RFR(L) 8223320: [AOT] jck test api/javax_script/ScriptEngine/PutGet.html fails when test classes are AOTed
vladimir.kozlov at oracle.com
Thu May 30 21:15:33 UTC 2019
CCing to runtime group too.
So you went hard and correct way ;-)
deoptimization.cpp next #ifdef sequence is strange since we support only 64-bit on SPARC:
+ jlong res = (jlong)low->get_int();
Also the code here is guarded by INCLUDE_JVMCI but it should be applicable to AOT code too. Right? May be || INCLUDE_AOT?
Please, add more comment. For example add one in aotLoader.cpp for initialize_box_caches() to explain why we need to
eager initialize caches for AOT.
On 5/30/19 12:33 PM, Igor Veresov wrote:
> Graal models boxing (a call to valueOf()) as a BoxNode. If scalarized, it is encoded in the debug info as an allocation
> of a box object. However, for certain ranges of values the box object has to come from caches. The reason is that for
> these values JLS guarantees the identity of the boxes.
> The fix essentially propagates the information on whether the Box is a result of Box.valueOf() or new Box() to the
> deoptimization machinery that checks if the object is in the range that should be in a cache and gets it from there
> instead of allocating it.
> Mach5: tier1-6, tier2-6 with Graal
> Webrev: http://cr.openjdk.java.net/~iveresov/8223320/webrev.00/
> I’d like to push all this into JDK13 first and then follow up with a change to the upstream Graal.
More information about the hotspot-compiler-dev