From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-ed1-x531.google.com (mail-ed1-x531.google.com [IPv6:2a00:1450:4864:20::531]) by sourceware.org (Postfix) with ESMTPS id D833B385802E for ; Wed, 10 Nov 2021 08:59:48 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org D833B385802E Received: by mail-ed1-x531.google.com with SMTP id c8so7512186ede.13 for ; Wed, 10 Nov 2021 00:59:48 -0800 (PST) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=x-gm-message-state:mime-version:references:in-reply-to:from:date :message-id:subject:to:cc:content-transfer-encoding; bh=94nsKYuyR9jTB3tJqH9kLuc+b7qUEkQb8vOnVskfvqQ=; b=49K/1w2moCRuDfvJdrt3A3JWsfFvG1/TeBY8i4BzErO/lvKeRgYwk86y2WZwxJQliP Jz2kBETJY18zvXBPbTYPL47lrGppA88pIuctAkZfqOrk2jrpKWTsx/zwfv1+xkYptTSe 4emJj+nhynvWpdHipQrYXO1u8Yb5VQuWmelkR8b6On6SjXcM1wldS9ZMgOGHen7Axo2M r2oRvZoMbRXxBRJVqHLC8aF5pF2rMRO5wlpo3OoZDBmTE/YIcwtBQbHTbi6o9/3mSKoN XKHVfImc7Jdpxjm02BVRhN2KfMjqQmm3ReDKQGdra5HuKxqsTqo5aAh81lcTCOyLeETP b80w== X-Gm-Message-State: AOAM532tZjvEs8ngXYbZNy4YAzElKwGbZH2liWp4X4g1crlu/q3BkSIZ OF6WPKJ+3NraYwmuFvruttelP2ATf3hOlCDbWc8= X-Google-Smtp-Source: ABdhPJwSsBqd9g7FUncCsY3sigQSAHD52v0T4BkCgvRtZ+BvCZSOEIX3lw24brdTdgoHiFwF65355+GvwPCEE8N71ns= X-Received: by 2002:a50:e0c3:: with SMTP id j3mr19462981edl.97.1636534787928; Wed, 10 Nov 2021 00:59:47 -0800 (PST) MIME-Version: 1.0 References: <7791e850-f74d-8c1c-f67c-e02f3f6e007d@redhat.com> <5c6c91d4-ed8b-8d98-2cd9-bafc84e6f2a4@suse.cz> <8da24825-19ec-56a6-a68c-5c37c7acc3e1@redhat.com> <59763e1a-8432-5f23-c399-a9b4dd6c6dff@suse.cz> In-Reply-To: <59763e1a-8432-5f23-c399-a9b4dd6c6dff@suse.cz> From: Richard Biener Date: Wed, 10 Nov 2021 09:59:37 +0100 Message-ID: Subject: Re: [PATCH] Loop unswitching: support gswitch statements. To: =?UTF-8?Q?Martin_Li=C5=A1ka?= Cc: Andrew MacLeod , GCC Patches Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: quoted-printable X-Spam-Status: No, score=-2.5 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, FREEMAIL_FROM, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_PASS, TXREP autolearn=ham autolearn_force=no version=3.4.4 X-Spam-Checker-Version: SpamAssassin 3.4.4 (2020-01-24) on server2.sourceware.org X-BeenThere: gcc-patches@gcc.gnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Gcc-patches mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Wed, 10 Nov 2021 08:59:50 -0000 On Tue, Nov 9, 2021 at 5:44 PM Martin Li=C5=A1ka wrote: > > On 11/9/21 14:37, Richard Biener wrote: > > On Mon, Nov 8, 2021 at 8:45 PM Andrew MacLeod wro= te: > >> > >> On 11/8/21 10:05 AM, Martin Li=C5=A1ka wrote: > >>> On 9/28/21 22:39, Andrew MacLeod wrote: > >>>> In Theory, modifying the IL should be fine, it happens already in > >>>> places, but its not extensively tested under those conditions yet. > >>> > >>> Hello Andrew. > >>> > >>> I've just tried using a global gimple_ranger and it crashes when loop > >>> unswitching duplicates > >>> some BBs. > >>> > >>> Please try the attached patch for: > >> > >> hey Martin, > >> > >> try using this in your tree. Since nothing else is using a growing BB > >> right now, I'll let you work with it and see if everything works as > >> expected before checking it in, just in case we need more tweaking. > >> With this, > >> > >> make RUNTESTFLAGS=3Ddg.exp=3Dloop-unswitch*.c check-gcc > >> > >> runs clean. > >> > >> > >> basically, I tried to grow it by either a factor of 10% for the curren= t > >> BB size when the grow is requested, or some double the needed extra > >> size, or 128... whichever value is "maximum" That means it shoudnt = be > >> asking for tooo much each time, but also not a minimum amount. > >> > >> Im certainly open to suggestion on how much to grow it each time. > >> Note the vector being grown is ONLY fo the SSA_NAme being asked for.. = so > >> it really an on-demand thing just for specific names, in your case, > >> mostly just the switch index. > >> > >> Let me know how this works for you, and if you have any other issues. > > > > So I think in the end we shouldn't need the growing. Ideally we'd do a= ll > > the analysis before the first transform, but for that we'd need ranger = to > > be able to "simplify" conditions based on a known true/false predicate > > that's not yet in the IL. Consider > > > > for (;;) > > { > > if (invariant < 3) // A > > { > > ... > > } > > if (invariant < 5) // B > > { > > ... > > } > > } > > > > unswitch analysis will run into the condition 'A' and determine the loo= p > > can be unswitched with the condition 'invariant < 3'. To be able to > > perform cost assessment and to avoid redundant unswitching we > > want to determine that if we unswitch with 'invariant < 3' being > > true then the condition at 'B' is true as well before actually insertin= g > > the if (invariant < 3) outside of the loop. > > > > So I'm thinking of assigning a gimple_uid to each condition we want to > > unswitch on and have an array indexed by the uid with meta-data on > > the unswitch opportunity, the "related" conditions could be marked with > > the same uid (or some other), and the folding result recorded so that > > at transform time we can just do the appropriate replacement without > > invoking ranger again. > > Calculating all this before transformation is quite ambitious based on th= e code > we have now. > > Note one can have in a loop: > > if (a > 100) > ... > > switch (a) > case 1000: > ... > case 20: > ... > case 200: > ... > > which means the first predicate effectively makes some cases unreachable.= Moreover > one can have > > if (a > 100 && b < 300) > ... > > and more complex conditions. True - I guess we should do two things. 1) keep simplify_using_entry_checks like code for symbolic conditions 2) add integer ranges for unswitch conditions producing them, that includes all unswitching of switch stmts - we might be able to use the ranger queries (with global ranges) to simplify stmts with the known ranges as noted by Andrew I do think that pre-computing the simplifications is what we should do to be able to make the cost modeling sane. What we can avoid trying is evaluating multiple unswitch possibilities to pick the "best". I think changing the code do to the analysis first should be done before wiring in gcond support, even adding the additional 'range' capability will be useful without that since the current code wont figure out a > 5 is true when we unswitch on a > 3. > > > > Now, but how do we arrange for the ranger analysis here? > > That's likely something we need support from ranger, yes. > > > > > We might also somehow want to remember that on the > > 'invariant < 3' =3D=3D false copy of the loop there's still the > > unswitching opportunity on 'invariant < 5', but not on the > > 'invariant < 5' =3D=3D true copy. > > > > Currently unswitching uses a custom simplify_using_entry_checks > > which tries to do simplification only after the fact (and so costing > > also is far from costing the true cost and ordering of the opportunitie= s > > to do the best first is not implemented either). > > I'm sending updated version of the patch where I changed: > - simplify_using_entry_checks is put back for the floating point expressi= ons > - all scans utilize scan-tree-dump-times > - some new tests were added > - global ranger is used (I rely on the growing patch provided by Andrew) > - better ranger API is used for gcond expression: ranger.range_of_stmt (r= , stmt) && r.singleton_p (&result)) > - auto_edge_flag is used now > > Patch can bootstrap on x86_64-linux-gnu and survives regression tests. > > Thoughts? > Martin > > > > > Richard. > > > >> Andrew > >>