llvm

bc8d3dfd - [FPBuiltinFnSelection] Fold fdiv-by-exact-reciprocal-constant to fmul (#22948)

[FPBuiltinFnSelection] Fold fdiv-by-exact-reciprocal-constant to fmul (#22948) llvm.fpbuiltin.fdiv hides the division from InstCombine's foldFDivConstantDivisor until this pass lowers it to a real fdiv late in the pipeline, so InstCombine never gets a second look. Replicate the exact-reciprocal case of that fold here so X / C becomes X * (1/C) whenever C has an exact FP inverse. JIRA: CMPLRLLVM-77622