Re: [PATCH] Fix do_div() for all architectures

[Richard Henderson]

On Thu, Jul 10, 2003 at 06:18:59PM +0200, Andrea Arcangeli wrote:
> On Thu, Jul 10, 2003 at 08:40:19AM -0700, Richard Henderson wrote:
> > On Tue, Jul 08, 2003 at 08:27:26PM +0200, Bernardo Innocenti wrote:
> > > +extern uint32_t __div64_32(uint64_t *dividend, uint32_t divisor)
> > > __attribute_pure__;
> > ...
> > > +		__rem = __div64_32(&(n), __base);	\
> > 
> > The pure declaration is very incorrect.  You're writing to N.
> 
> now pure sounds more reasonable, I wondered how could gcc keep track of
> the stuff pointed by the parameters (especially if this stuff points to
> other stuff etc.. ;).

Bernardo mis-interpreted the documentation.

Define "local memory" as memory from the current stack frame.

Define "non-local memory" as anything else (including stack memory from
another function, or a different instantiation of the current function).

Any function can read/write local memory (since that is not visible to
anyone outside the function).

A "const" function cannot read or write to non-local memory.  There are
further constraints on not returning abnormally or not returning at all
that I'll not go into now.

A "pure" function can read non-local memory, but cannot write to it.

We use those conditions to determine if two invocations of a function
can be collapsed or moved or removed.

r~