From: Martin Jambor <mjambor@suse.cz>
To: Jan Hubicka <hubicka@ucw.cz>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>
Subject: Re: [PATCH 4/6] ipa: Multiple predicates for loop properties, with frequencies
Date: Fri, 02 Oct 2020 14:31:39 +0200 [thread overview]
Message-ID: <ri6wo08dd1w.fsf@suse.cz> (raw)
In-Reply-To: <20200929221815.GE7702@kam.mff.cuni.cz>
Hi,
On Wed, Sep 30 2020, Jan Hubicka wrote:
>> This patch enhances the ability of IPA to reason under what conditions
>> loops in a function have known iteration counts or strides because it
>> replaces single predicates which currently hold conjunction of
>> predicates for all loops with vectors capable of holding multiple
>> predicates, each with a cumulative frequency of loops with the
>> property.
>>
>> This second property is then used by IPA-CP to much more aggressively
>> boost its heuristic score for cloning opportunities which make
>> iteration counts or strides of frequent loops compile time constant.
>>
>> gcc/ChangeLog:
>>
>> 2020-09-03 Martin Jambor <mjambor@suse.cz>
>>
>> * ipa-fnsummary.h (ipa_freqcounting_predicate): New type.
>> (ipa_fn_summary): Change the type of loop_iterations and loop_strides
>> to vectors of ipa_freqcounting_predicate.
>> (ipa_fn_summary::ipa_fn_summary): Construct the new vectors.
>> (ipa_call_estimates): New fields loops_with_known_iterations and
>> loops_with_known_strides.
>> * ipa-cp.c (hint_time_bonus): Multiply param_ipa_cp_loop_hint_bonus
>> with the expected frequencies of loops with known iteration count or
>> stride.
>> * ipa-fnsummary.c (add_freqcounting_predicate): New function.
>> (ipa_fn_summary::~ipa_fn_summary): Release the new vectors instead of
>> just two predicates.
>> (remap_hint_predicate_after_duplication): Replace with function
>> remap_freqcounting_preds_after_dup.
>> (ipa_fn_summary_t::duplicate): Use it or duplicate new vectors.
>> (ipa_dump_fn_summary): Dump the new vectors.
>> (analyze_function_body): Compute the loop property vectors.
>> (ipa_call_context::estimate_size_and_time): Calculate also
>> loops_with_known_iterations and loops_with_known_strides. Adjusted
>> dumping accordinly.
>> (remap_hint_predicate): Replace with function
>> remap_freqcounting_predicate.
>> (ipa_merge_fn_summary_after_inlining): Use it.
>> (inline_read_section): Stream loopcounting vectors instead of two
>> simple predicates.
>> (ipa_fn_summary_write): Likewise.
>> * params.opt (ipa-max-loop-predicates): New parameter.
>> * doc/invoke.texi (ipa-max-loop-predicates): Document new param.
>>
>> diff --git a/gcc/ipa-fnsummary.c b/gcc/ipa-fnsummary.c
>> index 6082f34d63f..bbbb94aa930 100644
>> --- a/gcc/ipa-fnsummary.c
>> +++ b/gcc/ipa-fnsummary.c
>> @@ -310,6 +310,36 @@ set_hint_predicate (predicate **p, predicate new_predicate)
>> }
>> }
>>
>> +/* Find if NEW_PREDICATE is already in V and if so, increment its freq.
>> + Otherwise add a new item to the vector with this predicate and frerq equal
>> + to add_freq, unless the number of predicates would exceed MAX_NUM_PREDICATES
>> + in which case the function does nothing. */
>> +
>> +static void
>> +add_freqcounting_predicate (vec<ipa_freqcounting_predicate, va_gc> **v,
>> + const predicate &new_predicate, sreal add_freq,
>> + unsigned max_num_predicates)
>> +{
>> + if (new_predicate == false || new_predicate == true)
>> + return;
>> + ipa_freqcounting_predicate *f;
>> + for (int i = 0; vec_safe_iterate (*v, i, &f); i++)
>> + if (new_predicate == f->predicate)
>> + {
>> + f->freq += add_freq;
>> + return;
>> + }
>> + if (vec_safe_length (*v) >= max_num_predicates)
>> + /* Too many different predicates to account for. */
>> + return;
>> +
>> + ipa_freqcounting_predicate fcp;
>> + fcp.predicate = NULL;
>> + set_hint_predicate (&fcp.predicate, new_predicate);
>> + fcp.freq = add_freq;
>> + vec_safe_push (*v, fcp);
>> + return;
>> +}
>>
>> /* Compute what conditions may or may not hold given information about
>> parameters. RET_CLAUSE returns truths that may hold in a specialized copy,
>> @@ -710,13 +740,17 @@ ipa_call_summary::~ipa_call_summary ()
>>
>> ipa_fn_summary::~ipa_fn_summary ()
>> {
>> - if (loop_iterations)
>> - edge_predicate_pool.remove (loop_iterations);
>> - if (loop_stride)
>> - edge_predicate_pool.remove (loop_stride);
>> + unsigned len = vec_safe_length (loop_iterations);
>> + for (unsigned i = 0; i < len; i++)
>> + edge_predicate_pool.remove ((*loop_iterations)[i].predicate);
>> + len = vec_safe_length (loop_strides);
>> + for (unsigned i = 0; i < len; i++)
>> + edge_predicate_pool.remove ((*loop_strides)[i].predicate);
>
> For edges predicates are pointers since most of them have no interesting
> predicate and thus NULL is more compact. I guess here it would make
> snese to make predicates inline. Is there a problem with vectors not
> liking non-pods?
>> vec_free (conds);
>> vec_free (size_time_table);
>> vec_free (call_size_time_table);
>> + vec_free (loop_iterations);
>> + vec_free (loop_strides);
>
> However auto_vecs should work in the brave new C++ world.
Well, the summary lives in GC memory, so I don't think I can put
auto_vecs there.
I will add a note to look into putting a predicate directly instead as a
pointer to ipa_freqcounting_predicate as a follow-up patch.
>
> The patch looks reasonable to me. Did you check how much memory it
> consumes building bigger projects? Also I am bit worried about our
> ability to use it reasonably in the heuristics since it is quite
> complicated value...
Thanks!
Martin
next prev parent reply other threads:[~2020-10-02 12:31 UTC|newest]
Thread overview: 18+ messages / expand[flat|nested] mbox.gz Atom feed top
2020-09-29 18:12 [PATCH 0/6] IPA cleanups and IPA-CP improvements for 548.exchange2_r Martin Jambor
2020-09-21 14:25 ` [PATCH 6/6] ipa-cp: Separate and increase the large-unit parameter Martin Jambor
2020-09-29 19:30 ` Jan Hubicka
2020-09-30 6:35 ` Richard Biener
2020-09-30 16:39 ` Martin Jambor
2020-10-26 11:00 ` Tamar Christina
2020-09-21 14:25 ` [PATCH 4/6] ipa: Multiple predicates for loop properties, with frequencies Martin Jambor
2020-09-29 22:18 ` Jan Hubicka
2020-10-02 12:31 ` Martin Jambor [this message]
2020-09-21 14:25 ` [PATCH 5/6] ipa-cp: Add dumping of overall_size after cloning Martin Jambor
2020-09-29 18:39 ` Jan Hubicka
2020-09-28 18:47 ` [PATCH 1/6] ipa: Bundle vectors describing argument values Martin Jambor
2020-10-02 11:54 ` Jan Hubicka
2020-09-28 18:47 ` [PATCH 3/6] ipa: Bundle estimates of ipa_call_context::estimate_size_and_time Martin Jambor
2020-09-29 18:39 ` Jan Hubicka
2020-09-28 18:47 ` [PATCH 2/6] ipa: Introduce ipa_cached_call_context Martin Jambor
2020-09-29 18:27 ` Jan Hubicka
-- strict thread matches above, loose matches on Subject: below --
2020-09-07 19:36 [PATCH 4/6] ipa: Multiple predicates for loop properties, with frequencies Martin Jambor
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=ri6wo08dd1w.fsf@suse.cz \
--to=mjambor@suse.cz \
--cc=gcc-patches@gcc.gnu.org \
--cc=hubicka@ucw.cz \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).