rsandifo-arm wrote:
> > Thinking about it a bit more, maybe we can just do some magic to make
> > things work? Say, if you specify `__attribute__((target("sve")))
> > __arm_streaming_compatible`, and the caller is in streaming mode, allow the
> > call even if the caller doesn't have SVE proper.
>
> Thinking a bit more, this is probably not quite what we want: even if the
> function body itself is streaming compatible, it might call non-streaming
> functions that require SVE. Maybe spell this something like
> `__attribute__((target("sve-or-streaming"))) __arm_streaming_compatible`.
I suppose the idea here is that:
```
__attribute__((target("sve"))) void f() { … }
void g() { … f(); … }
```
should be diagnosed, on the basis that, when compiled with default flags, `g`
doesn't guarantee the availability of SVE, whereas `f` requires it? If so, I
don't think we should do that, for two reasons:
First, it's IMO valid to do:
```
__attribute__((target("sve"))) void sve_routine() { … }
void main_interface() {
if (SVE_is_available() && problem_has_certain_characteristics())
sve_routine();
else
…
}
```
That is, feature gating can be dynamic. It doesn't need to be a load-time
thing.
Second, at least in GCC, the `target` attribute is not part of a function's
type, so it's not an error to do:
```
foo.h: void f();
foo.cc: __attribute__((target("sve"))) void f() { … }
```
And I'd argue that that's a feature rather than a bug. It allows load-time
selection of DSOs based on the target.
https://github.com/llvm/llvm-project/pull/92427
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
