[Bug c++/82793] New: __attribute__((target("sse"))) causes call throught ifunc

[vorfeed.canal at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82793

            Bug ID: 82793
           Summary: __attribute__((target("sse"))) causes call throught
                    ifunc
           Product: gcc
           Version: 7.2.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: vorfeed.canal at gmail dot com
  Target Milestone: ---

The following example illustrates the problem:

#include "xmmintrin.h"

extern __attribute__((target("avx"),visibility("hidden"))) __m128 foo(__m128 a,
__m128 b);
extern __attribute__((target("sse4.2"),visibility("hidden"))) __m128 foo(__m128
a, __m128 b);
extern __attribute__((target("default"),visibility("hidden"))) __m128
foo(__m128 a, __m128 b);

__attribute__((target("sse4.2"))) __m128 bar(__m128 a, __m128 b) {
  return foo(a, b);
}

__attribute__((target("avx"))) __m128 bar(__m128 a, __m128 b) {
  return foo(a, b);
}

__attribute__((target("default"))) __m128 bar(__m128 a, __m128 b) {
  return foo(a, b);
}

All versions of GCC which I've tried produced the following code:

...
       .type   _Z3barDv4_fS_.avx, @function
_Z3barDv4_fS_.avx:
.LFB526:
        .cfi_startproc
        jmp     _Z3fooDv4_fS_.avx
        .cfi_endproc
...
        .type   _Z3barDv4_fS_.sse4.2, @function
_Z3barDv4_fS_.sse4.2:
.LFB525:
        .cfi_startproc
        jmp     _Z19_Z3fooDv4_fS_.ifuncDv4_fS_
        .cfi_endproc


That is: AVX functions are calling AVX functions directly while SSE4.2
functions call SSE4.2 functions via ifunc-resolver. Needless to day this kills
the performance quite throughly - and this IS the reason to use function
multiversioning with AVS/SSE attributes!