abhinavgaba wrote:
> Even for cases where both the pointer, pointee are mapped, the expectation is
> for the pointer to still be captured by-ref.
>Why?
Because we don't know if the pointer/pointee were both already existing on the
device, in which case, there should be no pointer-attachment for them.
```c
int x = 111;
#pragma omp target enter data map(x) // map the pointee
int *p = &x;
#pragma omp target enter data map(p) // map the pointer
#pragma omp target map(p, p[0:1]) // No pointer-attachment between p and
p[0], since both
// already existed on the device
{
p = (int*) 111;
}
printf("%ld\n", (long) p); // CHECK: 111
```
And as I mentioend, even if there is a pointer attachment, and the value of p
cannot be modified in the region, a user may still assert that the address of p
in the device matches its mapped address, like:
```c
int x;
int *p = &x;
#pragma omp target enter data map(p)
int64_t mapped_p = (int64_t) omp_get_mapped_ptr(&p, omp_get_default_device());
#pragma omp target map(p, p[0:1]) // pointer-attachment succeeds here
{
assert(mapped_p == (int64_t) &p && "address of p in device should match the
mapped address before the region.");
}
```
https://github.com/llvm/llvm-project/pull/145454
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
