On 5/20/19 2:04 AM, Richard Biener wrote:
> On Sun, May 19, 2019 at 6:16 PM Marc Glisse <marc.gli...@inria.fr> wrote:
>>
>> Hello,
>>
>> 2 pieces:
>>
>> - the first one handles the case where the denominator is negative. It
>> doesn't happen often with exact_div, so I don't handle it everywhere, but
>> this one looked trivial
>>
>> - handle the case where a pointer difference is cast to an unsigned type
>> before being compared to a constant (I hit this in std::vector). With some
>> range info we could probably handle some non-constant cases as well...
>>
>> The second piece breaks Walloca-13.c (-Walloca-larger-than=100 -O2)
>>
>> void f (void*);
>> void g (int *p, int *q)
>> {
>> __SIZE_TYPE__ n = (__SIZE_TYPE__)(p - q);
>> if (n < 100)
>> f (__builtin_alloca (n));
>> }
>>
>> At the time of walloca2, we have
>>
>> _1 = p_5(D) - q_6(D);
>> # RANGE [-2305843009213693952, 2305843009213693951]
>> _2 = _1 /[ex] 4;
>> # RANGE ~[2305843009213693952, 16140901064495857663]
>> n_7 = (long unsigned intD.10) _2;
>> _11 = (long unsigned intD.10) _1;
>> if (_11 <= 396)
>> [...]
>> _3 = allocaD.1059 (n_7);
>>
>> and warn.
>
> That's indeed to complicated relation of _11 to n_7 for
> VRP predicate discovery.
>
>> However, DOM3 later produces
>>
>> _1 = p_5(D) - q_6(D);
>> _11 = (long unsigned intD.10) _1;
>> if (_11 <= 396)
>
> while _11 vs. _1 works fine.
>
>> [...]
>> # RANGE [0, 99] NONZERO 127
>> _2 = _1 /[ex] 4;
>> # RANGE [0, 99] NONZERO 127
>> n_7 = (long unsigned intD.10) _2;
>> _3 = allocaD.1059 (n_7);
>>
>> so I am tempted to say that the walloca2 pass is too early, xfail the
>> testcase and file an issue...
>
> Hmm, there's a DOM pass before walloca2 already and moving
> walloca2 after loop opts doesn't look like the best thing to do?
> I suppose it's not DOM but sinking that does the important transform
> here? That is,
It could well be sinking that's critical here. It certainly was in
cases I looked at in our <vec> code and how we needed to transform the
pointer subtraction and subsequent testing to discover infeasible paths.
Jeff
