[llvm-bugs] [Bug 42023] New: [X86] Failure to use HADD for reduction add patterns

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=42023

            Bug ID: 42023
           Summary: [X86] Failure to use HADD for reduction add patterns
           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: craig.topper at gmail.com, llvm-bugs at lists.llvm.org,
                    llvm-dev at redking.me.uk, spatel+llvm at rotateright.com

https://godbolt.org/z/g1wL8n

define i32 @hadd32_4(<4 x i32> %x225) optsize {
  %x226 = shufflevector <4 x i32> %x225, <4 x i32> undef, <4 x i32> <i32 2, i32
3, i32 undef, i32 undef>
  %x227 = add <4 x i32> %x225, %x226
  %x228 = shufflevector <4 x i32> %x227, <4 x i32> undef, <4 x i32> <i32 1, i32
undef, i32 undef, i32 undef>
  %x229 = add <4 x i32> %x227, %x228
  %x230 = extractelement <4 x i32> %x229, i32 0
  ret i32 %x230
}

define i16 @hadd16_8(<8 x i16> %x223) optsize {
  %x224 = shufflevector <8 x i16> %x223, <8 x i16> undef, <8 x i32> <i32 4, i32
5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
  %x225 = add <8 x i16> %x223, %x224
  %x226 = shufflevector <8 x i16> %x225, <8 x i16> undef, <8 x i32> <i32 2, i32
3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %x227 = add <8 x i16> %x225, %x226
  %x228 = shufflevector <8 x i16> %x227, <8 x i16> undef, <8 x i32> <i32 1, i32
undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
  %x229 = add <8 x i16> %x227, %x228
  %x230 = extractelement <8 x i16> %x229, i32 0
  ret i16 %x230
}

llc -mcpu=btver2

hadd32_4:
  vpshufd $78, %xmm0, %xmm1 # xmm1 = xmm0[2,3,0,1]
  vpaddd %xmm1, %xmm0, %xmm0
  vphaddd %xmm0, %xmm0, %xmm0
  vmovd %xmm0, %eax
  retq

hadd16_8: # @hadd16_8
  vpshufd $78, %xmm0, %xmm1 # xmm1 = xmm0[2,3,0,1]
  vpaddw %xmm1, %xmm0, %xmm0
  vpshufd $229, %xmm0, %xmm1 # xmm1 = xmm0[1,1,2,3]
  vpaddw %xmm1, %xmm0, %xmm0
  vphaddw %xmm0, %xmm0, %xmm0
  vmovd %xmm0, %eax
  retq

We should always be able to fold integer horizontal reduction patterns like
this, and fadd patterns if we are safe to reassociate.

Pulled out of [Bug #33758] where we are still missing some (optsize) HADD cases
(see also https://godbolt.org/z/yP5CgS ).

-- 
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/20190526/fc9b768c/attachment.html>