union *func()

[Chris Torek]

... and union func():

>In article <14172 at mimsy.UUCP> I noted that
>>	a = function().member;
>>
>>is conforming, etc., if function() returns a structure or union and the
>>`member' field corresponds, etc., with one additional restriction:
>>the member must not be an array.

In article <1585 at solo8.cs.vu.nl> maart at cs.vu.nl (Maarten Litmaath) asks
why the restriction exists.  `function().member' above is used in an
rvalue context; as such, if the member is an array, this amounts to
taking the address of part of a structure or union valued function.
If, as is common in some compilers, the return value is in fact copied
into local temporary storage, it may not have an address.  In
particular, consider the following:

	struct arr {
		int arr[10];
	};
	struct arr fn();

	main() {
		int *p;

		p = &fn().arr[8];
		printf("fn 3,4 = %d,%d; p = %p\n", p[0], p[1], (void *)p);
	}

This might `want' to compile into something like this:

	main_:
		| {
		mov	fp,-(sp)	| make stack frame
		mov	sp,fp

		| int *p;
		sub	#2,sp		| make stack space for local `p'

		| fn()
		sub	#20,sp		| make room
		lea	(sp),r0		| address of return space
		call	fn_		| fn()
		lea	(sp),r0		| compute address of return value
		| & .arr[8]
		lea	16(r0),r0
		| p =
		mov	r0,-2(fp)	| store in p
		| ;
		add	#20,sp		| done with return value

		| p
		mov	-2(fp),r0	| noop, hope optimiser pulls it out
		mov	r0,-(sp)	| push p
		| p[1]
		mov	-2(fp),r0	| noop, hope optimiser pulls it out
		mov	2(r0),-(sp)	| push p[1]
		| p[0]
		mov	(r0),-(sp)	| push p[0]
		| "string"
		pea	S1		| push address of string
		| printf(...)
		call	printf_		| printf the various variables
		add	#8,sp		| fix stack

		| }
		mov	fp,sp		| undo frame
		mov	(sp)+,fp
		ret

There is a problem here: pushing p clobbered p[1]!

Anyway, it is legislated out, as is such stuff as

	struct arr fn();
	main() {
		fn().arr[2] = 3;
	}

(At least, I *hope* stuff like the latter is illegal, given that the
point of the original restriction above is to allow structure return
values to be done using invisible temporaries....)
-- 
In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 454 7163)
Domain:	chris at mimsy.umd.edu	Path:	uunet!mimsy!chris