[CIR] Enable floating point casts (#144142)

We already had the code in place to emit CIR floating point cast ops
that get lowered to fpext or fptrunc, but we weren't calling the
function to emit that cast from ScalarExprEmitter::emitScalarCast. This
change adds that call.

Author: andykaylor

clang/lib/CIR/CodeGen/CIRGenExprScalar.cpp

@@ -346,8 +346,8 @@ class ScalarExprEmitter : public StmtVisitor<ScalarExprEmitter, mlir::Value> {
         assert(!cir::MissingFeatures::fpConstraints());
         castKind = cir::CastKind::float_to_int;
       } else if (mlir::isa<cir::CIRFPTypeInterface>(dstTy)) {
-        cgf.getCIRGenModule().errorNYI("floating point casts");
-        return cgf.createDummyValue(src.getLoc(), dstType);
+        // TODO: split this to createFPExt/createFPTrunc
+        return builder.createFloatingCast(src, fullDstTy);
       } else {
         llvm_unreachable("Internal error: Cast to unexpected type");
       }

clang/test/CIR/CodeGen/cast.cpp

@@ -73,6 +73,14 @@ int cStyleCasts_0(unsigned x1, int x2, float x3, short x4, double x5) {
   // LLVM: %{{[0-9]+}} = fcmp une float %{{[0-9]+}}, 0.000000e+00
   // LLVM: %{{[0-9]+}} = zext i1 %{{[0-9]+}} to i8
 
+  double d2 = f; // float to double
+  // CIR: %{{[0-9]+}} = cir.cast(floating, %{{[0-9]+}} : !cir.float), !cir.double
+  // LLVM: %{{[0-9]+}} = fpext float %{{[0-9]+}} to double
+
+  f = d2; // double to float
+  // CIR: %{{[0-9]+}} = cir.cast(floating, %{{[0-9]+}} : !cir.double), !cir.float
+  // LLVM: %{{[0-9]+}} = fptrunc double %{{[0-9]+}} to float
+
   return 0;
 }