On Tue, Apr 10, 2018 at 5:28 PM, Bin.Cheng <amker.ch...@gmail.com> wrote:
> On Tue, Apr 10, 2018 at 3:58 PM, Bin.Cheng <amker.ch...@gmail.com> wrote:
>> On Tue, Apr 10, 2018 at 2:26 PM, Jakub Jelinek <ja...@redhat.com> wrote:
>>> On Tue, Apr 10, 2018 at 09:55:35AM +0000, Bin Cheng wrote:
>>>> Hi Rainer, could you please help me double check that this solves the
>>>> issue?
>>>>
>>>> Thanks,
>>>> bin
>>>>
>>>> gcc/testsuite
>>>> 2018-04-10 Bin Cheng <bin.ch...@arm.com>
>>>>
>>>> PR testsuite/85190
>>>> * gcc.dg/vect/pr81196.c: Adjust pointer for aligned access.
>>>
>>>> diff --git a/gcc/testsuite/gcc.dg/vect/pr81196.c
>>>> b/gcc/testsuite/gcc.dg/vect/pr81196.c
>>>> index 46d7a9e..15320ae 100644
>>>> --- a/gcc/testsuite/gcc.dg/vect/pr81196.c
>>>> +++ b/gcc/testsuite/gcc.dg/vect/pr81196.c
>>>> @@ -4,14 +4,14 @@
>>>>
>>>> void f(short*p){
>>>> p=(short*)__builtin_assume_aligned(p,64);
>>>> - short*q=p+256;
>>>> + short*q=p+255;
>>>> for(;p!=q;++p,--q){
>>>> short t=*p;*p=*q;*q=t;
>>>
>>> This is UB then though, because p will never be equal to q.
>
> Hmm, though it's UB in this case, is it OK for niter analysis gives
> below results?
>
> Analyzing # of iterations of loop 1
> exit condition [126, + , 18446744073709551615] != 0
> bounds on difference of bases: -126 ... -126
> result:
> # of iterations 126, bounded by 126
>
> I don't really follow last piece of code in number_of_iterations_ne:
>
> /* Let nsd (step, size of mode) = d. If d does not divide c, the loop
> is infinite. Otherwise, the number of iterations is
> (inverse(s/d) * (c/d)) mod (size of mode/d). */
> bits = num_ending_zeros (s);
> bound = build_low_bits_mask (niter_type,
> (TYPE_PRECISION (niter_type)
> - tree_to_uhwi (bits)));
>
> d = fold_binary_to_constant (LSHIFT_EXPR, niter_type,
> build_int_cst (niter_type, 1), bits);
> s = fold_binary_to_constant (RSHIFT_EXPR, niter_type, s, bits);
>
> if (!exit_must_be_taken)
> {
> /* If we cannot assume that the exit is taken eventually, record the
> assumptions for divisibility of c. */
> assumption = fold_build2 (FLOOR_MOD_EXPR, niter_type, c, d);
> assumption = fold_build2 (EQ_EXPR, boolean_type_node,
> assumption, build_int_cst (niter_type, 0));
> if (!integer_nonzerop (assumption))
> niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
> niter->assumptions, assumption);
> }
>
> c = fold_build2 (EXACT_DIV_EXPR, niter_type, c, d);
> tmp = fold_build2 (MULT_EXPR, niter_type, c, inverse (s, bound));
> niter->niter = fold_build2 (BIT_AND_EXPR, niter_type, tmp, bound);
> return true;
>
> Though infinite niters is mentioned, I don't see it's handled?
So the default behavior we have for long time is to completely unroll
below loop:
void f(short*p){
p=(short*)__builtin_assume_aligned(p,64);
short*q=p+5;
for(;p!=q;++p,--q){
short t=*p;*p=*q;*q=t;
}
}
because:
Analyzing # of iterations of loop 1
exit condition [p_6, + , 4](no_overflow) != p_6 + 10
bounds on difference of bases: 10 ... 10
result:
# of iterations 2, bounded by 2
Thanks,
bin
>
> Thanks,
> bin
>> Sorry I already checked in, will try to correct it in another patch.
>>
>> Thanks,
>> bin
>>>
>>>> }
>>>> }
>>>> void b(short*p){
>>>> p=(short*)__builtin_assume_aligned(p,64);
>>>> - short*q=p+256;
>>>> + short*q=p+255;
>>>> for(;p<q;++p,--q){
>>>> short t=*p;*p=*q;*q=t;
>>>
>>> This one is fine, sure.
>>>
>>> Jakub
