x86ish *.S: Use `stalloc' consistently to allocate space on the stack.

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

diff --git a/symm/chacha-x86ish-sse2.S b/symm/chacha-x86ish-sse2.S
index 8688996..974ec5b 100644 (file)
--- a/symm/chacha-x86ish-sse2.S
+++ b/symm/chacha-x86ish-sse2.S
@@ -25,24 +25,24 @@
 /// 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)
 
@@ -50,12 +50,12 @@ 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.
@@ -66,15 +66,15 @@ FUNC(chacha_core_x86ish_sse2)
 #  define SAVE0 xmm5
 #  define SAVE1 xmm6
 #  define SAVE2 xmm7
-#  define SAVE3 [esp]
-
-       push    ebp
-       mov     ebp, esp
-       sub     esp, 16
-       mov     IN, [ebp + 12]
-       mov     OUT, [ebp + 16]
-       and     esp, ~15
-       mov     NR, [ebp + 8]
+#  define SAVE3 [SP]
+
+       pushreg BP
+       setfp
+       stalloc 16
+       mov     IN, [BP + 12]
+       mov     OUT, [BP + 16]
+       and     SP, ~15
+       mov     NR, [BP + 8]
 #endif
 
 #if CPUFAM_AMD64 && ABI_SYSV
@@ -105,13 +105,15 @@ FUNC(chacha_core_x86ish_sse2)
 #  define IN rdx
 #  define OUT r8
 #  define SAVE0 xmm5
-#  define SAVE1 [rsp +  0]
-#  define SAVE2 [rsp + 16]
-#  define SAVE3 [rsp + 32]
+#  define SAVE1 [SP +  0]
+#  define SAVE2 [SP + 16]
+#  define SAVE3 [SP + 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.
        //
@@ -162,9 +164,9 @@ FUNC(chacha_core_x86ish_sse2)
 
        // 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
@@ -182,7 +184,7 @@ 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(1, 2, 3, 0)
 
        // Apply the diagonal quarterround to each of the columns
        // simultaneously.
@@ -213,9 +215,9 @@ FUNC(chacha_core_x86ish_sse2)
 
        // 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
@@ -224,7 +226,7 @@ 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(3, 0, 1, 2)
 
        // Decrement the loop counter and see if we should go round again.
        sub     NR, 2
@@ -245,11 +247,11 @@ FUNC(chacha_core_x86ish_sse2)
 
        // Tidy things up.
 #if CPUFAM_X86
-       mov     esp, ebp
-       pop     ebp
+       dropfp
+       popreg  BP
 #endif
 #if CPUFAM_AMD64 && ABI_WIN
-       add     rsp, 48 + 8
+       stfree  48 + 8
 #endif
 
        // And with that, we're done.