Hello,
I did some debugging with bouds-checking and came up with attached patch.
I seriously doubt any one did use bounds checking in a large project
before.
Currently I can use this now in a large multi threaded project. It
still needs some more testing so do not apply the patch yet.
I disabled some errors. For example if a bounded pointer is not found
I give no error. I also relaxed printing free errors.
There were some off by 1 errors in lib/bcheck.c and I needed to make
the code thread safe.
I used the patch to not link in libtcc1.a in shared objects when
bounds checking so I have only one memory pool.
This has to be documented because you cannot use this with dlopen for
example.
I also added the pthread library when bounds checking so it is now
multi threaded.
I found another problem with nocode_wanted when using sizeof().
Also the push/pop trick needed to push some more registers when more
parameters are passed in registers.
I probably forget to mention a lot a other changes. See the patch.
I only tested this on linux x86_64. There are for sure problems on
other targets.
Regards,
Herman
On 2019-11-28 17:41, Michael Matz wrote:
Hello again,
but to maybe be a bit more constructive:
On Thu, 28 Nov 2019, Michael Matz wrote:
I fixed this with some push/pop trickery.
I see, yeah, expanding calls during calls is broken as gfunc_call in the
generators doesn't generally leave a trace in vtop[] which registers are
currently holding values. I think you only need so push/pop si/di, as
cx/dx aren't used intentionally during reg-param setup.
(I think i386-gen.c has a simila bug with fastcall functions).
This probably could be
improved. I have now added a minimum patch so bounds checking works a
little bit. We need still to fix the shared lib reloc problems and the
malloc/free hooks.
Do we? Can we perhaps also simply declare bounds checking to work only
with the main executable? Or remove that whole feature altogether?
And perhaps another compromise: only conditionally enable tracking of
locals: Invent a new cmdline option (say, '-bb'), which sets
do_bounds_checking to 2. And only if it's > 1 you would also track
locals, whereas with == 1 you would only track arrays and structs.
Your decision, I think you can push this patch either with that
change, or
without (but try to remove cx/dx from the push/pop). It doesn't make
tccs
source code larger or uglier in any meaningful way, but does fix
practical
bugs.
Ciao,
Michael.