X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/8276a9c565bc29afbc6cc120dfde0801c06c2bbd..7b0d1a63587f3cb1ae3bb8b248bbb1b82bdca7bd:/symm/chacha-x86ish-sse2.S

diff --git a/symm/chacha-x86ish-sse2.S b/symm/chacha-x86ish-sse2.S
index 2dab283b..3fb623af 100644
--- a/symm/chacha-x86ish-sse2.S
+++ b/symm/chacha-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(chacha_core_x86ish_avx)
+	.arch	.avx
+	vzeroupper
+  endprologue
+	// drop through...
+ENDFUNC
+
+	.arch	pentium4
 
 FUNC(chacha_core_x86ish_sse2)
 
@@ -61,7 +69,7 @@ FUNC(chacha_core_x86ish_sse2)
 #  define SAVE3 [esp]
 
 	pushreg	ebp
-	setfp	ebp
+	setfp
 	sub	esp, 16
 	mov	IN, [ebp + 12]
 	mov	OUT, [ebp + 16]
@@ -156,9 +164,9 @@ FUNC(chacha_core_x86ish_sse2)
 
 	// c += d; b ^= c; b <<<=  7
 	paddd	xmm2, xmm3
-	 pshufd	xmm3, xmm3, SHUF(2, 1, 0, 3)
+	 pshufd	xmm3, xmm3, SHUF(3, 0, 1, 2)
 	pxor	xmm1, xmm2
-	 pshufd	xmm2, xmm2, SHUF(1, 0, 3, 2)
+	 pshufd	xmm2, xmm2, SHUF(2, 3, 0, 1)
 	movdqa	xmm4, xmm1
 	pslld	xmm1, 7
 	psrld	xmm4, 25
@@ -176,7 +184,7 @@ FUNC(chacha_core_x86ish_sse2)
 	//
 	// The shuffles have quite high latency, so they've mostly been
 	// pushed upwards.  The remaining one can't be moved, though.
-	pshufd	xmm1, xmm1, SHUF(0, 3, 2, 1)
+	pshufd	xmm1, xmm1, SHUF(1, 2, 3, 0)
 
 	// Apply the diagonal quarterround to each of the columns
 	// simultaneously.
@@ -207,9 +215,9 @@ FUNC(chacha_core_x86ish_sse2)
 
 	// c += d; b ^= c; b <<<=  7
 	paddd	xmm2, xmm3
-	 pshufd	xmm3, xmm3, SHUF(0, 3, 2, 1)
+	 pshufd	xmm3, xmm3, SHUF(1, 2, 3, 0)
 	pxor	xmm1, xmm2
-	 pshufd	xmm2, xmm2, SHUF(1, 0, 3, 2)
+	 pshufd	xmm2, xmm2, SHUF(2, 3, 0, 1)
 	movdqa	xmm4, xmm1
 	pslld	xmm1, 7
 	psrld	xmm4, 25
@@ -218,7 +226,7 @@ FUNC(chacha_core_x86ish_sse2)
 	// Finally, finish off undoing the transpose, and we're done for this
 	// doubleround.  Again, most of this was done above so we don't have
 	// to wait for the shuffles.
-	pshufd	xmm1, xmm1, SHUF(2, 1, 0, 3)
+	pshufd	xmm1, xmm1, SHUF(3, 0, 1, 2)
 
 	// Decrement the loop counter and see if we should go round again.
 	sub	NR, 2