[PATCH] D67434: [InstCombine] Limit FMul constant folding for fma simplifications.

[Florian Hahn via Phabricator via llvm-commits]

fhahn created this revision.
fhahn added reviewers: spatel, lebedev.ri, reames.
Herald added a subscriber: hiraditya.
Herald added a project: LLVM.

As @reames pointed out post-commit, rL371518 <https://reviews.llvm.org/rL371518> adds additional rounding
in some cases, when doing constant folding of the multiplication.
This breaks a guarantee llvm.fma makes and must be avoided.

This patch guards the simplification by additional conditions, to
ensure we do not constant fold when there could be rounding.
It should be fine to use SimplifyFMulInst, if either

  - one of the operands is not constant (no constant folding)
  - Either operands is either 1.0 or 0.0. The result won't require rounding.
- we are processing a fmuladd intrinsic (no rounding guarantees)

It also expands on the legality reasoning in a comment.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D67434

Files:
  llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
  llvm/test/Transforms/InstCombine/fma.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D67434.219679.patch
Type: text/x-patch
Size: 5323 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20190911/3516ae63/attachment.bin>