[PATCH] D39642: [ValueTracking] readnone is a requirement for converting sqrt to llvm.sqrt; nnan is not

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In https://reviews.llvm.org/D39642#917009, @efriedma wrote:

> I'm not sure how the C standard fits into this?  "-ffinite-math-only" mode intentionally doesn't conform to the C standard.


It's moot after this fix, but I was imagining a strict-mode case where we somehow determined that the sqrt does not have nan input or output.

  define double @foo(double %x)  {
  entry:
    %cmp.i = fcmp ord double %x, 0.0
    br i1 %cmp.i, label %no_nan_input, label %end
  
  no_nan_input:  
    %sqrt = tail call double @sqrt(double %x) #2   <--- some pass can mark this call with 'nnan' if it knows the libcall won't return nan?
    br label %return
  
  end:  
    %r = phi double [ %sqrt, %no_nan_input ], [ 1.0, %entry ]
    ret double %r
  }

So if that sqrt got marked with nnan, the vectorizers or whoever else is using getIntrinsicForCallSite() would convert the libcall to the intrinsic, but that wouldn't be correct if %x is a negative number? We would fail to set errno when we should have.


https://reviews.llvm.org/D39642