(x86 asm): Zero the high parts of the ?MM registers if available.

[catacomb] / math / mpx-mul4-amd64-sse2.S

diff --git a/math/mpx-mul4-amd64-sse2.S b/math/mpx-mul4-amd64-sse2.S
index 2d78a99..d8f54e1 100644 (file)
--- a/math/mpx-mul4-amd64-sse2.S
+++ b/math/mpx-mul4-amd64-sse2.S
@@ -752,6 +752,13 @@ ENDFUNC
 ///--------------------------------------------------------------------------
 /// Bulk multipliers.
 
+FUNC(mpx_umul4_amd64_avx)
+       .arch   .avx
+       vzeroupper
+  endprologue
+       .arch   pentium4
+ENDFUNC
+
 FUNC(mpx_umul4_amd64_sse2)
        // void mpx_umul4_amd64_sse2(mpw *dv, const mpw *av, const mpw *avl,
        //                         const mpw *bv, const mpw *bvl);
@@ -901,6 +908,13 @@ FUNC(mpx_umul4_amd64_sse2)
 
 ENDFUNC
 
+FUNC(mpxmont_mul4_amd64_avx)
+       .arch   .avx
+       vzeroupper
+  endprologue
+       .arch   pentium4
+ENDFUNC
+
 FUNC(mpxmont_mul4_amd64_sse2)
        // void mpxmont_mul4_amd64_sse2(mpw *dv, const mpw *av, const mpw *bv,
        //                           const mpw *nv, size_t n, const mpw *mi);
@@ -1095,6 +1109,13 @@ FUNC(mpxmont_mul4_amd64_sse2)
 
 ENDFUNC
 
+FUNC(mpxmont_redc4_amd64_avx)
+       .arch   .avx
+       vzeroupper
+  endprologue
+       .arch   pentium4
+ENDFUNC
+
 FUNC(mpxmont_redc4_amd64_sse2)
        // void mpxmont_redc4_amd64_sse2(mpw *dv, mpw *dvl, const mpw *nv,
        //                             size_t n, const mpw *mi);