[clang] Miscompile at -O2/3

[cardigan1008]

This code prints 83619472 at -O0/1 and 83619480 at -O2/3:

int printf(const char *, ...);
int a, b;
int main() {
  int c = 7;
  float d = 0;
  while (c) {
    d++;
    c /= 2;
    a = 384 + 1;
  }
  float e = a + 2;
  b = d + 1.442695f * (e * e * e);
  printf("%d\n", b);
}

Compiler Explorer: https://godbolt.org/z/f5bf9os1P

Bisected to 17d9560, which was committed by @zahiraam

[antoniofrighetto]

When clang is passed -ffp-contract=off/fast, ConstantFold folds the following:

%mul2 = fmul float 0x418BA344E0000000, 0x3FF7154760000000
%add3 = fadd float 3.000000e+00, %mul

resulting in 0x4193EFBA40000000 via separate rounding. Conversely, with -ffp-contract=on, a FMA is emitted, which is expected to perform one round:

%0 = tail call float @llvm.fmuladd.f32(float 0x418BA344E0000000, float 0x3FF7154760000000, float 3.000000e+00)

resulting in 0x4193EFBA60000000. Should we check if the result of APFloat::fusedMultiplyAdd was rounded (opInexact), and fall back to double rounding, if that is the case? cc/ @arsenm, @dtcxzyw

[arsenm]

Not a bug. It is never wrong to treat fmuladd as an FMA.