Fix / workaround a couple of more 2 return value issues

Summary:
There's a couple of more spots where 2 return values are biting is:
1) JITRT_CallWithIncorrectArgcount - for now this is just working around things by making this return PyObject* on Windows. That means it won't work properly w/ static Python but gets everything else working.
2) LoadTypeMethodCache::lookupHelper needs the appendCall2RetValues treatment.

Reviewed By: alexmalyshev

Differential Revision: D103952453

fbshipit-source-id: 854f2d03d7a40c9eca909e2489712b6ca49e5d52

Author: DinoV

cinderx/Jit/jit_rt.cpp

@@ -366,7 +366,12 @@ JITRT_StaticCallFPReturn JITRT_CallWithIncorrectArgcountFPReturn(
       (PyObject*)defaulted_args);
 }
 
-JITRT_StaticCallReturn JITRT_CallWithIncorrectArgcount(
+#ifdef _WIN32
+PyObject*
+#else
+JITRT_StaticCallReturn
+#endif
+JITRT_CallWithIncorrectArgcount(
     PyFunctionObject* func,
     PyObject** args,
     size_t nargsf,
@@ -377,7 +382,11 @@ JITRT_StaticCallReturn JITRT_CallWithIncorrectArgcount(
     // Fallback to the default _PyFunction_Vectorcall implementation
     // to produce an appropriate exception.
     auto interpVectorcall = getInterpretedVectorcall(func);
+#ifdef _WIN32
+    return interpVectorcall((PyObject*)func, args, nargsf, nullptr);
+#else
     return {interpVectorcall((PyObject*)func, args, nargsf, nullptr), nullptr};
+#endif
   }
   Py_ssize_t defcount = PyTuple_GET_SIZE(defaults);
   Py_ssize_t nargs = PyVectorcall_NARGS(nargsf);
@@ -387,7 +396,11 @@ JITRT_StaticCallReturn JITRT_CallWithIncorrectArgcount(
   if (nargs + defcount < argcount || nargs > argcount) {
     // Not enough args with defaults, or too many args without defaults.
     auto interpVectorcall = getInterpretedVectorcall(func);
+#ifdef _WIN32
+    return interpVectorcall((PyObject*)func, args, nargsf, nullptr);
+#else
     return {interpVectorcall((PyObject*)func, args, nargsf, nullptr), nullptr};
+#endif
   }
 
   Py_ssize_t i;
@@ -403,6 +416,10 @@ JITRT_StaticCallReturn JITRT_CallWithIncorrectArgcount(
 
   size_t new_nargsf = argcount;
 
+#ifdef _WIN32
+  return JITRT_GET_REENTRY(func->vectorcall)(
+      (PyObject*)func, arg_space.get(), new_nargsf, (PyObject*)defaulted_args);
+#else
   return reinterpret_cast<staticvectorcallfunc>(
       JITRT_GET_REENTRY(func->vectorcall))(
       (PyObject*)func,
@@ -411,6 +428,7 @@ JITRT_StaticCallReturn JITRT_CallWithIncorrectArgcount(
       // We lie to C++ here, and smuggle in the number of defaulted args filled
       // in.
       (PyObject*)defaulted_args);
+#endif
 }
 
 bool JITRT_PackStaticArgs(

cinderx/Jit/jit_rt.h

@@ -96,7 +96,15 @@ PyObject* JITRT_CallWithKeywordArgsSimple(
     size_t nargsf,
     PyObject* kwnames);
 
-JITRT_StaticCallReturn JITRT_CallWithIncorrectArgcount(
+// On Windows x64, returning JITRT_StaticCallReturn (16 bytes) would use a
+// hidden first parameter for the return pointer, shifting all visible
+// arguments. Return PyObject* instead to keep register assignments correct.
+#ifdef _WIN32
+PyObject*
+#else
+JITRT_StaticCallReturn
+#endif
+JITRT_CallWithIncorrectArgcount(
     PyFunctionObject* func,
     PyObject** args,
     size_t nargsf,

cinderx/Jit/lir/generator.cpp

@@ -767,6 +767,7 @@ void GeneratePrimitiveArgsPrologueBlock(
   auto helper = returns_primitive_double
       ? reinterpret_cast<uint64_t>(JITRT_CallStaticallyWithPrimitiveSignatureFP)
       : reinterpret_cast<uint64_t>(JITRT_CallStaticallyWithPrimitiveSignature);
+
   block->allocateInstr(Instruction::kCall, nullptr, Imm{helper});
 
   // The helper either handled the call (result in return register) and we
@@ -2503,7 +2504,8 @@ LIRGenerator::TranslatedBlock LIRGenerator::TranslateOneBasicBlock(
               PyUnicode_AsUTF8(code->co_filename),
               PyUnicode_AsUTF8(code->co_name));
         }
-        bbb.appendCallInstruction(
+        appendCall2RetValues(
+            bbb,
             instr->output(),
             LoadTypeMethodCache::lookupHelper,
             cache_entry,