Hi Sandra,
Tobias Burnus wrote:
(I have now an errant to do - and will continue later with the review.)
First, something a bit unrelated to this patch but affecting related
code (quoting old, existing code):
int
omp_context_selector_matches (tree ctx)
...
case OMP_TRAIT_DEVICE_KIND:
if (set == OMP_TRAIT_SET_DEVICE)
for (tree p = OMP_TS_PROPERTIES (ts); p; p = TREE_CHAIN (p))
{
const char *prop = omp_context_name_list_prop (p);
if (prop == NULL)
return 0;
if (!strcmp (prop, "any"))
continue;
[Cf. also comment to 3/8] As OpenMP states:
"If trait-property 'any' is specified in the 'kind' trait-selector of
the device selector set or the target_device selector sets, no other
trait-property may be specified in the same selector set."
[From "Restrictions to context selectors are as follows:", here quoting
TR12]
It seems as if we can avoid run-time evaluation for 'device' if there is
a kind(any) - and likewise for 'target_device' if there is 'kind(any)',
but we want to have an error in that case if any other trait has been
specified, I guess.
* * *
Sandra Loosemore wrote:
@@ -2194,12 +2308,21 @@ omp_context_compute_score (tree ctx, score_wide_int
*score, bool declare_simd)
...
int *scores
= (int *) alloca ((2 * nconstructs + 2) * sizeof (int));
That's not new but I have the feeling it should be '+ 3' and not '+ 2'
for device or target_device + and having both device and target device,
it might even need to be + 6.
I also wonder whether 'alloca' will really work - or only when inlined
at all call sites.
Ignore the last sentence. ['scores' is used locally, while 'score' is
passed by the caller.] Still, I wonder about the + 2 vs. +3 or +6.
Otherwise, I have not spotted anything.
Tobias
