Closed GoogleCodeExporter closed 8 years ago
Can we automate the workaround somehow? Can we forcibly add this to the list of
command-line flags LLVM is parsing in
Python/global_llvm_data.cc:_PyLlvm_Init(), or
is there an LLVM option we can toggle programmatically?
Original comment by collinw
on 8 Jun 2009 at 6:06
Yes, you can get it passed in that way, but the way the code is written right
now
makes it tricky. We could read the env variable, append " --disable-fp-elim"
to it,
then write it back to the env variable, or we could read the variable, properly
split
it on spaces (blech), and pass that off to ParseCommandLineOptions.
The code to toggle that option at runtime isn't in LLVM right now, and Jeff
says that
if you just go in there and flip the value it has some assertions that will
complain
that the value was set twice.
Original comment by reid.kle...@gmail.com
on 8 Jun 2009 at 6:14
Original comment by collinw
on 10 Jun 2009 at 3:11
Original comment by collinw
on 13 Jul 2009 at 5:50
I committed the LLVM side of the GDB JIT debugging interface in r82418 of LLVM.
We
don't have to update LLVM right away to start using it, because I think right
now
unladen builds against LLVM TOT. All you have to do is install GDB from CVS
and link
unladen against LLVM TOT.
I'm leaving this bug open, and we can close it when we feel like updating LLVM
in
trunk again, since it's kind of a pain.
Original comment by reid.kle...@gmail.com
on 21 Sep 2009 at 12:01
Great, thanks Reid! Do you have a test script (shell will do) that we can use
to make
sure the gdb/LLVM integration is working smoothly, or should I just stick a
null
pointer dereference somewhere?
Original comment by collinw
on 21 Sep 2009 at 9:06
I just put a breakpoint on PyObject_Print and write a script like so:
def foo(): bar()
def bar(): baz()
def baz(): print 'Hello, World!'
foo()
... and run it with -S and -j always. The backtrace for unladen should show 4
JITed
frames starting with #u#, one for the module and one for each function. You
could
make an integration test out of this, but you'd have to guard it on the version
of GDB.
Original comment by reid.kle...@gmail.com
on 21 Sep 2009 at 11:48
Closed, since Jeff pulled the latest changes in from LLVM. Apparently GDB 7.0
is
coming around the corner, so this should work out of the box for everyone in
less
than a year.
Original comment by reid.kle...@gmail.com
on 24 Sep 2009 at 5:27
I actually need to check in the gdb test before this is fully closed.
Original comment by jyass...@gmail.com
on 24 Sep 2009 at 6:50
Tested with r841.
Original comment by jyass...@gmail.com
on 29 Sep 2009 at 12:22
Original issue reported on code.google.com by
reid.kle...@gmail.com
on 8 Jun 2009 at 5:50