From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from nikam.ms.mff.cuni.cz (nikam.ms.mff.cuni.cz [195.113.20.16]) by sourceware.org (Postfix) with ESMTPS id 05AC33858D28; Thu, 16 Dec 2021 11:18:20 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org 05AC33858D28 Received: by nikam.ms.mff.cuni.cz (Postfix, from userid 16202) id 7E213281EAA; Thu, 16 Dec 2021 12:18:18 +0100 (CET) Date: Thu, 16 Dec 2021 12:18:18 +0100 From: Jan Hubicka To: Xionghu Luo Cc: wschmidt@linux.ibm.com, dje.gcc@gmail.com, gcc-patches@gcc.gnu.org, linkw@gcc.gnu.org, segher@kernel.crashing.org Subject: Re: [PATCH 2/3] Fix incorrect loop exit edge probability [PR103270] Message-ID: <20211216111818.GE4516@kam.mff.cuni.cz> References: <20211208055416.1415283-1-luoxhu@linux.ibm.com> <20211208055416.1415283-3-luoxhu@linux.ibm.com> <20211213092548.GA91590@kam.mff.cuni.cz> <5a057da8-677c-b5e9-48b3-2cb434e68505@linux.ibm.com> <8af589d9-13c1-2ff8-08d3-7caf98fc037a@linux.ibm.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <8af589d9-13c1-2ff8-08d3-7caf98fc037a@linux.ibm.com> User-Agent: Mutt/1.10.1 (2018-07-13) X-Spam-Status: No, score=-5.5 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, RCVD_IN_MSPIKE_H3, RCVD_IN_MSPIKE_WL, SPF_HELO_NONE, SPF_NONE, 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: Thu, 16 Dec 2021 11:18:23 -0000 > > > > > > ./contrib/analyze_brprob.py ~/workspace/tests/spec2017/dump_file_all > > HEURISTICS BRANCHES (REL) BR. HITRATE HITRATE COVERAGE COVERAGE (REL) predict.def (REL) HOT branches (>10%) > > noreturn call 1 0.0% 100.00% 50.00% / 50.00% 2 2.00 0.0% 100%:1 > > Fortran zero-sized array 3 0.0% 66.67% 41.71% / 60.50% 362 362.00 0.0% 100%:3 > > loop iv compare 16 0.0% 93.75% 98.26% / 98.76% 279847 279.85k 0.0% 93%:4 > > __builtin_expect 35 0.0% 97.14% 78.09% / 78.35% 17079558 17.08M 0.0% > > loop guard with recursion 45 0.1% 86.67% 85.13% / 85.14% 6722424412 6.72G 1.3% 74%:4 > > extra loop exit 80 0.1% 58.75% 81.49% / 89.21% 438470261 438.47M 0.1% 86%:3 > > guess loop iv compare 235 0.3% 80.85% 52.83% / 73.97% 148558247 148.56M 0.0% 47%:3 > > negative return 241 0.3% 71.37% 25.33% / 92.61% 250402383 250.40M 0.0% 69%:2 > > loop exit with recursion 315 0.4% 74.60% 85.07% / 85.71% 9403136858 9.40G 1.8% 59%:4 > > const return 320 0.4% 51.88% 90.45% / 95.63% 925341727 925.34M 0.2% 76%:5 > > indirect call 377 0.5% 51.46% 84.72% / 91.14% 2133772848 2.13G 0.4% 69%:1 > > polymorphic call 410 0.5% 44.15% 31.26% / 79.37% 3272688244 3.27G 0.6% 53%:2 > > recursive call 506 0.7% 39.53% 44.97% / 83.92% 1211036806 1.21G 0.2% 10%:1 > > goto 618 0.8% 64.24% 65.37% / 83.57% 702446178 702.45M 0.1% 20%:1 > > null return 800 1.1% 64.62% 56.59% / 77.70% 603952067 603.95M 0.1% 28%:2 > > continue 956 1.3% 63.70% 65.65% / 79.97% 3780303799 3.78G 0.7% 52%:3 > > loop guard 1177 1.6% 56.33% 42.54% / 80.32% 7373601457 7.37G 1.4% 50%:2 > > opcode values positive (on trees) 2020 2.7% 62.38% 64.16% / 84.44% 31695571761 31.70G 6.0% 21%:2 > > loop exit 3293 4.4% 76.19% 87.18% / 88.35% 50377138963 50.38G 9.6% 18%:1 > > loop iterations 4761 6.3% 99.98% 84.27% / 84.27% 73463634555 73.46G 13.9% > > pointer (on trees) 8076 10.7% 56.23% 69.36% / 83.15% 12322099991 12.32G 2.3% > > call 11396 15.1% 64.14% 74.13% / 89.82% 25197949198 25.20G 4.8% 34%:1 > > opcode values nonequal (on trees) 12237 16.3% 70.70% 70.86% / 83.54% 36638772333 36.64G 6.9% > > guessed loop iterations 16760 22.3% 99.78% 91.49% / 91.49% 162952747918 162.95G 30.9% > > > > HEURISTICS BRANCHES (REL) BR. HITRATE HITRATE COVERAGE COVERAGE (REL) predict.def (REL) HOT branches (>10%) > > no prediction 12730 16.9% 39.29% 33.32% / 79.93% 121106031835 121.11G 23.0% > > first match 25261 33.6% 92.17% 88.33% / 88.98% 296652487962 296.65G 56.3% > > DS theory 28333 37.7% 63.03% 72.05% / 85.00% 109563734005 109.56G 20.8% > > combined 75232 100.0% 73.17% 72.32% / 86.08% 527351738575 527.35G 100.0% > > > > Loop count: 37870 > > avg. # of iter: 8444.77 > > median # of iter: 7.00 > > avg. (1% cutoff) # of iter: 174.68 > > avg. (5% cutoff) # of iter: 55.14 > > avg. (10% cutoff) # of iter: 35.21 > > avg. (20% cutoff) # of iter: 26.23 > > avg. (30% cutoff) # of iter: 21.70 > > This is the output data collected without the patch, as can be seen, no difference on "extra loop exit". > But this issue should be fixed. > > > ./contrib/analyze_brprob_spec.py ~/workspace/tests/spec2017/ > > benchspec > HEURISTICS BRANCHES (REL) BR. HITRATE HITRATE COVERAGE COVERAGE (REL) predict.def (REL) HOT branches (>10%) > noreturn call 1 0.0% 100.00% 50.00% / 50.00% 2 2.00 0.0% 100%:1 > Fortran zero-sized array 3 0.0% 66.67% 41.71% / 60.50% 362 362.00 0.0% 100%:3 > loop iv compare 16 0.0% 93.75% 98.26% / 98.76% 279847 279.85k 0.0% 93%:4 > __builtin_expect 35 0.0% 97.14% 78.09% / 78.35% 17079558 17.08M 0.0% > loop guard with recursion 45 0.1% 86.67% 85.13% / 85.14% 6722424412 6.72G 1.3% 74%:4 > extra loop exit 80 0.1% 58.75% 81.49% / 89.21% 438470261 438.47M 0.1% 86%:3 > guess loop iv compare 235 0.3% 80.85% 52.83% / 73.97% 148558247 148.56M 0.0% 47%:3 > negative return 241 0.3% 71.37% 25.33% / 92.61% 250402383 250.40M 0.0% 69%:2 > loop exit with recursion 315 0.4% 74.60% 85.07% / 85.71% 9403136858 9.40G 1.8% 59%:4 > const return 320 0.4% 51.88% 90.45% / 95.63% 925341727 925.34M 0.2% 76%:5 > indirect call 377 0.5% 51.46% 84.72% / 91.14% 2133772848 2.13G 0.4% 69%:1 > polymorphic call 410 0.5% 44.15% 31.26% / 79.37% 3272688238 3.27G 0.6% 53%:2 > recursive call 506 0.7% 39.53% 44.97% / 83.92% 1211036806 1.21G 0.2% 10%:1 > goto 618 0.8% 64.24% 65.37% / 83.57% 702446178 702.45M 0.1% 20%:1 > null return 800 1.1% 64.62% 56.59% / 77.70% 603952067 603.95M 0.1% 28%:2 > continue 956 1.3% 63.70% 65.65% / 79.97% 3780303795 3.78G 0.7% 52%:3 > loop guard 1178 1.6% 56.37% 42.54% / 80.32% 7373601533 7.37G 1.4% 50%:2 > opcode values positive (on trees) 2020 2.7% 62.38% 64.16% / 84.44% 31695571761 31.70G 5.9% 21%:2 > loop exit 3293 4.4% 76.19% 87.18% / 88.35% 50377138963 50.38G 9.4% 18%:1 > loop iterations 4772 6.3% 99.98% 84.27% / 84.27% 74045982111 74.05G 13.8% > pointer (on trees) 8076 10.7% 56.23% 69.36% / 83.15% 12322099991 12.32G 2.3% > call 11396 15.1% 64.14% 74.13% / 89.82% 25197949198 25.20G 4.7% 34%:1 > opcode values nonequal (on trees) 12240 16.2% 70.71% 70.86% / 83.54% 36638772682 36.64G 6.9% > guessed loop iterations 16854 22.4% 99.78% 91.21% / 91.22% 169765264401 169.77G 31.7% > > HEURISTICS BRANCHES (REL) BR. HITRATE HITRATE COVERAGE COVERAGE (REL) predict.def (REL) HOT branches (>10%) > no prediction 12731 16.9% 39.30% 33.32% / 79.93% 121106031963 121.11G 22.6% > first match 25366 33.7% 92.20% 88.24% / 88.88% 304047352001 304.05G 56.9% > DS theory 28337 37.6% 63.03% 72.05% / 85.00% 109563734430 109.56G 20.5% > combined 75342 100.0% 73.21% 72.49% / 86.06% 534746603167 534.75G 100.0% Thank you. So it seems that the problem does not trigger in Spec but I was also wondering if our current predict.def values are anywhere near to reality. THe table reads as follows: - BRANCHES is number of branches the heuristics hit on (so extra loop exit has 80 and therefore we do not have that good statistics on it) - HITRATE is the probability that the prediction goes given direction during the train run. after / is the value which would be reached by perfect predictor (which predict branch to the direction that dominates during train) Extra loop exit is 81% out of 89% so it is pretty close to optimum - COVERAGE is how many times the predicted branch was executed In general the idea is that for most heuristics (wihch can not determine exact value like loop iteraitons) HITRATE values can be put to predict.def so the Dempster-Shafer formula (DS theory) combines the hypothesis sort of realistically (it assumes that all the predictors are staistically independent which they are not). We have HITRATE 67 for extra loop exit which is bit off what we do have in the measured data, but I think our predict.def is still based on spec2006 numbers. So the patch is OK. Perhaps we could experiment with updating predict.def (It does develop even when run across same benchmark suite since early optimizations change - this stage1 I think the threading work definitly affects the situation substantially) Honza > > Loop count: 38058 > avg. # of iter: 8403.32 > median # of iter: 7.00 > avg. (1% cutoff) # of iter: 173.72 > avg. (5% cutoff) # of iter: 54.90 > avg. (10% cutoff) # of iter: 35.20 > avg. (20% cutoff) # of iter: 26.35 > avg. (30% cutoff) # of iter: 21.87 > > > -- > Thanks, > Xionghu