https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81430
Bug ID: 81430
Summary: nvptx acceleration compilation broken because of
running pass_partition_blocks
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: lto
Assignee: unassigned at gcc dot gnu.org
Reporter: vries at gcc dot gnu.org
CC: marxin at gcc dot gnu.org
Target Milestone: ---
[ tentatively setting component to lto ]
Atm, when running the libgomp testsuite for an x86_64 compiler with nvptx
acceleration, we run into a number of ICEs in lto1. I've investigated an ICE
in fix_crossing_unconditional_branches, which occurs for both openmp and
openacc. The basic problem is that we're running pass_partition_blocks for
nvptx, while
we don't want to do that.
Note that for standalone compilation for single-thread ptx execution, we don't
attempt to run pass_partition_blocks. This is because for nvptx,
TARGET_HAVE_NAMED_SECTIONS is set to false, and this bit in finish_options
switches off pass_partition_blocks:
...
/* If the target requested unwind info, then turn off the partitioning
optimization with a different message. Likewise, if the target does not
support named sections. */
if (opts->x_flag_reorder_blocks_and_partition
&& (!targetm_common.have_named_sections
|| (opts->x_flag_unwind_tables
&& targetm_common.unwind_tables_default
&& (ui_except == UI_SJLJ || ui_except >= UI_TARGET))))
{
if (opts_set->x_flag_reorder_blocks_and_partition)
inform (loc,
"-freorder-blocks-and-partition does not work "
"on this architecture");
opts->x_flag_reorder_blocks_and_partition = 0;
opts->x_flag_reorder_blocks = 1;
}
...
But that doesn't seem to work for the offloading setup. The scenario is as
follows:
- testcase is compiled with -O2
- ix86_option_optimization_table enables OPT_freorder_blocks_and_partition at
-O2
- cc1 writes out the flag as part of DECL_FUNCTION_SPECIFIC_OPTIMIZATION
- lto1 reads in the flag as part of DECL_FUNCTION_SPECIFIC_OPTIMIZATION
- lto1 uses the flag, and runs pass_partition_blocks
- pass_partition_blocks ICEs
