Merge branch '2.4.x' into 2.5.x

[catacomb] / math / mpx-mul4-x86-sse2.S

diff --git a/math/mpx-mul4-x86-sse2.S b/math/mpx-mul4-x86-sse2.S
index 591a7a8..904c0d0 100644 (file)
--- a/math/mpx-mul4-x86-sse2.S
+++ b/math/mpx-mul4-x86-sse2.S
@@ -24,15 +24,13 @@
 /// MA 02111-1307, USA.
 
 ///--------------------------------------------------------------------------
-/// External definitions.
+/// Preliminaries.
 
 #include "config.h"
 #include "asm-common.h"
 
-///--------------------------------------------------------------------------
-/// Prologue.
-
        .arch   pentium4
+
        .text
 
 ///--------------------------------------------------------------------------
@@ -103,7 +101,7 @@
   .ifnes "\d3", "nil"
        movdqa  \d3, [\s + 16]          // (s'_2, s'_3; s''_2, s''_3)
   .endif
-       pshufd  \d0, \d0, SHUF(3, 0, 3, 0) // (r_i, ?; r_i, ?)
+       pshufd  \d0, \d0, SHUF(0, 3, 0, 3) // (r_i, ?; r_i, ?)
   .ifnes "\d1", "nil"
        psrldq  \d1, 4                  // (s'_1, s''_0; s''_1, 0)
   .endif
@@ -171,7 +169,7 @@
        // carry registers.  On completion, XMM3 is clobbered.  If CC is
        // `nil', then the contribution which would have been added to it is
        // left in C.
-       pshufd  xmm3, \c, SHUF(2, 3, 3, 3) // (?, ?; ?, t = c'' mod B)
+       pshufd  xmm3, \c, SHUF(3, 3, 3, 2) // (?, ?; ?, t = c'' mod B)
        psrldq  xmm3, 12                // (t, 0; 0, 0) = (t, 0)
        pslldq  xmm3, 2                 // (t b; 0)
        paddq   \c, xmm3                // (c' + t b; c'')
@@ -209,11 +207,11 @@
        punpcklwd \c, \z                // (c'_0, c''_0; c'_1, c''_1)
        punpckhwd \d, \z                // (c'_2, c''_2; c'_3, c''_3)
   .endif
-       pshufd  \a, \a, SHUF(3, 1, 2, 0) // (a'_0, a'_1; a''_0, a''_1)
-       pshufd  \b, \b, SHUF(3, 1, 2, 0) // (a'_2, a'_3; a''_2, a''_3)
+       pshufd  \a, \a, SHUF(0, 2, 1, 3) // (a'_0, a'_1; a''_0, a''_1)
+       pshufd  \b, \b, SHUF(0, 2, 1, 3) // (a'_2, a'_3; a''_2, a''_3)
   .ifnes "\c", "nil"
-       pshufd  \c, \c, SHUF(3, 1, 2, 0) // (c'_0, c'_1; c''_0, c''_1)
-       pshufd  \d, \d, SHUF(3, 1, 2, 0) // (c'_2, c'_3; c''_2, c''_3)
+       pshufd  \c, \c, SHUF(0, 2, 1, 3) // (c'_0, c'_1; c''_0, c''_1)
+       pshufd  \d, \d, SHUF(0, 2, 1, 3) // (c'_2, c'_3; c''_2, c''_3)
   .endif
 .endm
 
@@ -707,7 +705,7 @@ FUNC(mpx_umul4_x86_sse2)
        pushreg ebx
        pushreg esi
        pushreg edi
-       setfp   ebp
+       setfp
        and     esp, ~15
        sub     esp, 32
   endprologue
@@ -822,7 +820,7 @@ FUNC(mpxmont_mul4_x86_sse2)
        pushreg ebx
        pushreg esi
        pushreg edi
-       setfp   ebp
+       setfp
        and     esp, ~15
        sub     esp, 112
   endprologue
@@ -968,7 +966,7 @@ FUNC(mpxmont_redc4_x86_sse2)
        pushreg ebx
        pushreg esi
        pushreg edi
-       setfp   ebp
+       setfp
        and     esp, ~15
        sub     esp, 76
   endprologue
@@ -1097,7 +1095,7 @@ ENDFUNC
        pushreg ebx
        pushreg esi
        pushreg edi
-       setfp   ebp
+       setfp
        and     esp, ~15
        sub     esp, 3*32 + 4*4
   endprologue
@@ -1214,6 +1212,18 @@ FUNC(test_mul4)
        testepilogue
 ENDFUNC
 
+FUNC(test_mul4zc)
+       testprologue [ebp + 36]
+       testldcarry [ebp + 24]
+       testexpand nil, [ebp + 32]
+       mov     edi, [ebp + 20]
+       testtop nil, [ebp + 28]
+       call    mul4zc
+       testtail [ebp + 40]
+       testcarryout [ebp + 24]
+       testepilogue
+ENDFUNC
+
 FUNC(test_mla4)
        testprologue [ebp + 36]
        testldcarry [ebp + 24]
@@ -1226,6 +1236,18 @@ FUNC(test_mla4)
        testepilogue
 ENDFUNC
 
+FUNC(test_mla4zc)
+       testprologue [ebp + 36]
+       testldcarry [ebp + 24]
+       testexpand nil, [ebp + 32]
+       mov     edi, [ebp + 20]
+       testtop nil, [ebp + 28]
+       call    mla4zc
+       testtail [ebp + 40]
+       testcarryout [ebp + 24]
+       testepilogue
+ENDFUNC
+
 FUNC(test_mmul4)
        testprologue [ebp + 48]
        testexpand [ebp + 40], [ebp + 44]