Jan Hubicka wrote:
Hi,
this patch resulted from attempt to solve regression we have in
gdb.opt/inline-locals.exp gdb testsuite and also problems with fact that when
clonning function by ipa-cp we lose any information on function argument.
(and yes, it solves it)
<...>
I know that this is one of problems being solved on debug branch, but
there is problem with DEBUG_INSN approach and scalability for C++
programs. In tramp3d and similar testcases we have >10 inlined function
scopes for every instruction, so adding about 20 DEBUG_INSN markers into
stream for every instruction would mean that we will need 20 times more
declaratoins, statements, RTL instruction, track 20 times more variable
locations and we would have to do that -g or not. So if we want to have
DEBUG_INSN approach, those aproaches should combine themselves well:
inline substitution tracking will handle most cases chaply and
DEBUG_INSN will help to keep the substitutions valid.
I'm also experimenting with an implementation prototype right now that
overloads the source_location in expressions. An enhanced debuglocus
tracks information regarding variables similar to kind of information
DEBUG_INSNs provide, but without the additional instruction overhead.
It shows promise at least within the tree-ssa infrastructure, but its
still a hair too early to tell if it will work as a general solution.
Its a few weeks from being able to compare to the example you have here
to see how coverage compares.
Andrew
