[ This is with an updated version of the patch series submitted here ( https://sourceware.org/pipermail/gdb-patches/2023-August/201887.html ). ] Consider the exec of test-case gdb.dwarf2/imported-unit-runto-main.exp, before adding the .gdb_index section. It has a backward inter-cu intra-shard reference: ... $ gdb -q -batch -iex "maint set worker-threads 0" imported-unit-runto-main handle_deferred_entries, intra-shard case parent map 0x00000000000000fb 0x2935500 (deferred) 0x00000000000000fc 0x0 parent valid map 0x000000000000002e 0x1 0x0000000000000045 0x0 0x0000000000000062 0x1 0x0000000000000084 0x0 0x00000000000000fb 0x1 0x0000000000000111 0x0 0x0000000000000149 0x1 0x00000000000002d4 0x0 handle_deferred_entries, inter-shard case 0x00000000000000fb 0x2935500 (deferred) 0x00000000000000fc 0x0 Resolve deferred: 0xfb -> 0xd5: no parent ... The dependency is resolved in the inter-shard case, which assumes that all CUs have been processed, and consequently doesn't check the parent valid map. But if we look at the parent valid map, we don't find 0xd5. So the "no parent" is not a case of, we've processed 0xd5 and found that is has no parent, but we've not processed the PU containing 0xd5. The problem is that we have a PU without import, and consequently no CU wins the race to read it. This doesn't cause a problem with the current implementation (because "not processed" and "no parent" look the same), but we can reproduce the same problem with test-case forward-spec-inter-cu.exp. By changing one of the cus into a pu: ... --- a/gdb/testsuite/gdb.dwarf2/forward-spec-inter-cu.exp +++ b/gdb/testsuite/gdb.dwarf2/forward-spec-inter-cu.exp @@ -48,7 +48,7 @@ Dwarf::assemble $asm_file { } cu {} { - DW_TAG_compile_unit { + DW_TAG_partial_unit { {DW_AT_language @DW_LANG_C_plus_plus} } { declare_labels myint ... we get: ... FAIL: gdb.dwarf2/forward-spec-inter-cu.exp: v has a parent ...
Note that readnow has no problem with the dwarf, we simply get: ... (gdb) p ns::v $1 = 23 ... but with the cooked index: ... (gdb) p ns::v No symbol "ns" in current context. ...
(In reply to Tom de Vries from comment #1) > Note that readnow has no problem with the dwarf, we simply get: > ... > (gdb) p ns::v > $1 = 23 > ... > but with the cooked index: > ... > (gdb) p ns::v > No symbol "ns" in current context. > ... Same problem with gdb-12-branch.
I have a hard time reasoning about how this dwarf should be interpreted. Perhaps this PR should be about unsupporting it for -readnow.
I tend to think a PU that is not included just indicates a bad DWARF writer. Maybe gdb could issue a complaint, if it's easy. Really a job for (the mythical) dwarflint.
The DWARF spec says: • A partial compilation unit describes a part of a compilation (generally corresponding to an imported module) which is imported into one or more related full compilation units. So I think that a non-imported PU can safely be ignored. The full reader could be changed to do this as well.