symm/salsa20-*.S: Optimize the output permutations.

[catacomb] / symm / rijndael-x86ish-aesni.S

diff --git a/symm/rijndael-x86ish-aesni.S b/symm/rijndael-x86ish-aesni.S
index 5194c17..5293d12 100644 (file)
--- a/symm/rijndael-x86ish-aesni.S
+++ b/symm/rijndael-x86ish-aesni.S
@@ -34,14 +34,6 @@
        .globl  F(rijndael_rcon)
 
 ///--------------------------------------------------------------------------
-/// Local utilities.
-
-// Magic constants for shuffling.
-#define ROTL 0x93
-#define ROT2 0x4e
-#define ROTR 0x39
-
-///--------------------------------------------------------------------------
 /// Main code.
 
        .arch   .aes
@@ -165,7 +157,10 @@ FUNC(rijndael_setup_x86ish_aesni)
        // We'll need the index registers, which belong to the caller in this
        // ABI.
        push    rsi
+         .seh_pushreg rsi
        push    rdi
+         .seh_pushreg rdi
+  .seh_endprologue
 
        // Move arguments to more useful places.
        mov     SI, r8                  // key material
@@ -229,16 +224,16 @@ 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, ROTL
+       pshufd  xmm0, xmm0, SHUF(2, 1, 0, 3)
        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, ROTR
+       pshufd  xmm0, xmm0, SHUF(0, 3, 2, 1)
        aeskeygenassist xmm1, xmm0, 0
-       pshufd  xmm1, xmm1, ROTL
+       pshufd  xmm1, xmm1, SHUF(2, 1, 0, 3)
        movd    eax, xmm1
        xor     al, [RCON]
        inc     RCON
@@ -391,18 +386,14 @@ ENDFUNC
 .macro encdec  op, aes, koff
   FUNC(rijndael_\op\()_x86ish_aesni)
 
-       // Find the magic endianness-swapping table.
-       ldgot   ecx
-       movdqa  xmm5, [INTADDR(endswap_tab, ecx)]
-
 #if CPUFAM_X86
        // Arguments come in on the stack, and need to be collected.  We
        // don't have a shortage of registers.
 
-#  define K ecx
+#  define K eax
 #  define SRC edx
 #  define DST edx
-#  define NR eax
+#  define NR ecx
 
        mov     K, [esp + 4]
        mov     SRC, [esp + 8]
@@ -424,8 +415,13 @@ ENDFUNC
 #  define SRC rdx
 #  define DST r8
 #  define NR eax
+  .seh_endprologue
 #endif
 
+       // Find the magic endianness-swapping table.
+       ldgot   ecx
+       movdqa  xmm5, [INTADDR(endswap_tab, ecx)]
+
        // Initial setup.
        movdqu  xmm0, [SRC]
        pshufb  xmm0, xmm5
@@ -436,6 +432,9 @@ ENDFUNC
        movdqu  xmm1, [K]
        add     K, 16
        pxor    xmm0, xmm1
+#if CPUFAM_X86
+       mov     DST, [esp + 12]
+#endif
 
        // Dispatch to the correct code.
        cmp     NR, 10
@@ -513,9 +512,6 @@ ENDFUNC
 
        // Unpermute the ciphertext block and store it.
        pshufb  xmm0, xmm5
-#if CPUFAM_X86
-       mov     DST, [esp + 12]
-#endif
        movdqu  [DST], xmm0
 
        // And we're done.
@@ -546,8 +542,6 @@ bogus:      callext F(abort)
 0:     hlt
        jmp     0b
 
-       gotaux  ecx
-
 ///--------------------------------------------------------------------------
 /// Data tables.