On 07/20/2017 12:10 PM, Richard Biener wrote:
On Thu, 20 Jul 2017, Tom de Vries wrote:
Hi,
this patch fixes PR81430, an ICE in the libgomp testsuite for both openmp and
openacc test-cases for x86_64 with nvptx accelerator.
The scenario how we hit the ICE is as follows:
- a 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, because it generates code that is not
supported by the nvptx target.
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;
}
...
The patch fixes this by calling finish_options in lto1 after
cl_optimization_restore.
Points for review:
1. I'm uncertain though about the placement of the call. Perhaps it should be
in cl_optimization_restore, before targetm.override_options_after_change?
2. I think that this is offloading specific, so perhaps this should be guarded
with lto_stream_offload_p or #ifdef ACCEL_COMPILER or some such.
Hmm, I agree with #2. I think it conceptually is a LTO stream adjustment
and thus we should do this at the time we stream in the
optimization/target nodes (like we remap modes for example). Not
sure if it's possible to do this at that point, but it looks like
finish_options takes two option structs and thus we should be able to
call it.
With what parameters? Patch below tries with same option struct, but ...
Do you get the inform note? I suppose we don't really want that, no?
... I think that way we'll get the inform note (while the previous
solution did not).
I've also tried with a tmp2 memset to 0, but that ran into problems when
doing a maybe_set_param_value.
Thanks,
- Tom
diff --git a/gcc/tree-streamer-in.c b/gcc/tree-streamer-in.c
index 7f7ea7f..e0e792b 100644
--- a/gcc/tree-streamer-in.c
+++ b/gcc/tree-streamer-in.c
@@ -33,6 +33,7 @@ along with GCC; see the file COPYING3. If not see
#include "ipa-chkp.h"
#include "gomp-constants.h"
#include "asan.h"
+#include "opts.h"
/* Read a STRING_CST from the string table in DATA_IN using input
@@ -769,6 +770,20 @@ lto_input_ts_function_decl_tree_pointers (struct
lto_input_block *ib,
DECL_FUNCTION_SPECIFIC_TARGET (expr) = stream_read_tree (ib, data_in);
#endif
DECL_FUNCTION_SPECIFIC_OPTIMIZATION (expr) = stream_read_tree (ib,
data_in);
+#ifdef ACCEL_COMPILER
+ {
+ tree opts = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (expr);
+ if (opts)
+ {
+ struct gcc_options tmp;
+ cl_optimization_restore (&tmp, TREE_OPTIMIZATION (opts));
+ finish_options (&tmp, &tmp, DECL_SOURCE_LOCATION (expr));
+ opts = build_optimization_node (&tmp);
+
+ DECL_FUNCTION_SPECIFIC_OPTIMIZATION (expr) = opts;
+ }
+ }
+#endif
/* If the file contains a function with an EH personality set,
then it was compiled with -fexceptions. In that case, initialize
