On Mon, Apr 21, 2014 at 1:03 PM, Yuri Gribov <tetra2...@gmail.com> wrote: >> You call __asan_report_load8 directly. >> This will only work if __asan_load8 does not have its own frame. > > You mean you don't want asan callbacks to be present in stacktraces? That too, but we also need to be able to unwind properly. Generally, unless we call GET_CALLER_PC_BP_SP directly from the function called from the user code we are not guaranteed to get correct stack trace (since asan-rt is built w/o frame pointers).
> >> My version lacked the UNLIKELY trick. fixed. > > Cool, upstream code seems to run really fast now: 16 sec. (vs my 17.8 > sec.). Looks like outline overhead is only 300% now (compared to 200% > for normal inline ASan) which is probably acceptable. And we'll have > to work around ABI to improve further. > > Some suggestions: > * can we move separate calls for begin/end for intrinsics (and larger > memory accesses) into callbacks ? > * can we remove check for size == 0 and perform it in callback instead ? > I can submit a patch if you are ok with those. Hold on a bit here. First, my recent patches have a bug around handling of memset&co, second I am considering to competely replace memset&co calls with another function call -- this will solve some other problems as well, but it needs performance evaluation... > > BTW note that we will suffer significant PLT penalty in sanitized shared libs. Don't use shared libs :) --kcc > > -Y > > -- > You received this message because you are subscribed to the Google Groups > "address-sanitizer" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to address-sanitizer+unsubscr...@googlegroups.com. > For more options, visit https://groups.google.com/d/optout. -- You received this message because you are subscribed to the Google Groups "address-sanitizer" group. To unsubscribe from this group and stop receiving emails from it, send an email to address-sanitizer+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
