[llvm-dev] [cfe-dev] [4.0.0 Release] 'final' has been tagged

[Hans Wennborg via llvm-dev]

On Mon, Mar 13, 2017 at 10:09 PM, Khem Raj <raj.khem at gmail.com> wrote:
> Hi Hans
>
> On 3/10/17 1:10 PM, Hans Wennborg wrote:
>> Hi Khem,
>>
>> On Fri, Mar 10, 2017 at 1:03 PM, Khem Raj via llvm-dev
>> <llvm-dev at lists.llvm.org> wrote:
>>> On Fri, Mar 10, 2017 at 7:01 AM, Renato Golin via llvm-dev
>>> <llvm-dev at lists.llvm.org> wrote:
>>>> ARM and AArch64 looking good, uploaded.
>>>>
>>>
>>> This is not particularly related to this RC but in general with 4.0.0
>>> I am seeing a failure quite common while building for arm
>>>
>>> | ./.libs/libsqlite3.so: undefined reference to `__mulodi4'
>>> | clang-4.0: error: linker command failed with exit code 1 (use -v to
>>> see invocation)
>>> | make: *** [Makefile:509: sqlite3] Error 1
>>>
>>> There are few bug reports in bugzilla also reporting similar issue
>>> e.g.
>>>
>>> https://bugs.llvm.org//show_bug.cgi?id=14469
>>
>> Many of those bugs are pre-4.0.0. Did you see the same issues with
>> earlier LLVM versions as well, or are you saying these errors are new
>> with 4.0.0?
>>
>
> I dug into a bit and here is a reproducer
>
> #include <stddef.h>
> size_t n;
> size_t s;
> ptrdiff_t z;
> main()
> {
>     return __builtin_mul_overflow(n,s, &z);
> }
>
> compiling with
> clang --target=armv7l -mfloat-abi=hard hello.c -S
>
> This generates the call to __mulodi4
>
> same code with gcc does not.
>
> I see that sqlite3 as well as m4 package pokes the compiler for these
> builtins before using them. For clang it uses __has_builtin() to
> determine if a given builtin is supported before using it.
>
> see
> http://git.savannah.gnu.org/cgit/gnulib.git/tree/lib/xalloc-oversized.h
>
> That is the source of trouble, since its in gnulib that explains why its
> showing up in lot of packages.
>
> perhaps this commit here
> http://git.savannah.gnu.org/cgit/gnulib.git/commit/lib/xalloc-oversized.h?id=9c919a06435266a88d91ffb9eb3f70a82aa65b4b
>
> brings this issue to fore.

I get a __mulodi4 call with your reproducer also with llvm 3.9 and
3.8, so I'm not sure this changed on our end. (I also added the
__has_builtin check from the code you linked, with the same result.)

In any case, it's no good generating calls to functions if they're not
defined in the runtime.

Can you file a bug about this in the Bugzilla and cc myself and Renato?

Thanks,
Hans