llvm dev - Nov 2020 - Questions about ctor/dtor attribute related diagnostics

Hi,

I noticed that there are some diagnostics discrepancies between clang and
gcc related to `__attribute__((constructor(n)))` and
`__attribute__((destructor(n)))`.

*[clang]*
It seems priorities >65535 and <0 on ctor/dtor functions are ignored and
are set back to default 65535 in clang and Clang only gives an error when >
32-bit unsigned value specified.

*[g++]*
g++ gives an error for any values <0 and >65535.

I am wondering should we let Clang's diagnostics match g++ to make things
clearer to the user? Or why Clang emits an error for *>32-bit unsigned
value* only?

Thank you,
Xiangling
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20201030/d7cb5ddc/attachment.html>

ping.

On Fri, Oct 30, 2020 at 5:01 PM Xiangling Liao <xiangxdh at gmail.com>
wrote:
> Hi,
>
> I noticed that there are some diagnostics discrepancies between clang and
> gcc related to `__attribute__((constructor(n)))` and
> `__attribute__((destructor(n)))`.
>
> *[clang]*
> It seems priorities >65535 and <0 on ctor/dtor functions are ignored
and
> are set back to default 65535 in clang and Clang only gives an error when
>
> 32-bit unsigned value specified.
>
> *[g++]*
> g++ gives an error for any values <0 and >65535.
>
> I am wondering should we let Clang's diagnostics match g++ to make
things
> clearer to the user? Or why Clang emits an error for *>32-bit unsigned
> value* only?
>
> Thank you,
> Xiangling
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20201105/bc20766c/attachment.html>

On Fri, Oct 30, 2020 at 5:03 PM Xiangling Liao via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
>
> Hi,
>
> I noticed that there are some diagnostics discrepancies between clang and
gcc related to `__attribute__((constructor(n)))` and
`__attribute__((destructor(n)))`.
>
> [clang]
> It seems priorities >65535 and <0 on ctor/dtor functions are ignored
and are set back to default 65535 in clang and Clang only gives an error when
> 32-bit unsigned value specified.
>
> [g++]
> g++ gives an error for any values <0 and >65535.
>
> I am wondering should we let Clang's diagnostics match g++ to make
things clearer to the user? Or why Clang emits an error for >32-bit unsigned
value only?
The code for this lives in SemaDeclAttr.cpp in
`handleConstructorAttr()` and `handleDestructorAttr()`, both of which
are defined similarly as trying to read the attribute argument as an
unsigned 32-bit value. If the value is outside of the range of a
32-bit unsigned number, the `checkUInt32Argument()` helper function
will diagnose it as an error. We don't have any further range checking
happening for the argument, though.

I think it's reasonable for us to match GCC's behavior with
diagnostics here. You could model the check after what is done in
`handleInitPriorityAttr()` which has to do a similar range check.
However, you should also verify whether values < 0 or > 65535 are
reserved to the implementation itself (perhaps are used as part of
system headers like glibc) as they are with `init_priority` values <
101 or > 65535. The `init_priority` attribute shows how to exclude
system headers from the range check if that's needed.

~Aaron
>
> Thank you,
> Xiangling
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

On Thu, Nov 5, 2020 at 6:31 AM Aaron Ballman via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
>
> On Fri, Oct 30, 2020 at 5:03 PM Xiangling Liao via llvm-dev
> <llvm-dev at lists.llvm.org> wrote:
> >
> > Hi,
> >
> > I noticed that there are some diagnostics discrepancies between clang
and gcc related to `__attribute__((constructor(n)))` and
`__attribute__((destructor(n)))`.
> >
> > [clang]
> > It seems priorities >65535 and <0 on ctor/dtor functions are
ignored and are set back to default 65535 in clang and Clang only gives an error
when > 32-bit unsigned value specified.
> >
> > [g++]
> > g++ gives an error for any values <0 and >65535.
> >
> > I am wondering should we let Clang's diagnostics match g++ to make
things clearer to the user? Or why Clang emits an error for >32-bit unsigned
value only?
>
> The code for this lives in SemaDeclAttr.cpp in
> `handleConstructorAttr()` and `handleDestructorAttr()`, both of which
> are defined similarly as trying to read the attribute argument as an
> unsigned 32-bit value. If the value is outside of the range of a
> 32-bit unsigned number, the `checkUInt32Argument()` helper function
> will diagnose it as an error. We don't have any further range checking
> happening for the argument, though.
>
> I think it's reasonable for us to match GCC's behavior with
> diagnostics here. You could model the check after what is done in
> `handleInitPriorityAttr()` which has to do a similar range check.
> However, you should also verify whether values < 0 or > 65535 are
> reserved to the implementation itself (perhaps are used as part of
> system headers like glibc) as they are with `init_priority` values <
> 101 or > 65535. The `init_priority` attribute shows how to exclude
> system headers from the range check if that's needed.
>
> ~Aaron
Supplement: the range [0,100] is reserved for use by system runtime
libraries and will be warned by -Wprio-ctor-dtor (unimplemented in
Clang).

I added #pragma GCC diagnostic ignored "-Wprio-ctor-dtor" for exactly
this diagnostic in compiler-rt/lib/profile/GCDAProfiling.c
> >
> > Thank you,
> > Xiangling



-- 
宋方睿