[all-commits] [llvm/llvm-project] 017016: Transforms: Clone distinct nodes in metadata mappe...

[Duncan P. N. Exon Smith via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 01701646d5555409a6fad920f0e4801d79c154ea
      https://github.com/llvm/llvm-project/commit/01701646d5555409a6fad920f0e4801d79c154ea
  Author: Duncan P. N. Exon Smith <dexonsmith at apple.com>
  Date:   2021-02-24 (Wed, 24 Feb 2021)

  Changed paths:
    M llvm/include/llvm/IR/DebugInfo.h
    M llvm/lib/Transforms/Utils/CloneFunction.cpp
    M llvm/lib/Transforms/Utils/ValueMapper.cpp
    A llvm/test/Transforms/ThinLTOBitcodeWriter/cfi-debug-info-cloned-type-references-global-value.ll

  Log Message:
  -----------
  Transforms: Clone distinct nodes in metadata mapper unless RF_ReuseAndMutateDistinctMDs

This is a follow up to 22a52dfddcefad4f275eb8ad1cc0e200074c2d8a and a
revert of df763188c9a1ecb1e7e5c4d4ea53a99fbb755903.

With this change, we only skip cloning distinct nodes in
MDNodeMapper::mapDistinct if RF_ReuseAndMutateDistinctMDs, dropping the
no-longer-needed local helper `cloneOrBuildODR()`.  Skipping cloning in
other cases is unsound and breaks CloneModule, which is why the textual
IR for PR48841 didn't pass previously. This commit adds the test as:
Transforms/ThinLTOBitcodeWriter/cfi-debug-info-cloned-type-references-global-value.ll

Cloning less often exposed a hole in subprogram cloning in
CloneFunctionInto thanks to df763188c9a1ecb1e7e5c4d4ea53a99fbb755903's
test ThinLTO/X86/Inputs/dicompositetype-unique-alias.ll. If a function
has a subprogram attachment whose scope is a DICompositeType that
shouldn't be cloned, but it has no internal debug info pointing at that
type, that composite type was being cloned. This commit plugs that hole,
calling DebugInfoFinder::processSubprogram from CloneFunctionInto.

As hinted at in 22a52dfddcefad4f275eb8ad1cc0e200074c2d8a's commit
message, I think we need to formalize ownership of metadata a bit more
so that ValueMapper/CloneFunctionInto (and similar functions) can deal
with cloning (or not) metadata in a more generic, less fragile way.

This fixes PR48841.

Differential Revision: https://reviews.llvm.org/D96734