[all-commits] [llvm/llvm-project] 3c77eb: [DAGCombiner] Only fold shift+mask compare to rota...

[Stanislav Bardyuk via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 3c77eb3a3f2793a75e0ce86d96076637d5989bea
      https://github.com/llvm/llvm-project/commit/3c77eb3a3f2793a75e0ce86d96076637d5989bea
  Author: Stanislav Bardyuk <6258615+kodlan at users.noreply.github.com>
  Date:   2026-09-13 (Sun, 13 Sep 2026)

  Changed paths:
    M llvm/include/llvm/CodeGen/TargetLowering.h
    M llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
    M llvm/lib/Target/X86/X86ISelLowering.cpp
    M llvm/test/CodeGen/X86/cmp-shiftX-maskX.ll

  Log Message:
  -----------
  [DAGCombiner] Only fold shift+mask compare to rotate if the shift amount divides the bit width (#220743)

DAGCombiner::visitSETCC can turn `(X << S) == (X & -(1 << S))` (or the
srl variant) into `rotl(X, S) == X`, and the other way around, when the
target prefers one form over the other. The two forms are only equivalent when
S divides the bit width. The check used `isPowerOf2(S)`, which is the same
thing for i8/i16/i32/i64 but not for odd widths like i3, where S = 2 is a
power of two that does not divide 3. On x86, which prefers the rotate,
the reporter's i3 function returned the wrong value for two of its eight
inputs.

This changes the `MayTransformRotate` flag passed to
`preferedOpcodeForCmpEqPiecesOfOperand` to check `NumBits % S == 0`
(with S != 0) and fixes the two comments that described the old rule. For legal
scalar types nothing changes. On illegal non power of two widths this
also lets a non power of two amount that divides the width use the rotate
(like i6 with S = 3); that is correct for the same reason and legalizes
through the usual rotate promotion, same as i6 with S = 2 did before. The x86
hook already honors the flag in both directions, so the fix lives in the
target independent code.

Tests: the i3 shl and lshr cases from the issue (must stay shift+mask), an
i6 / S = 3 case (3 divides 6, so the rotate is still allowed), and the
reporter's function.

Fixes #220542



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