[nvptx] Apply vector-partitionable routines workaround to default vl
authorTom de Vries <tdevries@suse.de>
Sat, 12 Jan 2019 10:27:34 +0000 (10:27 +0000)
committerTom de Vries <vries@gcc.gnu.org>
Sat, 12 Jan 2019 10:27:34 +0000 (10:27 +0000)
commitaf79605ec27c0db7dee9ee001cd7d768eb6fcf02
tree2a27c57227ee654b750724c6229e1cdfb72ca771
parentd6f528e70ad88ddcc9786e6050d92e003d19d0a3
[nvptx] Apply vector-partitionable routines workaround to default vl

Make "[nvptx] Force vl32 if calling vector-partitionable routines" work as well
if vector length is set by modifying PTX_DEFAULT_VECTOR_LENGTH.

2019-01-12  Tom de Vries  <tdevries@suse.de>

* config/nvptx/nvptx.c (nvptx_goacc_validate_dims_1): In offloading
region calling vector-partitionable routine, set default_vector_length
to WARP_SIZE.

From-SVN: r267879
gcc/ChangeLog
gcc/config/nvptx/nvptx.c