update asm files in openssl-1.0.2l

shigeki · May 29, 2017 · aaa09eb · aaa09eb
1 parent a389060
commit aaa09eb
Show file tree

Hide file tree

Showing 15 changed files with 111 additions and 105 deletions.
diff --git a/deps/openssl/asm/arm-void-gas/aes/bsaes-armv7.S b/deps/openssl/asm/arm-void-gas/aes/bsaes-armv7.S
@@ -1298,7 +1298,7 @@ bsaes_cbc_encrypt:
 	vmov	q4,q15		@ just in case ensure that IV
 	vmov	q5,q0			@ and input are preserved
 	bl	AES_decrypt
-	vld1.8	{q0}, [r9,:64]		@ load result
+	vld1.8	{q0}, [r9]		@ load result
 	veor	q0, q0, q4	@ ^= IV
 	vmov	q15, q5		@ q5 holds input
 	vst1.8	{q0}, [r10]		@ write output

diff --git a/deps/openssl/asm/x64-elf-gas/aes/aesni-sha1-x86_64.s b/deps/openssl/asm/x64-elf-gas/aes/aesni-sha1-x86_64.s
@@ -2692,6 +2692,7 @@ aesni_cbc_sha1_enc_shaext:
 	movl	240(%rcx),%r11d
 	subq	%rdi,%rsi
 	movups	(%rcx),%xmm15
+	movups	(%r8),%xmm2
 	movups	16(%rcx),%xmm0
 	leaq	112(%rcx),%rcx
 

diff --git a/deps/openssl/asm/x64-elf-gas/aes/aesni-sha256-x86_64.s b/deps/openssl/asm/x64-elf-gas/aes/aesni-sha256-x86_64.s
@@ -4012,6 +4012,7 @@ aesni_cbc_sha256_enc_shaext:
 	movl	240(%rcx),%r11d
 	subq	%rdi,%rsi
 	movups	(%rcx),%xmm15
+	movups	(%r8),%xmm6
 	movups	16(%rcx),%xmm4
 	leaq	112(%rcx),%rcx
 

diff --git a/deps/openssl/asm/x64-elf-gas/x86_64cpuid.s b/deps/openssl/asm/x64-elf-gas/x86_64cpuid.s
@@ -107,14 +107,6 @@ OPENSSL_ia32_cpuid:
 	shrl	$14,%r10d
 	andl	$0xfff,%r10d
 
-	cmpl	$7,%r11d
-	jb	.Lnocacheinfo
-
-	movl	$7,%eax
-	xorl	%ecx,%ecx
-	cpuid
-	movl	%ebx,8(%rdi)
-
 .Lnocacheinfo:
 	movl	$1,%eax
 	cpuid
@@ -144,6 +136,15 @@ OPENSSL_ia32_cpuid:
 	orl	%ecx,%r9d
 
 	movl	%edx,%r10d
+
+	cmpl	$7,%r11d
+	jb	.Lno_extended_info
+	movl	$7,%eax
+	xorl	%ecx,%ecx
+	cpuid
+	movl	%ebx,8(%rdi)
+.Lno_extended_info:
+
 	btl	$27,%r9d
 	jnc	.Lclear_avx
 	xorl	%ecx,%ecx

diff --git a/deps/openssl/asm/x64-macosx-gas/aes/aesni-sha1-x86_64.s b/deps/openssl/asm/x64-macosx-gas/aes/aesni-sha1-x86_64.s
@@ -2692,6 +2692,7 @@ aesni_cbc_sha1_enc_shaext:
 	movl	240(%rcx),%r11d
 	subq	%rdi,%rsi
 	movups	(%rcx),%xmm15
+	movups	(%r8),%xmm2
 	movups	16(%rcx),%xmm0
 	leaq	112(%rcx),%rcx
 

diff --git a/deps/openssl/asm/x64-macosx-gas/aes/aesni-sha256-x86_64.s b/deps/openssl/asm/x64-macosx-gas/aes/aesni-sha256-x86_64.s
@@ -4012,6 +4012,7 @@ aesni_cbc_sha256_enc_shaext:
 	movl	240(%rcx),%r11d
 	subq	%rdi,%rsi
 	movups	(%rcx),%xmm15
+	movups	(%r8),%xmm6
 	movups	16(%rcx),%xmm4
 	leaq	112(%rcx),%rcx
 

diff --git a/deps/openssl/asm/x64-macosx-gas/x86_64cpuid.s b/deps/openssl/asm/x64-macosx-gas/x86_64cpuid.s
@@ -108,14 +108,6 @@ L$intel:
 	shrl	$14,%r10d
 	andl	$0xfff,%r10d
 
-	cmpl	$7,%r11d
-	jb	L$nocacheinfo
-
-	movl	$7,%eax
-	xorl	%ecx,%ecx
-	cpuid
-	movl	%ebx,8(%rdi)
-
 L$nocacheinfo:
 	movl	$1,%eax
 	cpuid
@@ -145,6 +137,15 @@ L$generic:
 	orl	%ecx,%r9d
 
 	movl	%edx,%r10d
+
+	cmpl	$7,%r11d
+	jb	L$no_extended_info
+	movl	$7,%eax
+	xorl	%ecx,%ecx
+	cpuid
+	movl	%ebx,8(%rdi)
+L$no_extended_info:
+
 	btl	$27,%r9d
 	jnc	L$clear_avx
 	xorl	%ecx,%ecx

diff --git a/deps/openssl/asm/x64-win32-masm/aes/aesni-sha1-x86_64.asm b/deps/openssl/asm/x64-win32-masm/aes/aesni-sha1-x86_64.asm
@@ -2793,6 +2793,7 @@ $L$prologue_shaext::
 	mov	r11d,DWORD PTR[240+rcx]
 	sub	rsi,rdi
 	movups	xmm15,XMMWORD PTR[rcx]
+	movups	xmm2,XMMWORD PTR[r8]
 	movups	xmm0,XMMWORD PTR[16+rcx]
 	lea	rcx,QWORD PTR[112+rcx]
 

diff --git a/deps/openssl/asm/x64-win32-masm/aes/aesni-sha256-x86_64.asm b/deps/openssl/asm/x64-win32-masm/aes/aesni-sha256-x86_64.asm
@@ -4146,11 +4146,12 @@ $L$prologue_shaext::
 	mov	r11d,DWORD PTR[240+rcx]
 	sub	rsi,rdi
 	movups	xmm15,XMMWORD PTR[rcx]
+	movups	xmm6,XMMWORD PTR[r8]
 	movups	xmm4,XMMWORD PTR[16+rcx]
 	lea	rcx,QWORD PTR[112+rcx]
 
 	pshufd	xmm0,xmm1,01bh
-	pshufd	xmm1,xmm1,0b1h
+	pshufd	xmm1,xmm1,1h
 	pshufd	xmm2,xmm2,01bh
 	movdqa	xmm7,xmm3
 DB	102,15,58,15,202,8
@@ -4475,9 +4476,9 @@ $L$aesenclast4::
 	lea	rdi,QWORD PTR[64+rdi]
 	jnz	$L$oop_shaext
 
-	pshufd	xmm2,xmm2,0b1h
+	pshufd	xmm2,xmm2,1h
 	pshufd	xmm3,xmm1,01bh
-	pshufd	xmm1,xmm1,0b1h
+	pshufd	xmm1,xmm1,1h
 	punpckhqdq	xmm1,xmm2
 DB	102,15,58,15,211,8
 

diff --git a/deps/openssl/asm/x64-win32-masm/ec/ecp_nistz256-x86_64.asm b/deps/openssl/asm/x64-win32-masm/ec/ecp_nistz256-x86_64.asm
@@ -2051,7 +2051,7 @@ $L$SEH_begin_ecp_nistz256_point_add::
 	por	xmm5,xmm4
 
 	movdqu	xmm0,XMMWORD PTR[rsi]
-	pshufd	xmm3,xmm5,0b1h
+	pshufd	xmm3,xmm5,1h
 	movdqu	xmm1,XMMWORD PTR[16+rsi]
 	movdqu	xmm2,XMMWORD PTR[32+rsi]
 	por	xmm5,xmm3
@@ -2081,7 +2081,7 @@ DB	102,72,15,110,199
 	call	__ecp_nistz256_sqr_montq
 
 	pcmpeqd	xmm5,xmm4
-	pshufd	xmm4,xmm1,0b1h
+	pshufd	xmm4,xmm1,1h
 	por	xmm4,xmm1
 	pshufd	xmm5,xmm5,0
 	pshufd	xmm3,xmm4,01eh
@@ -2466,7 +2466,7 @@ $L$SEH_begin_ecp_nistz256_point_add_affine::
 	por	xmm5,xmm4
 
 	movdqu	xmm0,XMMWORD PTR[rbx]
-	pshufd	xmm3,xmm5,0b1h
+	pshufd	xmm3,xmm5,1h
 	movdqu	xmm1,XMMWORD PTR[16+rbx]
 	movdqu	xmm2,XMMWORD PTR[32+rbx]
 	por	xmm5,xmm3
@@ -2488,7 +2488,7 @@ DB	102,72,15,110,199
 	call	__ecp_nistz256_sqr_montq
 
 	pcmpeqd	xmm5,xmm4
-	pshufd	xmm4,xmm3,0b1h
+	pshufd	xmm4,xmm3,1h
 	mov	rax,QWORD PTR[rbx]
 
 	mov	r9,r12
@@ -3117,7 +3117,7 @@ $L$point_addx::
 	por	xmm5,xmm4
 
 	movdqu	xmm0,XMMWORD PTR[rsi]
-	pshufd	xmm3,xmm5,0b1h
+	pshufd	xmm3,xmm5,1h
 	movdqu	xmm1,XMMWORD PTR[16+rsi]
 	movdqu	xmm2,XMMWORD PTR[32+rsi]
 	por	xmm5,xmm3
@@ -3147,7 +3147,7 @@ DB	102,72,15,110,199
 	call	__ecp_nistz256_sqr_montx
 
 	pcmpeqd	xmm5,xmm4
-	pshufd	xmm4,xmm1,0b1h
+	pshufd	xmm4,xmm1,1h
 	por	xmm4,xmm1
 	pshufd	xmm5,xmm5,0
 	pshufd	xmm3,xmm4,01eh
@@ -3528,7 +3528,7 @@ $L$point_add_affinex::
 	por	xmm5,xmm4
 
 	movdqu	xmm0,XMMWORD PTR[rbx]
-	pshufd	xmm3,xmm5,0b1h
+	pshufd	xmm3,xmm5,1h
 	movdqu	xmm1,XMMWORD PTR[16+rbx]
 	movdqu	xmm2,XMMWORD PTR[32+rbx]
 	por	xmm5,xmm3
@@ -3550,7 +3550,7 @@ DB	102,72,15,110,199
 	call	__ecp_nistz256_sqr_montx
 
 	pcmpeqd	xmm5,xmm4
-	pshufd	xmm4,xmm3,0b1h
+	pshufd	xmm4,xmm3,1h
 	mov	rdx,QWORD PTR[rbx]
 
 	mov	r9,r12

diff --git a/deps/openssl/asm/x64-win32-masm/sha/sha256-x86_64.asm b/deps/openssl/asm/x64-win32-masm/sha/sha256-x86_64.asm
@@ -1796,7 +1796,7 @@ $L$prologue_shaext::
 	movdqa	xmm7,XMMWORD PTR[((512-128))+rcx]
 
 	pshufd	xmm0,xmm1,01bh
-	pshufd	xmm1,xmm1,0b1h
+	pshufd	xmm1,xmm1,1h
 	pshufd	xmm2,xmm2,01bh
 	movdqa	xmm8,xmm7
 DB	102,15,58,15,202,8
@@ -1983,9 +1983,9 @@ DB	15,56,203,202
 	paddd	xmm1,xmm9
 	jnz	$L$oop_shaext
 
-	pshufd	xmm2,xmm2,0b1h
+	pshufd	xmm2,xmm2,1h
 	pshufd	xmm7,xmm1,01bh
-	pshufd	xmm1,xmm1,0b1h
+	pshufd	xmm1,xmm1,1h
 	punpckhqdq	xmm1,xmm2
 DB	102,15,58,15,215,8
 

diff --git a/deps/openssl/asm/x64-win32-masm/x86_64cpuid.asm b/deps/openssl/asm/x64-win32-masm/x86_64cpuid.asm
@@ -118,14 +118,6 @@ $L$intel::
 	shr	r10d,14
 	and	r10d,0fffh
 
-	cmp	r11d,7
-	jb	$L$nocacheinfo
-
-	mov	eax,7
-	xor	ecx,ecx
-	cpuid
-	mov	DWORD PTR[8+rdi],ebx
-
 $L$nocacheinfo::
 	mov	eax,1
 	cpuid
@@ -155,6 +147,15 @@ $L$generic::
 	or	r9d,ecx
 
 	mov	r10d,edx
+
+	cmp	r11d,7
+	jb	$L$no_extended_info
+	mov	eax,7
+	xor	ecx,ecx
+	cpuid
+	mov	DWORD PTR[8+rdi],ebx
+$L$no_extended_info::
+
 	bt	r9d,27
 	jnc	$L$clear_avx
 	xor	ecx,ecx

diff --git a/deps/openssl/asm/x86-elf-gas/x86cpuid.s b/deps/openssl/asm/x86-elf-gas/x86cpuid.s
@@ -20,10 +20,10 @@ OPENSSL_ia32_cpuid:
 	popl	%eax
 	xorl	%eax,%ecx
 	xorl	%eax,%eax
-	btl	$21,%ecx
-	jnc	.L000nocpuid
 	movl	20(%esp),%esi
 	movl	%eax,8(%esi)
+	btl	$21,%ecx
+	jnc	.L000nocpuid
 	.byte	0x0f,0xa2
 	movl	%eax,%edi
 	xorl	%eax,%eax
@@ -74,40 +74,32 @@ OPENSSL_ia32_cpuid:
 	andl	$4026531839,%edx
 	jmp	.L002generic
 .L001intel:
-	cmpl	$7,%edi
-	jb	.L003cacheinfo
-	movl	20(%esp),%esi
-	movl	$7,%eax
-	xorl	%ecx,%ecx
-	.byte	0x0f,0xa2
-	movl	%ebx,8(%esi)
-.L003cacheinfo:
 	cmpl	$4,%edi
-	movl	$-1,%edi
-	jb	.L004nocacheinfo
+	movl	$-1,%esi
+	jb	.L003nocacheinfo
 	movl	$4,%eax
 	movl	$0,%ecx
 	.byte	0x0f,0xa2
-	movl	%eax,%edi
-	shrl	$14,%edi
-	andl	$4095,%edi
-.L004nocacheinfo:
+	movl	%eax,%esi
+	shrl	$14,%esi
+	andl	$4095,%esi
+.L003nocacheinfo:
 	movl	$1,%eax
 	xorl	%ecx,%ecx
 	.byte	0x0f,0xa2
 	andl	$3220176895,%edx
 	cmpl	$0,%ebp
-	jne	.L005notintel
+	jne	.L004notintel
 	orl	$1073741824,%edx
 	andb	$15,%ah
 	cmpb	$15,%ah
-	jne	.L005notintel
+	jne	.L004notintel
 	orl	$1048576,%edx
-.L005notintel:
+.L004notintel:
 	btl	$28,%edx
 	jnc	.L002generic
 	andl	$4026531839,%edx
-	cmpl	$0,%edi
+	cmpl	$0,%esi
 	je	.L002generic
 	orl	$268435456,%edx
 	shrl	$16,%ebx
@@ -119,7 +111,15 @@ OPENSSL_ia32_cpuid:
 	andl	$4294965247,%ecx
 	movl	%edx,%esi
 	orl	%ecx,%ebp
-	btl	$27,%ecx
+	cmpl	$7,%edi
+	movl	20(%esp),%edi
+	jb	.L005no_extended_info
+	movl	$7,%eax
+	xorl	%ecx,%ecx
+	.byte	0x0f,0xa2
+	movl	%ebx,8(%edi)
+.L005no_extended_info:
+	btl	$27,%ebp
 	jnc	.L006clear_avx
 	xorl	%ecx,%ecx
 .byte	15,1,208
@@ -133,7 +133,6 @@ OPENSSL_ia32_cpuid:
 	andl	$4278190079,%esi
 .L006clear_avx:
 	andl	$4026525695,%ebp
-	movl	20(%esp),%edi
 	andl	$4294967263,8(%edi)
 .L007done:
 	movl	%esi,%eax