[clang] [Clang] Add float type support to __builtin_reduce_add and __builtin_reduce_multipy (PR #120367)

[Csanád Hajdú via cfe-commits]

================
@@ -4274,12 +4274,37 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
         *this, E, GetIntrinsicID(E->getArg(0)->getType()), "rdx.min"));
   }
 
-  case Builtin::BI__builtin_reduce_add:
+  case Builtin::BI__builtin_reduce_add: {
+    // Note: vector_reduce_fadd takes two arguments a
+    // scalar start value and a vector. That would mean to
+    // correctly call it we would need emitBuiltinWithOneOverloadedType<2>
+    // To keep the  builtin sema behavior the same despite type we will
+    // popululate vector_reduce_fadd scalar value with a 0.
+    if (E->getArg(0)->getType()->hasFloatingRepresentation()) {
+      Value *X = EmitScalarExpr(E->getArg(0));
+      auto EltTy = X->getType()->getScalarType();
+      Value *Seed = ConstantFP::get(EltTy, 0);
----------------
Il-Capitano wrote:

Use `-0.0` as the initial value, as that's the neutral additive float, i.e. `-0.0 + x == x`. This can have an effect on code generation, see for example https://github.com/rust-lang/rust/issues/130028.

https://github.com/llvm/llvm-project/pull/120367