[llvm-bugs] [Bug 39132] New: [amdgpu] rL343329: [InstCombine] don't propagate wider shufflevector arguments to predecessors cause regression in storing char3 vectors

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=39132

            Bug ID: 39132
           Summary: [amdgpu]  rL343329: [InstCombine] don't propagate
                    wider shufflevector arguments to predecessors cause
                    regression in storing char3 vectors
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: AMDGPU
          Assignee: unassignedbugs at nondot.org
          Reporter: jan.vesely at rutgers.edu
                CC: llvm-bugs at lists.llvm.org

Created attachment 20942
  --> https://bugs.llvm.org/attachment.cgi?id=20942&action=edit
piglit test that fails

this change introduces failures when loading/storing char3 vectors in
OpenCL (amdgcn backend):
llc: /home/orome/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp:4159:
llvm::SDValue llvm::SelectionDAG::getNode(unsigned int, const
llvm::SDLoc&, llvm::EVT, llvm::SDValue, llvm::SDNodeFlags): Assertion
`VT.getSizeInBits() == Operand.getValueSizeInBits() && "Cannot BITCAST
between types of different sizes!"' failed.
Stack dump:
0.      Program arguments: /home/orome/.local/bin/llc -march=amdgcn 
1.      Running pass 'CallGraph Pass Manager' on module '<stdin>'.
2.      Running pass 'AMDGPU DAG->DAG Pattern Instruction Selection'
on function '@vload3_constant'

the patch changes how char3 vector is constructed. instead of starting
with <4xi8> (edited):

  %3 = insertelement <4 x i8> undef, i8 %2, i32 0
  %5 = insertelement <4 x i8> %3, i8 %4, i32 1
  %7 = insertelement <4 x i8> %5, i8 %6, i32 2
  store <4 x i8> %7, <4 x i8> addrspace(1)* %storetmp, align 4, !tbaa !11

it creates <3 x i8>, which is then shuffled into another <3 x i8>, which is
then bitcast to <4 x i8>:

  %vecinit.i = insertelement <3 x i8> undef, i8 %2, i32 0
  %vecinit1.i = insertelement <3 x i8> %vecinit.i, i8 %3, i32 1
  %vecinit4.i = insertelement <3 x i8> %vecinit1.i, i8 %4, i32 2

  %extractVec = shufflevector <3 x i8> %vecinit4.i, <3 x i8> undef, <4 x i32>
<i32 0, i32 1, i32 2, i32 undef>

  %storetmp = bitcast <3 x i8> addrspace(1)* %out to <4 x i8> addrspace(1)*
  ^^^ I originally claimed that this one fails, but it's a pointer cast so it
really shouldn't
  store <4 x i8> %extractVec, <4 x i8> addrspace(1)* %storetmp, align 4, !tbaa
!11

the attached .cl file will need libclc

-- 
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/20180929/7c09c1c0/attachment.html>