[all-commits] [llvm/llvm-project] 0a5793: [InstCombine] Preserve IR flags when reordering ic...

[ayrai-gb via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 0a5793ba557fb31ea1f847aa353397bb9b88cb18
      https://github.com/llvm/llvm-project/commit/0a5793ba557fb31ea1f847aa353397bb9b88cb18
  Author: ayrai-gb <ayrai at amd.com>
  Date:   2026-07-14 (Tue, 14 Jul 2026)

  Changed paths:
    M llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
    A llvm/test/Transforms/InstCombine/shuffle-cmp-cast-preserve-flags.ll

  Log Message:
  -----------
  [InstCombine] Preserve IR flags when reordering icmp/fcmp/cast through a shuffle (#208627)

InstCombine's `evaluateInDifferentElementOrder` pushes a lane-reordering
`shufflevector` *through* the instruction that produced its input,
rebuilding that instruction on the reordered operands via the `buildNew`
helper — a `switch` with one case per opcode. The binary-operator case
carefully re-applies the original's poison-generating flags
(`nuw`/`nsw`, `exact`, FMF), and the `GetElementPtr` case passes through
its `getNoWrapFlags()`.

The `ICmp`, `FCmp`, and cast cases are the odd ones out. Each is a terse
`return Builder.CreateICmp/CreateFCmp/CreateCast(...)` that copies only
the predicate/opcode, returning the rebuilt instruction with default
flags. As a result a reordered `icmp samesign`, `fcmp nnan ninf`, `zext
nneg`, or `trunc nuw`/`nsw` silently loses its flag, even though the
flag held on the original.

Route each rebuilt instruction through `copyIRFlags(I)`, bringing these
cases in line with the binary-op and GEP cases. `copyIRFlags` transfers
only the flags the source instruction actually carried, so this can
never introduce an unsound promise. This is a missed-optimization fix,
not a miscompile — dropping these flags only makes later passes more
conservative.

Added
`llvm/test/Transforms/InstCombine/shuffle-cmp-cast-preserve-flags.ll`
(`icmp_samesign`, `fcmp_fmf`, `zext_nneg`), which feed the compare/cast
from `insertelement` chains so that `canEvaluateShuffled` permits the
reorder and `buildNew` actually runs, and check that the flag survives
on the rebuilt instruction after the reversing shuffle.

Found via jlebar's X86 LLVM bug-hunt / FuzzX effort:
https://github.com/SemiAnalysisAI/FuzzX/tree/master/x86/bugs/241-instcombine-buildNew-shuffle-reorder-drops-cmp-cast-flags

---------

Co-authored-by: Ayush Rai <ayrai at amd.com>



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