[llvm] [GISel] Add KnownFPClass Analysis to GISelValueTrackingPass (PR #134611)

[Tim Gymnich via llvm-commits]

================
@@ -631,12 +655,1448 @@ void GISelValueTracking::computeKnownBitsImpl(Register R, KnownBits &Known,
   }
   }
 
-  LLVM_DEBUG(dumpResult(MI, Known, Depth));
+  LLVM_DEBUG(dumpKnownBitsResult(MI, Known, Depth));
 
   // Update the cache.
   ComputeKnownBitsCache[R] = Known;
 }
 
+/// Return true if it's possible to assume IEEE treatment of input denormals in
+/// \p MF for \p Ty.
+static bool inputDenormalIsIEEE(const MachineFunction &MF, LLT Ty) {
+  Ty = Ty.getScalarType();
+  return MF.getDenormalMode(getFltSemanticForLLT(Ty)).Input ==
+         DenormalMode::IEEE;
+}
+
+static bool outputDenormalIsIEEEOrPosZero(const MachineFunction &MF, LLT Ty) {
+  Ty = Ty.getScalarType();
+  DenormalMode Mode = MF.getDenormalMode(getFltSemanticForLLT(Ty));
+  return Mode.Output == DenormalMode::IEEE ||
+         Mode.Output == DenormalMode::PositiveZero;
+}
+
+std::pair<Register, FPClassTest> GISelValueTracking::fcmpToClassTest(
+    FCmpInst::Predicate Pred, const MachineFunction &MF, Register LHS,
+    const APFloat *ConstRHS, bool LookThroughSrc) {
+
+  auto [Src, ClassIfTrue, ClassIfFalse] =
+      fcmpImpliesClass(Pred, MF, LHS, *ConstRHS, LookThroughSrc);
+  if (Src && ClassIfTrue == ~ClassIfFalse)
+    return {Src, ClassIfTrue};
+
+  return {Register(), fcAllFlags};
+}
+
+/// Return the return value for fcmpImpliesClass for a compare that produces an
+/// exact class test.
+static std::tuple<Register, FPClassTest, FPClassTest>
+exactClass(Register V, FPClassTest M) {
+  return {V, M, ~M};
+}
+
+std::tuple<Register, FPClassTest, FPClassTest>
+GISelValueTracking::fcmpImpliesClass(CmpInst::Predicate Pred,
----------------
tgymnich wrote:

Here is my attempt of templating this: https://github.com/tgymnich/llvm-project/blob/tim/gisel-value-tracking/llvm/include/llvm/ADT/FloatingPointModeUtils.h

In the end I wasn't quite happy with how I handled the LookThrough.
- Just passing a lambda is not extensible enoguh imo. => e.g. what happens if we want to look through to more than just fabs.
- `if constexpr` is also not to nice since we'd need to inlcude all the MIR and IR headers.


https://github.com/llvm/llvm-project/pull/134611