Daniel Kraft wrote:
On 04/03/11 17:03, Tobias Burnus wrote:
For the support, the scalarizer had to be modified and the cobounds had
to be saved in the descriptor (and in TYPE_LANG_SPECIFIC(node)).
I'm all but an expert on this area (maybe you could ask Mikael for his
opinion on the scalarizer changes),
I am happy for any additional comment.
but isn't this TYPE_LANG_SPECIFIC stuff only for compile-time known
things?
Yes and no. gfortran only uses array descriptors for deferred-shape and
assumed-shape arrays ("A(:)") - for all other arrays (e.g. "A(4)",
"A(n)") the bounds only live as "tree" in the the array's
TYPE_LANG_SPECIFIC(node). If one needs them, e.g. to create an an array
descriptor on in the scalarizer, one uses the such stored "tree" for
assignment etc. Hence, the bounds are stored there - and live there only
at compile time - but the they are used to materialize if needed.
@@ -4129,7 +4129,12 @@ gfc_is_coindexed (gfc_expr *e)
It seems to me there's some tab-vs-space issue here (for the int and
for lines).
Indeed, a "tab" was missing in the "int" line.
@@ -3152,6 +3173,14 @@ gfc_conv_ss_startstride (gfc_loopinfo * loop)
That doesn't change anything, but I'd rather have a "break" for the
new cases (even though there was none before and fall-through just has
the same effect). It seems clearer to me as you don't do anything
"special" via fall-through.
Done.
@@ -3739,7 +3780,7 @@ gfc_conv_loop_setup (gfc_loopinfo * loop, locus
* where)
+ if (loopspec[n] != NULL /*|| n >= loop->dimen*/)
Or rather remove it and add later seems also like a good possibility
to me.
Removed outcommented code.
@@ -7386,8 +7484,9 @@ gfc_walk_variable_expr (gfc_ss * ss, gfc_expr *
expr)
+
newss->data.info.dim[newss->data.info.dimen+newss->data.info.codimen]
= n;
Nit: Please add appropriate spacing around the "+" in the "long line".
(As it is in the hunk above.)
Done - or actually: I also added a line break as the line was too long.
Committed as Rev. 171949.
Thanks for the review!
Tobias
