I have measured how much time it takes to run :tier1 w/ and and w/o the fix which removes @build for jdk.test.lib.** classes[1-2]: - w/o 8181391, i.e. w/ @build: real 33m4.624s, user 111m56.758s, sys 6m24.022s. [3] is a breakdown for jtreg actions - w/ 8181391, i.e. w/o @build: real 32m17.259s, user 109m18.236s, sys 6m9.669s. [4] is a breakdown for jtreg actions as you can see there is no much difference in execution time, and the run w/o @build action was even a bit faster. the total time spend on build was lower. hence I'd say removing @build actions does not impact overall execution time. Even if it did, I don't think I'd prefer us to choose isolation and determinism over small performance improvements.
as Ioi and I stated before, removing @build actions did not help in all cases in hotspot. the root cause of this is having @run actions whose target is a class from library, this is identical to have explicit @build action for this class. if this class has dependency on other classes from testlibrary, you can get a testlibrary split into different locations and as a results NCDFE in runtime due to CODETOOLS-7901986. Fortunately, it is not the case for jdk tests, the only test library class which is used in @run is ClassFileInstaller, which does not have any dependencies. Therefore I think removing explicit @build is a more reliable and clearer way to work around current problems and it does not have a big drawback if any. PS measurements were done on my mac 3.1 GHz Intel Core i7, 16 GB 1867 MHz DDR3, jtreg was run w/ "-conc:8 -agentvm" [1] http://scaab055.us.oracle.com:9502/vmsqe/home/iignatye/webrev//8181391/webrev.00/index.html <http://scaab055.us.oracle.com:9502/vmsqe/home/iignatye/webrev//8181391/webrev.00/index.html> [2] https://bugs.openjdk.java.net/browse/JDK-8181391 [3] compile: 826.206 build: 776.955 testng: 5362.58 junit: 640.705 shell: 861.206 main: 6823.19 clean: 0.004 driver: 6.578 [4] compile: 829.317 build: 774.904 testng: 5251 junit: 648.888 shell: 852.658 main: 6686.99 clean: 0.002 driver: 5.973 Thanks, -- Igor > On Jun 2, 2017, at 8:44 AM, Ioi Lam <ioi....@oracle.com> wrote: > > On 6/2/17 6:40 AM, Chris Hegarty wrote: >> On 02/06/17 00:14, Ioi Lam wrote: >>> ... >>> >>> The gem is hidden in the compile.0.jta file. It contains something like: >>> >>> -sourcepath <blahblah>:/jdk/foobar/test/lib:<blahblah> >>> >>> So if my test refers to a class under /test/lib, such as >>> jdk.test.lib.process.ProcessTools, javac will be able to locate it under >>> /jdk/foobar/test/lib/jdk/test/lib/process/ProcessTools.java, and will >>> build it automatically. >>> >>> So really, there's no reason why the test must explicitly do an @build >>> of the library classes that it uses. >> >> Sure, you're relying on the implicit compilation of dependencies >> by javac. Look at the output, where it compiles the library >> classes to. It is part of the classes directory for the >> individual test. That means that the library classes will need >> to be compiled many many times. The @build tag will compile >> the library classes to a common output directory, where they >> can be reused ( unless I'm missing something ). >> >> -Chris. > Yes, @build will compile classes so that they can be reused. But why should > it be the responsibility of every test to do this? > > To reuse my malloc metaphore -- is it reasonable for every program that uses > malloc to explicitly build libc? > > By the way, jtreg arranges the output directory of the test by the directory > they sit in, so > > jdk/test/foo/bar/XTest.java > jdk/test/foo/bar/YTest.java > > will all output their .class files to the same directory. Therefore, the > amount of duplicated classes is not as bad as you might think. We've been > omitting the @build tags in the hotspot tests and we haven't seen any > problems. > > - Ioi
