[PATCH] D137159: [LoopVectorize] Introduce CanSinkMemoryInstrs to isFixedOrderRecurrence

[Peter Waller via Phabricator via llvm-commits]

peterwaller-arm created this revision.
Herald added subscribers: shiva0217, hiraditya.
Herald added a project: All.
peterwaller-arm requested review of this revision.
Herald added subscribers: llvm-commits, pcwang-thead.
Herald added a project: LLVM.

Given a loop of the form:

  double w[1024], x[1024], a[1024];
  void f() {
    for (unsigned i = 0; i < 1024-1; i++) {
      w[i] += a[i];
      x[i] += a[i+1];
    }
  }

Currently GVN PRE creates a phi which feeds a[i+1] into the next
iteration.

Prior to this patch, this phi is reported as being 'unknown':

  remark: loop not vectorized: value that could not be identified as
          reduction is used outside the loop [-Rpass-analysis=loop-vectorize]
      w[i] += a[i];
              ^

The reason is that isFixedOrderRecurrence sees the store to w[i]
occurring before the read from a[i+1] and it knows nothing about how
these are related.

This patch introduces a query for the memory dependences of the phi
input. If the memory dependency analysis reports a nonlocal dependency,
then any memory instructions can be sunk past it.

NOTE: I'm putting this patch up for early comments. Some things I'm
currently continuing to look into: Does LoopVectorize need the
memory dependence analysis, or is there other infrastructure available
that can be reused? Can I find any examples of miscompiles and does it
have the desired benchmarking impact?


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D137159

Files:
  llvm/lib/Analysis/IVDescriptors.cpp
  llvm/test/Transforms/LoopVectorize/AArch64/fixed-order-recurrence.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D137159.472276.patch
Type: text/x-patch
Size: 7949 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20221101/564d49e7/attachment.bin>