[runtime] Fix numerous marshalling bugs in the dynamic registrar. (#22209)

For arm64:

* We didn't take into account that a register can contain more than one field of a struct (for fields where the total size is less than or equal to 8 bytes). Example: a struct with 8 separate bytes fields will be passed in a single register, instead
  of using 8 registers.
* We didn't take alignment / padding of elements in structs into account.
* We didn't correctly detect which structs can be fully passed in floating point registers.

For x64:

* We didn't take alignment / padding of elements in structs into account.
* We didn't correctly detect which structs can be fully passed in floating point registers.
* We didn't correctly detect when we can't pass any more values in registers.
* We didn't correctly marshal structs with non-standard small sizes (anything between 1 and 16, inclusive).

Fixes #22196.

Author: rolfbjarne

runtime/runtime-internal.h

@@ -44,6 +44,32 @@ void xamarin_dyn_objc_msgSend_stret ();
 void xamarin_dyn_objc_msgSendSuper_stret ();
 void xamarin_add_internal_call (const char *name, const void *method);
 
+static inline size_t align_size (size_t target, uint64_t alignment)
+{
+	target = (target + (alignment - 1)) & ~(alignment - 1);
+	return target;
+}
+
+static inline unsigned long align_ulong (unsigned long target, unsigned long alignment)
+{
+	target = (target + (alignment - 1)) & ~(alignment - 1);
+	return target;
+}
+
+static inline int32_t align_int32 (int32_t target, int32_t alignment)
+{
+	target = (target + (alignment - 1)) & ~(alignment - 1);
+	return target;
+}
+
+static inline void* align_ptr (void* target, uint64_t alignment)
+{
+	uint64_t ptr = (uint64_t) target;
+	ptr = (ptr + (alignment - 1)) & ~(alignment - 1);
+	return (void *) ptr;
+}
+
+
 #ifdef __cplusplus
 }
 #endif

runtime/runtime.m

@@ -1677,7 +1677,10 @@ -(void) xamarinSetFlags: (enum XamarinGCHandleFlags) flags;
 					xamarin_assertion_message ("Unsupported struct type: %s", original_type);
 				unsigned long item_size = xamarin_objc_type_size (type);
 
-				size += (item_size + (sizeof (void *) - 1)) & ~((sizeof (void *) - 1));
+				// first align the accumulated size to the current item's size (this would be any padding between struct elements)
+				size = align_ulong (size, item_size);
+				// accumulate this item's size
+				size += item_size;
 
 				type = objc_skip_type (type);
 			}

runtime/trampolines-arm64.h

@@ -9,11 +9,18 @@ struct BigDouble {
 	struct _f {
 		float f1;
 		float f2;
+		float f3;
+		float f4;
+	};
+	struct _d {
+		double d1;
+		double d2;
 	};
 	union {
 		__int128 bits;
-		long double d;
+		long double q;
 		_f f;
+		_d d;
 	};
 };
 
@@ -52,6 +59,7 @@ struct ParamIterator {
 	struct XamarinCallState *state;
 	int ngrn; // Next General-purpose Register Number
 	int nsrn; // Next SIMD and Floating-point Register Number
+	size_t ngrn_offset; // Offset into the current general purpose register number
 	uint8_t *nsaa; // Next stacked argument address.
 
 	// computed values