[all-commits] [llvm/llvm-project] cfdea8: [MLIR][Tosa] Fix fp canonicalization for `clamp`

[Rik Huijzer via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: cfdea8f5bb6b0138719c5b4f3fe1616b6ea915a7
      https://github.com/llvm/llvm-project/commit/cfdea8f5bb6b0138719c5b4f3fe1616b6ea915a7
  Author: rikhuijzer <rikhuijzer at pm.me>
  Date:   2023-06-07 (Wed, 07 Jun 2023)

  Changed paths:
    M mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
    M mlir/test/Dialect/Tosa/canonicalize.mlir

  Log Message:
  -----------
  [MLIR][Tosa] Fix fp canonicalization for `clamp`

The canonicalization for `clamp` removed the `clamp` operation when the
chosen min and max values were outside the range of what is possible to
represent with the input type. For example, if the input type is `i8`,
then the min and max values must be between -128 and 127. If the min and
max are, say, -200 and 200, then the `clamp` operation can be safely
removed.

However, as pointed out by @wyanzhao, this is wrong for floating point
types since they can represent infinity.

Fixes #62341.

Reviewed By: jpienaar

Differential Revision: https://reviews.llvm.org/D152123