LLVM 21 moved the KnownFPClass definition from `ValueTracking.h` to its own file. I verified that specific version moved it, with 20 being the last without it.
Replaced the denormal/zero check in `fma32x4` with LLVM's `KnownFPClass` analysis.
That method now includes a overload which expects the caller to provide the known struct. This is used for memorization with the FMA instructions, where the range clamping and instruction specific special value handling are combined. `FMS` and `FNMS` has a larger search depth as it can potentially shorten the chain by a lot.
Replace constant checks in `FM`, `FCEQ`, and `FCMEQ` with `KnownFPClass`. The equality compares now also optimize for when the operands is infinity, which is safe as it behaves identically to the default path.
Avoid clamping if the value is known not to be Inf/NaN (overflow). Includes a optional precalculated argument, which only calculates the known if it has not been done already.
Optimize the float multiply and compare methods by taking advantage of the `denormals-are-zero` flag, removing duplicated logic, and generally doing things more efficiently.
Refactors SELB by changing "division accuracy correction" pattern match to reduce nesting, include more comments, and use more descriptive variable names. Also replaced duplicated granularity-check loops with single loop which determines byte granularity as a integer.
Elad
Jordan Jones
Ani
Walter
yahfz
Megamouse