https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128176
Bug ID: 128176
Summary: [14/15/16/17 Regression] Fortran multi-level mapping
with pointer attachment
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: openmp, wrong-code
Severity: normal
Priority: P3
Component: fortran
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
CC: sandra at gcc dot gnu.org
Target Milestone: ---
Created attachment 65964
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=65964&action=edit
test case, compile with -fopenmp and look at the dump and/or run with
offloading enabled
The program was working with GCC 13 but since GCC 14, there are two issues:
* The FE no longer maps ‘map(to:*n->p) but only ‘alloc’s space for the pointer
'p' in n%p
* The ME changes the ‘map(to: *n)’ to ‘map(alloc: *n)’
Kudos to Sandra for finding this issue.
* * *
Regarding the first issue:
For 'map(to:n), map(to:n%p), map(to:n%p%p)', the original dump is (since GCC
14):
- map(to:*n) map(alloc:n [pointer assign, bias: 0]) (looks ok)
- map(alloc:n->p [len: 8]) (okay but as '*n' is mapped should be a noop)
- (missing: map(to *n->p [len: 16]))
- map(to:*n->p->p [len: 16]) map(attach_detach:n->p->p [bias: 0]) (looks OK)
Note that this regressed between gfortran-13 and gfortran-14 (same as
mainline); GCC 13's original dump was:
- map(to:*n) map(alloc:n [pointer assign, bias: 0]) (same)
- map(to:*n->p [len: 16]) map(always_pointer:n->p [pointer assign, bias: 0])
(here ok, wrong in GCC 14+)
- map(to:*n->p->p [len: 16]) map(always_pointer:n->p->p [pointer assign, bias:
0]) (ok; ptr attach way changed in GCC 13→14)
* * *
When lowering the current (gcc 14 to gcc 17 mainline) code, the FE issue
obviously remains - but a new issue shows up. Namely, -fdump-tree-gimple code
is:
The gimple lowering turns the mainline version into:
- map(struct:*n->p [len: 1]) map(alloc:n->p [len: 8]) This one is wrong
- map(to:*n->p->p [len: 16]) This is fine, attachment happens in (4)
- map(alloc:*n [len: 16]) map(alloc:n [pointer assign, bias: 0]) This one maps
all of 'n', which looks fine – except that '*n' is only 'alloc' and no longer
'to' !
- map(attach: n->p->p [bias: 0]) map(attach: n->p [bias: 0]) looks fine
Namely, the map type in the third map clause changed from 'to' to 'alloc'!
