Re: [PATCH,TESTSUITE]Reduce malloc size used by ipa-sra-2.c

[Andrew Hutchinson <andrewhutchinson at cox dot net>]

You assumptions about sizes are correct.

int is 16 bit
pointers are 16bit
long are 32 bit



Richard Guenther wrote:
> On Sun, Nov 1, 2009 at 12:37 PM, Andrew Hutchinson
> <andrewhutchinson@cox.net> wrote:
>   
> > That's an interesting point !
> >
> > AVR does not reject this. It never gets to see it. Error is:
> >
> > error: size of array 'data' is too large
> >
> > which appear to come from c-decl.c or perhaps c-common.c
> >
> > The former has check for array size (in bytes) overflowing integer max+1.
> > Which for AVR would be 32768 bytes. (this actually raises another issue
> > where 'unsigned' size of  65536 bytes should be limit)
> >     
>
> Hmm.  I think the constant is parsed as a long (which is 32bits on avr?),
> but the C frontend uses c_common_signed_type (sizetype) as canonical
> index type here which is only 16 bits (you have 16bits sizetype I guess).
>
> I suspect that choosing such a small sizetype is the problem, but the
> C frontend rejecting the program is probably non-conforming as well
> (it could warn and put error_mark_node in the size or so).
>
> Maybe Joseph can clarify this.
>
> Richard.
>
>   
> > Andy
> >
> >
> > Richard Guenther wrote:
> >     
> > > On Sun, Nov 1, 2009 at 11:40 AM, Andrew Hutchinson
> > > <andrewhutchinson@cox.net> wrote:
> > >
> > >       
> > > > Its the array size:
> > > >
> > > > int data[1000000]
> > > >
> > > > Yes, I could equally skip this.
> > > >
> > > >         
> > > But the array is not allocated - does AVR really reject a program just
> > > because there appears a possibly unused large type?
> > >
> > > Richard.
> > >
> > >
> > >       
> > > > Richard Guenther wrote:
> > > >
> > > >         
> > > > > On Sat, Oct 31, 2009 at 8:27 PM, Andrew Hutchinson
> > > > > <andrewhutchinson@cox.net> wrote:
> > > > >
> > > > >
> > > > >           
> > > > > > Thanks - patch was not as tested. Correction below.
> > > > > >
> > > > > > This patch corrects a testcase where malloc size 1,000,000 is way too
> > > > > > big
> > > > > > for target.
> > > > > > Size requested is reduce if stack size is below 16000.
> > > > > >
> > > > > > The patch removes failures for AVR target.
> > > > > > Ok to commit?
> > > > > >
> > > > > >
> > > > > >             
> > > > > I think for this kind of tests you should simply skip AVR.  Note that it
> > > > > mallocs sizeof (struct small), which is small.  So I don't see why
> > > > > you get a failure here anyway.
> > > > >
> > > > > Richard.
> > > > >
> > > > >
> > > > >
> > > > >           
> > > > > > 2009-10-31  Andy Hutchinson <hutchinsonandy@gcc.gnu.org>
> > > > > >
> > > > > >  *gcc.c-torture/execute/ipa-sra-2.c: Reduce allocation if stack is
> > > > > > small.
> > > > > >
> > > > > >
> > > > > >
> > > > > > Index: ipa-sra-2.c
> > > > > > ===================================================================
> > > > > > --- ipa-sra-2.c    (revision 153773)
> > > > > > +++ ipa-sra-2.c    (working copy)
> > > > > > @@ -1,6 +1,11 @@
> > > > > > +#if defined(STACK_SIZE) && STACK_SIZE < 16000
> > > > > > +#define ARRAY_SIZE (STACK_SIZE)
> > > > > > +#else
> > > > > > +#define ARRAY_SIZE 1000000
> > > > > > +#endif
> > > > > > struct big
> > > > > > {
> > > > > > -  int data[1000000];
> > > > > > +  int data[ARRAY_SIZE];
> > > > > > };
> > > > > >
> > > > > > struct small
> > > > > > @@ -23,7 +28,7 @@
> > > > > > {
> > > > > >  int r;
> > > > > >  if (fail)
> > > > > > -    r = agg->big.data[999999];
> > > > > > +    r = agg->big.data[ARRAY_SIZE - 1];
> > > > > >  else
> > > > > >  r = agg->small.data[0];
> > > > > >  return r;
> > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > >             
> > > > >           
> > >       
>
>