https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
--- Comment #12 from vries at gcc dot gnu.org ---
Created attachment 36063
-- https://gcc.gnu.org/bugzilla/attachment.cgi?id=36063action=edit
autopar/outer-7.c
C example to reproduce the same problem
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
--- Comment #11 from vries at gcc dot gnu.org ---
Created attachment 36057
-- https://gcc.gnu.org/bugzilla/attachment.cgi?id=36057action=edit
Updated tentative patch
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
vries at gcc dot gnu.org changed:
What|Removed |Added
Keywords|patch |
--- Comment #10 from vries
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
--- Comment #9 from vries at gcc dot gnu.org ---
(In reply to vries from comment #7)
Created attachment 35986 [details]
Updated tentative patch
I found that always doing graphite before parloops resulted in failures to
parallelize reduction
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
vries at gcc dot gnu.org changed:
What|Removed |Added
Attachment #35983|0 |1
is obsolete|
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
--- Comment #6 from vries at gcc dot gnu.org ---
(In reply to kargl from comment #5)
Does the loop ordering matter? Fortran is a column major language,
so your nested loops are backwards. One would normally write.
do jj = 0, n - 1
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
vries at gcc dot gnu.org changed:
What|Removed |Added
Keywords||patch
--- Comment #8 from
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
vries at gcc dot gnu.org changed:
What|Removed |Added
Keywords||missed-optimization
---
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
kargl at gcc dot gnu.org changed:
What|Removed |Added
CC||kargl at gcc dot gnu.org
---
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
--- Comment #2 from vries at gcc dot gnu.org ---
Another obvious difference is that the fortran 2-dimensional array access is
rewritten into a single dimension array access:
...
bb 3:
# ii_7 = PHI 0(2), ii_16(7)
pretmp_52 =
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
--- Comment #4 from vries at gcc dot gnu.org ---
Created attachment 35983
-- https://gcc.gnu.org/bugzilla/attachment.cgi?id=35983action=edit
Tentative patch
Using this tentative patch, we use graphite analysis (if available) by default
for
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66873
--- Comment #3 from vries at gcc dot gnu.org ---
The fortran example succeeds when floop-parallelize-all is used.
Even though the access function in graphite seems the same:
...
Access function 0: {{0, +, 1}_3, +, 500}_4
...
12 matches
Mail list logo