[Bug bootstrap/63888] [5 Regression] bootstrap failed when configured with -with-build-config=bootstrap-asan --disable-werror

[kcc at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63888

--- Comment #8 from Kostya Serebryany <kcc at gcc dot gnu.org> ---
(sorry for delay, I missed the last comment)
> Generally, we do want to instrument even artificial variables, and on many
> of them buffer overflow is possible.

Yea, agree. 

> 
> > I.e. is a buffer overflow on these variables possible? 
> > 
> > If we don't instrument them we won't report an ODR violation. 
> > 
> > >> GCC with -fmerge-all-constants also merges many read-only constants, even addressable ones, when they have the same content.
> > 
> > Then I guess we may have other false positives. 
> > We may want to disable -fmerge-all-constants under asan
> 
> You haven't responded about the language thing, there is no such thing as
> ODR in C or Fortran, so you shouldn't report it.

In LLVM, I do not (and should not) know what source language is being compiled.
The differences between languages are represented in the linkage types 
of the globals. E.g. a regular global in C will not be instrumented at all 
unless -fno-common is given. I.e. the difference is not in the source language 
but in the linkage type of the globals. 

% clang -S -o - -emit-llvm glob.c 
@GLOBAL = common global i32 0, align 4
% clang -S -o - -emit-llvm glob.c -fno-common
@GLOBAL = global i32 0, align 4
% clang++ -x c++ -S -o - -emit-llvm glob.c 
@GLOBAL = global i32 0, align 4


Can you give an example in C where we do not want to check for ODRs, yet
the basic asan instrumentation is legal? 

> There are 31 (or 63) bits left in __has_dynamic_init field, can't one bit be
> used for whether ODR should be reported or not?
That will still mean an API change, something I like to avoid when possible.