{"thread":{"id":"27547","subject":"[PATCH] Profile feedback based optimization","startedAt":"2011-06-04T15:38:21Z","lastAt":"2011-06-04T15:38:21Z","messageCount":1,"participants":["Cesar Eduardo Barros"],"isPatch":true,"patchVersion":1,"patchTotal":null},"messages":[{"id":"169302","messageId":"1307201901-8298-1-git-send-email-cesarb@cesarb.net","threadId":"27547","inReplyTo":null,"subject":"[PATCH] Profile feedback based optimization","fromName":"Cesar Eduardo Barros","fromEmail":"cesarb@cesarb.net","sentAt":"2011-06-04T15:38:21Z","receivedAt":"2011-06-04T15:38:21Z","isPatch":true,"sender":{"key":"cesarb@cesarb.net","avatar":"https://avatars.githubusercontent.com/u/31324?v=4"},"body":"gcc has options to instrument the generated executable and use the\ncollected information to better optimize the generated code.\n\nTo use the profile feedback options, you have to compile twice. After\nthe first compilation, you have to run the executable with a\nrepresentative set of data to generate the statistics for the second\ncompilation.\n\nA benchmark was chosen as the representative set of data because a\nbenchmark focuses on measuring the code speed. A good benchmark should\nbe designed considering how git is commonly used, emphasizing the most\nused parts, but without neglecting the less used parts.\n\nTo simplify the patch and focus on the Makefile rules, a dummy benchmark\nwas used, which does only a few simple operations a hundred times.\n\nEven with just that simple benchmark, it seems to already have some\neffect. These were the results of a test run of \"./git fsck --full\" on\nthe git repository, recorded via perf stat, ran twice without profile\nfeedback and then twice with profile feedback:\n\nWithout profile feedback:\n\n Performance counter stats for './git fsck --full':\n\n 84159,595732 task-clock-msecs # 0,998 CPUs\n 8470 context-switches # 0,000 M/sec\n 7 CPU-migrations # 0,000 M/sec\n 263682 page-faults # 0,003 M/sec\n 179028259343 cycles # 2127,247 M/sec\n 222624952497 instructions # 1,244 IPC\n 22651880335 branches # 269,154 M/sec\n 1798298585 branch-misses # 7,939 %\n 1380706193 cache-references # 16,406 M/sec\n 588704688 cache-misses # 6,995 M/sec\n\n 84,360035244 seconds time elapsed\n\n Performance counter stats for './git fsck --full':\n\n 84207,836012 task-clock-msecs # 0,998 CPUs\n 8450 context-switches # 0,000 M/sec\n 2 CPU-migrations # 0,000 M/sec\n 211878 page-faults # 0,003 M/sec\n 179131407549 cycles # 2127,253 M/sec\n 222395625483 instructions # 1,242 IPC\n 22614432765 branches # 268,555 M/sec\n 1798786420 branch-misses # 7,954 %\n 1384878082 cache-references # 16,446 M/sec\n 594115144 cache-misses # 7,055 M/sec\n\n 84,404733868 seconds time elapsed\n\nWith profile feedback:\n\n Performance counter stats for './git fsck --full':\n\n 83576,534090 task-clock-msecs # 0,998 CPUs\n 8610 context-switches # 0,000 M/sec\n 10 CPU-migrations # 0,000 M/sec\n 211882 page-faults # 0,003 M/sec\n 177777232502 cycles # 2127,119 M/sec\n 200278330301 instructions # 1,127 IPC\n 21014495344 branches # 251,440 M/sec\n 1659583500 branch-misses # 7,897 %\n 2297603485 cache-references # 27,491 M/sec\n 1314536406 cache-misses # 15,729 M/sec\n\n 83,783009768 seconds time elapsed\n\n Performance counter stats for './git fsck --full':\n\n 83288,834113 task-clock-msecs # 0,998 CPUs\n 8365 context-switches # 0,000 M/sec\n 5 CPU-migrations # 0,000 M/sec\n 174403 page-faults # 0,002 M/sec\n 177173914267 cycles # 2127,223 M/sec\n 200234762879 instructions # 1,130 IPC\n 21006452953 branches # 252,212 M/sec\n 1659600980 branch-misses # 7,900 %\n 2298086212 cache-references # 27,592 M/sec\n 1309056989 cache-misses # 15,717 M/sec\n\n 83,482945132 seconds time elapsed\n\nSigned-off-by: Cesar Eduardo Barros \n---\n Makefile | 20 ++++++++++++++++++++\n bench.sh | 45 +++++++++++++++++++++++++++++++++++++++++++++\n 2 files changed, 65 insertions(+), 0 deletions(-)\n create mode 100755 bench.sh\n\ndiff --git a/Makefile b/Makefile\nindex e40ac0c..8eaa8a0 100644\n--- a/Makefile\n+++ b/Makefile\n@@ -2486,3 +2486,23 @@ cover_db: coverage-report\n \n cover_db_html: cover_db\n \tcover -report html -outputdir cover_db_html cover_db\n+\n+### Profile feedback based optimization\n+#\n+.PHONY: benchmark profile-generate profile-use profile-optimized-build\n+\n+benchmark:\n+\ttime ./bench.sh\n+\n+profile-generate:\n+\t$(MAKE) clean coverage-clean\n+\t$(MAKE) CFLAGS=\"$(CFLAGS) -fprofile-generate\" LDFLAGS=\"$(LDFLAGS) -fprofile-generate\" all\n+\n+profile-use:\n+\t$(MAKE) clean\n+\t$(MAKE) CFLAGS=\"$(CFLAGS) -fprofile-use\" LDFLAGS=\"$(LDFLAGS) -fprofile-use\" all\n+\n+profile-optimized-build:\n+\t$(MAKE) profile-generate\n+\t$(MAKE) -j1 benchmark\n+\t$(MAKE) profile-use\ndiff --git a/bench.sh b/bench.sh\nnew file mode 100755\nindex 0000000..939f3b3\n--- /dev/null\n+++ b/bench.sh\n@@ -0,0 +1,45 @@\n+#!/bin/bash\n+set -eu\n+\n+BENCHDIR=\"$PWD/bench\"\n+GIT_BUILD_DIR=\"$BENCHDIR/..\"\n+export PATH=\"$GIT_BUILD_DIR/bin-wrappers:$PATH\"\n+\n+mkdir \"$BENCHDIR\"\n+cd \"$BENCHDIR\"\n+\n+git --version\n+exec > bench.log 2>&1\n+\n+for count in `seq 100`; do\n+\n+mkdir \"$BENCHDIR/test\"\n+cd \"$BENCHDIR/test\"\n+git init\n+echo a > a\n+echo b > b\n+git add a b\n+git commit -m 'Commit 1'\n+echo c > c\n+echo d > d\n+git add c d\n+git commit -m 'Commit 2'\n+git checkout -b branch\n+echo e > e\n+echo f > f\n+git add e f\n+git commit -m 'Commit 3'\n+git checkout -\n+echo g > g\n+echo h > h\n+git add g h\n+git commit -m 'Commit 4'\n+git merge branch\n+git fsck --full\n+cd \"$BENCHDIR\"\n+rm -rf test\n+\n+done\n+\n+cd ..\n+rm -rf \"$BENCHDIR\"\n-- \n1.7.4.4\n"}]}