[PATCH] D144089: [IndVarSimplify] Transform the trip count to a simpler form

[Tiehu Zhang via Phabricator via llvm-commits]

TiehuZhang created this revision.
TiehuZhang added reviewers: fhahn, dmgreen, sdesmalen, lebedev.ri.
Herald added subscribers: StephenFan, javed.absar, hiraditya.
Herald added a project: All.
TiehuZhang requested review of this revision.
Herald added subscribers: llvm-commits, pcwang-thead.
Herald added a project: LLVM.

LoopVectorizer will create a loop guard based on the SCEV expression of ExitCount, and complex ExitCount will introduce redundant instructions. This patch simplifies the ExitCount by offsetting the trip count and induction variable so that there will be fewer redundant conditions after vectorization, e.g., llvm.smax.i64 as shown in test.

      

Before the optimization

  int initial = 0;
  for(int i = 0; i < nOut; i++) {
    double temp_value = 0.0;
    for (int j = initial; j < initial + nIn; j++) {
      temp_value += values[j] * x[idx[j]];
    }
    initial += nIn;
    b[i] = temp_value;
  }


After the optimization

  int initial = 0;
  for(int i = 0; i < nOut; i++) {
    double temp_value = 0.0;
    for (int k = 0; k < nIn; k++) {
      temp_value += values[k + initial] * x[idx[k + initial]];
    }
    initial += nIn;
    b[i] = temp_value;
  }


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D144089

Files:
  llvm/lib/Transforms/Scalar/IndVarSimplify.cpp
  llvm/test/Transforms/IndVarSimplify/simplify-tripcount.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D144089.497625.patch
Type: text/x-patch
Size: 29126 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230215/9440a000/attachment.bin>