[llvm] [ConstantFolding] Add constant folding support for nextafter/nexttoward (PR #167324)

[Sayan Sivakumaran via llvm-commits]

================
@@ -3221,6 +3223,26 @@ static Constant *ConstantFoldLibCall2(StringRef Name, Type *Ty,
     if (TLI->has(Func))
       return ConstantFoldBinaryFP(atan2, Op1V, Op2V, Ty);
     break;
+  case LibFunc_nextafter:
+  case LibFunc_nextafterf:
+  case LibFunc_nexttoward:
+  case LibFunc_nexttowardf:
+    if (TLI->has(Func)) {
+      if (Op1V.isNaN() || Op2V.isNaN()) {
+        return ConstantFP::get(Ty->getContext(),
+                               APFloat::getNaN(Ty->getFltSemantics()));
+      }
+
+      APFloat PromotedOp1V = Op1V.getPromoted(APFloat::IEEEquad());
+      APFloat PromotedOp2V = Op2V.getPromoted(APFloat::IEEEquad());
+      if (PromotedOp1V == PromotedOp2V) {
+        return ConstantFP::get(Ty->getContext(), Op1V);
+      }
+
+      APFloat Next(Op1V);
+      Next.next(/*nextDown=*/PromotedOp1V > PromotedOp2V);
+      return ConstantFP::get(Ty->getContext(), Next);
----------------
sivakusayan wrote:

We need to promote here before doing any of these checks, as otherwise the operands have different types in the case of `nexttoward`. I believe that converting to `IEEEQuad` is strictly widening no matter what the argument types are, and the assertions seem to agree with that.

https://github.com/llvm/llvm-project/pull/167324