[Bug target/41084] New: Filling xmm register with all bit set is not optimized

["etjq78kl at free dot fr" <gcc-bugzilla at gcc dot gnu dot org>]

Hi,

A way to fill an xmm register with all one, is to use _mm_cmpeq_epi{8,16,32}
wiht itself.

However, if you write:

  __m128i r; r = _mm_cmpeq_epi32 (r, r);

gcc absolutely wants to clear the register before and generates (this is output
of objdump -d, compiled with -O3 -march=core2):

  401484:       66 0f ef c0             pxor   %xmm0,%xmm0
  401488:       66 0f 74 c0             pcmpeqw %xmm0,%xmm0

It does not discover that the result is independant of the initial value of r,
and wants to clear it before.

Similarly, if one writes (code adapted from _mm_setzero_si128 (void) in
emmintrin.h):

  __m128i r = __extension__ (__m128i)(__v4si){ 0xffffffff, 0xffffffff,
0xffffffff, 0xffffffff }; 

then this will generate a memory load operation, instead of the optimized
pcmpeqw instruction.

I would expect both
  __m128i r; r = _mm_cmpeq_epi32 (r, r);
and
  __m128i r = __extension__ (__m128i)(__v4si){ 0xffffffff, 0xffffffff, 
0xffffffff, 0xffffffff }; 

generate the same unique instruction: pcmpeqw %xmm0, %xmm0

exactly as:
  __m128i r; r = _mm_xor_si128 (r, r);
and
  __m128i r = __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 }; 

outputs pxor %xmm0, %xmm0 in both cases.

Best regards.

Antoine


-- 
           Summary: Filling xmm register with all bit set is not optimized
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: target
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: etjq78kl at free dot fr
GCC target triplet: i?86


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41084