On Sat, Sep 24, 2011 at 2:15 PM, Jakub Jelinek <[email protected]> wrote:
> On Sat, Sep 24, 2011 at 01:26:36PM +0200, Richard Guenther wrote:
>> > +int
>> > +f3 (S &__restrict x, S &__restrict y)
>> > +{
>> > + x.p[0] = 3;
>> > + y.p[0] = 0;
>> > +// { dg-final { scan-tree-dump-times "return 3" 1 "optimized" } }
>> > + return x.p[0];
>> > +}
>> > +
>> > +int
>> > +f4 (S &x, S &y)
>> > +{
>> > + x.p[0] = 4;
>> > + y.p[0] = 0;
>> > +// { dg-final { scan-tree-dump-times "return 4" 0 "optimized" } }
>> > + return x.p[0];
>> > +}
>
>> I don't see why
>>
>> f4 (s, s)
>>
>> would be invalid. But you would miscompile it.
>
> f3 (s, s) you mean? I believe it is invalid. For f4 it would be valid
> and not optimized out.
Ah, I misread the dump-test.
>> (I'm not sure that a restrict qualified component is properly defined
>> by the C standard - we're just making this extension in a very constrained
>> case to allow Fortran array descriptors to work).
>
> Well, C standard doesn't have references, and C++ doesn't have restrict.
> So it is all about extensions.
> But what else would be & __restrict for than similar to *__restrict
> to say that the pointed (resp. referenced) object must not be accessed
> through other means than the reference or references/pointers derived from
> it, in the spirit of ISO C99 6.7.3.1.
> So, before jumping to __restrict fields, consider
> int a[10], b[10];
> int *
> f8 (S &__restrict x, S &__restrict y)
> {
> x.p = a;
> y.p = b;
> return x.p;
> }
> which we already optimize even before the patch.
> It is certainly invalid to call f8 (s, s).
>
> And the restricted fields, it is a straightforward extension to the restrict
> definition of ISO C99. We don't use it just for Fortran descriptors, but
> e.g. std::valarray uses __restrict fields too and has that backed up by the
> C++ standard requirements. Two different std::valarray objects will have
> different pointers inside of the structure.
>
> My intent currently is to be able to vectorize:
> #include <valarray>
>
> std::valarray<int>
> f9 (std::valarray<int> a, std::valarray<int> b, std::valarray<int> c, int z)
> {
> int i;
> for (i = 0; i < z; i++)
> {
> a[i] = b[i] + c[i];
> a[i] += b[i] * c[i];
> }
> return a;
> }
>
> void
> f10 (std::valarray<int> &__restrict a, std::valarray<int> &__restrict b,
> std::valarray<int> &__restrict c, int z)
> {
> int i;
> for (i = 0; i < z; i++)
> {
> a[i] = b[i] + c[i];
> a[i] += b[i] * c[i];
> }
> }
>
> In f9 we currently handle it differently than in f10, while IMHO it should
> be the same thing, a is guaranteed in both cases not to alias b nor c and b
> is guaranteed not to alias c, furthermore, a._M_data[0] through a._M_data[z-1]
> is guaranteed not to alias b._M_data[0] through b._M_data[z-1] and
> c._M_data[0]
> through c._M_data[z-1] and similarly for b vs. c. The __restrict on the
> _M_data field in std::valarray is a hint that different std::valarray
> objects will have different pointers.
Ok, I'm just worried that people get bitten by this (given the two existing
wrong-code issues we still have with restrict tracking and inlining).
But yes, your patch looks safe to me. Maybe we can document how GCC
treats restrict qualification of structure members?
>
> In f9 we have:
> size_tD.1850 D.53593;
> intD.9 * restrict D.53592;
> intD.9 & D.53591;
> ...
> D.53592_7 = MEM[(struct valarrayD.50086 *)aD.50087_6(D) + 8B];
> D.53593_42 = D.53456_5 * 4;
> # PT = nonlocal escaped { D.53660 } (restr)
> D.53591_43 = D.53592_7 + D.53593_42;
> ...
> *D.53591_43 = D.53462_19;
> and while PTA computes the restricted property here, we unfortunately still
> don't use it, because D.53591 (which comes from all the inlined wrappers)
> isn't TYPE_RESTRICT.
Yeah, that's extra safety checks because we ignore nonlocal/escaped when
applying the restrict tag match. Otherwise the restrict tags are prone to leak
into other variables solutions. Maybe finally fixing that two
wrong-code restrict
bugs will solve this issue though.
> Shouldn't we propagate that property to either
> SSA_NAMEs initialized from restricted pointers resp. POINTER_PLUS_EXPRs,
> or if it is common to all VAR_DECLs underlying such SSA_NAMEs, to the
> VAR_DECLs?
I'm not sure where to best do that. In principle we shouldn't need to check
TYPE_RESTRICT at all, but that requires some thoughts.
> But in f10 we don't get even that far, the a._M_data (which is actually
> a->_M_data, since a is a (restricted) reference) load is already itself
> not considered as restricted by PTA.
>
> It is nice that we optimize Fortran arrays well, but it would be nice if we
> did the same for C++ too.
Yeah, I agree.
Your patch is ok.
Thanks,
Richard.
> Jakub
>
