[PATCH] D69629: [AMDGPU] Improve code size cost model (part 2)

[Jay Foad via Phabricator via llvm-commits]

foad added a comment.

@dfukalov this is breaking some of our internal tests. Here's an example:

  define amdgpu_kernel void @shufflevector_xxx(<4 x i8> addrspace(1)* %out, <2 x i8> addrspace(1)* %vaddr) {
    %vec = load <2 x i8>, <2 x i8> addrspace(1)* %vaddr
    %shuf = shufflevector <2 x i8> %vec, <2 x i8> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
    store <4 x i8> %shuf, <4 x i8> addrspace(1)* %out
    ret void
  }

I get:

  opt: /home/jayfoad2/git/llvm-project/llvm/lib/Analysis/TargetTransformInfo.cpp:206: int llvm::TargetTransformInfo::getUserCost(const llvm::User *, ArrayRef<const llvm::Value *>) const: Assertion `Cost >= 0 && "TTI should not produce negative costs!"' failed.

Can you take a look please?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69629/new/

https://reviews.llvm.org/D69629