On Mar 22, 2007, at 3:11 PM, der Mouse wrote:

>
>>
>>>
>>>> [...volatile...]
>>
>>> No, the compiler is not allowed to keep cached in a register a
>>
>>> memory access across a function call, precisely because it cannot
>>
>>> know if that function call modifies the memory.
>> don't callee vs. caller saved regs give the compiler a way around
>> this?
>
> Not really.  The point is, if you write, for example,
>
>    x = ...something...;
>    somefunction();
>    y = x;
>
> then, in terms of the C abstract machine, the compiler must re-fetch x
> when assigning to y rather than re-using the value from the former
> assignment (which it may well have lying around in a register, whether
> caller-saved, callee-saved, hardware-saved, or what, is irrelevant).
> This is because, in terms of the C abstract machine, somefunction()
> could modify x.

Uh, no, I don't think so.  In the case of:

void
foo(int *addr)
{
	int x, y;

	x = *addr;
	somefunction();
	y = x;

	...
}

"x" is not memory, "*addr" is memory.  So, it WOULD have to refetch if:

void
foo(int *addr)
{
	int x, y;

	x = *addr;
	somefunction();
	y = *addr;

	...
}

i.e. the assignment of y could NOT be optimized into "y = x;"

-- thorpej