math/mpx-mul4-*.S: Output expanded Montgomery factor in a sensible order.

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

diff --git a/math/mpx-mul4-x86-sse2.S b/math/mpx-mul4-x86-sse2.S
index 7e7173f..cdbdfaf 100644 (file)
--- a/math/mpx-mul4-x86-sse2.S
+++ b/math/mpx-mul4-x86-sse2.S
@@ -1286,6 +1286,8 @@ FUNC(test_mmul4)
        mov     edi, [BP + 28]
        movdqa  xmm0, [SP + 64]
        movdqa  xmm1, [SP + 80]
+       pshufd  xmm0, xmm0, SHUF(0, 2, 1, 3)
+       pshufd  xmm1, xmm1, SHUF(0, 2, 1, 3)
        movdqu  [edi], xmm0
        movdqu  [edi + 16], xmm1
        testcarryout [BP + 24]
@@ -1302,6 +1304,8 @@ FUNC(test_mmla4)
        mov     edi, [BP + 28]
        movdqa  xmm0, [SP + 64]
        movdqa  xmm1, [SP + 80]
+       pshufd  xmm0, xmm0, SHUF(0, 2, 1, 3)
+       pshufd  xmm1, xmm1, SHUF(0, 2, 1, 3)
        movdqu  [edi], xmm0
        movdqu  [edi + 16], xmm1
        testcarryout [BP + 24]
@@ -1318,6 +1322,8 @@ FUNC(test_mont4)
        mov     edi, [BP + 28]
        movdqa  xmm0, [SP + 64]
        movdqa  xmm1, [SP + 80]
+       pshufd  xmm0, xmm0, SHUF(0, 2, 1, 3)
+       pshufd  xmm1, xmm1, SHUF(0, 2, 1, 3)
        movdqu  [edi], xmm0
        movdqu  [edi + 16], xmm1
        testcarryout [BP + 24]