[Bug debug/87039] [8/9 Regression] DW_OP_fbreg used without a frame base on a C++ code w/ -fopenmp
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87039 --- Comment #5 from Jakub Jelinek --- Author: jakub Date: Mon Nov 19 13:44:13 2018 New Revision: 266272 URL: https://gcc.gnu.org/viewcvs?rev=266272=gcc=rev Log: PR debug/87039 * omp-expand.c: Don't include debug.h. (adjust_context_and_scope): Add REGION argument. Find DECL_CONTEXT from innermost outer parallel, task, teams or target that has a child_fn set, or, if there is no such outer region, use current_function_decl. Do the DECL_CONTEXT adjustment regardless of whether a suitable BLOCK is found or not. (expand_parallel_call, expand_teams_call): Don't call adjust_context_and_scope here. (grid_expand_target_grid_body): Revert 2017-01-25 changes. (expand_omp_taskreg, expand_omp_target): Likewise. Call adjust_context_and_scope. * dwarf2out.c (dwarf2out_early_global_decl): For decl_function_context recurse instead of calling dwarf2out_decl. * g++.dg/gomp/pr78363-4.C: New test. * g++.dg/gomp/pr78363-5.C: New test. * g++.dg/gomp/pr78363-6.C: New test. * g++.dg/gomp/pr78363-7.C: New test. Added: trunk/gcc/testsuite/g++.dg/gomp/pr78363-4.C trunk/gcc/testsuite/g++.dg/gomp/pr78363-5.C trunk/gcc/testsuite/g++.dg/gomp/pr78363-6.C trunk/gcc/testsuite/g++.dg/gomp/pr78363-7.C Modified: trunk/gcc/ChangeLog trunk/gcc/dwarf2out.c trunk/gcc/omp-expand.c trunk/gcc/testsuite/ChangeLog
[Bug debug/87039] [8/9 Regression] DW_OP_fbreg used without a frame base on a C++ code w/ -fopenmp
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87039 Jakub Jelinek changed: What|Removed |Added Attachment #45006|0 |1 is obsolete|| --- Comment #4 from Jakub Jelinek --- Created attachment 45020 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=45020=edit gcc9-pr87039.patch Unfortunately the previous patch only fixed the pr78363-1.C testcase and not -2.C or -3.C. The problem is that r253335 only tweaked parallel regions (and later I've copied that to host teams), not others (task, target). There is another thing I've noticed. Because OMP expansion works from leaf regions to outer regions, for parallel/task/target nested in another parallel/task/target/host teams DECL_CONTEXT would be set incorrectly, pointing at the ultimate containing function (the user function), rather than the immediate parent. And, after fixing that I've noticed dwarf2out.c needs tweaking too, because it handles just one decl_function_context, doesn't work recursively.
[Bug debug/87039] [8/9 Regression] DW_OP_fbreg used without a frame base on a C++ code w/ -fopenmp
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87039
Jakub Jelinek changed:
What|Removed |Added
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |jakub at gcc dot gnu.org
--- Comment #3 from Jakub Jelinek ---
Created attachment 45006
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=45006=edit
gcc9-pr87039.patch
Untested fix. As mentioned on IRC, starting with r253335 we set DECL_CONTEXT
that the r241023 code then uses to make sure we emit early debug info of main
before main.omp_fn.0. But it is undesirable to emit early debug info before
actually outlining the regions. So, by reverting the omp-expand.c part of
r244892 the pr78363-*.C testcases still work and:
$ ./cc1plus.vanilla -quiet -g -O2 -fopenmp pr87039.C ; g++ -c -fopenmp -o
pr87039{.o,.s}; readelf -wi pr87039.o | grep without; echo ==
<13a> DW_AT_location: 2 byte block: 91 68 (DW_OP_fbreg: -24) [without
DW_AT_frame_base]
<157> DW_AT_GNU_call_site_value: 2 byte block: 91 68 (DW_OP_fbreg: -24)
[without DW_AT_frame_base]
==
$ ./cc1plus -quiet -g -O2 -fopenmp pr87039.C ; g++ -c -fopenmp -o
pr87039{.o,.s}; readelf -wi pr87039.o | grep without; echo ==
==
No testcase in this patch, we don't have the guality infrastructure in libgomp
testing.
[Bug debug/87039] [8/9 Regression] DW_OP_fbreg used without a frame base on a C++ code w/ -fopenmp
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87039 --- Comment #2 from Jakub Jelinek --- Actually, I think this got broken with PR78363 r244892. Before that the early_global_decl hooks are only called on the functions when all the functions are lowered and thus the BLOCKs moved.
[Bug debug/87039] [8/9 Regression] DW_OP_fbreg used without a frame base on a C++ code w/ -fopenmp
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87039 --- Comment #1 from Jakub Jelinek --- I think this shows a pretty major bug caused by the early dwarf changes - it happens too early. When OpenMP/OpenACC etc. does move_sese_region_to_fn, it duplicates the BLOCK which is on the boundary and should stay in both parent and child, but the other BLOCKs are just moved. With early dwarf, we've already created the DW_TAG_lexical_block for those blocks and added them under their parent block. So, either we shouldn't do that during parsing and only create the lexical blocks with early dwarf at the end of the lowering passes, or move_sese_region_to_fn needs to use some debug hook that would reparent the lexical blocks being moved. Also note that for the omp_fn outlined functions we don't really emit any DWARF early at all, dwarf2out_function_decl is called only late from rest_of_handle_final.
[Bug debug/87039] [8/9 Regression] DW_OP_fbreg used without a frame base on a C++ code w/ -fopenmp
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87039 Richard Biener changed: What|Removed |Added Target Milestone|--- |8.3
[Bug debug/87039] [8/9 Regression] DW_OP_fbreg used without a frame base on a C++ code w/ -fopenmp
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87039 Martin Liška changed: What|Removed |Added Status|UNCONFIRMED |NEW Last reconfirmed||2018-08-21 CC||aoliva at gcc dot gnu.org, ||jakub at gcc dot gnu.org, ||mark at gcc dot gnu.org Ever confirmed|0 |1
