Source release 14.0.0

third_party/boringssl/legacy_kit/mac-x86/crypto/aes/vpaes-x86.S

@@ -0,0 +1,650 @@
+#if defined(__i386__)
+.file	"vpaes-x86.S"
+.text
+.align	6,0x90
+L_vpaes_consts:
+.long	218628480,235210255,168496130,67568393
+.long	252381056,17041926,33884169,51187212
+.long	252645135,252645135,252645135,252645135
+.long	1512730624,3266504856,1377990664,3401244816
+.long	830229760,1275146365,2969422977,3447763452
+.long	3411033600,2979783055,338359620,2782886510
+.long	4209124096,907596821,221174255,1006095553
+.long	191964160,3799684038,3164090317,1589111125
+.long	182528256,1777043520,2877432650,3265356744
+.long	1874708224,3503451415,3305285752,363511674
+.long	1606117888,3487855781,1093350906,2384367825
+.long	197121,67569157,134941193,202313229
+.long	67569157,134941193,202313229,197121
+.long	134941193,202313229,197121,67569157
+.long	202313229,197121,67569157,134941193
+.long	33619971,100992007,168364043,235736079
+.long	235736079,33619971,100992007,168364043
+.long	168364043,235736079,33619971,100992007
+.long	100992007,168364043,235736079,33619971
+.long	50462976,117835012,185207048,252579084
+.long	252314880,51251460,117574920,184942860
+.long	184682752,252054788,50987272,118359308
+.long	118099200,185467140,251790600,50727180
+.long	2946363062,528716217,1300004225,1881839624
+.long	1532713819,1532713819,1532713819,1532713819
+.long	3602276352,4288629033,3737020424,4153884961
+.long	1354558464,32357713,2958822624,3775749553
+.long	1201988352,132424512,1572796698,503232858
+.long	2213177600,1597421020,4103937655,675398315
+.long	2749646592,4273543773,1511898873,121693092
+.long	3040248576,1103263732,2871565598,1608280554
+.long	2236667136,2588920351,482954393,64377734
+.long	3069987328,291237287,2117370568,3650299247
+.long	533321216,3573750986,2572112006,1401264716
+.long	1339849704,2721158661,548607111,3445553514
+.long	2128193280,3054596040,2183486460,1257083700
+.long	655635200,1165381986,3923443150,2344132524
+.long	190078720,256924420,290342170,357187870
+.long	1610966272,2263057382,4103205268,309794674
+.long	2592527872,2233205587,1335446729,3402964816
+.long	3973531904,3225098121,3002836325,1918774430
+.long	3870401024,2102906079,2284471353,4117666579
+.long	617007872,1021508343,366931923,691083277
+.long	2528395776,3491914898,2968704004,1613121270
+.long	3445188352,3247741094,844474987,4093578302
+.long	651481088,1190302358,1689581232,574775300
+.long	4289380608,206939853,2555985458,2489840491
+.long	2130264064,327674451,3566485037,3349835193
+.long	2470714624,316102159,3636825756,3393945945
+.byte	86,101,99,116,111,114,32,80,101,114,109,117,116,97,116,105
+.byte	111,110,32,65,69,83,32,102,111,114,32,120,56,54,47,83
+.byte	83,83,69,51,44,32,77,105,107,101,32,72,97,109,98,117
+.byte	114,103,32,40,83,116,97,110,102,111,114,100,32,85,110,105
+.byte	118,101,114,115,105,116,121,41,0
+.align	6,0x90
+.private_extern	__vpaes_preheat
+.align	4
+__vpaes_preheat:
+	addl	(%esp),%ebp
+	movdqa	-48(%ebp),%xmm7
+	movdqa	-16(%ebp),%xmm6
+	ret
+.private_extern	__vpaes_encrypt_core
+.align	4
+__vpaes_encrypt_core:
+	movl	$16,%ecx
+	movl	240(%edx),%eax
+	movdqa	%xmm6,%xmm1
+	movdqa	(%ebp),%xmm2
+	pandn	%xmm0,%xmm1
+	pand	%xmm6,%xmm0
+	movdqu	(%edx),%xmm5
+.byte	102,15,56,0,208
+	movdqa	16(%ebp),%xmm0
+	pxor	%xmm5,%xmm2
+	psrld	$4,%xmm1
+	addl	$16,%edx
+.byte	102,15,56,0,193
+	leal	192(%ebp),%ebx
+	pxor	%xmm2,%xmm0
+	jmp	L000enc_entry
+.align	4,0x90
+L001enc_loop:
+	movdqa	32(%ebp),%xmm4
+	movdqa	48(%ebp),%xmm0
+.byte	102,15,56,0,226
+.byte	102,15,56,0,195
+	pxor	%xmm5,%xmm4
+	movdqa	64(%ebp),%xmm5
+	pxor	%xmm4,%xmm0
+	movdqa	-64(%ebx,%ecx,1),%xmm1
+.byte	102,15,56,0,234
+	movdqa	80(%ebp),%xmm2
+	movdqa	(%ebx,%ecx,1),%xmm4
+.byte	102,15,56,0,211
+	movdqa	%xmm0,%xmm3
+	pxor	%xmm5,%xmm2
+.byte	102,15,56,0,193
+	addl	$16,%edx
+	pxor	%xmm2,%xmm0
+.byte	102,15,56,0,220
+	addl	$16,%ecx
+	pxor	%xmm0,%xmm3
+.byte	102,15,56,0,193
+	andl	$48,%ecx
+	subl	$1,%eax
+	pxor	%xmm3,%xmm0
+L000enc_entry:
+	movdqa	%xmm6,%xmm1
+	movdqa	-32(%ebp),%xmm5
+	pandn	%xmm0,%xmm1
+	psrld	$4,%xmm1
+	pand	%xmm6,%xmm0
+.byte	102,15,56,0,232
+	movdqa	%xmm7,%xmm3
+	pxor	%xmm1,%xmm0
+.byte	102,15,56,0,217
+	movdqa	%xmm7,%xmm4
+	pxor	%xmm5,%xmm3
+.byte	102,15,56,0,224
+	movdqa	%xmm7,%xmm2
+	pxor	%xmm5,%xmm4
+.byte	102,15,56,0,211
+	movdqa	%xmm7,%xmm3
+	pxor	%xmm0,%xmm2
+.byte	102,15,56,0,220
+	movdqu	(%edx),%xmm5
+	pxor	%xmm1,%xmm3
+	jnz	L001enc_loop
+	movdqa	96(%ebp),%xmm4
+	movdqa	112(%ebp),%xmm0
+.byte	102,15,56,0,226
+	pxor	%xmm5,%xmm4
+.byte	102,15,56,0,195
+	movdqa	64(%ebx,%ecx,1),%xmm1
+	pxor	%xmm4,%xmm0
+.byte	102,15,56,0,193
+	ret
+.private_extern	__vpaes_decrypt_core
+.align	4
+__vpaes_decrypt_core:
+	leal	608(%ebp),%ebx
+	movl	240(%edx),%eax
+	movdqa	%xmm6,%xmm1
+	movdqa	-64(%ebx),%xmm2
+	pandn	%xmm0,%xmm1
+	movl	%eax,%ecx
+	psrld	$4,%xmm1
+	movdqu	(%edx),%xmm5
+	shll	$4,%ecx
+	pand	%xmm6,%xmm0
+.byte	102,15,56,0,208
+	movdqa	-48(%ebx),%xmm0
+	xorl	$48,%ecx
+.byte	102,15,56,0,193
+	andl	$48,%ecx
+	pxor	%xmm5,%xmm2
+	movdqa	176(%ebp),%xmm5
+	pxor	%xmm2,%xmm0
+	addl	$16,%edx
+	leal	-352(%ebx,%ecx,1),%ecx
+	jmp	L002dec_entry
+.align	4,0x90
+L003dec_loop:
+	movdqa	-32(%ebx),%xmm4
+	movdqa	-16(%ebx),%xmm1
+.byte	102,15,56,0,226
+.byte	102,15,56,0,203
+	pxor	%xmm4,%xmm0
+	movdqa	(%ebx),%xmm4
+	pxor	%xmm1,%xmm0
+	movdqa	16(%ebx),%xmm1
+.byte	102,15,56,0,226
+.byte	102,15,56,0,197
+.byte	102,15,56,0,203
+	pxor	%xmm4,%xmm0
+	movdqa	32(%ebx),%xmm4
+	pxor	%xmm1,%xmm0
+	movdqa	48(%ebx),%xmm1
+.byte	102,15,56,0,226
+.byte	102,15,56,0,197
+.byte	102,15,56,0,203
+	pxor	%xmm4,%xmm0
+	movdqa	64(%ebx),%xmm4
+	pxor	%xmm1,%xmm0
+	movdqa	80(%ebx),%xmm1
+.byte	102,15,56,0,226
+.byte	102,15,56,0,197
+.byte	102,15,56,0,203
+	pxor	%xmm4,%xmm0
+	addl	$16,%edx
+.byte	102,15,58,15,237,12
+	pxor	%xmm1,%xmm0
+	subl	$1,%eax
+L002dec_entry:
+	movdqa	%xmm6,%xmm1
+	movdqa	-32(%ebp),%xmm2
+	pandn	%xmm0,%xmm1
+	pand	%xmm6,%xmm0
+	psrld	$4,%xmm1
+.byte	102,15,56,0,208
+	movdqa	%xmm7,%xmm3
+	pxor	%xmm1,%xmm0
+.byte	102,15,56,0,217
+	movdqa	%xmm7,%xmm4
+	pxor	%xmm2,%xmm3
+.byte	102,15,56,0,224
+	pxor	%xmm2,%xmm4
+	movdqa	%xmm7,%xmm2
+.byte	102,15,56,0,211
+	movdqa	%xmm7,%xmm3
+	pxor	%xmm0,%xmm2
+.byte	102,15,56,0,220
+	movdqu	(%edx),%xmm0
+	pxor	%xmm1,%xmm3
+	jnz	L003dec_loop
+	movdqa	96(%ebx),%xmm4
+.byte	102,15,56,0,226
+	pxor	%xmm0,%xmm4
+	movdqa	112(%ebx),%xmm0
+	movdqa	(%ecx),%xmm2
+.byte	102,15,56,0,195
+	pxor	%xmm4,%xmm0
+.byte	102,15,56,0,194
+	ret
+.private_extern	__vpaes_schedule_core
+.align	4
+__vpaes_schedule_core:
+	addl	(%esp),%ebp
+	movdqu	(%esi),%xmm0
+	movdqa	320(%ebp),%xmm2
+	movdqa	%xmm0,%xmm3
+	leal	(%ebp),%ebx
+	movdqa	%xmm2,4(%esp)
+	call	__vpaes_schedule_transform
+	movdqa	%xmm0,%xmm7
+	testl	%edi,%edi
+	jnz	L004schedule_am_decrypting
+	movdqu	%xmm0,(%edx)
+	jmp	L005schedule_go
+L004schedule_am_decrypting:
+	movdqa	256(%ebp,%ecx,1),%xmm1
+.byte	102,15,56,0,217
+	movdqu	%xmm3,(%edx)
+	xorl	$48,%ecx
+L005schedule_go:
+	cmpl	$192,%eax
+	ja	L006schedule_256
+	je	L007schedule_192
+L008schedule_128:
+	movl	$10,%eax
+L009loop_schedule_128:
+	call	__vpaes_schedule_round
+	decl	%eax
+	jz	L010schedule_mangle_last
+	call	__vpaes_schedule_mangle
+	jmp	L009loop_schedule_128
+.align	4,0x90
+L007schedule_192:
+	movdqu	8(%esi),%xmm0
+	call	__vpaes_schedule_transform
+	movdqa	%xmm0,%xmm6
+	pxor	%xmm4,%xmm4
+	movhlps	%xmm4,%xmm6
+	movl	$4,%eax
+L011loop_schedule_192:
+	call	__vpaes_schedule_round
+.byte	102,15,58,15,198,8
+	call	__vpaes_schedule_mangle
+	call	__vpaes_schedule_192_smear
+	call	__vpaes_schedule_mangle
+	call	__vpaes_schedule_round
+	decl	%eax
+	jz	L010schedule_mangle_last
+	call	__vpaes_schedule_mangle
+	call	__vpaes_schedule_192_smear
+	jmp	L011loop_schedule_192
+.align	4,0x90
+L006schedule_256:
+	movdqu	16(%esi),%xmm0
+	call	__vpaes_schedule_transform
+	movl	$7,%eax
+L012loop_schedule_256:
+	call	__vpaes_schedule_mangle
+	movdqa	%xmm0,%xmm6
+	call	__vpaes_schedule_round
+	decl	%eax
+	jz	L010schedule_mangle_last
+	call	__vpaes_schedule_mangle
+	pshufd	$255,%xmm0,%xmm0
+	movdqa	%xmm7,20(%esp)
+	movdqa	%xmm6,%xmm7
+	call	L_vpaes_schedule_low_round
+	movdqa	20(%esp),%xmm7
+	jmp	L012loop_schedule_256
+.align	4,0x90
+L010schedule_mangle_last:
+	leal	384(%ebp),%ebx
+	testl	%edi,%edi
+	jnz	L013schedule_mangle_last_dec
+	movdqa	256(%ebp,%ecx,1),%xmm1
+.byte	102,15,56,0,193
+	leal	352(%ebp),%ebx
+	addl	$32,%edx
+L013schedule_mangle_last_dec:
+	addl	$-16,%edx
+	pxor	336(%ebp),%xmm0
+	call	__vpaes_schedule_transform
+	movdqu	%xmm0,(%edx)
+	pxor	%xmm0,%xmm0
+	pxor	%xmm1,%xmm1
+	pxor	%xmm2,%xmm2
+	pxor	%xmm3,%xmm3
+	pxor	%xmm4,%xmm4
+	pxor	%xmm5,%xmm5
+	pxor	%xmm6,%xmm6
+	pxor	%xmm7,%xmm7
+	ret
+.private_extern	__vpaes_schedule_192_smear
+.align	4
+__vpaes_schedule_192_smear:
+	pshufd	$128,%xmm6,%xmm1
+	pshufd	$254,%xmm7,%xmm0
+	pxor	%xmm1,%xmm6
+	pxor	%xmm1,%xmm1
+	pxor	%xmm0,%xmm6
+	movdqa	%xmm6,%xmm0
+	movhlps	%xmm1,%xmm6
+	ret
+.private_extern	__vpaes_schedule_round
+.align	4
+__vpaes_schedule_round:
+	movdqa	8(%esp),%xmm2
+	pxor	%xmm1,%xmm1
+.byte	102,15,58,15,202,15
+.byte	102,15,58,15,210,15
+	pxor	%xmm1,%xmm7
+	pshufd	$255,%xmm0,%xmm0
+.byte	102,15,58,15,192,1
+	movdqa	%xmm2,8(%esp)
+L_vpaes_schedule_low_round:
+	movdqa	%xmm7,%xmm1
+	pslldq	$4,%xmm7
+	pxor	%xmm1,%xmm7
+	movdqa	%xmm7,%xmm1
+	pslldq	$8,%xmm7
+	pxor	%xmm1,%xmm7
+	pxor	336(%ebp),%xmm7
+	movdqa	-16(%ebp),%xmm4
+	movdqa	-48(%ebp),%xmm5
+	movdqa	%xmm4,%xmm1
+	pandn	%xmm0,%xmm1
+	psrld	$4,%xmm1
+	pand	%xmm4,%xmm0
+	movdqa	-32(%ebp),%xmm2
+.byte	102,15,56,0,208
+	pxor	%xmm1,%xmm0
+	movdqa	%xmm5,%xmm3
+.byte	102,15,56,0,217
+	pxor	%xmm2,%xmm3
+	movdqa	%xmm5,%xmm4
+.byte	102,15,56,0,224
+	pxor	%xmm2,%xmm4
+	movdqa	%xmm5,%xmm2
+.byte	102,15,56,0,211
+	pxor	%xmm0,%xmm2
+	movdqa	%xmm5,%xmm3
+.byte	102,15,56,0,220
+	pxor	%xmm1,%xmm3
+	movdqa	32(%ebp),%xmm4
+.byte	102,15,56,0,226
+	movdqa	48(%ebp),%xmm0
+.byte	102,15,56,0,195
+	pxor	%xmm4,%xmm0
+	pxor	%xmm7,%xmm0
+	movdqa	%xmm0,%xmm7
+	ret
+.private_extern	__vpaes_schedule_transform
+.align	4
+__vpaes_schedule_transform:
+	movdqa	-16(%ebp),%xmm2
+	movdqa	%xmm2,%xmm1
+	pandn	%xmm0,%xmm1
+	psrld	$4,%xmm1
+	pand	%xmm2,%xmm0
+	movdqa	(%ebx),%xmm2
+.byte	102,15,56,0,208
+	movdqa	16(%ebx),%xmm0
+.byte	102,15,56,0,193
+	pxor	%xmm2,%xmm0
+	ret
+.private_extern	__vpaes_schedule_mangle
+.align	4
+__vpaes_schedule_mangle:
+	movdqa	%xmm0,%xmm4
+	movdqa	128(%ebp),%xmm5
+	testl	%edi,%edi
+	jnz	L014schedule_mangle_dec
+	addl	$16,%edx
+	pxor	336(%ebp),%xmm4
+.byte	102,15,56,0,229
+	movdqa	%xmm4,%xmm3
+.byte	102,15,56,0,229
+	pxor	%xmm4,%xmm3
+.byte	102,15,56,0,229
+	pxor	%xmm4,%xmm3
+	jmp	L015schedule_mangle_both
+.align	4,0x90
+L014schedule_mangle_dec:
+	movdqa	-16(%ebp),%xmm2
+	leal	416(%ebp),%esi
+	movdqa	%xmm2,%xmm1
+	pandn	%xmm4,%xmm1
+	psrld	$4,%xmm1
+	pand	%xmm2,%xmm4
+	movdqa	(%esi),%xmm2
+.byte	102,15,56,0,212
+	movdqa	16(%esi),%xmm3
+.byte	102,15,56,0,217
+	pxor	%xmm2,%xmm3
+.byte	102,15,56,0,221
+	movdqa	32(%esi),%xmm2
+.byte	102,15,56,0,212
+	pxor	%xmm3,%xmm2
+	movdqa	48(%esi),%xmm3
+.byte	102,15,56,0,217
+	pxor	%xmm2,%xmm3
+.byte	102,15,56,0,221
+	movdqa	64(%esi),%xmm2
+.byte	102,15,56,0,212
+	pxor	%xmm3,%xmm2
+	movdqa	80(%esi),%xmm3
+.byte	102,15,56,0,217
+	pxor	%xmm2,%xmm3
+.byte	102,15,56,0,221
+	movdqa	96(%esi),%xmm2
+.byte	102,15,56,0,212
+	pxor	%xmm3,%xmm2
+	movdqa	112(%esi),%xmm3
+.byte	102,15,56,0,217
+	pxor	%xmm2,%xmm3
+	addl	$-16,%edx
+L015schedule_mangle_both:
+	movdqa	256(%ebp,%ecx,1),%xmm1
+.byte	102,15,56,0,217
+	addl	$-16,%ecx
+	andl	$48,%ecx
+	movdqu	%xmm3,(%edx)
+	ret
+.globl	_vpaes_set_encrypt_key
+.private_extern	_vpaes_set_encrypt_key
+.align	4
+_vpaes_set_encrypt_key:
+L_vpaes_set_encrypt_key_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	leal	-56(%esp),%ebx
+	movl	24(%esp),%eax
+	andl	$-16,%ebx
+	movl	28(%esp),%edx
+	xchgl	%esp,%ebx
+	movl	%ebx,48(%esp)
+	movl	%eax,%ebx
+	shrl	$5,%ebx
+	addl	$5,%ebx
+	movl	%ebx,240(%edx)
+	movl	$48,%ecx
+	movl	$0,%edi
+	leal	L_vpaes_consts+0x30-L016pic_point,%ebp
+	call	__vpaes_schedule_core
+L016pic_point:
+	movl	48(%esp),%esp
+	xorl	%eax,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_vpaes_set_decrypt_key
+.private_extern	_vpaes_set_decrypt_key
+.align	4
+_vpaes_set_decrypt_key:
+L_vpaes_set_decrypt_key_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	leal	-56(%esp),%ebx
+	movl	24(%esp),%eax
+	andl	$-16,%ebx
+	movl	28(%esp),%edx
+	xchgl	%esp,%ebx
+	movl	%ebx,48(%esp)
+	movl	%eax,%ebx
+	shrl	$5,%ebx
+	addl	$5,%ebx
+	movl	%ebx,240(%edx)
+	shll	$4,%ebx
+	leal	16(%edx,%ebx,1),%edx
+	movl	$1,%edi
+	movl	%eax,%ecx
+	shrl	$1,%ecx
+	andl	$32,%ecx
+	xorl	$32,%ecx
+	leal	L_vpaes_consts+0x30-L017pic_point,%ebp
+	call	__vpaes_schedule_core
+L017pic_point:
+	movl	48(%esp),%esp
+	xorl	%eax,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_vpaes_encrypt
+.private_extern	_vpaes_encrypt
+.align	4
+_vpaes_encrypt:
+L_vpaes_encrypt_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	leal	L_vpaes_consts+0x30-L018pic_point,%ebp
+	call	__vpaes_preheat
+L018pic_point:
+	movl	20(%esp),%esi
+	leal	-56(%esp),%ebx
+	movl	24(%esp),%edi
+	andl	$-16,%ebx
+	movl	28(%esp),%edx
+	xchgl	%esp,%ebx
+	movl	%ebx,48(%esp)
+	movdqu	(%esi),%xmm0
+	call	__vpaes_encrypt_core
+	movdqu	%xmm0,(%edi)
+	movl	48(%esp),%esp
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_vpaes_decrypt
+.private_extern	_vpaes_decrypt
+.align	4
+_vpaes_decrypt:
+L_vpaes_decrypt_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	leal	L_vpaes_consts+0x30-L019pic_point,%ebp
+	call	__vpaes_preheat
+L019pic_point:
+	movl	20(%esp),%esi
+	leal	-56(%esp),%ebx
+	movl	24(%esp),%edi
+	andl	$-16,%ebx
+	movl	28(%esp),%edx
+	xchgl	%esp,%ebx
+	movl	%ebx,48(%esp)
+	movdqu	(%esi),%xmm0
+	call	__vpaes_decrypt_core
+	movdqu	%xmm0,(%edi)
+	movl	48(%esp),%esp
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_vpaes_cbc_encrypt
+.private_extern	_vpaes_cbc_encrypt
+.align	4
+_vpaes_cbc_encrypt:
+L_vpaes_cbc_encrypt_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	20(%esp),%esi
+	movl	24(%esp),%edi
+	movl	28(%esp),%eax
+	movl	32(%esp),%edx
+	subl	$16,%eax
+	jc	L020cbc_abort
+	leal	-56(%esp),%ebx
+	movl	36(%esp),%ebp
+	andl	$-16,%ebx
+	movl	40(%esp),%ecx
+	xchgl	%esp,%ebx
+	movdqu	(%ebp),%xmm1
+	subl	%esi,%edi
+	movl	%ebx,48(%esp)
+	movl	%edi,(%esp)
+	movl	%edx,4(%esp)
+	movl	%ebp,8(%esp)
+	movl	%eax,%edi
+	leal	L_vpaes_consts+0x30-L021pic_point,%ebp
+	call	__vpaes_preheat
+L021pic_point:
+	cmpl	$0,%ecx
+	je	L022cbc_dec_loop
+	jmp	L023cbc_enc_loop
+.align	4,0x90
+L023cbc_enc_loop:
+	movdqu	(%esi),%xmm0
+	pxor	%xmm1,%xmm0
+	call	__vpaes_encrypt_core
+	movl	(%esp),%ebx
+	movl	4(%esp),%edx
+	movdqa	%xmm0,%xmm1
+	movdqu	%xmm0,(%ebx,%esi,1)
+	leal	16(%esi),%esi
+	subl	$16,%edi
+	jnc	L023cbc_enc_loop
+	jmp	L024cbc_done
+.align	4,0x90
+L022cbc_dec_loop:
+	movdqu	(%esi),%xmm0
+	movdqa	%xmm1,16(%esp)
+	movdqa	%xmm0,32(%esp)
+	call	__vpaes_decrypt_core
+	movl	(%esp),%ebx
+	movl	4(%esp),%edx
+	pxor	16(%esp),%xmm0
+	movdqa	32(%esp),%xmm1
+	movdqu	%xmm0,(%ebx,%esi,1)
+	leal	16(%esi),%esi
+	subl	$16,%edi
+	jnc	L022cbc_dec_loop
+L024cbc_done:
+	movl	8(%esp),%ebx
+	movl	48(%esp),%esp
+	movdqu	%xmm1,(%ebx)
+L020cbc_abort:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+#endif

third_party/boringssl/legacy_kit/mac-x86/crypto/bn/x86-mont.S

@@ -0,0 +1,462 @@
+#if defined(__i386__)
+.file	"src/crypto/bn/asm/x86-mont.S"
+.text
+.globl	_bn_mul_mont
+.private_extern	_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	$-64,%esp
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	16(%esi),%esi
+	movl	(%esi),%esi
+	movl	%eax,4(%esp)
+	movl	%ebx,8(%esp)
+	movl	%ecx,12(%esp)
+	movl	%edx,16(%esp)
+	movl	%esi,20(%esp)
+	leal	-3(%edi),%ebx
+	movl	%ebp,24(%esp)
+	call	L001PIC_me_up
+L001PIC_me_up:
+	popl	%eax
+	movl	L_OPENSSL_ia32cap_P$non_lazy_ptr-L001PIC_me_up(%eax),%eax
+	btl	$26,(%eax)
+	jnc	L002non_sse2
+	movl	$-1,%eax
+	movd	%eax,%mm7
+	movl	8(%esp),%esi
+	movl	12(%esp),%edi
+	movl	16(%esp),%ebp
+	xorl	%edx,%edx
+	xorl	%ecx,%ecx
+	movd	(%edi),%mm4
+	movd	(%esi),%mm5
+	movd	(%ebp),%mm3
+	pmuludq	%mm4,%mm5
+	movq	%mm5,%mm2
+	movq	%mm5,%mm0
+	pand	%mm7,%mm0
+	pmuludq	20(%esp),%mm5
+	pmuludq	%mm5,%mm3
+	paddq	%mm0,%mm3
+	movd	4(%ebp),%mm1
+	movd	4(%esi),%mm0
+	psrlq	$32,%mm2
+	psrlq	$32,%mm3
+	incl	%ecx
+.align	4,0x90
+L0031st:
+	pmuludq	%mm4,%mm0
+	pmuludq	%mm5,%mm1
+	paddq	%mm0,%mm2
+	paddq	%mm1,%mm3
+	movq	%mm2,%mm0
+	pand	%mm7,%mm0
+	movd	4(%ebp,%ecx,4),%mm1
+	paddq	%mm0,%mm3
+	movd	4(%esi,%ecx,4),%mm0
+	psrlq	$32,%mm2
+	movd	%mm3,28(%esp,%ecx,4)
+	psrlq	$32,%mm3
+	leal	1(%ecx),%ecx
+	cmpl	%ebx,%ecx
+	jl	L0031st
+	pmuludq	%mm4,%mm0
+	pmuludq	%mm5,%mm1
+	paddq	%mm0,%mm2
+	paddq	%mm1,%mm3
+	movq	%mm2,%mm0
+	pand	%mm7,%mm0
+	paddq	%mm0,%mm3
+	movd	%mm3,28(%esp,%ecx,4)
+	psrlq	$32,%mm2
+	psrlq	$32,%mm3
+	paddq	%mm2,%mm3
+	movq	%mm3,32(%esp,%ebx,4)
+	incl	%edx
+L004outer:
+	xorl	%ecx,%ecx
+	movd	(%edi,%edx,4),%mm4
+	movd	(%esi),%mm5
+	movd	32(%esp),%mm6
+	movd	(%ebp),%mm3
+	pmuludq	%mm4,%mm5
+	paddq	%mm6,%mm5
+	movq	%mm5,%mm0
+	movq	%mm5,%mm2
+	pand	%mm7,%mm0
+	pmuludq	20(%esp),%mm5
+	pmuludq	%mm5,%mm3
+	paddq	%mm0,%mm3
+	movd	36(%esp),%mm6
+	movd	4(%ebp),%mm1
+	movd	4(%esi),%mm0
+	psrlq	$32,%mm2
+	psrlq	$32,%mm3
+	paddq	%mm6,%mm2
+	incl	%ecx
+	decl	%ebx
+L005inner:
+	pmuludq	%mm4,%mm0
+	pmuludq	%mm5,%mm1
+	paddq	%mm0,%mm2
+	paddq	%mm1,%mm3
+	movq	%mm2,%mm0
+	movd	36(%esp,%ecx,4),%mm6
+	pand	%mm7,%mm0
+	movd	4(%ebp,%ecx,4),%mm1
+	paddq	%mm0,%mm3
+	movd	4(%esi,%ecx,4),%mm0
+	psrlq	$32,%mm2
+	movd	%mm3,28(%esp,%ecx,4)
+	psrlq	$32,%mm3
+	paddq	%mm6,%mm2
+	decl	%ebx
+	leal	1(%ecx),%ecx
+	jnz	L005inner
+	movl	%ecx,%ebx
+	pmuludq	%mm4,%mm0
+	pmuludq	%mm5,%mm1
+	paddq	%mm0,%mm2
+	paddq	%mm1,%mm3
+	movq	%mm2,%mm0
+	pand	%mm7,%mm0
+	paddq	%mm0,%mm3
+	movd	%mm3,28(%esp,%ecx,4)
+	psrlq	$32,%mm2
+	psrlq	$32,%mm3
+	movd	36(%esp,%ebx,4),%mm6
+	paddq	%mm2,%mm3
+	paddq	%mm6,%mm3
+	movq	%mm3,32(%esp,%ebx,4)
+	leal	1(%edx),%edx
+	cmpl	%ebx,%edx
+	jle	L004outer
+	emms
+	jmp	L006common_tail
+.align	4,0x90
+L002non_sse2:
+	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	L007bn_sqr_mont
+	movl	%eax,28(%esp)
+	movl	(%esi),%eax
+	xorl	%edx,%edx
+.align	4,0x90
+L008mull:
+	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	L008mull
+	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	L0092ndmadd
+.align	4,0x90
+L0101stmadd:
+	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	L0101stmadd
+	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
+L0092ndmadd:
+	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	L0092ndmadd
+	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	L006common_tail
+	movl	(%ecx),%edi
+	movl	8(%esp),%esi
+	movl	%ecx,12(%esp)
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	movl	(%esi),%eax
+	jmp	L0101stmadd
+.align	4,0x90
+L007bn_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
+L011sqr:
+	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	L011sqr
+	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
+L0123rdmadd:
+	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	L0123rdmadd
+	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	L006common_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	L013sqrlast
+	movl	%edx,%ebx
+	shrl	$1,%edx
+	andl	$1,%ebx
+.align	4,0x90
+L014sqradd:
+	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	L014sqradd
+	movl	%edx,%ebp
+	addl	%edx,%edx
+	shrl	$31,%ebp
+	addl	%ebx,%edx
+	adcl	$0,%ebp
+L013sqrlast:
+	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	L0123rdmadd
+.align	4,0x90
+L006common_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
+L015sub:
+	sbbl	(%ebp,%edx,4),%eax
+	movl	%eax,(%edi,%edx,4)
+	decl	%ecx
+	movl	4(%esi,%edx,4),%eax
+	leal	1(%edx),%edx
+	jge	L015sub
+	sbbl	$0,%eax
+.align	4,0x90
+L016copy:
+	movl	(%esi,%ebx,4),%edx
+	movl	(%edi,%ebx,4),%ebp
+	xorl	%ebp,%edx
+	andl	%eax,%edx
+	xorl	%ebp,%edx
+	movl	%ecx,(%esi,%ebx,4)
+	movl	%edx,(%edi,%ebx,4)
+	decl	%ebx
+	jge	L016copy
+	movl	24(%esp),%esp
+	movl	$1,%eax
+L000just_leave:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.byte	77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
+.byte	112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
+.byte	54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
+.byte	32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
+.byte	111,114,103,62,0
+.section __IMPORT,__pointers,non_lazy_symbol_pointers
+L_OPENSSL_ia32cap_P$non_lazy_ptr:
+.indirect_symbol	_OPENSSL_ia32cap_P
+.long	0
+#endif

third_party/boringssl/legacy_kit/mac-x86/crypto/chacha/chacha-x86.S

@@ -0,0 +1,969 @@
+#if defined(__i386__)
+.file	"chacha-x86.S"
+.text
+.globl	_ChaCha20_ctr32
+.private_extern	_ChaCha20_ctr32
+.align	4
+_ChaCha20_ctr32:
+L_ChaCha20_ctr32_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	xorl	%eax,%eax
+	cmpl	28(%esp),%eax
+	je	L000no_data
+	call	Lpic_point
+Lpic_point:
+	popl	%eax
+	movl	L_OPENSSL_ia32cap_P$non_lazy_ptr-Lpic_point(%eax),%ebp
+	testl	$16777216,(%ebp)
+	jz	L001x86
+	testl	$512,4(%ebp)
+	jz	L001x86
+	jmp	Lssse3_shortcut
+L001x86:
+	movl	32(%esp),%esi
+	movl	36(%esp),%edi
+	subl	$132,%esp
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	%eax,80(%esp)
+	movl	%ebx,84(%esp)
+	movl	%ecx,88(%esp)
+	movl	%edx,92(%esp)
+	movl	16(%esi),%eax
+	movl	20(%esi),%ebx
+	movl	24(%esi),%ecx
+	movl	28(%esi),%edx
+	movl	%eax,96(%esp)
+	movl	%ebx,100(%esp)
+	movl	%ecx,104(%esp)
+	movl	%edx,108(%esp)
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	subl	$1,%eax
+	movl	%eax,112(%esp)
+	movl	%ebx,116(%esp)
+	movl	%ecx,120(%esp)
+	movl	%edx,124(%esp)
+	jmp	L002entry
+.align	4,0x90
+L003outer_loop:
+	movl	%ebx,156(%esp)
+	movl	%eax,152(%esp)
+	movl	%ecx,160(%esp)
+L002entry:
+	movl	$1634760805,%eax
+	movl	$857760878,4(%esp)
+	movl	$2036477234,8(%esp)
+	movl	$1797285236,12(%esp)
+	movl	84(%esp),%ebx
+	movl	88(%esp),%ebp
+	movl	104(%esp),%ecx
+	movl	108(%esp),%esi
+	movl	116(%esp),%edx
+	movl	120(%esp),%edi
+	movl	%ebx,20(%esp)
+	movl	%ebp,24(%esp)
+	movl	%ecx,40(%esp)
+	movl	%esi,44(%esp)
+	movl	%edx,52(%esp)
+	movl	%edi,56(%esp)
+	movl	92(%esp),%ebx
+	movl	124(%esp),%edi
+	movl	112(%esp),%edx
+	movl	80(%esp),%ebp
+	movl	96(%esp),%ecx
+	movl	100(%esp),%esi
+	addl	$1,%edx
+	movl	%ebx,28(%esp)
+	movl	%edi,60(%esp)
+	movl	%edx,112(%esp)
+	movl	$10,%ebx
+	jmp	L004loop
+.align	4,0x90
+L004loop:
+	addl	%ebp,%eax
+	movl	%ebx,128(%esp)
+	movl	%ebp,%ebx
+	xorl	%eax,%edx
+	roll	$16,%edx
+	addl	%edx,%ecx
+	xorl	%ecx,%ebx
+	movl	52(%esp),%edi
+	roll	$12,%ebx
+	movl	20(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,(%esp)
+	roll	$8,%edx
+	movl	4(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,48(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	movl	%ecx,32(%esp)
+	roll	$16,%edi
+	movl	%ebx,16(%esp)
+	addl	%edi,%esi
+	movl	40(%esp),%ecx
+	xorl	%esi,%ebp
+	movl	56(%esp),%edx
+	roll	$12,%ebp
+	movl	24(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,4(%esp)
+	roll	$8,%edi
+	movl	8(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,52(%esp)
+	xorl	%esi,%ebp
+	addl	%ebx,%eax
+	roll	$7,%ebp
+	xorl	%eax,%edx
+	movl	%esi,36(%esp)
+	roll	$16,%edx
+	movl	%ebp,20(%esp)
+	addl	%edx,%ecx
+	movl	44(%esp),%esi
+	xorl	%ecx,%ebx
+	movl	60(%esp),%edi
+	roll	$12,%ebx
+	movl	28(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,8(%esp)
+	roll	$8,%edx
+	movl	12(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,56(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	roll	$16,%edi
+	movl	%ebx,24(%esp)
+	addl	%edi,%esi
+	xorl	%esi,%ebp
+	roll	$12,%ebp
+	movl	20(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,12(%esp)
+	roll	$8,%edi
+	movl	(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,%edx
+	xorl	%esi,%ebp
+	addl	%ebx,%eax
+	roll	$7,%ebp
+	xorl	%eax,%edx
+	roll	$16,%edx
+	movl	%ebp,28(%esp)
+	addl	%edx,%ecx
+	xorl	%ecx,%ebx
+	movl	48(%esp),%edi
+	roll	$12,%ebx
+	movl	24(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,(%esp)
+	roll	$8,%edx
+	movl	4(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,60(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	movl	%ecx,40(%esp)
+	roll	$16,%edi
+	movl	%ebx,20(%esp)
+	addl	%edi,%esi
+	movl	32(%esp),%ecx
+	xorl	%esi,%ebp
+	movl	52(%esp),%edx
+	roll	$12,%ebp
+	movl	28(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,4(%esp)
+	roll	$8,%edi
+	movl	8(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,48(%esp)
+	xorl	%esi,%ebp
+	addl	%ebx,%eax
+	roll	$7,%ebp
+	xorl	%eax,%edx
+	movl	%esi,44(%esp)
+	roll	$16,%edx
+	movl	%ebp,24(%esp)
+	addl	%edx,%ecx
+	movl	36(%esp),%esi
+	xorl	%ecx,%ebx
+	movl	56(%esp),%edi
+	roll	$12,%ebx
+	movl	16(%esp),%ebp
+	addl	%ebx,%eax
+	xorl	%eax,%edx
+	movl	%eax,8(%esp)
+	roll	$8,%edx
+	movl	12(%esp),%eax
+	addl	%edx,%ecx
+	movl	%edx,52(%esp)
+	xorl	%ecx,%ebx
+	addl	%ebp,%eax
+	roll	$7,%ebx
+	xorl	%eax,%edi
+	roll	$16,%edi
+	movl	%ebx,28(%esp)
+	addl	%edi,%esi
+	xorl	%esi,%ebp
+	movl	48(%esp),%edx
+	roll	$12,%ebp
+	movl	128(%esp),%ebx
+	addl	%ebp,%eax
+	xorl	%eax,%edi
+	movl	%eax,12(%esp)
+	roll	$8,%edi
+	movl	(%esp),%eax
+	addl	%edi,%esi
+	movl	%edi,56(%esp)
+	xorl	%esi,%ebp
+	roll	$7,%ebp
+	decl	%ebx
+	jnz	L004loop
+	movl	160(%esp),%ebx
+	addl	$1634760805,%eax
+	addl	80(%esp),%ebp
+	addl	96(%esp),%ecx
+	addl	100(%esp),%esi
+	cmpl	$64,%ebx
+	jb	L005tail
+	movl	156(%esp),%ebx
+	addl	112(%esp),%edx
+	addl	120(%esp),%edi
+	xorl	(%ebx),%eax
+	xorl	16(%ebx),%ebp
+	movl	%eax,(%esp)
+	movl	152(%esp),%eax
+	xorl	32(%ebx),%ecx
+	xorl	36(%ebx),%esi
+	xorl	48(%ebx),%edx
+	xorl	56(%ebx),%edi
+	movl	%ebp,16(%eax)
+	movl	%ecx,32(%eax)
+	movl	%esi,36(%eax)
+	movl	%edx,48(%eax)
+	movl	%edi,56(%eax)
+	movl	4(%esp),%ebp
+	movl	8(%esp),%ecx
+	movl	12(%esp),%esi
+	movl	20(%esp),%edx
+	movl	24(%esp),%edi
+	addl	$857760878,%ebp
+	addl	$2036477234,%ecx
+	addl	$1797285236,%esi
+	addl	84(%esp),%edx
+	addl	88(%esp),%edi
+	xorl	4(%ebx),%ebp
+	xorl	8(%ebx),%ecx
+	xorl	12(%ebx),%esi
+	xorl	20(%ebx),%edx
+	xorl	24(%ebx),%edi
+	movl	%ebp,4(%eax)
+	movl	%ecx,8(%eax)
+	movl	%esi,12(%eax)
+	movl	%edx,20(%eax)
+	movl	%edi,24(%eax)
+	movl	28(%esp),%ebp
+	movl	40(%esp),%ecx
+	movl	44(%esp),%esi
+	movl	52(%esp),%edx
+	movl	60(%esp),%edi
+	addl	92(%esp),%ebp
+	addl	104(%esp),%ecx
+	addl	108(%esp),%esi
+	addl	116(%esp),%edx
+	addl	124(%esp),%edi
+	xorl	28(%ebx),%ebp
+	xorl	40(%ebx),%ecx
+	xorl	44(%ebx),%esi
+	xorl	52(%ebx),%edx
+	xorl	60(%ebx),%edi
+	leal	64(%ebx),%ebx
+	movl	%ebp,28(%eax)
+	movl	(%esp),%ebp
+	movl	%ecx,40(%eax)
+	movl	160(%esp),%ecx
+	movl	%esi,44(%eax)
+	movl	%edx,52(%eax)
+	movl	%edi,60(%eax)
+	movl	%ebp,(%eax)
+	leal	64(%eax),%eax
+	subl	$64,%ecx
+	jnz	L003outer_loop
+	jmp	L006done
+L005tail:
+	addl	112(%esp),%edx
+	addl	120(%esp),%edi
+	movl	%eax,(%esp)
+	movl	%ebp,16(%esp)
+	movl	%ecx,32(%esp)
+	movl	%esi,36(%esp)
+	movl	%edx,48(%esp)
+	movl	%edi,56(%esp)
+	movl	4(%esp),%ebp
+	movl	8(%esp),%ecx
+	movl	12(%esp),%esi
+	movl	20(%esp),%edx
+	movl	24(%esp),%edi
+	addl	$857760878,%ebp
+	addl	$2036477234,%ecx
+	addl	$1797285236,%esi
+	addl	84(%esp),%edx
+	addl	88(%esp),%edi
+	movl	%ebp,4(%esp)
+	movl	%ecx,8(%esp)
+	movl	%esi,12(%esp)
+	movl	%edx,20(%esp)
+	movl	%edi,24(%esp)
+	movl	28(%esp),%ebp
+	movl	40(%esp),%ecx
+	movl	44(%esp),%esi
+	movl	52(%esp),%edx
+	movl	60(%esp),%edi
+	addl	92(%esp),%ebp
+	addl	104(%esp),%ecx
+	addl	108(%esp),%esi
+	addl	116(%esp),%edx
+	addl	124(%esp),%edi
+	movl	%ebp,28(%esp)
+	movl	156(%esp),%ebp
+	movl	%ecx,40(%esp)
+	movl	152(%esp),%ecx
+	movl	%esi,44(%esp)
+	xorl	%esi,%esi
+	movl	%edx,52(%esp)
+	movl	%edi,60(%esp)
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+L007tail_loop:
+	movb	(%esi,%ebp,1),%al
+	movb	(%esp,%esi,1),%dl
+	leal	1(%esi),%esi
+	xorb	%dl,%al
+	movb	%al,-1(%ecx,%esi,1)
+	decl	%ebx
+	jnz	L007tail_loop
+L006done:
+	addl	$132,%esp
+L000no_data:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_ChaCha20_ssse3
+.private_extern	_ChaCha20_ssse3
+.align	4
+_ChaCha20_ssse3:
+L_ChaCha20_ssse3_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+Lssse3_shortcut:
+	movl	20(%esp),%edi
+	movl	24(%esp),%esi
+	movl	28(%esp),%ecx
+	movl	32(%esp),%edx
+	movl	36(%esp),%ebx
+	movl	%esp,%ebp
+	subl	$524,%esp
+	andl	$-64,%esp
+	movl	%ebp,512(%esp)
+	leal	Lssse3_data-Lpic_point(%eax),%eax
+	movdqu	(%ebx),%xmm3
+	cmpl	$256,%ecx
+	jb	L0081x
+	movl	%edx,516(%esp)
+	movl	%ebx,520(%esp)
+	subl	$256,%ecx
+	leal	384(%esp),%ebp
+	movdqu	(%edx),%xmm7
+	pshufd	$0,%xmm3,%xmm0
+	pshufd	$85,%xmm3,%xmm1
+	pshufd	$170,%xmm3,%xmm2
+	pshufd	$255,%xmm3,%xmm3
+	paddd	48(%eax),%xmm0
+	pshufd	$0,%xmm7,%xmm4
+	pshufd	$85,%xmm7,%xmm5
+	psubd	64(%eax),%xmm0
+	pshufd	$170,%xmm7,%xmm6
+	pshufd	$255,%xmm7,%xmm7
+	movdqa	%xmm0,64(%ebp)
+	movdqa	%xmm1,80(%ebp)
+	movdqa	%xmm2,96(%ebp)
+	movdqa	%xmm3,112(%ebp)
+	movdqu	16(%edx),%xmm3
+	movdqa	%xmm4,-64(%ebp)
+	movdqa	%xmm5,-48(%ebp)
+	movdqa	%xmm6,-32(%ebp)
+	movdqa	%xmm7,-16(%ebp)
+	movdqa	32(%eax),%xmm7
+	leal	128(%esp),%ebx
+	pshufd	$0,%xmm3,%xmm0
+	pshufd	$85,%xmm3,%xmm1
+	pshufd	$170,%xmm3,%xmm2
+	pshufd	$255,%xmm3,%xmm3
+	pshufd	$0,%xmm7,%xmm4
+	pshufd	$85,%xmm7,%xmm5
+	pshufd	$170,%xmm7,%xmm6
+	pshufd	$255,%xmm7,%xmm7
+	movdqa	%xmm0,(%ebp)
+	movdqa	%xmm1,16(%ebp)
+	movdqa	%xmm2,32(%ebp)
+	movdqa	%xmm3,48(%ebp)
+	movdqa	%xmm4,-128(%ebp)
+	movdqa	%xmm5,-112(%ebp)
+	movdqa	%xmm6,-96(%ebp)
+	movdqa	%xmm7,-80(%ebp)
+	leal	128(%esi),%esi
+	leal	128(%edi),%edi
+	jmp	L009outer_loop
+.align	4,0x90
+L009outer_loop:
+	movdqa	-112(%ebp),%xmm1
+	movdqa	-96(%ebp),%xmm2
+	movdqa	-80(%ebp),%xmm3
+	movdqa	-48(%ebp),%xmm5
+	movdqa	-32(%ebp),%xmm6
+	movdqa	-16(%ebp),%xmm7
+	movdqa	%xmm1,-112(%ebx)
+	movdqa	%xmm2,-96(%ebx)
+	movdqa	%xmm3,-80(%ebx)
+	movdqa	%xmm5,-48(%ebx)
+	movdqa	%xmm6,-32(%ebx)
+	movdqa	%xmm7,-16(%ebx)
+	movdqa	32(%ebp),%xmm2
+	movdqa	48(%ebp),%xmm3
+	movdqa	64(%ebp),%xmm4
+	movdqa	80(%ebp),%xmm5
+	movdqa	96(%ebp),%xmm6
+	movdqa	112(%ebp),%xmm7
+	paddd	64(%eax),%xmm4
+	movdqa	%xmm2,32(%ebx)
+	movdqa	%xmm3,48(%ebx)
+	movdqa	%xmm4,64(%ebx)
+	movdqa	%xmm5,80(%ebx)
+	movdqa	%xmm6,96(%ebx)
+	movdqa	%xmm7,112(%ebx)
+	movdqa	%xmm4,64(%ebp)
+	movdqa	-128(%ebp),%xmm0
+	movdqa	%xmm4,%xmm6
+	movdqa	-64(%ebp),%xmm3
+	movdqa	(%ebp),%xmm4
+	movdqa	16(%ebp),%xmm5
+	movl	$10,%edx
+	nop
+.align	4,0x90
+L010loop:
+	paddd	%xmm3,%xmm0
+	movdqa	%xmm3,%xmm2
+	pxor	%xmm0,%xmm6
+	pshufb	(%eax),%xmm6
+	paddd	%xmm6,%xmm4
+	pxor	%xmm4,%xmm2
+	movdqa	-48(%ebx),%xmm3
+	movdqa	%xmm2,%xmm1
+	pslld	$12,%xmm2
+	psrld	$20,%xmm1
+	por	%xmm1,%xmm2
+	movdqa	-112(%ebx),%xmm1
+	paddd	%xmm2,%xmm0
+	movdqa	80(%ebx),%xmm7
+	pxor	%xmm0,%xmm6
+	movdqa	%xmm0,-128(%ebx)
+	pshufb	16(%eax),%xmm6
+	paddd	%xmm6,%xmm4
+	movdqa	%xmm6,64(%ebx)
+	pxor	%xmm4,%xmm2
+	paddd	%xmm3,%xmm1
+	movdqa	%xmm2,%xmm0
+	pslld	$7,%xmm2
+	psrld	$25,%xmm0
+	pxor	%xmm1,%xmm7
+	por	%xmm0,%xmm2
+	movdqa	%xmm4,(%ebx)
+	pshufb	(%eax),%xmm7
+	movdqa	%xmm2,-64(%ebx)
+	paddd	%xmm7,%xmm5
+	movdqa	32(%ebx),%xmm4
+	pxor	%xmm5,%xmm3
+	movdqa	-32(%ebx),%xmm2
+	movdqa	%xmm3,%xmm0
+	pslld	$12,%xmm3
+	psrld	$20,%xmm0
+	por	%xmm0,%xmm3
+	movdqa	-96(%ebx),%xmm0
+	paddd	%xmm3,%xmm1
+	movdqa	96(%ebx),%xmm6
+	pxor	%xmm1,%xmm7
+	movdqa	%xmm1,-112(%ebx)
+	pshufb	16(%eax),%xmm7
+	paddd	%xmm7,%xmm5
+	movdqa	%xmm7,80(%ebx)
+	pxor	%xmm5,%xmm3
+	paddd	%xmm2,%xmm0
+	movdqa	%xmm3,%xmm1
+	pslld	$7,%xmm3
+	psrld	$25,%xmm1
+	pxor	%xmm0,%xmm6
+	por	%xmm1,%xmm3
+	movdqa	%xmm5,16(%ebx)
+	pshufb	(%eax),%xmm6
+	movdqa	%xmm3,-48(%ebx)
+	paddd	%xmm6,%xmm4
+	movdqa	48(%ebx),%xmm5
+	pxor	%xmm4,%xmm2
+	movdqa	-16(%ebx),%xmm3
+	movdqa	%xmm2,%xmm1
+	pslld	$12,%xmm2
+	psrld	$20,%xmm1
+	por	%xmm1,%xmm2
+	movdqa	-80(%ebx),%xmm1
+	paddd	%xmm2,%xmm0
+	movdqa	112(%ebx),%xmm7
+	pxor	%xmm0,%xmm6
+	movdqa	%xmm0,-96(%ebx)
+	pshufb	16(%eax),%xmm6
+	paddd	%xmm6,%xmm4
+	movdqa	%xmm6,96(%ebx)
+	pxor	%xmm4,%xmm2
+	paddd	%xmm3,%xmm1
+	movdqa	%xmm2,%xmm0
+	pslld	$7,%xmm2
+	psrld	$25,%xmm0
+	pxor	%xmm1,%xmm7
+	por	%xmm0,%xmm2
+	pshufb	(%eax),%xmm7
+	movdqa	%xmm2,-32(%ebx)
+	paddd	%xmm7,%xmm5
+	pxor	%xmm5,%xmm3
+	movdqa	-48(%ebx),%xmm2
+	movdqa	%xmm3,%xmm0
+	pslld	$12,%xmm3
+	psrld	$20,%xmm0
+	por	%xmm0,%xmm3
+	movdqa	-128(%ebx),%xmm0
+	paddd	%xmm3,%xmm1
+	pxor	%xmm1,%xmm7
+	movdqa	%xmm1,-80(%ebx)
+	pshufb	16(%eax),%xmm7
+	paddd	%xmm7,%xmm5
+	movdqa	%xmm7,%xmm6
+	pxor	%xmm5,%xmm3
+	paddd	%xmm2,%xmm0
+	movdqa	%xmm3,%xmm1
+	pslld	$7,%xmm3
+	psrld	$25,%xmm1
+	pxor	%xmm0,%xmm6
+	por	%xmm1,%xmm3
+	pshufb	(%eax),%xmm6
+	movdqa	%xmm3,-16(%ebx)
+	paddd	%xmm6,%xmm4
+	pxor	%xmm4,%xmm2
+	movdqa	-32(%ebx),%xmm3
+	movdqa	%xmm2,%xmm1
+	pslld	$12,%xmm2
+	psrld	$20,%xmm1
+	por	%xmm1,%xmm2
+	movdqa	-112(%ebx),%xmm1
+	paddd	%xmm2,%xmm0
+	movdqa	64(%ebx),%xmm7
+	pxor	%xmm0,%xmm6
+	movdqa	%xmm0,-128(%ebx)
+	pshufb	16(%eax),%xmm6
+	paddd	%xmm6,%xmm4
+	movdqa	%xmm6,112(%ebx)
+	pxor	%xmm4,%xmm2
+	paddd	%xmm3,%xmm1
+	movdqa	%xmm2,%xmm0
+	pslld	$7,%xmm2
+	psrld	$25,%xmm0
+	pxor	%xmm1,%xmm7
+	por	%xmm0,%xmm2
+	movdqa	%xmm4,32(%ebx)
+	pshufb	(%eax),%xmm7
+	movdqa	%xmm2,-48(%ebx)
+	paddd	%xmm7,%xmm5
+	movdqa	(%ebx),%xmm4
+	pxor	%xmm5,%xmm3
+	movdqa	-16(%ebx),%xmm2
+	movdqa	%xmm3,%xmm0
+	pslld	$12,%xmm3
+	psrld	$20,%xmm0
+	por	%xmm0,%xmm3
+	movdqa	-96(%ebx),%xmm0
+	paddd	%xmm3,%xmm1
+	movdqa	80(%ebx),%xmm6
+	pxor	%xmm1,%xmm7
+	movdqa	%xmm1,-112(%ebx)
+	pshufb	16(%eax),%xmm7
+	paddd	%xmm7,%xmm5
+	movdqa	%xmm7,64(%ebx)
+	pxor	%xmm5,%xmm3
+	paddd	%xmm2,%xmm0
+	movdqa	%xmm3,%xmm1
+	pslld	$7,%xmm3
+	psrld	$25,%xmm1
+	pxor	%xmm0,%xmm6
+	por	%xmm1,%xmm3
+	movdqa	%xmm5,48(%ebx)
+	pshufb	(%eax),%xmm6
+	movdqa	%xmm3,-32(%ebx)
+	paddd	%xmm6,%xmm4
+	movdqa	16(%ebx),%xmm5
+	pxor	%xmm4,%xmm2
+	movdqa	-64(%ebx),%xmm3
+	movdqa	%xmm2,%xmm1
+	pslld	$12,%xmm2
+	psrld	$20,%xmm1
+	por	%xmm1,%xmm2
+	movdqa	-80(%ebx),%xmm1
+	paddd	%xmm2,%xmm0
+	movdqa	96(%ebx),%xmm7
+	pxor	%xmm0,%xmm6
+	movdqa	%xmm0,-96(%ebx)
+	pshufb	16(%eax),%xmm6
+	paddd	%xmm6,%xmm4
+	movdqa	%xmm6,80(%ebx)
+	pxor	%xmm4,%xmm2
+	paddd	%xmm3,%xmm1
+	movdqa	%xmm2,%xmm0
+	pslld	$7,%xmm2
+	psrld	$25,%xmm0
+	pxor	%xmm1,%xmm7
+	por	%xmm0,%xmm2
+	pshufb	(%eax),%xmm7
+	movdqa	%xmm2,-16(%ebx)
+	paddd	%xmm7,%xmm5
+	pxor	%xmm5,%xmm3
+	movdqa	%xmm3,%xmm0
+	pslld	$12,%xmm3
+	psrld	$20,%xmm0
+	por	%xmm0,%xmm3
+	movdqa	-128(%ebx),%xmm0
+	paddd	%xmm3,%xmm1
+	movdqa	64(%ebx),%xmm6
+	pxor	%xmm1,%xmm7
+	movdqa	%xmm1,-80(%ebx)
+	pshufb	16(%eax),%xmm7
+	paddd	%xmm7,%xmm5
+	movdqa	%xmm7,96(%ebx)
+	pxor	%xmm5,%xmm3
+	movdqa	%xmm3,%xmm1
+	pslld	$7,%xmm3
+	psrld	$25,%xmm1
+	por	%xmm1,%xmm3
+	decl	%edx
+	jnz	L010loop
+	movdqa	%xmm3,-64(%ebx)
+	movdqa	%xmm4,(%ebx)
+	movdqa	%xmm5,16(%ebx)
+	movdqa	%xmm6,64(%ebx)
+	movdqa	%xmm7,96(%ebx)
+	movdqa	-112(%ebx),%xmm1
+	movdqa	-96(%ebx),%xmm2
+	movdqa	-80(%ebx),%xmm3
+	paddd	-128(%ebp),%xmm0
+	paddd	-112(%ebp),%xmm1
+	paddd	-96(%ebp),%xmm2
+	paddd	-80(%ebp),%xmm3
+	movdqa	%xmm0,%xmm6
+	punpckldq	%xmm1,%xmm0
+	movdqa	%xmm2,%xmm7
+	punpckldq	%xmm3,%xmm2
+	punpckhdq	%xmm1,%xmm6
+	punpckhdq	%xmm3,%xmm7
+	movdqa	%xmm0,%xmm1
+	punpcklqdq	%xmm2,%xmm0
+	movdqa	%xmm6,%xmm3
+	punpcklqdq	%xmm7,%xmm6
+	punpckhqdq	%xmm2,%xmm1
+	punpckhqdq	%xmm7,%xmm3
+	movdqu	-128(%esi),%xmm4
+	movdqu	-64(%esi),%xmm5
+	movdqu	(%esi),%xmm2
+	movdqu	64(%esi),%xmm7
+	leal	16(%esi),%esi
+	pxor	%xmm0,%xmm4
+	movdqa	-64(%ebx),%xmm0
+	pxor	%xmm1,%xmm5
+	movdqa	-48(%ebx),%xmm1
+	pxor	%xmm2,%xmm6
+	movdqa	-32(%ebx),%xmm2
+	pxor	%xmm3,%xmm7
+	movdqa	-16(%ebx),%xmm3
+	movdqu	%xmm4,-128(%edi)
+	movdqu	%xmm5,-64(%edi)
+	movdqu	%xmm6,(%edi)
+	movdqu	%xmm7,64(%edi)
+	leal	16(%edi),%edi
+	paddd	-64(%ebp),%xmm0
+	paddd	-48(%ebp),%xmm1
+	paddd	-32(%ebp),%xmm2
+	paddd	-16(%ebp),%xmm3
+	movdqa	%xmm0,%xmm6
+	punpckldq	%xmm1,%xmm0
+	movdqa	%xmm2,%xmm7
+	punpckldq	%xmm3,%xmm2
+	punpckhdq	%xmm1,%xmm6
+	punpckhdq	%xmm3,%xmm7
+	movdqa	%xmm0,%xmm1
+	punpcklqdq	%xmm2,%xmm0
+	movdqa	%xmm6,%xmm3
+	punpcklqdq	%xmm7,%xmm6
+	punpckhqdq	%xmm2,%xmm1
+	punpckhqdq	%xmm7,%xmm3
+	movdqu	-128(%esi),%xmm4
+	movdqu	-64(%esi),%xmm5
+	movdqu	(%esi),%xmm2
+	movdqu	64(%esi),%xmm7
+	leal	16(%esi),%esi
+	pxor	%xmm0,%xmm4
+	movdqa	(%ebx),%xmm0
+	pxor	%xmm1,%xmm5
+	movdqa	16(%ebx),%xmm1
+	pxor	%xmm2,%xmm6
+	movdqa	32(%ebx),%xmm2
+	pxor	%xmm3,%xmm7
+	movdqa	48(%ebx),%xmm3
+	movdqu	%xmm4,-128(%edi)
+	movdqu	%xmm5,-64(%edi)
+	movdqu	%xmm6,(%edi)
+	movdqu	%xmm7,64(%edi)
+	leal	16(%edi),%edi
+	paddd	(%ebp),%xmm0
+	paddd	16(%ebp),%xmm1
+	paddd	32(%ebp),%xmm2
+	paddd	48(%ebp),%xmm3
+	movdqa	%xmm0,%xmm6
+	punpckldq	%xmm1,%xmm0
+	movdqa	%xmm2,%xmm7
+	punpckldq	%xmm3,%xmm2
+	punpckhdq	%xmm1,%xmm6
+	punpckhdq	%xmm3,%xmm7
+	movdqa	%xmm0,%xmm1
+	punpcklqdq	%xmm2,%xmm0
+	movdqa	%xmm6,%xmm3
+	punpcklqdq	%xmm7,%xmm6
+	punpckhqdq	%xmm2,%xmm1
+	punpckhqdq	%xmm7,%xmm3
+	movdqu	-128(%esi),%xmm4
+	movdqu	-64(%esi),%xmm5
+	movdqu	(%esi),%xmm2
+	movdqu	64(%esi),%xmm7
+	leal	16(%esi),%esi
+	pxor	%xmm0,%xmm4
+	movdqa	64(%ebx),%xmm0
+	pxor	%xmm1,%xmm5
+	movdqa	80(%ebx),%xmm1
+	pxor	%xmm2,%xmm6
+	movdqa	96(%ebx),%xmm2
+	pxor	%xmm3,%xmm7
+	movdqa	112(%ebx),%xmm3
+	movdqu	%xmm4,-128(%edi)
+	movdqu	%xmm5,-64(%edi)
+	movdqu	%xmm6,(%edi)
+	movdqu	%xmm7,64(%edi)
+	leal	16(%edi),%edi
+	paddd	64(%ebp),%xmm0
+	paddd	80(%ebp),%xmm1
+	paddd	96(%ebp),%xmm2
+	paddd	112(%ebp),%xmm3
+	movdqa	%xmm0,%xmm6
+	punpckldq	%xmm1,%xmm0
+	movdqa	%xmm2,%xmm7
+	punpckldq	%xmm3,%xmm2
+	punpckhdq	%xmm1,%xmm6
+	punpckhdq	%xmm3,%xmm7
+	movdqa	%xmm0,%xmm1
+	punpcklqdq	%xmm2,%xmm0
+	movdqa	%xmm6,%xmm3
+	punpcklqdq	%xmm7,%xmm6
+	punpckhqdq	%xmm2,%xmm1
+	punpckhqdq	%xmm7,%xmm3
+	movdqu	-128(%esi),%xmm4
+	movdqu	-64(%esi),%xmm5
+	movdqu	(%esi),%xmm2
+	movdqu	64(%esi),%xmm7
+	leal	208(%esi),%esi
+	pxor	%xmm0,%xmm4
+	pxor	%xmm1,%xmm5
+	pxor	%xmm2,%xmm6
+	pxor	%xmm3,%xmm7
+	movdqu	%xmm4,-128(%edi)
+	movdqu	%xmm5,-64(%edi)
+	movdqu	%xmm6,(%edi)
+	movdqu	%xmm7,64(%edi)
+	leal	208(%edi),%edi
+	subl	$256,%ecx
+	jnc	L009outer_loop
+	addl	$256,%ecx
+	jz	L011done
+	movl	520(%esp),%ebx
+	leal	-128(%esi),%esi
+	movl	516(%esp),%edx
+	leal	-128(%edi),%edi
+	movd	64(%ebp),%xmm2
+	movdqu	(%ebx),%xmm3
+	paddd	96(%eax),%xmm2
+	pand	112(%eax),%xmm3
+	por	%xmm2,%xmm3
+L0081x:
+	movdqa	32(%eax),%xmm0
+	movdqu	(%edx),%xmm1
+	movdqu	16(%edx),%xmm2
+	movdqa	(%eax),%xmm6
+	movdqa	16(%eax),%xmm7
+	movl	%ebp,48(%esp)
+	movdqa	%xmm0,(%esp)
+	movdqa	%xmm1,16(%esp)
+	movdqa	%xmm2,32(%esp)
+	movdqa	%xmm3,48(%esp)
+	movl	$10,%edx
+	jmp	L012loop1x
+.align	4,0x90
+L013outer1x:
+	movdqa	80(%eax),%xmm3
+	movdqa	(%esp),%xmm0
+	movdqa	16(%esp),%xmm1
+	movdqa	32(%esp),%xmm2
+	paddd	48(%esp),%xmm3
+	movl	$10,%edx
+	movdqa	%xmm3,48(%esp)
+	jmp	L012loop1x
+.align	4,0x90
+L012loop1x:
+	paddd	%xmm1,%xmm0
+	pxor	%xmm0,%xmm3
+.byte	102,15,56,0,222
+	paddd	%xmm3,%xmm2
+	pxor	%xmm2,%xmm1
+	movdqa	%xmm1,%xmm4
+	psrld	$20,%xmm1
+	pslld	$12,%xmm4
+	por	%xmm4,%xmm1
+	paddd	%xmm1,%xmm0
+	pxor	%xmm0,%xmm3
+.byte	102,15,56,0,223
+	paddd	%xmm3,%xmm2
+	pxor	%xmm2,%xmm1
+	movdqa	%xmm1,%xmm4
+	psrld	$25,%xmm1
+	pslld	$7,%xmm4
+	por	%xmm4,%xmm1
+	pshufd	$78,%xmm2,%xmm2
+	pshufd	$57,%xmm1,%xmm1
+	pshufd	$147,%xmm3,%xmm3
+	nop
+	paddd	%xmm1,%xmm0
+	pxor	%xmm0,%xmm3
+.byte	102,15,56,0,222
+	paddd	%xmm3,%xmm2
+	pxor	%xmm2,%xmm1
+	movdqa	%xmm1,%xmm4
+	psrld	$20,%xmm1
+	pslld	$12,%xmm4
+	por	%xmm4,%xmm1
+	paddd	%xmm1,%xmm0
+	pxor	%xmm0,%xmm3
+.byte	102,15,56,0,223
+	paddd	%xmm3,%xmm2
+	pxor	%xmm2,%xmm1
+	movdqa	%xmm1,%xmm4
+	psrld	$25,%xmm1
+	pslld	$7,%xmm4
+	por	%xmm4,%xmm1
+	pshufd	$78,%xmm2,%xmm2
+	pshufd	$147,%xmm1,%xmm1
+	pshufd	$57,%xmm3,%xmm3
+	decl	%edx
+	jnz	L012loop1x
+	paddd	(%esp),%xmm0
+	paddd	16(%esp),%xmm1
+	paddd	32(%esp),%xmm2
+	paddd	48(%esp),%xmm3
+	cmpl	$64,%ecx
+	jb	L014tail
+	movdqu	(%esi),%xmm4
+	movdqu	16(%esi),%xmm5
+	pxor	%xmm4,%xmm0
+	movdqu	32(%esi),%xmm4
+	pxor	%xmm5,%xmm1
+	movdqu	48(%esi),%xmm5
+	pxor	%xmm4,%xmm2
+	pxor	%xmm5,%xmm3
+	leal	64(%esi),%esi
+	movdqu	%xmm0,(%edi)
+	movdqu	%xmm1,16(%edi)
+	movdqu	%xmm2,32(%edi)
+	movdqu	%xmm3,48(%edi)
+	leal	64(%edi),%edi
+	subl	$64,%ecx
+	jnz	L013outer1x
+	jmp	L011done
+L014tail:
+	movdqa	%xmm0,(%esp)
+	movdqa	%xmm1,16(%esp)
+	movdqa	%xmm2,32(%esp)
+	movdqa	%xmm3,48(%esp)
+	xorl	%eax,%eax
+	xorl	%edx,%edx
+	xorl	%ebp,%ebp
+L015tail_loop:
+	movb	(%esp,%ebp,1),%al
+	movb	(%esi,%ebp,1),%dl
+	leal	1(%ebp),%ebp
+	xorb	%dl,%al
+	movb	%al,-1(%edi,%ebp,1)
+	decl	%ecx
+	jnz	L015tail_loop
+L011done:
+	movl	512(%esp),%esp
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.align	6,0x90
+Lssse3_data:
+.byte	2,3,0,1,6,7,4,5,10,11,8,9,14,15,12,13
+.byte	3,0,1,2,7,4,5,6,11,8,9,10,15,12,13,14
+.long	1634760805,857760878,2036477234,1797285236
+.long	0,1,2,3
+.long	4,4,4,4
+.long	1,0,0,0
+.long	4,0,0,0
+.long	0,-1,-1,-1
+.align	6,0x90
+.byte	67,104,97,67,104,97,50,48,32,102,111,114,32,120,56,54
+.byte	44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32
+.byte	60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111
+.byte	114,103,62,0
+.section __IMPORT,__pointers,non_lazy_symbol_pointers
+L_OPENSSL_ia32cap_P$non_lazy_ptr:
+.indirect_symbol	_OPENSSL_ia32cap_P
+.long	0
+#endif

third_party/boringssl/legacy_kit/mac-x86/crypto/md5/md5-586.S

@@ -0,0 +1,680 @@
+#if defined(__i386__)
+.file	"src/crypto/md5/asm/md5-586.S"
+.text
+.globl	_md5_block_asm_data_order
+.private_extern	_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
+#endif