Re: RFR (12) JDK-8218025: disable pop_frame and force_early_return caps for Graal

[serguei.spit...@oracle.com]

On 1/30/19 20:46, serguei.spit...@oracle.com wrote: Hi Alex, Vladimir I. also mentioned the same as Igor that it is safe to use UseJVMCICompiler under "ifdef INCLUDE_JVMCI". One example from the runtime/thread.cpp: #if INCLUDE_JVMCI   bool force_JVMCI_intialization = false;   if (EnableJVMCI) {     // Initialize JVMCI eagerly when it is explicitly requested.     // Or when JVMCIPrintProperties is enabled.     // The JVMCI Java initialization code will read this flag and     // do the printing if it's set.     force_JVMCI_intialization = EagerJVMCI || JVMCIPrintProperties;     if (!force_JVMCI_intialization) {       // 8145270: Force initialization of JVMCI runtime otherwise requests for blocking       // compilations via JVMCI will not actually block until JVMCI is initialized.       force_JVMCI_intialization = UseJVMCICompiler && (!UseInterpreter || !BackgroundCompilation);     }   } #endif One more example from prims/jni.cpp: #if INCLUDE_JVMCI     if (EnableJVMCI) {       if (UseJVMCICompiler) {         . . .       }     } #endif Sorry, I've accidentally pressed 'Send' button. Continuing ... So, the fix needs to be more like this: + // Workaround for 8195635: + // disable pop_frame and force_early_return capabilities with Graal + #if INCLUDE_JVMCI + if (!(EnableJVMCI && UseJVMCICompiler)) { jc.can_pop_frame = 1; jc.can_force_early_return = 1; + } + #endif Not sure, if the check for EnableJVMCI can be removed above. Otherwise, it looks good to me. Also, I think, it is good to have updates for the ProblemList-graal.txt files. Thanks, Serguei On 1/30/19 17:34, Igor Ignatyev wrote: Hi Alex, UseJVMCICompiler is declared only if INCLUDE_JVMCI is defined, so your current patch will break builds where it's undefined. the rest (esp. problem list ;) ) looks good to me. adding hotspot-compiler alias. Thanks, -- Igor On Jan 30, 2019, at 5:27 PM, Alex Menkov <alexey.men...@oracle.com> wrote: Hi all, Please review a fix for tck-red bug: https://bugs.openjdk.java.net/browse/JDK-8218025 webrev: http://cr.openjdk.java.net/~amenkov/tck_red_disable_caps/webrev/ ForceEarlyReturn and PopFrame JCK tests intermittently fail with Graal. Real fix for the issue is too risky for jdk12, so we have to disable pop_frame and force_early_return capabilities running with Graal (the capabilities are optional). Currently Graal is the only compiler, so the fix checks if JVMCI compiler is enabled. JCK test passes with disabled capabilities. A number of hotspot tests do not check if the capabilities are enabled (as hotspot is expected to support all capabilities) and fail with Graal - they are problem-listed until the real problem (JDK-8195635) is resolved. --alex