yaxunl added a comment.
In D76937#1950077 <https://reviews.llvm.org/D76937#1950077>, @rjmccall wrote:
> Can you explain what exactly the emission/semantic model is for variables?
> Normal code-generation absolutely triggers the emission of many variables
> lazily (e.g. internal-linkage globals, C++ inline variables); and any
> variable that's *not* being emitted lazily actually needs to be treated as a
> potential root into the delayed-diagnostic graph.
Currently only in the following case a global variable can change the emission
state of a function:
int foobar3() { return 1; }
#pragma omp declare target
int (*C)() = &foobar3;
#pragma omp end declare target
The global variable needs to be in a omp declare target directive. Its
initializer references a host function.
This will cause the function emitted in device compilation.
This is not transitive for variable declaration/references, e.g. the following
case will not cause foobar3 to be emitted in device compilation, unless
variable C itself is enclosed in omp declare target directive.
int foobar3() { return 1; }
int (*C)() = &foobar3;
#pragma omp declare target
int (*D)() = C;
#pragma omp end declare target
Since we logged all such variable declarations in DeclsToCheckForDeferredDiags,
we only need to check variables decls in DeclsToCheckForDeferredDiags and do
not need to check variable decls in DeclRefExpr.
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D76937/new/
https://reviews.llvm.org/D76937
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
