[PATCH] D151937: InstCombine: Recognize fabs as bitcasted integer

[Bjorn Pettersson via Phabricator via llvm-commits]

bjope added a comment.

Having some problems understanding how this interacts with `-fdenormal-fp-math`. Our target is using `preserve-sign`. And the fabs is lowered into a floating point instructions that would give a normalized output.

So a test case that is doing

  float input = -0x1.fffffcp-127 + 0.0f;
  
  int main() {
    union {
      float f;
      ui32_t i;
    } rep;
    rep.f = input;
  
    const ui32_t aInt = rep.i;
    const ui32_t aAbs = aInt & 0x7ffffffful;
  
    if (aAbs == 0) {
      printf("masked %f is zero\n", input);
    } else {
      printf("masked %f is not zero\n", input);
    }
    ...
  }

used to result in "masked -0.000000 is not zero" being printed (since the store to rep.f is storing a denormal number to memory).

But with this patch the compiler introduces and fabs operation. And then our backend picks an floating point instruction that has flushed the result to zero. So with this patch it starts printing "masked -0.000000 is zero" instead.

So is this transform bad when denormal-fp-math isn't set to ieee?
Or is it something with the frontend (or test program) that is bad here. What should happen when storing that denormal literal to memory. Should that involve a flush-to-zero as well?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D151937/new/

https://reviews.llvm.org/D151937