[PATCH] D118126: Correct instcombine of fcmp+select.

[Christian Sigg via Phabricator via llvm-commits]

csigg created this revision.
Herald added subscribers: sanjoy.google, hiraditya.
csigg requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

1. Zero fcmp arguments were previously replaced with zero select arguments without considering or adjusting the predicate. However, without no-signed-zeros, only the following are equivalent:

     >= +0  <=>     > -0
     <= -0  <=>     < +0
  -0 >=     <=>  +0 >
  +0 <=     <=>  -0 <



2. Combining fcmp+select into fminnum is not permitted when the arguments might be two zeros of opposite sign. E.g., fminnum(-0, +0) can return any of the two operands and is therefore not equivalent to '-0 < +0 ? -0 : +0'. This was handled for >= and <= but marked as 'FIXME' for > and <.

3. fminimum does treat -0 as smaller than +0, so fcmp+select can be combined if the NaN-propagating bahaviour of fminimum is permitted. Previously, the handling of signed zeros applied to combining fcmp+select into both fminnum and fminimum.

The same applies for fmaxnum/fmaximum.

Hopefully 3) can make up for the performance lost by fixing 1) and 2).


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D118126

Files:
  llvm/lib/Analysis/ValueTracking.cpp
  llvm/test/CodeGen/AArch64/arm64-fmax-safe.ll
  llvm/test/CodeGen/AArch64/arm64-fmax.ll
  llvm/test/CodeGen/ARM/fp16-vminmaxnm-safe.ll
  llvm/test/CodeGen/ARM/neon_minmax.ll
  llvm/test/CodeGen/ARM/vminmaxnm-safe.ll
  llvm/test/Transforms/InstCombine/minmax-fp.ll
  llvm/unittests/Analysis/ValueTrackingTest.cpp

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D118126.402853.patch
Type: text/x-patch
Size: 12947 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220125/dc48f4be/attachment.bin>