[Mlir-commits] [mlir] [MLIR][Math] add canonicalize-f32-promotion pass (PR #92482)

[Krzysztof Drewniak]

================
@@ -109,4 +135,14 @@ void LegalizeToF32Pass::runOnOperation() {
   math::populateLegalizeToF32Patterns(patterns, typeConverter);
   if (failed(applyPartialConversion(op, target, std::move(patterns))))
     return signalPassFailure();
+  
+  if (useCanonicalizeF32Promotion) {
----------------
krzysz00 wrote:

I don't really like this approach.

How about going up to `matchAndRewrite()` and doing this:

```c++
SmallVector<Value> extendedWidthOperands(operands);
for (auto [extended, original] : llvm::zip_equal(extendedWidthOperands, op->getOperands()) {
  // match trunc/ext pair. The inelegant version is.
  if (auto short = extended.getDefiningOp<arith::TruncFOp>()) {
    auto maybeOriginal = extended.getIn().getDefiningOp<arith;:ExtFOp>());
    if (maybeOriginal && maybeOriginal.getIn() == original)
      extended = original;
  }
  convertOpResultTypes(..., extendedWidthOperands, ...);
```

Now, you don't need a pass option, and all you're doing is "if this is the extension of the truncation of my original argument, use that original argument instead".

https://github.com/llvm/llvm-project/pull/92482