RFR: 8274983: Pattern.matcher performance regression after JDK-8238358

[Dean Long]

On Thu, 18 Nov 2021 06:13:31 GMT, Xin Liu <xliu at openjdk.org> wrote:

> The root cause of the C1 regression is that some regex generate multiple classes which all implement
> an interface. In SlowStartupTest.java, the following **invokeinterface** happens frequently with different receivers. the target is a private interface method. 
> 
> 
> 9: invokeinterface #25,  3           // InterfaceMethod java/util/regex/Pattern$BmpCharPredicate.lambda$union$2:(Ljava/util/regex/Pattern$CharPredicate;I)Z
> 
> 
> This patch allows c1 to generate the optimized virtual call for invokeinterface
> whose targets are the private interface methods.
> 
> Before JDK-823835, LambdaMetaFactory generates invokespecial in this case. Because the private
> interface methods can not be overrided, c1 generates the optimized virtual call. After JDK-823835,
> LambdaMetaFactory generates invokeinterface instead. C1 generates the regular virtual call because
> it can not recognize the new pattern. If a multiple of subclasses all implement a same interface,
> it is possible that they trash the IC stub using their own concrete klass in runtime.
> 
> Optimized virtual call uses relocInfo::opt_virtual_call_type(3), It will call VM
> 'resolve_opt_virtual_call_C' once and resolve the target to the VEP of the nmethod.
> Therefore, this patch can prevent the callsite from trashing.
> 
> Before this patch,   SlowStartupTest had 38770 times  _resolve_invoke_virtual_cnt and 38695  _handle_wrong_method_cnt per 10k iterations. To dump `C1Statistics`, we use fastdebug build for comparison.
> 
> 
> $java -XX:TieredStopAtLevel=1 -XX:+PrintC1Statistics SlowStartupTest 1
> Executed 10000 iterations in 736ms
> C1 Runtime statistics:
>  _resolve_invoke_virtual_cnt:     38770
>  _resolve_invoke_opt_virtual_cnt: 186
>  _resolve_invoke_static_cnt:      44
>  _handle_wrong_method_cnt:        38695
>  _ic_miss_cnt:                    35
> 
> 
> With this patch,  only 1 _handle_wrong_method_cnt is triggered but we have 3 more `_resolve_invoke_opt_virtual_cnt` events instead.  The total runtime reduces from 736ms to 9ms. 
> 
> 
> $java -XX:TieredStopAtLevel=1 -XX:+PrintC1Statistics SlowStartupTest 1
> Executed 10000 iterations in 9ms
> C1 Runtime statistics:
>  _resolve_invoke_virtual_cnt:     77
>  _resolve_invoke_opt_virtual_cnt: 189
>  _resolve_invoke_static_cnt:      45
>  _handle_wrong_method_cnt:        1
>  _ic_miss_cnt:                    39
> 
> 
> Codegen wise,  before the patch, C1 generates LIR for the invokeinterface whose target is a private interface methoda as follows.
> 
> __bci__use__tid____instr____________________________________
> . 1    0    v2     a1.invokeinterface()
>                    InvokePrivateInterfaceMethod$I.bar()V
> . 6    0    v3     return
> 
> 
> With this patch, C1 generates LIR as follows. it first check a1 is a subtype of `InvokePrivateInterfaceMethod$I`. if so, an optimized virtual call is generated. The callsite will be fixed up once and only one time in runtime. 
> 
> __bci__use__tid____instr____________________________________
> . 1    1    a2     checkcast(a1) InvokePrivateInterfaceMethod$I
>                    stack [0:a1]
> . 1    0    v3     a2.invokeinterface()
>                    InvokePrivateInterfaceMethod$I.bar()V
> . 6    0    v4     return

It looks like this does what we want for private interface methods, but I'm wondering if we handle all combinations of private/final and invokevirtual/invokeinterface, or are we missing some cases where can_be_statically_bound() would return true?

-------------

PR: https://git.openjdk.java.net/jdk/pull/6445