================
@@ -3371,6 +3379,20 @@ Sema::ActOnBreakStmt(SourceLocation BreakLoc, Scope
*CurScope) {
if (S->isOpenMPLoopScope())
return StmtError(Diag(BreakLoc, diag::err_omp_loop_cannot_use_stmt)
<< "break");
+
+ // OpenACC doesn't allow 'break'ing from a compute construct, so diagnose if
+ // we are trying to do so. This can come in 2 flavors: 1-the break'able
thing
+ // (besides the compute construct) 'contains' the compute construct, at which
+ // point the 'break' scope will be the compute construct. Else it could be a
+ // loop of some sort that has a direct parent of the compute construct.
+ // However, a 'break' in a 'switch' marked as a compute construct doesn't
+ // count as 'branch out of' the compute construct.
+ if (S->isOpenACCComputeConstructScope() ||
+ (!S->isDirectlySwitchScope() && S->getParent() &&
----------------
erichkeane wrote:
>What is "ignorability"? Sorry, just asking to understand the context more
>correctly.
The idea is that we should be able to 'ignore' the pragmas and have the program
still have meaning. A big part of the 'theory' here is that you should be able
to add pragmas to existing code and get 'descriptive' messages that show what
you did wrong, and a pragma introducing a scope isnt really 'true' (so having
it observably act as a scope seems wrong).
>These constructs, if enabled, should be pretty similar to lambdas, so I assume
>it should be fine to diagnose dangling break/continue here. The diagnostics
>might be improved, though.
I think the difference is that the lambda introduces a lookup scope/etc,
whereas these pragmas aren't supposed to. For example, setting the 'function'
scope prevents something like:
```
auto foo(int i) {
#pragma acc parallel
int j = function_call();
return i + j;
}
```
from working. See : https://godbolt.org/z/Gq8Yen1vr (for the omp example).
This is actually something that is supposed to work in OpenACC, so introducing
the scope makes that not work.
https://github.com/llvm/llvm-project/pull/82543
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
