With valgrind checking, there are various errors reported on some C++26
libstdc++ tests, like:
==
2009913== Conditional jump or move depends on uninitialised value(s)
==
2009913== at 0x914C59: gt_ggc_mx_lang_tree_node(void*) (gt-cp-tree.h:107)
==
2009913== by 0x8AB7A5: gt_ggc_mx_tinst_level(void*) (gt-cp-pt.h:32)
==
2009913== by 0xB89B25: ggc_mark_root_tab(ggc_root_tab const*) (ggc-common.cc:75)
==
2009913== by 0xB89DF4: ggc_mark_roots() (ggc-common.cc:104)
==
2009913== by 0x9D6311: ggc_collect(ggc_collect) (ggc-page.cc:2227)
==
2009913== by 0xDB70F6: execute_one_pass(opt_pass*) (passes.cc:2738)
==
2009913== by 0xDB721F: execute_pass_list_1(opt_pass*) (passes.cc:2755)
==
2009913== by 0xDB7258: execute_pass_list(function*, opt_pass*) (passes.cc:2766)
==
2009913== by 0xA55525: cgraph_node::analyze() (cgraphunit.cc:695)
==
2009913== by 0xA57CC7: analyze_functions(bool) (cgraphunit.cc:1248)
==
2009913== by 0xA5890D: symbol_table::finalize_compilation_unit() (cgraphunit.cc:2555)
==
2009913== by 0xEB02A1: compile_file() (toplev.cc:473)
I think the problem is in the tinst_level::to_list optimization from 2018.
That function returns a TREE_LIST with TREE_PURPOSE/TREE_VALUE filled in.
Either it freshly allocates using build_tree_list (NULL, NULL); + stores
TREE_PURPOSE/TREE_VALUE, that case is fine (the whole tree_list object
is zeros, except for TREE_CODE set to TREE_LIST and TREE_PURPOSE/TREE_VALUE
modified later; the above also means in particular TREE_TYPE of it is NULL
and TREE_CHAIN is NULL and both are accessible/initialized even in valgrind
annotations.
Or it grabs a TREE_LIST node from a freelist.
If defined(ENABLE_GC_CHECKING), the object is still all zeros except
for TREE_CODE/TREE_PURPOSE/TREE_VALUE like in the fresh allocation case
(but unlike the build_tree_list case in the valgrind annotations
TREE_TYPE and TREE_CHAIN are marked as uninitialized).
If !defined(ENABLE_GC_CHECKING), I believe the actual memory content
is that everything but TREE_CODE/TREE_PURPOSE/TREE_VALUE/TREE_CHAIN is
zeros and TREE_CHAIN is something random (whatever next entry is in the
freelist, nothing overwrote it) and from valgrind POV again,
TREE_TYPE and TREE_CHAIN are marked as uninitialized.
When using the other freelist instantiations (pending_template and
tinst_level) I believe everything is correct, from valgrind POV it marks
the whole pending_template or tinst_level as uninitialized, but the
caller initializes it all).
One way to fix this would be let tinst_level::to_list not store just
TREE_PURPOSE (ret) = tldcl;
TREE_VALUE (ret) = targs;
but also
TREE_TYPE (ret) = NULL_TREE;
TREE_CHAIN (ret) = NULL_TREE;
Though, that seems like wasted effort in the build_tree_list case to me.
So, the following patch instead does that TREE_CHAIN = NULL_TREE store only
in the case where it isn't already done (and likewise for TREE_TYPE just to
be sure) and marks both TREE_CHAIN and TREE_TYPE as initialized (the latter
is at that spot, the former is because we never really touch TREE_TYPE of a
TREE_LIST anywhere and so the NULL gets stored into the freelist and
restored from there (except for ENABLE_GC_CHECKING where it is poisoned
and then cleared again).
2023-12-14 Jakub Jelinek <jakub@redhat.com>
PR c++/112968
* pt.cc (freelist<tree_node>::reinit): Make whole obj->common
defined for valgrind annotations rather than just obj->base,
and do it even for ENABLE_GC_CHECKING. If not ENABLE_GC_CHECKING,
clear TREE_CHAIN (obj) and TREE_TYPE (obj).