RFR: 8352415: x86: Tighten up template interpreter method entry code

[Aleksey Shipilev]

On Wed, 19 Mar 2025 13:44:40 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

> Interpreter performance is the still important for faster startup, since it would carry application until compilers kick in. After looking at Leyden scenarios in Xint mode, I believe incremental improvements are possible in template interpreter to make it faster.
> 
> One of those improvements is tightening up method entry code. Profiling shows the hottest path in the whole ordeal for non-native methods is resolving the Java mirror to store the GC root for currently executing Method*. It involves 4-5 chained memory accesses, which incurs significant latency.
> 
> We can massage the code to reuse some memory accesses and also spread them out to allow more latency-hiding hardware mechanisms to kick in.
> 
> Additional testing:
>  - [x] Ad-hoc `-Xint` benchmarks
>  - [ ] Linux x86_64 server fastdebug, `all`

Motivational improvements on 5950X, about 1.8% faster interpreted code.


Benchmark 1: build/linux-x86_64-server-release/images/jdk/bin/java -Xms1g -Xmx1g -XX:+UseSerialGC \
   -Xint -cp JavacBenchApp.jar JavacBenchApp 1

# Before
  Time (mean ± σ):      1.533 s ±  0.013 s    [User: 1.479 s, System: 0.051 s]
  Range (min … max):    1.517 s …  1.551 s    10 runs

# After
  Time (mean ± σ):      1.506 s ±  0.012 s    [User: 1.451 s, System: 0.051 s]
  Range (min … max):    1.493 s …  1.528 s    10 runs

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

PR Comment: https://git.openjdk.org/jdk/pull/24114#issuecomment-2736705387