[llvm-bugs] [Bug 44509] New: Redundant vxorps instruction

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=44509

            Bug ID: 44509
           Summary: Redundant vxorps instruction
           Product: clang
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: LLVM Codegen
          Assignee: unassignedclangbugs at nondot.org
          Reporter: zamazan4ik at tut.by
                CC: llvm-bugs at lists.llvm.org, neeilans at live.com,
                    richard-llvm at metafoo.co.uk

With fast-math on, clang generates less optimal (and less precise) code.

Consider following code:

__m128 foo(__m128 rounded_cam_pos, __m128 remainder_cam_pos, __m128 pos)
{
  __m128 ofs = _mm_sub_ps(pos, rounded_cam_pos);
  return _mm_sub_ps(ofs, remainder_cam_pos);
}

compiled with -O3 -ffast-math -march=native
Clang (trunk) will generate more instructions

        vxorps  xmm0, xmm0, dword ptr [rip + .LCPI0_0]{1to4}
        vsubps  xmm0, xmm0, xmm1
        vaddps  xmm0, xmm0, xmm2

while also uselessly re-arrange the code.
gcc with fast-math, or clang without -ffast-math, generates simple and optimal

        vsubps  xmm0, xmm2, xmm0
        vsubps  xmm0, xmm0, xmm1

https://godbolt.org/z/y8rmap

-- 
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/20200110/0a802e91/attachment.html>