[llvm-bugs] [Bug 51091] New: Failure to widen loads from wider dereferenceable pointers

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=51091

            Bug ID: 51091
           Summary: Failure to widen loads from wider dereferenceable
                    pointers
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: llvm-dev at redking.me.uk
                CC: a.bataev at hotmail.com, craig.topper at gmail.com,
                    lebedev.ri at gmail.com, llvm-bugs at lists.llvm.org,
                    llvm-dev at redking.me.uk, pengfei.wang at intel.com,
                    spatel+llvm at rotateright.com

If we have generated this:

define float @Dot01(float* dereferenceable(16) %a0, float* dereferenceable(16)
%a1) {
  %bcx01 = bitcast float* %a0 to <2 x float>*
  %bcy01 = bitcast float* %a1 to <2 x float>*
  %x01 = load <2 x float>, <2 x float>* %bcx01, align 4
  %y01 = load <2 x float>, <2 x float>* %bcy01, align 4
  %mul01 = fmul <2 x float> %x01, %y01
  %mul0 = extractelement <2 x float> %mul01, i32 0
  %mul1 = extractelement <2 x float> %mul01, i32 1
  %dot01 = fadd float %mul0, %mul1
  ret float %dot01
}

llc -mcpu=znver2 

Dot01:
        vmovsd  (%rdi), %xmm0                   # xmm0 = mem[0],zero
        vmovsd  (%rsi), %xmm1                   # xmm1 = mem[0],zero
        vmulps  %xmm1, %xmm0, %xmm0
        vmovshdup       %xmm0, %xmm1            # xmm1 = xmm0[1,1,3,3]
        vaddss  %xmm1, %xmm0, %xmm0
        retq

As we know that the pointers at (%rdi) and (%rsi) are both dereferenceable
across the full 16 bytes - why can't we load full float4 loads, which would let
us fold at least one of the loads:

Dot01:
        vmovups   (%rdi), %xmm0
        vmulps    (%rsi), %xmm0, %xmm0
        vmovshdup %xmm0, %xmm1            # xmm1 = xmm0[1,1,3,3]
        vaddss    %xmm1, %xmm0, %xmm0
        retq

(I'm not sure if we can perform this generally in DAG or VectorCombine or
whether we should just handle it inside x86's EltsFromConsecutiveLoads or
something).

-- 
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/20210714/fa6399e4/attachment.html>