[PATCH] D147373: [VectorCombine] fold vector reverse loops

[Zhengyang Liu via Phabricator via llvm-commits]

liuz created this revision.
liuz added reviewers: spatel, RKSimon, hfinkel, efriedma, craig.topper, lebedev.ri, nemanjai.
Herald added subscribers: StephenFan, hiraditya.
Herald added a project: All.
liuz requested review of this revision.
Herald added subscribers: llvm-commits, pcwang-thead.
Herald added a project: LLVM.

LLVM's auto-vectorizer sometimes generates silly reverse loops, where a value is reversed by a shuffle, followed by some arithmetic, and then get reversed back.

For example, LLVM compiles this code sequence:

  do {
    if (*--p == '.') *p = '_';
  } while (p != name);

into

  %1 = shufflevector <32 x i8> %0, poison, <31, 30, 29, 28, 27, ... 4, 3, 2, 1, 0>
  %2 = icmp eq <32 x i8> %1, <46, 46, 46, 46, 46, ... 46, 46, 46, 46, 46>
  %3 = shufflevector <32 x i1> %2, poison, <31, 30, 29, 28, 27, ... 4, 3, 2, 1, 0>

Obviously reverse shuffle %1 and %3 are redundant, and the sequence can be optimized into

  %1 = icmp eq <32 x i8> %0, <46, 46, 46, 46, 46, ... 46, 46, 46, 46, 46>

This sequence is found in gzip's `make_simple_name` function.

This patch implements this rule in VectorCombine. Currently it works when %2 is a `cmp` instruction comparing with an splat constant. I will add other cases in the future passes.

The test cases in this patch are validated with Alive2.

This rewrite pattern is found by the Minotaur superoptimizer (https://github.com/minotaur-toolkit/minotaur).


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D147373

Files:
  llvm/lib/Transforms/Vectorize/VectorCombine.cpp
  llvm/test/Transforms/VectorCombine/reverse-loop.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D147373.510199.patch
Type: text/x-patch
Size: 20455 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230401/46c3d6ac/attachment.bin>