On Wed, Apr 25, 2012 at 4:32 PM, Igor Zamyatin <izamya...@gmail.com> wrote:
> On Wed, Apr 25, 2012 at 1:14 PM, Richard Guenther
> <richard.guent...@gmail.com> wrote:
>> On Wed, Apr 25, 2012 at 10:56 AM, Igor Zamyatin <izamya...@gmail.com> wrote:
>>> Hi all!
>>>
>>> I'd like to post for review the patch which makes some costs adjusting
>>> in get_computation_cost_at routine in ivopts part.
>>> As mentioned in the PR changes also fix the bwaves regression from PR 52272.
>>> Also changes introduce no degradations on spec2000/2006 and
>>> EEMBC1.1/2.0(this was measured on Atom) on x86
>>>
>>>
>>> Bootstrapped and regtested on x86. Ok to commit?
>>
>> I can't make sense of the patch and the comment does not help.
>>
>> + diff_cost = cost.cost;
>> cost.cost /= avg_loop_niter (data->current_loop);
>> + add_cost_val = add_cost (TYPE_MODE (ctype), data->speed);
>> + /* Do cost correction if address cost is small enough
>> + and difference cost is high enough. */
>> + if (address_p && diff_cost > add_cost_val
>> + && get_address_cost (symbol_present, var_present,
>> + offset, ratio, cstepi,
>> + TYPE_MODE (TREE_TYPE (utype)),
>> + TYPE_ADDR_SPACE (TREE_TYPE (utype)),
>> + speed, stmt_is_after_inc,
>> + can_autoinc).cost <= add_cost_val)
>> + cost.cost += add_cost_val;
>>
>> Please explain more thoroughly. It also would seem to be better to add
>> an extra case, as later code does
>
> For example for such code
>
> for (j=0; j<M;j++) {
> for (i=0; i<N; i++)
> sum += ptr->a[j][i] * ptr->c[k][i];
> }
> we currently have following gimple on x86 target (I provided a piece
> of all phase output):
>
> # ivtmp.13_30 = PHI <ivtmp.13_31(3), ivtmp.13_33(7)>
> D.1748_34 = (void *) ivtmp.13_30;
> D.1722_7 = MEM[base: D.1748_34, offset: 0B];
> D.1750_36 = ivtmp.27_28;
> D.1751_37 = D.1750_36 + ivtmp.13_30; <-- we got
> non-invariant add which is not taken into account currently in cost
> model
> D.1752_38 = (void *) D.1751_37;
> D.1753_39 = (sizetype) k_8(D);
> D.1754_40 = D.1753_39 * 800;
> D.1723_9 = MEM[base: D.1752_38, index: D.1754_40, offset: 16000B];
> ...
>
> With proposed fix we produce:
>
> # ivtmp.14_30 = PHI <ivtmp.14_31(3), 0(7)>
> D.1749_34 = (struct S *) ivtmp.25_28;
> D.1722_7 = MEM[base: D.1749_34, index: ivtmp.14_30, offset: 0B];
> D.1750_35 = (sizetype) k_8(D);
> D.1751_36 = D.1750_35 * 800;
> D.1752_37 = ptr_6(D) + D.1751_36;
> D.1723_9 = MEM[base: D.1752_37, index: ivtmp.14_30, offset: 16000B];
>
> which is more effective on platforms where address cost is cheaper
> than cost of addition operation. That's basically what this adjustment
> is for.
If we generally miss to account for the add then why is the adjustment
conditional on diff_cost > add_cost and address_cost <= add_cost?
Is this a new heuristic or a fix for not accurately computing the cost for the
stmts we generate?
Richard.
> So comment in the source code now looks as follows
>
> /* Do cost correction when address difference produces
> additional non-invariant add operation which is less
> profitable if address cost is cheaper than cost of add. */
>
>>
>> /* Now the computation is in shape symbol + var1 + const + ratio * var2.
>> (symbol/var1/const parts may be omitted). If we are looking for an
>> address, find the cost of addressing this. */
>> if (address_p)
>> return add_costs (cost,
>> get_address_cost (symbol_present, var_present,
>> offset, ratio, cstepi,
>> TYPE_MODE (TREE_TYPE (utype)),
>> TYPE_ADDR_SPACE (TREE_TYPE (utype)),
>> speed, stmt_is_after_inc,
>> can_autoinc));
>>
>> thus refactoring the code a bit would make it possible to CSE the
>> get_address_cost
>> call and eventually make it clearer what the code does.
>
> 'offset' could be changed beetween two calls of get_address_cost so
> such refactoring looks useless.
>
> New patch (only the comment was changed) attached. Changelog was
> changed as well.
>
>>
>> Richard.
>>
>
> Changelog:
>
> 2012-04-26 Yuri Rumyantsev <yuri.rumyant...@intel.com>
>
> * tree-ssa-loop-ivopts.c (get_computation_cost_at): Adjust
> cost model when address difference produces additional
> non-invariant add operation which is less profitable if
> address cost is cheaper than cost of add.
>
> Thanks,
> Igor
