Any significance for m_OneUse in (X / Y) / Z => X / (Y * Z) ??

Dear All,

The InstCombine pass performs the following transformation.

Z / (X / Y) => (Y * Z) / X

This is performed only when operand Op1 ( (X/Y) in this case) has only one use in future. The code snippet is shown below.

if (match(Op1, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) &&
(!isa(Y) || !isa(Op0))) {
// Z / (X / Y) => (Y * Z) / X
Value *YZ = Builder.CreateFMulFMF(Y, Op0, &I);
return BinaryOperator::CreateFDivFMF(YZ, X, &I);
}

It would be great if someone explains if there is any issue (correctness/performance-wise) if we avoid the m_OueUse check. What if we perform the transformation even if there are multiple uses?

There are similar transformations which perform the m_OueUse check. We may avoid those too if there is no particular reason for the check.

Regards,

As a general rule, InstCombine tries not increase the total number of instructions. If X/Y has another use other than this one, then it still ends up being calculated. Without the one use check you’d trade 2 fdivs, for 1 mul (Y * Z), and 2 fdivs ((X*Y)/Z) and the original (X / Y).

Thanks a lot Craig. That is convincing.

Regards,

A couple more general comments:

  1. There shouldn’t be any correctness issues removing one-use checks (the transform should be safe independently of use-counts).
  2. Ideally, you can remove a m_OneUse() from the code and run ‘make check’ or similar, and you will see a regression test failure because we have a ‘negative’ test to cover that pattern. That should make it clear that the one-use check is there intentionally and what the effect of removing it is. We’ve gotten better about including those kinds of regression tests over time, but I don’t know what percentage of all instcombine transforms actually have that test coverage.

Thanks Sanjay for your comments.

So, if we want to add a transformation avoiding the one-use check, which one is the ideal pass? Shall we do it in -aggressive-instcombine? I came to know that if the pattern search complexity is O(n) [1] we should go for aggressive-instcombine. If it is O(1) we must do that in -instcombine. However, in my case, the complexity is still O(1) and want to avoid the one-use check.

[1] n is the number of instructions in the function.

Regards,

Is your case the case mentioned in the subject or a different case?

Here is my case.

Z / (1.0 / Y) => (Y * Z)

This is similar to Z / (X / Y) => (Y * Z) / X. Currently, the former one is prohibitted because of the one-use check. In the latter, as you explained earlier, the number of instructions are increased from 2 to 3. However, in the former case (where X = 1.0), the number of instructions remain the same as the division by 1.0 is avoided. Additionally, instead of a division, now we have a multiplication. This potentially may reduce the number of instruction cycles.

Regards,

You can match that in InstCombine without a one use check if you explicitly match the 1.0. I think we generally understand that multiply is better than divide.

If the transform is a performance optimization despite potentially increasing the number of instructions, that should go in the backend. Most likely that will be in DAGCombiner or a target-specific lowering file.

But as noted in the later reply, if this question is really about a reciprocal special-case, that becomes an extension of what we already have in instcombine.

To be clear - we’re talking about FP math, and all of these transforms require some form of fast-math-flags to be valid (no matter where they are implemented).

Thank You.

Yes. It is a reciprocal special-case and expected to be invoked only in -fast-math.

Regards,