[Mono-devel-list] Trampolines...
Willibald Krenn
Willibald.Krenn at gmx.at
Mon Feb 21 15:26:20 EST 2005
Hi guys!
I'm a little bit lost on trampolines (esp AMD64): Currently, if a method
is compiled the first time within the magic service routine (I'll refer
to this trampoline 'servic routine' for better separation from the
others) it is checked if the caller can be patched to directly jump to
the new code. At the same time, the stub (I'll refer to the 'first'
trampoline 'stub', because it just emits a call to the real trampoline
which then calls the C-service routine) is patched not to call into the
trampoline anymore, but to directly call the newly emitted method.
(While I can see the benefits of doing this, it also means leaving any
further possibilities of avoiding the stub-jump by inlining the new
method's address in other callers unused..)
Is this subject to change?
The reason I'm asking is for stub-reference counting: Upon
recompilation, the method is changed to a stub that calls a new
trampoline and ends in a new service routine which will look up the
recompiled code position and patch the callers accordingly. Thereby
decreasing some refcount in the stub. Goal is to free the space used by
the first-time compilation of a method (where the stub is located after
recompilation)..
If the magic service routine won't be changed, it means that all
callers, except one will go through the stub - meaning that a newly
compiled method has at maximum 2 different callers (one is the stub, one
is the patched call). If the service routine will be changed, this
assumtion does not hold anymore..
BTW: Recompiling/Patching of virtual and non-virtual methods is working
here - just the refcount part does not work as it should... But I'll
change that :-)
Be warned: If this is fixed, I'll clean up the code and post it for
comments here.
Thanks,
Willi
More information about the Mono-devel-list
mailing list