[all-commits] [llvm/llvm-project] 076266: [SLP] Partially revert #224931 for alternate nodes

[hidekisaito via All-commits]

  Branch: refs/heads/users/hidekisaito/partial-revert-224931
  Home:   https://github.com/llvm/llvm-project
  Commit: 076266fc9704c7efc786d6e020a69cfc6967b809
      https://github.com/llvm/llvm-project/commit/076266fc9704c7efc786d6e020a69cfc6967b809
  Author: Hideki Saito <hidekido at amd.com>
  Date:   2026-10-03 (Sat, 03 Oct 2026)

  Changed paths:
    M llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
    A llvm/test/Transforms/SLPVectorizer/AMDGPU/alt-fmul-fadd-cost.ll

  Log Message:
  -----------
  [SLP] Partially revert #224931 for alternate nodes

Do not pass scalar context to alternate node vector cost queries.

PR #224931 passes the scalar main/alternate op as the context instruction when
costing the vector ops of an alternate node. Those vector ops only feed the
lane-select shuffle, so use-based discounts derived from the scalar's users
do not apply. On AMDGPU this priced an [fmul | fadd] node's vector fmul as
fused (free) where the sole user of the scalar fmul is fadd/fsub, keeping
unprofitable subtrees vectorized. In a rocFFT kernel on gfx950 this raised
VGPR usage from 160 to 178 and caused a performance regression. Restore the
null context for these queries.



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