[PATCH] D61917: [IR] allow fast-math-flags on select of FP values

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In D61917#1510408 <https://reviews.llvm.org/D61917#1510408>, @hfinkel wrote:

> Just to make sure that I understand the general direction here: We want to make all fast-math flags apply only to inputs, not outputs, and so we're going to end up adding fast-math flags on all instructions that might take floating-point inputs?


I think we want to have FMF apply to the value itself. That should allow removing FMF from fcmp by adjusting our analysis/pattern matching to detect the FMF on fcmp's operands. That would simplify the definition of FPMathOperator too - if it's an FP value (produces an FP value), it's an FPMO. To make that complete, I think we'd eventually allow adding FMF as an attribute to FP function arguments.

Some motivating examples:

  define float @PR39535(float %x) {
    %cmp = fcmp nsz oeq float %x, 0.0           ; 'nsz' behavior is implicit in any fcmp, so this is strange
    %cond = select i1 %cmp, float %x, float 0.0 ; if this has 'nsz', we can simplify to '0.0' because both arms of the select are the same
    ret float %cond
  }
  
  define i1 @orderedLessZeroUIToFP_nnan(i32 %x) {
    %a = uitofp i32 %x to float                 ; if this has 'nnan', the result is a number and >=0.0 (the fcmp must be true)
    %uge = fcmp nnan oge float %a, 0.000000e+00
    ret i1 %uge
  }


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D61917/new/

https://reviews.llvm.org/D61917