X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/81bc2bb831d1fea04ea819b573bd6ba9ea08f3cf..8f2287ef5c05d496fcb9b012629af007fe56f897:/symm/rijndael-x86ish-aesni.S

diff --git a/symm/rijndael-x86ish-aesni.S b/symm/rijndael-x86ish-aesni.S
index e556aa53..6d9b3b22 100644
--- a/symm/rijndael-x86ish-aesni.S
+++ b/symm/rijndael-x86ish-aesni.S
@@ -25,20 +25,21 @@
 /// MA 02111-1307, USA.
 
 ///--------------------------------------------------------------------------
-/// External definitions.
+/// Preliminaries.
 
 #include "config.h"
 #include "asm-common.h"
 
+	.arch	.aes
+
 	.extern	F(abort)
 	.extern	F(rijndael_rcon)
 
+	.text
+
 ///--------------------------------------------------------------------------
 /// Main code.
 
-	.arch	.aes
-	.text
-
 /// The AESNI instructions implement a little-endian version of AES, but
 /// Catacomb's internal interface presents as big-endian so as to work better
 /// with things like GCM.  We therefore maintain the round keys in
@@ -61,6 +62,12 @@
 ///--------------------------------------------------------------------------
 /// Key setup.
 
+FUNC(rijndael_setup_x86ish_aesni_avx)
+	vzeroupper		      // avoid penalty on `legacy' XMM access
+  endprologue
+	// and drop through...
+ENDFUNC
+
 FUNC(rijndael_setup_x86ish_aesni)
 
 #define SI WHOLE(si)
@@ -205,16 +212,16 @@ FUNC(rijndael_setup_x86ish_aesni)
 	// Fourth word of the cycle, and seven or eight words of key.  Do a
 	// byte substitution.
 	movd	xmm0, eax
-	pshufd	xmm0, xmm0, SHUF(2, 1, 0, 3)
+	pshufd	xmm0, xmm0, SHUF(3, 0, 1, 2)
 	aeskeygenassist xmm1, xmm0, 0
 	movd	eax, xmm1
 	jmp	2f
 
 	// First word of the cycle.  This is the complicated piece.
 1:	movd	xmm0, eax
-	pshufd	xmm0, xmm0, SHUF(0, 3, 2, 1)
+	pshufd	xmm0, xmm0, SHUF(1, 2, 3, 0)
 	aeskeygenassist xmm1, xmm0, 0
-	pshufd	xmm1, xmm1, SHUF(2, 1, 0, 3)
+	pshufd	xmm1, xmm1, SHUF(3, 0, 1, 2)
 	movd	eax, xmm1
 	xor	al, [RCON]
 	inc	RCON
@@ -365,6 +372,12 @@ ENDFUNC
 /// Encrypting and decrypting blocks.
 
 .macro	encdec	op, aes, koff
+  FUNC(rijndael_\op\()_x86ish_aesni_avx)
+	vzeroupper			// avoid XMM penalties
+  endprologue
+	// and drop through...
+  ENDFUNC
+
   FUNC(rijndael_\op\()_x86ish_aesni)
 
 #if CPUFAM_X86