On Wed, Nov 02, 2016 at 11:31:25AM -0400, Jason Merrill wrote: > On Wed, Nov 2, 2016 at 10:31 AM, Jakub Jelinek <ja...@redhat.com> wrote: > > It uses Alex' LANG_HOOKS_GET_PTRMEMFN_TYPE langhook. I've tried > > to think about https://gcc.gnu.org/ml/gcc-patches/2011-05/msg00227.html > > and we even have such a langhook now, modified_type_die > > uses lang_hooks.types.get_debug_type, but > > 1) it is just called in modified_type_die and not in > > gen_type_die_with_usage, that looks weird > > How much of a problem is that? modified_type_die calls gen_type_die, > does that not cover the cases needed here?
If e.g. on the ref-3.C testcase from the patch I put breakpoint on both gen_ptr_to_mbr_type_die and modified_type_die (the latter only for type->base.code == RECORD_TYPE), then I see first: #0 gen_ptr_to_mbr_type_die (type=<record_type 0x7ffff1727150>, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>, class_type=<record_type 0x7ffff1709540 S>, member_type=<method_type 0x7ffff1709738>) at ../../gcc/dwarf2out.c:23128 #1 0x0000000000c3a7e5 in gen_type_die_with_usage (type=<record_type 0x7ffff1727150>, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>, usage=DINFO_USAGE_DIR_USE) at ../../gcc/dwarf2out.c:24428 #2 0x0000000000c3ab92 in gen_type_die (type=<record_type 0x7ffff1727150>, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:24491 #3 0x0000000000c3caed in gen_decl_die (decl=<var_decl 0x7ffff7ff9ea0 o1>, origin=<tree 0x0>, ctx=0x0, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:25117 #4 0x0000000000c3b13e in process_scope_var (stmt=<block 0x7ffff17129c0>, decl=<var_decl 0x7ffff7ff9ea0 o1>, origin=<tree 0x0>, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:24620 #5 0x0000000000c3b1bb in decls_for_scope (stmt=<block 0x7ffff17129c0>, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:24645 and only afterwards: #0 modified_type_die (type=<record_type 0x7ffff1727150>, cv_quals=0, reverse=false, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:12328 #1 0x0000000000c2f03b in add_type_attribute (object_die=<dw_die_ref 0x7ffff1738aa0 DW_TAG_variable <parent=0x7ffff1738910 DW_TAG_subprogram>>, type=<record_type 0x7ffff1727150>, cv_quals=0, reverse=false, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:20346 #2 0x0000000000c354f4 in gen_variable_die (decl=<var_decl 0x7ffff7ff9ea0 o1>, origin=<tree 0x0>, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:22688 #3 0x0000000000c3cb8e in gen_decl_die (decl=<var_decl 0x7ffff7ff9ea0 o1>, origin=<tree 0x0>, ctx=0x0, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:25138 #4 0x0000000000c3b13e in process_scope_var (stmt=<block 0x7ffff17129c0>, decl=<var_decl 0x7ffff7ff9ea0 o1>, origin=<tree 0x0>, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:24620 #5 0x0000000000c3b1bb in decls_for_scope (stmt=<block 0x7ffff17129c0>, context_die=<dw_die_ref 0x7ffff1738910 DW_TAG_subprogram <parent=0x7ffff15c7000 DW_TAG_compile_unit>>) at ../../gcc/dwarf2out.c:24645 which means if gen_type_die or gen_type_die_with_usage doesn't use the langhook, then we'd emit a completely useless { __pfn; __delta } struct into debug info first, and then in modified_type_die used the langhook, get OFFSET_TYPE and probably create the DW_TAG_ptr_to_member_type. So I think we really need that. > > 2) it is used for something Ada-ish I really don't know how to test etc. > > to be able to find out if it is safe to call it in > > gen_type_die_with_usage too > > You could find an Ada test that uses the code and verify that the > output stays the same? I can try to find the patch that introduced it and if it contained any testcases. > > 3) most importantly, if the C++ version of this langhook would create > > OFFSET_TYPE on the fly, I don't know how to ensure effective sharing > > of DW_TAG_ptr_to_member_type nodes with the same DW_AT_type > > and DW_AT_containing_type; unless the C++ langhook adds some extra > > hash table that caches already created OFFSET_TYPEs or something similar, > > it would create a new OFFSET_TYPE each time it is called > > build_offset_type already uses a hash table. Ah, ok. > > Also, I really don't know how well does GDB (especially older releases) > > handle DW_TAG_ptr_to_member_type for PMF, so the patch wraps that currently > > with if (dwarf_version >= 5). Quick grep revealed that GDB has code to > > handle the __pfn/__delta fields. So, can I ask somebody from the GDB > > team to test this patch with that if (dwarf_version >= 5) replaced > > with if (1) and see if it works properly with current GDB as well as say > > 4-5 years old one (e.g. with -gdwarf-2 or -gdwarf-3)? If yes, we > > should emit it unconditionally. > > This all makes sense to me. > > > + if (dwarf_version >= 5) > > + { > > + tree class_type = lang_hooks.types.get_ptrmemfn_type (type, 0); > > + if (class_type != NULL_TREE) > > This can be > > if (dwarf_version >= 5) > if (tree class_type = lang_hooks.types.get_ptrmemfn_type (type, 0)) Ok. Jakub