We can't propagate FMF partially without breaking DAG-level CSE. We either need to
relax CSE to account for mismatched FMF as a temporary work-around or fully propagate
FMF throughout the DAG.
Surprisingly, there are no existing regression tests for this, but here's an example:
define float @fmf(float %a, float %b) {
%mul1 = fmul fast float %a, %b
%nega = fsub fast float 0.0, %a
%mul2 = fmul fast float %nega, %b
%abx2 = fsub fast float %mul1, %mul2
ret float %abx2
}
$ llc -o - badflags.ll -march=x86-64 -mattr=fma -enable-unsafe-fp-math -enable-fmf-dag=0
...
vmulss %xmm1, %xmm0, %xmm0
vaddss %xmm0, %xmm0, %xmm0
retq
$ llc -o - badflags.ll -march=x86-64 -mattr=fma -enable-unsafe-fp-math -enable-fmf-dag=1
...
vmulss %xmm1, %xmm0, %xmm2
vfmadd213ss %xmm2, %xmm1, %xmm0 <--- failed to recognize that (a * b) was already calculated
retq
llvm-svn: 244053
cl::init(0));
static cl::opt<bool>
-EnableFMFInDAG("enable-fmf-dag", cl::init(true), cl::Hidden,
+EnableFMFInDAG("enable-fmf-dag", cl::init(false), cl::Hidden,
cl::desc("Enable fast-math-flags for DAG nodes"));
// Limit the width of DAG chains. This is important in general to prevent