math/mpx-mul4-x86-sse2.S: Optimize `squash'.

We can use `punpckldq' to assemble the 32-bit pieces, rather than a lot
of shifting to clear bits and then `por'.

diff --git a/math/mpx-mul4-x86-sse2.S b/math/mpx-mul4-x86-sse2.S
index 0e87ff58263e086c39c919d94afbcd2dc2d861e1..471b3f1f90cbc69cea42cad6ecdecd82285a081b 100644 (file)
--- a/math/mpx-mul4-x86-sse2.S
+++ b/math/mpx-mul4-x86-sse2.S
@@ -262,32 +262,24 @@
        psrldq  \t, 8                   // (y_2, 0)
        psrlq   \c3, 32                 // (floor(y_0/B), ?)
        paddq   \c3, \u                 // (y_1 + floor(y_0/B), ?)
-       pslldq  \c0, 12                 // (0, 0, 0, y^*_0)
        movdqa  \c1, \c3                // (y^*_1, ?, ?, ?)
        psrldq  \u, 8                   // (y_3, 0)
        psrlq   \c3, 32                 // (floor((y_1 B + y_0)/B^2, ?)
        paddq   \c3, \t                 // (y_2 + floor((y_1 B + y_0)/B^2, ?)
-       pslldq  \c1, 12                 // (0, 0, 0, y^*_1)
-       psrldq  \c0, 12                 // (y^*_0, 0, 0, 0)
-       movdqa  \c2, \c3                // (y^*_2, ?, ?, ?)
+       punpckldq \c0, \c3              // (y^*_0, y^*_2, ?, ?)
        psrlq   \c3, 32             // (floor((y_2 B^2 + y_1 B + y_0)/B^3, ?)
        paddq   \c3, \u       // (y_3 + floor((y_2 B^2 + y_1 B + y_0)/B^3, ?)
-       pslldq  \c2, 12                 // (0, 0, 0, y^*_2)
-       psrldq  \c1, 8                  // (0, y^*_1, 0, 0)
-       psrldq  \c2, 4                  // (0, 0, y^*_2, 0)
   .ifnes "\h", "nil"
        movdqa  \t, \c3
        pxor    \u, \u
   .endif
-       pslldq  \c3, 12                 // (0, 0, 0, y^*_3)
-       por     \c0, \c1                // (y^*_0, y^*_1, 0, 0)
-       por     \c2, \c3                // (0, 0, y^*_2, y^*_3)
-       por     \c0, \c2                // y mod B^4
+       punpckldq \c1, \c3              // (y^*_1, y^*_3, ?, ?)
   .ifnes "\h", "nil"
        psrlq   \t, 32                  // very high bits of y
        paddq   \h, \t
        punpcklqdq \h, \u               // carry up
   .endif
+       punpckldq \c0, \c1              // y mod B^4
 .endm
 
 .macro carryadd