base/asm-common.h, *-x86ish-*.S: Centralize SSE shuffling constants.

[catacomb] / symm / chacha-x86ish-sse2.S

diff --git a/symm/chacha-x86ish-sse2.S b/symm/chacha-x86ish-sse2.S
index b58cdcea68cfa84a32e867ef309cac8d8ccb9e18..a7ff68b502fb2fbd28dd327bbd982ad45137d429 100644 (file)
--- a/symm/chacha-x86ish-sse2.S
+++ b/symm/chacha-x86ish-sse2.S
@@ -30,14 +30,6 @@
 #include "config.h"
 #include "asm-common.h"
 
-///--------------------------------------------------------------------------
-/// Local utilities.
-
-// Magic constants for shuffling.
-#define ROTL 0x93
-#define ROT2 0x4e
-#define ROTR 0x39
-
 ///--------------------------------------------------------------------------
 /// Main code.
 
@@ -164,9 +156,9 @@ FUNC(chacha_core_x86ish_sse2)
 
        // 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
@@ -184,7 +176,7 @@ 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, ROTR
+       pshufd  xmm1, xmm1, SHUF(0, 3, 2, 1)
 
        // Apply the diagonal quarterround to each of the columns
        // simultaneously.
@@ -215,9 +207,9 @@ FUNC(chacha_core_x86ish_sse2)
 
        // 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
@@ -226,7 +218,7 @@ 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, ROTL
+       pshufd  xmm1, xmm1, SHUF(2, 1, 0, 3)
 
        // Decrement the loop counter and see if we should go round again.
        sub     NR, 2