWORD PTR[rsi]
+	mov	r14,QWORD PTR[8+rsi]
+	mov	r13,QWORD PTR[16+rsi]
+	mov	r12,QWORD PTR[24+rsi]
+	mov	rbp,QWORD PTR[32+rsi]
+	mov	rbx,QWORD PTR[40+rsi]
+	lea	rsp,QWORD PTR[48+rsi]
+$L$epilogue::
+	mov	rdi,QWORD PTR[8+rsp]	;WIN64 epilogue
+	mov	rsi,QWORD PTR[16+rsp]
+	DB	0F3h,0C3h		;repret
+$L$SEH_end_bn_mul_mont::
+bn_mul_mont	ENDP
+DB	77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
+DB	112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
+DB	54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83
+DB	32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
+DB	115,108,46,111,114,103,62,0
+ALIGN	16
+EXTERN	__imp_RtlVirtualUnwind:NEAR
+
+ALIGN	16
+se_handler	PROC PRIVATE
+	push	rsi
+	push	rdi
+	push	rbx
+	push	rbp
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+	pushfq
+	sub	rsp,64
+
+	mov	rax,QWORD PTR[120+r8]
+	mov	rbx,QWORD PTR[248+r8]
+
+	lea	r10,QWORD PTR[$L$prologue]
+	cmp	rbx,r10
+	jb	$L$in_prologue
+
+	mov	rax,QWORD PTR[152+r8]
+
+	lea	r10,QWORD PTR[$L$epilogue]
+	cmp	rbx,r10
+	jae	$L$in_prologue
+
+	mov	r10,QWORD PTR[192+r8]
+	mov	rax,QWORD PTR[8+r10*8+rax]
+	lea	rax,QWORD PTR[48+rax]
+
+	mov	rbx,QWORD PTR[((-8))+rax]
+	mov	rbp,QWORD PTR[((-16))+rax]
+	mov	r12,QWORD PTR[((-24))+rax]
+	mov	r13,QWORD PTR[((-32))+rax]
+	mov	r14,QWORD PTR[((-40))+rax]
+	mov	r15,QWORD PTR[((-48))+rax]
+	mov	QWORD PTR[144+r8],rbx
+	mov	QWORD PTR[160+r8],rbp
+	mov	QWORD PTR[216+r8],r12
+	mov	QWORD PTR[224+r8],r13
+	mov	QWORD PTR[232+r8],r14
+	mov	QWORD PTR[240+r8],r15
+
+$L$in_prologue::
+	mov	rdi,QWORD PTR[8+rax]
+	mov	rsi,QWORD PTR[16+rax]
+	mov	QWORD PTR[152+r8],rax
+	mov	QWORD PTR[168+r8],rsi
+	mov	QWORD PTR[176+r8],rdi
+
+	mov	rdi,QWORD PTR[40+r9]
+	mov	rsi,r8
+	mov	ecx,154
+	DD	0a548f3fch
+
+
+	mov	rsi,r9
+	xor	rcx,rcx
+	mov	rdx,QWORD PTR[8+rsi]
+	mov	r8,QWORD PTR[rsi]
+	mov	r9,QWORD PTR[16+rsi]
+	mov	r10,QWORD PTR[40+rsi]
+	lea	r11,QWORD PTR[56+rsi]
+	lea	r12,QWORD PTR[24+rsi]
+	mov	QWORD PTR[32+rsp],r10
+	mov	QWORD PTR[40+rsp],r11
+	mov	QWORD PTR[48+rsp],r12
+	mov	QWORD PTR[56+rsp],rcx
+	call	QWORD PTR[__imp_RtlVirtualUnwind]
+
+	mov	eax,1
+	add	rsp,64
+	popfq
+	pop	r15
+	pop	r14
+	pop	r13
+	pop	r12
+	pop	rbp
+	pop	rbx
+	pop	rdi
+	pop	rsi
+	DB	0F3h,0C3h		;repret
+se_handler	ENDP
+
+.text$	ENDS
+.pdata	SEGMENT READONLY ALIGN(4)
+ALIGN	4
+	DD	imagerel $L$SEH_begin_bn_mul_mont
+	DD	imagerel $L$SEH_end_bn_mul_mont
+	DD	imagerel $L$SEH_info_bn_mul_mont
+
+.pdata	ENDS
+.xdata	SEGMENT READONLY ALIGN(8)
+ALIGN	8
+$L$SEH_info_bn_mul_mont::
+DB	9,0,0,0
+	DD	imagerel se_handler
+
+.xdata	ENDS
+END

deps/openssl/asm/x64-win32-masm/md5/md5-x86_64.asm

@@ -0,0 +1,781 @@
+OPTION	DOTNAME
+.text$	SEGMENT ALIGN(64) 'CODE'
+ALIGN	16
+
+PUBLIC	md5_block_asm_data_order
+
+md5_block_asm_data_order	PROC PUBLIC
+	mov	QWORD PTR[8+rsp],rdi	;WIN64 prologue
+	mov	QWORD PTR[16+rsp],rsi
+	mov	rax,rsp
+$L$SEH_begin_md5_block_asm_data_order::
+	mov	rdi,rcx
+	mov	rsi,rdx
+	mov	rdx,r8
+
+
+	push	rbp
+	push	rbx
+	push	r12
+	push	r14
+	push	r15
+$L$prologue::
+
+
+
+
+	mov	rbp,rdi
+	shl	rdx,6
+	lea	rdi,QWORD PTR[rdx*1+rsi]
+	mov	eax,DWORD PTR[((0*4))+rbp]
+	mov	ebx,DWORD PTR[((1*4))+rbp]
+	mov	ecx,DWORD PTR[((2*4))+rbp]
+	mov	edx,DWORD PTR[((3*4))+rbp]
+
+
+
+
+
+
+
+	cmp	rsi,rdi
+	je	$L$end
+
+
+
+$L$loop::
+	mov	r8d,eax
+	mov	r9d,ebx
+	mov	r14d,ecx
+	mov	r15d,edx
+	mov	r10d,DWORD PTR[((0*4))+rsi]
+	mov	r11d,edx
+	xor	r11d,ecx
+	lea	eax,DWORD PTR[0d76aa478h+r10*1+rax]
+	and	r11d,ebx
+	xor	r11d,edx
+	mov	r10d,DWORD PTR[((1*4))+rsi]
+	add	eax,r11d
+	rol	eax,7
+	mov	r11d,ecx
+	add	eax,ebx
+	xor	r11d,ebx
+	lea	edx,DWORD PTR[0e8c7b756h+r10*1+rdx]
+	and	r11d,eax
+	xor	r11d,ecx
+	mov	r10d,DWORD PTR[((2*4))+rsi]
+	add	edx,r11d
+	rol	edx,12
+	mov	r11d,ebx
+	add	edx,eax
+	xor	r11d,eax
+	lea	ecx,DWORD PTR[0242070dbh+r10*1+rcx]
+	and	r11d,edx
+	xor	r11d,ebx
+	mov	r10d,DWORD PTR[((3*4))+rsi]
+	add	ecx,r11d
+	rol	ecx,17
+	mov	r11d,eax
+	add	ecx,edx
+	xor	r11d,edx
+	lea	ebx,DWORD PTR[0c1bdceeeh+r10*1+rbx]
+	and	r11d,ecx
+	xor	r11d,eax
+	mov	r10d,DWORD PTR[((4*4))+rsi]
+	add	ebx,r11d
+	rol	ebx,22
+	mov	r11d,edx
+	add	ebx,ecx
+	xor	r11d,ecx
+	lea	eax,DWORD PTR[0f57c0fafh+r10*1+rax]
+	and	r11d,ebx
+	xor	r11d,edx
+	mov	r10d,DWORD PTR[((5*4))+rsi]
+	add	eax,r11d
+	rol	eax,7
+	mov	r11d,ecx
+	add	eax,ebx
+	xor	r11d,ebx
+	lea	edx,DWORD PTR[04787c62ah+r10*1+rdx]
+	and	r11d,eax
+	xor	r11d,ecx
+	mov	r10d,DWORD PTR[((6*4))+rsi]
+	add	edx,r11d
+	rol	edx,12
+	mov	r11d,ebx
+	add	edx,eax
+	xor	r11d,eax
+	lea	ecx,DWORD PTR[0a8304613h+r10*1+rcx]
+	and	r11d,edx
+	xor	r11d,ebx
+	mov	r10d,DWORD PTR[((7*4))+rsi]
+	add	ecx,r11d
+	rol	ecx,17
+	mov	r11d,eax
+	add	ecx,edx
+	xor	r11d,edx
+	lea	ebx,DWORD PTR[0fd469501h+r10*1+rbx]
+	and	r11d,ecx
+	xor	r11d,eax
+	mov	r10d,DWORD PTR[((8*4))+rsi]
+	add	ebx,r11d
+	rol	ebx,22
+	mov	r11d,edx
+	add	ebx,ecx
+	xor	r11d,ecx
+	lea	eax,DWORD PTR[0698098d8h+r10*1+rax]
+	and	r11d,ebx
+	xor	r11d,edx
+	mov	r10d,DWORD PTR[((9*4))+rsi]
+	add	eax,r11d
+	rol	eax,7
+	mov	r11d,ecx
+	add	eax,ebx
+	xor	r11d,ebx
+	lea	edx,DWORD PTR[08b44f7afh+r10*1+rdx]
+	and	r11d,eax
+	xor	r11d,ecx
+	mov	r10d,DWORD PTR[((10*4))+rsi]
+	add	edx,r11d
+	rol	edx,12
+	mov	r11d,ebx
+	add	edx,eax
+	xor	r11d,eax
+	lea	ecx,DWORD PTR[0ffff5bb1h+r10*1+rcx]
+	and	r11d,edx
+	xor	r11d,ebx
+	mov	r10d,DWORD PTR[((11*4))+rsi]
+	add	ecx,r11d
+	rol	ecx,17
+	mov	r11d,eax
+	add	ecx,edx
+	xor	r11d,edx
+	lea	ebx,DWORD PTR[0895cd7beh+r10*1+rbx]
+	and	r11d,ecx
+	xor	r11d,eax
+	mov	r10d,DWORD PTR[((12*4))+rsi]
+	add	ebx,r11d
+	rol	ebx,22
+	mov	r11d,edx
+	add	ebx,ecx
+	xor	r11d,ecx
+	lea	eax,DWORD PTR[06b901122h+r10*1+rax]
+	and	r11d,ebx
+	xor	r11d,edx
+	mov	r10d,DWORD PTR[((13*4))+rsi]
+	add	eax,r11d
+	rol	eax,7
+	mov	r11d,ecx
+	add	eax,ebx
+	xor	r11d,ebx
+	lea	edx,DWORD PTR[0fd987193h+r10*1+rdx]
+	and	r11d,eax
+	xor	r11d,ecx
+	mov	r10d,DWORD PTR[((14*4))+rsi]
+	add	edx,r11d
+	rol	edx,12
+	mov	r11d,ebx
+	add	edx,eax
+	xor	r11d,eax
+	lea	ecx,DWORD PTR[0a679438eh+r10*1+rcx]
+	and	r11d,edx
+	xor	r11d,ebx
+	mov	r10d,DWORD PTR[((15*4))+rsi]
+	add	ecx,r11d
+	rol	ecx,17
+	mov	r11d,eax
+	add	ecx,edx
+	xor	r11d,edx
+	lea	ebx,DWORD PTR[049b40821h+r10*1+rbx]
+	and	r11d,ecx
+	xor	r11d,eax
+	mov	r10d,DWORD PTR[((0*4))+rsi]
+	add	ebx,r11d
+	rol	ebx,22
+	mov	r11d,edx
+	add	ebx,ecx
+	mov	r10d,DWORD PTR[((1*4))+rsi]
+	mov	r11d,edx
+	mov	r12d,edx
+	not	r11d
+	lea	eax,DWORD PTR[0f61e2562h+r10*1+rax]
+	and	r12d,ebx
+	and	r11d,ecx
+	mov	r10d,DWORD PTR[((6*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,ecx
+	add	eax,r12d
+	mov	r12d,ecx
+	rol	eax,5
+	add	eax,ebx
+	not	r11d
+	lea	edx,DWORD PTR[0c040b340h+r10*1+rdx]
+	and	r12d,eax
+	and	r11d,ebx
+	mov	r10d,DWORD PTR[((11*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,ebx
+	add	edx,r12d
+	mov	r12d,ebx
+	rol	edx,9
+	add	edx,eax
+	not	r11d
+	lea	ecx,DWORD PTR[0265e5a51h+r10*1+rcx]
+	and	r12d,edx
+	and	r11d,eax
+	mov	r10d,DWORD PTR[((0*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,eax
+	add	ecx,r12d
+	mov	r12d,eax
+	rol	ecx,14
+	add	ecx,edx
+	not	r11d
+	lea	ebx,DWORD PTR[0e9b6c7aah+r10*1+rbx]
+	and	r12d,ecx
+	and	r11d,edx
+	mov	r10d,DWORD PTR[((5*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,edx
+	add	ebx,r12d
+	mov	r12d,edx
+	rol	ebx,20
+	add	ebx,ecx
+	not	r11d
+	lea	eax,DWORD PTR[0d62f105dh+r10*1+rax]
+	and	r12d,ebx
+	and	r11d,ecx
+	mov	r10d,DWORD PTR[((10*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,ecx
+	add	eax,r12d
+	mov	r12d,ecx
+	rol	eax,5
+	add	eax,ebx
+	not	r11d
+	lea	edx,DWORD PTR[02441453h+r10*1+rdx]
+	and	r12d,eax
+	and	r11d,ebx
+	mov	r10d,DWORD PTR[((15*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,ebx
+	add	edx,r12d
+	mov	r12d,ebx
+	rol	edx,9
+	add	edx,eax
+	not	r11d
+	lea	ecx,DWORD PTR[0d8a1e681h+r10*1+rcx]
+	and	r12d,edx
+	and	r11d,eax
+	mov	r10d,DWORD PTR[((4*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,eax
+	add	ecx,r12d
+	mov	r12d,eax
+	rol	ecx,14
+	add	ecx,edx
+	not	r11d
+	lea	ebx,DWORD PTR[0e7d3fbc8h+r10*1+rbx]
+	and	r12d,ecx
+	and	r11d,edx
+	mov	r10d,DWORD PTR[((9*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,edx
+	add	ebx,r12d
+	mov	r12d,edx
+	rol	ebx,20
+	add	ebx,ecx
+	not	r11d
+	lea	eax,DWORD PTR[021e1cde6h+r10*1+rax]
+	and	r12d,ebx
+	and	r11d,ecx
+	mov	r10d,DWORD PTR[((14*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,ecx
+	add	eax,r12d
+	mov	r12d,ecx
+	rol	eax,5
+	add	eax,ebx
+	not	r11d
+	lea	edx,DWORD PTR[0c33707d6h+r10*1+rdx]
+	and	r12d,eax
+	and	r11d,ebx
+	mov	r10d,DWORD PTR[((3*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,ebx
+	add	edx,r12d
+	mov	r12d,ebx
+	rol	edx,9
+	add	edx,eax
+	not	r11d
+	lea	ecx,DWORD PTR[0f4d50d87h+r10*1+rcx]
+	and	r12d,edx
+	and	r11d,eax
+	mov	r10d,DWORD PTR[((8*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,eax
+	add	ecx,r12d
+	mov	r12d,eax
+	rol	ecx,14
+	add	ecx,edx
+	not	r11d
+	lea	ebx,DWORD PTR[0455a14edh+r10*1+rbx]
+	and	r12d,ecx
+	and	r11d,edx
+	mov	r10d,DWORD PTR[((13*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,edx
+	add	ebx,r12d
+	mov	r12d,edx
+	rol	ebx,20
+	add	ebx,ecx
+	not	r11d
+	lea	eax,DWORD PTR[0a9e3e905h+r10*1+rax]
+	and	r12d,ebx
+	and	r11d,ecx
+	mov	r10d,DWORD PTR[((2*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,ecx
+	add	eax,r12d
+	mov	r12d,ecx
+	rol	eax,5
+	add	eax,ebx
+	not	r11d
+	lea	edx,DWORD PTR[0fcefa3f8h+r10*1+rdx]
+	and	r12d,eax
+	and	r11d,ebx
+	mov	r10d,DWORD PTR[((7*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,ebx
+	add	edx,r12d
+	mov	r12d,ebx
+	rol	edx,9
+	add	edx,eax
+	not	r11d
+	lea	ecx,DWORD PTR[0676f02d9h+r10*1+rcx]
+	and	r12d,edx
+	and	r11d,eax
+	mov	r10d,DWORD PTR[((12*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,eax
+	add	ecx,r12d
+	mov	r12d,eax
+	rol	ecx,14
+	add	ecx,edx
+	not	r11d
+	lea	ebx,DWORD PTR[08d2a4c8ah+r10*1+rbx]
+	and	r12d,ecx
+	and	r11d,edx
+	mov	r10d,DWORD PTR[((0*4))+rsi]
+	or	r12d,r11d
+	mov	r11d,edx
+	add	ebx,r12d
+	mov	r12d,edx
+	rol	ebx,20
+	add	ebx,ecx
+	mov	r10d,DWORD PTR[((5*4))+rsi]
+	mov	r11d,ecx
+	lea	eax,DWORD PTR[0fffa3942h+r10*1+rax]
+	mov	r10d,DWORD PTR[((8*4))+rsi]
+	xor	r11d,edx
+	xor	r11d,ebx
+	add	eax,r11d
+	rol	eax,4
+	mov	r11d,ebx
+	add	eax,ebx
+	lea	edx,DWORD PTR[08771f681h+r10*1+rdx]
+	mov	r10d,DWORD PTR[((11*4))+rsi]
+	xor	r11d,ecx
+	xor	r11d,eax
+	add	edx,r11d
+	rol	edx,11
+	mov	r11d,eax
+	add	edx,eax
+	lea	ecx,DWORD PTR[06d9d6122h+r10*1+rcx]
+	mov	r10d,DWORD PTR[((14*4))+rsi]
+	xor	r11d,ebx
+	xor	r11d,edx
+	add	ecx,r11d
+	rol	ecx,16
+	mov	r11d,edx
+	add	ecx,edx
+	lea	ebx,DWORD PTR[0fde5380ch+r10*1+rbx]
+	mov	r10d,DWORD PTR[((1*4))+rsi]
+	xor	r11d,eax
+	xor	r11d,ecx
+	add	ebx,r11d
+	rol	ebx,23
+	mov	r11d,ecx
+	add	ebx,ecx
+	lea	eax,DWORD PTR[0a4beea44h+r10*1+rax]
+	mov	r10d,DWORD PTR[((4*4))+rsi]
+	xor	r11d,edx
+	xor	r11d,ebx
+	add	eax,r11d
+	rol	eax,4
+	mov	r11d,ebx
+	add	eax,ebx
+	lea	edx,DWORD PTR[04bdecfa9h+r10*1+rdx]
+	mov	r10d,DWORD PTR[((7*4))+rsi]
+	xor	r11d,ecx
+	xor	r11d,eax
+	add	edx,r11d
+	rol	edx,11
+	mov	r11d,eax
+	add	edx,eax
+	lea	ecx,DWORD PTR[0f6bb4b60h+r10*1+rcx]
+	mov	r10d,DWORD PTR[((10*4))+rsi]
+	xor	r11d,ebx
+	xor	r11d,edx
+	add	ecx,r11d
+	rol	ecx,16
+	mov	r11d,edx
+	add	ecx,edx
+	lea	ebx,DWORD PTR[0bebfbc70h+r10*1+rbx]
+	mov	r10d,DWORD PTR[((13*4))+rsi]
+	xor	r11d,eax
+	xor	r11d,ecx
+	add	ebx,r11d
+	rol	ebx,23
+	mov	r11d,ecx
+	add	ebx,ecx
+	lea	eax,DWORD PTR[0289b7ec6h+r10*1+rax]
+	mov	r10d,DWORD PTR[((0*4))+rsi]
+	xor	r11d,edx
+	xor	r11d,ebx
+	add	eax,r11d
+	rol	eax,4
+	mov	r11d,ebx
+	add	eax,ebx
+	lea	edx,DWORD PTR[0eaa127fah+r10*1+rdx]
+	mov	r10d,DWORD PTR[((3*4))+rsi]
+	xor	r11d,ecx
+	xor	r11d,eax
+	add	edx,r11d
+	rol	edx,11
+	mov	r11d,eax
+	add	edx,eax
+	lea	ecx,DWORD PTR[0d4ef3085h+r10*1+rcx]
+	mov	r10d,DWORD PTR[((6*4))+rsi]
+	xor	r11d,ebx
+	xor	r11d,edx
+	add	ecx,r11d
+	rol	ecx,16
+	mov	r11d,edx
+	add	ecx,edx
+	lea	ebx,DWORD PTR[04881d05h+r10*1+rbx]
+	mov	r10d,DWORD PTR[((9*4))+rsi]
+	xor	r11d,eax
+	xor	r11d,ecx
+	add	ebx,r11d
+	rol	ebx,23
+	mov	r11d,ecx
+	add	ebx,ecx
+	lea	eax,DWORD PTR[0d9d4d039h+r10*1+rax]
+	mov	r10d,DWORD PTR[((12*4))+rsi]
+	xor	r11d,edx
+	xor	r11d,ebx
+	add	eax,r11d
+	rol	eax,4
+	mov	r11d,ebx
+	add	eax,ebx
+	lea	edx,DWORD PTR[0e6db99e5h+r10*1+rdx]
+	mov	r10d,DWORD PTR[((15*4))+rsi]
+	xor	r11d,ecx
+	xor	r11d,eax
+	add	edx,r11d
+	rol	edx,11
+	mov	r11d,eax
+	add	edx,eax
+	lea	ecx,DWORD PTR[01fa27cf8h+r10*1+rcx]
+	mov	r10d,DWORD PTR[((2*4))+rsi]
+	xor	r11d,ebx
+	xor	r11d,edx
+	add	ecx,r11d
+	rol	ecx,16
+	mov	r11d,edx
+	add	ecx,edx
+	lea	ebx,DWORD PTR[0c4ac5665h+r10*1+rbx]
+	mov	r10d,DWORD PTR[((0*4))+rsi]
+	xor	r11d,eax
+	xor	r11d,ecx
+	add	ebx,r11d
+	rol	ebx,23
+	mov	r11d,ecx
+	add	ebx,ecx
+	mov	r10d,DWORD PTR[((0*4))+rsi]
+	mov	r11d,0ffffffffh
+	xor	r11d,edx
+	lea	eax,DWORD PTR[0f4292244h+r10*1+rax]
+	or	r11d,ebx
+	xor	r11d,ecx
+	add	eax,r11d
+	mov	r10d,DWORD PTR[((7*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	eax,6
+	xor	r11d,ecx
+	add	eax,ebx
+	lea	edx,DWORD PTR[0432aff97h+r10*1+rdx]
+	or	r11d,eax
+	xor	r11d,ebx
+	add	edx,r11d
+	mov	r10d,DWORD PTR[((14*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	edx,10
+	xor	r11d,ebx
+	add	edx,eax
+	lea	ecx,DWORD PTR[0ab9423a7h+r10*1+rcx]
+	or	r11d,edx
+	xor	r11d,eax
+	add	ecx,r11d
+	mov	r10d,DWORD PTR[((5*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	ecx,15
+	xor	r11d,eax
+	add	ecx,edx
+	lea	ebx,DWORD PTR[0fc93a039h+r10*1+rbx]
+	or	r11d,ecx
+	xor	r11d,edx
+	add	ebx,r11d
+	mov	r10d,DWORD PTR[((12*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	ebx,21
+	xor	r11d,edx
+	add	ebx,ecx
+	lea	eax,DWORD PTR[0655b59c3h+r10*1+rax]
+	or	r11d,ebx
+	xor	r11d,ecx
+	add	eax,r11d
+	mov	r10d,DWORD PTR[((3*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	eax,6
+	xor	r11d,ecx
+	add	eax,ebx
+	lea	edx,DWORD PTR[08f0ccc92h+r10*1+rdx]
+	or	r11d,eax
+	xor	r11d,ebx
+	add	edx,r11d
+	mov	r10d,DWORD PTR[((10*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	edx,10
+	xor	r11d,ebx
+	add	edx,eax
+	lea	ecx,DWORD PTR[0ffeff47dh+r10*1+rcx]
+	or	r11d,edx
+	xor	r11d,eax
+	add	ecx,r11d
+	mov	r10d,DWORD PTR[((1*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	ecx,15
+	xor	r11d,eax
+	add	ecx,edx
+	lea	ebx,DWORD PTR[085845dd1h+r10*1+rbx]
+	or	r11d,ecx
+	xor	r11d,edx
+	add	ebx,r11d
+	mov	r10d,DWORD PTR[((8*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	ebx,21
+	xor	r11d,edx
+	add	ebx,ecx
+	lea	eax,DWORD PTR[06fa87e4fh+r10*1+rax]
+	or	r11d,ebx
+	xor	r11d,ecx
+	add	eax,r11d
+	mov	r10d,DWORD PTR[((15*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	eax,6
+	xor	r11d,ecx
+	add	eax,ebx
+	lea	edx,DWORD PTR[0fe2ce6e0h+r10*1+rdx]
+	or	r11d,eax
+	xor	r11d,ebx
+	add	edx,r11d
+	mov	r10d,DWORD PTR[((6*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	edx,10
+	xor	r11d,ebx
+	add	edx,eax
+	lea	ecx,DWORD PTR[0a3014314h+r10*1+rcx]
+	or	r11d,edx
+	xor	r11d,eax
+	add	ecx,r11d
+	mov	r10d,DWORD PTR[((13*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	ecx,15
+	xor	r11d,eax
+	add	ecx,edx
+	lea	ebx,DWORD PTR[04e0811a1h+r10*1+rbx]
+	or	r11d,ecx
+	xor	r11d,edx
+	add	ebx,r11d
+	mov	r10d,DWORD PTR[((4*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	ebx,21
+	xor	r11d,edx
+	add	ebx,ecx
+	lea	eax,DWORD PTR[0f7537e82h+r10*1+rax]
+	or	r11d,ebx
+	xor	r11d,ecx
+	add	eax,r11d
+	mov	r10d,DWORD PTR[((11*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	eax,6
+	xor	r11d,ecx
+	add	eax,ebx
+	lea	edx,DWORD PTR[0bd3af235h+r10*1+rdx]
+	or	r11d,eax
+	xor	r11d,ebx
+	add	edx,r11d
+	mov	r10d,DWORD PTR[((2*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	edx,10
+	xor	r11d,ebx
+	add	edx,eax
+	lea	ecx,DWORD PTR[02ad7d2bbh+r10*1+rcx]
+	or	r11d,edx
+	xor	r11d,eax
+	add	ecx,r11d
+	mov	r10d,DWORD PTR[((9*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	ecx,15
+	xor	r11d,eax
+	add	ecx,edx
+	lea	ebx,DWORD PTR[0eb86d391h+r10*1+rbx]
+	or	r11d,ecx
+	xor	r11d,edx
+	add	ebx,r11d
+	mov	r10d,DWORD PTR[((0*4))+rsi]
+	mov	r11d,0ffffffffh
+	rol	ebx,21
+	xor	r11d,edx
+	add	ebx,ecx
+
+	add	eax,r8d
+	add	ebx,r9d
+	add	ecx,r14d
+	add	edx,r15d
+
+
+	add	rsi,64
+	cmp	rsi,rdi
+	jb	$L$loop
+
+
+
+$L$end::
+	mov	DWORD PTR[((0*4))+rbp],eax
+	mov	DWORD PTR[((1*4))+rbp],ebx
+	mov	DWORD PTR[((2*4))+rbp],ecx
+	mov	DWORD PTR[((3*4))+rbp],edx
+
+	mov	r15,QWORD PTR[rsp]
+	mov	r14,QWORD PTR[8+rsp]
+	mov	r12,QWORD PTR[16+rsp]
+	mov	rbx,QWORD PTR[24+rsp]
+	mov	rbp,QWORD PTR[32+rsp]
+	add	rsp,40
+$L$epilogue::
+	mov	rdi,QWORD PTR[8+rsp]	;WIN64 epilogue
+	mov	rsi,QWORD PTR[16+rsp]
+	DB	0F3h,0C3h		;repret
+$L$SEH_end_md5_block_asm_data_order::
+md5_block_asm_data_order	ENDP
+EXTERN	__imp_RtlVirtualUnwind:NEAR
+
+ALIGN	16
+se_handler	PROC PRIVATE
+	push	rsi
+	push	rdi
+	push	rbx
+	push	rbp
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+	pushfq
+	sub	rsp,64
+
+	mov	rax,QWORD PTR[120+r8]
+	mov	rbx,QWORD PTR[248+r8]
+
+	lea	r10,QWORD PTR[$L$prologue]
+	cmp	rbx,r10
+	jb	$L$in_prologue
+
+	mov	rax,QWORD PTR[152+r8]
+
+	lea	r10,QWORD PTR[$L$epilogue]
+	cmp	rbx,r10
+	jae	$L$in_prologue
+
+	lea	rax,QWORD PTR[40+rax]
+
+	mov	rbp,QWORD PTR[((-8))+rax]
+	mov	rbx,QWORD PTR[((-16))+rax]
+	mov	r12,QWORD PTR[((-24))+rax]
+	mov	r14,QWORD PTR[((-32))+rax]
+	mov	r15,QWORD PTR[((-40))+rax]
+	mov	QWORD PTR[144+r8],rbx
+	mov	QWORD PTR[160+r8],rbp
+	mov	QWORD PTR[216+r8],r12
+	mov	QWORD PTR[232+r8],r14
+	mov	QWORD PTR[240+r8],r15
+
+$L$in_prologue::
+	mov	rdi,QWORD PTR[8+rax]
+	mov	rsi,QWORD PTR[16+rax]
+	mov	QWORD PTR[152+r8],rax
+	mov	QWORD PTR[168+r8],rsi
+	mov	QWORD PTR[176+r8],rdi
+
+	mov	rdi,QWORD PTR[40+r9]
+	mov	rsi,r8
+	mov	ecx,154
+	DD	0a548f3fch
+
+
+	mov	rsi,r9
+	xor	rcx,rcx
+	mov	rdx,QWORD PTR[8+rsi]
+	mov	r8,QWORD PTR[rsi]
+	mov	r9,QWORD PTR[16+rsi]
+	mov	r10,QWORD PTR[40+rsi]
+	lea	r11,QWORD PTR[56+rsi]
+	lea	r12,QWORD PTR[24+rsi]
+	mov	QWORD PTR[32+rsp],r10
+	mov	QWORD PTR[40+rsp],r11
+	mov	QWORD PTR[48+rsp],r12
+	mov	QWORD PTR[56+rsp],rcx
+	call	QWORD PTR[__imp_RtlVirtualUnwind]
+
+	mov	eax,1
+	add	rsp,64
+	popfq
+	pop	r15
+	pop	r14
+	pop	r13
+	pop	r12
+	pop	rbp
+	pop	rbx
+	pop	rdi
+	pop	rsi
+	DB	0F3h,0C3h		;repret
+se_handler	ENDP
+
+.text$	ENDS
+.pdata	SEGMENT READONLY ALIGN(4)
+ALIGN	4
+	DD	imagerel $L$SEH_begin_md5_block_asm_data_order
+	DD	imagerel $L$SEH_end_md5_block_asm_data_order
+	DD	imagerel $L$SEH_info_md5_block_asm_data_order
+
+.pdata	ENDS
+.xdata	SEGMENT READONLY ALIGN(8)
+ALIGN	8
+$L$SEH_info_md5_block_asm_data_order::
+DB	9,0,0,0
+	DD	imagerel se_handler
+
+.xdata	ENDS
+END

deps/openssl/asm/x64-win32-masm/rc4/rc4-x86_64.asm

@@ -0,0 +1,586 @@
+OPTION	DOTNAME
+.text$	SEGMENT ALIGN(64) 'CODE'
+
+PUBLIC	RC4
+
+ALIGN	16
+RC4	PROC PUBLIC
+	mov	QWORD PTR[8+rsp],rdi	;WIN64 prologue
+	mov	QWORD PTR[16+rsp],rsi
+	mov	rax,rsp
+$L$SEH_begin_RC4::
+	mov	rdi,rcx
+	mov	rsi,rdx
+	mov	rdx,r8
+	mov	rcx,r9
+
+	or	rsi,rsi
+	jne	$L$entry
+	mov	rdi,QWORD PTR[8+rsp]	;WIN64 epilogue
+	mov	rsi,QWORD PTR[16+rsp]
+	DB	0F3h,0C3h		;repret
+$L$entry::
+	push	rbx
+	push	r12
+	push	r13
+$L$prologue::
+
+	add	rdi,8
+	mov	r8d,DWORD PTR[((-8))+rdi]
+	mov	r12d,DWORD PTR[((-4))+rdi]
+	cmp	DWORD PTR[256+rdi],-1
+	je	$L$RC4_CHAR
+	inc	r8b
+	mov	r9d,DWORD PTR[r8*4+rdi]
+	test	rsi,-8
+	jz	$L$loop1
+	jmp	$L$loop8
+ALIGN	16
+$L$loop8::
+	add	r12b,r9b
+	mov	r10,r8
+	mov	r13d,DWORD PTR[r12*4+rdi]
+	ror	rax,8
+	inc	r10b
+	mov	r11d,DWORD PTR[r10*4+rdi]
+	cmp	r12,r10
+	mov	DWORD PTR[r12*4+rdi],r9d
+	cmove	r11,r9
+	mov	DWORD PTR[r8*4+rdi],r13d
+	add	r13b,r9b
+	mov	al,BYTE PTR[r13*4+rdi]
+	add	r12b,r11b
+	mov	r8,r10
+	mov	r13d,DWORD PTR[r12*4+rdi]
+	ror	rax,8
+	inc	r8b
+	mov	r9d,DWORD PTR[r8*4+rdi]
+	cmp	r12,r8
+	mov	DWORD PTR[r12*4+rdi],r11d
+	cmove	r9,r11
+	mov	DWORD PTR[r10*4+rdi],r13d
+	add	r13b,r11b
+	mov	al,BYTE PTR[r13*4+rdi]
+	add	r12b,r9b
+	mov	r10,r8
+	mov	r13d,DWORD PTR[r12*4+rdi]
+	ror	rax,8
+	inc	r10b
+	mov	r11d,DWORD PTR[r10*4+rdi]
+	cmp	r12,r10
+	mov	DWORD PTR[r12*4+rdi],r9d
+	cmove	r11,r9
+	mov	DWORD PTR[r8*4+rdi],r13d
+	add	r13b,r9b
+	mov	al,BYTE PTR[r13*4+rdi]
+	add	r12b,r11b
+	mov	r8,r10
+	mov	r13d,DWORD PTR[r12*4+rdi]
+	ror	rax,8
+	inc	r8b
+	mov	r9d,DWORD PTR[r8*4+rdi]
+	cmp	r12,r8
+	mov	DWORD PTR[r12*4+rdi],r11d
+	cmove	r9,r11
+	mov	DWORD PTR[r10*4+rdi],r13d
+	add	r13b,r11b
+	mov	al,BYTE PTR[r13*4+rdi]
+	add	r12b,r9b
+	mov	r10,r8
+	mov	r13d,DWORD PTR[r12*4+rdi]
+	ror	rax,8
+	inc	r10b
+	mov	r11d,DWORD PTR[r10*4+rdi]
+	cmp	r12,r10
+	mov	DWORD PTR[r12*4+rdi],r9d
+	cmove	r11,r9
+	mov	DWORD PTR[r8*4+rdi],r13d
+	add	r13b,r9b
+	mov	al,BYTE PTR[r13*4+rdi]
+	add	r12b,r11b
+	mov	r8,r10
+	mov	r13d,DWORD PTR[r12*4+rdi]
+	ror	rax,8
+	inc	r8b
+	mov	r9d,DWORD PTR[r8*4+rdi]
+	cmp	r12,r8
+	mov	DWORD PTR[r12*4+rdi],r11d
+	cmove	r9,r11
+	mov	DWORD PTR[r10*4+rdi],r13d
+	add	r13b,r11b
+	mov	al,BYTE PTR[r13*4+rdi]
+	add	r12b,r9b
+	mov	r10,r8
+	mov	r13d,DWORD PTR[r12*4+rdi]
+	ror	rax,8
+	inc	r10b
+	mov	r11d,DWORD PTR[r10*4+rdi]
+	cmp	r12,r10
+	mov	DWORD PTR[r12*4+rdi],r9d
+	cmove	r11,r9
+	mov	DWORD PTR[r8*4+rdi],r13d
+	add	r13b,r9b
+	mov	al,BYTE PTR[r13*4+rdi]
+	add	r12b,r11b
+	mov	r8,r10
+	mov	r13d,DWORD PTR[r12*4+rdi]
+	ror	rax,8
+	inc	r8b
+	mov	r9d,DWORD PTR[r8*4+rdi]
+	cmp	r12,r8
+	mov	DWORD PTR[r12*4+rdi],r11d
+	cmove	r9,r11
+	mov	DWORD PTR[r10*4+rdi],r13d
+	add	r13b,r11b
+	mov	al,BYTE PTR[r13*4+rdi]
+	ror	rax,8
+	sub	rsi,8
+
+	xor	rax,QWORD PTR[rdx]
+	add	rdx,8
+	mov	QWORD PTR[rcx],rax
+	add	rcx,8
+
+	test	rsi,-8
+	jnz	$L$loop8
+	cmp	rsi,0
+	jne	$L$loop1
+	jmp	$L$exit
+
+ALIGN	16
+$L$loop1::
+	add	r12b,r9b
+	mov	r13d,DWORD PTR[r12*4+rdi]
+	mov	DWORD PTR[r12*4+rdi],r9d
+	mov	DWORD PTR[r8*4+rdi],r13d
+	add	r9b,r13b
+	inc	r8b
+	mov	r13d,DWORD PTR[r9*4+rdi]
+	mov	r9d,DWORD PTR[r8*4+rdi]
+	xor	r13b,BYTE PTR[rdx]
+	inc	rdx
+	mov	BYTE PTR[rcx],r13b
+	inc	rcx
+	dec	rsi
+	jnz	$L$loop1
+	jmp	$L$exit
+
+ALIGN	16
+$L$RC4_CHAR::
+	add	r8b,1
+	movzx	r9d,BYTE PTR[r8*1+rdi]
+	test	rsi,-8
+	jz	$L$cloop1
+	cmp	DWORD PTR[260+rdi],0
+	jnz	$L$cloop1
+	jmp	$L$cloop8
+ALIGN	16
+$L$cloop8::
+	mov	eax,DWORD PTR[rdx]
+	mov	ebx,DWORD PTR[4+rdx]
+	add	r12b,r9b
+	lea	r10,QWORD PTR[1+r8]
+	movzx	r13d,BYTE PTR[r12*1+rdi]
+	movzx	r10d,r10b
+	movzx	r11d,BYTE PTR[r10*1+rdi]
+	mov	BYTE PTR[r12*1+rdi],r9b
+	cmp	r12,r10
+	mov	BYTE PTR[r8*1+rdi],r13b
+	jne	$L$cmov0
+
+	mov	r11,r9
+$L$cmov0::
+	add	r13b,r9b
+	xor	al,BYTE PTR[r13*1+rdi]
+	ror	eax,8
+	add	r12b,r11b
+	lea	r8,QWORD PTR[1+r10]
+	movzx	r13d,BYTE PTR[r12*1+rdi]
+	movzx	r8d,r8b
+	movzx	r9d,BYTE PTR[r8*1+rdi]
+	mov	BYTE PTR[r12*1+rdi],r11b
+	cmp	r12,r8
+	mov	BYTE PTR[r10*1+rdi],r13b
+	jne	$L$cmov1
+
+	mov	r9,r11
+$L$cmov1::
+	add	r13b,r11b
+	xor	al,BYTE PTR[r13*1+rdi]
+	ror	eax,8
+	add	r12b,r9b
+	lea	r10,QWORD PTR[1+r8]
+	movzx	r13d,BYTE PTR[r12*1+rdi]
+	movzx	r10d,r10b
+	movzx	r11d,BYTE PTR[r10*1+rdi]
+	mov	BYTE PTR[r12*1+rdi],r9b
+	cmp	r12,r10
+	mov	BYTE PTR[r8*1+rdi],r13b
+	jne	$L$cmov2
+
+	mov	r11,r9
+$L$cmov2::
+	add	r13b,r9b
+	xor	al,BYTE PTR[r13*1+rdi]
+	ror	eax,8
+	add	r12b,r11b
+	lea	r8,QWORD PTR[1+r10]
+	movzx	r13d,BYTE PTR[r12*1+rdi]
+	movzx	r8d,r8b
+	movzx	r9d,BYTE PTR[r8*1+rdi]
+	mov	BYTE PTR[r12*1+rdi],r11b
+	cmp	r12,r8
+	mov	BYTE PTR[r10*1+rdi],r13b
+	jne	$L$cmov3
+
+	mov	r9,r11
+$L$cmov3::
+	add	r13b,r11b
+	xor	al,BYTE PTR[r13*1+rdi]
+	ror	eax,8
+	add	r12b,r9b
+	lea	r10,QWORD PTR[1+r8]
+	movzx	r13d,BYTE PTR[r12*1+rdi]
+	movzx	r10d,r10b
+	movzx	r11d,BYTE PTR[r10*1+rdi]
+	mov	BYTE PTR[r12*1+rdi],r9b
+	cmp	r12,r10
+	mov	BYTE PTR[r8*1+rdi],r13b
+	jne	$L$cmov4
+
+	mov	r11,r9
+$L$cmov4::
+	add	r13b,r9b
+	xor	bl,BYTE PTR[r13*1+rdi]
+	ror	ebx,8
+	add	r12b,r11b
+	lea	r8,QWORD PTR[1+r10]
+	movzx	r13d,BYTE PTR[r12*1+rdi]
+	movzx	r8d,r8b
+	movzx	r9d,BYTE PTR[r8*1+rdi]
+	mov	BYTE PTR[r12*1+rdi],r11b
+	cmp	r12,r8
+	mov	BYTE PTR[r10*1+rdi],r13b
+	jne	$L$cmov5
+
+	mov	r9,r11
+$L$cmov5::
+	add	r13b,r11b
+	xor	bl,BYTE PTR[r13*1+rdi]
+	ror	ebx,8
+	add	r12b,r9b
+	lea	r10,QWORD PTR[1+r8]
+	movzx	r13d,BYTE PTR[r12*1+rdi]
+	movzx	r10d,r10b
+	movzx	r11d,BYTE PTR[r10*1+rdi]
+	mov	BYTE PTR[r12*1+rdi],r9b
+	cmp	r12,r10
+	mov	BYTE PTR[r8*1+rdi],r13b
+	jne	$L$cmov6
+
+	mov	r11,r9
+$L$cmov6::
+	add	r13b,r9b
+	xor	bl,BYTE PTR[r13*1+rdi]
+	ror	ebx,8
+	add	r12b,r11b
+	lea	r8,QWORD PTR[1+r10]
+	movzx	r13d,BYTE PTR[r12*1+rdi]
+	movzx	r8d,r8b
+	movzx	r9d,BYTE PTR[r8*1+rdi]
+	mov	BYTE PTR[r12*1+rdi],r11b
+	cmp	r12,r8
+	mov	BYTE PTR[r10*1+rdi],r13b
+	jne	$L$cmov7
+
+	mov	r9,r11
+$L$cmov7::
+	add	r13b,r11b
+	xor	bl,BYTE PTR[r13*1+rdi]
+	ror	ebx,8
+	lea	rsi,QWORD PTR[((-8))+rsi]
+	mov	DWORD PTR[rcx],eax
+	lea	rdx,QWORD PTR[8+rdx]
+	mov	DWORD PTR[4+rcx],ebx
+	lea	rcx,QWORD PTR[8+rcx]
+
+	test	rsi,-8
+	jnz	$L$cloop8
+	cmp	rsi,0
+	jne	$L$cloop1
+	jmp	$L$exit
+ALIGN	16
+$L$cloop1::
+	add	r12b,r9b
+	movzx	r13d,BYTE PTR[r12*1+rdi]
+	mov	BYTE PTR[r12*1+rdi],r9b
+	mov	BYTE PTR[r8*1+rdi],r13b
+	add	r13b,r9b
+	add	r8b,1
+	movzx	r13d,r13b
+	movzx	r8d,r8b
+	movzx	r13d,BYTE PTR[r13*1+rdi]
+	movzx	r9d,BYTE PTR[r8*1+rdi]
+	xor	r13b,BYTE PTR[rdx]
+	lea	rdx,QWORD PTR[1+rdx]
+	mov	BYTE PTR[rcx],r13b
+	lea	rcx,QWORD PTR[1+rcx]
+	sub	rsi,1
+	jnz	$L$cloop1
+	jmp	$L$exit
+
+ALIGN	16
+$L$exit::
+	sub	r8b,1
+	mov	DWORD PTR[((-8))+rdi],r8d
+	mov	DWORD PTR[((-4))+rdi],r12d
+
+	mov	r13,QWORD PTR[rsp]
+	mov	r12,QWORD PTR[8+rsp]
+	mov	rbx,QWORD PTR[16+rsp]
+	add	rsp,24
+$L$epilogue::
+	mov	rdi,QWORD PTR[8+rsp]	;WIN64 epilogue
+	mov	rsi,QWORD PTR[16+rsp]
+	DB	0F3h,0C3h		;repret
+$L$SEH_end_RC4::
+RC4	ENDP
+EXTERN	OPENSSL_ia32cap_P:NEAR
+PUBLIC	RC4_set_key
+
+ALIGN	16
+RC4_set_key	PROC PUBLIC
+	mov	QWORD PTR[8+rsp],rdi	;WIN64 prologue
+	mov	QWORD PTR[16+rsp],rsi
+	mov	rax,rsp
+$L$SEH_begin_RC4_set_key::
+	mov	rdi,rcx
+	mov	rsi,rdx
+	mov	rdx,r8
+
+
+	lea	rdi,QWORD PTR[8+rdi]
+	lea	rdx,QWORD PTR[rsi*1+rdx]
+	neg	rsi
+	mov	rcx,rsi
+	xor	eax,eax
+	xor	r9,r9
+	xor	r10,r10
+	xor	r11,r11
+
+	mov	r8d,DWORD PTR[OPENSSL_ia32cap_P]
+	bt	r8d,20
+	jnc	$L$w1stloop
+	bt	r8d,30
+	setc	r9b
+	mov	DWORD PTR[260+rdi],r9d
+	jmp	$L$c1stloop
+
+ALIGN	16
+$L$w1stloop::
+	mov	DWORD PTR[rax*4+rdi],eax
+	add	al,1
+	jnc	$L$w1stloop
+
+	xor	r9,r9
+	xor	r8,r8
+ALIGN	16
+$L$w2ndloop::
+	mov	r10d,DWORD PTR[r9*4+rdi]
+	add	r8b,BYTE PTR[rsi*1+rdx]
+	add	r8b,r10b
+	add	rsi,1
+	mov	r11d,DWORD PTR[r8*4+rdi]
+	cmovz	rsi,rcx
+	mov	DWORD PTR[r8*4+rdi],r10d
+	mov	DWORD PTR[r9*4+rdi],r11d
+	add	r9b,1
+	jnc	$L$w2ndloop
+	jmp	$L$exit_key
+
+ALIGN	16
+$L$c1stloop::
+	mov	BYTE PTR[rax*1+rdi],al
+	add	al,1
+	jnc	$L$c1stloop
+
+	xor	r9,r9
+	xor	r8,r8
+ALIGN	16
+$L$c2ndloop::
+	mov	r10b,BYTE PTR[r9*1+rdi]
+	add	r8b,BYTE PTR[rsi*1+rdx]
+	add	r8b,r10b
+	add	rsi,1
+	mov	r11b,BYTE PTR[r8*1+rdi]
+	jnz	$L$cnowrap
+	mov	rsi,rcx
+$L$cnowrap::
+	mov	BYTE PTR[r8*1+rdi],r10b
+	mov	BYTE PTR[r9*1+rdi],r11b
+	add	r9b,1
+	jnc	$L$c2ndloop
+	mov	DWORD PTR[256+rdi],-1
+
+ALIGN	16
+$L$exit_key::
+	xor	eax,eax
+	mov	DWORD PTR[((-8))+rdi],eax
+	mov	DWORD PTR[((-4))+rdi],eax
+	mov	rdi,QWORD PTR[8+rsp]	;WIN64 epilogue
+	mov	rsi,QWORD PTR[16+rsp]
+	DB	0F3h,0C3h		;repret
+$L$SEH_end_RC4_set_key::
+RC4_set_key	ENDP
+
+PUBLIC	RC4_options
+
+ALIGN	16
+RC4_options	PROC PUBLIC
+	lea	rax,QWORD PTR[$L$opts]
+	mov	edx,DWORD PTR[OPENSSL_ia32cap_P]
+	bt	edx,20
+	jnc	$L$done
+	add	rax,12
+	bt	edx,30
+	jnc	$L$done
+	add	rax,13
+$L$done::
+	DB	0F3h,0C3h		;repret
+ALIGN	64
+$L$opts::
+DB	114,99,52,40,56,120,44,105,110,116,41,0
+DB	114,99,52,40,56,120,44,99,104,97,114,41,0
+DB	114,99,52,40,49,120,44,99,104,97,114,41,0
+DB	82,67,52,32,102,111,114,32,120,56,54,95,54,52,44,32
+DB	67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+DB	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+DB	62,0
+ALIGN	64
+RC4_options	ENDP
+EXTERN	__imp_RtlVirtualUnwind:NEAR
+
+ALIGN	16
+stream_se_handler	PROC PRIVATE
+	push	rsi
+	push	rdi
+	push	rbx
+	push	rbp
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+	pushfq
+	sub	rsp,64
+
+	mov	rax,QWORD PTR[120+r8]
+	mov	rbx,QWORD PTR[248+r8]
+
+	lea	r10,QWORD PTR[$L$prologue]
+	cmp	rbx,r10
+	jb	$L$in_prologue
+
+	mov	rax,QWORD PTR[152+r8]
+
+	lea	r10,QWORD PTR[$L$epilogue]
+	cmp	rbx,r10
+	jae	$L$in_prologue
+
+	lea	rax,QWORD PTR[24+rax]
+
+	mov	rbx,QWORD PTR[((-8))+rax]
+	mov	r12,QWORD PTR[((-16))+rax]
+	mov	r13,QWORD PTR[((-24))+rax]
+	mov	QWORD PTR[144+r8],rbx
+	mov	QWORD PTR[216+r8],r12
+	mov	QWORD PTR[224+r8],r13
+
+$L$in_prologue::
+	mov	rdi,QWORD PTR[8+rax]
+	mov	rsi,QWORD PTR[16+rax]
+	mov	QWORD PTR[152+r8],rax
+	mov	QWORD PTR[168+r8],rsi
+	mov	QWORD PTR[176+r8],rdi
+
+	jmp	$L$common_seh_exit
+stream_se_handler	ENDP
+
+
+ALIGN	16
+key_se_handler	PROC PRIVATE
+	push	rsi
+	push	rdi
+	push	rbx
+	push	rbp
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+	pushfq
+	sub	rsp,64
+
+	mov	rax,QWORD PTR[152+r8]
+	mov	rdi,QWORD PTR[8+rax]
+	mov	rsi,QWORD PTR[16+rax]
+	mov	QWORD PTR[168+r8],rsi
+	mov	QWORD PTR[176+r8],rdi
+
+$L$common_seh_exit::
+
+	mov	rdi,QWORD PTR[40+r9]
+	mov	rsi,r8
+	mov	ecx,154
+	DD	0a548f3fch
+
+
+	mov	rsi,r9
+	xor	rcx,rcx
+	mov	rdx,QWORD PTR[8+rsi]
+	mov	r8,QWORD PTR[rsi]
+	mov	r9,QWORD PTR[16+rsi]
+	mov	r10,QWORD PTR[40+rsi]
+	lea	r11,QWORD PTR[56+rsi]
+	lea	r12,QWORD PTR[24+rsi]
+	mov	QWORD PTR[32+rsp],r10
+	mov	QWORD PTR[40+rsp],r11
+	mov	QWORD PTR[48+rsp],r12
+	mov	QWORD PTR[56+rsp],rcx
+	call	QWORD PTR[__imp_RtlVirtualUnwind]
+
+	mov	eax,1
+	add	rsp,64
+	popfq
+	pop	r15
+	pop	r14
+	pop	r13
+	pop	r12
+	pop	rbp
+	pop	rbx
+	pop	rdi
+	pop	rsi
+	DB	0F3h,0C3h		;repret
+key_se_handler	ENDP
+
+.text$	ENDS
+.pdata	SEGMENT READONLY ALIGN(4)
+ALIGN	4
+	DD	imagerel $L$SEH_begin_RC4
+	DD	imagerel $L$SEH_end_RC4
+	DD	imagerel $L$SEH_info_RC4
+
+	DD	imagerel $L$SEH_begin_RC4_set_key
+	DD	imagerel $L$SEH_end_RC4_set_key
+	DD	imagerel $L$SEH_info_RC4_set_key
+
+.pdata	ENDS
+.xdata	SEGMENT READONLY ALIGN(8)
+ALIGN	8
+$L$SEH_info_RC4::
+DB	9,0,0,0
+	DD	imagerel stream_se_handler
+$L$SEH_info_RC4_set_key::
+DB	9,0,0,0
+	DD	imagerel key_se_handler
+
+.xdata	ENDS
+END

deps/openssl/asm/x64-win32-masm/whrlpool/wp-x86_64.asm

@@ -0,0 +1,972 @@
+OPTION	DOTNAME
+.text$	SEGMENT ALIGN(64) 'CODE'
+
+PUBLIC	whirlpool_block
+
+ALIGN	16
+whirlpool_block	PROC PUBLIC
+	mov	QWORD PTR[8+rsp],rdi	;WIN64 prologue
+	mov	QWORD PTR[16+rsp],rsi
+	mov	rax,rsp
+$L$SEH_begin_whirlpool_block::
+	mov	rdi,rcx
+	mov	rsi,rdx
+	mov	rdx,r8
+
+
+	push	rbx
+	push	rbp
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+
+	mov	r11,rsp
+	sub	rsp,128+40
+	and	rsp,-64
+
+	lea	r10,QWORD PTR[128+rsp]
+	mov	QWORD PTR[r10],rdi
+	mov	QWORD PTR[8+r10],rsi
+	mov	QWORD PTR[16+r10],rdx
+	mov	QWORD PTR[32+r10],r11
+$L$prologue::
+
+	mov	rbx,r10
+	lea	rbp,QWORD PTR[$L$table]
+
+	xor	rcx,rcx
+	xor	rdx,rdx
+	mov	r8,QWORD PTR[((0*8))+rdi]
+	mov	r9,QWORD PTR[((1*8))+rdi]
+	mov	r10,QWORD PTR[((2*8))+rdi]
+	mov	r11,QWORD PTR[((3*8))+rdi]
+	mov	r12,QWORD PTR[((4*8))+rdi]
+	mov	r13,QWORD PTR[((5*8))+rdi]
+	mov	r14,QWORD PTR[((6*8))+rdi]
+	mov	r15,QWORD PTR[((7*8))+rdi]
+$L$outerloop::
+	mov	QWORD PTR[((0*8))+rsp],r8
+	mov	QWORD PTR[((1*8))+rsp],r9
+	mov	QWORD PTR[((2*8))+rsp],r10
+	mov	QWORD PTR[((3*8))+rsp],r11
+	mov	QWORD PTR[((4*8))+rsp],r12
+	mov	QWORD PTR[((5*8))+rsp],r13
+	mov	QWORD PTR[((6*8))+rsp],r14
+	mov	QWORD PTR[((7*8))+rsp],r15
+	xor	r8,QWORD PTR[((0*8))+rsi]
+	xor	r9,QWORD PTR[((1*8))+rsi]
+	xor	r10,QWORD PTR[((2*8))+rsi]
+	xor	r11,QWORD PTR[((3*8))+rsi]
+	xor	r12,QWORD PTR[((4*8))+rsi]
+	xor	r13,QWORD PTR[((5*8))+rsi]
+	xor	r14,QWORD PTR[((6*8))+rsi]
+	xor	r15,QWORD PTR[((7*8))+rsi]
+	mov	QWORD PTR[((64+0*8))+rsp],r8
+	mov	QWORD PTR[((64+1*8))+rsp],r9
+	mov	QWORD PTR[((64+2*8))+rsp],r10
+	mov	QWORD PTR[((64+3*8))+rsp],r11
+	mov	QWORD PTR[((64+4*8))+rsp],r12
+	mov	QWORD PTR[((64+5*8))+rsp],r13
+	mov	QWORD PTR[((64+6*8))+rsp],r14
+	mov	QWORD PTR[((64+7*8))+rsp],r15
+	xor	rsi,rsi
+	mov	QWORD PTR[24+rbx],rsi
+ALIGN	16
+$L$round::
+	mov	r8,QWORD PTR[4096+rsi*8+rbp]
+	mov	eax,DWORD PTR[rsp]
+	mov	ebx,DWORD PTR[4+rsp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r8,QWORD PTR[rsi*8+rbp]
+	mov	r9,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((0*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	mov	r10,QWORD PTR[6+rsi*8+rbp]
+	mov	r11,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	mov	r12,QWORD PTR[4+rsi*8+rbp]
+	mov	r13,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((0*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	mov	r14,QWORD PTR[2+rsi*8+rbp]
+	mov	r15,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r9,QWORD PTR[rsi*8+rbp]
+	xor	r10,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((1*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r11,QWORD PTR[6+rsi*8+rbp]
+	xor	r12,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r13,QWORD PTR[4+rsi*8+rbp]
+	xor	r14,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((1*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r15,QWORD PTR[2+rsi*8+rbp]
+	xor	r8,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r10,QWORD PTR[rsi*8+rbp]
+	xor	r11,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((2*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r12,QWORD PTR[6+rsi*8+rbp]
+	xor	r13,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r14,QWORD PTR[4+rsi*8+rbp]
+	xor	r15,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((2*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r8,QWORD PTR[2+rsi*8+rbp]
+	xor	r9,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r11,QWORD PTR[rsi*8+rbp]
+	xor	r12,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((3*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r13,QWORD PTR[6+rsi*8+rbp]
+	xor	r14,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r15,QWORD PTR[4+rsi*8+rbp]
+	xor	r8,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((3*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r9,QWORD PTR[2+rsi*8+rbp]
+	xor	r10,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r12,QWORD PTR[rsi*8+rbp]
+	xor	r13,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((4*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r14,QWORD PTR[6+rsi*8+rbp]
+	xor	r15,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r8,QWORD PTR[4+rsi*8+rbp]
+	xor	r9,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((4*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r10,QWORD PTR[2+rsi*8+rbp]
+	xor	r11,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r13,QWORD PTR[rsi*8+rbp]
+	xor	r14,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((5*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r15,QWORD PTR[6+rsi*8+rbp]
+	xor	r8,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r9,QWORD PTR[4+rsi*8+rbp]
+	xor	r10,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((5*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r11,QWORD PTR[2+rsi*8+rbp]
+	xor	r12,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r14,QWORD PTR[rsi*8+rbp]
+	xor	r15,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((6*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r8,QWORD PTR[6+rsi*8+rbp]
+	xor	r9,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r10,QWORD PTR[4+rsi*8+rbp]
+	xor	r11,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((6*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r12,QWORD PTR[2+rsi*8+rbp]
+	xor	r13,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r15,QWORD PTR[rsi*8+rbp]
+	xor	r8,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((7*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r9,QWORD PTR[6+rsi*8+rbp]
+	xor	r10,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r11,QWORD PTR[4+rsi*8+rbp]
+	xor	r12,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((7*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r13,QWORD PTR[2+rsi*8+rbp]
+	xor	r14,QWORD PTR[1+rdi*8+rbp]
+	mov	QWORD PTR[((0*8))+rsp],r8
+	mov	QWORD PTR[((1*8))+rsp],r9
+	mov	QWORD PTR[((2*8))+rsp],r10
+	mov	QWORD PTR[((3*8))+rsp],r11
+	mov	QWORD PTR[((4*8))+rsp],r12
+	mov	QWORD PTR[((5*8))+rsp],r13
+	mov	QWORD PTR[((6*8))+rsp],r14
+	mov	QWORD PTR[((7*8))+rsp],r15
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r8,QWORD PTR[rsi*8+rbp]
+	xor	r9,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((64+0*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r10,QWORD PTR[6+rsi*8+rbp]
+	xor	r11,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r12,QWORD PTR[4+rsi*8+rbp]
+	xor	r13,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((64+0*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r14,QWORD PTR[2+rsi*8+rbp]
+	xor	r15,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r9,QWORD PTR[rsi*8+rbp]
+	xor	r10,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((64+1*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r11,QWORD PTR[6+rsi*8+rbp]
+	xor	r12,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r13,QWORD PTR[4+rsi*8+rbp]
+	xor	r14,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((64+1*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r15,QWORD PTR[2+rsi*8+rbp]
+	xor	r8,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r10,QWORD PTR[rsi*8+rbp]
+	xor	r11,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((64+2*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r12,QWORD PTR[6+rsi*8+rbp]
+	xor	r13,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r14,QWORD PTR[4+rsi*8+rbp]
+	xor	r15,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((64+2*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r8,QWORD PTR[2+rsi*8+rbp]
+	xor	r9,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r11,QWORD PTR[rsi*8+rbp]
+	xor	r12,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((64+3*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r13,QWORD PTR[6+rsi*8+rbp]
+	xor	r14,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r15,QWORD PTR[4+rsi*8+rbp]
+	xor	r8,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((64+3*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r9,QWORD PTR[2+rsi*8+rbp]
+	xor	r10,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r12,QWORD PTR[rsi*8+rbp]
+	xor	r13,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((64+4*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r14,QWORD PTR[6+rsi*8+rbp]
+	xor	r15,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r8,QWORD PTR[4+rsi*8+rbp]
+	xor	r9,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((64+4*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r10,QWORD PTR[2+rsi*8+rbp]
+	xor	r11,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r13,QWORD PTR[rsi*8+rbp]
+	xor	r14,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((64+5*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r15,QWORD PTR[6+rsi*8+rbp]
+	xor	r8,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r9,QWORD PTR[4+rsi*8+rbp]
+	xor	r10,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((64+5*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r11,QWORD PTR[2+rsi*8+rbp]
+	xor	r12,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r14,QWORD PTR[rsi*8+rbp]
+	xor	r15,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	mov	eax,DWORD PTR[((64+6*8+8))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r8,QWORD PTR[6+rsi*8+rbp]
+	xor	r9,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r10,QWORD PTR[4+rsi*8+rbp]
+	xor	r11,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	mov	ebx,DWORD PTR[((64+6*8+8+4))+rsp]
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r12,QWORD PTR[2+rsi*8+rbp]
+	xor	r13,QWORD PTR[1+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	eax,16
+	xor	r15,QWORD PTR[rsi*8+rbp]
+	xor	r8,QWORD PTR[7+rdi*8+rbp]
+	mov	cl,al
+	mov	dl,ah
+
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r9,QWORD PTR[6+rsi*8+rbp]
+	xor	r10,QWORD PTR[5+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	shr	ebx,16
+	xor	r11,QWORD PTR[4+rsi*8+rbp]
+	xor	r12,QWORD PTR[3+rdi*8+rbp]
+	mov	cl,bl
+	mov	dl,bh
+
+	lea	rsi,QWORD PTR[rcx*1+rcx]
+	lea	rdi,QWORD PTR[rdx*1+rdx]
+	xor	r13,QWORD PTR[2+rsi*8+rbp]
+	xor	r14,QWORD PTR[1+rdi*8+rbp]
+	lea	rbx,QWORD PTR[128+rsp]
+	mov	rsi,QWORD PTR[24+rbx]
+	add	rsi,1
+	cmp	rsi,10
+	je	$L$roundsdone
+
+	mov	QWORD PTR[24+rbx],rsi
+	mov	QWORD PTR[((64+0*8))+rsp],r8
+	mov	QWORD PTR[((64+1*8))+rsp],r9
+	mov	QWORD PTR[((64+2*8))+rsp],r10
+	mov	QWORD PTR[((64+3*8))+rsp],r11
+	mov	QWORD PTR[((64+4*8))+rsp],r12
+	mov	QWORD PTR[((64+5*8))+rsp],r13
+	mov	QWORD PTR[((64+6*8))+rsp],r14
+	mov	QWORD PTR[((64+7*8))+rsp],r15
+	jmp	$L$round
+ALIGN	16
+$L$roundsdone::
+	mov	rdi,QWORD PTR[rbx]
+	mov	rsi,QWORD PTR[8+rbx]
+	mov	rax,QWORD PTR[16+rbx]
+	xor	r8,QWORD PTR[((0*8))+rsi]
+	xor	r9,QWORD PTR[((1*8))+rsi]
+	xor	r10,QWORD PTR[((2*8))+rsi]
+	xor	r11,QWORD PTR[((3*8))+rsi]
+	xor	r12,QWORD PTR[((4*8))+rsi]
+	xor	r13,QWORD PTR[((5*8))+rsi]
+	xor	r14,QWORD PTR[((6*8))+rsi]
+	xor	r15,QWORD PTR[((7*8))+rsi]
+	xor	r8,QWORD PTR[((0*8))+rdi]
+	xor	r9,QWORD PTR[((1*8))+rdi]
+	xor	r10,QWORD PTR[((2*8))+rdi]
+	xor	r11,QWORD PTR[((3*8))+rdi]
+	xor	r12,QWORD PTR[((4*8))+rdi]
+	xor	r13,QWORD PTR[((5*8))+rdi]
+	xor	r14,QWORD PTR[((6*8))+rdi]
+	xor	r15,QWORD PTR[((7*8))+rdi]
+	mov	QWORD PTR[((0*8))+rdi],r8
+	mov	QWORD PTR[((1*8))+rdi],r9
+	mov	QWORD PTR[((2*8))+rdi],r10
+	mov	QWORD PTR[((3*8))+rdi],r11
+	mov	QWORD PTR[((4*8))+rdi],r12
+	mov	QWORD PTR[((5*8))+rdi],r13
+	mov	QWORD PTR[((6*8))+rdi],r14
+	mov	QWORD PTR[((7*8))+rdi],r15
+	lea	rsi,QWORD PTR[64+rsi]
+	sub	rax,1
+	jz	$L$alldone
+	mov	QWORD PTR[8+rbx],rsi
+	mov	QWORD PTR[16+rbx],rax
+	jmp	$L$outerloop
+$L$alldone::
+	mov	rsi,QWORD PTR[32+rbx]
+	mov	r15,QWORD PTR[rsi]
+	mov	r14,QWORD PTR[8+rsi]
+	mov	r13,QWORD PTR[16+rsi]
+	mov	r12,QWORD PTR[24+rsi]
+	mov	rbp,QWORD PTR[32+rsi]
+	mov	rbx,QWORD PTR[40+rsi]
+	lea	rsp,QWORD PTR[48+rsi]
+$L$epilogue::
+	mov	rdi,QWORD PTR[8+rsp]	;WIN64 epilogue
+	mov	rsi,QWORD PTR[16+rsp]
+	DB	0F3h,0C3h		;repret
+$L$SEH_end_whirlpool_block::
+whirlpool_block	ENDP
+
+ALIGN	64
+
+$L$table::
+DB	24,24,96,24,192,120,48,216,24,24,96,24,192,120,48,216
+DB	35,35,140,35,5,175,70,38,35,35,140,35,5,175,70,38
+DB	198,198,63,198,126,249,145,184,198,198,63,198,126,249,145,184
+DB	232,232,135,232,19,111,205,251,232,232,135,232,19,111,205,251
+DB	135,135,38,135,76,161,19,203,135,135,38,135,76,161,19,203
+DB	184,184,218,184,169,98,109,17,184,184,218,184,169,98,109,17
+DB	1,1,4,1,8,5,2,9,1,1,4,1,8,5,2,9
+DB	79,79,33,79,66,110,158,13,79,79,33,79,66,110,158,13
+DB	54,54,216,54,173,238,108,155,54,54,216,54,173,238,108,155
+DB	166,166,162,166,89,4,81,255,166,166,162,166,89,4,81,255
+DB	210,210,111,210,222,189,185,12,210,210,111,210,222,189,185,12
+DB	245,245,243,245,251,6,247,14,245,245,243,245,251,6,247,14
+DB	121,121,249,121,239,128,242,150,121,121,249,121,239,128,242,150
+DB	111,111,161,111,95,206,222,48,111,111,161,111,95,206,222,48
+DB	145,145,126,145,252,239,63,109,145,145,126,145,252,239,63,109
+DB	82,82,85,82,170,7,164,248,82,82,85,82,170,7,164,248
+DB	96,96,157,96,39,253,192,71,96,96,157,96,39,253,192,71
+DB	188,188,202,188,137,118,101,53,188,188,202,188,137,118,101,53
+DB	155,155,86,155,172,205,43,55,155,155,86,155,172,205,43,55
+DB	142,142,2,142,4,140,1,138,142,142,2,142,4,140,1,138
+DB	163,163,182,163,113,21,91,210,163,163,182,163,113,21,91,210
+DB	12,12,48,12,96,60,24,108,12,12,48,12,96,60,24,108
+DB	123,123,241,123,255,138,246,132,123,123,241,123,255,138,246,132
+DB	53,53,212,53,181,225,106,128,53,53,212,53,181,225,106,128
+DB	29,29,116,29,232,105,58,245,29,29,116,29,232,105,58,245
+DB	224,224,167,224,83,71,221,179,224,224,167,224,83,71,221,179
+DB	215,215,123,215,246,172,179,33,215,215,123,215,246,172,179,33
+DB	194,194,47,194,94,237,153,156,194,194,47,194,94,237,153,156
+DB	46,46,184,46,109,150,92,67,46,46,184,46,109,150,92,67
+DB	75,75,49,75,98,122,150,41,75,75,49,75,98,122,150,41
+DB	254,254,223,254,163,33,225,93,254,254,223,254,163,33,225,93
+DB	87,87,65,87,130,22,174,213,87,87,65,87,130,22,174,213
+DB	21,21,84,21,168,65,42,189,21,21,84,21,168,65,42,189
+DB	119,119,193,119,159,182,238,232,119,119,193,119,159,182,238,232
+DB	55,55,220,55,165,235,110,146,55,55,220,55,165,235,110,146
+DB	229,229,179,229,123,86,215,158,229,229,179,229,123,86,215,158
+DB	159,159,70,159,140,217,35,19,159,159,70,159,140,217,35,19
+DB	240,240,231,240,211,23,253,35,240,240,231,240,211,23,253,35
+DB	74,74,53,74,106,127,148,32,74,74,53,74,106,127,148,32
+DB	218,218,79,218,158,149,169,68,218,218,79,218,158,149,169,68
+DB	88,88,125,88,250,37,176,162,88,88,125,88,250,37,176,162
+DB	201,201,3,201,6,202,143,207,201,201,3,201,6,202,143,207
+DB	41,41,164,41,85,141,82,124,41,41,164,41,85,141,82,124
+DB	10,10,40,10,80,34,20,90,10,10,40,10,80,34,20,90
+DB	177,177,254,177,225,79,127,80,177,177,254,177,225,79,127,80
+DB	160,160,186,160,105,26,93,201,160,160,186,160,105,26,93,201
+DB	107,107,177,107,127,218,214,20,107,107,177,107,127,218,214,20
+DB	133,133,46,133,92,171,23,217,133,133,46,133,92,171,23,217
+DB	189,189,206,189,129,115,103,60,189,189,206,189,129,115,103,60
+DB	93,93,105,93,210,52,186,143,93,93,105,93,210,52,186,143
+DB	16,16,64,16,128,80,32,144,16,16,64,16,128,80,32,144
+DB	244,244,247,244,243,3,245,7,244,244,247,244,243,3,245,7
+DB	203,203,11,203,22,192,139,221,203,203,11,203,22,192,139,221
+DB	62,62,248,62,237,198,124,211,62,62,248,62,237,198,124,211
+DB	5,5,20,5,40,17,10,45,5,5,20,5,40,17,10,45
+DB	103,103,129,103,31,230,206,120,103,103,129,103,31,230,206,120
+DB	228,228,183,228,115,83,213,151,228,228,183,228,115,83,213,151
+DB	39,39,156,39,37,187,78,2,39,39,156,39,37,187,78,2
+DB	65,65,25,65,50,88,130,115,65,65,25,65,50,88,130,115
+DB	139,139,22,139,44,157,11,167,139,139,22,139,44,157,11,167
+DB	167,167,166,167,81,1,83,246,167,167,166,167,81,1,83,246
+DB	125,125,233,125,207,148,250,178,125,125,233,125,207,148,250,178
+DB	149,149,110,149,220,251,55,73,149,149,110,149,220,251,55,73
+DB	216,216,71,216,142,159,173,86,216,216,71,216,142,159,173,86
+DB	251,251,203,251,139,48,235,112,251,251,203,251,139,48,235,112
+DB	238,238,159,238,35,113,193,205,238,238,159,238,35,113,193,205
+DB	124,124,237,124,199,145,248,187,124,124,237,124,199,145,248,187
+DB	102,102,133,102,23,227,204,113,102,102,133,102,23,227,204,113
+DB	221,221,83,221,166,142,167,123,221,221,83,221,166,142,167,123
+DB	23,23,92,23,184,75,46,175,23,23,92,23,184,75,46,175
+DB	71,71,1,71,2,70,142,69,71,71,1,71,2,70,142,69
+DB	158,158,66,158,132,220,33,26,158,158,66,158,132,220,33,26
+DB	202,202,15,202,30,197,137,212,202,202,15,202,30,197,137,212
+DB	45,45,180,45,117,153,90,88,45,45,180,45,117,153,90,88
+DB	191,191,198,191,145,121,99,46,191,191,198,191,145,121,99,46
+DB	7,7,28,7,56,27,14,63,7,7,28,7,56,27,14,63
+DB	173,173,142,173,1,35,71,172,173,173,142,173,1,35,71,172
+DB	90,90,117,90,234,47,180,176,90,90,117,90,234,47,180,176
+DB	131,131,54,131,108,181,27,239,131,131,54,131,108,181,27,239
+DB	51,51,204,51,133,255,102,182,51,51,204,51,133,255,102,182
+DB	99,99,145,99,63,242,198,92,99,99,145,99,63,242,198,92
+DB	2,2,8,2,16,10,4,18,2,2,8,2,16,10,4,18
+DB	170,170,146,170,57,56,73,147,170,170,146,170,57,56,73,147
+DB	113,113,217,113,175,168,226,222,113,113,217,113,175,168,226,222
+DB	200,200,7,200,14,207,141,198,200,200,7,200,14,207,141,198
+DB	25,25,100,25,200,125,50,209,25,25,100,25,200,125,50,209
+DB	73,73,57,73,114,112,146,59,73,73,57,73,114,112,146,59
+DB	217,217,67,217,134,154,175,95,217,217,67,217,134,154,175,95
+DB	242,242,239,242,195,29,249,49,242,242,239,242,195,29,249,49
+DB	227,227,171,227,75,72,219,168,227,227,171,227,75,72,219,168
+DB	91,91,113,91,226,42,182,185,91,91,113,91,226,42,182,185
+DB	136,136,26,136,52,146,13,188,136,136,26,136,52,146,13,188
+DB	154,154,82,154,164,200,41,62,154,154,82,154,164,200,41,62
+DB	38,38,152,38,45,190,76,11,38,38,152,38,45,190,76,11
+DB	50,50,200,50,141,250,100,191,50,50,200,50,141,250,100,191
+DB	176,176,250,176,233,74,125,89,176,176,250,176,233,74,125,89
+DB	233,233,131,233,27,106,207,242,233,233,131,233,27,106,207,242
+DB	15,15,60,15,120,51,30,119,15,15,60,15,120,51,30,119
+DB	213,213,115,213,230,166,183,51,213,213,115,213,230,166,183,51
+DB	128,128,58,128,116,186,29,244,128,128,58,128,116,186,29,244
+DB	190,190,194,190,153,124,97,39,190,190,194,190,153,124,97,39
+DB	205,205,19,205,38,222,135,235,205,205,19,205,38,222,135,235
+DB	52,52,208,52,189,228,104,137,52,52,208,52,189,228,104,137
+DB	72,72,61,72,122,117,144,50,72,72,61,72,122,117,144,50
+DB	255,255,219,255,171,36,227,84,255,255,219,255,171,36,227,84
+DB	122,122,245,122,247,143,244,141,122,122,245,122,247,143,244,141
+DB	144,144,122,144,244,234,61,100,144,144,122,144,244,234,61,100
+DB	95,95,97,95,194,62,190,157,95,95,97,95,194,62,190,157
+DB	32,32,128,32,29,160,64,61,32,32,128,32,29,160,64,61
+DB	104,104,189,104,103,213,208,15,104,104,189,104,103,213,208,15
+DB	26,26,104,26,208,114,52,202,26,26,104,26,208,114,52,202
+DB	174,174,130,174,25,44,65,183,174,174,130,174,25,44,65,183
+DB	180,180,234,180,201,94,117,125,180,180,234,180,201,94,117,125
+DB	84,84,77,84,154,25,168,206,84,84,77,84,154,25,168,206
+DB	147,147,118,147,236,229,59,127,147,147,118,147,236,229,59,127
+DB	34,34,136,34,13,170,68,47,34,34,136,34,13,170,68,47
+DB	100,100,141,100,7,233,200,99,100,100,141,100,7,233,200,99
+DB	241,241,227,241,219,18,255,42,241,241,227,241,219,18,255,42
+DB	115,115,209,115,191,162,230,204,115,115,209,115,191,162,230,204
+DB	18,18,72,18,144,90,36,130,18,18,72,18,144,90,36,130
+DB	64,64,29,64,58,93,128,122,64,64,29,64,58,93,128,122
+DB	8,8,32,8,64,40,16,72,8,8,32,8,64,40,16,72
+DB	195,195,43,195,86,232,155,149,195,195,43,195,86,232,155,149
+DB	236,236,151,236,51,123,197,223,236,236,151,236,51,123,197,223
+DB	219,219,75,219,150,144,171,77,219,219,75,219,150,144,171,77
+DB	161,161,190,161,97,31,95,192,161,161,190,161,97,31,95,192
+DB	141,141,14,141,28,131,7,145,141,141,14,141,28,131,7,145
+DB	61,61,244,61,245,201,122,200,61,61,244,61,245,201,122,200
+DB	151,151,102,151,204,241,51,91,151,151,102,151,204,241,51,91
+DB	0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+DB	207,207,27,207,54,212,131,249,207,207,27,207,54,212,131,249
+DB	43,43,172,43,69,135,86,110,43,43,172,43,69,135,86,110
+DB	118,118,197,118,151,179,236,225,118,118,197,118,151,179,236,225
+DB	130,130,50,130,100,176,25,230,130,130,50,130,100,176,25,230
+DB	214,214,127,214,254,169,177,40,214,214,127,214,254,169,177,40
+DB	27,27,108,27,216,119,54,195,27,27,108,27,216,119,54,195
+DB	181,181,238,181,193,91,119,116,181,181,238,181,193,91,119,116
+DB	175,175,134,175,17,41,67,190,175,175,134,175,17,41,67,190
+DB	106,106,181,106,119,223,212,29,106,106,181,106,119,223,212,29
+DB	80,80,93,80,186,13,160,234,80,80,93,80,186,13,160,234
+DB	69,69,9,69,18,76,138,87,69,69,9,69,18,76,138,87
+DB	243,243,235,243,203,24,251,56,243,243,235,243,203,24,251,56
+DB	48,48,192,48,157,240,96,173,48,48,192,48,157,240,96,173
+DB	239,239,155,239,43,116,195,196,239,239,155,239,43,116,195,196
+DB	63,63,252,63,229,195,126,218,63,63,252,63,229,195,126,218
+DB	85,85,73,85,146,28,170,199,85,85,73,85,146,28,170,199
+DB	162,162,178,162,121,16,89,219,162,162,178,162,121,16,89,219
+DB	234,234,143,234,3,101,201,233,234,234,143,234,3,101,201,233
+DB	101,101,137,101,15,236,202,106,101,101,137,101,15,236,202,106
+DB	186,186,210,186,185,104,105,3,186,186,210,186,185,104,105,3
+DB	47,47,188,47,101,147,94,74,47,47,188,47,101,147,94,74
+DB	192,192,39,192,78,231,157,142,192,192,39,192,78,231,157,142
+DB	222,222,95,222,190,129,161,96,222,222,95,222,190,129,161,96
+DB	28,28,112,28,224,108,56,252,28,28,112,28,224,108,56,252
+DB	253,253,211,253,187,46,231,70,253,253,211,253,187,46,231,70
+DB	77,77,41,77,82,100,154,31,77,77,41,77,82,100,154,31
+DB	146,146,114,146,228,224,57,118,146,146,114,146,228,224,57,118
+DB	117,117,201,117,143,188,234,250,117,117,201,117,143,188,234,250
+DB	6,6,24,6,48,30,12,54,6,6,24,6,48,30,12,54
+DB	138,138,18,138,36,152,9,174,138,138,18,138,36,152,9,174
+DB	178,178,242,178,249,64,121,75,178,178,242,178,249,64,121,75
+DB	230,230,191,230,99,89,209,133,230,230,191,230,99,89,209,133
+DB	14,14,56,14,112,54,28,126,14,14,56,14,112,54,28,126
+DB	31,31,124,31,248,99,62,231,31,31,124,31,248,99,62,231
+DB	98,98,149,98,55,247,196,85,98,98,149,98,55,247,196,85
+DB	212,212,119,212,238,163,181,58,212,212,119,212,238,163,181,58
+DB	168,168,154,168,41,50,77,129,168,168,154,168,41,50,77,129
+DB	150,150,98,150,196,244,49,82,150,150,98,150,196,244,49,82
+DB	249,249,195,249,155,58,239,98,249,249,195,249,155,58,239,98
+DB	197,197,51,197,102,246,151,163,197,197,51,197,102,246,151,163
+DB	37,37,148,37,53,177,74,16,37,37,148,37,53,177,74,16
+DB	89,89,121,89,242,32,178,171,89,89,121,89,242,32,178,171
+DB	132,132,42,132,84,174,21,208,132,132,42,132,84,174,21,208
+DB	114,114,213,114,183,167,228,197,114,114,213,114,183,167,228,197
+DB	57,57,228,57,213,221,114,236,57,57,228,57,213,221,114,236
+DB	76,76,45,76,90,97,152,22,76,76,45,76,90,97,152,22
+DB	94,94,101,94,202,59,188,148,94,94,101,94,202,59,188,148
+DB	120,120,253,120,231,133,240,159,120,120,253,120,231,133,240,159
+DB	56,56,224,56,221,216,112,229,56,56,224,56,221,216,112,229
+DB	140,140,10,140,20,134,5,152,140,140,10,140,20,134,5,152
+DB	209,209,99,209,198,178,191,23,209,209,99,209,198,178,191,23
+DB	165,165,174,165,65,11,87,228,165,165,174,165,65,11,87,228
+DB	226,226,175,226,67,77,217,161,226,226,175,226,67,77,217,161
+DB	97,97,153,97,47,248,194,78,97,97,153,97,47,248,194,78
+DB	179,179,246,179,241,69,123,66,179,179,246,179,241,69,123,66
+DB	33,33,132,33,21,165,66,52,33,33,132,33,21,165,66,52
+DB	156,156,74,156,148,214,37,8,156,156,74,156,148,214,37,8
+DB	30,30,120,30,240,102,60,238,30,30,120,30,240,102,60,238
+DB	67,67,17,67,34,82,134,97,67,67,17,67,34,82,134,97
+DB	199,199,59,199,118,252,147,177,199,199,59,199,118,252,147,177
+DB	252,252,215,252,179,43,229,79,252,252,215,252,179,43,229,79
+DB	4,4,16,4,32,20,8,36,4,4,16,4,32,20,8,36
+DB	81,81,89,81,178,8,162,227,81,81,89,81,178,8,162,227
+DB	153,153,94,153,188,199,47,37,153,153,94,153,188,199,47,37
+DB	109,109,169,109,79,196,218,34,109,109,169,109,79,196,218,34
+DB	13,13,52,13,104,57,26,101,13,13,52,13,104,57,26,101
+DB	250,250,207,250,131,53,233,121,250,250,207,250,131,53,233,121
+DB	223,223,91,223,182,132,163,105,223,223,91,223,182,132,163,105
+DB	126,126,229,126,215,155,252,169,126,126,229,126,215,155,252,169
+DB	36,36,144,36,61,180,72,25,36,36,144,36,61,180,72,25
+DB	59,59,236,59,197,215,118,254,59,59,236,59,197,215,118,254
+DB	171,171,150,171,49,61,75,154,171,171,150,171,49,61,75,154
+DB	206,206,31,206,62,209,129,240,206,206,31,206,62,209,129,240
+DB	17,17,68,17,136,85,34,153,17,17,68,17,136,85,34,153
+DB	143,143,6,143,12,137,3,131,143,143,6,143,12,137,3,131
+DB	78,78,37,78,74,107,156,4,78,78,37,78,74,107,156,4
+DB	183,183,230,183,209,81,115,102,183,183,230,183,209,81,115,102
+DB	235,235,139,235,11,96,203,224,235,235,139,235,11,96,203,224
+DB	60,60,240,60,253,204,120,193,60,60,240,60,253,204,120,193
+DB	129,129,62,129,124,191,31,253,129,129,62,129,124,191,31,253
+DB	148,148,106,148,212,254,53,64,148,148,106,148,212,254,53,64
+DB	247,247,251,247,235,12,243,28,247,247,251,247,235,12,243,28
+DB	185,185,222,185,161,103,111,24,185,185,222,185,161,103,111,24
+DB	19,19,76,19,152,95,38,139,19,19,76,19,152,95,38,139
+DB	44,44,176,44,125,156,88,81,44,44,176,44,125,156,88,81
+DB	211,211,107,211,214,184,187,5,211,211,107,211,214,184,187,5
+DB	231,231,187,231,107,92,211,140,231,231,187,231,107,92,211,140
+DB	110,110,165,110,87,203,220,57,110,110,165,110,87,203,220,57
+DB	196,196,55,196,110,243,149,170,196,196,55,196,110,243,149,170
+DB	3,3,12,3,24,15,6,27,3,3,12,3,24,15,6,27
+DB	86,86,69,86,138,19,172,220,86,86,69,86,138,19,172,220
+DB	68,68,13,68,26,73,136,94,68,68,13,68,26,73,136,94
+DB	127,127,225,127,223,158,254,160,127,127,225,127,223,158,254,160
+DB	169,169,158,169,33,55,79,136,169,169,158,169,33,55,79,136
+DB	42,42,168,42,77,130,84,103,42,42,168,42,77,130,84,103
+DB	187,187,214,187,177,109,107,10,187,187,214,187,177,109,107,10
+DB	193,193,35,193,70,226,159,135,193,193,35,193,70,226,159,135
+DB	83,83,81,83,162,2,166,241,83,83,81,83,162,2,166,241
+DB	220,220,87,220,174,139,165,114,220,220,87,220,174,139,165,114
+DB	11,11,44,11,88,39,22,83,11,11,44,11,88,39,22,83
+DB	157,157,78,157,156,211,39,1,157,157,78,157,156,211,39,1
+DB	108,108,173,108,71,193,216,43,108,108,173,108,71,193,216,43
+DB	49,49,196,49,149,245,98,164,49,49,196,49,149,245,98,164
+DB	116,116,205,116,135,185,232,243,116,116,205,116,135,185,232,243
+DB	246,246,255,246,227,9,241,21,246,246,255,246,227,9,241,21
+DB	70,70,5,70,10,67,140,76,70,70,5,70,10,67,140,76
+DB	172,172,138,172,9,38,69,165,172,172,138,172,9,38,69,165
+DB	137,137,30,137,60,151,15,181,137,137,30,137,60,151,15,181
+DB	20,20,80,20,160,68,40,180,20,20,80,20,160,68,40,180
+DB	225,225,163,225,91,66,223,186,225,225,163,225,91,66,223,186
+DB	22,22,88,22,176,78,44,166,22,22,88,22,176,78,44,166
+DB	58,58,232,58,205,210,116,247,58,58,232,58,205,210,116,247
+DB	105,105,185,105,111,208,210,6,105,105,185,105,111,208,210,6
+DB	9,9,36,9,72,45,18,65,9,9,36,9,72,45,18,65
+DB	112,112,221,112,167,173,224,215,112,112,221,112,167,173,224,215
+DB	182,182,226,182,217,84,113,111,182,182,226,182,217,84,113,111
+DB	208,208,103,208,206,183,189,30,208,208,103,208,206,183,189,30
+DB	237,237,147,237,59,126,199,214,237,237,147,237,59,126,199,214
+DB	204,204,23,204,46,219,133,226,204,204,23,204,46,219,133,226
+DB	66,66,21,66,42,87,132,104,66,66,21,66,42,87,132,104
+DB	152,152,90,152,180,194,45,44,152,152,90,152,180,194,45,44
+DB	164,164,170,164,73,14,85,237,164,164,170,164,73,14,85,237
+DB	40,40,160,40,93,136,80,117,40,40,160,40,93,136,80,117
+DB	92,92,109,92,218,49,184,134,92,92,109,92,218,49,184,134
+DB	248,248,199,248,147,63,237,107,248,248,199,248,147,63,237,107
+DB	134,134,34,134,68,164,17,194,134,134,34,134,68,164,17,194
+DB	24,35,198,232,135,184,1,79
+DB	54,166,210,245,121,111,145,82
+DB	96,188,155,142,163,12,123,53
+DB	29,224,215,194,46,75,254,87
+DB	21,119,55,229,159,240,74,218
+DB	88,201,41,10,177,160,107,133
+DB	189,93,16,244,203,62,5,103
+DB	228,39,65,139,167,125,149,216
+DB	251,238,124,102,221,23,71,158
+DB	202,45,191,7,173,90,131,51
+EXTERN	__imp_RtlVirtualUnwind:NEAR
+
+ALIGN	16
+se_handler	PROC PRIVATE
+	push	rsi
+	push	rdi
+	push	rbx
+	push	rbp
+	push	r12
+	push	r13
+	push	r14
+	push	r15
+	pushfq
+	sub	rsp,64
+
+	mov	rax,QWORD PTR[120+r8]
+	mov	rbx,QWORD PTR[248+r8]
+
+	lea	r10,QWORD PTR[$L$prologue]
+	cmp	rbx,r10
+	jb	$L$in_prologue
+
+	mov	rax,QWORD PTR[152+r8]
+
+	lea	r10,QWORD PTR[$L$epilogue]
+	cmp	rbx,r10
+	jae	$L$in_prologue
+
+	mov	rax,QWORD PTR[((128+32))+rax]
+	lea	rax,QWORD PTR[48+rax]
+
+	mov	rbx,QWORD PTR[((-8))+rax]
+	mov	rbp,QWORD PTR[((-16))+rax]
+	mov	r12,QWORD PTR[((-24))+rax]
+	mov	r13,QWORD PTR[((-32))+rax]
+	mov	r14,QWORD PTR[((-40))+rax]
+	mov	r15,QWORD PTR[((-48))+rax]
+	mov	QWORD PTR[144+r8],rbx
+	mov	QWORD PTR[160+r8],rbp
+	mov	QWORD PTR[216+r8],r12
+	mov	QWORD PTR[224+r8],r13
+	mov	QWORD PTR[232+r8],r14
+	mov	QWORD PTR[240+r8],r15
+
+$L$in_prologue::
+	mov	rdi,QWORD PTR[8+rax]
+	mov	rsi,QWORD PTR[16+rax]
+	mov	QWORD PTR[152+r8],rax
+	mov	QWORD PTR[168+r8],rsi
+	mov	QWORD PTR[176+r8],rdi
+
+	mov	rdi,QWORD PTR[40+r9]
+	mov	rsi,r8
+	mov	ecx,154
+	DD	0a548f3fch
+
+
+	mov	rsi,r9
+	xor	rcx,rcx
+	mov	rdx,QWORD PTR[8+rsi]
+	mov	r8,QWORD PTR[rsi]
+	mov	r9,QWORD PTR[16+rsi]
+	mov	r10,QWORD PTR[40+rsi]
+	lea	r11,QWORD PTR[56+rsi]
+	lea	r12,QWORD PTR[24+rsi]
+	mov	QWORD PTR[32+rsp],r10
+	mov	QWORD PTR[40+rsp],r11
+	mov	QWORD PTR[48+rsp],r12
+	mov	QWORD PTR[56+rsp],rcx
+	call	QWORD PTR[__imp_RtlVirtualUnwind]
+
+	mov	eax,1
+	add	rsp,64
+	popfq
+	pop	r15
+	pop	r14
+	pop	r13
+	pop	r12
+	pop	rbp
+	pop	rbx
+	pop	rdi
+	pop	rsi
+	DB	0F3h,0C3h		;repret
+se_handler	ENDP
+
+.text$	ENDS
+.pdata	SEGMENT READONLY ALIGN(4)
+ALIGN	4
+	DD	imagerel $L$SEH_begin_whirlpool_block
+	DD	imagerel $L$SEH_end_whirlpool_block
+	DD	imagerel $L$SEH_info_whirlpool_block
+
+.pdata	ENDS
+.xdata	SEGMENT READONLY ALIGN(8)
+ALIGN	8
+$L$SEH_info_whirlpool_block::
+DB	9,0,0,0
+	DD	imagerel se_handler
+
+.xdata	ENDS
+END

deps/openssl/asm/x64-win32-masm/x86_64cpuid.asm

@@ -0,0 +1,186 @@
+OPTION	DOTNAME
+EXTERN	OPENSSL_cpuid_setup:NEAR
+.CRT$XCU	SEGMENT READONLY DWORD
+		DQ	OPENSSL_cpuid_setup
+
+.CRT$XCU	ENDS
+.text$	SEGMENT ALIGN(64) 'CODE'
+
+PUBLIC	OPENSSL_atomic_add
+
+ALIGN	16
+OPENSSL_atomic_add	PROC PUBLIC
+	mov	eax,DWORD PTR[rcx]
+$L$spin::	lea	r8,QWORD PTR[rax*1+rdx]
+DB	0f0h
+
+	cmpxchg	DWORD PTR[rcx],r8d
+	jne	$L$spin
+	mov	eax,r8d
+DB	048h,098h
+
+	DB	0F3h,0C3h		;repret
+OPENSSL_atomic_add	ENDP
+
+PUBLIC	OPENSSL_rdtsc
+
+ALIGN	16
+OPENSSL_rdtsc	PROC PUBLIC
+	rdtsc
+	shl	rdx,32
+	or	rax,rdx
+	DB	0F3h,0C3h		;repret
+OPENSSL_rdtsc	ENDP
+
+PUBLIC	OPENSSL_ia32_cpuid
+
+ALIGN	16
+OPENSSL_ia32_cpuid	PROC PUBLIC
+	mov	r8,rbx
+
+	xor	eax,eax
+	cpuid
+	mov	r11d,eax
+
+	xor	eax,eax
+	cmp	ebx,0756e6547h
+	setne	al
+	mov	r9d,eax
+	cmp	edx,049656e69h
+	setne	al
+	or	r9d,eax
+	cmp	ecx,06c65746eh
+	setne	al
+	or	r9d,eax
+	jz	$L$intel
+
+	cmp	ebx,068747541h
+	setne	al
+	mov	r10d,eax
+	cmp	edx,069746E65h
+	setne	al
+	or	r10d,eax
+	cmp	ecx,0444D4163h
+	setne	al
+	or	r10d,eax
+	jnz	$L$intel
+
+
+	mov	eax,080000000h
+	cpuid
+	cmp	eax,080000008h
+	jb	$L$intel
+
+	mov	eax,080000008h
+	cpuid
+	movzx	r10,cl
+	inc	r10
+
+	mov	eax,1
+	cpuid
+	bt	edx,28
+	jnc	$L$done
+	shr	ebx,16
+	cmp	bl,r10b
+	ja	$L$done
+	and	edx,0efffffffh
+	jmp	$L$done
+
+$L$intel::
+	cmp	r11d,4
+	mov	r10d,-1
+	jb	$L$nocacheinfo
+
+	mov	eax,4
+	mov	ecx,0
+	cpuid
+	mov	r10d,eax
+	shr	r10d,14
+	and	r10d,0fffh
+
+$L$nocacheinfo::
+	mov	eax,1
+	cpuid
+	cmp	r9d,0
+	jne	$L$notintel
+	or	edx,000100000h
+	and	ah,15
+	cmp	ah,15
+	je	$L$notintel
+	or	edx,040000000h
+$L$notintel::
+	bt	edx,28
+	jnc	$L$done
+	and	edx,0efffffffh
+	cmp	r10d,0
+	je	$L$done
+
+	or	edx,010000000h
+	shr	ebx,16
+	cmp	bl,1
+	ja	$L$done
+	and	edx,0efffffffh
+$L$done::
+	shl	rcx,32
+	mov	eax,edx
+	mov	rbx,r8
+	or	rax,rcx
+	DB	0F3h,0C3h		;repret
+OPENSSL_ia32_cpuid	ENDP
+
+PUBLIC	OPENSSL_cleanse
+
+ALIGN	16
+OPENSSL_cleanse	PROC PUBLIC
+	xor	rax,rax
+	cmp	rdx,15
+	jae	$L$ot
+	cmp	rdx,0
+	je	$L$ret
+$L$ittle::
+	mov	BYTE PTR[rcx],al
+	sub	rdx,1
+	lea	rcx,QWORD PTR[1+rcx]
+	jnz	$L$ittle
+$L$ret::
+	DB	0F3h,0C3h		;repret
+ALIGN	16
+$L$ot::
+	test	rcx,7
+	jz	$L$aligned
+	mov	BYTE PTR[rcx],al
+	lea	rdx,QWORD PTR[((-1))+rdx]
+	lea	rcx,QWORD PTR[1+rcx]
+	jmp	$L$ot
+$L$aligned::
+	mov	QWORD PTR[rcx],rax
+	lea	rdx,QWORD PTR[((-8))+rdx]
+	test	rdx,-8
+	lea	rcx,QWORD PTR[8+rcx]
+	jnz	$L$aligned
+	cmp	rdx,0
+	jne	$L$ittle
+	DB	0F3h,0C3h		;repret
+OPENSSL_cleanse	ENDP
+PUBLIC	OPENSSL_wipe_cpu
+
+ALIGN	16
+OPENSSL_wipe_cpu	PROC PUBLIC
+	pxor	xmm0,xmm0
+	pxor	xmm1,xmm1
+	pxor	xmm2,xmm2
+	pxor	xmm3,xmm3
+	pxor	xmm4,xmm4
+	pxor	xmm5,xmm5
+	xor	rcx,rcx
+	xor	rdx,rdx
+	xor	r8,r8
+	xor	r9,r9
+	xor	r10,r10
+	xor	r11,r11
+	lea	rax,QWORD PTR[8+rsp]
+	DB	0F3h,0C3h		;repret
+OPENSSL_wipe_cpu	ENDP
+
+.text$	ENDS
+END

deps/openssl/asm/x86-elf-gas/bf/bf-686.s

@@ -0,0 +1,864 @@
+.file	"bf-686.s"
+.text
+.globl	BF_encrypt
+.type	BF_encrypt,@function
+.align	16
+BF_encrypt:
+.L_BF_encrypt_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+
+	movl	20(%esp),%eax
+	movl	(%eax),%ecx
+	movl	4(%eax),%edx
+
+
+	movl	24(%esp),%edi
+	xorl	%eax,%eax
+	xorl	%ebx,%ebx
+	xorl	(%edi),%ecx
+
+
+	rorl	$16,%ecx
+	movl	4(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	8(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	12(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	16(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	20(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	24(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	28(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	32(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	36(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	40(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	44(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	48(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	52(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	56(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	60(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	64(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+	xorl	68(%edi),%edx
+	movl	20(%esp),%eax
+	movl	%edx,(%eax)
+	movl	%ecx,4(%eax)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	BF_encrypt,.-.L_BF_encrypt_begin
+.globl	BF_decrypt
+.type	BF_decrypt,@function
+.align	16
+BF_decrypt:
+.L_BF_decrypt_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+
+	movl	20(%esp),%eax
+	movl	(%eax),%ecx
+	movl	4(%eax),%edx
+
+
+	movl	24(%esp),%edi
+	xorl	%eax,%eax
+	xorl	%ebx,%ebx
+	xorl	68(%edi),%ecx
+
+
+	rorl	$16,%ecx
+	movl	64(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	60(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	56(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	52(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	48(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	44(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	40(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	36(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	32(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	28(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	24(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	20(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	16(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	12(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+
+	rorl	$16,%ecx
+	movl	8(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+
+	rorl	$16,%edx
+	movl	4(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+	xorl	(%edi),%edx
+	movl	20(%esp),%eax
+	movl	%edx,(%eax)
+	movl	%ecx,4(%eax)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	BF_decrypt,.-.L_BF_decrypt_begin
+.globl	BF_cbc_encrypt
+.type	BF_cbc_encrypt,@function
+.align	16
+BF_cbc_encrypt:
+.L_BF_cbc_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%ebp
+
+	movl	36(%esp),%ebx
+	movl	(%ebx),%esi
+	movl	4(%ebx),%edi
+	pushl	%edi
+	pushl	%esi
+	pushl	%edi
+	pushl	%esi
+	movl	%esp,%ebx
+	movl	36(%esp),%esi
+	movl	40(%esp),%edi
+
+	movl	56(%esp),%ecx
+
+	movl	48(%esp),%eax
+	pushl	%eax
+	pushl	%ebx
+	cmpl	$0,%ecx
+	jz	.L000decrypt
+	andl	$4294967288,%ebp
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	jz	.L001encrypt_finish
+.L002encrypt_loop:
+	movl	(%esi),%ecx
+	movl	4(%esi),%edx
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_BF_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	.L002encrypt_loop
+.L001encrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	.L003finish
+	call	.L004PIC_point
+.L004PIC_point:
+	popl	%edx
+	leal	.L005cbc_enc_jmp_table-.L004PIC_point(%edx),%ecx
+	movl	(%ecx,%ebp,4),%ebp
+	addl	%edx,%ebp
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	jmp	*%ebp
+.L006ej7:
+	movb	6(%esi),%dh
+	shll	$8,%edx
+.L007ej6:
+	movb	5(%esi),%dh
+.L008ej5:
+	movb	4(%esi),%dl
+.L009ej4:
+	movl	(%esi),%ecx
+	jmp	.L010ejend
+.L011ej3:
+	movb	2(%esi),%ch
+	shll	$8,%ecx
+.L012ej2:
+	movb	1(%esi),%ch
+.L013ej1:
+	movb	(%esi),%cl
+.L010ejend:
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_BF_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	jmp	.L003finish
+.L000decrypt:
+	andl	$4294967288,%ebp
+	movl	16(%esp),%eax
+	movl	20(%esp),%ebx
+	jz	.L014decrypt_finish
+.L015decrypt_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_BF_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%ecx,(%edi)
+	movl	%edx,4(%edi)
+	movl	%eax,16(%esp)
+	movl	%ebx,20(%esp)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	.L015decrypt_loop
+.L014decrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	.L003finish
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_BF_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+.L016dj7:
+	rorl	$16,%edx
+	movb	%dl,6(%edi)
+	shrl	$16,%edx
+.L017dj6:
+	movb	%dh,5(%edi)
+.L018dj5:
+	movb	%dl,4(%edi)
+.L019dj4:
+	movl	%ecx,(%edi)
+	jmp	.L020djend
+.L021dj3:
+	rorl	$16,%ecx
+	movb	%cl,2(%edi)
+	shll	$16,%ecx
+.L022dj2:
+	movb	%ch,1(%esi)
+.L023dj1:
+	movb	%cl,(%esi)
+.L020djend:
+	jmp	.L003finish
+.L003finish:
+	movl	60(%esp),%ecx
+	addl	$24,%esp
+	movl	%eax,(%ecx)
+	movl	%ebx,4(%ecx)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	64
+.L005cbc_enc_jmp_table:
+.long	0
+.long	.L013ej1-.L004PIC_point
+.long	.L012ej2-.L004PIC_point
+.long	.L011ej3-.L004PIC_point
+.long	.L009ej4-.L004PIC_point
+.long	.L008ej5-.L004PIC_point
+.long	.L007ej6-.L004PIC_point
+.long	.L006ej7-.L004PIC_point
+.align	64
+.size	BF_cbc_encrypt,.-.L_BF_cbc_encrypt_begin

deps/openssl/asm/x86-elf-gas/bn/x86-mont.s

@@ -0,0 +1,338 @@
+.file	"../openssl/crypto/bn/asm/x86-mont.s"
+.text
+.globl	bn_mul_mont
+.type	bn_mul_mont,@function
+.align	16
+bn_mul_mont:
+.L_bn_mul_mont_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	xorl	%eax,%eax
+	movl	40(%esp),%edi
+	cmpl	$4,%edi
+	jl	.L000just_leave
+	leal	20(%esp),%esi
+	leal	24(%esp),%edx
+	movl	%esp,%ebp
+	addl	$2,%edi
+	negl	%edi
+	leal	-32(%esp,%edi,4),%esp
+	negl	%edi
+	movl	%esp,%eax
+	subl	%edx,%eax
+	andl	$2047,%eax
+	subl	%eax,%esp
+	xorl	%esp,%edx
+	andl	$2048,%edx
+	xorl	$2048,%edx
+	subl	%edx,%esp
+	andl	$-64,%esp
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	16(%esi),%esi
+	movl	(%esi),%esi
+	movl	%eax,4(%esp)
+	movl	%ebx,8(%esp)
+	movl	%ecx,12(%esp)
+	movl	%edx,16(%esp)
+	movl	%esi,20(%esp)
+	leal	-3(%edi),%ebx
+	movl	%ebp,24(%esp)
+	movl	8(%esp),%esi
+	leal	1(%ebx),%ebp
+	movl	12(%esp),%edi
+	xorl	%ecx,%ecx
+	movl	%esi,%edx
+	andl	$1,%ebp
+	subl	%edi,%edx
+	leal	4(%edi,%ebx,4),%eax
+	orl	%edx,%ebp
+	movl	(%edi),%edi
+	jz	.L001bn_sqr_mont
+	movl	%eax,28(%esp)
+	movl	(%esi),%eax
+	xorl	%edx,%edx
+.align	16
+.L002mull:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%eax,%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	movl	(%esi,%ecx,4),%eax
+	cmpl	%ebx,%ecx
+	movl	%ebp,28(%esp,%ecx,4)
+	jl	.L002mull
+	movl	%edx,%ebp
+	mull	%edi
+	movl	20(%esp),%edi
+	addl	%ebp,%eax
+	movl	16(%esp),%esi
+	adcl	$0,%edx
+	imull	32(%esp),%edi
+	movl	%eax,32(%esp,%ebx,4)
+	xorl	%ecx,%ecx
+	movl	%edx,36(%esp,%ebx,4)
+	movl	%ecx,40(%esp,%ebx,4)
+	movl	(%esi),%eax
+	mull	%edi
+	addl	32(%esp),%eax
+	movl	4(%esi),%eax
+	adcl	$0,%edx
+	incl	%ecx
+	jmp	.L0032ndmadd
+.align	16
+.L0041stmadd:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ecx,4),%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	cmpl	%ebx,%ecx
+	movl	%ebp,28(%esp,%ecx,4)
+	jl	.L0041stmadd
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ebx,4),%eax
+	movl	20(%esp),%edi
+	adcl	$0,%edx
+	movl	16(%esp),%esi
+	addl	%eax,%ebp
+	adcl	$0,%edx
+	imull	32(%esp),%edi
+	xorl	%ecx,%ecx
+	addl	36(%esp,%ebx,4),%edx
+	movl	%ebp,32(%esp,%ebx,4)
+	adcl	$0,%ecx
+	movl	(%esi),%eax
+	movl	%edx,36(%esp,%ebx,4)
+	movl	%ecx,40(%esp,%ebx,4)
+	mull	%edi
+	addl	32(%esp),%eax
+	movl	4(%esi),%eax
+	adcl	$0,%edx
+	movl	$1,%ecx
+.align	16
+.L0032ndmadd:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ecx,4),%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	cmpl	%ebx,%ecx
+	movl	%ebp,24(%esp,%ecx,4)
+	jl	.L0032ndmadd
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ebx,4),%ebp
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	adcl	$0,%edx
+	movl	%ebp,28(%esp,%ebx,4)
+	xorl	%eax,%eax
+	movl	12(%esp),%ecx
+	addl	36(%esp,%ebx,4),%edx
+	adcl	40(%esp,%ebx,4),%eax
+	leal	4(%ecx),%ecx
+	movl	%edx,32(%esp,%ebx,4)
+	cmpl	28(%esp),%ecx
+	movl	%eax,36(%esp,%ebx,4)
+	je	.L005common_tail
+	movl	(%ecx),%edi
+	movl	8(%esp),%esi
+	movl	%ecx,12(%esp)
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	movl	(%esi),%eax
+	jmp	.L0041stmadd
+.align	16
+.L001bn_sqr_mont:
+	movl	%ebx,(%esp)
+	movl	%ecx,12(%esp)
+	movl	%edi,%eax
+	mull	%edi
+	movl	%eax,32(%esp)
+	movl	%edx,%ebx
+	shrl	$1,%edx
+	andl	$1,%ebx
+	incl	%ecx
+.align	16
+.L006sqr:
+	movl	(%esi,%ecx,4),%eax
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%ebp,%eax
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	leal	(%ebx,%eax,2),%ebp
+	shrl	$31,%eax
+	cmpl	(%esp),%ecx
+	movl	%eax,%ebx
+	movl	%ebp,28(%esp,%ecx,4)
+	jl	.L006sqr
+	movl	(%esi,%ecx,4),%eax
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%ebp,%eax
+	movl	20(%esp),%edi
+	adcl	$0,%edx
+	movl	16(%esp),%esi
+	leal	(%ebx,%eax,2),%ebp
+	imull	32(%esp),%edi
+	shrl	$31,%eax
+	movl	%ebp,32(%esp,%ecx,4)
+	leal	(%eax,%edx,2),%ebp
+	movl	(%esi),%eax
+	shrl	$31,%edx
+	movl	%ebp,36(%esp,%ecx,4)
+	movl	%edx,40(%esp,%ecx,4)
+	mull	%edi
+	addl	32(%esp),%eax
+	movl	%ecx,%ebx
+	adcl	$0,%edx
+	movl	4(%esi),%eax
+	movl	$1,%ecx
+.align	16
+.L0073rdmadd:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ecx,4),%ebp
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	4(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	movl	%ebp,28(%esp,%ecx,4)
+	movl	%edx,%ebp
+	mull	%edi
+	addl	36(%esp,%ecx,4),%ebp
+	leal	2(%ecx),%ecx
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	cmpl	%ebx,%ecx
+	movl	%ebp,24(%esp,%ecx,4)
+	jl	.L0073rdmadd
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ebx,4),%ebp
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	adcl	$0,%edx
+	movl	%ebp,28(%esp,%ebx,4)
+	movl	12(%esp),%ecx
+	xorl	%eax,%eax
+	movl	8(%esp),%esi
+	addl	36(%esp,%ebx,4),%edx
+	adcl	40(%esp,%ebx,4),%eax
+	movl	%edx,32(%esp,%ebx,4)
+	cmpl	%ebx,%ecx
+	movl	%eax,36(%esp,%ebx,4)
+	je	.L005common_tail
+	movl	4(%esi,%ecx,4),%edi
+	leal	1(%ecx),%ecx
+	movl	%edi,%eax
+	movl	%ecx,12(%esp)
+	mull	%edi
+	addl	32(%esp,%ecx,4),%eax
+	adcl	$0,%edx
+	movl	%eax,32(%esp,%ecx,4)
+	xorl	%ebp,%ebp
+	cmpl	%ebx,%ecx
+	leal	1(%ecx),%ecx
+	je	.L008sqrlast
+	movl	%edx,%ebx
+	shrl	$1,%edx
+	andl	$1,%ebx
+.align	16
+.L009sqradd:
+	movl	(%esi,%ecx,4),%eax
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%ebp,%eax
+	leal	(%eax,%eax,1),%ebp
+	adcl	$0,%edx
+	shrl	$31,%eax
+	addl	32(%esp,%ecx,4),%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%eax
+	addl	%ebx,%ebp
+	adcl	$0,%eax
+	cmpl	(%esp),%ecx
+	movl	%ebp,28(%esp,%ecx,4)
+	movl	%eax,%ebx
+	jle	.L009sqradd
+	movl	%edx,%ebp
+	addl	%edx,%edx
+	shrl	$31,%ebp
+	addl	%ebx,%edx
+	adcl	$0,%ebp
+.L008sqrlast:
+	movl	20(%esp),%edi
+	movl	16(%esp),%esi
+	imull	32(%esp),%edi
+	addl	32(%esp,%ecx,4),%edx
+	movl	(%esi),%eax
+	adcl	$0,%ebp
+	movl	%edx,32(%esp,%ecx,4)
+	movl	%ebp,36(%esp,%ecx,4)
+	mull	%edi
+	addl	32(%esp),%eax
+	leal	-1(%ecx),%ebx
+	adcl	$0,%edx
+	movl	$1,%ecx
+	movl	4(%esi),%eax
+	jmp	.L0073rdmadd
+.align	16
+.L005common_tail:
+	movl	16(%esp),%ebp
+	movl	4(%esp),%edi
+	leal	32(%esp),%esi
+	movl	(%esi),%eax
+	movl	%ebx,%ecx
+	xorl	%edx,%edx
+.align	16
+.L010sub:
+	sbbl	(%ebp,%edx,4),%eax
+	movl	%eax,(%edi,%edx,4)
+	decl	%ecx
+	movl	4(%esi,%edx,4),%eax
+	leal	1(%edx),%edx
+	jge	.L010sub
+	sbbl	$0,%eax
+	andl	%eax,%esi
+	notl	%eax
+	movl	%edi,%ebp
+	andl	%eax,%ebp
+	orl	%ebp,%esi
+.align	16
+.L011copy:
+	movl	(%esi,%ebx,4),%eax
+	movl	%eax,(%edi,%ebx,4)
+	movl	%ecx,32(%esp,%ebx,4)
+	decl	%ebx
+	jge	.L011copy
+	movl	24(%esp),%esp
+	movl	$1,%eax
+.L000just_leave:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	bn_mul_mont,.-.L_bn_mul_mont_begin
+.byte	77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
+.byte	112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
+.byte	54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
+.byte	32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
+.byte	111,114,103,62,0

deps/openssl/asm/x86-elf-gas/cast/cast-586.s

@@ -0,0 +1,933 @@
+.file	"cast-586.s"
+.text
+.globl	CAST_encrypt
+.type	CAST_encrypt,@function
+.align	16
+CAST_encrypt:
+.L_CAST_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	movl	12(%esp),%ebx
+	movl	16(%esp),%ebp
+	pushl	%esi
+	pushl	%edi
+
+	movl	(%ebx),%edi
+	movl	4(%ebx),%esi
+
+	movl	128(%ebp),%eax
+	pushl	%eax
+	xorl	%eax,%eax
+
+	movl	(%ebp),%edx
+	movl	4(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	8(%ebp),%edx
+	movl	12(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	16(%ebp),%edx
+	movl	20(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	24(%ebp),%edx
+	movl	28(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	32(%ebp),%edx
+	movl	36(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	40(%ebp),%edx
+	movl	44(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	48(%ebp),%edx
+	movl	52(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	56(%ebp),%edx
+	movl	60(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	64(%ebp),%edx
+	movl	68(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	72(%ebp),%edx
+	movl	76(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	80(%ebp),%edx
+	movl	84(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	88(%ebp),%edx
+	movl	92(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	popl	%edx
+	orl	%edx,%edx
+	jnz	.L000cast_enc_done
+
+	movl	96(%ebp),%edx
+	movl	100(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	104(%ebp),%edx
+	movl	108(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	112(%ebp),%edx
+	movl	116(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	120(%ebp),%edx
+	movl	124(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+.L000cast_enc_done:
+	nop
+	movl	20(%esp),%eax
+	movl	%edi,4(%eax)
+	movl	%esi,(%eax)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	CAST_encrypt,.-.L_CAST_encrypt_begin
+.globl	CAST_decrypt
+.type	CAST_decrypt,@function
+.align	16
+CAST_decrypt:
+.L_CAST_decrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	movl	12(%esp),%ebx
+	movl	16(%esp),%ebp
+	pushl	%esi
+	pushl	%edi
+
+	movl	(%ebx),%edi
+	movl	4(%ebx),%esi
+
+	movl	128(%ebp),%eax
+	orl	%eax,%eax
+	jnz	.L001cast_dec_skip
+	xorl	%eax,%eax
+
+	movl	120(%ebp),%edx
+	movl	124(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	112(%ebp),%edx
+	movl	116(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	104(%ebp),%edx
+	movl	108(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	96(%ebp),%edx
+	movl	100(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+.L001cast_dec_skip:
+
+	movl	88(%ebp),%edx
+	movl	92(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	80(%ebp),%edx
+	movl	84(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	72(%ebp),%edx
+	movl	76(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	64(%ebp),%edx
+	movl	68(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	56(%ebp),%edx
+	movl	60(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	48(%ebp),%edx
+	movl	52(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	40(%ebp),%edx
+	movl	44(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	32(%ebp),%edx
+	movl	36(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	24(%ebp),%edx
+	movl	28(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	16(%ebp),%edx
+	movl	20(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+
+	movl	8(%ebp),%edx
+	movl	12(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+
+	movl	(%ebp),%edx
+	movl	4(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	CAST_S_table0(,%ecx,4),%ecx
+	movl	CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+	nop
+	movl	20(%esp),%eax
+	movl	%edi,4(%eax)
+	movl	%esi,(%eax)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	CAST_decrypt,.-.L_CAST_decrypt_begin
+.globl	CAST_cbc_encrypt
+.type	CAST_cbc_encrypt,@function
+.align	16
+CAST_cbc_encrypt:
+.L_CAST_cbc_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%ebp
+
+	movl	36(%esp),%ebx
+	movl	(%ebx),%esi
+	movl	4(%ebx),%edi
+	pushl	%edi
+	pushl	%esi
+	pushl	%edi
+	pushl	%esi
+	movl	%esp,%ebx
+	movl	36(%esp),%esi
+	movl	40(%esp),%edi
+
+	movl	56(%esp),%ecx
+
+	movl	48(%esp),%eax
+	pushl	%eax
+	pushl	%ebx
+	cmpl	$0,%ecx
+	jz	.L002decrypt
+	andl	$4294967288,%ebp
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	jz	.L003encrypt_finish
+.L004encrypt_loop:
+	movl	(%esi),%ecx
+	movl	4(%esi),%edx
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_CAST_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	.L004encrypt_loop
+.L003encrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	.L005finish
+	call	.L006PIC_point
+.L006PIC_point:
+	popl	%edx
+	leal	.L007cbc_enc_jmp_table-.L006PIC_point(%edx),%ecx
+	movl	(%ecx,%ebp,4),%ebp
+	addl	%edx,%ebp
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	jmp	*%ebp
+.L008ej7:
+	movb	6(%esi),%dh
+	shll	$8,%edx
+.L009ej6:
+	movb	5(%esi),%dh
+.L010ej5:
+	movb	4(%esi),%dl
+.L011ej4:
+	movl	(%esi),%ecx
+	jmp	.L012ejend
+.L013ej3:
+	movb	2(%esi),%ch
+	shll	$8,%ecx
+.L014ej2:
+	movb	1(%esi),%ch
+.L015ej1:
+	movb	(%esi),%cl
+.L012ejend:
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_CAST_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	jmp	.L005finish
+.L002decrypt:
+	andl	$4294967288,%ebp
+	movl	16(%esp),%eax
+	movl	20(%esp),%ebx
+	jz	.L016decrypt_finish
+.L017decrypt_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_CAST_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%ecx,(%edi)
+	movl	%edx,4(%edi)
+	movl	%eax,16(%esp)
+	movl	%ebx,20(%esp)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	.L017decrypt_loop
+.L016decrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	.L005finish
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_CAST_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+.L018dj7:
+	rorl	$16,%edx
+	movb	%dl,6(%edi)
+	shrl	$16,%edx
+.L019dj6:
+	movb	%dh,5(%edi)
+.L020dj5:
+	movb	%dl,4(%edi)
+.L021dj4:
+	movl	%ecx,(%edi)
+	jmp	.L022djend
+.L023dj3:
+	rorl	$16,%ecx
+	movb	%cl,2(%edi)
+	shll	$16,%ecx
+.L024dj2:
+	movb	%ch,1(%esi)
+.L025dj1:
+	movb	%cl,(%esi)
+.L022djend:
+	jmp	.L005finish
+.L005finish:
+	movl	60(%esp),%ecx
+	addl	$24,%esp
+	movl	%eax,(%ecx)
+	movl	%ebx,4(%ecx)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	64
+.L007cbc_enc_jmp_table:
+.long	0
+.long	.L015ej1-.L006PIC_point
+.long	.L014ej2-.L006PIC_point
+.long	.L013ej3-.L006PIC_point
+.long	.L011ej4-.L006PIC_point
+.long	.L010ej5-.L006PIC_point
+.long	.L009ej6-.L006PIC_point
+.long	.L008ej7-.L006PIC_point
+.align	64
+.size	CAST_cbc_encrypt,.-.L_CAST_cbc_encrypt_begin

deps/openssl/asm/x86-elf-gas/des/crypt586.s

@@ -0,0 +1,875 @@
+.file	"crypt586.s"
+.text
+.globl	fcrypt_body
+.type	fcrypt_body,@function
+.align	16
+fcrypt_body:
+.L_fcrypt_body_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+
+	xorl	%edi,%edi
+	xorl	%esi,%esi
+	leal	DES_SPtrans,%edx
+	pushl	%edx
+	movl	28(%esp),%ebp
+	pushl	$25
+.L000start:
+
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	4(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	8(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	12(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	16(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	20(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	24(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	28(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	32(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	36(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	40(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	44(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	48(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	52(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	56(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	60(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	64(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	68(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	72(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	76(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	80(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	84(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	88(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	92(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	96(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	100(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	104(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	108(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	112(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	116(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	120(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	124(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+	movl	(%esp),%ebx
+	movl	%edi,%eax
+	decl	%ebx
+	movl	%esi,%edi
+	movl	%eax,%esi
+	movl	%ebx,(%esp)
+	jnz	.L000start
+
+
+	movl	28(%esp),%edx
+	rorl	$1,%edi
+	movl	%esi,%eax
+	xorl	%edi,%esi
+	andl	$0xaaaaaaaa,%esi
+	xorl	%esi,%eax
+	xorl	%esi,%edi
+
+	roll	$23,%eax
+	movl	%eax,%esi
+	xorl	%edi,%eax
+	andl	$0x03fc03fc,%eax
+	xorl	%eax,%esi
+	xorl	%eax,%edi
+
+	roll	$10,%esi
+	movl	%esi,%eax
+	xorl	%edi,%esi
+	andl	$0x33333333,%esi
+	xorl	%esi,%eax
+	xorl	%esi,%edi
+
+	roll	$18,%edi
+	movl	%edi,%esi
+	xorl	%eax,%edi
+	andl	$0xfff0000f,%edi
+	xorl	%edi,%esi
+	xorl	%edi,%eax
+
+	roll	$12,%esi
+	movl	%esi,%edi
+	xorl	%eax,%esi
+	andl	$0xf0f0f0f0,%esi
+	xorl	%esi,%edi
+	xorl	%esi,%eax
+
+	rorl	$4,%eax
+	movl	%eax,(%edx)
+	movl	%edi,4(%edx)
+	addl	$8,%esp
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	fcrypt_body,.-.L_fcrypt_body_begin

deps/openssl/asm/x86-elf-gas/md5/md5-586.s

@@ -0,0 +1,679 @@
+.file	"../openssl/crypto/md5/asm/md5-586.s"
+.text
+.globl	md5_block_asm_data_order
+.type	md5_block_asm_data_order,@function
+.align	16
+md5_block_asm_data_order:
+.L_md5_block_asm_data_order_begin:
+	pushl	%esi
+	pushl	%edi
+	movl	12(%esp),%edi
+	movl	16(%esp),%esi
+	movl	20(%esp),%ecx
+	pushl	%ebp
+	shll	$6,%ecx
+	pushl	%ebx
+	addl	%esi,%ecx
+	subl	$64,%ecx
+	movl	(%edi),%eax
+	pushl	%ecx
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+.L000start:
+
+
+	movl	%ecx,%edi
+	movl	(%esi),%ebp
+
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	leal	3614090360(%eax,%ebp,1),%eax
+	xorl	%edx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$7,%eax
+	movl	4(%esi),%ebp
+	addl	%ebx,%eax
+
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	leal	3905402710(%edx,%ebp,1),%edx
+	xorl	%ecx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$12,%edx
+	movl	8(%esi),%ebp
+	addl	%eax,%edx
+
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	leal	606105819(%ecx,%ebp,1),%ecx
+	xorl	%ebx,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$17,%ecx
+	movl	12(%esi),%ebp
+	addl	%edx,%ecx
+
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	leal	3250441966(%ebx,%ebp,1),%ebx
+	xorl	%eax,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$22,%ebx
+	movl	16(%esi),%ebp
+	addl	%ecx,%ebx
+
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	leal	4118548399(%eax,%ebp,1),%eax
+	xorl	%edx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$7,%eax
+	movl	20(%esi),%ebp
+	addl	%ebx,%eax
+
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	leal	1200080426(%edx,%ebp,1),%edx
+	xorl	%ecx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$12,%edx
+	movl	24(%esi),%ebp
+	addl	%eax,%edx
+
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	leal	2821735955(%ecx,%ebp,1),%ecx
+	xorl	%ebx,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$17,%ecx
+	movl	28(%esi),%ebp
+	addl	%edx,%ecx
+
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	leal	4249261313(%ebx,%ebp,1),%ebx
+	xorl	%eax,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$22,%ebx
+	movl	32(%esi),%ebp
+	addl	%ecx,%ebx
+
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	leal	1770035416(%eax,%ebp,1),%eax
+	xorl	%edx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$7,%eax
+	movl	36(%esi),%ebp
+	addl	%ebx,%eax
+
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	leal	2336552879(%edx,%ebp,1),%edx
+	xorl	%ecx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$12,%edx
+	movl	40(%esi),%ebp
+	addl	%eax,%edx
+
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	leal	4294925233(%ecx,%ebp,1),%ecx
+	xorl	%ebx,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$17,%ecx
+	movl	44(%esi),%ebp
+	addl	%edx,%ecx
+
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	leal	2304563134(%ebx,%ebp,1),%ebx
+	xorl	%eax,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$22,%ebx
+	movl	48(%esi),%ebp
+	addl	%ecx,%ebx
+
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	leal	1804603682(%eax,%ebp,1),%eax
+	xorl	%edx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$7,%eax
+	movl	52(%esi),%ebp
+	addl	%ebx,%eax
+
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	leal	4254626195(%edx,%ebp,1),%edx
+	xorl	%ecx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$12,%edx
+	movl	56(%esi),%ebp
+	addl	%eax,%edx
+
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	leal	2792965006(%ecx,%ebp,1),%ecx
+	xorl	%ebx,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$17,%ecx
+	movl	60(%esi),%ebp
+	addl	%edx,%ecx
+
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	leal	1236535329(%ebx,%ebp,1),%ebx
+	xorl	%eax,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$22,%ebx
+	movl	4(%esi),%ebp
+	addl	%ecx,%ebx
+
+
+
+	leal	4129170786(%eax,%ebp,1),%eax
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	movl	24(%esi),%ebp
+	xorl	%ecx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$5,%eax
+	addl	%ebx,%eax
+
+	leal	3225465664(%edx,%ebp,1),%edx
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	movl	44(%esi),%ebp
+	xorl	%ebx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$9,%edx
+	addl	%eax,%edx
+
+	leal	643717713(%ecx,%ebp,1),%ecx
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	movl	(%esi),%ebp
+	xorl	%eax,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$14,%ecx
+	addl	%edx,%ecx
+
+	leal	3921069994(%ebx,%ebp,1),%ebx
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	movl	20(%esi),%ebp
+	xorl	%edx,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+
+	leal	3593408605(%eax,%ebp,1),%eax
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	movl	40(%esi),%ebp
+	xorl	%ecx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$5,%eax
+	addl	%ebx,%eax
+
+	leal	38016083(%edx,%ebp,1),%edx
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	movl	60(%esi),%ebp
+	xorl	%ebx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$9,%edx
+	addl	%eax,%edx
+
+	leal	3634488961(%ecx,%ebp,1),%ecx
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	movl	16(%esi),%ebp
+	xorl	%eax,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$14,%ecx
+	addl	%edx,%ecx
+
+	leal	3889429448(%ebx,%ebp,1),%ebx
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	movl	36(%esi),%ebp
+	xorl	%edx,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+
+	leal	568446438(%eax,%ebp,1),%eax
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	movl	56(%esi),%ebp
+	xorl	%ecx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$5,%eax
+	addl	%ebx,%eax
+
+	leal	3275163606(%edx,%ebp,1),%edx
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	movl	12(%esi),%ebp
+	xorl	%ebx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$9,%edx
+	addl	%eax,%edx
+
+	leal	4107603335(%ecx,%ebp,1),%ecx
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	movl	32(%esi),%ebp
+	xorl	%eax,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$14,%ecx
+	addl	%edx,%ecx
+
+	leal	1163531501(%ebx,%ebp,1),%ebx
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	movl	52(%esi),%ebp
+	xorl	%edx,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+
+	leal	2850285829(%eax,%ebp,1),%eax
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	movl	8(%esi),%ebp
+	xorl	%ecx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$5,%eax
+	addl	%ebx,%eax
+
+	leal	4243563512(%edx,%ebp,1),%edx
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	movl	28(%esi),%ebp
+	xorl	%ebx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$9,%edx
+	addl	%eax,%edx
+
+	leal	1735328473(%ecx,%ebp,1),%ecx
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	movl	48(%esi),%ebp
+	xorl	%eax,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$14,%ecx
+	addl	%edx,%ecx
+
+	leal	2368359562(%ebx,%ebp,1),%ebx
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	movl	20(%esi),%ebp
+	xorl	%edx,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+
+
+
+	xorl	%edx,%edi
+	xorl	%ebx,%edi
+	leal	4294588738(%eax,%ebp,1),%eax
+	addl	%edi,%eax
+	roll	$4,%eax
+	movl	32(%esi),%ebp
+	movl	%ebx,%edi
+
+	leal	2272392833(%edx,%ebp,1),%edx
+	addl	%ebx,%eax
+	xorl	%ecx,%edi
+	xorl	%eax,%edi
+	movl	44(%esi),%ebp
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$11,%edx
+	addl	%eax,%edx
+
+	xorl	%ebx,%edi
+	xorl	%edx,%edi
+	leal	1839030562(%ecx,%ebp,1),%ecx
+	addl	%edi,%ecx
+	roll	$16,%ecx
+	movl	56(%esi),%ebp
+	movl	%edx,%edi
+
+	leal	4259657740(%ebx,%ebp,1),%ebx
+	addl	%edx,%ecx
+	xorl	%eax,%edi
+	xorl	%ecx,%edi
+	movl	4(%esi),%ebp
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$23,%ebx
+	addl	%ecx,%ebx
+
+	xorl	%edx,%edi
+	xorl	%ebx,%edi
+	leal	2763975236(%eax,%ebp,1),%eax
+	addl	%edi,%eax
+	roll	$4,%eax
+	movl	16(%esi),%ebp
+	movl	%ebx,%edi
+
+	leal	1272893353(%edx,%ebp,1),%edx
+	addl	%ebx,%eax
+	xorl	%ecx,%edi
+	xorl	%eax,%edi
+	movl	28(%esi),%ebp
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$11,%edx
+	addl	%eax,%edx
+
+	xorl	%ebx,%edi
+	xorl	%edx,%edi
+	leal	4139469664(%ecx,%ebp,1),%ecx
+	addl	%edi,%ecx
+	roll	$16,%ecx
+	movl	40(%esi),%ebp
+	movl	%edx,%edi
+
+	leal	3200236656(%ebx,%ebp,1),%ebx
+	addl	%edx,%ecx
+	xorl	%eax,%edi
+	xorl	%ecx,%edi
+	movl	52(%esi),%ebp
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$23,%ebx
+	addl	%ecx,%ebx
+
+	xorl	%edx,%edi
+	xorl	%ebx,%edi
+	leal	681279174(%eax,%ebp,1),%eax
+	addl	%edi,%eax
+	roll	$4,%eax
+	movl	(%esi),%ebp
+	movl	%ebx,%edi
+
+	leal	3936430074(%edx,%ebp,1),%edx
+	addl	%ebx,%eax
+	xorl	%ecx,%edi
+	xorl	%eax,%edi
+	movl	12(%esi),%ebp
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$11,%edx
+	addl	%eax,%edx
+
+	xorl	%ebx,%edi
+	xorl	%edx,%edi
+	leal	3572445317(%ecx,%ebp,1),%ecx
+	addl	%edi,%ecx
+	roll	$16,%ecx
+	movl	24(%esi),%ebp
+	movl	%edx,%edi
+
+	leal	76029189(%ebx,%ebp,1),%ebx
+	addl	%edx,%ecx
+	xorl	%eax,%edi
+	xorl	%ecx,%edi
+	movl	36(%esi),%ebp
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$23,%ebx
+	addl	%ecx,%ebx
+
+	xorl	%edx,%edi
+	xorl	%ebx,%edi
+	leal	3654602809(%eax,%ebp,1),%eax
+	addl	%edi,%eax
+	roll	$4,%eax
+	movl	48(%esi),%ebp
+	movl	%ebx,%edi
+
+	leal	3873151461(%edx,%ebp,1),%edx
+	addl	%ebx,%eax
+	xorl	%ecx,%edi
+	xorl	%eax,%edi
+	movl	60(%esi),%ebp
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$11,%edx
+	addl	%eax,%edx
+
+	xorl	%ebx,%edi
+	xorl	%edx,%edi
+	leal	530742520(%ecx,%ebp,1),%ecx
+	addl	%edi,%ecx
+	roll	$16,%ecx
+	movl	8(%esi),%ebp
+	movl	%edx,%edi
+
+	leal	3299628645(%ebx,%ebp,1),%ebx
+	addl	%edx,%ecx
+	xorl	%eax,%edi
+	xorl	%ecx,%edi
+	movl	(%esi),%ebp
+	addl	%edi,%ebx
+	movl	$-1,%edi
+	roll	$23,%ebx
+	addl	%ecx,%ebx
+
+
+
+	xorl	%edx,%edi
+	orl	%ebx,%edi
+	leal	4096336452(%eax,%ebp,1),%eax
+	xorl	%ecx,%edi
+	movl	28(%esi),%ebp
+	addl	%edi,%eax
+	movl	$-1,%edi
+	roll	$6,%eax
+	xorl	%ecx,%edi
+	addl	%ebx,%eax
+
+	orl	%eax,%edi
+	leal	1126891415(%edx,%ebp,1),%edx
+	xorl	%ebx,%edi
+	movl	56(%esi),%ebp
+	addl	%edi,%edx
+	movl	$-1,%edi
+	roll	$10,%edx
+	xorl	%ebx,%edi
+	addl	%eax,%edx
+
+	orl	%edx,%edi
+	leal	2878612391(%ecx,%ebp,1),%ecx
+	xorl	%eax,%edi
+	movl	20(%esi),%ebp
+	addl	%edi,%ecx
+	movl	$-1,%edi
+	roll	$15,%ecx
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+
+	orl	%ecx,%edi
+	leal	4237533241(%ebx,%ebp,1),%ebx
+	xorl	%edx,%edi
+	movl	48(%esi),%ebp
+	addl	%edi,%ebx
+	movl	$-1,%edi
+	roll	$21,%ebx
+	xorl	%edx,%edi
+	addl	%ecx,%ebx
+
+	orl	%ebx,%edi
+	leal	1700485571(%eax,%ebp,1),%eax
+	xorl	%ecx,%edi
+	movl	12(%esi),%ebp
+	addl	%edi,%eax
+	movl	$-1,%edi
+	roll	$6,%eax
+	xorl	%ecx,%edi
+	addl	%ebx,%eax
+
+	orl	%eax,%edi
+	leal	2399980690(%edx,%ebp,1),%edx
+	xorl	%ebx,%edi
+	movl	40(%esi),%ebp
+	addl	%edi,%edx
+	movl	$-1,%edi
+	roll	$10,%edx
+	xorl	%ebx,%edi
+	addl	%eax,%edx
+
+	orl	%edx,%edi
+	leal	4293915773(%ecx,%ebp,1),%ecx
+	xorl	%eax,%edi
+	movl	4(%esi),%ebp
+	addl	%edi,%ecx
+	movl	$-1,%edi
+	roll	$15,%ecx
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+
+	orl	%ecx,%edi
+	leal	2240044497(%ebx,%ebp,1),%ebx
+	xorl	%edx,%edi
+	movl	32(%esi),%ebp
+	addl	%edi,%ebx
+	movl	$-1,%edi
+	roll	$21,%ebx
+	xorl	%edx,%edi
+	addl	%ecx,%ebx
+
+	orl	%ebx,%edi
+	leal	1873313359(%eax,%ebp,1),%eax
+	xorl	%ecx,%edi
+	movl	60(%esi),%ebp
+	addl	%edi,%eax
+	movl	$-1,%edi
+	roll	$6,%eax
+	xorl	%ecx,%edi
+	addl	%ebx,%eax
+
+	orl	%eax,%edi
+	leal	4264355552(%edx,%ebp,1),%edx
+	xorl	%ebx,%edi
+	movl	24(%esi),%ebp
+	addl	%edi,%edx
+	movl	$-1,%edi
+	roll	$10,%edx
+	xorl	%ebx,%edi
+	addl	%eax,%edx
+
+	orl	%edx,%edi
+	leal	2734768916(%ecx,%ebp,1),%ecx
+	xorl	%eax,%edi
+	movl	52(%esi),%ebp
+	addl	%edi,%ecx
+	movl	$-1,%edi
+	roll	$15,%ecx
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+
+	orl	%ecx,%edi
+	leal	1309151649(%ebx,%ebp,1),%ebx
+	xorl	%edx,%edi
+	movl	16(%esi),%ebp
+	addl	%edi,%ebx
+	movl	$-1,%edi
+	roll	$21,%ebx
+	xorl	%edx,%edi
+	addl	%ecx,%ebx
+
+	orl	%ebx,%edi
+	leal	4149444226(%eax,%ebp,1),%eax
+	xorl	%ecx,%edi
+	movl	44(%esi),%ebp
+	addl	%edi,%eax
+	movl	$-1,%edi
+	roll	$6,%eax
+	xorl	%ecx,%edi
+	addl	%ebx,%eax
+
+	orl	%eax,%edi
+	leal	3174756917(%edx,%ebp,1),%edx
+	xorl	%ebx,%edi
+	movl	8(%esi),%ebp
+	addl	%edi,%edx
+	movl	$-1,%edi
+	roll	$10,%edx
+	xorl	%ebx,%edi
+	addl	%eax,%edx
+
+	orl	%edx,%edi
+	leal	718787259(%ecx,%ebp,1),%ecx
+	xorl	%eax,%edi
+	movl	36(%esi),%ebp
+	addl	%edi,%ecx
+	movl	$-1,%edi
+	roll	$15,%ecx
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+
+	orl	%ecx,%edi
+	leal	3951481745(%ebx,%ebp,1),%ebx
+	xorl	%edx,%edi
+	movl	24(%esp),%ebp
+	addl	%edi,%ebx
+	addl	$64,%esi
+	roll	$21,%ebx
+	movl	(%ebp),%edi
+	addl	%ecx,%ebx
+	addl	%edi,%eax
+	movl	4(%ebp),%edi
+	addl	%edi,%ebx
+	movl	8(%ebp),%edi
+	addl	%edi,%ecx
+	movl	12(%ebp),%edi
+	addl	%edi,%edx
+	movl	%eax,(%ebp)
+	movl	%ebx,4(%ebp)
+	movl	(%esp),%edi
+	movl	%ecx,8(%ebp)
+	movl	%edx,12(%ebp)
+	cmpl	%esi,%edi
+	jae	.L000start
+	popl	%eax
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret
+.size	md5_block_asm_data_order,.-.L_md5_block_asm_data_order_begin

deps/openssl/asm/x86-elf-gas/rc4/rc4-586.s

@@ -0,0 +1,230 @@
+.file	"rc4-586.s"
+.text
+.globl	RC4
+.type	RC4,@function
+.align	16
+RC4:
+.L_RC4_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	movl	24(%esp),%edx
+	movl	28(%esp),%esi
+	movl	32(%esp),%ebp
+	xorl	%eax,%eax
+	xorl	%ebx,%ebx
+	cmpl	$0,%edx
+	je	.L000abort
+	movb	(%edi),%al
+	movb	4(%edi),%bl
+	addl	$8,%edi
+	leal	(%esi,%edx,1),%ecx
+	subl	%esi,%ebp
+	movl	%ecx,24(%esp)
+	incb	%al
+	cmpl	$-1,256(%edi)
+	je	.L001RC4_CHAR
+	movl	(%edi,%eax,4),%ecx
+	andl	$-4,%edx
+	jz	.L002loop1
+	leal	-4(%esi,%edx,1),%edx
+	movl	%edx,28(%esp)
+	movl	%ebp,32(%esp)
+.align	16
+.L003loop4:
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	movl	(%edi,%eax,4),%ecx
+	movl	(%edi,%edx,4),%ebp
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	rorl	$8,%ebp
+	movl	(%edi,%eax,4),%ecx
+	orl	(%edi,%edx,4),%ebp
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	rorl	$8,%ebp
+	movl	(%edi,%eax,4),%ecx
+	orl	(%edi,%edx,4),%ebp
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	rorl	$8,%ebp
+	movl	32(%esp),%ecx
+	orl	(%edi,%edx,4),%ebp
+	rorl	$8,%ebp
+	xorl	(%esi),%ebp
+	cmpl	28(%esp),%esi
+	movl	%ebp,(%ecx,%esi,1)
+	leal	4(%esi),%esi
+	movl	(%edi,%eax,4),%ecx
+	jb	.L003loop4
+	cmpl	24(%esp),%esi
+	je	.L004done
+	movl	32(%esp),%ebp
+.align	16
+.L002loop1:
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	movl	(%edi,%edx,4),%edx
+	xorb	(%esi),%dl
+	leal	1(%esi),%esi
+	movl	(%edi,%eax,4),%ecx
+	cmpl	24(%esp),%esi
+	movb	%dl,-1(%ebp,%esi,1)
+	jb	.L002loop1
+	jmp	.L004done
+.align	16
+.L001RC4_CHAR:
+	movzbl	(%edi,%eax,1),%ecx
+.L005cloop1:
+	addb	%cl,%bl
+	movzbl	(%edi,%ebx,1),%edx
+	movb	%cl,(%edi,%ebx,1)
+	movb	%dl,(%edi,%eax,1)
+	addb	%cl,%dl
+	movzbl	(%edi,%edx,1),%edx
+	addb	$1,%al
+	xorb	(%esi),%dl
+	leal	1(%esi),%esi
+	movzbl	(%edi,%eax,1),%ecx
+	cmpl	24(%esp),%esi
+	movb	%dl,-1(%ebp,%esi,1)
+	jb	.L005cloop1
+.L004done:
+	decb	%al
+	movb	%bl,-4(%edi)
+	movb	%al,-8(%edi)
+.L000abort:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	RC4,.-.L_RC4_begin
+.globl	RC4_set_key
+.type	RC4_set_key,@function
+.align	16
+RC4_set_key:
+.L_RC4_set_key_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	movl	24(%esp),%ebp
+	movl	28(%esp),%esi
+	leal	OPENSSL_ia32cap_P,%edx
+	leal	8(%edi),%edi
+	leal	(%esi,%ebp,1),%esi
+	negl	%ebp
+	xorl	%eax,%eax
+	movl	%ebp,-4(%edi)
+	btl	$20,(%edx)
+	jc	.L006c1stloop
+.align	16
+.L007w1stloop:
+	movl	%eax,(%edi,%eax,4)
+	addb	$1,%al
+	jnc	.L007w1stloop
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+.align	16
+.L008w2ndloop:
+	movl	(%edi,%ecx,4),%eax
+	addb	(%esi,%ebp,1),%dl
+	addb	%al,%dl
+	addl	$1,%ebp
+	movl	(%edi,%edx,4),%ebx
+	jnz	.L009wnowrap
+	movl	-4(%edi),%ebp
+.L009wnowrap:
+	movl	%eax,(%edi,%edx,4)
+	movl	%ebx,(%edi,%ecx,4)
+	addb	$1,%cl
+	jnc	.L008w2ndloop
+	jmp	.L010exit
+.align	16
+.L006c1stloop:
+	movb	%al,(%edi,%eax,1)
+	addb	$1,%al
+	jnc	.L006c1stloop
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	xorl	%ebx,%ebx
+.align	16
+.L011c2ndloop:
+	movb	(%edi,%ecx,1),%al
+	addb	(%esi,%ebp,1),%dl
+	addb	%al,%dl
+	addl	$1,%ebp
+	movb	(%edi,%edx,1),%bl
+	jnz	.L012cnowrap
+	movl	-4(%edi),%ebp
+.L012cnowrap:
+	movb	%al,(%edi,%edx,1)
+	movb	%bl,(%edi,%ecx,1)
+	addb	$1,%cl
+	jnc	.L011c2ndloop
+	movl	$-1,256(%edi)
+.L010exit:
+	xorl	%eax,%eax
+	movl	%eax,-8(%edi)
+	movl	%eax,-4(%edi)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	RC4_set_key,.-.L_RC4_set_key_begin
+.globl	RC4_options
+.type	RC4_options,@function
+.align	16
+RC4_options:
+.L_RC4_options_begin:
+	call	.L013pic_point
+.L013pic_point:
+	popl	%eax
+	leal	.L014opts-.L013pic_point(%eax),%eax
+	leal	OPENSSL_ia32cap_P,%edx
+	btl	$20,(%edx)
+	jnc	.L015skip
+	addl	$12,%eax
+.L015skip:
+	ret
+.align	64
+.L014opts:
+.byte	114,99,52,40,52,120,44,105,110,116,41,0
+.byte	114,99,52,40,49,120,44,99,104,97,114,41,0
+.byte	82,67,52,32,102,111,114,32,120,56,54,44,32,67,82,89
+.byte	80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114
+.byte	111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.align	64
+.size	RC4_options,.-.L_RC4_options_begin
+.comm	OPENSSL_ia32cap_P,4,4

deps/openssl/asm/x86-elf-gas/rc5/rc5-586.s

@@ -0,0 +1,564 @@
+.file	"rc5-586.s"
+.text
+.globl	RC5_32_encrypt
+.type	RC5_32_encrypt,@function
+.align	16
+RC5_32_encrypt:
+.L_RC5_32_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%esi
+	pushl	%edi
+	movl	16(%esp),%edx
+	movl	20(%esp),%ebp
+
+	movl	(%edx),%edi
+	movl	4(%edx),%esi
+	pushl	%ebx
+	movl	(%ebp),%ebx
+	addl	4(%ebp),%edi
+	addl	8(%ebp),%esi
+	xorl	%esi,%edi
+	movl	12(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	16(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	20(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	24(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	28(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	32(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	36(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	40(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	44(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	48(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	52(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	56(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	60(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	64(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	68(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	72(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	cmpl	$8,%ebx
+	je	.L000rc5_exit
+	xorl	%esi,%edi
+	movl	76(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	80(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	84(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	88(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	92(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	96(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	100(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	104(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	cmpl	$12,%ebx
+	je	.L000rc5_exit
+	xorl	%esi,%edi
+	movl	108(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	112(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	116(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	120(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	124(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	128(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	132(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	136(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+.L000rc5_exit:
+	movl	%edi,(%edx)
+	movl	%esi,4(%edx)
+	popl	%ebx
+	popl	%edi
+	popl	%esi
+	popl	%ebp
+	ret
+.size	RC5_32_encrypt,.-.L_RC5_32_encrypt_begin
+.globl	RC5_32_decrypt
+.type	RC5_32_decrypt,@function
+.align	16
+RC5_32_decrypt:
+.L_RC5_32_decrypt_begin:
+
+	pushl	%ebp
+	pushl	%esi
+	pushl	%edi
+	movl	16(%esp),%edx
+	movl	20(%esp),%ebp
+
+	movl	(%edx),%edi
+	movl	4(%edx),%esi
+	pushl	%ebx
+	movl	(%ebp),%ebx
+	cmpl	$12,%ebx
+	je	.L001rc5_dec_12
+	cmpl	$8,%ebx
+	je	.L002rc5_dec_8
+	movl	136(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	132(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	128(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	124(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	120(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	116(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	112(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	108(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+.L001rc5_dec_12:
+	movl	104(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	100(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	96(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	92(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	88(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	84(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	80(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	76(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+.L002rc5_dec_8:
+	movl	72(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	68(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	64(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	60(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	56(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	52(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	48(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	44(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	40(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	36(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	32(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	28(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	24(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	20(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	16(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	12(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	subl	8(%ebp),%esi
+	subl	4(%ebp),%edi
+.L003rc5_exit:
+	movl	%edi,(%edx)
+	movl	%esi,4(%edx)
+	popl	%ebx
+	popl	%edi
+	popl	%esi
+	popl	%ebp
+	ret
+.size	RC5_32_decrypt,.-.L_RC5_32_decrypt_begin
+.globl	RC5_32_cbc_encrypt
+.type	RC5_32_cbc_encrypt,@function
+.align	16
+RC5_32_cbc_encrypt:
+.L_RC5_32_cbc_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%ebp
+
+	movl	36(%esp),%ebx
+	movl	(%ebx),%esi
+	movl	4(%ebx),%edi
+	pushl	%edi
+	pushl	%esi
+	pushl	%edi
+	pushl	%esi
+	movl	%esp,%ebx
+	movl	36(%esp),%esi
+	movl	40(%esp),%edi
+
+	movl	56(%esp),%ecx
+
+	movl	48(%esp),%eax
+	pushl	%eax
+	pushl	%ebx
+	cmpl	$0,%ecx
+	jz	.L004decrypt
+	andl	$4294967288,%ebp
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	jz	.L005encrypt_finish
+.L006encrypt_loop:
+	movl	(%esi),%ecx
+	movl	4(%esi),%edx
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_RC5_32_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	.L006encrypt_loop
+.L005encrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	.L007finish
+	call	.L008PIC_point
+.L008PIC_point:
+	popl	%edx
+	leal	.L009cbc_enc_jmp_table-.L008PIC_point(%edx),%ecx
+	movl	(%ecx,%ebp,4),%ebp
+	addl	%edx,%ebp
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	jmp	*%ebp
+.L010ej7:
+	movb	6(%esi),%dh
+	shll	$8,%edx
+.L011ej6:
+	movb	5(%esi),%dh
+.L012ej5:
+	movb	4(%esi),%dl
+.L013ej4:
+	movl	(%esi),%ecx
+	jmp	.L014ejend
+.L015ej3:
+	movb	2(%esi),%ch
+	shll	$8,%ecx
+.L016ej2:
+	movb	1(%esi),%ch
+.L017ej1:
+	movb	(%esi),%cl
+.L014ejend:
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_RC5_32_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	jmp	.L007finish
+.L004decrypt:
+	andl	$4294967288,%ebp
+	movl	16(%esp),%eax
+	movl	20(%esp),%ebx
+	jz	.L018decrypt_finish
+.L019decrypt_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_RC5_32_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%ecx,(%edi)
+	movl	%edx,4(%edi)
+	movl	%eax,16(%esp)
+	movl	%ebx,20(%esp)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	.L019decrypt_loop
+.L018decrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	.L007finish
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	.L_RC5_32_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+.L020dj7:
+	rorl	$16,%edx
+	movb	%dl,6(%edi)
+	shrl	$16,%edx
+.L021dj6:
+	movb	%dh,5(%edi)
+.L022dj5:
+	movb	%dl,4(%edi)
+.L023dj4:
+	movl	%ecx,(%edi)
+	jmp	.L024djend
+.L025dj3:
+	rorl	$16,%ecx
+	movb	%cl,2(%edi)
+	shll	$16,%ecx
+.L026dj2:
+	movb	%ch,1(%esi)
+.L027dj1:
+	movb	%cl,(%esi)
+.L024djend:
+	jmp	.L007finish
+.L007finish:
+	movl	60(%esp),%ecx
+	addl	$24,%esp
+	movl	%eax,(%ecx)
+	movl	%ebx,4(%ecx)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	64
+.L009cbc_enc_jmp_table:
+.long	0
+.long	.L017ej1-.L008PIC_point
+.long	.L016ej2-.L008PIC_point
+.long	.L015ej3-.L008PIC_point
+.long	.L013ej4-.L008PIC_point
+.long	.L012ej5-.L008PIC_point
+.long	.L011ej6-.L008PIC_point
+.long	.L010ej7-.L008PIC_point
+.align	64
+.size	RC5_32_cbc_encrypt,.-.L_RC5_32_cbc_encrypt_begin

deps/openssl/asm/x86-elf-gas/sha/sha256-586.s

@@ -0,0 +1,261 @@
+.file	"sha512-586.s"
+.text
+.globl	sha256_block_data_order
+.type	sha256_block_data_order,@function
+.align	16
+sha256_block_data_order:
+.L_sha256_block_data_order_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	movl	24(%esp),%edi
+	movl	28(%esp),%eax
+	movl	%esp,%ebx
+	call	.L000pic_point
+.L000pic_point:
+	popl	%ebp
+	leal	.L001K256-.L000pic_point(%ebp),%ebp
+	subl	$16,%esp
+	andl	$-64,%esp
+	shll	$6,%eax
+	addl	%edi,%eax
+	movl	%esi,(%esp)
+	movl	%edi,4(%esp)
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+.align	16
+.L002loop:
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	16(%edi),%eax
+	movl	20(%edi),%ebx
+	movl	24(%edi),%ecx
+	movl	28(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	32(%edi),%eax
+	movl	36(%edi),%ebx
+	movl	40(%edi),%ecx
+	movl	44(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	48(%edi),%eax
+	movl	52(%edi),%ebx
+	movl	56(%edi),%ecx
+	movl	60(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	addl	$64,%edi
+	subl	$32,%esp
+	movl	%edi,100(%esp)
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edi
+	movl	%ebx,4(%esp)
+	movl	%ecx,8(%esp)
+	movl	%edi,12(%esp)
+	movl	16(%esi),%edx
+	movl	20(%esi),%ebx
+	movl	24(%esi),%ecx
+	movl	28(%esi),%edi
+	movl	%ebx,20(%esp)
+	movl	%ecx,24(%esp)
+	movl	%edi,28(%esp)
+.align	16
+.L00300_15:
+	movl	92(%esp),%ebx
+	movl	%edx,%ecx
+	rorl	$6,%ecx
+	movl	%edx,%edi
+	rorl	$11,%edi
+	movl	20(%esp),%esi
+	xorl	%edi,%ecx
+	rorl	$14,%edi
+	xorl	%edi,%ecx
+	movl	24(%esp),%edi
+	addl	%ecx,%ebx
+	movl	%edx,16(%esp)
+	xorl	%edi,%esi
+	movl	%eax,%ecx
+	andl	%edx,%esi
+	movl	12(%esp),%edx
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	addl	%esi,%ebx
+	rorl	$2,%ecx
+	addl	28(%esp),%ebx
+	rorl	$13,%edi
+	movl	4(%esp),%esi
+	xorl	%edi,%ecx
+	rorl	$9,%edi
+	addl	%ebx,%edx
+	xorl	%edi,%ecx
+	movl	8(%esp),%edi
+	addl	%ecx,%ebx
+	movl	%eax,(%esp)
+	movl	%eax,%ecx
+	subl	$4,%esp
+	orl	%esi,%eax
+	andl	%esi,%ecx
+	andl	%edi,%eax
+	movl	(%ebp),%esi
+	orl	%ecx,%eax
+	addl	$4,%ebp
+	addl	%ebx,%eax
+	addl	%esi,%edx
+	addl	%esi,%eax
+	cmpl	$3248222580,%esi
+	jne	.L00300_15
+	movl	152(%esp),%ebx
+.align	16
+.L00416_63:
+	movl	%ebx,%esi
+	movl	100(%esp),%ecx
+	shrl	$3,%ebx
+	rorl	$7,%esi
+	xorl	%esi,%ebx
+	rorl	$11,%esi
+	movl	%ecx,%edi
+	xorl	%esi,%ebx
+	shrl	$10,%ecx
+	movl	156(%esp),%esi
+	rorl	$17,%edi
+	xorl	%edi,%ecx
+	rorl	$2,%edi
+	addl	%esi,%ebx
+	xorl	%ecx,%edi
+	addl	%edi,%ebx
+	movl	%edx,%ecx
+	addl	120(%esp),%ebx
+	rorl	$6,%ecx
+	movl	%edx,%edi
+	rorl	$11,%edi
+	movl	20(%esp),%esi
+	xorl	%edi,%ecx
+	rorl	$14,%edi
+	movl	%ebx,92(%esp)
+	xorl	%edi,%ecx
+	movl	24(%esp),%edi
+	addl	%ecx,%ebx
+	movl	%edx,16(%esp)
+	xorl	%edi,%esi
+	movl	%eax,%ecx
+	andl	%edx,%esi
+	movl	12(%esp),%edx
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	addl	%esi,%ebx
+	rorl	$2,%ecx
+	addl	28(%esp),%ebx
+	rorl	$13,%edi
+	movl	4(%esp),%esi
+	xorl	%edi,%ecx
+	rorl	$9,%edi
+	addl	%ebx,%edx
+	xorl	%edi,%ecx
+	movl	8(%esp),%edi
+	addl	%ecx,%ebx
+	movl	%eax,(%esp)
+	movl	%eax,%ecx
+	subl	$4,%esp
+	orl	%esi,%eax
+	andl	%esi,%ecx
+	andl	%edi,%eax
+	movl	(%ebp),%esi
+	orl	%ecx,%eax
+	addl	$4,%ebp
+	addl	%ebx,%eax
+	movl	152(%esp),%ebx
+	addl	%esi,%edx
+	addl	%esi,%eax
+	cmpl	$3329325298,%esi
+	jne	.L00416_63
+	movl	352(%esp),%esi
+	movl	4(%esp),%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edi
+	addl	(%esi),%eax
+	addl	4(%esi),%ebx
+	addl	8(%esi),%ecx
+	addl	12(%esi),%edi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edi,12(%esi)
+	movl	20(%esp),%eax
+	movl	24(%esp),%ebx
+	movl	28(%esp),%ecx
+	movl	356(%esp),%edi
+	addl	16(%esi),%edx
+	addl	20(%esi),%eax
+	addl	24(%esi),%ebx
+	addl	28(%esi),%ecx
+	movl	%edx,16(%esi)
+	movl	%eax,20(%esi)
+	movl	%ebx,24(%esi)
+	movl	%ecx,28(%esi)
+	addl	$352,%esp
+	subl	$256,%ebp
+	cmpl	8(%esp),%edi
+	jb	.L002loop
+	movl	12(%esp),%esp
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	64
+.L001K256:
+.long	1116352408,1899447441,3049323471,3921009573
+.long	961987163,1508970993,2453635748,2870763221
+.long	3624381080,310598401,607225278,1426881987
+.long	1925078388,2162078206,2614888103,3248222580
+.long	3835390401,4022224774,264347078,604807628
+.long	770255983,1249150122,1555081692,1996064986
+.long	2554220882,2821834349,2952996808,3210313671
+.long	3336571891,3584528711,113926993,338241895
+.long	666307205,773529912,1294757372,1396182291
+.long	1695183700,1986661051,2177026350,2456956037
+.long	2730485921,2820302411,3259730800,3345764771
+.long	3516065817,3600352804,4094571909,275423344
+.long	430227734,506948616,659060556,883997877
+.long	958139571,1322822218,1537002063,1747873779
+.long	1955562222,2024104815,2227730452,2361852424
+.long	2428436474,2756734187,3204031479,3329325298
+.size	sha256_block_data_order,.-.L_sha256_block_data_order_begin
+.byte	83,72,65,50,53,54,32,98,108,111,99,107,32,116,114,97
+.byte	110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+.byte	67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+.byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+.byte	62,0

deps/openssl/asm/x86-elf-gas/sha/sha512-586.s

@@ -0,0 +1,563 @@
+.file	"sha512-586.s"
+.text
+.globl	sha512_block_data_order
+.type	sha512_block_data_order,@function
+.align	16
+sha512_block_data_order:
+.L_sha512_block_data_order_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	movl	24(%esp),%edi
+	movl	28(%esp),%eax
+	movl	%esp,%ebx
+	call	.L000pic_point
+.L000pic_point:
+	popl	%ebp
+	leal	.L001K512-.L000pic_point(%ebp),%ebp
+	subl	$16,%esp
+	andl	$-64,%esp
+	shll	$7,%eax
+	addl	%edi,%eax
+	movl	%esi,(%esp)
+	movl	%edi,4(%esp)
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+.align	16
+.L002loop_x86:
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	16(%edi),%eax
+	movl	20(%edi),%ebx
+	movl	24(%edi),%ecx
+	movl	28(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	32(%edi),%eax
+	movl	36(%edi),%ebx
+	movl	40(%edi),%ecx
+	movl	44(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	48(%edi),%eax
+	movl	52(%edi),%ebx
+	movl	56(%edi),%ecx
+	movl	60(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	64(%edi),%eax
+	movl	68(%edi),%ebx
+	movl	72(%edi),%ecx
+	movl	76(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	80(%edi),%eax
+	movl	84(%edi),%ebx
+	movl	88(%edi),%ecx
+	movl	92(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	96(%edi),%eax
+	movl	100(%edi),%ebx
+	movl	104(%edi),%ecx
+	movl	108(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	112(%edi),%eax
+	movl	116(%edi),%ebx
+	movl	120(%edi),%ecx
+	movl	124(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	addl	$128,%edi
+	subl	$72,%esp
+	movl	%edi,204(%esp)
+	leal	8(%esp),%edi
+	movl	$16,%ecx
+.long	2784229001
+.align	16
+.L00300_15_x86:
+	movl	40(%esp),%ecx
+	movl	44(%esp),%edx
+	movl	%ecx,%esi
+	shrl	$9,%ecx
+	movl	%edx,%edi
+	shrl	$9,%edx
+	movl	%ecx,%ebx
+	shll	$14,%esi
+	movl	%edx,%eax
+	shll	$14,%edi
+	xorl	%esi,%ebx
+	shrl	$5,%ecx
+	xorl	%edi,%eax
+	shrl	$5,%edx
+	xorl	%ecx,%eax
+	shll	$4,%esi
+	xorl	%edx,%ebx
+	shll	$4,%edi
+	xorl	%esi,%ebx
+	shrl	$4,%ecx
+	xorl	%edi,%eax
+	shrl	$4,%edx
+	xorl	%ecx,%eax
+	shll	$5,%esi
+	xorl	%edx,%ebx
+	shll	$5,%edi
+	xorl	%esi,%eax
+	xorl	%edi,%ebx
+	movl	48(%esp),%ecx
+	movl	52(%esp),%edx
+	movl	56(%esp),%esi
+	movl	60(%esp),%edi
+	addl	64(%esp),%eax
+	adcl	68(%esp),%ebx
+	xorl	%esi,%ecx
+	xorl	%edi,%edx
+	andl	40(%esp),%ecx
+	andl	44(%esp),%edx
+	addl	192(%esp),%eax
+	adcl	196(%esp),%ebx
+	xorl	%esi,%ecx
+	xorl	%edi,%edx
+	movl	(%ebp),%esi
+	movl	4(%ebp),%edi
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	32(%esp),%ecx
+	movl	36(%esp),%edx
+	addl	%esi,%eax
+	adcl	%edi,%ebx
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edx
+	movl	%eax,32(%esp)
+	movl	%ebx,36(%esp)
+	movl	%ecx,%esi
+	shrl	$2,%ecx
+	movl	%edx,%edi
+	shrl	$2,%edx
+	movl	%ecx,%ebx
+	shll	$4,%esi
+	movl	%edx,%eax
+	shll	$4,%edi
+	xorl	%esi,%ebx
+	shrl	$5,%ecx
+	xorl	%edi,%eax
+	shrl	$5,%edx
+	xorl	%ecx,%ebx
+	shll	$21,%esi
+	xorl	%edx,%eax
+	shll	$21,%edi
+	xorl	%esi,%eax
+	shrl	$21,%ecx
+	xorl	%edi,%ebx
+	shrl	$21,%edx
+	xorl	%ecx,%eax
+	shll	$5,%esi
+	xorl	%edx,%ebx
+	shll	$5,%edi
+	xorl	%esi,%eax
+	xorl	%edi,%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edx
+	movl	16(%esp),%esi
+	movl	20(%esp),%edi
+	addl	(%esp),%eax
+	adcl	4(%esp),%ebx
+	orl	%esi,%ecx
+	orl	%edi,%edx
+	andl	24(%esp),%ecx
+	andl	28(%esp),%edx
+	andl	8(%esp),%esi
+	andl	12(%esp),%edi
+	orl	%esi,%ecx
+	orl	%edi,%edx
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	movb	(%ebp),%dl
+	subl	$8,%esp
+	leal	8(%ebp),%ebp
+	cmpb	$148,%dl
+	jne	.L00300_15_x86
+.align	16
+.L00416_79_x86:
+	movl	312(%esp),%ecx
+	movl	316(%esp),%edx
+	movl	%ecx,%esi
+	shrl	$1,%ecx
+	movl	%edx,%edi
+	shrl	$1,%edx
+	movl	%ecx,%eax
+	shll	$24,%esi
+	movl	%edx,%ebx
+	shll	$24,%edi
+	xorl	%esi,%ebx
+	shrl	$6,%ecx
+	xorl	%edi,%eax
+	shrl	$6,%edx
+	xorl	%ecx,%eax
+	shll	$7,%esi
+	xorl	%edx,%ebx
+	shll	$1,%edi
+	xorl	%esi,%ebx
+	shrl	$1,%ecx
+	xorl	%edi,%eax
+	shrl	$1,%edx
+	xorl	%ecx,%eax
+	shll	$6,%edi
+	xorl	%edx,%ebx
+	xorl	%edi,%eax
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	movl	208(%esp),%ecx
+	movl	212(%esp),%edx
+	movl	%ecx,%esi
+	shrl	$6,%ecx
+	movl	%edx,%edi
+	shrl	$6,%edx
+	movl	%ecx,%eax
+	shll	$3,%esi
+	movl	%edx,%ebx
+	shll	$3,%edi
+	xorl	%esi,%eax
+	shrl	$13,%ecx
+	xorl	%edi,%ebx
+	shrl	$13,%edx
+	xorl	%ecx,%eax
+	shll	$10,%esi
+	xorl	%edx,%ebx
+	shll	$10,%edi
+	xorl	%esi,%ebx
+	shrl	$10,%ecx
+	xorl	%edi,%eax
+	shrl	$10,%edx
+	xorl	%ecx,%ebx
+	shll	$13,%edi
+	xorl	%edx,%eax
+	xorl	%edi,%eax
+	movl	320(%esp),%ecx
+	movl	324(%esp),%edx
+	addl	(%esp),%eax
+	adcl	4(%esp),%ebx
+	movl	248(%esp),%esi
+	movl	252(%esp),%edi
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	addl	%esi,%eax
+	adcl	%edi,%ebx
+	movl	%eax,192(%esp)
+	movl	%ebx,196(%esp)
+	movl	40(%esp),%ecx
+	movl	44(%esp),%edx
+	movl	%ecx,%esi
+	shrl	$9,%ecx
+	movl	%edx,%edi
+	shrl	$9,%edx
+	movl	%ecx,%ebx
+	shll	$14,%esi
+	movl	%edx,%eax
+	shll	$14,%edi
+	xorl	%esi,%ebx
+	shrl	$5,%ecx
+	xorl	%edi,%eax
+	shrl	$5,%edx
+	xorl	%ecx,%eax
+	shll	$4,%esi
+	xorl	%edx,%ebx
+	shll	$4,%edi
+	xorl	%esi,%ebx
+	shrl	$4,%ecx
+	xorl	%edi,%eax
+	shrl	$4,%edx
+	xorl	%ecx,%eax
+	shll	$5,%esi
+	xorl	%edx,%ebx
+	shll	$5,%edi
+	xorl	%esi,%eax
+	xorl	%edi,%ebx
+	movl	48(%esp),%ecx
+	movl	52(%esp),%edx
+	movl	56(%esp),%esi
+	movl	60(%esp),%edi
+	addl	64(%esp),%eax
+	adcl	68(%esp),%ebx
+	xorl	%esi,%ecx
+	xorl	%edi,%edx
+	andl	40(%esp),%ecx
+	andl	44(%esp),%edx
+	addl	192(%esp),%eax
+	adcl	196(%esp),%ebx
+	xorl	%esi,%ecx
+	xorl	%edi,%edx
+	movl	(%ebp),%esi
+	movl	4(%ebp),%edi
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	32(%esp),%ecx
+	movl	36(%esp),%edx
+	addl	%esi,%eax
+	adcl	%edi,%ebx
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edx
+	movl	%eax,32(%esp)
+	movl	%ebx,36(%esp)
+	movl	%ecx,%esi
+	shrl	$2,%ecx
+	movl	%edx,%edi
+	shrl	$2,%edx
+	movl	%ecx,%ebx
+	shll	$4,%esi
+	movl	%edx,%eax
+	shll	$4,%edi
+	xorl	%esi,%ebx
+	shrl	$5,%ecx
+	xorl	%edi,%eax
+	shrl	$5,%edx
+	xorl	%ecx,%ebx
+	shll	$21,%esi
+	xorl	%edx,%eax
+	shll	$21,%edi
+	xorl	%esi,%eax
+	shrl	$21,%ecx
+	xorl	%edi,%ebx
+	shrl	$21,%edx
+	xorl	%ecx,%eax
+	shll	$5,%esi
+	xorl	%edx,%ebx
+	shll	$5,%edi
+	xorl	%esi,%eax
+	xorl	%edi,%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edx
+	movl	16(%esp),%esi
+	movl	20(%esp),%edi
+	addl	(%esp),%eax
+	adcl	4(%esp),%ebx
+	orl	%esi,%ecx
+	orl	%edi,%edx
+	andl	24(%esp),%ecx
+	andl	28(%esp),%edx
+	andl	8(%esp),%esi
+	andl	12(%esp),%edi
+	orl	%esi,%ecx
+	orl	%edi,%edx
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	movb	(%ebp),%dl
+	subl	$8,%esp
+	leal	8(%ebp),%ebp
+	cmpb	$23,%dl
+	jne	.L00416_79_x86
+	movl	840(%esp),%esi
+	movl	844(%esp),%edi
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	addl	8(%esp),%eax
+	adcl	12(%esp),%ebx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	addl	16(%esp),%ecx
+	adcl	20(%esp),%edx
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	movl	16(%esi),%eax
+	movl	20(%esi),%ebx
+	movl	24(%esi),%ecx
+	movl	28(%esi),%edx
+	addl	24(%esp),%eax
+	adcl	28(%esp),%ebx
+	movl	%eax,16(%esi)
+	movl	%ebx,20(%esi)
+	addl	32(%esp),%ecx
+	adcl	36(%esp),%edx
+	movl	%ecx,24(%esi)
+	movl	%edx,28(%esi)
+	movl	32(%esi),%eax
+	movl	36(%esi),%ebx
+	movl	40(%esi),%ecx
+	movl	44(%esi),%edx
+	addl	40(%esp),%eax
+	adcl	44(%esp),%ebx
+	movl	%eax,32(%esi)
+	movl	%ebx,36(%esi)
+	addl	48(%esp),%ecx
+	adcl	52(%esp),%edx
+	movl	%ecx,40(%esi)
+	movl	%edx,44(%esi)
+	movl	48(%esi),%eax
+	movl	52(%esi),%ebx
+	movl	56(%esi),%ecx
+	movl	60(%esi),%edx
+	addl	56(%esp),%eax
+	adcl	60(%esp),%ebx
+	movl	%eax,48(%esi)
+	movl	%ebx,52(%esi)
+	addl	64(%esp),%ecx
+	adcl	68(%esp),%edx
+	movl	%ecx,56(%esi)
+	movl	%edx,60(%esi)
+	addl	$840,%esp
+	subl	$640,%ebp
+	cmpl	8(%esp),%edi
+	jb	.L002loop_x86
+	movl	12(%esp),%esp
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	64
+.L001K512:
+.long	3609767458,1116352408
+.long	602891725,1899447441
+.long	3964484399,3049323471
+.long	2173295548,3921009573
+.long	4081628472,961987163
+.long	3053834265,1508970993
+.long	2937671579,2453635748
+.long	3664609560,2870763221
+.long	2734883394,3624381080
+.long	1164996542,310598401
+.long	1323610764,607225278
+.long	3590304994,1426881987
+.long	4068182383,1925078388
+.long	991336113,2162078206
+.long	633803317,2614888103
+.long	3479774868,3248222580
+.long	2666613458,3835390401
+.long	944711139,4022224774
+.long	2341262773,264347078
+.long	2007800933,604807628
+.long	1495990901,770255983
+.long	1856431235,1249150122
+.long	3175218132,1555081692
+.long	2198950837,1996064986
+.long	3999719339,2554220882
+.long	766784016,2821834349
+.long	2566594879,2952996808
+.long	3203337956,3210313671
+.long	1034457026,3336571891
+.long	2466948901,3584528711
+.long	3758326383,113926993
+.long	168717936,338241895
+.long	1188179964,666307205
+.long	1546045734,773529912
+.long	1522805485,1294757372
+.long	2643833823,1396182291
+.long	2343527390,1695183700
+.long	1014477480,1986661051
+.long	1206759142,2177026350
+.long	344077627,2456956037
+.long	1290863460,2730485921
+.long	3158454273,2820302411
+.long	3505952657,3259730800
+.long	106217008,3345764771
+.long	3606008344,3516065817
+.long	1432725776,3600352804
+.long	1467031594,4094571909
+.long	851169720,275423344
+.long	3100823752,430227734
+.long	1363258195,506948616
+.long	3750685593,659060556
+.long	3785050280,883997877
+.long	3318307427,958139571
+.long	3812723403,1322822218
+.long	2003034995,1537002063
+.long	3602036899,1747873779
+.long	1575990012,1955562222
+.long	1125592928,2024104815
+.long	2716904306,2227730452
+.long	442776044,2361852424
+.long	593698344,2428436474
+.long	3733110249,2756734187
+.long	2999351573,3204031479
+.long	3815920427,3329325298
+.long	3928383900,3391569614
+.long	566280711,3515267271
+.long	3454069534,3940187606
+.long	4000239992,4118630271
+.long	1914138554,116418474
+.long	2731055270,174292421
+.long	3203993006,289380356
+.long	320620315,460393269
+.long	587496836,685471733
+.long	1086792851,852142971
+.long	365543100,1017036298
+.long	2618297676,1126000580
+.long	3409855158,1288033470
+.long	4234509866,1501505948
+.long	987167468,1607167915
+.long	1246189591,1816402316
+.size	sha512_block_data_order,.-.L_sha512_block_data_order_begin
+.byte	83,72,65,53,49,50,32,98,108,111,99,107,32,116,114,97
+.byte	110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+.byte	67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+.byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+.byte	62,0

deps/openssl/asm/x86-elf-gas/x86cpuid.s

@@ -0,0 +1,279 @@
+.file	"x86cpuid.s"
+.text
+.globl	OPENSSL_ia32_cpuid
+.type	OPENSSL_ia32_cpuid,@function
+.align	16
+OPENSSL_ia32_cpuid:
+.L_OPENSSL_ia32_cpuid_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	xorl	%edx,%edx
+	pushfl
+	popl	%eax
+	movl	%eax,%ecx
+	xorl	$2097152,%eax
+	pushl	%eax
+	popfl
+	pushfl
+	popl	%eax
+	xorl	%eax,%ecx
+	btl	$21,%ecx
+	jnc	.L000done
+	xorl	%eax,%eax
+	.byte	0x0f,0xa2
+	movl	%eax,%edi
+	xorl	%eax,%eax
+	cmpl	$1970169159,%ebx
+	setne	%al
+	movl	%eax,%ebp
+	cmpl	$1231384169,%edx
+	setne	%al
+	orl	%eax,%ebp
+	cmpl	$1818588270,%ecx
+	setne	%al
+	orl	%eax,%ebp
+	jz	.L001intel
+	cmpl	$1752462657,%ebx
+	setne	%al
+	movl	%eax,%esi
+	cmpl	$1769238117,%edx
+	setne	%al
+	orl	%eax,%esi
+	cmpl	$1145913699,%ecx
+	setne	%al
+	orl	%eax,%esi
+	jnz	.L001intel
+	movl	$2147483648,%eax
+	.byte	0x0f,0xa2
+	cmpl	$2147483656,%eax
+	jb	.L001intel
+	movl	$2147483656,%eax
+	.byte	0x0f,0xa2
+	movzbl	%cl,%esi
+	incl	%esi
+	movl	$1,%eax
+	.byte	0x0f,0xa2
+	btl	$28,%edx
+	jnc	.L000done
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	cmpl	%esi,%ebx
+	ja	.L000done
+	andl	$4026531839,%edx
+	jmp	.L000done
+.L001intel:
+	cmpl	$4,%edi
+	movl	$-1,%edi
+	jb	.L002nocacheinfo
+	movl	$4,%eax
+	movl	$0,%ecx
+	.byte	0x0f,0xa2
+	movl	%eax,%edi
+	shrl	$14,%edi
+	andl	$4095,%edi
+.L002nocacheinfo:
+	movl	$1,%eax
+	.byte	0x0f,0xa2
+	cmpl	$0,%ebp
+	jne	.L003notP4
+	andb	$15,%ah
+	cmpb	$15,%ah
+	jne	.L003notP4
+	orl	$1048576,%edx
+.L003notP4:
+	btl	$28,%edx
+	jnc	.L000done
+	andl	$4026531839,%edx
+	cmpl	$0,%edi
+	je	.L000done
+	orl	$268435456,%edx
+	shrl	$16,%ebx
+	cmpb	$1,%bl
+	ja	.L000done
+	andl	$4026531839,%edx
+.L000done:
+	movl	%edx,%eax
+	movl	%ecx,%edx
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	OPENSSL_ia32_cpuid,.-.L_OPENSSL_ia32_cpuid_begin
+.globl	OPENSSL_rdtsc
+.type	OPENSSL_rdtsc,@function
+.align	16
+OPENSSL_rdtsc:
+.L_OPENSSL_rdtsc_begin:
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	leal	OPENSSL_ia32cap_P,%ecx
+	btl	$4,(%ecx)
+	jnc	.L004notsc
+	.byte	0x0f,0x31
+.L004notsc:
+	ret
+.size	OPENSSL_rdtsc,.-.L_OPENSSL_rdtsc_begin
+.globl	OPENSSL_instrument_halt
+.type	OPENSSL_instrument_halt,@function
+.align	16
+OPENSSL_instrument_halt:
+.L_OPENSSL_instrument_halt_begin:
+	leal	OPENSSL_ia32cap_P,%ecx
+	btl	$4,(%ecx)
+	jnc	.L005nohalt
+.long	2421723150
+	andl	$3,%eax
+	jnz	.L005nohalt
+	pushfl
+	popl	%eax
+	btl	$9,%eax
+	jnc	.L005nohalt
+	.byte	0x0f,0x31
+	pushl	%edx
+	pushl	%eax
+	hlt
+	.byte	0x0f,0x31
+	subl	(%esp),%eax
+	sbbl	4(%esp),%edx
+	addl	$8,%esp
+	ret
+.L005nohalt:
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	ret
+.size	OPENSSL_instrument_halt,.-.L_OPENSSL_instrument_halt_begin
+.globl	OPENSSL_far_spin
+.type	OPENSSL_far_spin,@function
+.align	16
+OPENSSL_far_spin:
+.L_OPENSSL_far_spin_begin:
+	pushfl
+	popl	%eax
+	btl	$9,%eax
+	jnc	.L006nospin
+	movl	4(%esp),%eax
+	movl	8(%esp),%ecx
+.long	2430111262
+	xorl	%eax,%eax
+	movl	(%ecx),%edx
+	jmp	.L007spin
+.align	16
+.L007spin:
+	incl	%eax
+	cmpl	(%ecx),%edx
+	je	.L007spin
+.long	529567888
+	ret
+.L006nospin:
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	ret
+.size	OPENSSL_far_spin,.-.L_OPENSSL_far_spin_begin
+.globl	OPENSSL_wipe_cpu
+.type	OPENSSL_wipe_cpu,@function
+.align	16
+OPENSSL_wipe_cpu:
+.L_OPENSSL_wipe_cpu_begin:
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	leal	OPENSSL_ia32cap_P,%ecx
+	movl	(%ecx),%ecx
+	btl	$1,(%ecx)
+	jnc	.L008no_x87
+.long	4007259865,4007259865,4007259865,4007259865,2430851995
+.L008no_x87:
+	leal	4(%esp),%eax
+	ret
+.size	OPENSSL_wipe_cpu,.-.L_OPENSSL_wipe_cpu_begin
+.globl	OPENSSL_atomic_add
+.type	OPENSSL_atomic_add,@function
+.align	16
+OPENSSL_atomic_add:
+.L_OPENSSL_atomic_add_begin:
+	movl	4(%esp),%edx
+	movl	8(%esp),%ecx
+	pushl	%ebx
+	nop
+	movl	(%edx),%eax
+.L009spin:
+	leal	(%eax,%ecx,1),%ebx
+	nop
+.long	447811568
+	jne	.L009spin
+	movl	%ebx,%eax
+	popl	%ebx
+	ret
+.size	OPENSSL_atomic_add,.-.L_OPENSSL_atomic_add_begin
+.globl	OPENSSL_indirect_call
+.type	OPENSSL_indirect_call,@function
+.align	16
+OPENSSL_indirect_call:
+.L_OPENSSL_indirect_call_begin:
+	pushl	%ebp
+	movl	%esp,%ebp
+	subl	$28,%esp
+	movl	12(%ebp),%ecx
+	movl	%ecx,(%esp)
+	movl	16(%ebp),%edx
+	movl	%edx,4(%esp)
+	movl	20(%ebp),%eax
+	movl	%eax,8(%esp)
+	movl	24(%ebp),%eax
+	movl	%eax,12(%esp)
+	movl	28(%ebp),%eax
+	movl	%eax,16(%esp)
+	movl	32(%ebp),%eax
+	movl	%eax,20(%esp)
+	movl	36(%ebp),%eax
+	movl	%eax,24(%esp)
+	call	*8(%ebp)
+	movl	%ebp,%esp
+	popl	%ebp
+	ret
+.size	OPENSSL_indirect_call,.-.L_OPENSSL_indirect_call_begin
+.globl	OPENSSL_cleanse
+.type	OPENSSL_cleanse,@function
+.align	16
+OPENSSL_cleanse:
+.L_OPENSSL_cleanse_begin:
+	movl	4(%esp),%edx
+	movl	8(%esp),%ecx
+	xorl	%eax,%eax
+	cmpl	$7,%ecx
+	jae	.L010lot
+	cmpl	$0,%ecx
+	je	.L011ret
+.L012little:
+	movb	%al,(%edx)
+	subl	$1,%ecx
+	leal	1(%edx),%edx
+	jnz	.L012little
+.L011ret:
+	ret
+.align	16
+.L010lot:
+	testl	$3,%edx
+	jz	.L013aligned
+	movb	%al,(%edx)
+	leal	-1(%ecx),%ecx
+	leal	1(%edx),%edx
+	jmp	.L010lot
+.L013aligned:
+	movl	%eax,(%edx)
+	leal	-4(%ecx),%ecx
+	testl	$-4,%ecx
+	leal	4(%edx),%edx
+	jnz	.L013aligned
+	cmpl	$0,%ecx
+	jne	.L012little
+	ret
+.size	OPENSSL_cleanse,.-.L_OPENSSL_cleanse_begin
+.comm	OPENSSL_ia32cap_P,4,4
+.section	.init
+	call	OPENSSL_cpuid_setup
+	jmp	.Linitalign
+.align	16
+.Linitalign:

deps/openssl/asm/x86-macosx-gas/bf/bf-686.s

@@ -0,0 +1,897 @@
+.file	"bf-686.s"
+.text
+.globl	_BF_encrypt
+.align	4
+_BF_encrypt:
+L_BF_encrypt_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	# Load the 2 words
+
+	movl	20(%esp),%eax
+	movl	(%eax),%ecx
+	movl	4(%eax),%edx
+
+	# P pointer, s and enc flag
+
+	movl	24(%esp),%edi
+	xorl	%eax,%eax
+	xorl	%ebx,%ebx
+	xorl	(%edi),%ecx
+
+	# Round 0
+
+	rorl	$16,%ecx
+	movl	4(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 1
+
+	rorl	$16,%edx
+	movl	8(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 2
+
+	rorl	$16,%ecx
+	movl	12(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 3
+
+	rorl	$16,%edx
+	movl	16(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 4
+
+	rorl	$16,%ecx
+	movl	20(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 5
+
+	rorl	$16,%edx
+	movl	24(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 6
+
+	rorl	$16,%ecx
+	movl	28(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 7
+
+	rorl	$16,%edx
+	movl	32(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 8
+
+	rorl	$16,%ecx
+	movl	36(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 9
+
+	rorl	$16,%edx
+	movl	40(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 10
+
+	rorl	$16,%ecx
+	movl	44(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 11
+
+	rorl	$16,%edx
+	movl	48(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 12
+
+	rorl	$16,%ecx
+	movl	52(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 13
+
+	rorl	$16,%edx
+	movl	56(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 14
+
+	rorl	$16,%ecx
+	movl	60(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 15
+
+	rorl	$16,%edx
+	movl	64(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+	xorl	68(%edi),%edx
+	movl	20(%esp),%eax
+	movl	%edx,(%eax)
+	movl	%ecx,4(%eax)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_BF_decrypt
+.align	4
+_BF_decrypt:
+L_BF_decrypt_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	# Load the 2 words
+
+	movl	20(%esp),%eax
+	movl	(%eax),%ecx
+	movl	4(%eax),%edx
+
+	# P pointer, s and enc flag
+
+	movl	24(%esp),%edi
+	xorl	%eax,%eax
+	xorl	%ebx,%ebx
+	xorl	68(%edi),%ecx
+
+	# Round 16
+
+	rorl	$16,%ecx
+	movl	64(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 15
+
+	rorl	$16,%edx
+	movl	60(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 14
+
+	rorl	$16,%ecx
+	movl	56(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 13
+
+	rorl	$16,%edx
+	movl	52(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 12
+
+	rorl	$16,%ecx
+	movl	48(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 11
+
+	rorl	$16,%edx
+	movl	44(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 10
+
+	rorl	$16,%ecx
+	movl	40(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 9
+
+	rorl	$16,%edx
+	movl	36(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 8
+
+	rorl	$16,%ecx
+	movl	32(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 7
+
+	rorl	$16,%edx
+	movl	28(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 6
+
+	rorl	$16,%ecx
+	movl	24(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 5
+
+	rorl	$16,%edx
+	movl	20(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 4
+
+	rorl	$16,%ecx
+	movl	16(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 3
+
+	rorl	$16,%edx
+	movl	12(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+
+	# Round 2
+
+	rorl	$16,%ecx
+	movl	8(%edi),%esi
+	movb	%ch,%al
+	movb	%cl,%bl
+	rorl	$16,%ecx
+	xorl	%esi,%edx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%ch,%al
+	movb	%cl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%edx
+
+	# Round 1
+
+	rorl	$16,%edx
+	movl	4(%edi),%esi
+	movb	%dh,%al
+	movb	%dl,%bl
+	rorl	$16,%edx
+	xorl	%esi,%ecx
+	movl	72(%edi,%eax,4),%esi
+	movl	1096(%edi,%ebx,4),%ebp
+	movb	%dh,%al
+	movb	%dl,%bl
+	addl	%ebp,%esi
+	movl	2120(%edi,%eax,4),%eax
+	xorl	%eax,%esi
+	movl	3144(%edi,%ebx,4),%ebp
+	addl	%ebp,%esi
+	xorl	%eax,%eax
+	xorl	%esi,%ecx
+	xorl	(%edi),%edx
+	movl	20(%esp),%eax
+	movl	%edx,(%eax)
+	movl	%ecx,4(%eax)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_BF_cbc_encrypt
+.align	4
+_BF_cbc_encrypt:
+L_BF_cbc_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%ebp
+	# getting iv ptr from parameter 4
+
+	movl	36(%esp),%ebx
+	movl	(%ebx),%esi
+	movl	4(%ebx),%edi
+	pushl	%edi
+	pushl	%esi
+	pushl	%edi
+	pushl	%esi
+	movl	%esp,%ebx
+	movl	36(%esp),%esi
+	movl	40(%esp),%edi
+	# getting encrypt flag from parameter 5
+
+	movl	56(%esp),%ecx
+	# get and push parameter 3
+
+	movl	48(%esp),%eax
+	pushl	%eax
+	pushl	%ebx
+	cmpl	$0,%ecx
+	jz	L000decrypt
+	andl	$4294967288,%ebp
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	jz	L001encrypt_finish
+L002encrypt_loop:
+	movl	(%esi),%ecx
+	movl	4(%esi),%edx
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_BF_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	L002encrypt_loop
+L001encrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	L003finish
+	call	L004PIC_point
+L004PIC_point:
+	popl	%edx
+	leal	L005cbc_enc_jmp_table-L004PIC_point(%edx),%ecx
+	movl	(%ecx,%ebp,4),%ebp
+	addl	%edx,%ebp
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	jmp	*%ebp
+L006ej7:
+	movb	6(%esi),%dh
+	shll	$8,%edx
+L007ej6:
+	movb	5(%esi),%dh
+L008ej5:
+	movb	4(%esi),%dl
+L009ej4:
+	movl	(%esi),%ecx
+	jmp	L010ejend
+L011ej3:
+	movb	2(%esi),%ch
+	shll	$8,%ecx
+L012ej2:
+	movb	1(%esi),%ch
+L013ej1:
+	movb	(%esi),%cl
+L010ejend:
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_BF_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	jmp	L003finish
+L000decrypt:
+	andl	$4294967288,%ebp
+	movl	16(%esp),%eax
+	movl	20(%esp),%ebx
+	jz	L014decrypt_finish
+L015decrypt_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_BF_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%ecx,(%edi)
+	movl	%edx,4(%edi)
+	movl	%eax,16(%esp)
+	movl	%ebx,20(%esp)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	L015decrypt_loop
+L014decrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	L003finish
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_BF_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+L016dj7:
+	rorl	$16,%edx
+	movb	%dl,6(%edi)
+	shrl	$16,%edx
+L017dj6:
+	movb	%dh,5(%edi)
+L018dj5:
+	movb	%dl,4(%edi)
+L019dj4:
+	movl	%ecx,(%edi)
+	jmp	L020djend
+L021dj3:
+	rorl	$16,%ecx
+	movb	%cl,2(%edi)
+	shll	$16,%ecx
+L022dj2:
+	movb	%ch,1(%esi)
+L023dj1:
+	movb	%cl,(%esi)
+L020djend:
+	jmp	L003finish
+L003finish:
+	movl	60(%esp),%ecx
+	addl	$24,%esp
+	movl	%eax,(%ecx)
+	movl	%ebx,4(%ecx)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	6,0x90
+L005cbc_enc_jmp_table:
+.long	0
+.long	L013ej1-L004PIC_point
+.long	L012ej2-L004PIC_point
+.long	L011ej3-L004PIC_point
+.long	L009ej4-L004PIC_point
+.long	L008ej5-L004PIC_point
+.long	L007ej6-L004PIC_point
+.long	L006ej7-L004PIC_point
+.align	6,0x90

deps/openssl/asm/x86-macosx-gas/bn/x86-mont.s

@@ -0,0 +1,336 @@
+.file	"../openssl/crypto/bn/asm/x86-mont.s"
+.text
+.globl	_bn_mul_mont
+.align	4
+_bn_mul_mont:
+L_bn_mul_mont_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	xorl	%eax,%eax
+	movl	40(%esp),%edi
+	cmpl	$4,%edi
+	jl	L000just_leave
+	leal	20(%esp),%esi
+	leal	24(%esp),%edx
+	movl	%esp,%ebp
+	addl	$2,%edi
+	negl	%edi
+	leal	-32(%esp,%edi,4),%esp
+	negl	%edi
+	movl	%esp,%eax
+	subl	%edx,%eax
+	andl	$2047,%eax
+	subl	%eax,%esp
+	xorl	%esp,%edx
+	andl	$2048,%edx
+	xorl	$2048,%edx
+	subl	%edx,%esp
+	andl	$-64,%esp
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	16(%esi),%esi
+	movl	(%esi),%esi
+	movl	%eax,4(%esp)
+	movl	%ebx,8(%esp)
+	movl	%ecx,12(%esp)
+	movl	%edx,16(%esp)
+	movl	%esi,20(%esp)
+	leal	-3(%edi),%ebx
+	movl	%ebp,24(%esp)
+	movl	8(%esp),%esi
+	leal	1(%ebx),%ebp
+	movl	12(%esp),%edi
+	xorl	%ecx,%ecx
+	movl	%esi,%edx
+	andl	$1,%ebp
+	subl	%edi,%edx
+	leal	4(%edi,%ebx,4),%eax
+	orl	%edx,%ebp
+	movl	(%edi),%edi
+	jz	L001bn_sqr_mont
+	movl	%eax,28(%esp)
+	movl	(%esi),%eax
+	xorl	%edx,%edx
+.align	4,0x90
+L002mull:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%eax,%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	movl	(%esi,%ecx,4),%eax
+	cmpl	%ebx,%ecx
+	movl	%ebp,28(%esp,%ecx,4)
+	jl	L002mull
+	movl	%edx,%ebp
+	mull	%edi
+	movl	20(%esp),%edi
+	addl	%ebp,%eax
+	movl	16(%esp),%esi
+	adcl	$0,%edx
+	imull	32(%esp),%edi
+	movl	%eax,32(%esp,%ebx,4)
+	xorl	%ecx,%ecx
+	movl	%edx,36(%esp,%ebx,4)
+	movl	%ecx,40(%esp,%ebx,4)
+	movl	(%esi),%eax
+	mull	%edi
+	addl	32(%esp),%eax
+	movl	4(%esi),%eax
+	adcl	$0,%edx
+	incl	%ecx
+	jmp	L0032ndmadd
+.align	4,0x90
+L0041stmadd:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ecx,4),%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	cmpl	%ebx,%ecx
+	movl	%ebp,28(%esp,%ecx,4)
+	jl	L0041stmadd
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ebx,4),%eax
+	movl	20(%esp),%edi
+	adcl	$0,%edx
+	movl	16(%esp),%esi
+	addl	%eax,%ebp
+	adcl	$0,%edx
+	imull	32(%esp),%edi
+	xorl	%ecx,%ecx
+	addl	36(%esp,%ebx,4),%edx
+	movl	%ebp,32(%esp,%ebx,4)
+	adcl	$0,%ecx
+	movl	(%esi),%eax
+	movl	%edx,36(%esp,%ebx,4)
+	movl	%ecx,40(%esp,%ebx,4)
+	mull	%edi
+	addl	32(%esp),%eax
+	movl	4(%esi),%eax
+	adcl	$0,%edx
+	movl	$1,%ecx
+.align	4,0x90
+L0032ndmadd:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ecx,4),%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	cmpl	%ebx,%ecx
+	movl	%ebp,24(%esp,%ecx,4)
+	jl	L0032ndmadd
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ebx,4),%ebp
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	adcl	$0,%edx
+	movl	%ebp,28(%esp,%ebx,4)
+	xorl	%eax,%eax
+	movl	12(%esp),%ecx
+	addl	36(%esp,%ebx,4),%edx
+	adcl	40(%esp,%ebx,4),%eax
+	leal	4(%ecx),%ecx
+	movl	%edx,32(%esp,%ebx,4)
+	cmpl	28(%esp),%ecx
+	movl	%eax,36(%esp,%ebx,4)
+	je	L005common_tail
+	movl	(%ecx),%edi
+	movl	8(%esp),%esi
+	movl	%ecx,12(%esp)
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	movl	(%esi),%eax
+	jmp	L0041stmadd
+.align	4,0x90
+L001bn_sqr_mont:
+	movl	%ebx,(%esp)
+	movl	%ecx,12(%esp)
+	movl	%edi,%eax
+	mull	%edi
+	movl	%eax,32(%esp)
+	movl	%edx,%ebx
+	shrl	$1,%edx
+	andl	$1,%ebx
+	incl	%ecx
+.align	4,0x90
+L006sqr:
+	movl	(%esi,%ecx,4),%eax
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%ebp,%eax
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	leal	(%ebx,%eax,2),%ebp
+	shrl	$31,%eax
+	cmpl	(%esp),%ecx
+	movl	%eax,%ebx
+	movl	%ebp,28(%esp,%ecx,4)
+	jl	L006sqr
+	movl	(%esi,%ecx,4),%eax
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%ebp,%eax
+	movl	20(%esp),%edi
+	adcl	$0,%edx
+	movl	16(%esp),%esi
+	leal	(%ebx,%eax,2),%ebp
+	imull	32(%esp),%edi
+	shrl	$31,%eax
+	movl	%ebp,32(%esp,%ecx,4)
+	leal	(%eax,%edx,2),%ebp
+	movl	(%esi),%eax
+	shrl	$31,%edx
+	movl	%ebp,36(%esp,%ecx,4)
+	movl	%edx,40(%esp,%ecx,4)
+	mull	%edi
+	addl	32(%esp),%eax
+	movl	%ecx,%ebx
+	adcl	$0,%edx
+	movl	4(%esi),%eax
+	movl	$1,%ecx
+.align	4,0x90
+L0073rdmadd:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ecx,4),%ebp
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	4(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	movl	%ebp,28(%esp,%ecx,4)
+	movl	%edx,%ebp
+	mull	%edi
+	addl	36(%esp,%ecx,4),%ebp
+	leal	2(%ecx),%ecx
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	cmpl	%ebx,%ecx
+	movl	%ebp,24(%esp,%ecx,4)
+	jl	L0073rdmadd
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ebx,4),%ebp
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	adcl	$0,%edx
+	movl	%ebp,28(%esp,%ebx,4)
+	movl	12(%esp),%ecx
+	xorl	%eax,%eax
+	movl	8(%esp),%esi
+	addl	36(%esp,%ebx,4),%edx
+	adcl	40(%esp,%ebx,4),%eax
+	movl	%edx,32(%esp,%ebx,4)
+	cmpl	%ebx,%ecx
+	movl	%eax,36(%esp,%ebx,4)
+	je	L005common_tail
+	movl	4(%esi,%ecx,4),%edi
+	leal	1(%ecx),%ecx
+	movl	%edi,%eax
+	movl	%ecx,12(%esp)
+	mull	%edi
+	addl	32(%esp,%ecx,4),%eax
+	adcl	$0,%edx
+	movl	%eax,32(%esp,%ecx,4)
+	xorl	%ebp,%ebp
+	cmpl	%ebx,%ecx
+	leal	1(%ecx),%ecx
+	je	L008sqrlast
+	movl	%edx,%ebx
+	shrl	$1,%edx
+	andl	$1,%ebx
+.align	4,0x90
+L009sqradd:
+	movl	(%esi,%ecx,4),%eax
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%ebp,%eax
+	leal	(%eax,%eax,1),%ebp
+	adcl	$0,%edx
+	shrl	$31,%eax
+	addl	32(%esp,%ecx,4),%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%eax
+	addl	%ebx,%ebp
+	adcl	$0,%eax
+	cmpl	(%esp),%ecx
+	movl	%ebp,28(%esp,%ecx,4)
+	movl	%eax,%ebx
+	jle	L009sqradd
+	movl	%edx,%ebp
+	addl	%edx,%edx
+	shrl	$31,%ebp
+	addl	%ebx,%edx
+	adcl	$0,%ebp
+L008sqrlast:
+	movl	20(%esp),%edi
+	movl	16(%esp),%esi
+	imull	32(%esp),%edi
+	addl	32(%esp,%ecx,4),%edx
+	movl	(%esi),%eax
+	adcl	$0,%ebp
+	movl	%edx,32(%esp,%ecx,4)
+	movl	%ebp,36(%esp,%ecx,4)
+	mull	%edi
+	addl	32(%esp),%eax
+	leal	-1(%ecx),%ebx
+	adcl	$0,%edx
+	movl	$1,%ecx
+	movl	4(%esi),%eax
+	jmp	L0073rdmadd
+.align	4,0x90
+L005common_tail:
+	movl	16(%esp),%ebp
+	movl	4(%esp),%edi
+	leal	32(%esp),%esi
+	movl	(%esi),%eax
+	movl	%ebx,%ecx
+	xorl	%edx,%edx
+.align	4,0x90
+L010sub:
+	sbbl	(%ebp,%edx,4),%eax
+	movl	%eax,(%edi,%edx,4)
+	decl	%ecx
+	movl	4(%esi,%edx,4),%eax
+	leal	1(%edx),%edx
+	jge	L010sub
+	sbbl	$0,%eax
+	andl	%eax,%esi
+	notl	%eax
+	movl	%edi,%ebp
+	andl	%eax,%ebp
+	orl	%ebp,%esi
+.align	4,0x90
+L011copy:
+	movl	(%esi,%ebx,4),%eax
+	movl	%eax,(%edi,%ebx,4)
+	movl	%ecx,32(%esp,%ebx,4)
+	decl	%ebx
+	jge	L011copy
+	movl	24(%esp),%esp
+	movl	$1,%eax
+L000just_leave:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.byte	77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
+.byte	112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
+.byte	54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
+.byte	32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
+.byte	111,114,103,62,0

deps/openssl/asm/x86-macosx-gas/cast/cast-586.s

@@ -0,0 +1,967 @@
+.file	"cast-586.s"
+.text
+.globl	_CAST_encrypt
+.align	4
+_CAST_encrypt:
+L_CAST_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	movl	12(%esp),%ebx
+	movl	16(%esp),%ebp
+	pushl	%esi
+	pushl	%edi
+	# Load the 2 words
+
+	movl	(%ebx),%edi
+	movl	4(%ebx),%esi
+	# Get short key flag
+
+	movl	128(%ebp),%eax
+	pushl	%eax
+	xorl	%eax,%eax
+	# round 0
+
+	movl	(%ebp),%edx
+	movl	4(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 1
+
+	movl	8(%ebp),%edx
+	movl	12(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 2
+
+	movl	16(%ebp),%edx
+	movl	20(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 3
+
+	movl	24(%ebp),%edx
+	movl	28(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 4
+
+	movl	32(%ebp),%edx
+	movl	36(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 5
+
+	movl	40(%ebp),%edx
+	movl	44(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 6
+
+	movl	48(%ebp),%edx
+	movl	52(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 7
+
+	movl	56(%ebp),%edx
+	movl	60(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 8
+
+	movl	64(%ebp),%edx
+	movl	68(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 9
+
+	movl	72(%ebp),%edx
+	movl	76(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 10
+
+	movl	80(%ebp),%edx
+	movl	84(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 11
+
+	movl	88(%ebp),%edx
+	movl	92(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# test short key flag
+
+	popl	%edx
+	orl	%edx,%edx
+	jnz	L000cast_enc_done
+	# round 12
+
+	movl	96(%ebp),%edx
+	movl	100(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 13
+
+	movl	104(%ebp),%edx
+	movl	108(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 14
+
+	movl	112(%ebp),%edx
+	movl	116(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 15
+
+	movl	120(%ebp),%edx
+	movl	124(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+L000cast_enc_done:
+	nop
+	movl	20(%esp),%eax
+	movl	%edi,4(%eax)
+	movl	%esi,(%eax)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_CAST_decrypt
+.align	4
+_CAST_decrypt:
+L_CAST_decrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	movl	12(%esp),%ebx
+	movl	16(%esp),%ebp
+	pushl	%esi
+	pushl	%edi
+	# Load the 2 words
+
+	movl	(%ebx),%edi
+	movl	4(%ebx),%esi
+	# Get short key flag
+
+	movl	128(%ebp),%eax
+	orl	%eax,%eax
+	jnz	L001cast_dec_skip
+	xorl	%eax,%eax
+	# round 15
+
+	movl	120(%ebp),%edx
+	movl	124(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 14
+
+	movl	112(%ebp),%edx
+	movl	116(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 13
+
+	movl	104(%ebp),%edx
+	movl	108(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 12
+
+	movl	96(%ebp),%edx
+	movl	100(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+L001cast_dec_skip:
+	# round 11
+
+	movl	88(%ebp),%edx
+	movl	92(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 10
+
+	movl	80(%ebp),%edx
+	movl	84(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 9
+
+	movl	72(%ebp),%edx
+	movl	76(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 8
+
+	movl	64(%ebp),%edx
+	movl	68(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 7
+
+	movl	56(%ebp),%edx
+	movl	60(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 6
+
+	movl	48(%ebp),%edx
+	movl	52(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 5
+
+	movl	40(%ebp),%edx
+	movl	44(%ebp),%ecx
+	subl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 4
+
+	movl	32(%ebp),%edx
+	movl	36(%ebp),%ecx
+	xorl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 3
+
+	movl	24(%ebp),%edx
+	movl	28(%ebp),%ecx
+	addl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 2
+
+	movl	16(%ebp),%edx
+	movl	20(%ebp),%ecx
+	subl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	subl	%ebx,%ecx
+	xorl	%ecx,%esi
+	# round 1
+
+	movl	8(%ebp),%edx
+	movl	12(%ebp),%ecx
+	xorl	%esi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	addl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	xorl	%ebx,%ecx
+	xorl	%ecx,%edi
+	# round 0
+
+	movl	(%ebp),%edx
+	movl	4(%ebp),%ecx
+	addl	%edi,%edx
+	roll	%cl,%edx
+	movl	%edx,%ebx
+	xorl	%ecx,%ecx
+	movb	%dh,%cl
+	andl	$255,%ebx
+	shrl	$16,%edx
+	xorl	%eax,%eax
+	movb	%dh,%al
+	andl	$255,%edx
+	movl	_CAST_S_table0(,%ecx,4),%ecx
+	movl	_CAST_S_table1(,%ebx,4),%ebx
+	xorl	%ebx,%ecx
+	movl	_CAST_S_table2(,%eax,4),%ebx
+	subl	%ebx,%ecx
+	movl	_CAST_S_table3(,%edx,4),%ebx
+	addl	%ebx,%ecx
+	xorl	%ecx,%esi
+	nop
+	movl	20(%esp),%eax
+	movl	%edi,4(%eax)
+	movl	%esi,(%eax)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_CAST_cbc_encrypt
+.align	4
+_CAST_cbc_encrypt:
+L_CAST_cbc_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%ebp
+	# getting iv ptr from parameter 4
+
+	movl	36(%esp),%ebx
+	movl	(%ebx),%esi
+	movl	4(%ebx),%edi
+	pushl	%edi
+	pushl	%esi
+	pushl	%edi
+	pushl	%esi
+	movl	%esp,%ebx
+	movl	36(%esp),%esi
+	movl	40(%esp),%edi
+	# getting encrypt flag from parameter 5
+
+	movl	56(%esp),%ecx
+	# get and push parameter 3
+
+	movl	48(%esp),%eax
+	pushl	%eax
+	pushl	%ebx
+	cmpl	$0,%ecx
+	jz	L002decrypt
+	andl	$4294967288,%ebp
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	jz	L003encrypt_finish
+L004encrypt_loop:
+	movl	(%esi),%ecx
+	movl	4(%esi),%edx
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_CAST_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	L004encrypt_loop
+L003encrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	L005finish
+	call	L006PIC_point
+L006PIC_point:
+	popl	%edx
+	leal	L007cbc_enc_jmp_table-L006PIC_point(%edx),%ecx
+	movl	(%ecx,%ebp,4),%ebp
+	addl	%edx,%ebp
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	jmp	*%ebp
+L008ej7:
+	movb	6(%esi),%dh
+	shll	$8,%edx
+L009ej6:
+	movb	5(%esi),%dh
+L010ej5:
+	movb	4(%esi),%dl
+L011ej4:
+	movl	(%esi),%ecx
+	jmp	L012ejend
+L013ej3:
+	movb	2(%esi),%ch
+	shll	$8,%ecx
+L014ej2:
+	movb	1(%esi),%ch
+L015ej1:
+	movb	(%esi),%cl
+L012ejend:
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_CAST_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	jmp	L005finish
+L002decrypt:
+	andl	$4294967288,%ebp
+	movl	16(%esp),%eax
+	movl	20(%esp),%ebx
+	jz	L016decrypt_finish
+L017decrypt_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_CAST_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%ecx,(%edi)
+	movl	%edx,4(%edi)
+	movl	%eax,16(%esp)
+	movl	%ebx,20(%esp)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	L017decrypt_loop
+L016decrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	L005finish
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_CAST_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	bswap	%eax
+	bswap	%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+L018dj7:
+	rorl	$16,%edx
+	movb	%dl,6(%edi)
+	shrl	$16,%edx
+L019dj6:
+	movb	%dh,5(%edi)
+L020dj5:
+	movb	%dl,4(%edi)
+L021dj4:
+	movl	%ecx,(%edi)
+	jmp	L022djend
+L023dj3:
+	rorl	$16,%ecx
+	movb	%cl,2(%edi)
+	shll	$16,%ecx
+L024dj2:
+	movb	%ch,1(%esi)
+L025dj1:
+	movb	%cl,(%esi)
+L022djend:
+	jmp	L005finish
+L005finish:
+	movl	60(%esp),%ecx
+	addl	$24,%esp
+	movl	%eax,(%ecx)
+	movl	%ebx,4(%ecx)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	6,0x90
+L007cbc_enc_jmp_table:
+.long	0
+.long	L015ej1-L006PIC_point
+.long	L014ej2-L006PIC_point
+.long	L013ej3-L006PIC_point
+.long	L011ej4-L006PIC_point
+.long	L010ej5-L006PIC_point
+.long	L009ej6-L006PIC_point
+.long	L008ej7-L006PIC_point
+.align	6,0x90

deps/openssl/asm/x86-macosx-gas/des/crypt586.s

@@ -0,0 +1,891 @@
+.file	"crypt586.s"
+.text
+.globl	_fcrypt_body
+.align	4
+_fcrypt_body:
+L_fcrypt_body_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	# Load the 2 words
+
+	xorl	%edi,%edi
+	xorl	%esi,%esi
+	leal	_DES_SPtrans,%edx
+	pushl	%edx
+	movl	28(%esp),%ebp
+	pushl	$25
+L000start:
+
+	# Round 0
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	4(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+	# Round 1
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	8(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	12(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+	# Round 2
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	16(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	20(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+	# Round 3
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	24(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	28(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+	# Round 4
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	32(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	36(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+	# Round 5
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	40(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	44(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+	# Round 6
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	48(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	52(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+	# Round 7
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	56(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	60(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+	# Round 8
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	64(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	68(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+	# Round 9
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	72(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	76(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+	# Round 10
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	80(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	84(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+	# Round 11
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	88(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	92(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+	# Round 12
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	96(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	100(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+	# Round 13
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	104(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	108(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+
+	# Round 14
+
+	movl	36(%esp),%eax
+	movl	%esi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%esi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	112(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	116(%ebp),%ecx
+	xorl	%esi,%eax
+	xorl	%esi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%edi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%edi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%edi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%edi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%edi
+	movl	32(%esp),%ebp
+
+	# Round 15
+
+	movl	36(%esp),%eax
+	movl	%edi,%edx
+	shrl	$16,%edx
+	movl	40(%esp),%ecx
+	xorl	%edi,%edx
+	andl	%edx,%eax
+	andl	%ecx,%edx
+	movl	%eax,%ebx
+	shll	$16,%ebx
+	movl	%edx,%ecx
+	shll	$16,%ecx
+	xorl	%ebx,%eax
+	xorl	%ecx,%edx
+	movl	120(%ebp),%ebx
+	xorl	%ebx,%eax
+	movl	124(%ebp),%ecx
+	xorl	%edi,%eax
+	xorl	%edi,%edx
+	xorl	%ecx,%edx
+	andl	$0xfcfcfcfc,%eax
+	xorl	%ebx,%ebx
+	andl	$0xcfcfcfcf,%edx
+	xorl	%ecx,%ecx
+	movb	%al,%bl
+	movb	%ah,%cl
+	rorl	$4,%edx
+	movl	4(%esp),%ebp
+	xorl	(%ebp,%ebx,1),%esi
+	movb	%dl,%bl
+	xorl	0x200(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	shrl	$16,%eax
+	xorl	0x100(%ebp,%ebx,1),%esi
+	movb	%ah,%bl
+	shrl	$16,%edx
+	xorl	0x300(%ebp,%ecx,1),%esi
+	movb	%dh,%cl
+	andl	$0xff,%eax
+	andl	$0xff,%edx
+	movl	0x600(%ebp,%ebx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x700(%ebp,%ecx,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x400(%ebp,%eax,1),%ebx
+	xorl	%ebx,%esi
+	movl	0x500(%ebp,%edx,1),%ebx
+	xorl	%ebx,%esi
+	movl	32(%esp),%ebp
+	movl	(%esp),%ebx
+	movl	%edi,%eax
+	decl	%ebx
+	movl	%esi,%edi
+	movl	%eax,%esi
+	movl	%ebx,(%esp)
+	jnz	L000start
+
+	# FP
+
+	movl	28(%esp),%edx
+	rorl	$1,%edi
+	movl	%esi,%eax
+	xorl	%edi,%esi
+	andl	$0xaaaaaaaa,%esi
+	xorl	%esi,%eax
+	xorl	%esi,%edi
+
+	roll	$23,%eax
+	movl	%eax,%esi
+	xorl	%edi,%eax
+	andl	$0x03fc03fc,%eax
+	xorl	%eax,%esi
+	xorl	%eax,%edi
+
+	roll	$10,%esi
+	movl	%esi,%eax
+	xorl	%edi,%esi
+	andl	$0x33333333,%esi
+	xorl	%esi,%eax
+	xorl	%esi,%edi
+
+	roll	$18,%edi
+	movl	%edi,%esi
+	xorl	%eax,%edi
+	andl	$0xfff0000f,%edi
+	xorl	%edi,%esi
+	xorl	%edi,%eax
+
+	roll	$12,%esi
+	movl	%esi,%edi
+	xorl	%eax,%esi
+	andl	$0xf0f0f0f0,%esi
+	xorl	%esi,%edi
+	xorl	%esi,%eax
+
+	rorl	$4,%eax
+	movl	%eax,(%edx)
+	movl	%edi,4(%edx)
+	addl	$8,%esp
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret

deps/openssl/asm/x86-macosx-gas/md5/md5-586.s

@@ -0,0 +1,745 @@
+.file	"../openssl/crypto/md5/asm/md5-586.s"
+.text
+.globl	_md5_block_asm_data_order
+.align	4
+_md5_block_asm_data_order:
+L_md5_block_asm_data_order_begin:
+	pushl	%esi
+	pushl	%edi
+	movl	12(%esp),%edi
+	movl	16(%esp),%esi
+	movl	20(%esp),%ecx
+	pushl	%ebp
+	shll	$6,%ecx
+	pushl	%ebx
+	addl	%esi,%ecx
+	subl	$64,%ecx
+	movl	(%edi),%eax
+	pushl	%ecx
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+L000start:
+
+	# R0 section
+
+	movl	%ecx,%edi
+	movl	(%esi),%ebp
+	# R0 0
+
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	leal	3614090360(%eax,%ebp,1),%eax
+	xorl	%edx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$7,%eax
+	movl	4(%esi),%ebp
+	addl	%ebx,%eax
+	# R0 1
+
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	leal	3905402710(%edx,%ebp,1),%edx
+	xorl	%ecx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$12,%edx
+	movl	8(%esi),%ebp
+	addl	%eax,%edx
+	# R0 2
+
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	leal	606105819(%ecx,%ebp,1),%ecx
+	xorl	%ebx,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$17,%ecx
+	movl	12(%esi),%ebp
+	addl	%edx,%ecx
+	# R0 3
+
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	leal	3250441966(%ebx,%ebp,1),%ebx
+	xorl	%eax,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$22,%ebx
+	movl	16(%esi),%ebp
+	addl	%ecx,%ebx
+	# R0 4
+
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	leal	4118548399(%eax,%ebp,1),%eax
+	xorl	%edx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$7,%eax
+	movl	20(%esi),%ebp
+	addl	%ebx,%eax
+	# R0 5
+
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	leal	1200080426(%edx,%ebp,1),%edx
+	xorl	%ecx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$12,%edx
+	movl	24(%esi),%ebp
+	addl	%eax,%edx
+	# R0 6
+
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	leal	2821735955(%ecx,%ebp,1),%ecx
+	xorl	%ebx,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$17,%ecx
+	movl	28(%esi),%ebp
+	addl	%edx,%ecx
+	# R0 7
+
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	leal	4249261313(%ebx,%ebp,1),%ebx
+	xorl	%eax,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$22,%ebx
+	movl	32(%esi),%ebp
+	addl	%ecx,%ebx
+	# R0 8
+
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	leal	1770035416(%eax,%ebp,1),%eax
+	xorl	%edx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$7,%eax
+	movl	36(%esi),%ebp
+	addl	%ebx,%eax
+	# R0 9
+
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	leal	2336552879(%edx,%ebp,1),%edx
+	xorl	%ecx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$12,%edx
+	movl	40(%esi),%ebp
+	addl	%eax,%edx
+	# R0 10
+
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	leal	4294925233(%ecx,%ebp,1),%ecx
+	xorl	%ebx,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$17,%ecx
+	movl	44(%esi),%ebp
+	addl	%edx,%ecx
+	# R0 11
+
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	leal	2304563134(%ebx,%ebp,1),%ebx
+	xorl	%eax,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$22,%ebx
+	movl	48(%esi),%ebp
+	addl	%ecx,%ebx
+	# R0 12
+
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	leal	1804603682(%eax,%ebp,1),%eax
+	xorl	%edx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$7,%eax
+	movl	52(%esi),%ebp
+	addl	%ebx,%eax
+	# R0 13
+
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	leal	4254626195(%edx,%ebp,1),%edx
+	xorl	%ecx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$12,%edx
+	movl	56(%esi),%ebp
+	addl	%eax,%edx
+	# R0 14
+
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	leal	2792965006(%ecx,%ebp,1),%ecx
+	xorl	%ebx,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$17,%ecx
+	movl	60(%esi),%ebp
+	addl	%edx,%ecx
+	# R0 15
+
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	leal	1236535329(%ebx,%ebp,1),%ebx
+	xorl	%eax,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$22,%ebx
+	movl	4(%esi),%ebp
+	addl	%ecx,%ebx
+
+	# R1 section
+
+	# R1 16
+
+	leal	4129170786(%eax,%ebp,1),%eax
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	movl	24(%esi),%ebp
+	xorl	%ecx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$5,%eax
+	addl	%ebx,%eax
+	# R1 17
+
+	leal	3225465664(%edx,%ebp,1),%edx
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	movl	44(%esi),%ebp
+	xorl	%ebx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$9,%edx
+	addl	%eax,%edx
+	# R1 18
+
+	leal	643717713(%ecx,%ebp,1),%ecx
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	movl	(%esi),%ebp
+	xorl	%eax,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	# R1 19
+
+	leal	3921069994(%ebx,%ebp,1),%ebx
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	movl	20(%esi),%ebp
+	xorl	%edx,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	# R1 20
+
+	leal	3593408605(%eax,%ebp,1),%eax
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	movl	40(%esi),%ebp
+	xorl	%ecx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$5,%eax
+	addl	%ebx,%eax
+	# R1 21
+
+	leal	38016083(%edx,%ebp,1),%edx
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	movl	60(%esi),%ebp
+	xorl	%ebx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$9,%edx
+	addl	%eax,%edx
+	# R1 22
+
+	leal	3634488961(%ecx,%ebp,1),%ecx
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	movl	16(%esi),%ebp
+	xorl	%eax,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	# R1 23
+
+	leal	3889429448(%ebx,%ebp,1),%ebx
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	movl	36(%esi),%ebp
+	xorl	%edx,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	# R1 24
+
+	leal	568446438(%eax,%ebp,1),%eax
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	movl	56(%esi),%ebp
+	xorl	%ecx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$5,%eax
+	addl	%ebx,%eax
+	# R1 25
+
+	leal	3275163606(%edx,%ebp,1),%edx
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	movl	12(%esi),%ebp
+	xorl	%ebx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$9,%edx
+	addl	%eax,%edx
+	# R1 26
+
+	leal	4107603335(%ecx,%ebp,1),%ecx
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	movl	32(%esi),%ebp
+	xorl	%eax,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	# R1 27
+
+	leal	1163531501(%ebx,%ebp,1),%ebx
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	movl	52(%esi),%ebp
+	xorl	%edx,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+	# R1 28
+
+	leal	2850285829(%eax,%ebp,1),%eax
+	xorl	%ebx,%edi
+	andl	%edx,%edi
+	movl	8(%esi),%ebp
+	xorl	%ecx,%edi
+	addl	%edi,%eax
+	movl	%ebx,%edi
+	roll	$5,%eax
+	addl	%ebx,%eax
+	# R1 29
+
+	leal	4243563512(%edx,%ebp,1),%edx
+	xorl	%eax,%edi
+	andl	%ecx,%edi
+	movl	28(%esi),%ebp
+	xorl	%ebx,%edi
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$9,%edx
+	addl	%eax,%edx
+	# R1 30
+
+	leal	1735328473(%ecx,%ebp,1),%ecx
+	xorl	%edx,%edi
+	andl	%ebx,%edi
+	movl	48(%esi),%ebp
+	xorl	%eax,%edi
+	addl	%edi,%ecx
+	movl	%edx,%edi
+	roll	$14,%ecx
+	addl	%edx,%ecx
+	# R1 31
+
+	leal	2368359562(%ebx,%ebp,1),%ebx
+	xorl	%ecx,%edi
+	andl	%eax,%edi
+	movl	20(%esi),%ebp
+	xorl	%edx,%edi
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$20,%ebx
+	addl	%ecx,%ebx
+
+	# R2 section
+
+	# R2 32
+
+	xorl	%edx,%edi
+	xorl	%ebx,%edi
+	leal	4294588738(%eax,%ebp,1),%eax
+	addl	%edi,%eax
+	roll	$4,%eax
+	movl	32(%esi),%ebp
+	movl	%ebx,%edi
+	# R2 33
+
+	leal	2272392833(%edx,%ebp,1),%edx
+	addl	%ebx,%eax
+	xorl	%ecx,%edi
+	xorl	%eax,%edi
+	movl	44(%esi),%ebp
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$11,%edx
+	addl	%eax,%edx
+	# R2 34
+
+	xorl	%ebx,%edi
+	xorl	%edx,%edi
+	leal	1839030562(%ecx,%ebp,1),%ecx
+	addl	%edi,%ecx
+	roll	$16,%ecx
+	movl	56(%esi),%ebp
+	movl	%edx,%edi
+	# R2 35
+
+	leal	4259657740(%ebx,%ebp,1),%ebx
+	addl	%edx,%ecx
+	xorl	%eax,%edi
+	xorl	%ecx,%edi
+	movl	4(%esi),%ebp
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$23,%ebx
+	addl	%ecx,%ebx
+	# R2 36
+
+	xorl	%edx,%edi
+	xorl	%ebx,%edi
+	leal	2763975236(%eax,%ebp,1),%eax
+	addl	%edi,%eax
+	roll	$4,%eax
+	movl	16(%esi),%ebp
+	movl	%ebx,%edi
+	# R2 37
+
+	leal	1272893353(%edx,%ebp,1),%edx
+	addl	%ebx,%eax
+	xorl	%ecx,%edi
+	xorl	%eax,%edi
+	movl	28(%esi),%ebp
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$11,%edx
+	addl	%eax,%edx
+	# R2 38
+
+	xorl	%ebx,%edi
+	xorl	%edx,%edi
+	leal	4139469664(%ecx,%ebp,1),%ecx
+	addl	%edi,%ecx
+	roll	$16,%ecx
+	movl	40(%esi),%ebp
+	movl	%edx,%edi
+	# R2 39
+
+	leal	3200236656(%ebx,%ebp,1),%ebx
+	addl	%edx,%ecx
+	xorl	%eax,%edi
+	xorl	%ecx,%edi
+	movl	52(%esi),%ebp
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$23,%ebx
+	addl	%ecx,%ebx
+	# R2 40
+
+	xorl	%edx,%edi
+	xorl	%ebx,%edi
+	leal	681279174(%eax,%ebp,1),%eax
+	addl	%edi,%eax
+	roll	$4,%eax
+	movl	(%esi),%ebp
+	movl	%ebx,%edi
+	# R2 41
+
+	leal	3936430074(%edx,%ebp,1),%edx
+	addl	%ebx,%eax
+	xorl	%ecx,%edi
+	xorl	%eax,%edi
+	movl	12(%esi),%ebp
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$11,%edx
+	addl	%eax,%edx
+	# R2 42
+
+	xorl	%ebx,%edi
+	xorl	%edx,%edi
+	leal	3572445317(%ecx,%ebp,1),%ecx
+	addl	%edi,%ecx
+	roll	$16,%ecx
+	movl	24(%esi),%ebp
+	movl	%edx,%edi
+	# R2 43
+
+	leal	76029189(%ebx,%ebp,1),%ebx
+	addl	%edx,%ecx
+	xorl	%eax,%edi
+	xorl	%ecx,%edi
+	movl	36(%esi),%ebp
+	addl	%edi,%ebx
+	movl	%ecx,%edi
+	roll	$23,%ebx
+	addl	%ecx,%ebx
+	# R2 44
+
+	xorl	%edx,%edi
+	xorl	%ebx,%edi
+	leal	3654602809(%eax,%ebp,1),%eax
+	addl	%edi,%eax
+	roll	$4,%eax
+	movl	48(%esi),%ebp
+	movl	%ebx,%edi
+	# R2 45
+
+	leal	3873151461(%edx,%ebp,1),%edx
+	addl	%ebx,%eax
+	xorl	%ecx,%edi
+	xorl	%eax,%edi
+	movl	60(%esi),%ebp
+	addl	%edi,%edx
+	movl	%eax,%edi
+	roll	$11,%edx
+	addl	%eax,%edx
+	# R2 46
+
+	xorl	%ebx,%edi
+	xorl	%edx,%edi
+	leal	530742520(%ecx,%ebp,1),%ecx
+	addl	%edi,%ecx
+	roll	$16,%ecx
+	movl	8(%esi),%ebp
+	movl	%edx,%edi
+	# R2 47
+
+	leal	3299628645(%ebx,%ebp,1),%ebx
+	addl	%edx,%ecx
+	xorl	%eax,%edi
+	xorl	%ecx,%edi
+	movl	(%esi),%ebp
+	addl	%edi,%ebx
+	movl	$-1,%edi
+	roll	$23,%ebx
+	addl	%ecx,%ebx
+
+	# R3 section
+
+	# R3 48
+
+	xorl	%edx,%edi
+	orl	%ebx,%edi
+	leal	4096336452(%eax,%ebp,1),%eax
+	xorl	%ecx,%edi
+	movl	28(%esi),%ebp
+	addl	%edi,%eax
+	movl	$-1,%edi
+	roll	$6,%eax
+	xorl	%ecx,%edi
+	addl	%ebx,%eax
+	# R3 49
+
+	orl	%eax,%edi
+	leal	1126891415(%edx,%ebp,1),%edx
+	xorl	%ebx,%edi
+	movl	56(%esi),%ebp
+	addl	%edi,%edx
+	movl	$-1,%edi
+	roll	$10,%edx
+	xorl	%ebx,%edi
+	addl	%eax,%edx
+	# R3 50
+
+	orl	%edx,%edi
+	leal	2878612391(%ecx,%ebp,1),%ecx
+	xorl	%eax,%edi
+	movl	20(%esi),%ebp
+	addl	%edi,%ecx
+	movl	$-1,%edi
+	roll	$15,%ecx
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	# R3 51
+
+	orl	%ecx,%edi
+	leal	4237533241(%ebx,%ebp,1),%ebx
+	xorl	%edx,%edi
+	movl	48(%esi),%ebp
+	addl	%edi,%ebx
+	movl	$-1,%edi
+	roll	$21,%ebx
+	xorl	%edx,%edi
+	addl	%ecx,%ebx
+	# R3 52
+
+	orl	%ebx,%edi
+	leal	1700485571(%eax,%ebp,1),%eax
+	xorl	%ecx,%edi
+	movl	12(%esi),%ebp
+	addl	%edi,%eax
+	movl	$-1,%edi
+	roll	$6,%eax
+	xorl	%ecx,%edi
+	addl	%ebx,%eax
+	# R3 53
+
+	orl	%eax,%edi
+	leal	2399980690(%edx,%ebp,1),%edx
+	xorl	%ebx,%edi
+	movl	40(%esi),%ebp
+	addl	%edi,%edx
+	movl	$-1,%edi
+	roll	$10,%edx
+	xorl	%ebx,%edi
+	addl	%eax,%edx
+	# R3 54
+
+	orl	%edx,%edi
+	leal	4293915773(%ecx,%ebp,1),%ecx
+	xorl	%eax,%edi
+	movl	4(%esi),%ebp
+	addl	%edi,%ecx
+	movl	$-1,%edi
+	roll	$15,%ecx
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	# R3 55
+
+	orl	%ecx,%edi
+	leal	2240044497(%ebx,%ebp,1),%ebx
+	xorl	%edx,%edi
+	movl	32(%esi),%ebp
+	addl	%edi,%ebx
+	movl	$-1,%edi
+	roll	$21,%ebx
+	xorl	%edx,%edi
+	addl	%ecx,%ebx
+	# R3 56
+
+	orl	%ebx,%edi
+	leal	1873313359(%eax,%ebp,1),%eax
+	xorl	%ecx,%edi
+	movl	60(%esi),%ebp
+	addl	%edi,%eax
+	movl	$-1,%edi
+	roll	$6,%eax
+	xorl	%ecx,%edi
+	addl	%ebx,%eax
+	# R3 57
+
+	orl	%eax,%edi
+	leal	4264355552(%edx,%ebp,1),%edx
+	xorl	%ebx,%edi
+	movl	24(%esi),%ebp
+	addl	%edi,%edx
+	movl	$-1,%edi
+	roll	$10,%edx
+	xorl	%ebx,%edi
+	addl	%eax,%edx
+	# R3 58
+
+	orl	%edx,%edi
+	leal	2734768916(%ecx,%ebp,1),%ecx
+	xorl	%eax,%edi
+	movl	52(%esi),%ebp
+	addl	%edi,%ecx
+	movl	$-1,%edi
+	roll	$15,%ecx
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	# R3 59
+
+	orl	%ecx,%edi
+	leal	1309151649(%ebx,%ebp,1),%ebx
+	xorl	%edx,%edi
+	movl	16(%esi),%ebp
+	addl	%edi,%ebx
+	movl	$-1,%edi
+	roll	$21,%ebx
+	xorl	%edx,%edi
+	addl	%ecx,%ebx
+	# R3 60
+
+	orl	%ebx,%edi
+	leal	4149444226(%eax,%ebp,1),%eax
+	xorl	%ecx,%edi
+	movl	44(%esi),%ebp
+	addl	%edi,%eax
+	movl	$-1,%edi
+	roll	$6,%eax
+	xorl	%ecx,%edi
+	addl	%ebx,%eax
+	# R3 61
+
+	orl	%eax,%edi
+	leal	3174756917(%edx,%ebp,1),%edx
+	xorl	%ebx,%edi
+	movl	8(%esi),%ebp
+	addl	%edi,%edx
+	movl	$-1,%edi
+	roll	$10,%edx
+	xorl	%ebx,%edi
+	addl	%eax,%edx
+	# R3 62
+
+	orl	%edx,%edi
+	leal	718787259(%ecx,%ebp,1),%ecx
+	xorl	%eax,%edi
+	movl	36(%esi),%ebp
+	addl	%edi,%ecx
+	movl	$-1,%edi
+	roll	$15,%ecx
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	# R3 63
+
+	orl	%ecx,%edi
+	leal	3951481745(%ebx,%ebp,1),%ebx
+	xorl	%edx,%edi
+	movl	24(%esp),%ebp
+	addl	%edi,%ebx
+	addl	$64,%esi
+	roll	$21,%ebx
+	movl	(%ebp),%edi
+	addl	%ecx,%ebx
+	addl	%edi,%eax
+	movl	4(%ebp),%edi
+	addl	%edi,%ebx
+	movl	8(%ebp),%edi
+	addl	%edi,%ecx
+	movl	12(%ebp),%edi
+	addl	%edi,%edx
+	movl	%eax,(%ebp)
+	movl	%ebx,4(%ebp)
+	movl	(%esp),%edi
+	movl	%ecx,8(%ebp)
+	movl	%edx,12(%ebp)
+	cmpl	%esi,%edi
+	jae	L000start
+	popl	%eax
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret

deps/openssl/asm/x86-macosx-gas/rc4/rc4-586.s

@@ -0,0 +1,224 @@
+.file	"rc4-586.s"
+.text
+.globl	_RC4
+.align	4
+_RC4:
+L_RC4_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	movl	24(%esp),%edx
+	movl	28(%esp),%esi
+	movl	32(%esp),%ebp
+	xorl	%eax,%eax
+	xorl	%ebx,%ebx
+	cmpl	$0,%edx
+	je	L000abort
+	movb	(%edi),%al
+	movb	4(%edi),%bl
+	addl	$8,%edi
+	leal	(%esi,%edx,1),%ecx
+	subl	%esi,%ebp
+	movl	%ecx,24(%esp)
+	incb	%al
+	cmpl	$-1,256(%edi)
+	je	L001RC4_CHAR
+	movl	(%edi,%eax,4),%ecx
+	andl	$-4,%edx
+	jz	L002loop1
+	leal	-4(%esi,%edx,1),%edx
+	movl	%edx,28(%esp)
+	movl	%ebp,32(%esp)
+.align	4,0x90
+L003loop4:
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	movl	(%edi,%eax,4),%ecx
+	movl	(%edi,%edx,4),%ebp
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	rorl	$8,%ebp
+	movl	(%edi,%eax,4),%ecx
+	orl	(%edi,%edx,4),%ebp
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	rorl	$8,%ebp
+	movl	(%edi,%eax,4),%ecx
+	orl	(%edi,%edx,4),%ebp
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	rorl	$8,%ebp
+	movl	32(%esp),%ecx
+	orl	(%edi,%edx,4),%ebp
+	rorl	$8,%ebp
+	xorl	(%esi),%ebp
+	cmpl	28(%esp),%esi
+	movl	%ebp,(%ecx,%esi,1)
+	leal	4(%esi),%esi
+	movl	(%edi,%eax,4),%ecx
+	jb	L003loop4
+	cmpl	24(%esp),%esi
+	je	L004done
+	movl	32(%esp),%ebp
+.align	4,0x90
+L002loop1:
+	addb	%cl,%bl
+	movl	(%edi,%ebx,4),%edx
+	movl	%ecx,(%edi,%ebx,4)
+	movl	%edx,(%edi,%eax,4)
+	addl	%ecx,%edx
+	incb	%al
+	andl	$255,%edx
+	movl	(%edi,%edx,4),%edx
+	xorb	(%esi),%dl
+	leal	1(%esi),%esi
+	movl	(%edi,%eax,4),%ecx
+	cmpl	24(%esp),%esi
+	movb	%dl,-1(%ebp,%esi,1)
+	jb	L002loop1
+	jmp	L004done
+.align	4,0x90
+L001RC4_CHAR:
+	movzbl	(%edi,%eax,1),%ecx
+L005cloop1:
+	addb	%cl,%bl
+	movzbl	(%edi,%ebx,1),%edx
+	movb	%cl,(%edi,%ebx,1)
+	movb	%dl,(%edi,%eax,1)
+	addb	%cl,%dl
+	movzbl	(%edi,%edx,1),%edx
+	addb	$1,%al
+	xorb	(%esi),%dl
+	leal	1(%esi),%esi
+	movzbl	(%edi,%eax,1),%ecx
+	cmpl	24(%esp),%esi
+	movb	%dl,-1(%ebp,%esi,1)
+	jb	L005cloop1
+L004done:
+	decb	%al
+	movb	%bl,-4(%edi)
+	movb	%al,-8(%edi)
+L000abort:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_RC4_set_key
+.align	4
+_RC4_set_key:
+L_RC4_set_key_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	movl	24(%esp),%ebp
+	movl	28(%esp),%esi
+	leal	_OPENSSL_ia32cap_P,%edx
+	leal	8(%edi),%edi
+	leal	(%esi,%ebp,1),%esi
+	negl	%ebp
+	xorl	%eax,%eax
+	movl	%ebp,-4(%edi)
+	btl	$20,(%edx)
+	jc	L006c1stloop
+.align	4,0x90
+L007w1stloop:
+	movl	%eax,(%edi,%eax,4)
+	addb	$1,%al
+	jnc	L007w1stloop
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+.align	4,0x90
+L008w2ndloop:
+	movl	(%edi,%ecx,4),%eax
+	addb	(%esi,%ebp,1),%dl
+	addb	%al,%dl
+	addl	$1,%ebp
+	movl	(%edi,%edx,4),%ebx
+	jnz	L009wnowrap
+	movl	-4(%edi),%ebp
+L009wnowrap:
+	movl	%eax,(%edi,%edx,4)
+	movl	%ebx,(%edi,%ecx,4)
+	addb	$1,%cl
+	jnc	L008w2ndloop
+	jmp	L010exit
+.align	4,0x90
+L006c1stloop:
+	movb	%al,(%edi,%eax,1)
+	addb	$1,%al
+	jnc	L006c1stloop
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	xorl	%ebx,%ebx
+.align	4,0x90
+L011c2ndloop:
+	movb	(%edi,%ecx,1),%al
+	addb	(%esi,%ebp,1),%dl
+	addb	%al,%dl
+	addl	$1,%ebp
+	movb	(%edi,%edx,1),%bl
+	jnz	L012cnowrap
+	movl	-4(%edi),%ebp
+L012cnowrap:
+	movb	%al,(%edi,%edx,1)
+	movb	%bl,(%edi,%ecx,1)
+	addb	$1,%cl
+	jnc	L011c2ndloop
+	movl	$-1,256(%edi)
+L010exit:
+	xorl	%eax,%eax
+	movl	%eax,-8(%edi)
+	movl	%eax,-4(%edi)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_RC4_options
+.align	4
+_RC4_options:
+L_RC4_options_begin:
+	call	L013pic_point
+L013pic_point:
+	popl	%eax
+	leal	L014opts-L013pic_point(%eax),%eax
+	leal	_OPENSSL_ia32cap_P,%edx
+	btl	$20,(%edx)
+	jnc	L015skip
+	addl	$12,%eax
+L015skip:
+	ret
+.align	6,0x90
+L014opts:
+.byte	114,99,52,40,52,120,44,105,110,116,41,0
+.byte	114,99,52,40,49,120,44,99,104,97,114,41,0
+.byte	82,67,52,32,102,111,114,32,120,56,54,44,32,67,82,89
+.byte	80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114
+.byte	111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.align	6,0x90
+.comm	_OPENSSL_ia32cap_P,4

deps/openssl/asm/x86-macosx-gas/rc5/rc5-586.s

@@ -0,0 +1,563 @@
+.file	"rc5-586.s"
+.text
+.globl	_RC5_32_encrypt
+.align	4
+_RC5_32_encrypt:
+L_RC5_32_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%esi
+	pushl	%edi
+	movl	16(%esp),%edx
+	movl	20(%esp),%ebp
+	# Load the 2 words
+
+	movl	(%edx),%edi
+	movl	4(%edx),%esi
+	pushl	%ebx
+	movl	(%ebp),%ebx
+	addl	4(%ebp),%edi
+	addl	8(%ebp),%esi
+	xorl	%esi,%edi
+	movl	12(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	16(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	20(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	24(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	28(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	32(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	36(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	40(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	44(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	48(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	52(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	56(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	60(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	64(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	68(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	72(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	cmpl	$8,%ebx
+	je	L000rc5_exit
+	xorl	%esi,%edi
+	movl	76(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	80(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	84(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	88(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	92(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	96(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	100(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	104(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	cmpl	$12,%ebx
+	je	L000rc5_exit
+	xorl	%esi,%edi
+	movl	108(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	112(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	116(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	120(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	124(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	128(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+	xorl	%esi,%edi
+	movl	132(%ebp),%eax
+	movl	%esi,%ecx
+	roll	%cl,%edi
+	addl	%eax,%edi
+	xorl	%edi,%esi
+	movl	136(%ebp),%eax
+	movl	%edi,%ecx
+	roll	%cl,%esi
+	addl	%eax,%esi
+L000rc5_exit:
+	movl	%edi,(%edx)
+	movl	%esi,4(%edx)
+	popl	%ebx
+	popl	%edi
+	popl	%esi
+	popl	%ebp
+	ret
+.globl	_RC5_32_decrypt
+.align	4
+_RC5_32_decrypt:
+L_RC5_32_decrypt_begin:
+
+	pushl	%ebp
+	pushl	%esi
+	pushl	%edi
+	movl	16(%esp),%edx
+	movl	20(%esp),%ebp
+	# Load the 2 words
+
+	movl	(%edx),%edi
+	movl	4(%edx),%esi
+	pushl	%ebx
+	movl	(%ebp),%ebx
+	cmpl	$12,%ebx
+	je	L001rc5_dec_12
+	cmpl	$8,%ebx
+	je	L002rc5_dec_8
+	movl	136(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	132(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	128(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	124(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	120(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	116(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	112(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	108(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+L001rc5_dec_12:
+	movl	104(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	100(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	96(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	92(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	88(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	84(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	80(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	76(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+L002rc5_dec_8:
+	movl	72(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	68(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	64(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	60(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	56(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	52(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	48(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	44(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	40(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	36(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	32(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	28(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	24(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	20(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	movl	16(%ebp),%eax
+	subl	%eax,%esi
+	movl	%edi,%ecx
+	rorl	%cl,%esi
+	xorl	%edi,%esi
+	movl	12(%ebp),%eax
+	subl	%eax,%edi
+	movl	%esi,%ecx
+	rorl	%cl,%edi
+	xorl	%esi,%edi
+	subl	8(%ebp),%esi
+	subl	4(%ebp),%edi
+L003rc5_exit:
+	movl	%edi,(%edx)
+	movl	%esi,4(%edx)
+	popl	%ebx
+	popl	%edi
+	popl	%esi
+	popl	%ebp
+	ret
+.globl	_RC5_32_cbc_encrypt
+.align	4
+_RC5_32_cbc_encrypt:
+L_RC5_32_cbc_encrypt_begin:
+
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	28(%esp),%ebp
+	# getting iv ptr from parameter 4
+
+	movl	36(%esp),%ebx
+	movl	(%ebx),%esi
+	movl	4(%ebx),%edi
+	pushl	%edi
+	pushl	%esi
+	pushl	%edi
+	pushl	%esi
+	movl	%esp,%ebx
+	movl	36(%esp),%esi
+	movl	40(%esp),%edi
+	# getting encrypt flag from parameter 5
+
+	movl	56(%esp),%ecx
+	# get and push parameter 3
+
+	movl	48(%esp),%eax
+	pushl	%eax
+	pushl	%ebx
+	cmpl	$0,%ecx
+	jz	L004decrypt
+	andl	$4294967288,%ebp
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	jz	L005encrypt_finish
+L006encrypt_loop:
+	movl	(%esi),%ecx
+	movl	4(%esi),%edx
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_RC5_32_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	L006encrypt_loop
+L005encrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	L007finish
+	call	L008PIC_point
+L008PIC_point:
+	popl	%edx
+	leal	L009cbc_enc_jmp_table-L008PIC_point(%edx),%ecx
+	movl	(%ecx,%ebp,4),%ebp
+	addl	%edx,%ebp
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	jmp	*%ebp
+L010ej7:
+	movb	6(%esi),%dh
+	shll	$8,%edx
+L011ej6:
+	movb	5(%esi),%dh
+L012ej5:
+	movb	4(%esi),%dl
+L013ej4:
+	movl	(%esi),%ecx
+	jmp	L014ejend
+L015ej3:
+	movb	2(%esi),%ch
+	shll	$8,%ecx
+L016ej2:
+	movb	1(%esi),%ch
+L017ej1:
+	movb	(%esi),%cl
+L014ejend:
+	xorl	%ecx,%eax
+	xorl	%edx,%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_RC5_32_encrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	movl	%eax,(%edi)
+	movl	%ebx,4(%edi)
+	jmp	L007finish
+L004decrypt:
+	andl	$4294967288,%ebp
+	movl	16(%esp),%eax
+	movl	20(%esp),%ebx
+	jz	L018decrypt_finish
+L019decrypt_loop:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_RC5_32_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%ecx,(%edi)
+	movl	%edx,4(%edi)
+	movl	%eax,16(%esp)
+	movl	%ebx,20(%esp)
+	addl	$8,%esi
+	addl	$8,%edi
+	subl	$8,%ebp
+	jnz	L019decrypt_loop
+L018decrypt_finish:
+	movl	52(%esp),%ebp
+	andl	$7,%ebp
+	jz	L007finish
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+	call	L_RC5_32_decrypt_begin
+	movl	8(%esp),%eax
+	movl	12(%esp),%ebx
+	movl	16(%esp),%ecx
+	movl	20(%esp),%edx
+	xorl	%eax,%ecx
+	xorl	%ebx,%edx
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+L020dj7:
+	rorl	$16,%edx
+	movb	%dl,6(%edi)
+	shrl	$16,%edx
+L021dj6:
+	movb	%dh,5(%edi)
+L022dj5:
+	movb	%dl,4(%edi)
+L023dj4:
+	movl	%ecx,(%edi)
+	jmp	L024djend
+L025dj3:
+	rorl	$16,%ecx
+	movb	%cl,2(%edi)
+	shll	$16,%ecx
+L026dj2:
+	movb	%ch,1(%esi)
+L027dj1:
+	movb	%cl,(%esi)
+L024djend:
+	jmp	L007finish
+L007finish:
+	movl	60(%esp),%ecx
+	addl	$24,%esp
+	movl	%eax,(%ecx)
+	movl	%ebx,4(%ecx)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	6,0x90
+L009cbc_enc_jmp_table:
+.long	0
+.long	L017ej1-L008PIC_point
+.long	L016ej2-L008PIC_point
+.long	L015ej3-L008PIC_point
+.long	L013ej4-L008PIC_point
+.long	L012ej5-L008PIC_point
+.long	L011ej6-L008PIC_point
+.long	L010ej7-L008PIC_point
+.align	6,0x90

deps/openssl/asm/x86-macosx-gas/sha/sha256-586.s

@@ -0,0 +1,259 @@
+.file	"sha512-586.s"
+.text
+.globl	_sha256_block_data_order
+.align	4
+_sha256_block_data_order:
+L_sha256_block_data_order_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	movl	24(%esp),%edi
+	movl	28(%esp),%eax
+	movl	%esp,%ebx
+	call	L000pic_point
+L000pic_point:
+	popl	%ebp
+	leal	L001K256-L000pic_point(%ebp),%ebp
+	subl	$16,%esp
+	andl	$-64,%esp
+	shll	$6,%eax
+	addl	%edi,%eax
+	movl	%esi,(%esp)
+	movl	%edi,4(%esp)
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+.align	4,0x90
+L002loop:
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	16(%edi),%eax
+	movl	20(%edi),%ebx
+	movl	24(%edi),%ecx
+	movl	28(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	32(%edi),%eax
+	movl	36(%edi),%ebx
+	movl	40(%edi),%ecx
+	movl	44(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	48(%edi),%eax
+	movl	52(%edi),%ebx
+	movl	56(%edi),%ecx
+	movl	60(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	addl	$64,%edi
+	subl	$32,%esp
+	movl	%edi,100(%esp)
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edi
+	movl	%ebx,4(%esp)
+	movl	%ecx,8(%esp)
+	movl	%edi,12(%esp)
+	movl	16(%esi),%edx
+	movl	20(%esi),%ebx
+	movl	24(%esi),%ecx
+	movl	28(%esi),%edi
+	movl	%ebx,20(%esp)
+	movl	%ecx,24(%esp)
+	movl	%edi,28(%esp)
+.align	4,0x90
+L00300_15:
+	movl	92(%esp),%ebx
+	movl	%edx,%ecx
+	rorl	$6,%ecx
+	movl	%edx,%edi
+	rorl	$11,%edi
+	movl	20(%esp),%esi
+	xorl	%edi,%ecx
+	rorl	$14,%edi
+	xorl	%edi,%ecx
+	movl	24(%esp),%edi
+	addl	%ecx,%ebx
+	movl	%edx,16(%esp)
+	xorl	%edi,%esi
+	movl	%eax,%ecx
+	andl	%edx,%esi
+	movl	12(%esp),%edx
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	addl	%esi,%ebx
+	rorl	$2,%ecx
+	addl	28(%esp),%ebx
+	rorl	$13,%edi
+	movl	4(%esp),%esi
+	xorl	%edi,%ecx
+	rorl	$9,%edi
+	addl	%ebx,%edx
+	xorl	%edi,%ecx
+	movl	8(%esp),%edi
+	addl	%ecx,%ebx
+	movl	%eax,(%esp)
+	movl	%eax,%ecx
+	subl	$4,%esp
+	orl	%esi,%eax
+	andl	%esi,%ecx
+	andl	%edi,%eax
+	movl	(%ebp),%esi
+	orl	%ecx,%eax
+	addl	$4,%ebp
+	addl	%ebx,%eax
+	addl	%esi,%edx
+	addl	%esi,%eax
+	cmpl	$3248222580,%esi
+	jne	L00300_15
+	movl	152(%esp),%ebx
+.align	4,0x90
+L00416_63:
+	movl	%ebx,%esi
+	movl	100(%esp),%ecx
+	shrl	$3,%ebx
+	rorl	$7,%esi
+	xorl	%esi,%ebx
+	rorl	$11,%esi
+	movl	%ecx,%edi
+	xorl	%esi,%ebx
+	shrl	$10,%ecx
+	movl	156(%esp),%esi
+	rorl	$17,%edi
+	xorl	%edi,%ecx
+	rorl	$2,%edi
+	addl	%esi,%ebx
+	xorl	%ecx,%edi
+	addl	%edi,%ebx
+	movl	%edx,%ecx
+	addl	120(%esp),%ebx
+	rorl	$6,%ecx
+	movl	%edx,%edi
+	rorl	$11,%edi
+	movl	20(%esp),%esi
+	xorl	%edi,%ecx
+	rorl	$14,%edi
+	movl	%ebx,92(%esp)
+	xorl	%edi,%ecx
+	movl	24(%esp),%edi
+	addl	%ecx,%ebx
+	movl	%edx,16(%esp)
+	xorl	%edi,%esi
+	movl	%eax,%ecx
+	andl	%edx,%esi
+	movl	12(%esp),%edx
+	xorl	%edi,%esi
+	movl	%eax,%edi
+	addl	%esi,%ebx
+	rorl	$2,%ecx
+	addl	28(%esp),%ebx
+	rorl	$13,%edi
+	movl	4(%esp),%esi
+	xorl	%edi,%ecx
+	rorl	$9,%edi
+	addl	%ebx,%edx
+	xorl	%edi,%ecx
+	movl	8(%esp),%edi
+	addl	%ecx,%ebx
+	movl	%eax,(%esp)
+	movl	%eax,%ecx
+	subl	$4,%esp
+	orl	%esi,%eax
+	andl	%esi,%ecx
+	andl	%edi,%eax
+	movl	(%ebp),%esi
+	orl	%ecx,%eax
+	addl	$4,%ebp
+	addl	%ebx,%eax
+	movl	152(%esp),%ebx
+	addl	%esi,%edx
+	addl	%esi,%eax
+	cmpl	$3329325298,%esi
+	jne	L00416_63
+	movl	352(%esp),%esi
+	movl	4(%esp),%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edi
+	addl	(%esi),%eax
+	addl	4(%esi),%ebx
+	addl	8(%esi),%ecx
+	addl	12(%esi),%edi
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	movl	%ecx,8(%esi)
+	movl	%edi,12(%esi)
+	movl	20(%esp),%eax
+	movl	24(%esp),%ebx
+	movl	28(%esp),%ecx
+	movl	356(%esp),%edi
+	addl	16(%esi),%edx
+	addl	20(%esi),%eax
+	addl	24(%esi),%ebx
+	addl	28(%esi),%ecx
+	movl	%edx,16(%esi)
+	movl	%eax,20(%esi)
+	movl	%ebx,24(%esi)
+	movl	%ecx,28(%esi)
+	addl	$352,%esp
+	subl	$256,%ebp
+	cmpl	8(%esp),%edi
+	jb	L002loop
+	movl	12(%esp),%esp
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	6,0x90
+L001K256:
+.long	1116352408,1899447441,3049323471,3921009573
+.long	961987163,1508970993,2453635748,2870763221
+.long	3624381080,310598401,607225278,1426881987
+.long	1925078388,2162078206,2614888103,3248222580
+.long	3835390401,4022224774,264347078,604807628
+.long	770255983,1249150122,1555081692,1996064986
+.long	2554220882,2821834349,2952996808,3210313671
+.long	3336571891,3584528711,113926993,338241895
+.long	666307205,773529912,1294757372,1396182291
+.long	1695183700,1986661051,2177026350,2456956037
+.long	2730485921,2820302411,3259730800,3345764771
+.long	3516065817,3600352804,4094571909,275423344
+.long	430227734,506948616,659060556,883997877
+.long	958139571,1322822218,1537002063,1747873779
+.long	1955562222,2024104815,2227730452,2361852424
+.long	2428436474,2756734187,3204031479,3329325298
+.byte	83,72,65,50,53,54,32,98,108,111,99,107,32,116,114,97
+.byte	110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+.byte	67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+.byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+.byte	62,0

deps/openssl/asm/x86-macosx-gas/sha/sha512-586.s

@@ -0,0 +1,561 @@
+.file	"sha512-586.s"
+.text
+.globl	_sha512_block_data_order
+.align	4
+_sha512_block_data_order:
+L_sha512_block_data_order_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	movl	24(%esp),%edi
+	movl	28(%esp),%eax
+	movl	%esp,%ebx
+	call	L000pic_point
+L000pic_point:
+	popl	%ebp
+	leal	L001K512-L000pic_point(%ebp),%ebp
+	subl	$16,%esp
+	andl	$-64,%esp
+	shll	$7,%eax
+	addl	%edi,%eax
+	movl	%esi,(%esp)
+	movl	%edi,4(%esp)
+	movl	%eax,8(%esp)
+	movl	%ebx,12(%esp)
+.align	4,0x90
+L002loop_x86:
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	16(%edi),%eax
+	movl	20(%edi),%ebx
+	movl	24(%edi),%ecx
+	movl	28(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	32(%edi),%eax
+	movl	36(%edi),%ebx
+	movl	40(%edi),%ecx
+	movl	44(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	48(%edi),%eax
+	movl	52(%edi),%ebx
+	movl	56(%edi),%ecx
+	movl	60(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	64(%edi),%eax
+	movl	68(%edi),%ebx
+	movl	72(%edi),%ecx
+	movl	76(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	80(%edi),%eax
+	movl	84(%edi),%ebx
+	movl	88(%edi),%ecx
+	movl	92(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	96(%edi),%eax
+	movl	100(%edi),%ebx
+	movl	104(%edi),%ecx
+	movl	108(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	movl	112(%edi),%eax
+	movl	116(%edi),%ebx
+	movl	120(%edi),%ecx
+	movl	124(%edi),%edx
+	bswap	%eax
+	bswap	%ebx
+	bswap	%ecx
+	bswap	%edx
+	pushl	%eax
+	pushl	%ebx
+	pushl	%ecx
+	pushl	%edx
+	addl	$128,%edi
+	subl	$72,%esp
+	movl	%edi,204(%esp)
+	leal	8(%esp),%edi
+	movl	$16,%ecx
+.long	2784229001
+.align	4,0x90
+L00300_15_x86:
+	movl	40(%esp),%ecx
+	movl	44(%esp),%edx
+	movl	%ecx,%esi
+	shrl	$9,%ecx
+	movl	%edx,%edi
+	shrl	$9,%edx
+	movl	%ecx,%ebx
+	shll	$14,%esi
+	movl	%edx,%eax
+	shll	$14,%edi
+	xorl	%esi,%ebx
+	shrl	$5,%ecx
+	xorl	%edi,%eax
+	shrl	$5,%edx
+	xorl	%ecx,%eax
+	shll	$4,%esi
+	xorl	%edx,%ebx
+	shll	$4,%edi
+	xorl	%esi,%ebx
+	shrl	$4,%ecx
+	xorl	%edi,%eax
+	shrl	$4,%edx
+	xorl	%ecx,%eax
+	shll	$5,%esi
+	xorl	%edx,%ebx
+	shll	$5,%edi
+	xorl	%esi,%eax
+	xorl	%edi,%ebx
+	movl	48(%esp),%ecx
+	movl	52(%esp),%edx
+	movl	56(%esp),%esi
+	movl	60(%esp),%edi
+	addl	64(%esp),%eax
+	adcl	68(%esp),%ebx
+	xorl	%esi,%ecx
+	xorl	%edi,%edx
+	andl	40(%esp),%ecx
+	andl	44(%esp),%edx
+	addl	192(%esp),%eax
+	adcl	196(%esp),%ebx
+	xorl	%esi,%ecx
+	xorl	%edi,%edx
+	movl	(%ebp),%esi
+	movl	4(%ebp),%edi
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	32(%esp),%ecx
+	movl	36(%esp),%edx
+	addl	%esi,%eax
+	adcl	%edi,%ebx
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edx
+	movl	%eax,32(%esp)
+	movl	%ebx,36(%esp)
+	movl	%ecx,%esi
+	shrl	$2,%ecx
+	movl	%edx,%edi
+	shrl	$2,%edx
+	movl	%ecx,%ebx
+	shll	$4,%esi
+	movl	%edx,%eax
+	shll	$4,%edi
+	xorl	%esi,%ebx
+	shrl	$5,%ecx
+	xorl	%edi,%eax
+	shrl	$5,%edx
+	xorl	%ecx,%ebx
+	shll	$21,%esi
+	xorl	%edx,%eax
+	shll	$21,%edi
+	xorl	%esi,%eax
+	shrl	$21,%ecx
+	xorl	%edi,%ebx
+	shrl	$21,%edx
+	xorl	%ecx,%eax
+	shll	$5,%esi
+	xorl	%edx,%ebx
+	shll	$5,%edi
+	xorl	%esi,%eax
+	xorl	%edi,%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edx
+	movl	16(%esp),%esi
+	movl	20(%esp),%edi
+	addl	(%esp),%eax
+	adcl	4(%esp),%ebx
+	orl	%esi,%ecx
+	orl	%edi,%edx
+	andl	24(%esp),%ecx
+	andl	28(%esp),%edx
+	andl	8(%esp),%esi
+	andl	12(%esp),%edi
+	orl	%esi,%ecx
+	orl	%edi,%edx
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	movb	(%ebp),%dl
+	subl	$8,%esp
+	leal	8(%ebp),%ebp
+	cmpb	$148,%dl
+	jne	L00300_15_x86
+.align	4,0x90
+L00416_79_x86:
+	movl	312(%esp),%ecx
+	movl	316(%esp),%edx
+	movl	%ecx,%esi
+	shrl	$1,%ecx
+	movl	%edx,%edi
+	shrl	$1,%edx
+	movl	%ecx,%eax
+	shll	$24,%esi
+	movl	%edx,%ebx
+	shll	$24,%edi
+	xorl	%esi,%ebx
+	shrl	$6,%ecx
+	xorl	%edi,%eax
+	shrl	$6,%edx
+	xorl	%ecx,%eax
+	shll	$7,%esi
+	xorl	%edx,%ebx
+	shll	$1,%edi
+	xorl	%esi,%ebx
+	shrl	$1,%ecx
+	xorl	%edi,%eax
+	shrl	$1,%edx
+	xorl	%ecx,%eax
+	shll	$6,%edi
+	xorl	%edx,%ebx
+	xorl	%edi,%eax
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	movl	208(%esp),%ecx
+	movl	212(%esp),%edx
+	movl	%ecx,%esi
+	shrl	$6,%ecx
+	movl	%edx,%edi
+	shrl	$6,%edx
+	movl	%ecx,%eax
+	shll	$3,%esi
+	movl	%edx,%ebx
+	shll	$3,%edi
+	xorl	%esi,%eax
+	shrl	$13,%ecx
+	xorl	%edi,%ebx
+	shrl	$13,%edx
+	xorl	%ecx,%eax
+	shll	$10,%esi
+	xorl	%edx,%ebx
+	shll	$10,%edi
+	xorl	%esi,%ebx
+	shrl	$10,%ecx
+	xorl	%edi,%eax
+	shrl	$10,%edx
+	xorl	%ecx,%ebx
+	shll	$13,%edi
+	xorl	%edx,%eax
+	xorl	%edi,%eax
+	movl	320(%esp),%ecx
+	movl	324(%esp),%edx
+	addl	(%esp),%eax
+	adcl	4(%esp),%ebx
+	movl	248(%esp),%esi
+	movl	252(%esp),%edi
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	addl	%esi,%eax
+	adcl	%edi,%ebx
+	movl	%eax,192(%esp)
+	movl	%ebx,196(%esp)
+	movl	40(%esp),%ecx
+	movl	44(%esp),%edx
+	movl	%ecx,%esi
+	shrl	$9,%ecx
+	movl	%edx,%edi
+	shrl	$9,%edx
+	movl	%ecx,%ebx
+	shll	$14,%esi
+	movl	%edx,%eax
+	shll	$14,%edi
+	xorl	%esi,%ebx
+	shrl	$5,%ecx
+	xorl	%edi,%eax
+	shrl	$5,%edx
+	xorl	%ecx,%eax
+	shll	$4,%esi
+	xorl	%edx,%ebx
+	shll	$4,%edi
+	xorl	%esi,%ebx
+	shrl	$4,%ecx
+	xorl	%edi,%eax
+	shrl	$4,%edx
+	xorl	%ecx,%eax
+	shll	$5,%esi
+	xorl	%edx,%ebx
+	shll	$5,%edi
+	xorl	%esi,%eax
+	xorl	%edi,%ebx
+	movl	48(%esp),%ecx
+	movl	52(%esp),%edx
+	movl	56(%esp),%esi
+	movl	60(%esp),%edi
+	addl	64(%esp),%eax
+	adcl	68(%esp),%ebx
+	xorl	%esi,%ecx
+	xorl	%edi,%edx
+	andl	40(%esp),%ecx
+	andl	44(%esp),%edx
+	addl	192(%esp),%eax
+	adcl	196(%esp),%ebx
+	xorl	%esi,%ecx
+	xorl	%edi,%edx
+	movl	(%ebp),%esi
+	movl	4(%ebp),%edi
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	32(%esp),%ecx
+	movl	36(%esp),%edx
+	addl	%esi,%eax
+	adcl	%edi,%ebx
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edx
+	movl	%eax,32(%esp)
+	movl	%ebx,36(%esp)
+	movl	%ecx,%esi
+	shrl	$2,%ecx
+	movl	%edx,%edi
+	shrl	$2,%edx
+	movl	%ecx,%ebx
+	shll	$4,%esi
+	movl	%edx,%eax
+	shll	$4,%edi
+	xorl	%esi,%ebx
+	shrl	$5,%ecx
+	xorl	%edi,%eax
+	shrl	$5,%edx
+	xorl	%ecx,%ebx
+	shll	$21,%esi
+	xorl	%edx,%eax
+	shll	$21,%edi
+	xorl	%esi,%eax
+	shrl	$21,%ecx
+	xorl	%edi,%ebx
+	shrl	$21,%edx
+	xorl	%ecx,%eax
+	shll	$5,%esi
+	xorl	%edx,%ebx
+	shll	$5,%edi
+	xorl	%esi,%eax
+	xorl	%edi,%ebx
+	movl	8(%esp),%ecx
+	movl	12(%esp),%edx
+	movl	16(%esp),%esi
+	movl	20(%esp),%edi
+	addl	(%esp),%eax
+	adcl	4(%esp),%ebx
+	orl	%esi,%ecx
+	orl	%edi,%edx
+	andl	24(%esp),%ecx
+	andl	28(%esp),%edx
+	andl	8(%esp),%esi
+	andl	12(%esp),%edi
+	orl	%esi,%ecx
+	orl	%edi,%edx
+	addl	%ecx,%eax
+	adcl	%edx,%ebx
+	movl	%eax,(%esp)
+	movl	%ebx,4(%esp)
+	movb	(%ebp),%dl
+	subl	$8,%esp
+	leal	8(%ebp),%ebp
+	cmpb	$23,%dl
+	jne	L00416_79_x86
+	movl	840(%esp),%esi
+	movl	844(%esp),%edi
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	addl	8(%esp),%eax
+	adcl	12(%esp),%ebx
+	movl	%eax,(%esi)
+	movl	%ebx,4(%esi)
+	addl	16(%esp),%ecx
+	adcl	20(%esp),%edx
+	movl	%ecx,8(%esi)
+	movl	%edx,12(%esi)
+	movl	16(%esi),%eax
+	movl	20(%esi),%ebx
+	movl	24(%esi),%ecx
+	movl	28(%esi),%edx
+	addl	24(%esp),%eax
+	adcl	28(%esp),%ebx
+	movl	%eax,16(%esi)
+	movl	%ebx,20(%esi)
+	addl	32(%esp),%ecx
+	adcl	36(%esp),%edx
+	movl	%ecx,24(%esi)
+	movl	%edx,28(%esi)
+	movl	32(%esi),%eax
+	movl	36(%esi),%ebx
+	movl	40(%esi),%ecx
+	movl	44(%esi),%edx
+	addl	40(%esp),%eax
+	adcl	44(%esp),%ebx
+	movl	%eax,32(%esi)
+	movl	%ebx,36(%esi)
+	addl	48(%esp),%ecx
+	adcl	52(%esp),%edx
+	movl	%ecx,40(%esi)
+	movl	%edx,44(%esi)
+	movl	48(%esi),%eax
+	movl	52(%esi),%ebx
+	movl	56(%esi),%ecx
+	movl	60(%esi),%edx
+	addl	56(%esp),%eax
+	adcl	60(%esp),%ebx
+	movl	%eax,48(%esi)
+	movl	%ebx,52(%esi)
+	addl	64(%esp),%ecx
+	adcl	68(%esp),%edx
+	movl	%ecx,56(%esi)
+	movl	%edx,60(%esi)
+	addl	$840,%esp
+	subl	$640,%ebp
+	cmpl	8(%esp),%edi
+	jb	L002loop_x86
+	movl	12(%esp),%esp
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	6,0x90
+L001K512:
+.long	3609767458,1116352408
+.long	602891725,1899447441
+.long	3964484399,3049323471
+.long	2173295548,3921009573
+.long	4081628472,961987163
+.long	3053834265,1508970993
+.long	2937671579,2453635748
+.long	3664609560,2870763221
+.long	2734883394,3624381080
+.long	1164996542,310598401
+.long	1323610764,607225278
+.long	3590304994,1426881987
+.long	4068182383,1925078388
+.long	991336113,2162078206
+.long	633803317,2614888103
+.long	3479774868,3248222580
+.long	2666613458,3835390401
+.long	944711139,4022224774
+.long	2341262773,264347078
+.long	2007800933,604807628
+.long	1495990901,770255983
+.long	1856431235,1249150122
+.long	3175218132,1555081692
+.long	2198950837,1996064986
+.long	3999719339,2554220882
+.long	766784016,2821834349
+.long	2566594879,2952996808
+.long	3203337956,3210313671
+.long	1034457026,3336571891
+.long	2466948901,3584528711
+.long	3758326383,113926993
+.long	168717936,338241895
+.long	1188179964,666307205
+.long	1546045734,773529912
+.long	1522805485,1294757372
+.long	2643833823,1396182291
+.long	2343527390,1695183700
+.long	1014477480,1986661051
+.long	1206759142,2177026350
+.long	344077627,2456956037
+.long	1290863460,2730485921
+.long	3158454273,2820302411
+.long	3505952657,3259730800
+.long	106217008,3345764771
+.long	3606008344,3516065817
+.long	1432725776,3600352804
+.long	1467031594,4094571909
+.long	851169720,275423344
+.long	3100823752,430227734
+.long	1363258195,506948616
+.long	3750685593,659060556
+.long	3785050280,883997877
+.long	3318307427,958139571
+.long	3812723403,1322822218
+.long	2003034995,1537002063
+.long	3602036899,1747873779
+.long	1575990012,1955562222
+.long	1125592928,2024104815
+.long	2716904306,2227730452
+.long	442776044,2361852424
+.long	593698344,2428436474
+.long	3733110249,2756734187
+.long	2999351573,3204031479
+.long	3815920427,3329325298
+.long	3928383900,3391569614
+.long	566280711,3515267271
+.long	3454069534,3940187606
+.long	4000239992,4118630271
+.long	1914138554,116418474
+.long	2731055270,174292421
+.long	3203993006,289380356
+.long	320620315,460393269
+.long	587496836,685471733
+.long	1086792851,852142971
+.long	365543100,1017036298
+.long	2618297676,1126000580
+.long	3409855158,1288033470
+.long	4234509866,1501505948
+.long	987167468,1607167915
+.long	1246189591,1816402316
+.byte	83,72,65,53,49,50,32,98,108,111,99,107,32,116,114,97
+.byte	110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+.byte	67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+.byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+.byte	62,0

deps/openssl/asm/x86-macosx-gas/x86cpuid.s

@@ -0,0 +1,261 @@
+.file	"x86cpuid.s"
+.text
+.globl	_OPENSSL_ia32_cpuid
+.align	4
+_OPENSSL_ia32_cpuid:
+L_OPENSSL_ia32_cpuid_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	xorl	%edx,%edx
+	pushfl
+	popl	%eax
+	movl	%eax,%ecx
+	xorl	$2097152,%eax
+	pushl	%eax
+	popfl
+	pushfl
+	popl	%eax
+	xorl	%eax,%ecx
+	btl	$21,%ecx
+	jnc	L000done
+	xorl	%eax,%eax
+	.byte	0x0f,0xa2
+	movl	%eax,%edi
+	xorl	%eax,%eax
+	cmpl	$1970169159,%ebx
+	setne	%al
+	movl	%eax,%ebp
+	cmpl	$1231384169,%edx
+	setne	%al
+	orl	%eax,%ebp
+	cmpl	$1818588270,%ecx
+	setne	%al
+	orl	%eax,%ebp
+	jz	L001intel
+	cmpl	$1752462657,%ebx
+	setne	%al
+	movl	%eax,%esi
+	cmpl	$1769238117,%edx
+	setne	%al
+	orl	%eax,%esi
+	cmpl	$1145913699,%ecx
+	setne	%al
+	orl	%eax,%esi
+	jnz	L001intel
+	movl	$2147483648,%eax
+	.byte	0x0f,0xa2
+	cmpl	$2147483656,%eax
+	jb	L001intel
+	movl	$2147483656,%eax
+	.byte	0x0f,0xa2
+	movzbl	%cl,%esi
+	incl	%esi
+	movl	$1,%eax
+	.byte	0x0f,0xa2
+	btl	$28,%edx
+	jnc	L000done
+	shrl	$16,%ebx
+	andl	$255,%ebx
+	cmpl	%esi,%ebx
+	ja	L000done
+	andl	$4026531839,%edx
+	jmp	L000done
+L001intel:
+	cmpl	$4,%edi
+	movl	$-1,%edi
+	jb	L002nocacheinfo
+	movl	$4,%eax
+	movl	$0,%ecx
+	.byte	0x0f,0xa2
+	movl	%eax,%edi
+	shrl	$14,%edi
+	andl	$4095,%edi
+L002nocacheinfo:
+	movl	$1,%eax
+	.byte	0x0f,0xa2
+	cmpl	$0,%ebp
+	jne	L003notP4
+	andb	$15,%ah
+	cmpb	$15,%ah
+	jne	L003notP4
+	orl	$1048576,%edx
+L003notP4:
+	btl	$28,%edx
+	jnc	L000done
+	andl	$4026531839,%edx
+	cmpl	$0,%edi
+	je	L000done
+	orl	$268435456,%edx
+	shrl	$16,%ebx
+	cmpb	$1,%bl
+	ja	L000done
+	andl	$4026531839,%edx
+L000done:
+	movl	%edx,%eax
+	movl	%ecx,%edx
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_OPENSSL_rdtsc
+.align	4
+_OPENSSL_rdtsc:
+L_OPENSSL_rdtsc_begin:
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	leal	_OPENSSL_ia32cap_P,%ecx
+	btl	$4,(%ecx)
+	jnc	L004notsc
+	.byte	0x0f,0x31
+L004notsc:
+	ret
+.globl	_OPENSSL_instrument_halt
+.align	4
+_OPENSSL_instrument_halt:
+L_OPENSSL_instrument_halt_begin:
+	leal	_OPENSSL_ia32cap_P,%ecx
+	btl	$4,(%ecx)
+	jnc	L005nohalt
+.long	2421723150
+	andl	$3,%eax
+	jnz	L005nohalt
+	pushfl
+	popl	%eax
+	btl	$9,%eax
+	jnc	L005nohalt
+	.byte	0x0f,0x31
+	pushl	%edx
+	pushl	%eax
+	hlt
+	.byte	0x0f,0x31
+	subl	(%esp),%eax
+	sbbl	4(%esp),%edx
+	addl	$8,%esp
+	ret
+L005nohalt:
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	ret
+.globl	_OPENSSL_far_spin
+.align	4
+_OPENSSL_far_spin:
+L_OPENSSL_far_spin_begin:
+	pushfl
+	popl	%eax
+	btl	$9,%eax
+	jnc	L006nospin
+	movl	4(%esp),%eax
+	movl	8(%esp),%ecx
+.long	2430111262
+	xorl	%eax,%eax
+	movl	(%ecx),%edx
+	jmp	L007spin
+.align	4,0x90
+L007spin:
+	incl	%eax
+	cmpl	(%ecx),%edx
+	je	L007spin
+.long	529567888
+	ret
+L006nospin:
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	ret
+.globl	_OPENSSL_wipe_cpu
+.align	4
+_OPENSSL_wipe_cpu:
+L_OPENSSL_wipe_cpu_begin:
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	leal	_OPENSSL_ia32cap_P,%ecx
+	movl	(%ecx),%ecx
+	btl	$1,(%ecx)
+	jnc	L008no_x87
+.long	4007259865,4007259865,4007259865,4007259865,2430851995
+L008no_x87:
+	leal	4(%esp),%eax
+	ret
+.globl	_OPENSSL_atomic_add
+.align	4
+_OPENSSL_atomic_add:
+L_OPENSSL_atomic_add_begin:
+	movl	4(%esp),%edx
+	movl	8(%esp),%ecx
+	pushl	%ebx
+	nop
+	movl	(%edx),%eax
+L009spin:
+	leal	(%eax,%ecx,1),%ebx
+	nop
+.long	447811568
+	jne	L009spin
+	movl	%ebx,%eax
+	popl	%ebx
+	ret
+.globl	_OPENSSL_indirect_call
+.align	4
+_OPENSSL_indirect_call:
+L_OPENSSL_indirect_call_begin:
+	pushl	%ebp
+	movl	%esp,%ebp
+	subl	$28,%esp
+	movl	12(%ebp),%ecx
+	movl	%ecx,(%esp)
+	movl	16(%ebp),%edx
+	movl	%edx,4(%esp)
+	movl	20(%ebp),%eax
+	movl	%eax,8(%esp)
+	movl	24(%ebp),%eax
+	movl	%eax,12(%esp)
+	movl	28(%ebp),%eax
+	movl	%eax,16(%esp)
+	movl	32(%ebp),%eax
+	movl	%eax,20(%esp)
+	movl	36(%ebp),%eax
+	movl	%eax,24(%esp)
+	call	*8(%ebp)
+	movl	%ebp,%esp
+	popl	%ebp
+	ret
+.globl	_OPENSSL_cleanse
+.align	4
+_OPENSSL_cleanse:
+L_OPENSSL_cleanse_begin:
+	movl	4(%esp),%edx
+	movl	8(%esp),%ecx
+	xorl	%eax,%eax
+	cmpl	$7,%ecx
+	jae	L010lot
+	cmpl	$0,%ecx
+	je	L011ret
+L012little:
+	movb	%al,(%edx)
+	subl	$1,%ecx
+	leal	1(%edx),%edx
+	jnz	L012little
+L011ret:
+	ret
+.align	4,0x90
+L010lot:
+	testl	$3,%edx
+	jz	L013aligned
+	movb	%al,(%edx)
+	leal	-1(%ecx),%ecx
+	leal	1(%edx),%edx
+	jmp	L010lot
+L013aligned:
+	movl	%eax,(%edx)
+	leal	-4(%ecx),%ecx
+	testl	$-4,%ecx
+	leal	4(%edx),%edx
+	jnz	L013aligned
+	cmpl	$0,%ecx
+	jne	L012little
+	ret
+.comm	_OPENSSL_ia32cap_P,4
+.mod_init_func
+.align 2
+.long   _OPENSSL_cpuid_setup

deps/openssl/asm/x86-win32-masm/bf/bf-686.asm

@@ -0,0 +1,907 @@
+TITLE	bf-686.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+ALIGN	16
+_BF_encrypt	PROC PUBLIC
+$L_BF_encrypt_begin::
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	;
+
+	; Load the 2 words
+	mov	eax,DWORD PTR 20[esp]
+	mov	ecx,DWORD PTR [eax]
+	mov	edx,DWORD PTR 4[eax]
+	;
+
+	; P pointer, s and enc flag
+	mov	edi,DWORD PTR 24[esp]
+	xor	eax,eax
+	xor	ebx,ebx
+	xor	ecx,DWORD PTR [edi]
+	;
+
+	; Round 0
+	ror	ecx,16
+	mov	esi,DWORD PTR 4[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 1
+	ror	edx,16
+	mov	esi,DWORD PTR 8[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 2
+	ror	ecx,16
+	mov	esi,DWORD PTR 12[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 3
+	ror	edx,16
+	mov	esi,DWORD PTR 16[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 4
+	ror	ecx,16
+	mov	esi,DWORD PTR 20[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 5
+	ror	edx,16
+	mov	esi,DWORD PTR 24[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 6
+	ror	ecx,16
+	mov	esi,DWORD PTR 28[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 7
+	ror	edx,16
+	mov	esi,DWORD PTR 32[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 8
+	ror	ecx,16
+	mov	esi,DWORD PTR 36[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 9
+	ror	edx,16
+	mov	esi,DWORD PTR 40[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 10
+	ror	ecx,16
+	mov	esi,DWORD PTR 44[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 11
+	ror	edx,16
+	mov	esi,DWORD PTR 48[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 12
+	ror	ecx,16
+	mov	esi,DWORD PTR 52[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 13
+	ror	edx,16
+	mov	esi,DWORD PTR 56[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 14
+	ror	ecx,16
+	mov	esi,DWORD PTR 60[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 15
+	ror	edx,16
+	mov	esi,DWORD PTR 64[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	xor	edx,DWORD PTR 68[edi]
+	mov	eax,DWORD PTR 20[esp]
+	mov	DWORD PTR [eax],edx
+	mov	DWORD PTR 4[eax],ecx
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_BF_encrypt ENDP
+ALIGN	16
+_BF_decrypt	PROC PUBLIC
+$L_BF_decrypt_begin::
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	;
+
+	; Load the 2 words
+	mov	eax,DWORD PTR 20[esp]
+	mov	ecx,DWORD PTR [eax]
+	mov	edx,DWORD PTR 4[eax]
+	;
+
+	; P pointer, s and enc flag
+	mov	edi,DWORD PTR 24[esp]
+	xor	eax,eax
+	xor	ebx,ebx
+	xor	ecx,DWORD PTR 68[edi]
+	;
+
+	; Round 16
+	ror	ecx,16
+	mov	esi,DWORD PTR 64[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 15
+	ror	edx,16
+	mov	esi,DWORD PTR 60[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 14
+	ror	ecx,16
+	mov	esi,DWORD PTR 56[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 13
+	ror	edx,16
+	mov	esi,DWORD PTR 52[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 12
+	ror	ecx,16
+	mov	esi,DWORD PTR 48[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 11
+	ror	edx,16
+	mov	esi,DWORD PTR 44[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 10
+	ror	ecx,16
+	mov	esi,DWORD PTR 40[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 9
+	ror	edx,16
+	mov	esi,DWORD PTR 36[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 8
+	ror	ecx,16
+	mov	esi,DWORD PTR 32[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 7
+	ror	edx,16
+	mov	esi,DWORD PTR 28[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 6
+	ror	ecx,16
+	mov	esi,DWORD PTR 24[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 5
+	ror	edx,16
+	mov	esi,DWORD PTR 20[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 4
+	ror	ecx,16
+	mov	esi,DWORD PTR 16[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 3
+	ror	edx,16
+	mov	esi,DWORD PTR 12[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	;
+
+	; Round 2
+	ror	ecx,16
+	mov	esi,DWORD PTR 8[edi]
+	mov	al,ch
+	mov	bl,cl
+	ror	ecx,16
+	xor	edx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,ch
+	mov	bl,cl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	edx,esi
+	;
+
+	; Round 1
+	ror	edx,16
+	mov	esi,DWORD PTR 4[edi]
+	mov	al,dh
+	mov	bl,dl
+	ror	edx,16
+	xor	ecx,esi
+	mov	esi,DWORD PTR 72[eax*4+edi]
+	mov	ebp,DWORD PTR 1096[ebx*4+edi]
+	mov	al,dh
+	mov	bl,dl
+	add	esi,ebp
+	mov	eax,DWORD PTR 2120[eax*4+edi]
+	xor	esi,eax
+	mov	ebp,DWORD PTR 3144[ebx*4+edi]
+	add	esi,ebp
+	xor	eax,eax
+	xor	ecx,esi
+	xor	edx,DWORD PTR [edi]
+	mov	eax,DWORD PTR 20[esp]
+	mov	DWORD PTR [eax],edx
+	mov	DWORD PTR 4[eax],ecx
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_BF_decrypt ENDP
+ALIGN	16
+_BF_cbc_encrypt	PROC PUBLIC
+$L_BF_cbc_encrypt_begin::
+	;
+
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	mov	ebp,DWORD PTR 28[esp]
+	; getting iv ptr from parameter 4
+	mov	ebx,DWORD PTR 36[esp]
+	mov	esi,DWORD PTR [ebx]
+	mov	edi,DWORD PTR 4[ebx]
+	push	edi
+	push	esi
+	push	edi
+	push	esi
+	mov	ebx,esp
+	mov	esi,DWORD PTR 36[esp]
+	mov	edi,DWORD PTR 40[esp]
+	; getting encrypt flag from parameter 5
+	mov	ecx,DWORD PTR 56[esp]
+	; get and push parameter 3
+	mov	eax,DWORD PTR 48[esp]
+	push	eax
+	push	ebx
+	cmp	ecx,0
+	jz	$L000decrypt
+	and	ebp,4294967288
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	jz	$L001encrypt_finish
+$L002encrypt_loop:
+	mov	ecx,DWORD PTR [esi]
+	mov	edx,DWORD PTR 4[esi]
+	xor	eax,ecx
+	xor	ebx,edx
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_BF_encrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR [edi],eax
+	mov	DWORD PTR 4[edi],ebx
+	add	esi,8
+	add	edi,8
+	sub	ebp,8
+	jnz	$L002encrypt_loop
+$L001encrypt_finish:
+	mov	ebp,DWORD PTR 52[esp]
+	and	ebp,7
+	jz	$L003finish
+	call	$L004PIC_point
+$L004PIC_point:
+	pop	edx
+	lea	ecx,DWORD PTR ($L005cbc_enc_jmp_table-$L004PIC_point)[edx]
+	mov	ebp,DWORD PTR [ebp*4+ecx]
+	add	ebp,edx
+	xor	ecx,ecx
+	xor	edx,edx
+	jmp	ebp
+$L006ej7:
+	mov	dh,BYTE PTR 6[esi]
+	shl	edx,8
+$L007ej6:
+	mov	dh,BYTE PTR 5[esi]
+$L008ej5:
+	mov	dl,BYTE PTR 4[esi]
+$L009ej4:
+	mov	ecx,DWORD PTR [esi]
+	jmp	$L010ejend
+$L011ej3:
+	mov	ch,BYTE PTR 2[esi]
+	shl	ecx,8
+$L012ej2:
+	mov	ch,BYTE PTR 1[esi]
+$L013ej1:
+	mov	cl,BYTE PTR [esi]
+$L010ejend:
+	xor	eax,ecx
+	xor	ebx,edx
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_BF_encrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR [edi],eax
+	mov	DWORD PTR 4[edi],ebx
+	jmp	$L003finish
+$L000decrypt:
+	and	ebp,4294967288
+	mov	eax,DWORD PTR 16[esp]
+	mov	ebx,DWORD PTR 20[esp]
+	jz	$L014decrypt_finish
+$L015decrypt_loop:
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_BF_decrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	ecx,DWORD PTR 16[esp]
+	mov	edx,DWORD PTR 20[esp]
+	xor	ecx,eax
+	xor	edx,ebx
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	mov	DWORD PTR [edi],ecx
+	mov	DWORD PTR 4[edi],edx
+	mov	DWORD PTR 16[esp],eax
+	mov	DWORD PTR 20[esp],ebx
+	add	esi,8
+	add	edi,8
+	sub	ebp,8
+	jnz	$L015decrypt_loop
+$L014decrypt_finish:
+	mov	ebp,DWORD PTR 52[esp]
+	and	ebp,7
+	jz	$L003finish
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_BF_decrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	ecx,DWORD PTR 16[esp]
+	mov	edx,DWORD PTR 20[esp]
+	xor	ecx,eax
+	xor	edx,ebx
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+$L016dj7:
+	ror	edx,16
+	mov	BYTE PTR 6[edi],dl
+	shr	edx,16
+$L017dj6:
+	mov	BYTE PTR 5[edi],dh
+$L018dj5:
+	mov	BYTE PTR 4[edi],dl
+$L019dj4:
+	mov	DWORD PTR [edi],ecx
+	jmp	$L020djend
+$L021dj3:
+	ror	ecx,16
+	mov	BYTE PTR 2[edi],cl
+	shl	ecx,16
+$L022dj2:
+	mov	BYTE PTR 1[esi],ch
+$L023dj1:
+	mov	BYTE PTR [esi],cl
+$L020djend:
+	jmp	$L003finish
+$L003finish:
+	mov	ecx,DWORD PTR 60[esp]
+	add	esp,24
+	mov	DWORD PTR [ecx],eax
+	mov	DWORD PTR 4[ecx],ebx
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+ALIGN	64
+$L005cbc_enc_jmp_table:
+DD	0
+DD	$L013ej1-$L004PIC_point
+DD	$L012ej2-$L004PIC_point
+DD	$L011ej3-$L004PIC_point
+DD	$L009ej4-$L004PIC_point
+DD	$L008ej5-$L004PIC_point
+DD	$L007ej6-$L004PIC_point
+DD	$L006ej7-$L004PIC_point
+ALIGN	64
+_BF_cbc_encrypt ENDP
+.text$	ENDS
+END

deps/openssl/asm/x86-win32-masm/bn/x86-mont.asm

@@ -0,0 +1,348 @@
+TITLE	../openssl/crypto/bn/asm/x86-mont.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+ALIGN	16
+_bn_mul_mont	PROC PUBLIC
+$L_bn_mul_mont_begin::
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	xor	eax,eax
+	mov	edi,DWORD PTR 40[esp]
+	cmp	edi,4
+	jl	$L000just_leave
+	lea	esi,DWORD PTR 20[esp]
+	lea	edx,DWORD PTR 24[esp]
+	mov	ebp,esp
+	add	edi,2
+	neg	edi
+	lea	esp,DWORD PTR [edi*4+esp-32]
+	neg	edi
+	mov	eax,esp
+	sub	eax,edx
+	and	eax,2047
+	sub	esp,eax
+	xor	edx,esp
+	and	edx,2048
+	xor	edx,2048
+	sub	esp,edx
+	and	esp,-64
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	mov	ecx,DWORD PTR 8[esi]
+	mov	edx,DWORD PTR 12[esi]
+	mov	esi,DWORD PTR 16[esi]
+	mov	esi,DWORD PTR [esi]
+	mov	DWORD PTR 4[esp],eax
+	mov	DWORD PTR 8[esp],ebx
+	mov	DWORD PTR 12[esp],ecx
+	mov	DWORD PTR 16[esp],edx
+	mov	DWORD PTR 20[esp],esi
+	lea	ebx,DWORD PTR [edi-3]
+	mov	DWORD PTR 24[esp],ebp
+	mov	esi,DWORD PTR 8[esp]
+	lea	ebp,DWORD PTR 1[ebx]
+	mov	edi,DWORD PTR 12[esp]
+	xor	ecx,ecx
+	mov	edx,esi
+	and	ebp,1
+	sub	edx,edi
+	lea	eax,DWORD PTR 4[ebx*4+edi]
+	or	ebp,edx
+	mov	edi,DWORD PTR [edi]
+	jz	$L001bn_sqr_mont
+	mov	DWORD PTR 28[esp],eax
+	mov	eax,DWORD PTR [esi]
+	xor	edx,edx
+ALIGN	16
+$L002mull:
+	mov	ebp,edx
+	mul	edi
+	add	ebp,eax
+	lea	ecx,DWORD PTR 1[ecx]
+	adc	edx,0
+	mov	eax,DWORD PTR [ecx*4+esi]
+	cmp	ecx,ebx
+	mov	DWORD PTR 28[ecx*4+esp],ebp
+	jl	$L002mull
+	mov	ebp,edx
+	mul	edi
+	mov	edi,DWORD PTR 20[esp]
+	add	eax,ebp
+	mov	esi,DWORD PTR 16[esp]
+	adc	edx,0
+	imul	edi,DWORD PTR 32[esp]
+	mov	DWORD PTR 32[ebx*4+esp],eax
+	xor	ecx,ecx
+	mov	DWORD PTR 36[ebx*4+esp],edx
+	mov	DWORD PTR 40[ebx*4+esp],ecx
+	mov	eax,DWORD PTR [esi]
+	mul	edi
+	add	eax,DWORD PTR 32[esp]
+	mov	eax,DWORD PTR 4[esi]
+	adc	edx,0
+	inc	ecx
+	jmp	$L0032ndmadd
+ALIGN	16
+$L0041stmadd:
+	mov	ebp,edx
+	mul	edi
+	add	ebp,DWORD PTR 32[ecx*4+esp]
+	lea	ecx,DWORD PTR 1[ecx]
+	adc	edx,0
+	add	ebp,eax
+	mov	eax,DWORD PTR [ecx*4+esi]
+	adc	edx,0
+	cmp	ecx,ebx
+	mov	DWORD PTR 28[ecx*4+esp],ebp
+	jl	$L0041stmadd
+	mov	ebp,edx
+	mul	edi
+	add	eax,DWORD PTR 32[ebx*4+esp]
+	mov	edi,DWORD PTR 20[esp]
+	adc	edx,0
+	mov	esi,DWORD PTR 16[esp]
+	add	ebp,eax
+	adc	edx,0
+	imul	edi,DWORD PTR 32[esp]
+	xor	ecx,ecx
+	add	edx,DWORD PTR 36[ebx*4+esp]
+	mov	DWORD PTR 32[ebx*4+esp],ebp
+	adc	ecx,0
+	mov	eax,DWORD PTR [esi]
+	mov	DWORD PTR 36[ebx*4+esp],edx
+	mov	DWORD PTR 40[ebx*4+esp],ecx
+	mul	edi
+	add	eax,DWORD PTR 32[esp]
+	mov	eax,DWORD PTR 4[esi]
+	adc	edx,0
+	mov	ecx,1
+ALIGN	16
+$L0032ndmadd:
+	mov	ebp,edx
+	mul	edi
+	add	ebp,DWORD PTR 32[ecx*4+esp]
+	lea	ecx,DWORD PTR 1[ecx]
+	adc	edx,0
+	add	ebp,eax
+	mov	eax,DWORD PTR [ecx*4+esi]
+	adc	edx,0
+	cmp	ecx,ebx
+	mov	DWORD PTR 24[ecx*4+esp],ebp
+	jl	$L0032ndmadd
+	mov	ebp,edx
+	mul	edi
+	add	ebp,DWORD PTR 32[ebx*4+esp]
+	adc	edx,0
+	add	ebp,eax
+	adc	edx,0
+	mov	DWORD PTR 28[ebx*4+esp],ebp
+	xor	eax,eax
+	mov	ecx,DWORD PTR 12[esp]
+	add	edx,DWORD PTR 36[ebx*4+esp]
+	adc	eax,DWORD PTR 40[ebx*4+esp]
+	lea	ecx,DWORD PTR 4[ecx]
+	mov	DWORD PTR 32[ebx*4+esp],edx
+	cmp	ecx,DWORD PTR 28[esp]
+	mov	DWORD PTR 36[ebx*4+esp],eax
+	je	$L005common_tail
+	mov	edi,DWORD PTR [ecx]
+	mov	esi,DWORD PTR 8[esp]
+	mov	DWORD PTR 12[esp],ecx
+	xor	ecx,ecx
+	xor	edx,edx
+	mov	eax,DWORD PTR [esi]
+	jmp	$L0041stmadd
+ALIGN	16
+$L001bn_sqr_mont:
+	mov	DWORD PTR [esp],ebx
+	mov	DWORD PTR 12[esp],ecx
+	mov	eax,edi
+	mul	edi
+	mov	DWORD PTR 32[esp],eax
+	mov	ebx,edx
+	shr	edx,1
+	and	ebx,1
+	inc	ecx
+ALIGN	16
+$L006sqr:
+	mov	eax,DWORD PTR [ecx*4+esi]
+	mov	ebp,edx
+	mul	edi
+	add	eax,ebp
+	lea	ecx,DWORD PTR 1[ecx]
+	adc	edx,0
+	lea	ebp,DWORD PTR [eax*2+ebx]
+	shr	eax,31
+	cmp	ecx,DWORD PTR [esp]
+	mov	ebx,eax
+	mov	DWORD PTR 28[ecx*4+esp],ebp
+	jl	$L006sqr
+	mov	eax,DWORD PTR [ecx*4+esi]
+	mov	ebp,edx
+	mul	edi
+	add	eax,ebp
+	mov	edi,DWORD PTR 20[esp]
+	adc	edx,0
+	mov	esi,DWORD PTR 16[esp]
+	lea	ebp,DWORD PTR [eax*2+ebx]
+	imul	edi,DWORD PTR 32[esp]
+	shr	eax,31
+	mov	DWORD PTR 32[ecx*4+esp],ebp
+	lea	ebp,DWORD PTR [edx*2+eax]
+	mov	eax,DWORD PTR [esi]
+	shr	edx,31
+	mov	DWORD PTR 36[ecx*4+esp],ebp
+	mov	DWORD PTR 40[ecx*4+esp],edx
+	mul	edi
+	add	eax,DWORD PTR 32[esp]
+	mov	ebx,ecx
+	adc	edx,0
+	mov	eax,DWORD PTR 4[esi]
+	mov	ecx,1
+ALIGN	16
+$L0073rdmadd:
+	mov	ebp,edx
+	mul	edi
+	add	ebp,DWORD PTR 32[ecx*4+esp]
+	adc	edx,0
+	add	ebp,eax
+	mov	eax,DWORD PTR 4[ecx*4+esi]
+	adc	edx,0
+	mov	DWORD PTR 28[ecx*4+esp],ebp
+	mov	ebp,edx
+	mul	edi
+	add	ebp,DWORD PTR 36[ecx*4+esp]
+	lea	ecx,DWORD PTR 2[ecx]
+	adc	edx,0
+	add	ebp,eax
+	mov	eax,DWORD PTR [ecx*4+esi]
+	adc	edx,0
+	cmp	ecx,ebx
+	mov	DWORD PTR 24[ecx*4+esp],ebp
+	jl	$L0073rdmadd
+	mov	ebp,edx
+	mul	edi
+	add	ebp,DWORD PTR 32[ebx*4+esp]
+	adc	edx,0
+	add	ebp,eax
+	adc	edx,0
+	mov	DWORD PTR 28[ebx*4+esp],ebp
+	mov	ecx,DWORD PTR 12[esp]
+	xor	eax,eax
+	mov	esi,DWORD PTR 8[esp]
+	add	edx,DWORD PTR 36[ebx*4+esp]
+	adc	eax,DWORD PTR 40[ebx*4+esp]
+	mov	DWORD PTR 32[ebx*4+esp],edx
+	cmp	ecx,ebx
+	mov	DWORD PTR 36[ebx*4+esp],eax
+	je	$L005common_tail
+	mov	edi,DWORD PTR 4[ecx*4+esi]
+	lea	ecx,DWORD PTR 1[ecx]
+	mov	eax,edi
+	mov	DWORD PTR 12[esp],ecx
+	mul	edi
+	add	eax,DWORD PTR 32[ecx*4+esp]
+	adc	edx,0
+	mov	DWORD PTR 32[ecx*4+esp],eax
+	xor	ebp,ebp
+	cmp	ecx,ebx
+	lea	ecx,DWORD PTR 1[ecx]
+	je	$L008sqrlast
+	mov	ebx,edx
+	shr	edx,1
+	and	ebx,1
+ALIGN	16
+$L009sqradd:
+	mov	eax,DWORD PTR [ecx*4+esi]
+	mov	ebp,edx
+	mul	edi
+	add	eax,ebp
+	lea	ebp,DWORD PTR [eax*1+eax]
+	adc	edx,0
+	shr	eax,31
+	add	ebp,DWORD PTR 32[ecx*4+esp]
+	lea	ecx,DWORD PTR 1[ecx]
+	adc	eax,0
+	add	ebp,ebx
+	adc	eax,0
+	cmp	ecx,DWORD PTR [esp]
+	mov	DWORD PTR 28[ecx*4+esp],ebp
+	mov	ebx,eax
+	jle	$L009sqradd
+	mov	ebp,edx
+	add	edx,edx
+	shr	ebp,31
+	add	edx,ebx
+	adc	ebp,0
+$L008sqrlast:
+	mov	edi,DWORD PTR 20[esp]
+	mov	esi,DWORD PTR 16[esp]
+	imul	edi,DWORD PTR 32[esp]
+	add	edx,DWORD PTR 32[ecx*4+esp]
+	mov	eax,DWORD PTR [esi]
+	adc	ebp,0
+	mov	DWORD PTR 32[ecx*4+esp],edx
+	mov	DWORD PTR 36[ecx*4+esp],ebp
+	mul	edi
+	add	eax,DWORD PTR 32[esp]
+	lea	ebx,DWORD PTR [ecx-1]
+	adc	edx,0
+	mov	ecx,1
+	mov	eax,DWORD PTR 4[esi]
+	jmp	$L0073rdmadd
+ALIGN	16
+$L005common_tail:
+	mov	ebp,DWORD PTR 16[esp]
+	mov	edi,DWORD PTR 4[esp]
+	lea	esi,DWORD PTR 32[esp]
+	mov	eax,DWORD PTR [esi]
+	mov	ecx,ebx
+	xor	edx,edx
+ALIGN	16
+$L010sub:
+	sbb	eax,DWORD PTR [edx*4+ebp]
+	mov	DWORD PTR [edx*4+edi],eax
+	dec	ecx
+	mov	eax,DWORD PTR 4[edx*4+esi]
+	lea	edx,DWORD PTR 1[edx]
+	jge	$L010sub
+	sbb	eax,0
+	and	esi,eax
+	not	eax
+	mov	ebp,edi
+	and	ebp,eax
+	or	esi,ebp
+ALIGN	16
+$L011copy:
+	mov	eax,DWORD PTR [ebx*4+esi]
+	mov	DWORD PTR [ebx*4+edi],eax
+	mov	DWORD PTR 32[ebx*4+esp],ecx
+	dec	ebx
+	jge	$L011copy
+	mov	esp,DWORD PTR 24[esp]
+	mov	eax,1
+$L000just_leave:
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_bn_mul_mont ENDP
+DB	77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
+DB	112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
+DB	54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
+DB	32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
+DB	111,114,103,62,0
+.text$	ENDS
+END

deps/openssl/asm/x86-win32-masm/cast/cast-586.asm

@@ -0,0 +1,950 @@
+TITLE	cast-586.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+EXTERN	_CAST_S_table0:NEAR
+EXTERN	_CAST_S_table1:NEAR
+EXTERN	_CAST_S_table2:NEAR
+EXTERN	_CAST_S_table3:NEAR
+ALIGN	16
+_CAST_encrypt	PROC PUBLIC
+$L_CAST_encrypt_begin::
+	;
+
+	push	ebp
+	push	ebx
+	mov	ebx,DWORD PTR 12[esp]
+	mov	ebp,DWORD PTR 16[esp]
+	push	esi
+	push	edi
+	; Load the 2 words
+	mov	edi,DWORD PTR [ebx]
+	mov	esi,DWORD PTR 4[ebx]
+	; Get short key flag
+	mov	eax,DWORD PTR 128[ebp]
+	push	eax
+	xor	eax,eax
+	; round 0
+	mov	edx,DWORD PTR [ebp]
+	mov	ecx,DWORD PTR 4[ebp]
+	add	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	edi,ecx
+	; round 1
+	mov	edx,DWORD PTR 8[ebp]
+	mov	ecx,DWORD PTR 12[ebp]
+	xor	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	esi,ecx
+	; round 2
+	mov	edx,DWORD PTR 16[ebp]
+	mov	ecx,DWORD PTR 20[ebp]
+	sub	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	edi,ecx
+	; round 3
+	mov	edx,DWORD PTR 24[ebp]
+	mov	ecx,DWORD PTR 28[ebp]
+	add	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	esi,ecx
+	; round 4
+	mov	edx,DWORD PTR 32[ebp]
+	mov	ecx,DWORD PTR 36[ebp]
+	xor	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	edi,ecx
+	; round 5
+	mov	edx,DWORD PTR 40[ebp]
+	mov	ecx,DWORD PTR 44[ebp]
+	sub	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	esi,ecx
+	; round 6
+	mov	edx,DWORD PTR 48[ebp]
+	mov	ecx,DWORD PTR 52[ebp]
+	add	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	edi,ecx
+	; round 7
+	mov	edx,DWORD PTR 56[ebp]
+	mov	ecx,DWORD PTR 60[ebp]
+	xor	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	esi,ecx
+	; round 8
+	mov	edx,DWORD PTR 64[ebp]
+	mov	ecx,DWORD PTR 68[ebp]
+	sub	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	edi,ecx
+	; round 9
+	mov	edx,DWORD PTR 72[ebp]
+	mov	ecx,DWORD PTR 76[ebp]
+	add	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	esi,ecx
+	; round 10
+	mov	edx,DWORD PTR 80[ebp]
+	mov	ecx,DWORD PTR 84[ebp]
+	xor	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	edi,ecx
+	; round 11
+	mov	edx,DWORD PTR 88[ebp]
+	mov	ecx,DWORD PTR 92[ebp]
+	sub	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	esi,ecx
+	; test short key flag
+	pop	edx
+	or	edx,edx
+	jnz	$L000cast_enc_done
+	; round 12
+	mov	edx,DWORD PTR 96[ebp]
+	mov	ecx,DWORD PTR 100[ebp]
+	add	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	edi,ecx
+	; round 13
+	mov	edx,DWORD PTR 104[ebp]
+	mov	ecx,DWORD PTR 108[ebp]
+	xor	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	esi,ecx
+	; round 14
+	mov	edx,DWORD PTR 112[ebp]
+	mov	ecx,DWORD PTR 116[ebp]
+	sub	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	edi,ecx
+	; round 15
+	mov	edx,DWORD PTR 120[ebp]
+	mov	ecx,DWORD PTR 124[ebp]
+	add	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	esi,ecx
+$L000cast_enc_done:
+	nop
+	mov	eax,DWORD PTR 20[esp]
+	mov	DWORD PTR 4[eax],edi
+	mov	DWORD PTR [eax],esi
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_CAST_encrypt ENDP
+EXTERN	_CAST_S_table0:NEAR
+EXTERN	_CAST_S_table1:NEAR
+EXTERN	_CAST_S_table2:NEAR
+EXTERN	_CAST_S_table3:NEAR
+ALIGN	16
+_CAST_decrypt	PROC PUBLIC
+$L_CAST_decrypt_begin::
+	;
+
+	push	ebp
+	push	ebx
+	mov	ebx,DWORD PTR 12[esp]
+	mov	ebp,DWORD PTR 16[esp]
+	push	esi
+	push	edi
+	; Load the 2 words
+	mov	edi,DWORD PTR [ebx]
+	mov	esi,DWORD PTR 4[ebx]
+	; Get short key flag
+	mov	eax,DWORD PTR 128[ebp]
+	or	eax,eax
+	jnz	$L001cast_dec_skip
+	xor	eax,eax
+	; round 15
+	mov	edx,DWORD PTR 120[ebp]
+	mov	ecx,DWORD PTR 124[ebp]
+	add	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	edi,ecx
+	; round 14
+	mov	edx,DWORD PTR 112[ebp]
+	mov	ecx,DWORD PTR 116[ebp]
+	sub	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	esi,ecx
+	; round 13
+	mov	edx,DWORD PTR 104[ebp]
+	mov	ecx,DWORD PTR 108[ebp]
+	xor	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	edi,ecx
+	; round 12
+	mov	edx,DWORD PTR 96[ebp]
+	mov	ecx,DWORD PTR 100[ebp]
+	add	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	esi,ecx
+$L001cast_dec_skip:
+	; round 11
+	mov	edx,DWORD PTR 88[ebp]
+	mov	ecx,DWORD PTR 92[ebp]
+	sub	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	edi,ecx
+	; round 10
+	mov	edx,DWORD PTR 80[ebp]
+	mov	ecx,DWORD PTR 84[ebp]
+	xor	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	esi,ecx
+	; round 9
+	mov	edx,DWORD PTR 72[ebp]
+	mov	ecx,DWORD PTR 76[ebp]
+	add	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	edi,ecx
+	; round 8
+	mov	edx,DWORD PTR 64[ebp]
+	mov	ecx,DWORD PTR 68[ebp]
+	sub	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	esi,ecx
+	; round 7
+	mov	edx,DWORD PTR 56[ebp]
+	mov	ecx,DWORD PTR 60[ebp]
+	xor	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	edi,ecx
+	; round 6
+	mov	edx,DWORD PTR 48[ebp]
+	mov	ecx,DWORD PTR 52[ebp]
+	add	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	esi,ecx
+	; round 5
+	mov	edx,DWORD PTR 40[ebp]
+	mov	ecx,DWORD PTR 44[ebp]
+	sub	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	edi,ecx
+	; round 4
+	mov	edx,DWORD PTR 32[ebp]
+	mov	ecx,DWORD PTR 36[ebp]
+	xor	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	esi,ecx
+	; round 3
+	mov	edx,DWORD PTR 24[ebp]
+	mov	ecx,DWORD PTR 28[ebp]
+	add	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	edi,ecx
+	; round 2
+	mov	edx,DWORD PTR 16[ebp]
+	mov	ecx,DWORD PTR 20[ebp]
+	sub	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	sub	ecx,ebx
+	xor	esi,ecx
+	; round 1
+	mov	edx,DWORD PTR 8[ebp]
+	mov	ecx,DWORD PTR 12[ebp]
+	xor	edx,esi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	add	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	xor	ecx,ebx
+	xor	edi,ecx
+	; round 0
+	mov	edx,DWORD PTR [ebp]
+	mov	ecx,DWORD PTR 4[ebp]
+	add	edx,edi
+	rol	edx,cl
+	mov	ebx,edx
+	xor	ecx,ecx
+	mov	cl,dh
+	and	ebx,255
+	shr	edx,16
+	xor	eax,eax
+	mov	al,dh
+	and	edx,255
+	mov	ecx,DWORD PTR _CAST_S_table0[ecx*4]
+	mov	ebx,DWORD PTR _CAST_S_table1[ebx*4]
+	xor	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table2[eax*4]
+	sub	ecx,ebx
+	mov	ebx,DWORD PTR _CAST_S_table3[edx*4]
+	add	ecx,ebx
+	xor	esi,ecx
+	nop
+	mov	eax,DWORD PTR 20[esp]
+	mov	DWORD PTR 4[eax],edi
+	mov	DWORD PTR [eax],esi
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_CAST_decrypt ENDP
+ALIGN	16
+_CAST_cbc_encrypt	PROC PUBLIC
+$L_CAST_cbc_encrypt_begin::
+	;
+
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	mov	ebp,DWORD PTR 28[esp]
+	; getting iv ptr from parameter 4
+	mov	ebx,DWORD PTR 36[esp]
+	mov	esi,DWORD PTR [ebx]
+	mov	edi,DWORD PTR 4[ebx]
+	push	edi
+	push	esi
+	push	edi
+	push	esi
+	mov	ebx,esp
+	mov	esi,DWORD PTR 36[esp]
+	mov	edi,DWORD PTR 40[esp]
+	; getting encrypt flag from parameter 5
+	mov	ecx,DWORD PTR 56[esp]
+	; get and push parameter 3
+	mov	eax,DWORD PTR 48[esp]
+	push	eax
+	push	ebx
+	cmp	ecx,0
+	jz	$L002decrypt
+	and	ebp,4294967288
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	jz	$L003encrypt_finish
+$L004encrypt_loop:
+	mov	ecx,DWORD PTR [esi]
+	mov	edx,DWORD PTR 4[esi]
+	xor	eax,ecx
+	xor	ebx,edx
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_CAST_encrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR [edi],eax
+	mov	DWORD PTR 4[edi],ebx
+	add	esi,8
+	add	edi,8
+	sub	ebp,8
+	jnz	$L004encrypt_loop
+$L003encrypt_finish:
+	mov	ebp,DWORD PTR 52[esp]
+	and	ebp,7
+	jz	$L005finish
+	call	$L006PIC_point
+$L006PIC_point:
+	pop	edx
+	lea	ecx,DWORD PTR ($L007cbc_enc_jmp_table-$L006PIC_point)[edx]
+	mov	ebp,DWORD PTR [ebp*4+ecx]
+	add	ebp,edx
+	xor	ecx,ecx
+	xor	edx,edx
+	jmp	ebp
+$L008ej7:
+	mov	dh,BYTE PTR 6[esi]
+	shl	edx,8
+$L009ej6:
+	mov	dh,BYTE PTR 5[esi]
+$L010ej5:
+	mov	dl,BYTE PTR 4[esi]
+$L011ej4:
+	mov	ecx,DWORD PTR [esi]
+	jmp	$L012ejend
+$L013ej3:
+	mov	ch,BYTE PTR 2[esi]
+	shl	ecx,8
+$L014ej2:
+	mov	ch,BYTE PTR 1[esi]
+$L015ej1:
+	mov	cl,BYTE PTR [esi]
+$L012ejend:
+	xor	eax,ecx
+	xor	ebx,edx
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_CAST_encrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR [edi],eax
+	mov	DWORD PTR 4[edi],ebx
+	jmp	$L005finish
+$L002decrypt:
+	and	ebp,4294967288
+	mov	eax,DWORD PTR 16[esp]
+	mov	ebx,DWORD PTR 20[esp]
+	jz	$L016decrypt_finish
+$L017decrypt_loop:
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_CAST_decrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	ecx,DWORD PTR 16[esp]
+	mov	edx,DWORD PTR 20[esp]
+	xor	ecx,eax
+	xor	edx,ebx
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	mov	DWORD PTR [edi],ecx
+	mov	DWORD PTR 4[edi],edx
+	mov	DWORD PTR 16[esp],eax
+	mov	DWORD PTR 20[esp],ebx
+	add	esi,8
+	add	edi,8
+	sub	ebp,8
+	jnz	$L017decrypt_loop
+$L016decrypt_finish:
+	mov	ebp,DWORD PTR 52[esp]
+	and	ebp,7
+	jz	$L005finish
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	bswap	eax
+	bswap	ebx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_CAST_decrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	bswap	eax
+	bswap	ebx
+	mov	ecx,DWORD PTR 16[esp]
+	mov	edx,DWORD PTR 20[esp]
+	xor	ecx,eax
+	xor	edx,ebx
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+$L018dj7:
+	ror	edx,16
+	mov	BYTE PTR 6[edi],dl
+	shr	edx,16
+$L019dj6:
+	mov	BYTE PTR 5[edi],dh
+$L020dj5:
+	mov	BYTE PTR 4[edi],dl
+$L021dj4:
+	mov	DWORD PTR [edi],ecx
+	jmp	$L022djend
+$L023dj3:
+	ror	ecx,16
+	mov	BYTE PTR 2[edi],cl
+	shl	ecx,16
+$L024dj2:
+	mov	BYTE PTR 1[esi],ch
+$L025dj1:
+	mov	BYTE PTR [esi],cl
+$L022djend:
+	jmp	$L005finish
+$L005finish:
+	mov	ecx,DWORD PTR 60[esp]
+	add	esp,24
+	mov	DWORD PTR [ecx],eax
+	mov	DWORD PTR 4[ecx],ebx
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+ALIGN	64
+$L007cbc_enc_jmp_table:
+DD	0
+DD	$L015ej1-$L006PIC_point
+DD	$L014ej2-$L006PIC_point
+DD	$L013ej3-$L006PIC_point
+DD	$L011ej4-$L006PIC_point
+DD	$L010ej5-$L006PIC_point
+DD	$L009ej6-$L006PIC_point
+DD	$L008ej7-$L006PIC_point
+ALIGN	64
+_CAST_cbc_encrypt ENDP
+.text$	ENDS
+END

deps/openssl/asm/x86-win32-masm/des/crypt586.asm

@@ -0,0 +1,909 @@
+TITLE	crypt586.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+EXTERN	_DES_SPtrans:NEAR
+ALIGN	16
+_fcrypt_body	PROC PUBLIC
+$L_fcrypt_body_begin::
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	;
+
+	; Load the 2 words
+	xor	edi,edi
+	xor	esi,esi
+	lea	edx,DWORD PTR _DES_SPtrans
+	push	edx
+	mov	ebp,DWORD PTR 28[esp]
+	push	25
+$L000start:
+	;
+
+	; Round 0
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,esi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,esi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR [ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 4[ebp]
+	xor	eax,esi
+	xor	edx,esi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	edi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	edi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	edi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	edi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 1
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,edi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,edi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 8[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 12[ebp]
+	xor	eax,edi
+	xor	edx,edi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	esi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	esi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	esi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	esi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	esi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 2
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,esi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,esi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 16[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 20[ebp]
+	xor	eax,esi
+	xor	edx,esi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	edi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	edi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	edi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	edi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 3
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,edi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,edi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 24[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 28[ebp]
+	xor	eax,edi
+	xor	edx,edi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	esi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	esi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	esi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	esi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	esi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 4
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,esi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,esi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 32[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 36[ebp]
+	xor	eax,esi
+	xor	edx,esi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	edi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	edi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	edi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	edi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 5
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,edi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,edi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 40[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 44[ebp]
+	xor	eax,edi
+	xor	edx,edi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	esi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	esi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	esi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	esi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	esi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 6
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,esi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,esi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 48[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 52[ebp]
+	xor	eax,esi
+	xor	edx,esi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	edi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	edi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	edi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	edi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 7
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,edi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,edi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 56[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 60[ebp]
+	xor	eax,edi
+	xor	edx,edi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	esi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	esi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	esi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	esi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	esi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 8
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,esi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,esi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 64[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 68[ebp]
+	xor	eax,esi
+	xor	edx,esi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	edi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	edi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	edi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	edi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 9
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,edi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,edi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 72[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 76[ebp]
+	xor	eax,edi
+	xor	edx,edi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	esi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	esi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	esi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	esi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	esi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 10
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,esi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,esi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 80[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 84[ebp]
+	xor	eax,esi
+	xor	edx,esi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	edi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	edi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	edi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	edi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 11
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,edi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,edi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 88[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 92[ebp]
+	xor	eax,edi
+	xor	edx,edi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	esi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	esi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	esi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	esi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	esi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 12
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,esi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,esi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 96[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 100[ebp]
+	xor	eax,esi
+	xor	edx,esi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	edi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	edi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	edi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	edi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 13
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,edi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,edi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 104[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 108[ebp]
+	xor	eax,edi
+	xor	edx,edi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	esi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	esi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	esi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	esi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	esi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 14
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,esi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,esi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 112[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 116[ebp]
+	xor	eax,esi
+	xor	edx,esi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	edi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	edi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	edi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	edi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	edi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	;
+
+	; Round 15
+	mov	eax,DWORD PTR 36[esp]
+	mov	edx,edi
+	shr	edx,16
+	mov	ecx,DWORD PTR 40[esp]
+	xor	edx,edi
+	and	eax,edx
+	and	edx,ecx
+	mov	ebx,eax
+	shl	ebx,16
+	mov	ecx,edx
+	shl	ecx,16
+	xor	eax,ebx
+	xor	edx,ecx
+	mov	ebx,DWORD PTR 120[ebp]
+	xor	eax,ebx
+	mov	ecx,DWORD PTR 124[ebp]
+	xor	eax,edi
+	xor	edx,edi
+	xor	edx,ecx
+	and	eax,0fcfcfcfch
+	xor	ebx,ebx
+	and	edx,0cfcfcfcfh
+	xor	ecx,ecx
+	mov	bl,al
+	mov	cl,ah
+	ror	edx,4
+	mov	ebp,DWORD PTR 4[esp]
+	xor	esi,DWORD PTR [ebx*1+ebp]
+	mov	bl,dl
+	xor	esi,DWORD PTR 0200h[ecx*1+ebp]
+	mov	cl,dh
+	shr	eax,16
+	xor	esi,DWORD PTR 0100h[ebx*1+ebp]
+	mov	bl,ah
+	shr	edx,16
+	xor	esi,DWORD PTR 0300h[ecx*1+ebp]
+	mov	cl,dh
+	and	eax,0ffh
+	and	edx,0ffh
+	mov	ebx,DWORD PTR 0600h[ebx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0700h[ecx*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0400h[eax*1+ebp]
+	xor	esi,ebx
+	mov	ebx,DWORD PTR 0500h[edx*1+ebp]
+	xor	esi,ebx
+	mov	ebp,DWORD PTR 32[esp]
+	mov	ebx,DWORD PTR [esp]
+	mov	eax,edi
+	dec	ebx
+	mov	edi,esi
+	mov	esi,eax
+	mov	DWORD PTR [esp],ebx
+	jnz	$L000start
+	;
+
+	; FP
+	mov	edx,DWORD PTR 28[esp]
+	ror	edi,1
+	mov	eax,esi
+	xor	esi,edi
+	and	esi,0aaaaaaaah
+	xor	eax,esi
+	xor	edi,esi
+	;
+
+	rol	eax,23
+	mov	esi,eax
+	xor	eax,edi
+	and	eax,003fc03fch
+	xor	esi,eax
+	xor	edi,eax
+	;
+
+	rol	esi,10
+	mov	eax,esi
+	xor	esi,edi
+	and	esi,033333333h
+	xor	eax,esi
+	xor	edi,esi
+	;
+
+	rol	edi,18
+	mov	esi,edi
+	xor	edi,eax
+	and	edi,0fff0000fh
+	xor	esi,edi
+	xor	eax,edi
+	;
+
+	rol	esi,12
+	mov	edi,esi
+	xor	esi,eax
+	and	esi,0f0f0f0f0h
+	xor	edi,esi
+	xor	eax,esi
+	;
+
+	ror	eax,4
+	mov	DWORD PTR [edx],eax
+	mov	DWORD PTR 4[edx],edi
+	add	esp,8
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_fcrypt_body ENDP
+.text$	ENDS
+END

deps/openssl/asm/x86-win32-masm/md5/md5-586.asm

@@ -0,0 +1,693 @@
+TITLE	../openssl/crypto/md5/asm/md5-586.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+ALIGN	16
+_md5_block_asm_data_order	PROC PUBLIC
+$L_md5_block_asm_data_order_begin::
+	push	esi
+	push	edi
+	mov	edi,DWORD PTR 12[esp]
+	mov	esi,DWORD PTR 16[esp]
+	mov	ecx,DWORD PTR 20[esp]
+	push	ebp
+	shl	ecx,6
+	push	ebx
+	add	ecx,esi
+	sub	ecx,64
+	mov	eax,DWORD PTR [edi]
+	push	ecx
+	mov	ebx,DWORD PTR 4[edi]
+	mov	ecx,DWORD PTR 8[edi]
+	mov	edx,DWORD PTR 12[edi]
+$L000start:
+	;
+
+	; R0 section
+	mov	edi,ecx
+	mov	ebp,DWORD PTR [esi]
+	; R0 0
+	xor	edi,edx
+	and	edi,ebx
+	lea	eax,DWORD PTR 3614090360[ebp*1+eax]
+	xor	edi,edx
+	add	eax,edi
+	mov	edi,ebx
+	rol	eax,7
+	mov	ebp,DWORD PTR 4[esi]
+	add	eax,ebx
+	; R0 1
+	xor	edi,ecx
+	and	edi,eax
+	lea	edx,DWORD PTR 3905402710[ebp*1+edx]
+	xor	edi,ecx
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,12
+	mov	ebp,DWORD PTR 8[esi]
+	add	edx,eax
+	; R0 2
+	xor	edi,ebx
+	and	edi,edx
+	lea	ecx,DWORD PTR 606105819[ebp*1+ecx]
+	xor	edi,ebx
+	add	ecx,edi
+	mov	edi,edx
+	rol	ecx,17
+	mov	ebp,DWORD PTR 12[esi]
+	add	ecx,edx
+	; R0 3
+	xor	edi,eax
+	and	edi,ecx
+	lea	ebx,DWORD PTR 3250441966[ebp*1+ebx]
+	xor	edi,eax
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,22
+	mov	ebp,DWORD PTR 16[esi]
+	add	ebx,ecx
+	; R0 4
+	xor	edi,edx
+	and	edi,ebx
+	lea	eax,DWORD PTR 4118548399[ebp*1+eax]
+	xor	edi,edx
+	add	eax,edi
+	mov	edi,ebx
+	rol	eax,7
+	mov	ebp,DWORD PTR 20[esi]
+	add	eax,ebx
+	; R0 5
+	xor	edi,ecx
+	and	edi,eax
+	lea	edx,DWORD PTR 1200080426[ebp*1+edx]
+	xor	edi,ecx
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,12
+	mov	ebp,DWORD PTR 24[esi]
+	add	edx,eax
+	; R0 6
+	xor	edi,ebx
+	and	edi,edx
+	lea	ecx,DWORD PTR 2821735955[ebp*1+ecx]
+	xor	edi,ebx
+	add	ecx,edi
+	mov	edi,edx
+	rol	ecx,17
+	mov	ebp,DWORD PTR 28[esi]
+	add	ecx,edx
+	; R0 7
+	xor	edi,eax
+	and	edi,ecx
+	lea	ebx,DWORD PTR 4249261313[ebp*1+ebx]
+	xor	edi,eax
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,22
+	mov	ebp,DWORD PTR 32[esi]
+	add	ebx,ecx
+	; R0 8
+	xor	edi,edx
+	and	edi,ebx
+	lea	eax,DWORD PTR 1770035416[ebp*1+eax]
+	xor	edi,edx
+	add	eax,edi
+	mov	edi,ebx
+	rol	eax,7
+	mov	ebp,DWORD PTR 36[esi]
+	add	eax,ebx
+	; R0 9
+	xor	edi,ecx
+	and	edi,eax
+	lea	edx,DWORD PTR 2336552879[ebp*1+edx]
+	xor	edi,ecx
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,12
+	mov	ebp,DWORD PTR 40[esi]
+	add	edx,eax
+	; R0 10
+	xor	edi,ebx
+	and	edi,edx
+	lea	ecx,DWORD PTR 4294925233[ebp*1+ecx]
+	xor	edi,ebx
+	add	ecx,edi
+	mov	edi,edx
+	rol	ecx,17
+	mov	ebp,DWORD PTR 44[esi]
+	add	ecx,edx
+	; R0 11
+	xor	edi,eax
+	and	edi,ecx
+	lea	ebx,DWORD PTR 2304563134[ebp*1+ebx]
+	xor	edi,eax
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,22
+	mov	ebp,DWORD PTR 48[esi]
+	add	ebx,ecx
+	; R0 12
+	xor	edi,edx
+	and	edi,ebx
+	lea	eax,DWORD PTR 1804603682[ebp*1+eax]
+	xor	edi,edx
+	add	eax,edi
+	mov	edi,ebx
+	rol	eax,7
+	mov	ebp,DWORD PTR 52[esi]
+	add	eax,ebx
+	; R0 13
+	xor	edi,ecx
+	and	edi,eax
+	lea	edx,DWORD PTR 4254626195[ebp*1+edx]
+	xor	edi,ecx
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,12
+	mov	ebp,DWORD PTR 56[esi]
+	add	edx,eax
+	; R0 14
+	xor	edi,ebx
+	and	edi,edx
+	lea	ecx,DWORD PTR 2792965006[ebp*1+ecx]
+	xor	edi,ebx
+	add	ecx,edi
+	mov	edi,edx
+	rol	ecx,17
+	mov	ebp,DWORD PTR 60[esi]
+	add	ecx,edx
+	; R0 15
+	xor	edi,eax
+	and	edi,ecx
+	lea	ebx,DWORD PTR 1236535329[ebp*1+ebx]
+	xor	edi,eax
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,22
+	mov	ebp,DWORD PTR 4[esi]
+	add	ebx,ecx
+	;
+
+	; R1 section
+	; R1 16
+	lea	eax,DWORD PTR 4129170786[ebp*1+eax]
+	xor	edi,ebx
+	and	edi,edx
+	mov	ebp,DWORD PTR 24[esi]
+	xor	edi,ecx
+	add	eax,edi
+	mov	edi,ebx
+	rol	eax,5
+	add	eax,ebx
+	; R1 17
+	lea	edx,DWORD PTR 3225465664[ebp*1+edx]
+	xor	edi,eax
+	and	edi,ecx
+	mov	ebp,DWORD PTR 44[esi]
+	xor	edi,ebx
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,9
+	add	edx,eax
+	; R1 18
+	lea	ecx,DWORD PTR 643717713[ebp*1+ecx]
+	xor	edi,edx
+	and	edi,ebx
+	mov	ebp,DWORD PTR [esi]
+	xor	edi,eax
+	add	ecx,edi
+	mov	edi,edx
+	rol	ecx,14
+	add	ecx,edx
+	; R1 19
+	lea	ebx,DWORD PTR 3921069994[ebp*1+ebx]
+	xor	edi,ecx
+	and	edi,eax
+	mov	ebp,DWORD PTR 20[esi]
+	xor	edi,edx
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,20
+	add	ebx,ecx
+	; R1 20
+	lea	eax,DWORD PTR 3593408605[ebp*1+eax]
+	xor	edi,ebx
+	and	edi,edx
+	mov	ebp,DWORD PTR 40[esi]
+	xor	edi,ecx
+	add	eax,edi
+	mov	edi,ebx
+	rol	eax,5
+	add	eax,ebx
+	; R1 21
+	lea	edx,DWORD PTR 38016083[ebp*1+edx]
+	xor	edi,eax
+	and	edi,ecx
+	mov	ebp,DWORD PTR 60[esi]
+	xor	edi,ebx
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,9
+	add	edx,eax
+	; R1 22
+	lea	ecx,DWORD PTR 3634488961[ebp*1+ecx]
+	xor	edi,edx
+	and	edi,ebx
+	mov	ebp,DWORD PTR 16[esi]
+	xor	edi,eax
+	add	ecx,edi
+	mov	edi,edx
+	rol	ecx,14
+	add	ecx,edx
+	; R1 23
+	lea	ebx,DWORD PTR 3889429448[ebp*1+ebx]
+	xor	edi,ecx
+	and	edi,eax
+	mov	ebp,DWORD PTR 36[esi]
+	xor	edi,edx
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,20
+	add	ebx,ecx
+	; R1 24
+	lea	eax,DWORD PTR 568446438[ebp*1+eax]
+	xor	edi,ebx
+	and	edi,edx
+	mov	ebp,DWORD PTR 56[esi]
+	xor	edi,ecx
+	add	eax,edi
+	mov	edi,ebx
+	rol	eax,5
+	add	eax,ebx
+	; R1 25
+	lea	edx,DWORD PTR 3275163606[ebp*1+edx]
+	xor	edi,eax
+	and	edi,ecx
+	mov	ebp,DWORD PTR 12[esi]
+	xor	edi,ebx
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,9
+	add	edx,eax
+	; R1 26
+	lea	ecx,DWORD PTR 4107603335[ebp*1+ecx]
+	xor	edi,edx
+	and	edi,ebx
+	mov	ebp,DWORD PTR 32[esi]
+	xor	edi,eax
+	add	ecx,edi
+	mov	edi,edx
+	rol	ecx,14
+	add	ecx,edx
+	; R1 27
+	lea	ebx,DWORD PTR 1163531501[ebp*1+ebx]
+	xor	edi,ecx
+	and	edi,eax
+	mov	ebp,DWORD PTR 52[esi]
+	xor	edi,edx
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,20
+	add	ebx,ecx
+	; R1 28
+	lea	eax,DWORD PTR 2850285829[ebp*1+eax]
+	xor	edi,ebx
+	and	edi,edx
+	mov	ebp,DWORD PTR 8[esi]
+	xor	edi,ecx
+	add	eax,edi
+	mov	edi,ebx
+	rol	eax,5
+	add	eax,ebx
+	; R1 29
+	lea	edx,DWORD PTR 4243563512[ebp*1+edx]
+	xor	edi,eax
+	and	edi,ecx
+	mov	ebp,DWORD PTR 28[esi]
+	xor	edi,ebx
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,9
+	add	edx,eax
+	; R1 30
+	lea	ecx,DWORD PTR 1735328473[ebp*1+ecx]
+	xor	edi,edx
+	and	edi,ebx
+	mov	ebp,DWORD PTR 48[esi]
+	xor	edi,eax
+	add	ecx,edi
+	mov	edi,edx
+	rol	ecx,14
+	add	ecx,edx
+	; R1 31
+	lea	ebx,DWORD PTR 2368359562[ebp*1+ebx]
+	xor	edi,ecx
+	and	edi,eax
+	mov	ebp,DWORD PTR 20[esi]
+	xor	edi,edx
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,20
+	add	ebx,ecx
+	;
+
+	; R2 section
+	; R2 32
+	xor	edi,edx
+	xor	edi,ebx
+	lea	eax,DWORD PTR 4294588738[ebp*1+eax]
+	add	eax,edi
+	rol	eax,4
+	mov	ebp,DWORD PTR 32[esi]
+	mov	edi,ebx
+	; R2 33
+	lea	edx,DWORD PTR 2272392833[ebp*1+edx]
+	add	eax,ebx
+	xor	edi,ecx
+	xor	edi,eax
+	mov	ebp,DWORD PTR 44[esi]
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,11
+	add	edx,eax
+	; R2 34
+	xor	edi,ebx
+	xor	edi,edx
+	lea	ecx,DWORD PTR 1839030562[ebp*1+ecx]
+	add	ecx,edi
+	rol	ecx,16
+	mov	ebp,DWORD PTR 56[esi]
+	mov	edi,edx
+	; R2 35
+	lea	ebx,DWORD PTR 4259657740[ebp*1+ebx]
+	add	ecx,edx
+	xor	edi,eax
+	xor	edi,ecx
+	mov	ebp,DWORD PTR 4[esi]
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,23
+	add	ebx,ecx
+	; R2 36
+	xor	edi,edx
+	xor	edi,ebx
+	lea	eax,DWORD PTR 2763975236[ebp*1+eax]
+	add	eax,edi
+	rol	eax,4
+	mov	ebp,DWORD PTR 16[esi]
+	mov	edi,ebx
+	; R2 37
+	lea	edx,DWORD PTR 1272893353[ebp*1+edx]
+	add	eax,ebx
+	xor	edi,ecx
+	xor	edi,eax
+	mov	ebp,DWORD PTR 28[esi]
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,11
+	add	edx,eax
+	; R2 38
+	xor	edi,ebx
+	xor	edi,edx
+	lea	ecx,DWORD PTR 4139469664[ebp*1+ecx]
+	add	ecx,edi
+	rol	ecx,16
+	mov	ebp,DWORD PTR 40[esi]
+	mov	edi,edx
+	; R2 39
+	lea	ebx,DWORD PTR 3200236656[ebp*1+ebx]
+	add	ecx,edx
+	xor	edi,eax
+	xor	edi,ecx
+	mov	ebp,DWORD PTR 52[esi]
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,23
+	add	ebx,ecx
+	; R2 40
+	xor	edi,edx
+	xor	edi,ebx
+	lea	eax,DWORD PTR 681279174[ebp*1+eax]
+	add	eax,edi
+	rol	eax,4
+	mov	ebp,DWORD PTR [esi]
+	mov	edi,ebx
+	; R2 41
+	lea	edx,DWORD PTR 3936430074[ebp*1+edx]
+	add	eax,ebx
+	xor	edi,ecx
+	xor	edi,eax
+	mov	ebp,DWORD PTR 12[esi]
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,11
+	add	edx,eax
+	; R2 42
+	xor	edi,ebx
+	xor	edi,edx
+	lea	ecx,DWORD PTR 3572445317[ebp*1+ecx]
+	add	ecx,edi
+	rol	ecx,16
+	mov	ebp,DWORD PTR 24[esi]
+	mov	edi,edx
+	; R2 43
+	lea	ebx,DWORD PTR 76029189[ebp*1+ebx]
+	add	ecx,edx
+	xor	edi,eax
+	xor	edi,ecx
+	mov	ebp,DWORD PTR 36[esi]
+	add	ebx,edi
+	mov	edi,ecx
+	rol	ebx,23
+	add	ebx,ecx
+	; R2 44
+	xor	edi,edx
+	xor	edi,ebx
+	lea	eax,DWORD PTR 3654602809[ebp*1+eax]
+	add	eax,edi
+	rol	eax,4
+	mov	ebp,DWORD PTR 48[esi]
+	mov	edi,ebx
+	; R2 45
+	lea	edx,DWORD PTR 3873151461[ebp*1+edx]
+	add	eax,ebx
+	xor	edi,ecx
+	xor	edi,eax
+	mov	ebp,DWORD PTR 60[esi]
+	add	edx,edi
+	mov	edi,eax
+	rol	edx,11
+	add	edx,eax
+	; R2 46
+	xor	edi,ebx
+	xor	edi,edx
+	lea	ecx,DWORD PTR 530742520[ebp*1+ecx]
+	add	ecx,edi
+	rol	ecx,16
+	mov	ebp,DWORD PTR 8[esi]
+	mov	edi,edx
+	; R2 47
+	lea	ebx,DWORD PTR 3299628645[ebp*1+ebx]
+	add	ecx,edx
+	xor	edi,eax
+	xor	edi,ecx
+	mov	ebp,DWORD PTR [esi]
+	add	ebx,edi
+	mov	edi,-1
+	rol	ebx,23
+	add	ebx,ecx
+	;
+
+	; R3 section
+	; R3 48
+	xor	edi,edx
+	or	edi,ebx
+	lea	eax,DWORD PTR 4096336452[ebp*1+eax]
+	xor	edi,ecx
+	mov	ebp,DWORD PTR 28[esi]
+	add	eax,edi
+	mov	edi,-1
+	rol	eax,6
+	xor	edi,ecx
+	add	eax,ebx
+	; R3 49
+	or	edi,eax
+	lea	edx,DWORD PTR 1126891415[ebp*1+edx]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 56[esi]
+	add	edx,edi
+	mov	edi,-1
+	rol	edx,10
+	xor	edi,ebx
+	add	edx,eax
+	; R3 50
+	or	edi,edx
+	lea	ecx,DWORD PTR 2878612391[ebp*1+ecx]
+	xor	edi,eax
+	mov	ebp,DWORD PTR 20[esi]
+	add	ecx,edi
+	mov	edi,-1
+	rol	ecx,15
+	xor	edi,eax
+	add	ecx,edx
+	; R3 51
+	or	edi,ecx
+	lea	ebx,DWORD PTR 4237533241[ebp*1+ebx]
+	xor	edi,edx
+	mov	ebp,DWORD PTR 48[esi]
+	add	ebx,edi
+	mov	edi,-1
+	rol	ebx,21
+	xor	edi,edx
+	add	ebx,ecx
+	; R3 52
+	or	edi,ebx
+	lea	eax,DWORD PTR 1700485571[ebp*1+eax]
+	xor	edi,ecx
+	mov	ebp,DWORD PTR 12[esi]
+	add	eax,edi
+	mov	edi,-1
+	rol	eax,6
+	xor	edi,ecx
+	add	eax,ebx
+	; R3 53
+	or	edi,eax
+	lea	edx,DWORD PTR 2399980690[ebp*1+edx]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 40[esi]
+	add	edx,edi
+	mov	edi,-1
+	rol	edx,10
+	xor	edi,ebx
+	add	edx,eax
+	; R3 54
+	or	edi,edx
+	lea	ecx,DWORD PTR 4293915773[ebp*1+ecx]
+	xor	edi,eax
+	mov	ebp,DWORD PTR 4[esi]
+	add	ecx,edi
+	mov	edi,-1
+	rol	ecx,15
+	xor	edi,eax
+	add	ecx,edx
+	; R3 55
+	or	edi,ecx
+	lea	ebx,DWORD PTR 2240044497[ebp*1+ebx]
+	xor	edi,edx
+	mov	ebp,DWORD PTR 32[esi]
+	add	ebx,edi
+	mov	edi,-1
+	rol	ebx,21
+	xor	edi,edx
+	add	ebx,ecx
+	; R3 56
+	or	edi,ebx
+	lea	eax,DWORD PTR 1873313359[ebp*1+eax]
+	xor	edi,ecx
+	mov	ebp,DWORD PTR 60[esi]
+	add	eax,edi
+	mov	edi,-1
+	rol	eax,6
+	xor	edi,ecx
+	add	eax,ebx
+	; R3 57
+	or	edi,eax
+	lea	edx,DWORD PTR 4264355552[ebp*1+edx]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 24[esi]
+	add	edx,edi
+	mov	edi,-1
+	rol	edx,10
+	xor	edi,ebx
+	add	edx,eax
+	; R3 58
+	or	edi,edx
+	lea	ecx,DWORD PTR 2734768916[ebp*1+ecx]
+	xor	edi,eax
+	mov	ebp,DWORD PTR 52[esi]
+	add	ecx,edi
+	mov	edi,-1
+	rol	ecx,15
+	xor	edi,eax
+	add	ecx,edx
+	; R3 59
+	or	edi,ecx
+	lea	ebx,DWORD PTR 1309151649[ebp*1+ebx]
+	xor	edi,edx
+	mov	ebp,DWORD PTR 16[esi]
+	add	ebx,edi
+	mov	edi,-1
+	rol	ebx,21
+	xor	edi,edx
+	add	ebx,ecx
+	; R3 60
+	or	edi,ebx
+	lea	eax,DWORD PTR 4149444226[ebp*1+eax]
+	xor	edi,ecx
+	mov	ebp,DWORD PTR 44[esi]
+	add	eax,edi
+	mov	edi,-1
+	rol	eax,6
+	xor	edi,ecx
+	add	eax,ebx
+	; R3 61
+	or	edi,eax
+	lea	edx,DWORD PTR 3174756917[ebp*1+edx]
+	xor	edi,ebx
+	mov	ebp,DWORD PTR 8[esi]
+	add	edx,edi
+	mov	edi,-1
+	rol	edx,10
+	xor	edi,ebx
+	add	edx,eax
+	; R3 62
+	or	edi,edx
+	lea	ecx,DWORD PTR 718787259[ebp*1+ecx]
+	xor	edi,eax
+	mov	ebp,DWORD PTR 36[esi]
+	add	ecx,edi
+	mov	edi,-1
+	rol	ecx,15
+	xor	edi,eax
+	add	ecx,edx
+	; R3 63
+	or	edi,ecx
+	lea	ebx,DWORD PTR 3951481745[ebp*1+ebx]
+	xor	edi,edx
+	mov	ebp,DWORD PTR 24[esp]
+	add	ebx,edi
+	add	esi,64
+	rol	ebx,21
+	mov	edi,DWORD PTR [ebp]
+	add	ebx,ecx
+	add	eax,edi
+	mov	edi,DWORD PTR 4[ebp]
+	add	ebx,edi
+	mov	edi,DWORD PTR 8[ebp]
+	add	ecx,edi
+	mov	edi,DWORD PTR 12[ebp]
+	add	edx,edi
+	mov	DWORD PTR [ebp],eax
+	mov	DWORD PTR 4[ebp],ebx
+	mov	edi,DWORD PTR [esp]
+	mov	DWORD PTR 8[ebp],ecx
+	mov	DWORD PTR 12[ebp],edx
+	cmp	edi,esi
+	jae	$L000start
+	pop	eax
+	pop	ebx
+	pop	ebp
+	pop	edi
+	pop	esi
+	ret
+_md5_block_asm_data_order ENDP
+.text$	ENDS
+END

deps/openssl/asm/x86-win32-masm/rc4/rc4-586.asm

@@ -0,0 +1,239 @@
+TITLE	rc4-586.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+ALIGN	16
+_RC4	PROC PUBLIC
+$L_RC4_begin::
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	mov	edi,DWORD PTR 20[esp]
+	mov	edx,DWORD PTR 24[esp]
+	mov	esi,DWORD PTR 28[esp]
+	mov	ebp,DWORD PTR 32[esp]
+	xor	eax,eax
+	xor	ebx,ebx
+	cmp	edx,0
+	je	$L000abort
+	mov	al,BYTE PTR [edi]
+	mov	bl,BYTE PTR 4[edi]
+	add	edi,8
+	lea	ecx,DWORD PTR [edx*1+esi]
+	sub	ebp,esi
+	mov	DWORD PTR 24[esp],ecx
+	inc	al
+	cmp	DWORD PTR 256[edi],-1
+	je	$L001RC4_CHAR
+	mov	ecx,DWORD PTR [eax*4+edi]
+	and	edx,-4
+	jz	$L002loop1
+	lea	edx,DWORD PTR [edx*1+esi-4]
+	mov	DWORD PTR 28[esp],edx
+	mov	DWORD PTR 32[esp],ebp
+ALIGN	16
+$L003loop4:
+	add	bl,cl
+	mov	edx,DWORD PTR [ebx*4+edi]
+	mov	DWORD PTR [ebx*4+edi],ecx
+	mov	DWORD PTR [eax*4+edi],edx
+	add	edx,ecx
+	inc	al
+	and	edx,255
+	mov	ecx,DWORD PTR [eax*4+edi]
+	mov	ebp,DWORD PTR [edx*4+edi]
+	add	bl,cl
+	mov	edx,DWORD PTR [ebx*4+edi]
+	mov	DWORD PTR [ebx*4+edi],ecx
+	mov	DWORD PTR [eax*4+edi],edx
+	add	edx,ecx
+	inc	al
+	and	edx,255
+	ror	ebp,8
+	mov	ecx,DWORD PTR [eax*4+edi]
+	or	ebp,DWORD PTR [edx*4+edi]
+	add	bl,cl
+	mov	edx,DWORD PTR [ebx*4+edi]
+	mov	DWORD PTR [ebx*4+edi],ecx
+	mov	DWORD PTR [eax*4+edi],edx
+	add	edx,ecx
+	inc	al
+	and	edx,255
+	ror	ebp,8
+	mov	ecx,DWORD PTR [eax*4+edi]
+	or	ebp,DWORD PTR [edx*4+edi]
+	add	bl,cl
+	mov	edx,DWORD PTR [ebx*4+edi]
+	mov	DWORD PTR [ebx*4+edi],ecx
+	mov	DWORD PTR [eax*4+edi],edx
+	add	edx,ecx
+	inc	al
+	and	edx,255
+	ror	ebp,8
+	mov	ecx,DWORD PTR 32[esp]
+	or	ebp,DWORD PTR [edx*4+edi]
+	ror	ebp,8
+	xor	ebp,DWORD PTR [esi]
+	cmp	esi,DWORD PTR 28[esp]
+	mov	DWORD PTR [esi*1+ecx],ebp
+	lea	esi,DWORD PTR 4[esi]
+	mov	ecx,DWORD PTR [eax*4+edi]
+	jb	$L003loop4
+	cmp	esi,DWORD PTR 24[esp]
+	je	$L004done
+	mov	ebp,DWORD PTR 32[esp]
+ALIGN	16
+$L002loop1:
+	add	bl,cl
+	mov	edx,DWORD PTR [ebx*4+edi]
+	mov	DWORD PTR [ebx*4+edi],ecx
+	mov	DWORD PTR [eax*4+edi],edx
+	add	edx,ecx
+	inc	al
+	and	edx,255
+	mov	edx,DWORD PTR [edx*4+edi]
+	xor	dl,BYTE PTR [esi]
+	lea	esi,DWORD PTR 1[esi]
+	mov	ecx,DWORD PTR [eax*4+edi]
+	cmp	esi,DWORD PTR 24[esp]
+	mov	BYTE PTR [esi*1+ebp-1],dl
+	jb	$L002loop1
+	jmp	$L004done
+ALIGN	16
+$L001RC4_CHAR:
+	movzx	ecx,BYTE PTR [eax*1+edi]
+$L005cloop1:
+	add	bl,cl
+	movzx	edx,BYTE PTR [ebx*1+edi]
+	mov	BYTE PTR [ebx*1+edi],cl
+	mov	BYTE PTR [eax*1+edi],dl
+	add	dl,cl
+	movzx	edx,BYTE PTR [edx*1+edi]
+	add	al,1
+	xor	dl,BYTE PTR [esi]
+	lea	esi,DWORD PTR 1[esi]
+	movzx	ecx,BYTE PTR [eax*1+edi]
+	cmp	esi,DWORD PTR 24[esp]
+	mov	BYTE PTR [esi*1+ebp-1],dl
+	jb	$L005cloop1
+$L004done:
+	dec	al
+	mov	BYTE PTR [edi-4],bl
+	mov	BYTE PTR [edi-8],al
+$L000abort:
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_RC4 ENDP
+;EXTERN	_OPENSSL_ia32cap_P:NEAR
+ALIGN	16
+_RC4_set_key	PROC PUBLIC
+$L_RC4_set_key_begin::
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	mov	edi,DWORD PTR 20[esp]
+	mov	ebp,DWORD PTR 24[esp]
+	mov	esi,DWORD PTR 28[esp]
+	lea	edx,DWORD PTR _OPENSSL_ia32cap_P
+	lea	edi,DWORD PTR 8[edi]
+	lea	esi,DWORD PTR [ebp*1+esi]
+	neg	ebp
+	xor	eax,eax
+	mov	DWORD PTR [edi-4],ebp
+	bt	DWORD PTR [edx],20
+	jc	$L006c1stloop
+ALIGN	16
+$L007w1stloop:
+	mov	DWORD PTR [eax*4+edi],eax
+	add	al,1
+	jnc	$L007w1stloop
+	xor	ecx,ecx
+	xor	edx,edx
+ALIGN	16
+$L008w2ndloop:
+	mov	eax,DWORD PTR [ecx*4+edi]
+	add	dl,BYTE PTR [ebp*1+esi]
+	add	dl,al
+	add	ebp,1
+	mov	ebx,DWORD PTR [edx*4+edi]
+	jnz	$L009wnowrap
+	mov	ebp,DWORD PTR [edi-4]
+$L009wnowrap:
+	mov	DWORD PTR [edx*4+edi],eax
+	mov	DWORD PTR [ecx*4+edi],ebx
+	add	cl,1
+	jnc	$L008w2ndloop
+	jmp	$L010exit
+ALIGN	16
+$L006c1stloop:
+	mov	BYTE PTR [eax*1+edi],al
+	add	al,1
+	jnc	$L006c1stloop
+	xor	ecx,ecx
+	xor	edx,edx
+	xor	ebx,ebx
+ALIGN	16
+$L011c2ndloop:
+	mov	al,BYTE PTR [ecx*1+edi]
+	add	dl,BYTE PTR [ebp*1+esi]
+	add	dl,al
+	add	ebp,1
+	mov	bl,BYTE PTR [edx*1+edi]
+	jnz	$L012cnowrap
+	mov	ebp,DWORD PTR [edi-4]
+$L012cnowrap:
+	mov	BYTE PTR [edx*1+edi],al
+	mov	BYTE PTR [ecx*1+edi],bl
+	add	cl,1
+	jnc	$L011c2ndloop
+	mov	DWORD PTR 256[edi],-1
+$L010exit:
+	xor	eax,eax
+	mov	DWORD PTR [edi-8],eax
+	mov	DWORD PTR [edi-4],eax
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_RC4_set_key ENDP
+ALIGN	16
+_RC4_options	PROC PUBLIC
+$L_RC4_options_begin::
+	call	$L013pic_point
+$L013pic_point:
+	pop	eax
+	lea	eax,DWORD PTR ($L014opts-$L013pic_point)[eax]
+	lea	edx,DWORD PTR _OPENSSL_ia32cap_P
+	bt	DWORD PTR [edx],20
+	jnc	$L015skip
+	add	eax,12
+$L015skip:
+	ret
+ALIGN	64
+$L014opts:
+DB	114,99,52,40,52,120,44,105,110,116,41,0
+DB	114,99,52,40,49,120,44,99,104,97,114,41,0
+DB	82,67,52,32,102,111,114,32,120,56,54,44,32,67,82,89
+DB	80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114
+DB	111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+ALIGN	64
+_RC4_options ENDP
+.text$	ENDS
+.bss	SEGMENT 'BSS'
+COMM	_OPENSSL_ia32cap_P:DWORD
+.bss	ENDS
+END

deps/openssl/asm/x86-win32-masm/rc5/rc5-586.asm

@@ -0,0 +1,573 @@
+TITLE	rc5-586.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+ALIGN	16
+_RC5_32_encrypt	PROC PUBLIC
+$L_RC5_32_encrypt_begin::
+	;
+
+	push	ebp
+	push	esi
+	push	edi
+	mov	edx,DWORD PTR 16[esp]
+	mov	ebp,DWORD PTR 20[esp]
+	; Load the 2 words
+	mov	edi,DWORD PTR [edx]
+	mov	esi,DWORD PTR 4[edx]
+	push	ebx
+	mov	ebx,DWORD PTR [ebp]
+	add	edi,DWORD PTR 4[ebp]
+	add	esi,DWORD PTR 8[ebp]
+	xor	edi,esi
+	mov	eax,DWORD PTR 12[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 16[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 20[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 24[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 28[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 32[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 36[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 40[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 44[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 48[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 52[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 56[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 60[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 64[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 68[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 72[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	cmp	ebx,8
+	je	$L000rc5_exit
+	xor	edi,esi
+	mov	eax,DWORD PTR 76[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 80[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 84[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 88[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 92[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 96[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 100[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 104[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	cmp	ebx,12
+	je	$L000rc5_exit
+	xor	edi,esi
+	mov	eax,DWORD PTR 108[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 112[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 116[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 120[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 124[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 128[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+	xor	edi,esi
+	mov	eax,DWORD PTR 132[ebp]
+	mov	ecx,esi
+	rol	edi,cl
+	add	edi,eax
+	xor	esi,edi
+	mov	eax,DWORD PTR 136[ebp]
+	mov	ecx,edi
+	rol	esi,cl
+	add	esi,eax
+$L000rc5_exit:
+	mov	DWORD PTR [edx],edi
+	mov	DWORD PTR 4[edx],esi
+	pop	ebx
+	pop	edi
+	pop	esi
+	pop	ebp
+	ret
+_RC5_32_encrypt ENDP
+ALIGN	16
+_RC5_32_decrypt	PROC PUBLIC
+$L_RC5_32_decrypt_begin::
+	;
+
+	push	ebp
+	push	esi
+	push	edi
+	mov	edx,DWORD PTR 16[esp]
+	mov	ebp,DWORD PTR 20[esp]
+	; Load the 2 words
+	mov	edi,DWORD PTR [edx]
+	mov	esi,DWORD PTR 4[edx]
+	push	ebx
+	mov	ebx,DWORD PTR [ebp]
+	cmp	ebx,12
+	je	$L001rc5_dec_12
+	cmp	ebx,8
+	je	$L002rc5_dec_8
+	mov	eax,DWORD PTR 136[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 132[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 128[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 124[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 120[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 116[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 112[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 108[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+$L001rc5_dec_12:
+	mov	eax,DWORD PTR 104[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 100[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 96[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 92[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 88[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 84[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 80[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 76[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+$L002rc5_dec_8:
+	mov	eax,DWORD PTR 72[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 68[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 64[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 60[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 56[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 52[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 48[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 44[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 40[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 36[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 32[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 28[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 24[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 20[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	mov	eax,DWORD PTR 16[ebp]
+	sub	esi,eax
+	mov	ecx,edi
+	ror	esi,cl
+	xor	esi,edi
+	mov	eax,DWORD PTR 12[ebp]
+	sub	edi,eax
+	mov	ecx,esi
+	ror	edi,cl
+	xor	edi,esi
+	sub	esi,DWORD PTR 8[ebp]
+	sub	edi,DWORD PTR 4[ebp]
+$L003rc5_exit:
+	mov	DWORD PTR [edx],edi
+	mov	DWORD PTR 4[edx],esi
+	pop	ebx
+	pop	edi
+	pop	esi
+	pop	ebp
+	ret
+_RC5_32_decrypt ENDP
+ALIGN	16
+_RC5_32_cbc_encrypt	PROC PUBLIC
+$L_RC5_32_cbc_encrypt_begin::
+	;
+
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	mov	ebp,DWORD PTR 28[esp]
+	; getting iv ptr from parameter 4
+	mov	ebx,DWORD PTR 36[esp]
+	mov	esi,DWORD PTR [ebx]
+	mov	edi,DWORD PTR 4[ebx]
+	push	edi
+	push	esi
+	push	edi
+	push	esi
+	mov	ebx,esp
+	mov	esi,DWORD PTR 36[esp]
+	mov	edi,DWORD PTR 40[esp]
+	; getting encrypt flag from parameter 5
+	mov	ecx,DWORD PTR 56[esp]
+	; get and push parameter 3
+	mov	eax,DWORD PTR 48[esp]
+	push	eax
+	push	ebx
+	cmp	ecx,0
+	jz	$L004decrypt
+	and	ebp,4294967288
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	jz	$L005encrypt_finish
+$L006encrypt_loop:
+	mov	ecx,DWORD PTR [esi]
+	mov	edx,DWORD PTR 4[esi]
+	xor	eax,ecx
+	xor	ebx,edx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_RC5_32_encrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	mov	DWORD PTR [edi],eax
+	mov	DWORD PTR 4[edi],ebx
+	add	esi,8
+	add	edi,8
+	sub	ebp,8
+	jnz	$L006encrypt_loop
+$L005encrypt_finish:
+	mov	ebp,DWORD PTR 52[esp]
+	and	ebp,7
+	jz	$L007finish
+	call	$L008PIC_point
+$L008PIC_point:
+	pop	edx
+	lea	ecx,DWORD PTR ($L009cbc_enc_jmp_table-$L008PIC_point)[edx]
+	mov	ebp,DWORD PTR [ebp*4+ecx]
+	add	ebp,edx
+	xor	ecx,ecx
+	xor	edx,edx
+	jmp	ebp
+$L010ej7:
+	mov	dh,BYTE PTR 6[esi]
+	shl	edx,8
+$L011ej6:
+	mov	dh,BYTE PTR 5[esi]
+$L012ej5:
+	mov	dl,BYTE PTR 4[esi]
+$L013ej4:
+	mov	ecx,DWORD PTR [esi]
+	jmp	$L014ejend
+$L015ej3:
+	mov	ch,BYTE PTR 2[esi]
+	shl	ecx,8
+$L016ej2:
+	mov	ch,BYTE PTR 1[esi]
+$L017ej1:
+	mov	cl,BYTE PTR [esi]
+$L014ejend:
+	xor	eax,ecx
+	xor	ebx,edx
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_RC5_32_encrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	mov	DWORD PTR [edi],eax
+	mov	DWORD PTR 4[edi],ebx
+	jmp	$L007finish
+$L004decrypt:
+	and	ebp,4294967288
+	mov	eax,DWORD PTR 16[esp]
+	mov	ebx,DWORD PTR 20[esp]
+	jz	$L018decrypt_finish
+$L019decrypt_loop:
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_RC5_32_decrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	mov	ecx,DWORD PTR 16[esp]
+	mov	edx,DWORD PTR 20[esp]
+	xor	ecx,eax
+	xor	edx,ebx
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	mov	DWORD PTR [edi],ecx
+	mov	DWORD PTR 4[edi],edx
+	mov	DWORD PTR 16[esp],eax
+	mov	DWORD PTR 20[esp],ebx
+	add	esi,8
+	add	edi,8
+	sub	ebp,8
+	jnz	$L019decrypt_loop
+$L018decrypt_finish:
+	mov	ebp,DWORD PTR 52[esp]
+	and	ebp,7
+	jz	$L007finish
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+	call	$L_RC5_32_decrypt_begin
+	mov	eax,DWORD PTR 8[esp]
+	mov	ebx,DWORD PTR 12[esp]
+	mov	ecx,DWORD PTR 16[esp]
+	mov	edx,DWORD PTR 20[esp]
+	xor	ecx,eax
+	xor	edx,ebx
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+$L020dj7:
+	ror	edx,16
+	mov	BYTE PTR 6[edi],dl
+	shr	edx,16
+$L021dj6:
+	mov	BYTE PTR 5[edi],dh
+$L022dj5:
+	mov	BYTE PTR 4[edi],dl
+$L023dj4:
+	mov	DWORD PTR [edi],ecx
+	jmp	$L024djend
+$L025dj3:
+	ror	ecx,16
+	mov	BYTE PTR 2[edi],cl
+	shl	ecx,16
+$L026dj2:
+	mov	BYTE PTR 1[esi],ch
+$L027dj1:
+	mov	BYTE PTR [esi],cl
+$L024djend:
+	jmp	$L007finish
+$L007finish:
+	mov	ecx,DWORD PTR 60[esp]
+	add	esp,24
+	mov	DWORD PTR [ecx],eax
+	mov	DWORD PTR 4[ecx],ebx
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+ALIGN	64
+$L009cbc_enc_jmp_table:
+DD	0
+DD	$L017ej1-$L008PIC_point
+DD	$L016ej2-$L008PIC_point
+DD	$L015ej3-$L008PIC_point
+DD	$L013ej4-$L008PIC_point
+DD	$L012ej5-$L008PIC_point
+DD	$L011ej6-$L008PIC_point
+DD	$L010ej7-$L008PIC_point
+ALIGN	64
+_RC5_32_cbc_encrypt ENDP
+.text$	ENDS
+END

deps/openssl/asm/x86-win32-masm/sha/sha256-586.asm

@@ -0,0 +1,271 @@
+TITLE	sha512-586.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+ALIGN	16
+_sha256_block_data_order	PROC PUBLIC
+$L_sha256_block_data_order_begin::
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	mov	esi,DWORD PTR 20[esp]
+	mov	edi,DWORD PTR 24[esp]
+	mov	eax,DWORD PTR 28[esp]
+	mov	ebx,esp
+	call	$L000pic_point
+$L000pic_point:
+	pop	ebp
+	lea	ebp,DWORD PTR ($L001K256-$L000pic_point)[ebp]
+	sub	esp,16
+	and	esp,-64
+	shl	eax,6
+	add	eax,edi
+	mov	DWORD PTR [esp],esi
+	mov	DWORD PTR 4[esp],edi
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+ALIGN	16
+$L002loop:
+	mov	eax,DWORD PTR [edi]
+	mov	ebx,DWORD PTR 4[edi]
+	mov	ecx,DWORD PTR 8[edi]
+	mov	edx,DWORD PTR 12[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 16[edi]
+	mov	ebx,DWORD PTR 20[edi]
+	mov	ecx,DWORD PTR 24[edi]
+	mov	edx,DWORD PTR 28[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 32[edi]
+	mov	ebx,DWORD PTR 36[edi]
+	mov	ecx,DWORD PTR 40[edi]
+	mov	edx,DWORD PTR 44[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 48[edi]
+	mov	ebx,DWORD PTR 52[edi]
+	mov	ecx,DWORD PTR 56[edi]
+	mov	edx,DWORD PTR 60[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	add	edi,64
+	sub	esp,32
+	mov	DWORD PTR 100[esp],edi
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	mov	ecx,DWORD PTR 8[esi]
+	mov	edi,DWORD PTR 12[esi]
+	mov	DWORD PTR 4[esp],ebx
+	mov	DWORD PTR 8[esp],ecx
+	mov	DWORD PTR 12[esp],edi
+	mov	edx,DWORD PTR 16[esi]
+	mov	ebx,DWORD PTR 20[esi]
+	mov	ecx,DWORD PTR 24[esi]
+	mov	edi,DWORD PTR 28[esi]
+	mov	DWORD PTR 20[esp],ebx
+	mov	DWORD PTR 24[esp],ecx
+	mov	DWORD PTR 28[esp],edi
+ALIGN	16
+$L00300_15:
+	mov	ebx,DWORD PTR 92[esp]
+	mov	ecx,edx
+	ror	ecx,6
+	mov	edi,edx
+	ror	edi,11
+	mov	esi,DWORD PTR 20[esp]
+	xor	ecx,edi
+	ror	edi,14
+	xor	ecx,edi
+	mov	edi,DWORD PTR 24[esp]
+	add	ebx,ecx
+	mov	DWORD PTR 16[esp],edx
+	xor	esi,edi
+	mov	ecx,eax
+	and	esi,edx
+	mov	edx,DWORD PTR 12[esp]
+	xor	esi,edi
+	mov	edi,eax
+	add	ebx,esi
+	ror	ecx,2
+	add	ebx,DWORD PTR 28[esp]
+	ror	edi,13
+	mov	esi,DWORD PTR 4[esp]
+	xor	ecx,edi
+	ror	edi,9
+	add	edx,ebx
+	xor	ecx,edi
+	mov	edi,DWORD PTR 8[esp]
+	add	ebx,ecx
+	mov	DWORD PTR [esp],eax
+	mov	ecx,eax
+	sub	esp,4
+	or	eax,esi
+	and	ecx,esi
+	and	eax,edi
+	mov	esi,DWORD PTR [ebp]
+	or	eax,ecx
+	add	ebp,4
+	add	eax,ebx
+	add	edx,esi
+	add	eax,esi
+	cmp	esi,3248222580
+	jne	$L00300_15
+	mov	ebx,DWORD PTR 152[esp]
+ALIGN	16
+$L00416_63:
+	mov	esi,ebx
+	mov	ecx,DWORD PTR 100[esp]
+	shr	ebx,3
+	ror	esi,7
+	xor	ebx,esi
+	ror	esi,11
+	mov	edi,ecx
+	xor	ebx,esi
+	shr	ecx,10
+	mov	esi,DWORD PTR 156[esp]
+	ror	edi,17
+	xor	ecx,edi
+	ror	edi,2
+	add	ebx,esi
+	xor	edi,ecx
+	add	ebx,edi
+	mov	ecx,edx
+	add	ebx,DWORD PTR 120[esp]
+	ror	ecx,6
+	mov	edi,edx
+	ror	edi,11
+	mov	esi,DWORD PTR 20[esp]
+	xor	ecx,edi
+	ror	edi,14
+	mov	DWORD PTR 92[esp],ebx
+	xor	ecx,edi
+	mov	edi,DWORD PTR 24[esp]
+	add	ebx,ecx
+	mov	DWORD PTR 16[esp],edx
+	xor	esi,edi
+	mov	ecx,eax
+	and	esi,edx
+	mov	edx,DWORD PTR 12[esp]
+	xor	esi,edi
+	mov	edi,eax
+	add	ebx,esi
+	ror	ecx,2
+	add	ebx,DWORD PTR 28[esp]
+	ror	edi,13
+	mov	esi,DWORD PTR 4[esp]
+	xor	ecx,edi
+	ror	edi,9
+	add	edx,ebx
+	xor	ecx,edi
+	mov	edi,DWORD PTR 8[esp]
+	add	ebx,ecx
+	mov	DWORD PTR [esp],eax
+	mov	ecx,eax
+	sub	esp,4
+	or	eax,esi
+	and	ecx,esi
+	and	eax,edi
+	mov	esi,DWORD PTR [ebp]
+	or	eax,ecx
+	add	ebp,4
+	add	eax,ebx
+	mov	ebx,DWORD PTR 152[esp]
+	add	edx,esi
+	add	eax,esi
+	cmp	esi,3329325298
+	jne	$L00416_63
+	mov	esi,DWORD PTR 352[esp]
+	mov	ebx,DWORD PTR 4[esp]
+	mov	ecx,DWORD PTR 8[esp]
+	mov	edi,DWORD PTR 12[esp]
+	add	eax,DWORD PTR [esi]
+	add	ebx,DWORD PTR 4[esi]
+	add	ecx,DWORD PTR 8[esi]
+	add	edi,DWORD PTR 12[esi]
+	mov	DWORD PTR [esi],eax
+	mov	DWORD PTR 4[esi],ebx
+	mov	DWORD PTR 8[esi],ecx
+	mov	DWORD PTR 12[esi],edi
+	mov	eax,DWORD PTR 20[esp]
+	mov	ebx,DWORD PTR 24[esp]
+	mov	ecx,DWORD PTR 28[esp]
+	mov	edi,DWORD PTR 356[esp]
+	add	edx,DWORD PTR 16[esi]
+	add	eax,DWORD PTR 20[esi]
+	add	ebx,DWORD PTR 24[esi]
+	add	ecx,DWORD PTR 28[esi]
+	mov	DWORD PTR 16[esi],edx
+	mov	DWORD PTR 20[esi],eax
+	mov	DWORD PTR 24[esi],ebx
+	mov	DWORD PTR 28[esi],ecx
+	add	esp,352
+	sub	ebp,256
+	cmp	edi,DWORD PTR 8[esp]
+	jb	$L002loop
+	mov	esp,DWORD PTR 12[esp]
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+ALIGN	64
+$L001K256:
+DD	1116352408,1899447441,3049323471,3921009573
+DD	961987163,1508970993,2453635748,2870763221
+DD	3624381080,310598401,607225278,1426881987
+DD	1925078388,2162078206,2614888103,3248222580
+DD	3835390401,4022224774,264347078,604807628
+DD	770255983,1249150122,1555081692,1996064986
+DD	2554220882,2821834349,2952996808,3210313671
+DD	3336571891,3584528711,113926993,338241895
+DD	666307205,773529912,1294757372,1396182291
+DD	1695183700,1986661051,2177026350,2456956037
+DD	2730485921,2820302411,3259730800,3345764771
+DD	3516065817,3600352804,4094571909,275423344
+DD	430227734,506948616,659060556,883997877
+DD	958139571,1322822218,1537002063,1747873779
+DD	1955562222,2024104815,2227730452,2361852424
+DD	2428436474,2756734187,3204031479,3329325298
+_sha256_block_data_order ENDP
+DB	83,72,65,50,53,54,32,98,108,111,99,107,32,116,114,97
+DB	110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+DB	67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+DB	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+DB	62,0
+.text$	ENDS
+END

deps/openssl/asm/x86-win32-masm/sha/sha512-586.asm

@@ -0,0 +1,573 @@
+TITLE	sha512-586.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+ALIGN	16
+_sha512_block_data_order	PROC PUBLIC
+$L_sha512_block_data_order_begin::
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	mov	esi,DWORD PTR 20[esp]
+	mov	edi,DWORD PTR 24[esp]
+	mov	eax,DWORD PTR 28[esp]
+	mov	ebx,esp
+	call	$L000pic_point
+$L000pic_point:
+	pop	ebp
+	lea	ebp,DWORD PTR ($L001K512-$L000pic_point)[ebp]
+	sub	esp,16
+	and	esp,-64
+	shl	eax,7
+	add	eax,edi
+	mov	DWORD PTR [esp],esi
+	mov	DWORD PTR 4[esp],edi
+	mov	DWORD PTR 8[esp],eax
+	mov	DWORD PTR 12[esp],ebx
+ALIGN	16
+$L002loop_x86:
+	mov	eax,DWORD PTR [edi]
+	mov	ebx,DWORD PTR 4[edi]
+	mov	ecx,DWORD PTR 8[edi]
+	mov	edx,DWORD PTR 12[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 16[edi]
+	mov	ebx,DWORD PTR 20[edi]
+	mov	ecx,DWORD PTR 24[edi]
+	mov	edx,DWORD PTR 28[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 32[edi]
+	mov	ebx,DWORD PTR 36[edi]
+	mov	ecx,DWORD PTR 40[edi]
+	mov	edx,DWORD PTR 44[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 48[edi]
+	mov	ebx,DWORD PTR 52[edi]
+	mov	ecx,DWORD PTR 56[edi]
+	mov	edx,DWORD PTR 60[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 64[edi]
+	mov	ebx,DWORD PTR 68[edi]
+	mov	ecx,DWORD PTR 72[edi]
+	mov	edx,DWORD PTR 76[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 80[edi]
+	mov	ebx,DWORD PTR 84[edi]
+	mov	ecx,DWORD PTR 88[edi]
+	mov	edx,DWORD PTR 92[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 96[edi]
+	mov	ebx,DWORD PTR 100[edi]
+	mov	ecx,DWORD PTR 104[edi]
+	mov	edx,DWORD PTR 108[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	mov	eax,DWORD PTR 112[edi]
+	mov	ebx,DWORD PTR 116[edi]
+	mov	ecx,DWORD PTR 120[edi]
+	mov	edx,DWORD PTR 124[edi]
+	bswap	eax
+	bswap	ebx
+	bswap	ecx
+	bswap	edx
+	push	eax
+	push	ebx
+	push	ecx
+	push	edx
+	add	edi,128
+	sub	esp,72
+	mov	DWORD PTR 204[esp],edi
+	lea	edi,DWORD PTR 8[esp]
+	mov	ecx,16
+DD	2784229001
+ALIGN	16
+$L00300_15_x86:
+	mov	ecx,DWORD PTR 40[esp]
+	mov	edx,DWORD PTR 44[esp]
+	mov	esi,ecx
+	shr	ecx,9
+	mov	edi,edx
+	shr	edx,9
+	mov	ebx,ecx
+	shl	esi,14
+	mov	eax,edx
+	shl	edi,14
+	xor	ebx,esi
+	shr	ecx,5
+	xor	eax,edi
+	shr	edx,5
+	xor	eax,ecx
+	shl	esi,4
+	xor	ebx,edx
+	shl	edi,4
+	xor	ebx,esi
+	shr	ecx,4
+	xor	eax,edi
+	shr	edx,4
+	xor	eax,ecx
+	shl	esi,5
+	xor	ebx,edx
+	shl	edi,5
+	xor	eax,esi
+	xor	ebx,edi
+	mov	ecx,DWORD PTR 48[esp]
+	mov	edx,DWORD PTR 52[esp]
+	mov	esi,DWORD PTR 56[esp]
+	mov	edi,DWORD PTR 60[esp]
+	add	eax,DWORD PTR 64[esp]
+	adc	ebx,DWORD PTR 68[esp]
+	xor	ecx,esi
+	xor	edx,edi
+	and	ecx,DWORD PTR 40[esp]
+	and	edx,DWORD PTR 44[esp]
+	add	eax,DWORD PTR 192[esp]
+	adc	ebx,DWORD PTR 196[esp]
+	xor	ecx,esi
+	xor	edx,edi
+	mov	esi,DWORD PTR [ebp]
+	mov	edi,DWORD PTR 4[ebp]
+	add	eax,ecx
+	adc	ebx,edx
+	mov	ecx,DWORD PTR 32[esp]
+	mov	edx,DWORD PTR 36[esp]
+	add	eax,esi
+	adc	ebx,edi
+	mov	DWORD PTR [esp],eax
+	mov	DWORD PTR 4[esp],ebx
+	add	eax,ecx
+	adc	ebx,edx
+	mov	ecx,DWORD PTR 8[esp]
+	mov	edx,DWORD PTR 12[esp]
+	mov	DWORD PTR 32[esp],eax
+	mov	DWORD PTR 36[esp],ebx
+	mov	esi,ecx
+	shr	ecx,2
+	mov	edi,edx
+	shr	edx,2
+	mov	ebx,ecx
+	shl	esi,4
+	mov	eax,edx
+	shl	edi,4
+	xor	ebx,esi
+	shr	ecx,5
+	xor	eax,edi
+	shr	edx,5
+	xor	ebx,ecx
+	shl	esi,21
+	xor	eax,edx
+	shl	edi,21
+	xor	eax,esi
+	shr	ecx,21
+	xor	ebx,edi
+	shr	edx,21
+	xor	eax,ecx
+	shl	esi,5
+	xor	ebx,edx
+	shl	edi,5
+	xor	eax,esi
+	xor	ebx,edi
+	mov	ecx,DWORD PTR 8[esp]
+	mov	edx,DWORD PTR 12[esp]
+	mov	esi,DWORD PTR 16[esp]
+	mov	edi,DWORD PTR 20[esp]
+	add	eax,DWORD PTR [esp]
+	adc	ebx,DWORD PTR 4[esp]
+	or	ecx,esi
+	or	edx,edi
+	and	ecx,DWORD PTR 24[esp]
+	and	edx,DWORD PTR 28[esp]
+	and	esi,DWORD PTR 8[esp]
+	and	edi,DWORD PTR 12[esp]
+	or	ecx,esi
+	or	edx,edi
+	add	eax,ecx
+	adc	ebx,edx
+	mov	DWORD PTR [esp],eax
+	mov	DWORD PTR 4[esp],ebx
+	mov	dl,BYTE PTR [ebp]
+	sub	esp,8
+	lea	ebp,DWORD PTR 8[ebp]
+	cmp	dl,148
+	jne	$L00300_15_x86
+ALIGN	16
+$L00416_79_x86:
+	mov	ecx,DWORD PTR 312[esp]
+	mov	edx,DWORD PTR 316[esp]
+	mov	esi,ecx
+	shr	ecx,1
+	mov	edi,edx
+	shr	edx,1
+	mov	eax,ecx
+	shl	esi,24
+	mov	ebx,edx
+	shl	edi,24
+	xor	ebx,esi
+	shr	ecx,6
+	xor	eax,edi
+	shr	edx,6
+	xor	eax,ecx
+	shl	esi,7
+	xor	ebx,edx
+	shl	edi,1
+	xor	ebx,esi
+	shr	ecx,1
+	xor	eax,edi
+	shr	edx,1
+	xor	eax,ecx
+	shl	edi,6
+	xor	ebx,edx
+	xor	eax,edi
+	mov	DWORD PTR [esp],eax
+	mov	DWORD PTR 4[esp],ebx
+	mov	ecx,DWORD PTR 208[esp]
+	mov	edx,DWORD PTR 212[esp]
+	mov	esi,ecx
+	shr	ecx,6
+	mov	edi,edx
+	shr	edx,6
+	mov	eax,ecx
+	shl	esi,3
+	mov	ebx,edx
+	shl	edi,3
+	xor	eax,esi
+	shr	ecx,13
+	xor	ebx,edi
+	shr	edx,13
+	xor	eax,ecx
+	shl	esi,10
+	xor	ebx,edx
+	shl	edi,10
+	xor	ebx,esi
+	shr	ecx,10
+	xor	eax,edi
+	shr	edx,10
+	xor	ebx,ecx
+	shl	edi,13
+	xor	eax,edx
+	xor	eax,edi
+	mov	ecx,DWORD PTR 320[esp]
+	mov	edx,DWORD PTR 324[esp]
+	add	eax,DWORD PTR [esp]
+	adc	ebx,DWORD PTR 4[esp]
+	mov	esi,DWORD PTR 248[esp]
+	mov	edi,DWORD PTR 252[esp]
+	add	eax,ecx
+	adc	ebx,edx
+	add	eax,esi
+	adc	ebx,edi
+	mov	DWORD PTR 192[esp],eax
+	mov	DWORD PTR 196[esp],ebx
+	mov	ecx,DWORD PTR 40[esp]
+	mov	edx,DWORD PTR 44[esp]
+	mov	esi,ecx
+	shr	ecx,9
+	mov	edi,edx
+	shr	edx,9
+	mov	ebx,ecx
+	shl	esi,14
+	mov	eax,edx
+	shl	edi,14
+	xor	ebx,esi
+	shr	ecx,5
+	xor	eax,edi
+	shr	edx,5
+	xor	eax,ecx
+	shl	esi,4
+	xor	ebx,edx
+	shl	edi,4
+	xor	ebx,esi
+	shr	ecx,4
+	xor	eax,edi
+	shr	edx,4
+	xor	eax,ecx
+	shl	esi,5
+	xor	ebx,edx
+	shl	edi,5
+	xor	eax,esi
+	xor	ebx,edi
+	mov	ecx,DWORD PTR 48[esp]
+	mov	edx,DWORD PTR 52[esp]
+	mov	esi,DWORD PTR 56[esp]
+	mov	edi,DWORD PTR 60[esp]
+	add	eax,DWORD PTR 64[esp]
+	adc	ebx,DWORD PTR 68[esp]
+	xor	ecx,esi
+	xor	edx,edi
+	and	ecx,DWORD PTR 40[esp]
+	and	edx,DWORD PTR 44[esp]
+	add	eax,DWORD PTR 192[esp]
+	adc	ebx,DWORD PTR 196[esp]
+	xor	ecx,esi
+	xor	edx,edi
+	mov	esi,DWORD PTR [ebp]
+	mov	edi,DWORD PTR 4[ebp]
+	add	eax,ecx
+	adc	ebx,edx
+	mov	ecx,DWORD PTR 32[esp]
+	mov	edx,DWORD PTR 36[esp]
+	add	eax,esi
+	adc	ebx,edi
+	mov	DWORD PTR [esp],eax
+	mov	DWORD PTR 4[esp],ebx
+	add	eax,ecx
+	adc	ebx,edx
+	mov	ecx,DWORD PTR 8[esp]
+	mov	edx,DWORD PTR 12[esp]
+	mov	DWORD PTR 32[esp],eax
+	mov	DWORD PTR 36[esp],ebx
+	mov	esi,ecx
+	shr	ecx,2
+	mov	edi,edx
+	shr	edx,2
+	mov	ebx,ecx
+	shl	esi,4
+	mov	eax,edx
+	shl	edi,4
+	xor	ebx,esi
+	shr	ecx,5
+	xor	eax,edi
+	shr	edx,5
+	xor	ebx,ecx
+	shl	esi,21
+	xor	eax,edx
+	shl	edi,21
+	xor	eax,esi
+	shr	ecx,21
+	xor	ebx,edi
+	shr	edx,21
+	xor	eax,ecx
+	shl	esi,5
+	xor	ebx,edx
+	shl	edi,5
+	xor	eax,esi
+	xor	ebx,edi
+	mov	ecx,DWORD PTR 8[esp]
+	mov	edx,DWORD PTR 12[esp]
+	mov	esi,DWORD PTR 16[esp]
+	mov	edi,DWORD PTR 20[esp]
+	add	eax,DWORD PTR [esp]
+	adc	ebx,DWORD PTR 4[esp]
+	or	ecx,esi
+	or	edx,edi
+	and	ecx,DWORD PTR 24[esp]
+	and	edx,DWORD PTR 28[esp]
+	and	esi,DWORD PTR 8[esp]
+	and	edi,DWORD PTR 12[esp]
+	or	ecx,esi
+	or	edx,edi
+	add	eax,ecx
+	adc	ebx,edx
+	mov	DWORD PTR [esp],eax
+	mov	DWORD PTR 4[esp],ebx
+	mov	dl,BYTE PTR [ebp]
+	sub	esp,8
+	lea	ebp,DWORD PTR 8[ebp]
+	cmp	dl,23
+	jne	$L00416_79_x86
+	mov	esi,DWORD PTR 840[esp]
+	mov	edi,DWORD PTR 844[esp]
+	mov	eax,DWORD PTR [esi]
+	mov	ebx,DWORD PTR 4[esi]
+	mov	ecx,DWORD PTR 8[esi]
+	mov	edx,DWORD PTR 12[esi]
+	add	eax,DWORD PTR 8[esp]
+	adc	ebx,DWORD PTR 12[esp]
+	mov	DWORD PTR [esi],eax
+	mov	DWORD PTR 4[esi],ebx
+	add	ecx,DWORD PTR 16[esp]
+	adc	edx,DWORD PTR 20[esp]
+	mov	DWORD PTR 8[esi],ecx
+	mov	DWORD PTR 12[esi],edx
+	mov	eax,DWORD PTR 16[esi]
+	mov	ebx,DWORD PTR 20[esi]
+	mov	ecx,DWORD PTR 24[esi]
+	mov	edx,DWORD PTR 28[esi]
+	add	eax,DWORD PTR 24[esp]
+	adc	ebx,DWORD PTR 28[esp]
+	mov	DWORD PTR 16[esi],eax
+	mov	DWORD PTR 20[esi],ebx
+	add	ecx,DWORD PTR 32[esp]
+	adc	edx,DWORD PTR 36[esp]
+	mov	DWORD PTR 24[esi],ecx
+	mov	DWORD PTR 28[esi],edx
+	mov	eax,DWORD PTR 32[esi]
+	mov	ebx,DWORD PTR 36[esi]
+	mov	ecx,DWORD PTR 40[esi]
+	mov	edx,DWORD PTR 44[esi]
+	add	eax,DWORD PTR 40[esp]
+	adc	ebx,DWORD PTR 44[esp]
+	mov	DWORD PTR 32[esi],eax
+	mov	DWORD PTR 36[esi],ebx
+	add	ecx,DWORD PTR 48[esp]
+	adc	edx,DWORD PTR 52[esp]
+	mov	DWORD PTR 40[esi],ecx
+	mov	DWORD PTR 44[esi],edx
+	mov	eax,DWORD PTR 48[esi]
+	mov	ebx,DWORD PTR 52[esi]
+	mov	ecx,DWORD PTR 56[esi]
+	mov	edx,DWORD PTR 60[esi]
+	add	eax,DWORD PTR 56[esp]
+	adc	ebx,DWORD PTR 60[esp]
+	mov	DWORD PTR 48[esi],eax
+	mov	DWORD PTR 52[esi],ebx
+	add	ecx,DWORD PTR 64[esp]
+	adc	edx,DWORD PTR 68[esp]
+	mov	DWORD PTR 56[esi],ecx
+	mov	DWORD PTR 60[esi],edx
+	add	esp,840
+	sub	ebp,640
+	cmp	edi,DWORD PTR 8[esp]
+	jb	$L002loop_x86
+	mov	esp,DWORD PTR 12[esp]
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+ALIGN	64
+$L001K512:
+DD	3609767458,1116352408
+DD	602891725,1899447441
+DD	3964484399,3049323471
+DD	2173295548,3921009573
+DD	4081628472,961987163
+DD	3053834265,1508970993
+DD	2937671579,2453635748
+DD	3664609560,2870763221
+DD	2734883394,3624381080
+DD	1164996542,310598401
+DD	1323610764,607225278
+DD	3590304994,1426881987
+DD	4068182383,1925078388
+DD	991336113,2162078206
+DD	633803317,2614888103
+DD	3479774868,3248222580
+DD	2666613458,3835390401
+DD	944711139,4022224774
+DD	2341262773,264347078
+DD	2007800933,604807628
+DD	1495990901,770255983
+DD	1856431235,1249150122
+DD	3175218132,1555081692
+DD	2198950837,1996064986
+DD	3999719339,2554220882
+DD	766784016,2821834349
+DD	2566594879,2952996808
+DD	3203337956,3210313671
+DD	1034457026,3336571891
+DD	2466948901,3584528711
+DD	3758326383,113926993
+DD	168717936,338241895
+DD	1188179964,666307205
+DD	1546045734,773529912
+DD	1522805485,1294757372
+DD	2643833823,1396182291
+DD	2343527390,1695183700
+DD	1014477480,1986661051
+DD	1206759142,2177026350
+DD	344077627,2456956037
+DD	1290863460,2730485921
+DD	3158454273,2820302411
+DD	3505952657,3259730800
+DD	106217008,3345764771
+DD	3606008344,3516065817
+DD	1432725776,3600352804
+DD	1467031594,4094571909
+DD	851169720,275423344
+DD	3100823752,430227734
+DD	1363258195,506948616
+DD	3750685593,659060556
+DD	3785050280,883997877
+DD	3318307427,958139571
+DD	3812723403,1322822218
+DD	2003034995,1537002063
+DD	3602036899,1747873779
+DD	1575990012,1955562222
+DD	1125592928,2024104815
+DD	2716904306,2227730452
+DD	442776044,2361852424
+DD	593698344,2428436474
+DD	3733110249,2756734187
+DD	2999351573,3204031479
+DD	3815920427,3329325298
+DD	3928383900,3391569614
+DD	566280711,3515267271
+DD	3454069534,3940187606
+DD	4000239992,4118630271
+DD	1914138554,116418474
+DD	2731055270,174292421
+DD	3203993006,289380356
+DD	320620315,460393269
+DD	587496836,685471733
+DD	1086792851,852142971
+DD	365543100,1017036298
+DD	2618297676,1126000580
+DD	3409855158,1288033470
+DD	4234509866,1501505948
+DD	987167468,1607167915
+DD	1246189591,1816402316
+_sha512_block_data_order ENDP
+DB	83,72,65,53,49,50,32,98,108,111,99,107,32,116,114,97
+DB	110,115,102,111,114,109,32,102,111,114,32,120,56,54,44,32
+DB	67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+DB	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+DB	62,0
+.text$	ENDS
+END

deps/openssl/asm/x86-win32-masm/x86cpuid.asm

@@ -0,0 +1,277 @@
+TITLE	x86cpuid.asm
+IF @Version LT 800
+ECHO MASM version 8.00 or later is strongly recommended.
+ENDIF
+.586
+.MODEL	FLAT
+OPTION	DOTNAME
+IF @Version LT 800
+.text$	SEGMENT PAGE 'CODE'
+ELSE
+.text$	SEGMENT ALIGN(64) 'CODE'
+ENDIF
+ALIGN	16
+_OPENSSL_ia32_cpuid	PROC PUBLIC
+$L_OPENSSL_ia32_cpuid_begin::
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	xor	edx,edx
+	pushfd
+	pop	eax
+	mov	ecx,eax
+	xor	eax,2097152
+	push	eax
+	popfd
+	pushfd
+	pop	eax
+	xor	ecx,eax
+	bt	ecx,21
+	jnc	$L000done
+	xor	eax,eax
+	cpuid
+	mov	edi,eax
+	xor	eax,eax
+	cmp	ebx,1970169159
+	setne	al
+	mov	ebp,eax
+	cmp	edx,1231384169
+	setne	al
+	or	ebp,eax
+	cmp	ecx,1818588270
+	setne	al
+	or	ebp,eax
+	jz	$L001intel
+	cmp	ebx,1752462657
+	setne	al
+	mov	esi,eax
+	cmp	edx,1769238117
+	setne	al
+	or	esi,eax
+	cmp	ecx,1145913699
+	setne	al
+	or	esi,eax
+	jnz	$L001intel
+	mov	eax,2147483648
+	cpuid
+	cmp	eax,2147483656
+	jb	$L001intel
+	mov	eax,2147483656
+	cpuid
+	movzx	esi,cl
+	inc	esi
+	mov	eax,1
+	cpuid
+	bt	edx,28
+	jnc	$L000done
+	shr	ebx,16
+	and	ebx,255
+	cmp	ebx,esi
+	ja	$L000done
+	and	edx,4026531839
+	jmp	$L000done
+$L001intel:
+	cmp	edi,4
+	mov	edi,-1
+	jb	$L002nocacheinfo
+	mov	eax,4
+	mov	ecx,0
+	cpuid
+	mov	edi,eax
+	shr	edi,14
+	and	edi,4095
+$L002nocacheinfo:
+	mov	eax,1
+	cpuid
+	cmp	ebp,0
+	jne	$L003notP4
+	and	ah,15
+	cmp	ah,15
+	jne	$L003notP4
+	or	edx,1048576
+$L003notP4:
+	bt	edx,28
+	jnc	$L000done
+	and	edx,4026531839
+	cmp	edi,0
+	je	$L000done
+	or	edx,268435456
+	shr	ebx,16
+	cmp	bl,1
+	ja	$L000done
+	and	edx,4026531839
+$L000done:
+	mov	eax,edx
+	mov	edx,ecx
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+_OPENSSL_ia32_cpuid ENDP
+;EXTERN	_OPENSSL_ia32cap_P:NEAR
+ALIGN	16
+_OPENSSL_rdtsc	PROC PUBLIC
+$L_OPENSSL_rdtsc_begin::
+	xor	eax,eax
+	xor	edx,edx
+	lea	ecx,DWORD PTR _OPENSSL_ia32cap_P
+	bt	DWORD PTR [ecx],4
+	jnc	$L004notsc
+	rdtsc
+$L004notsc:
+	ret
+_OPENSSL_rdtsc ENDP
+ALIGN	16
+_OPENSSL_instrument_halt	PROC PUBLIC
+$L_OPENSSL_instrument_halt_begin::
+	lea	ecx,DWORD PTR _OPENSSL_ia32cap_P
+	bt	DWORD PTR [ecx],4
+	jnc	$L005nohalt
+DD	2421723150
+	and	eax,3
+	jnz	$L005nohalt
+	pushfd
+	pop	eax
+	bt	eax,9
+	jnc	$L005nohalt
+	rdtsc
+	push	edx
+	push	eax
+	hlt
+	rdtsc
+	sub	eax,DWORD PTR [esp]
+	sbb	edx,DWORD PTR 4[esp]
+	add	esp,8
+	ret
+$L005nohalt:
+	xor	eax,eax
+	xor	edx,edx
+	ret
+_OPENSSL_instrument_halt ENDP
+ALIGN	16
+_OPENSSL_far_spin	PROC PUBLIC
+$L_OPENSSL_far_spin_begin::
+	pushfd
+	pop	eax
+	bt	eax,9
+	jnc	$L006nospin
+	mov	eax,DWORD PTR 4[esp]
+	mov	ecx,DWORD PTR 8[esp]
+DD	2430111262
+	xor	eax,eax
+	mov	edx,DWORD PTR [ecx]
+	jmp	$L007spin
+ALIGN	16
+$L007spin:
+	inc	eax
+	cmp	edx,DWORD PTR [ecx]
+	je	$L007spin
+DD	529567888
+	ret
+$L006nospin:
+	xor	eax,eax
+	xor	edx,edx
+	ret
+_OPENSSL_far_spin ENDP
+ALIGN	16
+_OPENSSL_wipe_cpu	PROC PUBLIC
+$L_OPENSSL_wipe_cpu_begin::
+	xor	eax,eax
+	xor	edx,edx
+	lea	ecx,DWORD PTR _OPENSSL_ia32cap_P
+	mov	ecx,DWORD PTR [ecx]
+	bt	DWORD PTR [ecx],1
+	jnc	$L008no_x87
+DD	4007259865,4007259865,4007259865,4007259865,2430851995
+$L008no_x87:
+	lea	eax,DWORD PTR 4[esp]
+	ret
+_OPENSSL_wipe_cpu ENDP
+ALIGN	16
+_OPENSSL_atomic_add	PROC PUBLIC
+$L_OPENSSL_atomic_add_begin::
+	mov	edx,DWORD PTR 4[esp]
+	mov	ecx,DWORD PTR 8[esp]
+	push	ebx
+	nop
+	mov	eax,DWORD PTR [edx]
+$L009spin:
+	lea	ebx,DWORD PTR [ecx*1+eax]
+	nop
+DD	447811568
+	jne	$L009spin
+	mov	eax,ebx
+	pop	ebx
+	ret
+_OPENSSL_atomic_add ENDP
+ALIGN	16
+_OPENSSL_indirect_call	PROC PUBLIC
+$L_OPENSSL_indirect_call_begin::
+	push	ebp
+	mov	ebp,esp
+	sub	esp,28
+	mov	ecx,DWORD PTR 12[ebp]
+	mov	DWORD PTR [esp],ecx
+	mov	edx,DWORD PTR 16[ebp]
+	mov	DWORD PTR 4[esp],edx
+	mov	eax,DWORD PTR 20[ebp]
+	mov	DWORD PTR 8[esp],eax
+	mov	eax,DWORD PTR 24[ebp]
+	mov	DWORD PTR 12[esp],eax
+	mov	eax,DWORD PTR 28[ebp]
+	mov	DWORD PTR 16[esp],eax
+	mov	eax,DWORD PTR 32[ebp]
+	mov	DWORD PTR 20[esp],eax
+	mov	eax,DWORD PTR 36[ebp]
+	mov	DWORD PTR 24[esp],eax
+	call	DWORD PTR 8[ebp]
+	mov	esp,ebp
+	pop	ebp
+	ret
+_OPENSSL_indirect_call ENDP
+ALIGN	16
+_OPENSSL_cleanse	PROC PUBLIC
+$L_OPENSSL_cleanse_begin::
+	mov	edx,DWORD PTR 4[esp]
+	mov	ecx,DWORD PTR 8[esp]
+	xor	eax,eax
+	cmp	ecx,7
+	jae	$L010lot
+	cmp	ecx,0
+	je	$L011ret
+$L012little:
+	mov	BYTE PTR [edx],al
+	sub	ecx,1
+	lea	edx,DWORD PTR 1[edx]
+	jnz	$L012little
+$L011ret:
+	ret
+ALIGN	16
+$L010lot:
+	test	edx,3
+	jz	$L013aligned
+	mov	BYTE PTR [edx],al
+	lea	ecx,DWORD PTR [ecx-1]
+	lea	edx,DWORD PTR 1[edx]
+	jmp	$L010lot
+$L013aligned:
+	mov	DWORD PTR [edx],eax
+	lea	ecx,DWORD PTR [ecx-4]
+	test	ecx,-4
+	lea	edx,DWORD PTR 4[edx]
+	jnz	$L013aligned
+	cmp	ecx,0
+	jne	$L012little
+	ret
+_OPENSSL_cleanse ENDP
+.text$	ENDS
+.bss	SEGMENT 'BSS'
+COMM	_OPENSSL_ia32cap_P:DWORD
+.bss	ENDS
+.CRT$XCU	SEGMENT DWORD PUBLIC 'DATA'
+EXTERN	_OPENSSL_cpuid_setup:NEAR
+DD	_OPENSSL_cpuid_setup
+.CRT$XCU	ENDS
+END