[LLVMbugs] [Bug 23349] New: [SSE] scalar intrinsics don't fold loads

[bugzilla-daemon at llvm.org]

https://llvm.org/bugs/show_bug.cgi?id=23349

            Bug ID: 23349
           Summary: [SSE] scalar intrinsics don't fold loads
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: spatel+llvm at rotateright.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

#include <xmmintrin.h>
__m128 add_fold(__m128 x, float *y) {
        __m128 ld = _mm_load_ss(y);
        return _mm_add_ss(x, ld);
}

or IR test case for llc:

define <4 x float> @add_fold(<4 x float> %x, float* %y) #0 {
  %0 = load float, float* %y, align 1, !tbaa !2
  %vecext1.i = extractelement <4 x float> %x, i32 0
  %add.i = fadd float %vecext1.i, %0
  %vecins.i = insertelement <4 x float> %x, float %add.i, i32 0
  ret <4 x float> %vecins.i
}


----------------------------------------------------------------------------

The load should be folded into the math op, but:

$ clang -O1 addfold.c -S -o -
...
    movss    (%rdi), %xmm1           ## xmm1 = mem[0],zero,zero,zero
    addss    %xmm1, %xmm0

----------------------------------------------------------------------------

The load does get folded with non-intrinsic C (IR won't have extract/insert in
this case) or with vector ops:

float add_fold2(float x, float *y) {
        return x + *y;
}

__m128 add_fold3(__m128 x, float *y) {
        __m128 ld = _mm_load_ps(y);
        return _mm_add_ps(x, ld);
}

...

    addss    (%rdi), %xmm0
...
    addps    (%rdi), %xmm0

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20150425/4d1483b0/attachment.html>