[aarch64-port-dev ] RFR: Add support for pipeline scheduling

[Edward Nevill]

Hi,

The following webrev is a first pass at adding pipeline scheduling to JDK8. The pipeline implementation is modeled on A53.

http://openjdk.linaro.org/webrev/141031/hotspot/

- The model only models 4 stages ISS1, EX1, EX2 and WR. It does not model the additional 4 stages in the fetch and decode stages of the pipeline because these are fixed and therefore make no difference to the scheduling.
- The A53 pipeline is dual issue and this is modeled by having 2 dummy resources 'INS0' and 'INS1' and a combined resource 'INS01' which means an instruction can issue as either instruction 0 or instruction 1. The resource INS0, INS1 or INS01 is used in the ISS stage.
- The different A53 pipelines (ALU0, ALU1, MAC, DIV, STORE & BRANCH) are modeled as resources (ALU, MAC, DEV, BRANCH, LDST) which are used in the final stage of the pipeline - ie. the stage where the result could be forwarded.
- I have made no attempt to model Late and Early pipeline stage. IE if a result can be available in Late_ISS I require it to be available in ISS and similarly if a result can be forwarded in Early_WR I say it is available at WR.
- I have not modeled the Neon pipeline. These are simply fixed latency.

The following is the performance improvement I see on A53 on a selection of benchmarks.

Grinderbench

Parallel: +16%
kXML: +1%
PNG: +2%
Chess: +4%
Crypto: +6%

SpecJVM2008

compiler.compiler: +5%
compress: +3%
crypto.aes: +26%
crypto.rsa: +2%
crypto.signverify: +6%
derby: -1%
mpegaudio: +5%
scimark.fft.large: +12%
scimark.lu.large: +0%
scimark.sor.large: -1%
scimark.sparse.large: +28%
scimark.fft.small: +2%
scimark.lu.small: +8%
scimark.sor.small: -1%
scimark.sparse.small: +0%
scimark.monte_carlo: +9%
serial: +1%
sunflow: +3%
xml.transform: +5%
xml.validation: +2%

There is quite a large variation on these results and I would be suspect about some of the larger improvements (+16%, +26%, +28%) however it does seem to offer a consistent improvement.

In the webrev I have turned the default for pipeline scheduling ('OptoScheduling'). It may be that we want to refine this and only turn it on for In-Order cores (not sure what you do about Big-Little). However for the moment I would like to make it the default so the overnight tests will run with it switched on.

OK to push to JDK8?

All the best,
Ed.