aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorJack Lloyd <[email protected]>2016-09-02 04:47:12 -0400
committerJack Lloyd <[email protected]>2016-09-02 04:47:12 -0400
commitac3d1ea17c0dfc0c279a6715a110c738ee615b3f (patch)
treeef4affadaf9df47232d502987e096fc64a526f99
parent3fc924b97934eda75fffff0c561b3d180ae290a6 (diff)
Avoid _mm_set_epi64x which is missing on 32-bit MSVC 12
-rw-r--r--src/lib/stream/chacha/chacha_sse2/chacha_sse2.cpp16
1 files changed, 8 insertions, 8 deletions
diff --git a/src/lib/stream/chacha/chacha_sse2/chacha_sse2.cpp b/src/lib/stream/chacha/chacha_sse2/chacha_sse2.cpp
index c881a2990..e39b285b3 100644
--- a/src/lib/stream/chacha/chacha_sse2/chacha_sse2.cpp
+++ b/src/lib/stream/chacha/chacha_sse2/chacha_sse2.cpp
@@ -37,19 +37,19 @@ void ChaCha::chacha_sse2_x4(byte output[64], u32bit input[16], size_t rounds)
__m128i r1_1 = input1;
__m128i r1_2 = input2;
__m128i r1_3 = input3;
- r1_3 = _mm_add_epi64(r0_3, _mm_set_epi64x(0, 1));
+ r1_3 = _mm_add_epi64(r0_3, _mm_set_epi32(0, 0, 0, 1));
__m128i r2_0 = input0;
__m128i r2_1 = input1;
__m128i r2_2 = input2;
__m128i r2_3 = input3;
- r2_3 = _mm_add_epi64(r0_3, _mm_set_epi64x(0, 2));
+ r2_3 = _mm_add_epi64(r0_3, _mm_set_epi32(0, 0, 0, 2));
__m128i r3_0 = input0;
__m128i r3_1 = input1;
__m128i r3_2 = input2;
__m128i r3_3 = input3;
- r3_3 = _mm_add_epi64(r0_3, _mm_set_epi64x(0, 3));
+ r3_3 = _mm_add_epi64(r0_3, _mm_set_epi32(0, 0, 0, 3));
for(size_t r = 0; r != rounds / 2; ++r)
{
@@ -215,26 +215,26 @@ void ChaCha::chacha_sse2_x4(byte output[64], u32bit input[16], size_t rounds)
r1_1 = _mm_add_epi32(r1_1, input1);
r1_2 = _mm_add_epi32(r1_2, input2);
r1_3 = _mm_add_epi32(r1_3, input3);
- r1_3 = _mm_add_epi64(r1_3, _mm_set_epi64x(0, 1));
+ r1_3 = _mm_add_epi64(r1_3, _mm_set_epi32(0, 0, 0, 1));
r2_0 = _mm_add_epi32(r2_0, input0);
r2_1 = _mm_add_epi32(r2_1, input1);
r2_2 = _mm_add_epi32(r2_2, input2);
r2_3 = _mm_add_epi32(r2_3, input3);
- r2_3 = _mm_add_epi64(r2_3, _mm_set_epi64x(0, 2));
+ r2_3 = _mm_add_epi64(r2_3, _mm_set_epi32(0, 0, 0, 2));
r3_0 = _mm_add_epi32(r3_0, input0);
r3_1 = _mm_add_epi32(r3_1, input1);
r3_2 = _mm_add_epi32(r3_2, input2);
r3_3 = _mm_add_epi32(r3_3, input3);
- r3_3 = _mm_add_epi64(r3_3, _mm_set_epi64x(0, 3));
+ r3_3 = _mm_add_epi64(r3_3, _mm_set_epi32(0, 0, 0, 3));
_mm_storeu_si128(output_mm + 0, r0_0);
_mm_storeu_si128(output_mm + 1, r0_1);
_mm_storeu_si128(output_mm + 2, r0_2);
_mm_storeu_si128(output_mm + 3, r0_3);
- _mm_storeu_si128(output_mm + 4 , r1_0);
+ _mm_storeu_si128(output_mm + 4, r1_0);
_mm_storeu_si128(output_mm + 5, r1_1);
_mm_storeu_si128(output_mm + 6, r1_2);
_mm_storeu_si128(output_mm + 7, r1_3);
@@ -244,7 +244,7 @@ void ChaCha::chacha_sse2_x4(byte output[64], u32bit input[16], size_t rounds)
_mm_storeu_si128(output_mm + 10, r2_2);
_mm_storeu_si128(output_mm + 11, r2_3);
- _mm_storeu_si128(output_mm + 12 , r3_0);
+ _mm_storeu_si128(output_mm + 12, r3_0);
_mm_storeu_si128(output_mm + 13, r3_1);
_mm_storeu_si128(output_mm + 14, r3_2);
_mm_storeu_si128(output_mm + 15, r3_3);