math/mpx-mul4-x86-sse2.S: Give `squash' an explicit destination argument.

Also, rearrange the arguments so the destination(s) are at the start.

diff --git a/math/mpx-mul4-x86-sse2.S b/math/mpx-mul4-x86-sse2.S
index 471b3f1f90cbc69cea42cad6ecdecd82285a081b..5d8c5b97d2e72c4a1f1adde83432f232dbd4a401 100644 (file)
--- a/math/mpx-mul4-x86-sse2.S
+++ b/math/mpx-mul4-x86-sse2.S
@@ -214,11 +214,11 @@
   .endif
 .endm
 
   .endif
 .endm
 

-.macro squash  c0, c1, c2, c3, h, t, u
+.macro squash  lo, hi, c0, c1, c2, c3, t, u

        // On entry, C0, C1, C2, C3 are carry registers representing a value
        // On entry, C0, C1, C2, C3 are carry registers representing a value

-       // Y.  On exit, C0 holds the low 128 bits of the carry value; C1, C2,
+       // Y.  On exit, LO holds the low 128 bits of the carry value; C1, C2,

        // C3, T, and U are clobbered; and the high bits of Y are stored in
        // C3, T, and U are clobbered; and the high bits of Y are stored in

-       // H, if this is not `nil'.
+       // HI, if this is not `nil'.

 
        // The first step is to eliminate the `double-prime' pieces -- i.e.,
        // the ones offset by 16 bytes from a 32-bit boundary -- by carrying
 
        // The first step is to eliminate the `double-prime' pieces -- i.e.,
        // the ones offset by 16 bytes from a 32-bit boundary -- by carrying


@@ -241,8 +241,8 @@
        psrlq   \c1, 16                 // high parts of (y''_1, y''_3)
        psrlq   \c2, 32                 // low parts of (y''_0, y''_2)
        psrlq   \c3, 32                 // low parts of (y''_1, y''_3)
        psrlq   \c1, 16                 // high parts of (y''_1, y''_3)
        psrlq   \c2, 32                 // low parts of (y''_0, y''_2)
        psrlq   \c3, 32                 // low parts of (y''_1, y''_3)

-  .ifnes "\h", "nil"
-       movdqa  \h, \c1
+  .ifnes "\hi", "nil"
+       movdqa  \hi, \c1

   .endif
        pslldq  \c1, 8                  // high part of (0, y''_1)
 
   .endif
        pslldq  \c1, 8                  // high part of (0, y''_1)
 


@@ -250,15 +250,15 @@
        paddq   \u, \c3
        paddq   \t, \c1                 // and up: (y_0, y_2)
        paddq   \u, \c0                 // (y_1, y_3)
        paddq   \u, \c3
        paddq   \t, \c1                 // and up: (y_0, y_2)
        paddq   \u, \c0                 // (y_1, y_3)

-  .ifnes "\h", "nil"
-       psrldq  \h, 8                   // high part of (y''_3, 0)
+  .ifnes "\hi", "nil"
+       psrldq  \hi, 8                  // high part of (y''_3, 0)

   .endif
 
        // Finally extract the answer.  This complicated dance is better than
        // storing to memory and loading, because the piecemeal stores
        // inhibit store forwarding.
        movdqa  \c3, \t                 // (y_0, y_1)
   .endif
 
        // Finally extract the answer.  This complicated dance is better than
        // storing to memory and loading, because the piecemeal stores
        // inhibit store forwarding.
        movdqa  \c3, \t                 // (y_0, y_1)

-       movdqa  \c0, \t                 // (y^*_0, ?, ?, ?)
+       movdqa  \lo, \t                 // (y^*_0, ?, ?, ?)

        psrldq  \t, 8                   // (y_2, 0)
        psrlq   \c3, 32                 // (floor(y_0/B), ?)
        paddq   \c3, \u                 // (y_1 + floor(y_0/B), ?)
        psrldq  \t, 8                   // (y_2, 0)
        psrlq   \c3, 32                 // (floor(y_0/B), ?)
        paddq   \c3, \u                 // (y_1 + floor(y_0/B), ?)


@@ -266,20 +266,20 @@
        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, ?)
        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, ?)

-       punpckldq \c0, \c3              // (y^*_0, y^*_2, ?, ?)
+       punpckldq \lo, \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, ?)
        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, ?)

-  .ifnes "\h", "nil"
+  .ifnes "\hi", "nil"

        movdqa  \t, \c3
        pxor    \u, \u
   .endif
        punpckldq \c1, \c3              // (y^*_1, y^*_3, ?, ?)
        movdqa  \t, \c3
        pxor    \u, \u
   .endif
        punpckldq \c1, \c3              // (y^*_1, y^*_3, ?, ?)

-  .ifnes "\h", "nil"
+  .ifnes "\hi", "nil"

        psrlq   \t, 32                  // very high bits of y
        psrlq   \t, 32                  // very high bits of y

-       paddq   \h, \t
-       punpcklqdq \h, \u               // carry up
+       paddq   \hi, \t
+       punpcklqdq \hi, \u              // carry up

   .endif
   .endif

-       punpckldq \c0, \c1              // y mod B^4
+       punpckldq \lo, \c1              // y mod B^4

 .endm
 
 .macro carryadd
 .endm
 
 .macro carryadd


@@ -576,7 +576,7 @@ INTFUNC(mmla4)
        accum                    xmm7, nil,  nil,  nil
 
        // That's lots of pieces.  Now we have to assemble the answer.
        accum                    xmm7, nil,  nil,  nil
 
        // That's lots of pieces.  Now we have to assemble the answer.

-       squash  xmm4, xmm5, xmm6, xmm7, nil, xmm0, xmm1
+       squash  xmm4, nil,  xmm4, xmm5, xmm6, xmm7,  xmm0, xmm1

 
        // Expand it.
        pxor    xmm2, xmm2
 
        // Expand it.
        pxor    xmm2, xmm2


@@ -639,7 +639,7 @@ INTFUNC(mont4)
        accum                    xmm7, nil,  nil,  nil
 
        // That's lots of pieces.  Now we have to assemble the answer.
        accum                    xmm7, nil,  nil,  nil
 
        // That's lots of pieces.  Now we have to assemble the answer.

-       squash  xmm4, xmm5, xmm6, xmm7, nil, xmm0, xmm1
+       squash  xmm4, nil,  xmm4, xmm5, xmm6, xmm7,  xmm0, xmm1

 
        // Expand it.
        pxor    xmm2, xmm2
 
        // Expand it.
        pxor    xmm2, xmm2