Re: [PATCH] Set bound/cmp/control for until wrap loop.

[Jiufu Guo <guojiufu@linux.ibm.com>]

在 2021/9/1 上午11:30, Jiufu Guo via Gcc-patches 写道:
> Richard Biener <rguenther@suse.de> writes:
>
>> On Tue, 31 Aug 2021, guojiufu wrote:
>>
>>> On 2021-08-30 20:02, Richard Biener wrote:
>>> > On Mon, 30 Aug 2021, guojiufu wrote:
>>> > >> On 2021-08-30 14:15, Jiufu Guo wrote:
>>> >> > Hi,
>>> >> >
>>> >> > In patch r12-3136, niter->control, niter->bound and >> > 
>>> niter->cmp are
>>> >> > derived from number_of_iterations_lt.  While for 'until >> > 
>>> wrap condition',
>>> >> > the calculation in number_of_iterations_lt is not align >> > 
>>> the requirements
>>> >> > on the define of them and requirements in >> > 
>>> determine_exit_conditions.
>>> >> >
>>> >> > This patch calculate niter->control, niter->bound and >> > 
>>> niter->cmp in
>>> >> > number_of_iterations_until_wrap.
>>> >> >
>>> >> > The ICEs in the PR are pass with this patch.
>>> >> > Bootstrap and reg-tests pass on ppc64/ppc64le and x86.
>>> >> > Is this ok for trunk?
>>> >> >
>>> >> > BR.
>>> >> > Jiufu Guo
>>> >> >
>>> >> Add ChangeLog:
>>> >> >  create mode 100644 gcc/testsuite/gcc.dg/pr102087.c
>>> >> >
>>> >> > diff --git a/gcc/tree-ssa-loop-niter.c >> > 
>>> b/gcc/tree-ssa-loop-niter.c
>>> >> > index 7af92d1c893..747f04d3ce0 100644
>>> >> > --- a/gcc/tree-ssa-loop-niter.c
>>> >> > +++ b/gcc/tree-ssa-loop-niter.c
>>> >> > @@ -1482,7 +1482,7 @@ number_of_iterations_until_wrap >> > 
>>> (class loop *,
>>> >> > tree type, affine_iv *iv0,
>>> >> >                   affine_iv *iv1, class >> >  tree_niter_desc 
>>> *niter)
>>> >> >  {
>>> >> >    tree niter_type = unsigned_type_for (type);
>>> >> > -  tree step, num, assumptions, may_be_zero;
>>> >> > +  tree step, num, assumptions, may_be_zero, span;
>>> >> >    wide_int high, low, max, min;
>>> >> >
>>> >> >    may_be_zero = fold_build2 (LE_EXPR, boolean_type_node, >> 
>>> >    iv1->base,
>>> >> > iv0->base);
>>> >> > @@ -1513,6 +1513,8 @@ number_of_iterations_until_wrap >> > 
>>> (class loop *,
>>> >> > tree type, affine_iv *iv0,
>>> >> >   low = wi::to_wide (iv0->base);
>>> >> >          else
>>> >> >     low = min;
>>> >> > +
>>> >> > +      niter->control = *iv1;
>>> >> >      }
>>> >> >    /* {base, -C} < n.  */
>>> >> >    else if (tree_int_cst_sign_bit (iv0->step) && >> >    
>>> integer_zerop
>>> >> > (iv1->step))
>>> >> > @@ -1533,6 +1535,8 @@ number_of_iterations_until_wrap >> > 
>>> (class loop *,
>>> >> > tree type, affine_iv *iv0,
>>> >> >   high = wi::to_wide (iv1->base);
>>> >> >          else
>>> >> >     high = max;
>>> >> > +
>>> >> > +      niter->control = *iv0;
>>> >> >      }
>>> >> >    else
>>> >> >      return false;
>>> > > it looks like the above two should already be in effect from > the
>>> > caller (guarding with integer_nozerop)?
>>>
>>> I add them just because set these fields in one function.
>>> Yes, they have been set in caller already,  I could remove them here.
>>>
>>> > >> > @@ -1556,6 +1560,14 @@ number_of_iterations_until_wrap >> > 
>>> (class loop *,
>>> >> > tree type, affine_iv *iv0,
>>> >> >            niter->assumptions, assumptions);
>>> >> >
>>> >> >    niter->control.no_overflow = false;
>>> >> > +  niter->control.base = fold_build2 (MINUS_EXPR, >> > niter_type,
>>> >> > +                     niter->control.base,
>>> >> > niter->control.step);
>>> > > how do we know IVn - STEP doesn't already wrap?
>>>
>>> The last IV value is just cross the max/min value of the type
>>> at the last iteration,  then IVn - STEP is the nearest value
>>> to max(or min) and not wrap.
>>>
>>> > A comment might be
>>> > good to explain you're turning the simplified exit condition > into
>>> > >    { IVbase - STEP, +, STEP } != niter * STEP + (IVbase - >    
>>> STEP)
>>> > > which, when mathematically looking at it makes me wonder why > 
>>> there's
>>> > the seemingly redundant '- STEP' term?  Also is NE_EXPR > really
>>> > correct since STEP might be not 1?  Only for non equality > compares
>>> > the '- STEP' should matter?
>>>
>>> I need to add comments for this.  This is a little tricky.
>>> The last value of the original IV just cross max/min at most one STEP,
>>> at there wrapping already happen.
>>> Using "{IVbase, +, STEP} != niter * STEP + IVbase" is not wrong
>>> in the aspect of exit condition.
>>>
>>> But this would not work well with existing code:
>>> like determine_exit_conditions, which will convert NE_EXP to
>>> LT_EXPR/GT_EXPR.  And so, the '- STEP' is added to adjust the
>>> IV.base and bound, with '- STEP' the bound will be the last value
>>> just before wrap.
>>
>> Hmm.  The control IV is documented as
>>
>>   /* The simplified shape of the exit condition.  The loop exits   if
>>      CONTROL CMP BOUND is false, where CMP is one of NE_EXPR,
>>      LT_EXPR, or GT_EXPR, and step of CONTROL is positive if CMP      is
>>      LE_EXPR and negative if CMP is GE_EXPR.  This information      
>> is used
>>      by loop unrolling.  */
>>   affine_iv control;
>>
>> but determine_exit_conditions seems to assume the IV does not wrap?
>
> Strictly speaking , I would say yes,  determine_exit_conditions assume
> IV does not wrap: there is code:
>
>  if (cmp == LT_EXPR)
>    assum = fold_build2 (GE_EXPR, boolean_type_node,
>              bound,
>              fold_build2 (PLUS_EXPR, type, min, delta));
>  else
>     xxxx
>     This means if 'bound' is the value after wrap, the 'assum' with be 
> false.
> This is also the reason that we may need to biase 'bound' and 'base' by
> 'step * 1'.  Because, in our case like "while(n<l++)",
> if we set 'bound' as 'iv.base + niter * step', the value of 'bound' will
> be '0(zero)' which crosses max of the type one STEP.
> So, we may transform the exit condition to
> "{IVbase - STEP, +, STEP } != niter * STEP + (IVbase - STEP)"
> And then new control IV and new bound does not wrap.
>
>> In fact determine_exit_conditions seems to just build ->base CMP bound
>> where bound is the IV bound biased by #unroll * step - step.  So how
>> does biasing by step * 1 help?
>
> determine_exit_conditions adjust bound by #unroll * step - step for more
> check, like transform to "base cmp new-bound"; for this checking, 'bound'
> is ok with wrapped value, but for the 'assum' as above (assum =
> bound >= min + #unroll * step - step), biasing "step * 1" would be fine.
>
>>
>> Does the control IV wrap in our case?
>
> I think, this is a key question, which may affect if it make sense for
> above transform.  Thanks!
> Let me explain my understanding: for original exit condition (like 
> n<l++),
> the final value of the IV(i), it is already cross max/min value of the 
> type.
> Or say, for the value of IV(i) "after exit loop", wraps if strictly
> speaking.
> While, for all the values of IV(i) inside the loop, they are in valid
> range of the type;  we may treat it as 'no wrap' for some usage like 
> unroll.

Correction: for all the "other" values of IV(i) inside the loop are not 
wrap.

But the issue is the 'final' value of IV(i) may be used in the last 
iteration.

So, I'm also wondering if we may need to keep it as:

"{IVbase, +, STEP } != niter * STEP + IVbase.

>
> Thanks for review and future comments and suggestions!

BTW, I notice the lines seems not well wrap-line, sorry.

>
> BR,
> Jiufu
>
>>
>> Richard.
>>
>>> Thanks again for your review!
>>>
>>> BR.
>>> Jiufu
>>>
>>> > > Richard.
>>> > >> > +  span = fold_build2 (MULT_EXPR, niter_type, >> > niter->niter,
>>> >> > +              fold_convert (niter_type, >> > 
>>> niter->control.step));
>>> >> > +  niter->bound = fold_build2 (PLUS_EXPR, niter_type, >> > span,
>>> >> > +                  fold_convert (niter_type, >> > 
>>> niter->control.base));
>>> >> > +  niter->bound = fold_convert (type, niter->bound);
>>> >> > +  niter->cmp = NE_EXPR;
>>> >> >
>>> >> >    return true;
>>> >> > }
>>> >> > diff --git a/gcc/testsuite/gcc.dg/pr102087.c
>>> >> > b/gcc/testsuite/gcc.dg/pr102087.c
>>> >> > new file mode 100644
>>> >> > index 00000000000..ef1f9f5cba9
>>> >> > --- /dev/null
>>> >> > +++ b/gcc/testsuite/gcc.dg/pr102087.c
>>> >> > @@ -0,0 +1,25 @@
>>> >> > +/* { dg-do compile } */
>>> >> > +/* { dg-options "-O3" } */
>>> >> > +
>>> >> > +unsigned __attribute__ ((noinline))
>>> >> > +foo (int *__restrict__ a, int *__restrict__ b, unsigned >> > 
>>> l, unsigned n)
>>> >> > +{
>>> >> > +  while (n < ++l)
>>> >> > +    *a++ = *b++ + 1;
>>> >> > +  return l;
>>> >> > +}
>>> >> > +
>>> >> > +volatile int a[1];
>>> >> > +unsigned b;
>>> >> > +int c;
>>> >> > +
>>> >> > +int
>>> >> > +check ()
>>> >> > +{
>>> >> > +  int d;
>>> >> > +  for (; b > 1; b++)
>>> >> > +    for (c = 0; c < 2; c++)
>>> >> > +      for (d = 0; d < 2; d++)
>>> >> > +    a[0];
>>> >> > +  return 0;
>>> >> > +}
>>> >> > diff --git a/gcc/testsuite/gcc.dg/vect/pr101145_3.c
>>> >> > b/gcc/testsuite/gcc.dg/vect/pr101145_3.c
>>> >> > index 99289afec0b..40cb0240aaa 100644
>>> >> > --- a/gcc/testsuite/gcc.dg/vect/pr101145_3.c
>>> >> > +++ b/gcc/testsuite/gcc.dg/vect/pr101145_3.c
>>> >> > @@ -1,5 +1,6 @@
>>> >> >  /* { dg-require-effective-target vect_int } */
>>> >> >  /* { dg-options "-O3 -fdump-tree-vect-details" } */
>>> >> > +
>>> >> >  #define TYPE int *
>>> >> >  #define MIN ((TYPE)0)
>>> >> >  #define MAX ((TYPE)((long long)-1))
>>> >> > @@ -10,4 +11,5 @@
>>> >> >
>>> >> >  #include "pr101145.inc"
>>> >> >
>>> >> > -/* { dg-final { scan-tree-dump-times "vectorized 1 loops" >> > 
>>> 2 "vect" } }
>>> >> > */
>>> >> > +/* pointer size may not be vectorized, checking niter is >> > 
>>> ok. */
>>> >> > +/* { dg-final { scan-tree-dump "Symbolic number of >> > 
>>> iterations is" "vect"
>>> >> > }
>>> >> > } */
>>> >>
>>>