RTC Thumb2 JIT methods containing exc handlers (take 2)

[Xerxes Rånby]

ed at camswl.com wrote:
> Hi folks,
>
> Well, overnight testing with the previous proposed patch to JIT methods containing
> exception handlers showed up a few problems.
>
> 1) With nested try catch the exception could be thrown to the outermost catch
> instead of the one it actually occured in.
>
> 2) The function Thumb2_lr_to_bci returned 0 if it could not find a matching
> entry in the exception table. Unfortunately 0 is a valid byte code index.
> Changed to return -1, and calling code updated to catch -1.
>
> 3) When an exception occurs some of the local variable may be cached in ARM
> registers because they have been register allocated. The exception handling
> code continues in the interpreter so these need to be decached. This is done
> by having a register map in the compiled method header to enable the
> JITs exception handle do determine which locals were in which registers.
>
> Other changes:
>
> - Copyright notices updated to 2010 and added to thumb2.cpp
> - assertions turned off in PRODUCT build
> - Compilation summary and statistics output to stderr instead of stdout
>
> I have put 3 binaries of the Thumb2 JIT at
>
> http://mint.camswl.com/libjvm-1.so.gz
> http://mint.camswl.com/libjvm-100.so.gz
> http://mint.camswl.com/libjvm-10000.so.gz
>
> The difference between the 3 binaries is that the -1 version has the compile
> threshold set to 1, the -100 has it set to 100 etc. The default for a PRODUCT
> build is to set it to 10000.
>
> It is possible to change the compile threshold with the -XX:CompileThreshold=N
> flag, however, sometimes it is difficult to do this if you are say running a
> plugin within a broswer.
>
> The most useful one to test is the -100 version. The -1 version 'compiles the
> world' so it quickly runs out of buffer space and simply stops compiling. In
> addition when you run -1 none of the invokes are resolved when compiling. In
> this case (for an unresolved invoke) the JIT simply calls out to the interpreter.
>
> OK to commit?
>
> Regards,
> Ed.
>   
Looks ok.
It compiles, runs, and bootstraps OpenJDK builds quickly on armv7 hardware,
please commit!

Xerxes