RFR: 8164525: Re-examine zero form link time pre-generation
Claes Redestad
claes.redestad at oracle.com
Sun Aug 21 09:59:43 UTC 2016
Hi,
the previous attempt to pre-generate zero LambdaForms[1] failed
due to the current implementation surprisingly emitting placeholder
constants to link to the associated identity function (which we can't
properly handle when pre-generating forms) rather than simply emitting
the corresponding bytecode directly (iconst_0, etc).
It turns out the setup of Zero and Identity functions aren't being made
intrinsic, which appears to be an oversight. Fixing this allows for more
optimal bytecode[1].
While this fix could have positive performance implications in general,
it also makes the startup optimization from JDK-8164451 re-apply
without issue (no test failures running RBT --test jdk/test/:tier1).
Bug: https://bugs.openjdk.java.net/browse/JDK-8164525
Webrev: http://cr.openjdk.java.net/~redestad/8164525/webrev.01/
Thanks!
/Claes
[1] https://bugs.openjdk.java.net/browse/JDK-8164451
[2] Before:
static java.lang.Object zero_000_L(java.lang.Object);
descriptor: (Ljava/lang/Object;)Ljava/lang/Object;
flags: ACC_STATIC
Code:
stack=2, locals=1, args_size=1
0: ldc #14 // String
CONSTANT_PLACEHOLDER_1 <<(Object)Object :
identity_000_L=Lambda(a0:L,a1:L)=>{a1:L}\n& Class=SimpleMethodHandle>>
2: checkcast #16 // class
java/lang/invoke/MethodHandle
5: aconst_null
6: invokevirtual #19 // Method
java/lang/invoke/MethodHandle.invokeBasic:(Ljava/lang/Object;)Ljava/lang/Object;
9: areturn
RuntimeVisibleAnnotations:
0: #8()
1: #9()
2: #10()
After:
static java.lang.Object zero_L(java.lang.Object);
descriptor: (Ljava/lang/Object;)Ljava/lang/Object;
flags: ACC_STATIC
Code:
stack=1, locals=1, args_size=1
0: aconst_null
1: areturn
RuntimeVisibleAnnotations:
0: #8()
1: #9()
2: #10()
More information about the core-libs-dev
mailing list