issue

Summary

InstCombine partially recognizes the common integer NaN test

(bitcast float to i32 & 0x7FFFFFFF) > 0x7F800000

but stops too early. It folds the sign-mask step to fabs, yet does not finish the job by converting the whole predicate to:

fcmp uno float %f, %f

That leaves x86 codegen at 4 instructions instead of the 2-instruction ucomiss plus setp sequence.

define i1 @float_is_nan(float %f) {
  %i = bitcast float %f to i32
  %a = and  i32 %i, 2147483647
  %r = icmp ugt i32 %a, 2139095040
  ret i1 %r
}

opt -O2 gets only part way there:

%1 = tail call float @llvm.fabs.f32(float %f)
%a = bitcast float %1 to i32
%r = icmp samesign ugt i32 %a, 2139095040
ret i1 %r

movd   %xmm0, %eax
andl   $2147483647, %eax
cmpl   $2139095041, %eax
setae  %al
retq

ucomiss  %xmm0, %xmm0
setp     %al
retq

That is exactly what LLVM produces for:

%r = fcmp uno float %f, %f

For IEEE single-precision:

So the two predicates are equivalent.