[nvptx] Force vl32 if calling vector-partitionable routines

[nvptx] Force vl32 if calling vector-partitionable routines

With PTX_MAX_VECTOR_LENGTH set to larger than PTX_WARP_SIZE, routines can be
called from offloading regions with vector-size set to larger than warp size.
OTOH, vector-partitionable routines assume warp-sized vector length.

Detect if we're calling a vector-partitionable routine from an offloading
region, and if so, fall back to warp-sized vector length in that region.

2019-01-07  Tom de Vries  <tdevries@suse.de>

	PR target/85486
	* config/nvptx/nvptx.c (has_vector_partitionable_routine_calls_p): New
	function.
	(nvptx_goacc_validate_dims): Force vl32 if calling vector-partitionable
	routines.

From-SVN: r267640