math/mpx-mul4-x86-sse2.S: Use `movdqa' to move between XMM registers.

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

diff --git a/math/mpx-mul4-x86-sse2.S b/math/mpx-mul4-x86-sse2.S
index a6613ed0da1fccd2b6bc2f0a0c95f64c26e9c0bf..0e87ff58263e086c39c919d94afbcd2dc2d861e1 100644 (file)
--- a/math/mpx-mul4-x86-sse2.S
+++ b/math/mpx-mul4-x86-sse2.S
@@ -276,7 +276,7 @@
        psrldq  \c1, 8                  // (0, y^*_1, 0, 0)
        psrldq  \c2, 4                  // (0, 0, y^*_2, 0)
   .ifnes "\h", "nil"
-       movdqu  \t, \c3
+       movdqa  \t, \c3
        pxor    \u, \u
   .endif
        pslldq  \c3, 12                 // (0, 0, 0, y^*_3)
@@ -302,6 +302,7 @@
        movd    xmm1, [edi +  4]        // (a_1, 0)
        movd    xmm2, [edi +  8]        // (a_2, 0)
        movd    xmm7, [edi + 12]        // (a_3, 0)
+
        paddq   xmm4, xmm0              // (c'_0 + a_0, c''_0)
        paddq   xmm5, xmm1              // (c'_1 + a_1, c''_1)
        paddq   xmm6, xmm2              // (c'_2 + a_2, c''_2 + a_3 b)
@@ -522,7 +523,7 @@ INTFUNC(mmul4)
        // of the sum U V + N Y to [EDI], leaving the remaining carry in
        // XMM4, XMM5, and XMM6.  The registers XMM0, XMM1, XMM2, XMM3, and
        // XMM7 are clobbered; the general-purpose registers are preserved.
-       stalloc 64                      // space for the carries
+       stalloc 48                      // space for the carries
   endprologue
 
        // Calculate W = U V, and leave it in the destination.  Stash the
@@ -547,7 +548,7 @@ INTFUNC(mmla4)
        // carry in XMM4, XMM5, and XMM6.  The registers XMM0, XMM1, XMM2,
        // XMM3, and XMM7 are clobbered; the general-purpose registers are
        // preserved.
-       stalloc 64                      // space for the carries
+       stalloc 48                      // space for the carries
   endprologue
 
        movd    xmm4, [edi +  0]
@@ -616,14 +617,14 @@ INTFUNC(mmla4)
        paddq   xmm6, [esp + 32]
 
        // And, with that, we're done.
-       stfree  64
+       stfree  48
        ret
 
 ENDFUNC
 
 INTFUNC(mont4)
        // On entry, EDI points to the destination buffer holding a packed
-       // value A; EBX points to a packed operand N; ESI points to an
+       // value W; EBX points to a packed operand N; ESI points to an
        // expanded operand M; and EDX points to a place to store an expanded
        // result Y (32 bytes, at a 16-byte boundary).
        //