#include "config.h"
#include "asm-common.h"
-///--------------------------------------------------------------------------
-/// Local utilities.
-
-// Magic constants for shuffling.
-#define ROTL 0x93
-#define ROT2 0x4e
-#define ROTR 0x39
-
///--------------------------------------------------------------------------
/// Main code.
# define SAVE2 xmm7
# define SAVE3 [esp]
- push ebp
- mov ebp, esp
+ pushreg ebp
+ setfp ebp
sub esp, 16
mov IN, [ebp + 12]
mov OUT, [ebp + 16]
# define SAVE2 [rsp + 16]
# define SAVE3 [rsp + 32]
- sub rsp, 48 + 8
- .seh_stackalloc 48 + 8
- .seh_endprologue
+ stalloc 48 + 8
#endif
+ endprologue
+
// First job is to slurp the matrix into XMM registers. Be careful:
// the input matrix isn't likely to be properly aligned.
//
// 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
//
// 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.
// 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
// 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
// Tidy things up.
#if CPUFAM_X86
- mov esp, ebp
- pop ebp
+ dropfp
+ popreg ebp
#endif
#if CPUFAM_AMD64 && ABI_WIN
- add rsp, 48 + 8
+ stfree 48 + 8
#endif
// And with that, we're done.