X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/0923a413958b0e778a3f059c76355ab58e5be414..1519ef6617fd7351bd40026b9eeb47c7f7f42234:/symm/salsa20-x86ish-sse2.S

diff --git a/symm/salsa20-x86ish-sse2.S b/symm/salsa20-x86ish-sse2.S
index ca677f17..26bab892 100644
--- a/symm/salsa20-x86ish-sse2.S
+++ b/symm/salsa20-x86ish-sse2.S
@@ -25,16 +25,24 @@
 /// MA 02111-1307, USA.
 
 ///--------------------------------------------------------------------------
-/// External definitions.
+/// Preliminaries.
 
 #include "config.h"
 #include "asm-common.h"
 
+	.text
+
 ///--------------------------------------------------------------------------
 /// Main code.
 
-	.arch pentium4
-	.text
+FUNC(salsa20_core_x86ish_avx)
+	.arch	.avx
+	vzeroupper
+  endprologue
+	// drop through...
+ENDFUNC
+
+	.arch	pentium4
 
 FUNC(salsa20_core_x86ish_sse2)
 
@@ -42,12 +50,12 @@ FUNC(salsa20_core_x86ish_sse2)
 
 #if CPUFAM_X86
 	// Arguments come in on the stack, and will need to be collected.  We
-	// we can get away with just the scratch registers for integer work,
-	// but we'll run out of XMM registers and will need some properly
-	// aligned space which we'll steal from the stack.  I don't trust the
-	// stack pointer's alignment, so I'll have to mask the stack pointer,
-	// which in turn means I'll need to keep track of the old value.
-	// Hence I'm making a full i386-style stack frame here.
+	// can get away with just the scratch registers for integer work, but
+	// we'll run out of XMM registers and will need some properly aligned
+	// space which we'll steal from the stack.  I don't trust the stack
+	// pointer's alignment, so I'll have to mask the stack pointer, which
+	// in turn means I'll need to keep track of the old value.  Hence I'm
+	// making a full i386-style stack frame here.
 	//
 	// The Windows and SysV ABIs are sufficiently similar that we don't
 	// need to worry about the differences here.
@@ -57,16 +65,16 @@ FUNC(salsa20_core_x86ish_sse2)
 #  define OUT edx
 #  define SAVE0 xmm6
 #  define SAVE1 xmm7
-#  define SAVE2 [esp + 0]
-#  define SAVE3 [esp + 16]
-
-	pushreg	ebp
-	setfp	ebp
-	sub	esp, 32
-	mov	IN, [ebp + 12]
-	mov	OUT, [ebp + 16]
-	and	esp, ~15
-	mov	NR, [ebp + 8]
+#  define SAVE2 [SP + 0]
+#  define SAVE3 [SP + 16]
+
+	pushreg	BP
+	setfp
+	stalloc	32
+	mov	IN, [BP + 12]
+	mov	OUT, [BP + 16]
+	and	SP, ~15
+	mov	NR, [BP + 8]
 #endif
 
 #if CPUFAM_AMD64 && ABI_SYSV
@@ -99,8 +107,8 @@ FUNC(salsa20_core_x86ish_sse2)
 #  define OUT r8
 #  define SAVE0 xmm6
 #  define SAVE1 xmm7
-#  define SAVE2 [rsp + 32]
-#  define SAVE3 [rsp + 48]
+#  define SAVE2 [SP + 32]
+#  define SAVE3 [SP + 48]
 
 	stalloc	64 + 8
 	savexmm	xmm6, 0
@@ -172,7 +180,7 @@ FUNC(salsa20_core_x86ish_sse2)
 	// d ^= (c + b) <<< 13
 	movdqa	xmm4, xmm2
 	paddd	xmm4, xmm1
-	 pshufd	xmm1, xmm1, SHUF(2, 1, 0, 3)
+	 pshufd	xmm1, xmm1, SHUF(3, 0, 1, 2)
 	movdqa	xmm5, xmm4
 	pslld	xmm4, 13
 	psrld	xmm5, 19
@@ -181,9 +189,9 @@ FUNC(salsa20_core_x86ish_sse2)
 
 	// a ^= (d + c) <<< 18
 	movdqa	xmm4, xmm3
-	 pshufd	xmm3, xmm3, SHUF(0, 3, 2, 1)
+	 pshufd	xmm3, xmm3, SHUF(1, 2, 3, 0)
 	paddd	xmm4, xmm2
-	 pshufd	xmm2, xmm2, SHUF(1, 0, 3, 2)
+	 pshufd	xmm2, xmm2, SHUF(2, 3, 0, 1)
 	movdqa	xmm5, xmm4
 	pslld	xmm4, 18
 	psrld	xmm5, 14
@@ -227,7 +235,7 @@ FUNC(salsa20_core_x86ish_sse2)
 	// d ^= (c + b) <<< 13
 	movdqa	xmm4, xmm2
 	paddd	xmm4, xmm3
-	 pshufd	xmm3, xmm3, SHUF(2, 1, 0, 3)
+	 pshufd	xmm3, xmm3, SHUF(3, 0, 1, 2)
 	movdqa	xmm5, xmm4
 	pslld	xmm4, 13
 	psrld	xmm5, 19
@@ -236,9 +244,9 @@ FUNC(salsa20_core_x86ish_sse2)
 
 	// a ^= (d + c) <<< 18
 	movdqa	xmm4, xmm1
-	 pshufd	xmm1, xmm1, SHUF(0, 3, 2, 1)
+	 pshufd	xmm1, xmm1, SHUF(1, 2, 3, 0)
 	paddd	xmm4, xmm2
-	 pshufd	xmm2, xmm2, SHUF(1, 0, 3, 2)
+	 pshufd	xmm2, xmm2, SHUF(2, 3, 0, 1)
 	movdqa	xmm5, xmm4
 	pslld	xmm4, 18
 	psrld	xmm5, 14
@@ -262,9 +270,9 @@ FUNC(salsa20_core_x86ish_sse2)
 	// input.  This can be done by juggling values in registers, with the
 	// following fancy footwork: some row rotations, a transpose, and
 	// some more rotations.
-	pshufd	xmm1, xmm1, SHUF(2, 1, 0, 3)	//  3,  4,  9, 14
-	pshufd	xmm2, xmm2, SHUF(1, 0, 3, 2)	//  2,  7,  8, 13
-	pshufd	xmm3, xmm3, SHUF(0, 3, 2, 1)	//  1,  6, 11, 12
+	pshufd	xmm1, xmm1, SHUF(3, 0, 1, 2)	//  3,  4,  9, 14
+	pshufd	xmm2, xmm2, SHUF(2, 3, 0, 1)	//  2,  7,  8, 13
+	pshufd	xmm3, xmm3, SHUF(1, 2, 3, 0)	//  1,  6, 11, 12
 
 	movdqa	xmm4, xmm0
 	movdqa	xmm5, xmm3
@@ -280,9 +288,9 @@ FUNC(salsa20_core_x86ish_sse2)
 	punpckhdq xmm1, xmm3			//  5,  6,  7,  4
 	punpckhdq xmm2, xmm5			// 15, 12, 13, 14
 
-	pshufd	xmm1, xmm1, SHUF(2, 1, 0, 3)	//  4,  5,  6,  7
-	pshufd	xmm4, xmm4, SHUF(1, 0, 3, 2)	//  8,  9, 10, 11
-	pshufd	xmm2, xmm2, SHUF(0, 3, 2, 1)	// 12, 13, 14, 15
+	pshufd	xmm1, xmm1, SHUF(3, 0, 1, 2)	//  4,  5,  6,  7
+	pshufd	xmm4, xmm4, SHUF(2, 3, 0, 1)	//  8,  9, 10, 11
+	pshufd	xmm2, xmm2, SHUF(1, 2, 3, 0)	// 12, 13, 14, 15
 
 	// Finally we have to write out the result.
 	movdqu	[OUT +  0], xmm0
@@ -293,11 +301,11 @@ FUNC(salsa20_core_x86ish_sse2)
 	// Tidy things up.
 #if CPUFAM_X86
 	dropfp
-	popreg	ebp
+	popreg	BP
 #endif
 #if CPUFAM_AMD64 && ABI_WIN
 	rstrxmm	xmm6, 0
-	rsrrxmm	xmm7, 16
+	rstrxmm	xmm7, 16
 	stfree	64 + 8
 #endif