diff mbox

[x86,4/n] Replace builtins with vector extensions

Message ID alpine.DEB.2.02.1411081156160.13616@stedding.saclay.inria.fr
State New
Headers show

Commit Message

Marc Glisse Nov. 8, 2014, 12:49 p.m. UTC 
Hello,

this patch uses &|^ for 128 bit integer vectors. I am doing the operations 
in type __v2du because __builtin_ia32_pand128 was apparently taking __v2di 
arguments, but using __v4su or any other should be equivalent. Even 
__int128 would in principle be ok, but since it is not usually stored in a 
vector register, it seems more likely to generate unexpected code (and we 
don't have __int256 so it would be inconsistent with other sizes).

Regtested with patch 3/n. Ok for the branch?

After that, I will post a last patch to generalize &|^ to sizes 256 and 
512, and I think that will be enough for gcc-5, we should discuss merging. 
"< > == abs min max" can wait until gcc-6, possibly after getting some 
feedback about +-*/&|^.



2014-11-10  Marc Glisse  <marc.glisse@inria.fr>

 	* config/i386/emmintrin.h (_mm_and_si128, _mm_or_si128, _mm_xor_si128):
 	Use vector extensions instead of builtins.

Comments

Uros Bizjak Nov. 9, 2014, 8:36 a.m. UTC | #1 
On Sat, Nov 8, 2014 at 1:49 PM, Marc Glisse <marc.glisse@inria.fr> wrote:
> Hello,
>
> this patch uses &|^ for 128 bit integer vectors. I am doing the operations
> in type __v2du because __builtin_ia32_pand128 was apparently taking __v2di
> arguments, but using __v4su or any other should be equivalent. Even __int128
> would in principle be ok, but since it is not usually stored in a vector
> register, it seems more likely to generate unexpected code (and we don't
> have __int256 so it would be inconsistent with other sizes).

Let's keep v2du for now.

> Regtested with patch 3/n. Ok for the branch?
>
> After that, I will post a last patch to generalize &|^ to sizes 256 and 512,
> and I think that will be enough for gcc-5, we should discuss merging. "< >
> == abs min max" can wait until gcc-6, possibly after getting some feedback
> about +-*/&|^.
>
>
>
> 2014-11-10  Marc Glisse  <marc.glisse@inria.fr>
>
>         * config/i386/emmintrin.h (_mm_and_si128, _mm_or_si128,
> _mm_xor_si128):
>
>         Use vector extensions instead of builtins.

OK.

Thanks,
Uros.
diff mbox

Patch

Index: config/i386/emmintrin.h
===================================================================
--- config/i386/emmintrin.h	(revision 217249)
+++ config/i386/emmintrin.h	(working copy)
@@ -1244,39 +1244,39 @@  _mm_srl_epi32 (__m128i __A, __m128i __B)
 
 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_srl_epi64 (__m128i __A, __m128i __B)
 {
   return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);
 }
 
 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_and_si128 (__m128i __A, __m128i __B)
 {
-  return (__m128i)__builtin_ia32_pand128 ((__v2di)__A, (__v2di)__B);
+  return (__m128i) ((__v2du)__A & (__v2du)__B);
 }
 
 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_andnot_si128 (__m128i __A, __m128i __B)
 {
   return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);
 }
 
 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_or_si128 (__m128i __A, __m128i __B)
 {
-  return (__m128i)__builtin_ia32_por128 ((__v2di)__A, (__v2di)__B);
+  return (__m128i) ((__v2du)__A | (__v2du)__B);
 }
 
 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_xor_si128 (__m128i __A, __m128i __B)
 {
-  return (__m128i)__builtin_ia32_pxor128 ((__v2di)__A, (__v2di)__B);
+  return (__m128i) ((__v2du)__A ^ (__v2du)__B);
 }
 
 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_cmpeq_epi8 (__m128i __A, __m128i __B)
 {
   return (__m128i)__builtin_ia32_pcmpeqb128 ((__v16qi)__A, (__v16qi)__B);
 }
 
 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_cmpeq_epi16 (__m128i __A, __m128i __B)