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

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

diff --git a/math/mpx-mul4-amd64-sse2.S b/math/mpx-mul4-amd64-sse2.S
index 1c205a7..1c344f4 100644 (file)
--- a/math/mpx-mul4-amd64-sse2.S
+++ b/math/mpx-mul4-amd64-sse2.S
@@ -1601,6 +1601,8 @@ FUNC(test_mmul4)
        testtop r11
        call    mmul4
        testtail
+       pshufd  xmm10, xmm10, SHUF(0, 2, 1, 3)
+       pshufd  xmm11, xmm11, SHUF(0, 2, 1, 3)
        movdqu  [r10 +  0], xmm10
        movdqu  [r10 + 16], xmm11
        testcarryout
@@ -1612,6 +1614,8 @@ FUNC(test_mmla4)
        testtop r11
        call    mmla4
        testtail
+       pshufd  xmm10, xmm10, SHUF(0, 2, 1, 3)
+       pshufd  xmm11, xmm11, SHUF(0, 2, 1, 3)
        movdqu  [r10 +  0], xmm10
        movdqu  [r10 + 16], xmm11
        testcarryout
@@ -1623,6 +1627,8 @@ FUNC(test_mont4)
        testtop
        call    mont4
        testtail
+       pshufd  xmm10, xmm10, SHUF(0, 2, 1, 3)
+       pshufd  xmm11, xmm11, SHUF(0, 2, 1, 3)
        movdqu  [r10 +  0], xmm10
        movdqu  [r10 + 16], xmm11
        testcarryout