Re: [PATCH GCC]Improve how we handle overflow for type conversion in scev/ivopts, part I

["Bin.Cheng" <amker dot cheng at gmail dot com>]

On Tue, Jun 2, 2015 at 11:37 AM, Bin.Cheng <amker.cheng@gmail.com> wrote:
> On Tue, May 26, 2015 at 5:04 PM, Richard Biener
> <richard.guenther@gmail.com> wrote:
>> On Sun, May 24, 2015 at 8:47 AM, Bin.Cheng <amker.cheng@gmail.com> wrote:
>>> On Fri, May 22, 2015 at 7:45 PM, Richard Biener
>>> <richard.guenther@gmail.com> wrote:
>>>> On Wed, May 20, 2015 at 11:41 AM, Bin Cheng <bin.cheng@arm.com> wrote:
>>>>> Hi,
>>>>> As we know, GCC is too conservative when checking overflow behavior in SCEV
>>>>> and loop related optimizers.  Result is some variable can't be recognized as
>>>>> scalar evolution and thus optimizations are missed.  To be specific,
>>>>> optimizers like ivopts and vectorizer are affected.
>>>>> This issue is more severe on 64 bit platforms, for example, PR62173 is
>>>>> failed on aarch64; scev-3.c and scev-4.c were marked as XFAIL on lp64
>>>>> platforms.
>>>>>
>>>>> As the first part to improve overflow checking in GCC, this patch does below
>>>>> improvements:
>>>>>   1) Ideally, chrec_convert should be responsible to convert scev like
>>>>> "(type){base, step}" to scev like "{(type)base, (type)step} when the result
>>>>> scev doesn't overflow; chrec_convert_aggressive should do the conversion if
>>>>> the result scev could overflow/wrap.  Unfortunately, current implementation
>>>>> may use chrec_convert_aggressive to return a scev that won't overflow.  This
>>>>> is because of a) the static parameter "fold_conversions" for
>>>>> instantiate_scev_convert can only tracks whether chrec_convert_aggressive
>>>>> may be called, rather than if it does some overflow conversion or not;  b)
>>>>> the implementation of instantiate_scev_convert sometimes shortcuts the call
>>>>> to chrec_convert and misses conversion opportunities.  This patch improves
>>>>> this.
>>>>>   2) iv->no_overflow computed in simple_iv is too conservative.  With 1)
>>>>> fixed, iv->no_overflow should reflects whether chrec_convert_aggressive does
>>>>> return an overflow scev.  This patch improves this.
>>>>>   3) chrec_convert should be able to prove the resulting scev won't overflow
>>>>> with loop niter information.  This patch doesn't finish this, but it
>>>>> factored a new interface out of scev_probably_wraps_p for future
>>>>> improvement.  And that will be the part II patch.
>>>>>
>>>>> With the improvements in SCEV, this patch also improves optimizer(IVOPT)
>>>>> that uses scev information like below:
>>>>>   For array reference in the form of arr[IV], GCC tries to derive new
>>>>> address iv {arr+iv.base, iv.step*elem_size} from IV.  If IV overflow wrto a
>>>>> type that is narrower than address space, this derivation is not true
>>>>> because &arr[IV] isn't a scev.  Root cause why scev-*.c are failed now is
>>>>> the overflow information of IV is too conservative.  IVOPT has to be
>>>>> conservative to reject &arr[IV] as a scev.  With more accurate overflow
>>>>> information, IVOPT can be improved too.  So this patch fixes the mentioned
>>>>> long standing issues.
>>>>>
>>>>> Bootstrap and test on x86_64, x86 and aarch64.
>>>>> BTW, test gcc.target/i386/pr49781-1.c failed on x86_64, but I can confirmed
>>>>> it's not this patch's fault.
>>>>>
>>>>> So what's your opinion on this?.
>>>>
>>>> I maybe mixing things up but does
>>>>
>>>> +chrec_convert_aggressive (tree type, tree chrec, bool *fold_conversions)
>>>>  {
>>>> ...
>>>> +  if (evolution_function_is_affine_p (chrec))
>>>> +    {
>>>> +      tree base, step;
>>>> +      struct loop *loop;
>>>> +
>>>> +      loop = get_chrec_loop (chrec);
>>>> +      base = CHREC_LEFT (chrec);
>>>> +      step = CHREC_RIGHT (chrec);
>>>> +      if (convert_affine_scev (loop, type, &base, &step, NULL, true))
>>>> +       return build_polynomial_chrec (loop->num, base, step);
>>>>
>>>> ^^^ not forget to set *fold_conversions to true?  Or we need to use
>>>> convert_affine_scev (..., false)?
>>>
>>> Nice catch.  It's supposed to be called only if source scev has no
>>> overflow behavior introduced by previous call to
>>> chrec_convert_aggressive.  In other words, it should be guarded by
>>> "!*fold_conversions" like below:
>>>
>>> +
>>> +  if (!*fold_conversions && evolution_function_is_affine_p (chrec))
>>> +    {
>>> +      tree base, step;
>>> +      struct loop *loop;
>>> +
>>> +      loop = get_chrec_loop (chrec);
>>> +      base = CHREC_LEFT (chrec);
>>> +      step = CHREC_RIGHT (chrec);
>>> +      if (convert_affine_scev (loop, type, &base, &step, NULL, true))
>>> +    return build_polynomial_chrec (loop->num, base, step);
>>> +    }
>>>
>>> The scenario is rare that didn't exposed in either bootstrap or reg-test.
>>>
>>> Here is the updated patch without any other difference.  Bootstrap and
>>> test on x86_64 & AArch64.
>>
>> Ok.
>>
>> Thanks,
>> Richard.
>>
>>> Thanks,
>>> bin
>>>>
>>>> +    }
>>>>
>>>> (bah, and the diff somehow messes up -p context :/  which is why I like
>>>> context diffs more)
>>>>
>>>> Other from the above the patch looks good to me.
>>>>
>>>> Thanks,
>>>> Richard.
>>>>
>>>>> Thanks,
>>>>> bin
>>>>>
>>>>> 2015-05-20  Bin Cheng  <bin.cheng@arm.com>
>>>>>
>>>>>         PR tree-optimization/62173
>>>>>         * tree-ssa-loop-ivopts.c (struct iv): New field.  Reorder fields.
>>>>>         (alloc_iv, set_iv): New parameter.
>>>>>         (determine_biv_step): Delete.
>>>>>         (find_bivs): Inline original determine_biv_step.  Pass new
>>>>>         argument to set_iv.
>>>>>         (idx_find_step): Use no_overflow information for conversion.
>>>>>         * tree-scalar-evolution.c (analyze_scalar_evolution_in_loop): Let
>>>>>         resolve_mixers handle folded_casts.
>>>>>         (instantiate_scev_name): Change bool parameter to bool pointer.
>>>>>         (instantiate_scev_poly, instantiate_scev_binary): Ditto.
>>>>>         (instantiate_array_ref, instantiate_scev_not): Ditto.
>>>>>         (instantiate_scev_3, instantiate_scev_2): Ditto.
>>>>>         (instantiate_scev_1, instantiate_scev_r): Ditto.
>>>>>         (instantiate_scev_convert, ): Change parameter.  Pass argument
>>>>>         to chrec_convert_aggressive.
>>>>>         (instantiate_scev): Change argument.
>>>>>         (resolve_mixers): New parameter and set it.
>>>>>         (scev_const_prop): New argument.
>>>>>         * tree-scalar-evolution.h (resolve_mixers): New parameter.
>>>>>         * tree-chrec.c (convert_affine_scev): Call chrec_convert instead
>>>>>         of chrec_conert_1.
>>>>>         (chrec_convert): New parameter.  Move definition below.
>>>>>         (chrec_convert_aggressive): New parameter and set it.  Call
>>>>>         convert_affine_scev.
>>>>>         * tree-chrec.h (chrec_convert): New parameter.
>>>>>         (chrec_convert_aggressive): Ditto.
>>>>>         * tree-ssa-loop-niter.c (loop_exits_before_overflow): New function.
>>>>>         (scev_probably_wraps_p): Factor loop niter related code into
>>>>>         loop_exits_before_overflow.
>>>>>
>>>>> gcc/testsuite/ChangeLog
>>>>> 2015-05-20  Bin Cheng  <bin.cheng@arm.com>
>>>>>
>>>>>         PR tree-optimization/62173
>>>>>         * gcc.dg/tree-ssa/scev-3.c: Remove xfail.
>>>>>         * gcc.dg/tree-ssa/scev-4.c: Ditto.
>>>>>         * gcc.dg/tree-ssa/scev-8.c: New.
>
> Attached patch applied as revision 224009.  I removed changes in
> tree-ssa-loop-niter.c (also the new test) in this patch because it's
> just code refactoring and will be covered by the second patch.

As noticed, gcc.target/i386/pr49781-1.c failed because of this patch.
Seems IVO made sub-optimal choice with even more IVs recognized.
Filed PR66388 for tracking and will investigate the failure.

Thanks,
bin