/// MA 02111-1307, USA.
///--------------------------------------------------------------------------
-/// External definitions.
+/// Preliminaries.
#include "config.h"
#include "asm-common.h"
-///--------------------------------------------------------------------------
-/// Local utilities.
-
-// Magic constants for shuffling.
-#define ROTL 0x93
-#define ROT2 0x4e
-#define ROTR 0x39
+ .text
///--------------------------------------------------------------------------
/// Main code.
- .arch pentium4
- .section .text
+FUNC(chacha_core_x86ish_avx)
+ .arch .avx
+ vzeroupper
+ endprologue
+ // drop through...
+ENDFUNC
+
+ .arch pentium4
FUNC(chacha_core_x86ish_sse2)
#if CPUFAM_X86
// Arguments come in on the stack, and will need to be collected. We
- // we can get away with just the scratch registers for integer work,
- // but we'll run out of XMM registers and will need some properly
- // aligned space which we'll steal from the stack. I don't trust the
- // stack pointer's alignment, so I'll have to mask the stack pointer,
- // which in turn means I'll need to keep track of the old value.
- // Hence I'm making a full i386-style stack frame here.
+ // can get away with just the scratch registers for integer work, but
+ // we'll run out of XMM registers and will need some properly aligned
+ // space which we'll steal from the stack. I don't trust the stack
+ // pointer's alignment, so I'll have to mask the stack pointer, which
+ // in turn means I'll need to keep track of the old value. Hence I'm
+ // making a full i386-style stack frame here.
//
// The Windows and SysV ABIs are sufficiently similar that we don't
// need to worry about the differences here.
# define SAVE2 xmm7
# define SAVE3 [esp]
- push ebp
- mov ebp, esp
+ pushreg ebp
+ setfp
sub esp, 16
mov IN, [ebp + 12]
mov OUT, [ebp + 16]
# define SAVE2 [rsp + 16]
# define SAVE3 [rsp + 32]
- sub rsp, 48 + 8
+ 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.
//
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.
// c += d; b ^= c; b <<<= 7
paddd xmm2, xmm3
- pshufd xmm3, xmm3, ROTL
+ pshufd xmm3, xmm3, SHUF(3, 0, 1, 2)
pxor xmm1, xmm2
- pshufd xmm2, xmm2, ROT2
+ pshufd xmm2, xmm2, SHUF(2, 3, 0, 1)
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(1, 2, 3, 0)
// 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(1, 2, 3, 0)
pxor xmm1, xmm2
- pshufd xmm2, xmm2, ROT2
+ pshufd xmm2, xmm2, SHUF(2, 3, 0, 1)
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(3, 0, 1, 2)
// 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
// 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.