Enabling -gsplit-dwarf by default and trying a build hits an assert in dw2_asm_output_delta_uleb128 because HAVE_AS_LEB128 is not defined. The problem appears to be in output_loc_list in dwarf2out.c which has in the dwarf_split_debug_info code /* FIXME: This will ICE ifndef HAVE_AS_LEB128. For that case we probably need to emit DW_LLE_startx_endx, but we'd need 2 .debug_addr entries rather than just one. */ riscv doesn't allow leb128 because of agressive linker relaxation, so we need the alternative solution here that works without HAVE_AS_LEB128.
Is there any reason why can't .{u,s}leb128 be supported in .debug* sections when referring to labels in .debug* sections (where I don't see how aggressive linker relaxation should apply to those sections, those sections heavily rely on no linker relaxation)? So essentially have partial as leb128 support instead of none at all?
Yes we could have partial uleb128 support. There is only a problem if at least one label is in the code section. There is another proposed solution to add special relaxable relocations for uleb128 but the initial proposal had flaws, and no one has reviewed the second proposal yet. Or we could change the -gsplit-dwarf support to work even when there is no uleb128 support.
Making it work is desirable in any case. But, the emitted debug sections will be in various spots less compact when .uleb128/.sleb128 don't support non-constant operands. Though, I'm afraid the /* FIXME: This will ICE ifndef HAVE_AS_LEB128. For that case we probably need to emit DW_LLE_startx_endx, but we'd need 2 .debug_addr entries rather than just one. */ dw2_asm_output_delta_uleb128 (curr->end, curr->begin, "Location list length (%s)", list_head->ll_symbol); case is actually code addresses and not .debug section labels, sorry for confusion.
Created attachment 50265 [details] gcc11-pr99090.patch Untested fix.
I tested it with a riscv-gnu-toolchain build and check. The 4 -gsplit-dwarf testcases that fail without the patch work with the patch. I also tried a build and check with -gsplit-dwarf enabled by default and discovered that there are a number of debug tests that fail simply because the output is a little different than what is expected. But nothing else appeared to fail.
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>: https://gcc.gnu.org/g:b5040344b9ca609e19ee59ba56cd4af9697a1692 commit r11-7472-gb5040344b9ca609e19ee59ba56cd4af9697a1692 Author: Jakub Jelinek <jakub@redhat.com> Date: Wed Mar 3 09:53:58 2021 +0100 dwarf2out: Fix -gsplit-dwarf on riscv or other non-.uleb128 targets [PR99090] As mentioned in the PR, riscv* only supports .uleb128 with constant arguments, doesn't support difference of two labels because of aggressive linker relaxations. But I bet various other targets, especially those not using GNU assembler, might suffer from the same problem. As the FIXME comment in output_loc_list indicates, we ICE on -gsplit-dwarf on those targets whenever we need .debug_loclists, because we only emit DW_LLE_startx_length which requires working .uleb128 delta of 2 code section labels. We can't use DW_LLE_base_addressx once followed by DW_LLE_offset_pair either because the latter suffers from the same issue - need .uleb128 difference of code section labels (and in that case not just for the second operand but also for the first). So, this patch implements what the comment said and emits DW_LLE_startx_endx instead, which wastes more space in .debug_addr, but will work. Bootstrapped/regtested on x86_64-linux and i686-linux and as written in the PR, Jim has tested it on riscv*linux. Ok for trunk? BTW, for HAVE_AS_LEB128 -gdwarf-5 -gsplit-dwarf, maybe we should consider instead of always emitting DW_LLE_startx_length do all the optimizations that we do for HAVE_AS_LEB128 -gdwarf-5, or at least a subset of them. For !have_multiple_function_sections, we in that case emit just DW_LLE_offset_pair (that can certainly be a win for small TUs, we wouldn't need any .debug_addr entry in that case; on the other side, just using DW_LLE_offset_pair can be harmful for very large TUs especially if the loclist has many entries, emitting in that case a single DW_LLE_base_address or for -gsplit-dwarf DW_LLE_base_addressx followed by DW_LLE_offset_pair might be much smaller), and for have_multiple_function_sections figuring out if DW_LLE_base_address followed by DW_LLE_offset_pair entries or DW_LLE_start_length is bettter. So perhaps a middle-ground for -gsplit-dwarf would be to always do the have_multiple_function_sections behavior, i.e. DW_LLE_base_addressx followed by DW_LLE_offset_pair vs. DW_LLE_startx_length decisions based on the ranges and their counts. And perhaps dwz could optimize afterwards, on linked binaries or shared libraries it knows all the offsets and could figure out optimal DW_LLE_*. 2021-03-03 Jakub Jelinek <jakub@redhat.com> PR debug/99090 * dwarf2out.c (dw_loc_list_struct): Add end_entry member. (new_loc_list): Clear end_entry. (output_loc_list): Only use DW_LLE_startx_length for -gsplit-dwarf if HAVE_AS_LEB128, otherwise use DW_LLE_startx_endx. Fix comment typo. (index_location_lists): For dwarf_version >= 5 without HAVE_AS_LEB128, initialize also end_entry.
Fixed on the trunk so far.
*** Bug 102290 has been marked as a duplicate of this bug. ***
It was fixed on the trunk, in time for gcc-12. I can't see that we're likely to backport to gcc-11 or earlier. So closing as fixed.