hg: mlvm/mlvm/hotspot: indy.compiler: First commit, not enabled in series file.

[Christian Thalinger]

John Rose wrote:
> I just pushed an update which fixed the stack pointer smashing problem  
> we ran into at JavaOne.  Since C2 stack frames do not use EBP for a  
> frame pointer (EBP is a regular register) there is no way to  
> checkpoint the current ESP just before a method handle call.  This is  
> unfortunate, since method handles can rearrange the stack as they  
> insert or delete arguments.
> 
> The solution is pretty simple:  Use EBP itself to hold a checkpointed  
> value of ESP.  (It is already the case that all calling sequences save  
> and restore EBP, for the sake of the interpreter and its adapters.)   
> This means a method handle call looks like this:
>    mov rbp, rsp
>    call MH.invoke.from_compiled_entry
>    mov rsp, rbp
> 
> That's not too bad, except of course that we need to inline these guys  
> routinely.
> 
> (I had considered more exotic solutions, like inventing new frame  
> types or checkpointing the ESP on a thread-local variable.  This  
> solution is pleasantly simple.)
> 
> The stack frame walking code had to be per-kludged (i.e., another  
> kludge on top of kludges) to know about this convention.  It wants a  
> clean-up.  And there's a deoptimization bug in there somewhere; yuck.
> 
> The resulting JVM runs the 1600 unit tests of MethodHandlesTest with  
> CompileThreshold set to 1 (usually it's 10000). It almost runs in - 
> Xcomp mode also, but crashes very late trying to deoptimize in a  
> catchException combinator.
> 
> In my previous push I put the compiler patch into the series file, but  
> it only applies if the "testable" guard is removed.  I suppose it's  
> actually testable now.
> 
> Your turn, Christian!

Cool!  Thanks a lot for fixing that problem!

-- Christian