[PATCH] D74185: Revert the revert of vectorization commits

[George Karpenkov via Phabricator via llvm-commits]

george.karpenkov created this revision.
george.karpenkov added reviewers: ABataev, bkramer, sanjoy.google.
Herald added subscribers: Charusso, jfb, mgrang, hiraditya.
Herald added a project: LLVM.

Revert of https://github.com/llvm/llvm-project/commit/bfa32573bf2d0ab587f9a5d933ea2144a382cf3c

The vectorization introduced by the commit causes a miscompile.

This is the input LLVM IR: https://gist.github.com/cheshire/9dff84d8fbb83278736278854c746d8d
This is IR with the mentioned commit reverted: https://gist.github.com/cheshire/fb06443834735c2fa04fa0eacb288606 (produces correct results)
This is IR with ToT LLVM: https://gist.github.com/cheshire/d0b4195b1be344f52d77b98c70b5241d

If just a diff in IR is not enough, the IR dump is actually runnable.

To run it, compile a standalone (no dependencies, only a single file) driver tool at https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/xla/tools/driver.cc and link with good/bad IR version above. Then run it with a single argument, containing a filename with the buffer assignment description (contents at https://gist.github.com/cheshire/f929aa118978f4bcad03a3da11209d27).

With the vectorization reverted, the driver produces this output:

  Output:
  (
  1.15675, 1.04532, 1.1549, 1.04836, 0.947364, 1.04667, 1.04491, 0.944249, 1.04323, 0.94699, 0.855764, 0.945469, 1.1536, 1.04247, 1.15175, 1.0455, 0.944784, 1.04382
  ,
  1.15675, 1.04532, 1.1549, 1.04836, 0.947364, 1.04667, 1.04491, 0.944249, 1.04323, 0.94699, 0.855764, 0.945469, 1.1536, 1.04247, 1.15175, 1.0455, 0.944784, 1.04382
  )

Without the revert, the output is different:

  Output:
  (
  1.15675, 1.04532, 1.1549, 1.04836, 0.947364, 1.04667, 1.04491, 0.944249, 1.04323, 0.94699, 0.855764, 0.945469, 1.1536, 1.04247, 1.15175, 1.0455, 0.944784, 1.04382
  ,
  1.15675, 1.10195, 1.1549, 1.04836, 0.947364, 1.04667, 1.04491, 0.899154, 1.04323, 0.94699, 0.855764, 0.945469, 1.1536, 1.09595, 1.15175, 1.0455, 0.944784, 1.04382
  )

(note the 8th number in the second row)


https://reviews.llvm.org/D74185

Files:
  llvm/include/llvm/Transforms/Vectorize/SLPVectorizer.h
  llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
  llvm/test/Transforms/SLPVectorizer/X86/arith-add-ssat.ll
  llvm/test/Transforms/SLPVectorizer/X86/arith-add-usat.ll
  llvm/test/Transforms/SLPVectorizer/X86/arith-add.ll
  llvm/test/Transforms/SLPVectorizer/X86/arith-fix.ll
  llvm/test/Transforms/SLPVectorizer/X86/arith-mul.ll
  llvm/test/Transforms/SLPVectorizer/X86/arith-sub-ssat.ll
  llvm/test/Transforms/SLPVectorizer/X86/arith-sub-usat.ll
  llvm/test/Transforms/SLPVectorizer/X86/arith-sub.ll
  llvm/test/Transforms/SLPVectorizer/X86/bitreverse.ll
  llvm/test/Transforms/SLPVectorizer/X86/ctlz.ll
  llvm/test/Transforms/SLPVectorizer/X86/ctpop.ll
  llvm/test/Transforms/SLPVectorizer/X86/cttz.ll
  llvm/test/Transforms/SLPVectorizer/X86/different-vec-widths.ll
  llvm/test/Transforms/SLPVectorizer/X86/jumbled_store_crash.ll
  llvm/test/Transforms/SLPVectorizer/X86/pr35497.ll
  llvm/test/Transforms/SLPVectorizer/X86/shift-ashr.ll
  llvm/test/Transforms/SLPVectorizer/X86/shift-lshr.ll
  llvm/test/Transforms/SLPVectorizer/X86/shift-shl.ll
  llvm/test/Transforms/SLPVectorizer/X86/store-jumbled.ll
  llvm/test/Transforms/SLPVectorizer/X86/stores_vectorize.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D74185.243056.patch
Type: text/x-patch
Size: 162985 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200207/038aff25/attachment-0001.bin>