[llvm-dev] Should the compiler-rt builtins be configured with CMAKE_TRY_COMPILE_TARGET_TYPE?

[Petr Hosek via llvm-dev]

I've been thinking about this a bit more and doing some experiments. I'm
no longer sure if setting CMAKE_TRY_COMPILE_TARGET_TYPE to STATIC_LIBRARY
is the best solution. While it could simplify checks in projects that only
ever produce static libraries like builtins or crt, using it more broadly
might be a problem because it breaks checks like check_function_exists and
check_library_exists.

The idea I'm considering would be to instead roll out our own checks. The
way checks like check_function_exists and check_library_exists work is that
CMake generates a little snippet which it then tries to compile. For
example when checking if dlopen in libdl exists, CMake would generate:

char dlopen(void);
int main(int ac, char* av[]) {
  dlopen();
}

It'd then try to compile this file with -ldl and see if the command
succeeds or fails. The problem with this snippet is that it implicitly uses
libc and other libraries like crt or builtins, which in the case of the
runtimes build, are libraries we're going to build introducing a cycle. The
alternative I came up with would be to generate a custom snippet that looks
like this:

char dlopen(void);
void _start() {
  dlopen();
}

When invoking the compiler, we would additionally pass -nostdlib. That way,
we would still check if libdl is available without any implicit
dependencies. The only downside to this approach I can see is that we would
need to maintain our own set of checks, but since we could rely on CMake's
try_compile to do most of the heavy lifting, I don't think it'd be too much
of a maintenance overhead.

If we switched to these custom checks across all runtimes, I hope that we
could eliminate the need to bootstrap builtins before we build the rest of
runtimes, and instead we could use a single CMake build for all runtimes
and rely on regular dependencies for things like builtins.

Does this make sense? Have I missed anything?

On Wed, Mar 10, 2021 at 11:34 PM Petr Hosek <phosek at google.com> wrote:

> I was actually wondering if we could go further and set this not just for
> builtins but also for the runtimes build. In the runtimes build, we link
> shared libraries like libunwind, libc++abi, libc++ and some of the
> sanitizers, but it's not clear to me if this would cause any issues. We
> have a few check_library_exists calls but most of them are checking
> libraries like libc, libm, libdl, librt? Even C libraries that don't
> provide these as separate libraries like musl typically provide linker
> scripts for backwards compatibility. Would any of the currently supported
> platforms break if we started linking those libraries unconditionally?
> Alternatively, we could also consider replacing the use of -nodefaultlibs
> with -nostdlib++ and letting the compiler driver handle this, but GCC as
> far as I'm aware doesn't yet support that flag so we would need a fallback
> (we could set CMAKE_TRY_COMPILE_TARGET_TYPE to STATIC_LIBRARY and
> -nostdlib++ only for Clang, and keeping the existing behavior for GCC, but
> it might increase the complexity).
>
> On Wed, Mar 10, 2021 at 11:27 PM Martin Storsjö via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> Hi,
>>
>> On Thu, 11 Mar 2021, Shoaib Meenai via llvm-dev wrote:
>>
>> > What’s happening is that the configuration for the builtins for Android
>> > tests for all supported architectures [1], using check_symbol_exists to
>> > check for the architecture’s preprocessor macro [2].
>> check_symbol_exists
>> > tries to link a test executable, which fails because my just-built
>> > compiler doesn’t have compiler-rt yet (since it’s trying to build
>> > compiler-rt). Consequently, the builtins don’t think that any
>> > architectures are supported, and just don’t build anything as a result.
>> >
>> >
>> >
>> > Other places in the builtins’ CMake logic use custom functions like
>> > try_compile_only to only test compilation and not linking, to avoid
>> issues
>> > like this. My understanding is that the builtins will never be built as
>> > shared libraries. The build already instructs its custom test macros to
>> > always only check compilation [3]; now that we’re on a newer CMake
>> version,
>> > would it be appropriate to always set CMAKE_TRY_COMPILE_TARGET_TYPE to
>> > STATIC_LIBRARY for the builtins as well, so that all CMake checks build
>> > static libraries (which has the same effect of only testing compilation
>> and
>> > not linking)? We could also probably clean up a lot of the custom logic
>> for
>> > only checking compilation afterwards.
>>
>> These kinds of chicken-and-egg problems are all over building the
>> runtimes, and when building the initial copy of runtime libraries,
>> there's
>> some amount of trickery needed to pass such tests.
>>
>> In most of the cases where I build runtimes, I've had to add
>> -DCMAKE_C_COMPILER_WORKS=TRUE -DCMAKE_CXX_COMPILER_WORKS=TRUE, but iirc
>> one of such cases for compiler-rt could be removed if we'd set
>> CMAKE_TRY_COMPILE_TARGET_TYPE to STATIC_LIBRARY. (For
>> libunwind/libcxxabi/libcxx, things were more complicated though, I think.)
>>
>> I actually sent a patch to this effect a couple months ago, have a look
>> at
>> [1]. That one was only aimed at standalone builds of
>> compiler-rt/lib/builtins, but maybe it could be generalized to some other
>> location as well. For non-standalone builds, I guess the option would
>> have
>> to be cleared after processing the current project?
>>
>> // Martin
>>
>> [1] https://reviews.llvm.org/D91334
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210315/1eecf055/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 3996 bytes
Desc: S/MIME Cryptographic Signature
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210315/1eecf055/attachment.bin>