[Bug middle-end/43514] New: [missed optimization] use of SSE shift intrinsics introduces unnecessary moves to the stack and back

[kretz at kde dot org]

Testcase:

#include <tmmintrin.h>

__m128i intrin()
{
    __m128i a = _mm_setzero_si128();
    a = _mm_cmpeq_epi8(a, a);
    __m128i count = _mm_sad_epu8(_mm_abs_epi8(a), _mm_setzero_si128());
    return _mm_sll_epi64(a, count);
}

__m128i assem()
{
    register __m128i r asm("xmm0");
    asm(
            "pxor %%xmm1,%%xmm1\n\t"       // 0x0000
            "pcmpeqb %%xmm0,%%xmm0\n\t"    // 0xffff                            
            "pabsb %%xmm0,%%xmm2\n\t"      // 0x0101
            "psadbw %%xmm1,%%xmm2\n\t"     // 0x0008                            
            "psllq %%xmm2,%%xmm0\n\t"                                           
            :::"xmm0", "xmm1", "xmm2");
    return r;                                                                   
}                                                                               

int main()                                                                      
{                                                                               
    const __m128i a = intrin();                                                 
    const __m128i b = assem();
    return (_mm_movemask_epi8(_mm_cmpeq_epi32(a, b)) == 0xffff) ? 0 : -1;
}

compiled with "g++ -march=core2 -O3"

GCC 4.4 and 4.5 (4.5.0 20100311) compile the intrin() function to include an
unnecessary movdqa -> mov -> movd:

psadbw %xmm1,%xmm2
movdqa %xmm2,-0x18(%rsp)
mov    -0x18(%rsp),%rax
movd   %eax,%xmm1
psllq  %xmm1,%xmm0

Which is not present in my inline asm version. Still both functions return the
same result, which is expected: The documentation for the psllq instruction
says that only the lower 64 bits of the shift argument are used. Thus the
zeroing of the upper 64 bits from the xmm2 register to the xmm1 register are
just wasted cycles.


-- 
           Summary: [missed optimization] use of SSE shift intrinsics
                    introduces unnecessary moves to the stack and back
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Severity: minor
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: kretz at kde dot org
 GCC build triplet: x86_64-unknown-linux-gnu
  GCC host triplet: x86_64-unknown-linux-gnu
GCC target triplet: x86_64-unknown-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43514