Re: need to focus on java performance?

[Bryce McKinlay <mckinlay at redhat dot com>]

David Daney wrote:
> I imagine that most of the stack traces are needed for security 
> checks, and not exceptions.

Most applications (excluding applets) do not use security managers, and 
thus do not perform many security checks (there's also some runtime 
access checking that occurs as part of reflection). Beyond EH/stack 
traces, most of the stack inspection that occurs in regular applications 
is actually to determine the calling class to maintain classloader 
semantics when loading classes and resources. Andrew recently made an 
improvement here that should reduce these checks significantly: the 
compiler now passes the calling class directly to certain common 
functions that need it.
> If stack traces are really done so much, would it make any sense to 
> compile all code with a frame pointer, and walking those instead of 
> using the DWARF unwinder?  In the recent past that was done on many 
> archetectures.  Perhaps changing it was a mistake.

That is certainly something to consider. Currently the DWARF data is the 
only way to properly map an IP address to a function, however if the 
compiler also produced a length-of-function field as part of the 
metadata, we could maintain our own data structure for this mapping and 
avoid the need to call into the unwinder to determine the function start 
address.

Whether this would be worth it, compared to the loss of code 
optimization from losing a register to the stack pointer, is unclear. I 
suspect it varies a lot by application, and platform.

Bryce