[llvm] [Analysis] Adding convert_from_arbitrary_fp support in computeKnownFPClass. (PR #208585)

[Matt Arsenault via llvm-commits]

================
@@ -5727,6 +5727,44 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
       Known.knownNot(fcNan | fcInf | fcNegative);
       break;
     }
+    case Intrinsic::convert_from_arbitrary_fp: {
+      auto *MD = cast<MetadataAsValue>(II->getArgOperand(1))->getMetadata();
+      StringRef FormatStr = cast<MDString>(MD)->getString();
+
+      const fltSemantics *SrcSemanticsPtr =
+          APFloat::getArbitraryFPSemantics(FormatStr);
+      // TODO: the unsupported formats should be handle way before this
+      // remove once the verifier is updated.
+      if (!SrcSemanticsPtr)
+        break;
+
+      const fltSemantics SrcSemantics = *SrcSemanticsPtr;
+      const fltSemantics DstSemantics =
+          II->getType()->getScalarType()->getFltSemantics();
+
+      if (!APFloat::semanticsHasNaN(SrcSemantics))
+        Known.knownNot(fcNan);
+      if (!APFloat::semanticsHasInf(SrcSemantics) &&
+          APFloat::isRepresentableBy(SrcSemantics, DstSemantics))
+        Known.knownNot(fcInf);
+
+      if (SrcSemantics.nanEncoding == fltNanEncoding::NegativeZero)
+        Known.knownNot(fcNegZero);
+
+      if (!APFloat::semanticsHasZero(SrcSemantics)) {
+        Known.knownNot(fcZero);
----------------
arsenm wrote:

This probably works in practice. The one possible edge case is if the source value would underflow in the destination format, but I think we'd have to support smaller result types in the IR or larger arbitrary input formats for it to matter 

https://github.com/llvm/llvm-project/pull/208585