RTC Thumb2 JIT enhancements

[Andrew John Hughes]

On 11 March 2010 11:23,  <ed at camswl.com> wrote:
> Hi folks,
>
> The following are some enhancements I have made to the T2JIT over the past
> month.
>
> I have tested the changes with various benchmarks, with Think Free Office,
> and with the hotspot and lang portions of JTreg which all pass. The jdk
> portion is still running.
>
> Unfortunately the diffs are quite large so I am submitting the diffs to
> the list as 4 separate sets of diffs.
>
> Has the b18 release been split off yet? I would prefer these changes not
> to go into the b18 release as they have not had sufficient testing yet.
>

Not yet.  I'd suggest you wait until it does fork in that case.  A 1.8
branch will be created when that happens.  I presume doko is handling
this.

> Regards,
> Ed.
>
> - Support for frameless compilation
>
>        Local variables and interpreter state are now addressed from Java SP
>        rather than separate 'istate' (R8) register. R8 now free for register
>        allocation giving 6 allocatable registers instead of 5.
>
> - Support for compilation of methods containing monitorenter / monitorexit
>
>        monitorenter / monitorexit moved from asm to C so they can be called
>        by JIT code compilation of monitorenter / monitorexit containinng
>        methods uses frame adressing because stack can expand dynamically to
>        accomodate monitors.
>
> - Interpreter caches 'thread' in a register
>
>        Interpreter reworked to cache 'thread' in a register. Involved
>        removal of one 'tmp' register in interpreter to make room for
>        'thread' register. Interpreter and JIT are now consistent in register
>        usage.
>
> - Compression of JIT 'header'
>
>        Removed the OSR entry from the JIT header. Previously, each JITted
>        method had an individual piece of code in the header to load the
>        appropriate registers from locals and the jump to the OSR entry
>        point. This is changed so the the OSR caller (the interpreter), uses
>        the 'regusage' array in the JIT header to load the correct registers
>        and then just jumps directly to the entry point in the method.
>        Saving is 24 bytes per method.
>
> - Backwards branch targets now doubleword aligned.
>
>        Pads with no ops before a backwards branch target to ensure backwards
>        branch target is 64 bit aligned. Small performance benefit.
>
> - Addition of OSPACE flag
>
>        Setting environment variable T2_OSPACE will cause the JIT to optimise
>        for space as opposed to time. OSPACE does the following.
>
>        o Tweaks the register allocation to allocate register based on space
>        o Enables tailing of returns. IE does a branch to the return.
>        o Disables doubleword alignment of backwards branch targets
>        o Disables inlining invoke virtual/static/special wrappers
>
>        OSPACE has a small performance effect on benchmarks (~10% on ECM) but
>        has no noticeable effect on large applications.
>
>        Default is compile for time (OSPACE disabled)
>
> - Removed a pass from the compiler
>
>        The JIT had an additional pass to detect 'zombie' code (ie code which
>        could be, but has never been called). Zombie code removal was
>        disabled in prior versions because it was unreliable, however the
>        additional pass was still there.
>
> - Removed vestigal and debug code
>
>        Removed #if 0 code and debug code which was no longer functional
>



-- 
Andrew :-)

Free Java Software Engineer
Red Hat, Inc. (http://www.redhat.com)

Support Free Java!
Contribute to GNU Classpath and the OpenJDK
http://www.gnu.org/software/classpath
http://openjdk.java.net

PGP Key: 94EFD9D8 (http://subkeys.pgp.net)
Fingerprint: F8EF F1EA 401E 2E60 15FA  7927 142C 2591 94EF D9D8