Question about canonicalization `tensor.log(tensor.exp(x)) -> x`

MLIR TOSA
1

Hi!

When I read canonicalization passes of tosa, I found this one:

OpFoldResult tosa::ExpOp::fold(FoldAdaptor adaptor) {
  auto input = getInput1();
  // Element-wise exp(log(x)) = x
  if (auto op = input.getDefiningOp<tosa::LogOp>()) {
    return op.getInput1();
  }

  return {};
}

However, if x is too large, then exp(x) would overflow and might not be x after applying log. Am I misunderstanding here? Why this optimization is valid?

Thanks for help!

All the best,

Yuyou

2

Hi @Hatsunespica ,

I suspect you’re correct here and that this is a bug, I think we should remove these canonicalizations. Possibly we could re-introduce these optimizations under a “fast-math“ sort of transformation if they are useful.

1 Like
3

This seems like a typical use of peephole transformation specifically to avoid such overflow cases as the OP had mentioned. Would like to know why this is wrong in a bit more detail.

From: Luke Hutton via LLVM Discussion Forums notifications@llvm.discoursemail.com
Date: Tuesday, August 26, 2025 at 11:07 AM
To: Muthu Annamalai mannamalai@cornami.com
Subject: [EXTERNAL] [LLVM] [MLIR/TOSA] Question about canonicalization tensor.log(tensor.exp(x)) -> x


lhutton1
August 26

Hi @Hatsunespica ,

I suspect you’re correct here and that this is a bug, I think we should remove these canonicalizations. Possibly we could re-introduce these optimizations under a “fast-math“ sort of transformation if they are useful.

Visit Topic or reply to this email to respond.

To unsubscribe from these emails, click here.

4

Potentially I’m not following the concern here, apologies.

My understanding of the OP was the canonicalization log(exp(x)) → x would alter the numerical semantics of the graph under TOSA’s floating-point behaviour rules. If we don’t run canonicalization, the result is inf but if canonicalization is run the result is x.

This seems undesirable to me particularly for a canonicalization which I believe should be semantics-preserving, and might be better suited to a separate, opt-in, transformation with relaxed FP expectations.