Le 06/07/2026 à 22:02, Harald Anlauf a écrit :
Hi Mikael,
On 7/6/26 11:35, Mikael Morin wrote:
This started as an attempt to remove this annoying usage of
gfc_conv_descriptor_rank in trans-array.cc(gfc_conv_array_parameter):
9278 tree idx = gfc_create_var (TREE_TYPE
(gfc_conv_descriptor_rank
9279 (old_desc)),
9280 "idx");
That snippet was creating a data reference, and then throwing it away to
only keep the type, which is the type of array descriptors' rank field.
I could have used just the rank field type directly, but I eventually
convinced myself that the snippet above showed a lack of a type readily
available to represent ranks and array dimensions. This patchs adds
such a type.
As usual, I wasn't very inspired when choosing the name for it. Feel
free to propose better.
Regression tested on powerpc64le-unknown-linux-gnu. OK for master?
this seems to be a nice improvement and looks fine to me.
A stupid question: have you considered a conversion of the result
of gfc_get_cfi_desc_rank already in the getter in trans-array.cc?
This could lead to simpler code in trans-decl.cc::gfc_conv_cfi_to_gfc
where it is used, unless I missed something.
After a quick investigation, it seems the function is also used to
assign to the field (see gfc_conv_gfc_desc_to_cfi_desc in
trans-expr.cc), so I prefer to keep unmodified the type of the data
reference tree returned by the function.
I received an aarch64 regression notification from the linaro CI bot,
but could not reproduce it, so I will push the patch as submitted and
see what happens.
Thanks for the review.
