[llvm] [KnownFPClass] Correct denormal handling for `KnownFPClass::roundToIntegral` (PR #219700)

[Eric Ross via llvm-commits]

================
@@ -467,10 +467,12 @@ struct KnownFPClass {
 
   /// Propagate known class for rounding intrinsics (trunc, floor, ceil, rint,
   /// nearbyint, round, roundeven). This is trunc if \p IsTrunc. \p
-  /// IsMultiUnitFPType if this is for a multi-unit floating-point type.
-  LLVM_ABI static KnownFPClass roundToIntegral(const KnownFPClass &Src,
-                                               bool IsTrunc,
-                                               bool IsMultiUnitFPType);
+  /// IsKnownNeverMultiUnitFPType if this is known to never be for a multi-unit
+  /// floating-point type.
+  LLVM_ABI static KnownFPClass
+  roundToIntegral(const KnownFPClass &Src, bool IsTrunc,
+                  bool IsKnownNeverMultiUnitFPType,
----------------
ZERICO2005 wrote:

My thinking was:
- `IsMultiUnitFPType` implies that additional deductions are possible for `ppcf128`.
- `IsKnownNeverMultiUnitFPType` implies that additional deductions are possible for types that are **not** `ppcf128`. Which I think better describes why we need this parameter for `KnownFPClass::roundToIntegeral`.

I was also planning on doing something similar for `KnownFPClass::log` to fix some incorrect deductions specifically for the `ppcf128` type.

An alternative would be to pass the `FltSemantics` into the function.



https://github.com/llvm/llvm-project/pull/219700