X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/1da1ed6a5815deef6c33d74f1eb3c856793df3e5..a13b573023a59dc1e2febef1142aa4b1d1af67b7:/symm/chacha-x86ish-sse2.S

diff --git a/symm/chacha-x86ish-sse2.S b/symm/chacha-x86ish-sse2.S
index f36bf90f..a7ff68b5 100644
--- a/symm/chacha-x86ish-sse2.S
+++ b/symm/chacha-x86ish-sse2.S
@@ -31,18 +31,10 @@
 #include "asm-common.h"
 
 ///--------------------------------------------------------------------------
-/// Local utilities.
-
-// Magic constants for shuffling.
-#define ROTL 0x93
-#define ROT2 0x4e
-#define ROTR 0x39
-
-///--------------------------------------------------------------------------
 /// Main code.
 
 	.arch pentium4
-	.section .text
+	.text
 
 FUNC(chacha_core_x86ish_sse2)
 
@@ -110,6 +102,8 @@ FUNC(chacha_core_x86ish_sse2)
 #  define SAVE3 [rsp + 32]
 
 	sub	rsp, 48 + 8
+	  .seh_stackalloc 48 + 8
+  .seh_endprologue
 #endif
 
 	// First job is to slurp the matrix into XMM registers.  Be careful:
@@ -131,7 +125,7 @@ FUNC(chacha_core_x86ish_sse2)
 	movdqa	SAVE2, xmm2
 	movdqa	SAVE3, xmm3
 
-loop:
+0:
 	// Apply a column quarterround to each of the columns simultaneously.
 	// Alas, there doesn't seem to be a packed doubleword rotate, so we
 	// have to synthesize it.
@@ -162,9 +156,9 @@ loop:
 
 	// c += d; b ^= c; b <<<=  7
 	paddd	xmm2, xmm3
-	pshufd	xmm3, xmm3, ROTL
+	 pshufd	xmm3, xmm3, SHUF(2, 1, 0, 3)
 	pxor	xmm1, xmm2
-	pshufd	xmm2, xmm2, ROT2
+	 pshufd	xmm2, xmm2, SHUF(1, 0, 3, 2)
 	movdqa	xmm4, xmm1
 	pslld	xmm1, 7
 	psrld	xmm4, 25
@@ -182,7 +176,7 @@ loop:
 	//
 	// The shuffles have quite high latency, so they've mostly been
 	// pushed upwards.  The remaining one can't be moved, though.
-	pshufd	xmm1, xmm1, ROTR
+	pshufd	xmm1, xmm1, SHUF(0, 3, 2, 1)
 
 	// Apply the diagonal quarterround to each of the columns
 	// simultaneously.
@@ -213,9 +207,9 @@ loop:
 
 	// c += d; b ^= c; b <<<=  7
 	paddd	xmm2, xmm3
-	pshufd	xmm3, xmm3, ROTR
+	 pshufd	xmm3, xmm3, SHUF(0, 3, 2, 1)
 	pxor	xmm1, xmm2
-	pshufd	xmm2, xmm2, ROT2
+	 pshufd	xmm2, xmm2, SHUF(1, 0, 3, 2)
 	movdqa	xmm4, xmm1
 	pslld	xmm1, 7
 	psrld	xmm4, 25
@@ -224,11 +218,11 @@ loop:
 	// 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, ROTL
+	pshufd	xmm1, xmm1, SHUF(2, 1, 0, 3)
 
 	// Decrement the loop counter and see if we should go round again.
 	sub	NR, 2
-	ja	loop
+	ja	0b
 
 	// Almost there.  Firstly, the feedforward addition.
 	paddd	xmm0, SAVE0