[all-commits] [llvm/llvm-project] 72e9d3: Reapply "[SelectionDAG] Recurse through mask expre...

[Valeriy Savchenko via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 72e9d33d2945504159c8d316d12c187fc706b156
      https://github.com/llvm/llvm-project/commit/72e9d33d2945504159c8d316d12c187fc706b156
  Author: Valeriy Savchenko <vsavchenko at apple.com>
  Date:   2026-09-21 (Mon, 21 Sep 2026)

  Changed paths:
    M llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
    M llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
    M llvm/test/CodeGen/AArch64/arm64-zip.ll
    A llvm/test/CodeGen/AArch64/vselect-widen-mask-tree.ll
    M llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-sext.ll
    M llvm/test/CodeGen/X86/freeze-vector.ll
    A llvm/test/CodeGen/X86/vselect-widen-mask-tree.ll

  Log Message:
  -----------
  Reapply "[SelectionDAG] Recurse through mask expression trees in WidenVSELECTMask" (#217307)

WidenVSELECTMask currently handles only two mask shapes: a bare SETCC
or a single AND/OR/XOR of exactly two SETCCs. Anything deeper bails out
to the generic condition widening path, which often introduces
unnecessary narrow/widen roundtrips (xtn+sshll on AArch64,
packssdw+vpmovsxwd on X86).

Replace the hand-coded cases with a recursive widenMaskTree that walks
through SETCC, AND/OR/XOR, FREEZE, VECTOR_SHUFFLE, SELECT/VSELECT, and
all-ones/all-zeros BUILD_VECTORs.

I had to revert #188085 earlier this year. This PR resolves the flagged
issue and cleans up suggested approach to avoid that class of potential
errors.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications