[Bug target/55461] New: _mm_loadu_si128 generates wrong instruction on 4.8

[andi-gcc at firstfloor dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55461



             Bug #: 55461

           Summary: _mm_loadu_si128 generates wrong instruction on 4.8

    Classification: Unclassified

           Product: gcc

           Version: 4.8.0

            Status: UNCONFIRMED

          Severity: normal

          Priority: P3

         Component: target

        AssignedTo: unassig...@gcc.gnu.org

        ReportedBy: andi-...@firstfloor.org





Following test case compiled with -Os -msse4.2 generates with 4.8:



.L3:

        movups  (%rdi), %xmm0         <--------------- wrong data type, should

be movupd

        pcmpistrm       $24, (%rdi,%rsi), %xmm0

        jna     .L2

        addq    $16, %rdi

        jmp     .L3



4.7 gives the correct (although it makes a mess out of the loop):



L3:

        movdqu  (%rdi,%rsi), %xmm1

        movdqu  (%rdi), %xmm2

        pcmpistrm       $24, %xmm1, %xmm2

        jna     .L2

        addq    $16, %rdi

        jmp     .L3





A simpler test case gives the correct movupd, so it must be related to the

funky pointer arithmetic the test case does. But in any case it should not turn

an integer vector into a float vector



#include <nmmintrin.h>



int c_strcmp(char *a, char *b)

{

        unsigned long diff = (unsigned long)b - (unsigned long)a;

        int r = 16;

        a -= r;

        for (;;) {



                if (_mm_cmpistra(_mm_loadu_si128((__m128i *)a),

                                 _mm_loadu_si128((__m128i *)((unsigned long)a +

diff)),

                                 0x18)) {

                        a += r;

                        continue;

                }

                /* check C here */

                return 0;

        }

        return 0;



}