Re: PATCH: Add --with-build-config=bootstrap-asan support

[Paolo Bonzini <bonzini at gnu dot org>]

On Mon, Nov 19, 2012 at 4:06 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Mon, Nov 19, 2012 at 12:01 AM, Paolo Bonzini <bonzini@gnu.org> wrote:
>> Il 19/11/2012 05:35, H.J. Lu ha scritto:
>>> On Sun, Nov 18, 2012 at 7:28 AM, Paolo Bonzini <bonzini@gnu.org> wrote:
>>>> Il 18/11/2012 00:54, H.J. Lu ha scritto:
>>>>> +@if gcc-bootstrap
>>>>> +ifneq ($(filter bootstrap-asan,$(BUILD_CONFIG)),)
>>>>> +LIBASAN_LIBS=-B$$r/prev-$(TARGET_SUBDIR)/libsanitizer/asan/.libs
>>>>> +endif
>>>>> +@endif gcc-bootstrap
>>>>
>>>> Do you need this to be here?  POSTSTAGE1_*_EXPORT is only used when
>>>> bootstrapping, and they are defined with "=" so the definition could be
>>>> placed (or so it seems at first look) in bootstrap-asan.mk.
>>>>
>>>
>>> It turns out to be quite tricky:
>>>
>>> 1. We can't use -faddress-sanitizer in libcpp due to
>>>
>>> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55380
>>
>> Does -faddress-sanitizer add a CPP symbol that lets you work around it?
>>  If not, it's a bug.  If so, please make libcpp revert to the dump
>> algorithm if said symbol is defined.
>
> No, it doesn't define a CPP symbol.

Then we have to fix this first.

>>> 2. We can't use  -faddress-sanitizer in libiberty since
>>> it will make all host binaries depend on libasan.
>>> 3. We can't use  -faddress-sanitizer in lto-plugin
>>> since it will make linker depends on libasan.
>>
>> These are features, not bugs.  You asked for that explicitly.  In fact,
>> I think that using bootstrap-asan in a combined tree should make all
>> host binaries use asan.
>
> That is true.  But we have to support separate ld. My patch
> will delegate this decision to each tool directory.

So if a plugin uses asan, it cannot be linked with a program that
doesn't use it?  Is this documented anywhere?  I suppose however that
a plugin compiled without asan can be linked to a program that uses
it.

So, if asan-enabled lto-plugin doesn't work together with normal ld,
the solution is to remove -faddress-sanitizer from the flags unless
you are in a combined tree.

Paolo