[PATCH] mpi/amd64: use SSE2 for shifting instead of MMX

[Jussi Kivilinna]

* mpi/amd64/mpih-lshift.S: Convert to SSE2.
* mpi/amd64/mpih-rshift.S: Ditto.
--

On current Intel processors, MMX instructions is slower than SSE2. Switch
lshift and rshift functions to use SSE2 registers instead of MMX.

Signed-off-by: Jussi Kivilinna <jussi.kivilinna at iki.fi>
---
 0 files changed

diff --git a/mpi/amd64/mpih-lshift.S b/mpi/amd64/mpih-lshift.S
index 9e8979b10..009bece64 100644
--- a/mpi/amd64/mpih-lshift.S
+++ b/mpi/amd64/mpih-lshift.S
@@ -43,37 +43,37 @@
 	.globl C_SYMBOL_NAME(_gcry_mpih_lshift)
 C_SYMBOL_NAME(_gcry_mpih_lshift:)
 	FUNC_ENTRY()
-	movq	-8(%rsi,%rdx,8), %mm7
-	movd	%ecx, %mm1
+	/* Note: %xmm6 and %xmm7 not used for WIN64 ABI compatibility. */
+	movq	-8(%rsi,%rdx,8), %xmm4
+	movd	%ecx, %xmm1
 	movl	$64, %eax
 	subl	%ecx, %eax
-	movd	%eax, %mm0
-	movq	%mm7, %mm3
-	psrlq	%mm0, %mm7
-	movd	%mm7, %rax
+	movd	%eax, %xmm0
+	movq	%xmm4, %xmm3
+	psrlq	%xmm0, %xmm4
+	movd	%xmm4, %rax
 	subq	$2, %rdx
 	jl	.Lendo
 
 	ALIGN(4)			/* minimal alignment for claimed speed */
-.Loop:	movq	(%rsi,%rdx,8), %mm6
-	movq	%mm6, %mm2
-	psrlq	%mm0, %mm6
-	psllq	%mm1, %mm3
-	por	%mm6, %mm3
-	movq	%mm3, 8(%rdi,%rdx,8)
+.Loop:	movq	(%rsi,%rdx,8), %xmm5
+	movq	%xmm5, %xmm2
+	psrlq	%xmm0, %xmm5
+	psllq	%xmm1, %xmm3
+	por	%xmm5, %xmm3
+	movq	%xmm3, 8(%rdi,%rdx,8)
 	je	.Lende
-	movq	-8(%rsi,%rdx,8), %mm7
-	movq	%mm7, %mm3
-	psrlq	%mm0, %mm7
-	psllq	%mm1, %mm2
-	por	%mm7, %mm2
-	movq	%mm2, (%rdi,%rdx,8)
+	movq	-8(%rsi,%rdx,8), %xmm4
+	movq	%xmm4, %xmm3
+	psrlq	%xmm0, %xmm4
+	psllq	%xmm1, %xmm2
+	por	%xmm4, %xmm2
+	movq	%xmm2, (%rdi,%rdx,8)
 	subq	$2, %rdx
 	jge	.Loop
 
-.Lendo:	movq	%mm3, %mm2
-.Lende:	psllq	%mm1, %mm2
-	movq	%mm2, (%rdi)
-	emms
+.Lendo:	movq	%xmm3, %xmm2
+.Lende:	psllq	%xmm1, %xmm2
+	movq	%xmm2, (%rdi)
 	FUNC_EXIT()
 	ret
diff --git a/mpi/amd64/mpih-rshift.S b/mpi/amd64/mpih-rshift.S
index 7bd594216..db31060de 100644
--- a/mpi/amd64/mpih-rshift.S
+++ b/mpi/amd64/mpih-rshift.S
@@ -43,14 +43,15 @@
 	.globl C_SYMBOL_NAME(_gcry_mpih_rshift)
 C_SYMBOL_NAME(_gcry_mpih_rshift:)
 	FUNC_ENTRY()
-	movq	(%rsi), %mm7
-	movd	%ecx, %mm1
+	/* Note: %xmm6 and %xmm7 not used for WIN64 ABI compatibility. */
+	movq	(%rsi), %xmm4
+	movd	%ecx, %xmm1
 	movl	$64, %eax
 	subl	%ecx, %eax
-	movd	%eax, %mm0
-	movq	%mm7, %mm3
-	psllq	%mm0, %mm7
-	movd	%mm7, %rax
+	movd	%eax, %xmm0
+	movq	%xmm4, %xmm3
+	psllq	%xmm0, %xmm4
+	movd	%xmm4, %rax
 	leaq	(%rsi,%rdx,8), %rsi
 	leaq	(%rdi,%rdx,8), %rdi
 	negq	%rdx
@@ -58,25 +59,24 @@ C_SYMBOL_NAME(_gcry_mpih_rshift:)
 	jg	.Lendo
 
 	ALIGN(4)			/* minimal alignment for claimed speed */
-.Loop:	movq	-8(%rsi,%rdx,8), %mm6
-	movq	%mm6, %mm2
-	psllq	%mm0, %mm6
-	psrlq	%mm1, %mm3
-	por	%mm6, %mm3
-	movq	%mm3, -16(%rdi,%rdx,8)
+.Loop:	movq	-8(%rsi,%rdx,8), %xmm5
+	movq	%xmm5, %xmm2
+	psllq	%xmm0, %xmm5
+	psrlq	%xmm1, %xmm3
+	por	%xmm5, %xmm3
+	movq	%xmm3, -16(%rdi,%rdx,8)
 	je	.Lende
-	movq	(%rsi,%rdx,8), %mm7
-	movq	%mm7, %mm3
-	psllq	%mm0, %mm7
-	psrlq	%mm1, %mm2
-	por	%mm7, %mm2
-	movq	%mm2, -8(%rdi,%rdx,8)
+	movq	(%rsi,%rdx,8), %xmm4
+	movq	%xmm4, %xmm3
+	psllq	%xmm0, %xmm4
+	psrlq	%xmm1, %xmm2
+	por	%xmm4, %xmm2
+	movq	%xmm2, -8(%rdi,%rdx,8)
 	addq	$2, %rdx
 	jle	.Loop
 
-.Lendo:	movq	%mm3, %mm2
-.Lende:	psrlq	%mm1, %mm2
-	movq	%mm2, -8(%rdi)
-	emms
+.Lendo:	movq	%xmm3, %xmm2
+.Lende:	psrlq	%xmm1, %xmm2
+	movq	%xmm2, -8(%rdi)
 	FUNC_EXIT()
 	ret