[PATCH] Patched clang to emit x86 blends as shufflevectors.

[Rafael Ávila de Espíndola]

Filipe points out that we cannot use a #define for the __builtin since it has to be available even when no .h is include. Optimizing the builtin would then require custom code in clang/llvm, which is probably not worth it.

================
Comment at: lib/Headers/avx2intrin.h:163
@@ -162,2 +162,3 @@
   __m256i __V2 = (V2); \
-  (__m256i)__builtin_ia32_pblendw256((__v16hi)__V1, (__v16hi)__V2, (M)); })
+  (__m256d)__builtin_shufflevector((__v16hi)__V1, (__v16hi)__V2, \
+                                   (((M) & 0x01) ? 16 : 0), \
----------------
Why the change to __m256d? The intel manual says the signature is

__m256i _mm256_blend_epi16 (__m256i v1, __m256i v2, const int mask)


================
Comment at: lib/Headers/avx2intrin.h:168
@@ +167,3 @@
+                                   (((M) & 0x08) ? 19 : 3), \
+                                   (((M) & 0x04) ? 20 : 4), \
+                                   (((M) & 0x04) ? 21 : 5), \
----------------
The masks looks wrong for the hight bits. Shouldn't this be

 (((M) & 0x01) ? 16 : 0), \
 (((M) & 0x02) ? 17 : 1), \
 (((M) & 0x04) ? 18 : 2), \
 (((M) & 0x08) ? 19 : 3), \
 (((M) & 0x10) ? 20 : 4), \
 (((M) & 0x20) ? 21 : 5), \
 (((M) & 0x40) ? 22 : 6), \
 (((M) & 0x80) ? 23 : 7), \
 (((M) & 0x01) ? 24 : 8), \
 (((M) & 0x02) ? 25 : 9), \
 (((M) & 0x04) ? 26 : 10), \
 (((M) & 0x08) ? 27 : 11), \
 (((M) & 0x10) ? 28 : 12), \
 (((M) & 0x20) ? 29 : 13), \
 (((M) & 0x40) ? 30 : 14), \
 (((M) & 0x80) ? 31 : 15), \

http://reviews.llvm.org/D3601