Unfortunately, I don't think @DontInline on an empty method is sufficient
here. If other code is relying on @DontInline for the same purpose then
we might need to reexamine that code. My understanding from discussing
with other compiler engineers is that using a native method is the safest
technique that the compilers can't see through. The problem with
@DontInline is that C2 looks at the bytecodes of the target method, even
if it isn't inlined (see BCEscapeAnalyzer and the EstimateArgEscape flag).
There may be a way to make it work, but that would require more
investigation, and I'm not sure the benefit outweighs the risk.
dl
On 12/15/18 6:48 AM, Claes Redestad wrote:
Hi Dean,
to avoid escape analysis-eliminated allocations in the past @DontInline
has been sufficient. This means a simpler patch (no changes to native
code needed - added assertions notwithstanding) and passes your tests
with C2 (it'd concern me if Graal's EA sees through this trick, as it
might break some existing places where DontInline is used to this
effect):
/**
* The value needs to be physically located in the frame, so that it
* can be found by a stack walk.
*/
@Hidden
@DontInline
private static void ensureMaterializedForStackWalk(Object o) {}
Thanks!
/Claes
On 2018-12-15 01:59, dean.l...@oracle.com wrote:
https://bugs.openjdk.java.net/browse/JDK-8214583
http://cr.openjdk.java.net/~dlong/8214583/webrev
This change includes two new regression test that demonstrate the
problem, and a fix that allows the tests
to pass.
The problem happens when the JIT compiler's escape analysis
eliminates the allocation of the AccessControlContext object passed
to doPrivileged. The compiler thinks this is safe because it does
not see that the object "escapes". However, getContext needs to be
able to find the object using a stack walk, so we need a way to tell
the compiler that it does indeed escape. To do this we pass the
value to a native method that does nothing.
Microbenchmark results:
jdk12-b18:
Benchmark Mode Cnt Score Error Units
DoPrivileged.test avgt 25 255.626 ± 6.446 ns/op
DoPrivileged.testInline avgt 25 250.968 ± 4.975 ns/op
jdk12-b19:
Benchmark Mode Cnt Score Error Units
DoPrivileged.test avgt 25 5.689 ± 0.001 ns/op
DoPrivileged.testInline avgt 25 2.765 ± 0.001 ns/op
this fix:
Benchmark Mode Cnt Score Error Units
DoPrivileged.test avgt 25 5.020 ± 0.001 ns/op
DoPrivileged.testInline avgt 25 2.774 ± 0.025 ns/op
dl
