[PATCH] D76838: [LV] WIP: Tail-folding counting down loops

[Sjoerd Meijer via Phabricator via llvm-commits]

SjoerdMeijer created this revision.
SjoerdMeijer added reviewers: Ayal, fhahn, samparker, hsaito.
Herald added subscribers: danielkiss, rkruppe, hiraditya, kristof.beyls.

This is a work-in-progress patch to support tail-folding for counting down loops, i.e. loops with a constant step value of -1.

The reason why tail-folding isn't triggered for these cases, is because induction variables with a -1 step value are not considered candidates as a primary induction variable. A minimal motivating example is as simple as this:

  void f (char *a, char *b, char *c, int N) {
    while (N-- > 0)
      *c++ = *a++ + *b++;
  }

This will (of course) be vectorised, but when tail-folding is requested quite early in the vectorisation pipeline, it hasn't discovered the primary induction variable, inhibiting tail-folding for counting down loops.

My motivating example is:

  test/Transforms/LoopVectorize/ARM/tail-folding-counting-down.ll

To make all differences explicit and visible, I have first locally regenerated all output for this example without this patch applied. This patch shows the differences compared to this local diff.

I am requesting early feedback on this patch because I need to deal with a handful of tests that need some special care. Motiving example for this is:

  test/Transforms/LoopVectorize/reverse_induction.ll

This needs the offset calculation as I have also commented and highlighted in that test:

      
  [[OFFSET_IDX:%.*]] = sub i64 [[STARTVAL]], [[INDEX]]

And to maintain this, I need to recognise the -1 step value in `CreateScalarIV()`, which I think this makes sense when we need a scalar IV and we're counting down, but this is what I wanted to check with you. As I said, this is work-progress, but if the approach is right, then I will clean up the code and modify some more tests.


https://reviews.llvm.org/D76838

Files:
  llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp
  llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
  llvm/test/Transforms/LoopVectorize/ARM/tail-folding-counting-down.ll
  llvm/test/Transforms/LoopVectorize/reverse_induction.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D76838.252802.patch
Type: text/x-patch
Size: 42888 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200326/ce37e9dd/attachment.bin>