[all-commits] [llvm/llvm-project] 7c8e05: [SCEV] Collect and merge loop guards through PHI n...

[Julian Nagele via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 7c8e05aa45f006401b71b37127537c4682fe16ee
      https://github.com/llvm/llvm-project/commit/7c8e05aa45f006401b71b37127537c4682fe16ee
  Author: Julian Nagele <j.nagele at apple.com>
  Date:   2024-11-15 (Fri, 15 Nov 2024)

  Changed paths:
    M llvm/include/llvm/Analysis/ScalarEvolution.h
    M llvm/lib/Analysis/ScalarEvolution.cpp
    A llvm/test/Analysis/ScalarEvolution/backedge-taken-count-guard-info-with-multiple-predecessors.ll
    M llvm/test/Transforms/PhaseOrdering/X86/pr38280.ll

  Log Message:
  -----------
  [SCEV] Collect and merge loop guards through PHI nodes with multiple incoming values (#113915)

This patch aims to strengthen collection of loop guards by processing
PHI nodes with multiple incoming values as follows: collect guards for
all incoming values/blocks and try to merge them into a single one for
the PHI node.

The goal is to determine tighter bounds on the trip counts of scalar
tail loops after vectorization, helping to avoid unnecessary transforms.
In particular we'd like to avoid vectorizing scalar tails of
hand-vectorized loops, for example in
[Transforms/PhaseOrdering/X86/pr38280.ll](https://github.com/llvm/llvm-project/blob/231e03ba7e82896847dbc27d457dbb208f04699c/llvm/test/Transforms/PhaseOrdering/X86/pr38280.ll),
discovered via https://github.com/llvm/llvm-project/pull/108190

Compile-time impact: https://llvm-compile-time-tracker.com/compare.php?from=a55248789ed3f653740e0723d016203b9d585f26&to=500e4c46e79f60b93b11a752698c520e345948e3&stat=instructions:u

PR: https://github.com/llvm/llvm-project/pull/113915



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