Request for reviews (S): 7011839: JSR 292 turn on escape analysis when using invokedynamic

[Christian Thalinger]

On Jan 13, 2011, at 12:00 AM, John Rose wrote:
> On Jan 12, 2011, at 5:46 AM, Christian Thalinger wrote:
> 
>> http://cr.openjdk.java.net/~twisti/7011839/webrev.01/
> 
> The arg_size computation is very fragile.  To start with, the arg_size value should be computed in one place.
> 
> As the code is written, arg_base points at the implicit argument to indy, which (I think) is not really on the stack.
> 
> I suggest this code:
> 
> 234   // compute size of arguments
> 235   int arg_size = target->arg_size();
>          if (code == Bytecodes::_invokedynamic) {
>            assert(!target->is_static(), "receiver explicit in method");
>            arg_size--;  // implicit, not really on stack
>          }
> 236   if (!target->is_loaded() && code == Bytecodes::_invokestatic) {
>            assert(!target->is_static(), "receiver explicit in unloaded method");
> 237     arg_size--;
> 238   }

I like that better.  I'd like to keep the invokedynamic check as I couldn't reproduce any escaping case with an invokedynamic call site (using Indify).  I think the reason is because they all bail before out as native call sites.

> 
> The second assertion is for good measure.  The bare "arg_size--" makes me nervous.

We can't use that assert since ciMethod::is_static uses check_is_loaded which asserts on is_loaded.

http://cr.openjdk.java.net/~twisti/7011839/webrev.02/

-- Christian