llvm dev - Jan 2021 - LLVM now assume functions may not return, unless they are marked as willreturn

Hi,

Just a quick heads up. Recently a few changes landed that updated code in LLVM
that implicitly assumed some functions will return. Now, the code explicitly
checks for the `willreturn` attribute. Please see the patches below
https://reviews.llvm.org/rG292077072ec1
<https://reviews.llvm.org/rG292077072ec1>
https://reviews.llvm.org/rG5d12b976b004
<https://reviews.llvm.org/rG5d12b976b004>

Among other things, this means LLVM now won’t remove function calls to readonly
functions, unless they are willreturn. This fixes some longstanding issues for
languages without C++-like forward-progress guarantees (see
https://bugs.llvm.org/show_bug.cgi?id=965 )

 For the simple example below, LLVM now won’t remove the call to loop(), if it
is compiled as a C program.

void loop() {
	while(1);
}
int main() {
	loop();
	return 1;
}


If your frontend provides C++-like forward progress guarantees, please make sure
functions/loops are marked as `mustprogress` accordingly. Mustprogress helps
with inferring `willlreturn`.

If your frontend can guarantee a function will always return, consider adding
the `willreturn` attribute to the generated functions.

Please audit the intrinsics definitions for your target. They also need to be
marked as willreturn, if possible.As an example, see the patch that updated
AArch64 https://reviews.llvm.org/rG50ae6a3ac9bd
<https://reviews.llvm.org/rG50ae6a3ac9bd> . At the moment, the code still
(wrongly) assumes that all intrinsics will return. Once people had time to
update their backends, the plan is to remove this assumption.

Cheers,
Florian
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210125/98053047/attachment.html>

Hi,

Should this doc be updated accordingly:
https://llvm.org/docs/Frontend/PerformanceTips.html ?

Thanks!

-- 
Mehdi


On Mon, Jan 25, 2021 at 2:23 AM Florian Hahn via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Hi,
>
> Just a quick heads up. Recently a few changes landed that updated code in
> LLVM that implicitly assumed some functions will return. Now, the code
> explicitly checks for the `willreturn` attribute. Please see the patches
> below
> https://reviews.llvm.org/rG292077072ec1
> https://reviews.llvm.org/rG5d12b976b004
>
> Among other things, this means LLVM now won’t remove function calls to
> readonly functions, unless they are willreturn. This fixes some
> longstanding issues for languages without C++-like forward-progress
> guarantees (see https://bugs.llvm.org/show_bug.cgi?id=965 )
>
>  For the simple example below, LLVM now won’t remove the call to loop(),
> if it is compiled as a C program.
>
> void loop() {
> while(1);
> }
> int main() {
> loop();
> return 1;
> }
>
>
> If your frontend provides C++-like forward progress guarantees, please
> make sure functions/loops are marked as `mustprogress` accordingly.
> Mustprogress helps with inferring `willlreturn`.
>
> If your frontend can guarantee a function will always return, consider
> adding the `willreturn` attribute to the generated functions.
>
> Please audit the intrinsics definitions for your target. They also need to
> be marked as willreturn, if possible.As an example, see the patch that
> updated AArch64 https://reviews.llvm.org/rG50ae6a3ac9bd . At the moment,
> the code still (wrongly) assumes that all intrinsics will return. Once
> people had time to update their backends, the plan is to remove this
> assumption.
>
> Cheers,
> Florian
> _______________________________________________
> 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/20210125/11944336/attachment.html>