Remove duplicate type definitions from f16cintrin.h

[jtsoftware]

Summary

These definitions can be obtained from avxintrin.h. Modularize flags them as duplicates.

[ygao]

LGTM. This is compatible with GCC headers.
By the way, this review did not appear on the mailing list for some reason,

[silvas]

Please include the header where those types are actually defined. We're trying to be modular here.

[probinson]

In D9557#170720, @silvas wrote:

Please include the header where those types are actually defined. We're trying to be modular here.

I suspect the correct fix there is to #include <f16cintrin.h> from immintrin.h instead of x86intrin.h; that's what gcc does.

[silvas]

Looks like there's more duplication to get rid of in this patch.

lib/Headers/avxintrin.h:typedef float __v8sf __attribute__ ((__vector_size__ (32)));
lib/Headers/f16cintrin.h:typedef float __v8sf __attribute__ ((__vector_size__ (32)));

lib/Headers/avxintrin.h:typedef float __m256 __attribute__ ((__vector_size__ (32)));
lib/Headers/f16cintrin.h:typedef float __m256 __attribute__ ((__vector_size__ (32)));

What is the right dependency order between these headers?

[silvas]

Gah, nevermind. Brain fart.

[jtsoftware]

If we want individual headers to be independent, we probably need a new header called something like. _intrinsic_types.h, as there are a few definitions like m256, m256d, and __m256i scattered among multiple headers. But actually, it seems to be the case that headers like f16cintrin.h are private headers under x86intrin.h, as indicated in the #error directive (#error "Never use <f16cintrin.h> directly; include <x86intrin.h> instead."), hence the need to be independent could be considered relaxed. If gnu is like this, it might be a good reason not to stray too far.

What do you think?

[jtsoftware]

It looks like this has already been addressed.

lib/Headers/f16cintrin.h

 
 #ifndef __F16C__
 # error "F16C instruction is not enabled"
 #endif /* __F16C__ */
 
 #ifndef __F16CINTRIN_H
 #define __F16CINTRIN_H
 
-typedef float __v8sf __attribute__ ((__vector_size__ (32)));
-typedef float __m256 __attribute__ ((__vector_size__ (32)));
-
 #define _mm_cvtps_ph(a, imm) __extension__ ({ \
   __m128 __a = (a); \
  (__m128i)__builtin_ia32_vcvtps2ph((__v4sf)__a, (imm)); })
 
 #define _mm256_cvtps_ph(a, imm) __extension__ ({ \
   __m256 __a = (a); \
  (__m128i)__builtin_ia32_vcvtps2ph256((__v8sf)__a, (imm)); })