[all-commits] [llvm/llvm-project] 5f01a6: [ConstantFold] Fix inbounds inference on mismatchi...

[Nikita Popov via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 5f01a626dd0615df49773d419c75aeb33666ee83
      https://github.com/llvm/llvm-project/commit/5f01a626dd0615df49773d419c75aeb33666ee83
  Author: Nikita Popov <npopov at redhat.com>
  Date:   2023-01-31 (Tue, 31 Jan 2023)

  Changed paths:
    M clang/test/CodeGenCXX/no-unique-address.cpp
    M llvm/lib/IR/ConstantFold.cpp
    M llvm/test/Assembler/opaque-ptr.ll
    M llvm/test/Transforms/InferAddressSpaces/AMDGPU/infer-getelementptr.ll
    M llvm/test/Transforms/InstCombine/getelementptr.ll
    M llvm/test/Transforms/InstCombine/memchr-9.ll
    M llvm/test/Transforms/InstCombine/wcslen-1.ll
    M llvm/test/Transforms/InstSimplify/compare.ll

  Log Message:
  -----------
  [ConstantFold] Fix inbounds inference on mismatching source element type

When inferring that a GEP of a global variable is inbounds because
there is no notional overindexing, we need to check that the
global value type and the GEP source element type match.

This was not necessary with typed pointers (because we would have
a bitcast in between), but is necessary with opaque pointers.

We should be able to recover some of the safe cases by performing
an offset based inbounds inference in DL-aware ConstantFolding.