[PATCH] D105238: GlobalISel/AArch64: don't optimize away redundant branches at -O0

[Adrian Prantl via Phabricator via llvm-commits]

aprantl added a comment.

In D105238#2853402 <https://reviews.llvm.org/D105238#2853402>, @aemerson wrote:

> @aprantl by the way, I'm pretty sure there are a fair number of tests that need updating with this change.

Yes that' right:

  Failed Tests (12):
    LLVM :: CodeGen/AArch64/GlobalISel/arm64-atomic.ll
    LLVM :: CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll
    LLVM :: CodeGen/AArch64/unwind-preserved.ll
    LLVM :: CodeGen/BPF/BTF/func-func-ptr.ll
    LLVM :: CodeGen/BPF/BTF/func-non-void.ll
    LLVM :: CodeGen/BPF/BTF/func-typedef.ll
    LLVM :: CodeGen/BPF/BTF/func-unused-arg.ll
    LLVM :: CodeGen/BPF/BTF/func-void.ll
    LLVM :: CodeGen/Mips/GlobalISel/llvm-ir/jump_table_and_brjt.ll
    LLVM :: CodeGen/Mips/GlobalISel/llvm-ir/long_ambiguous_chain_s32.ll
    LLVM :: CodeGen/Mips/GlobalISel/llvm-ir/long_ambiguous_chain_s64.ll
    LLVM :: CodeGen/Mips/GlobalISel/llvm-ir/phi.ll

Before I go through updating all the tests, do we think that this is a reasonable change?

Pros:

- Consistent with SelectionDAG behavior (fixes a "regression" in GlobalISEL)
- Enables debugging scenarios (setting breakpoints on early returns, etc)

Cons:

- Code size increase (-O0 only)

The patch improves debuggability, and debug builds are the primary reason why developers compile without optimizations. From my point of view this seems to be the right trade-off, but please let me know if there are any other interests we need to consider.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D105238/new/

https://reviews.llvm.org/D105238