Fix --gc-sections for C++ MIPS ELF
Sun Dec 2 20:15:00 GMT 2007
--Wl,--gc-sections generates wrong code for C++ on MIPS ELF. The problem
is that if:
(a) all C++ CIEs have relocations against the personality routine
itself, rather relocations against a pointer to the personality
(b) as usual, there are no references to the personality routine's
section outside .eh_frame
then we never mark the personality routine's section as needed.
The main code for marking .eh_frames is:
/* Keep .gcc_except_table.* if the associated .text.* (or the
associated .gnu.linkonce.t.* if .text.* doesn't exist) is
marked. This isn't very nice, but the proper solution,
splitting .eh_frame up and using comdat doesn't pan out
easily due to needing special relocs to handle the
difference of two symbols in separate sections.
Don't keep code sections referenced by .eh_frame. */
#define TEXT_PREFIX ".text."
#define TEXT_PREFIX2 ".gnu.linkonce.t."
#define GCC_EXCEPT_TABLE_PREFIX ".gcc_except_table."
for (o = sub->sections; o != NULL; o = o->next)
if (!o->gc_mark && o->gc_mark_from_eh && (o->flags & SEC_CODE) == 0)
and it ignores all relocations against code.
Non-PIC CIEs for targets like x86_64-linux-gnu also have direct references
to the personality routines. I suspect the only reason -Wl,--gc-sections
-static-libgcc works for them is that libgcc's own CIEs use an indirect
reference, so the section gets marked that way.
A simple fix would be to use indirect references for MIPS ELF too,
but it would be nice to avoid the overhead. It would also be nice to
make binutils work with older GCCs if possible.
Another reason to prefer a binutils change is that the current code
seems unsafe in ways that aren't explicitly mentioned in the comment
above. The marking code assumes that FDEs against discarded sections will
themselves be discarded, but if an .eh_frame section contains something
unexpected, like a future augmentation type, elf-eh-frame.c will keep
the section as-is. We would then get a link failure or silent wrong code.
Has there been any talk about parsing the CIEs and FDEs and marking
relocations against them individually? I couldn't find anything in the
archives, so I gave it a go, and it seems to work. The main problem was
that we currently don't parse the .eh_frame information until later in
the link process, so a fair amount of rejigging was needed.
Once we have access to the parsed information, we can mark an .eh_frame
(i) it is associated with an FDE that is itself associated with
a marked section or
(ii) it is associated with the CIE for such an FDE.
I tested the changes by running the C++ and libstdc++-v3 testsuites
on x86_64-linux-gnu and mipsisa64-elf with -Wl,--gc-sections.
The C++ PCH tests failed, but the results were otherwise identical
to those without -Wl,--gc-sections. There were also no regressions
in the binutils, gas and ld testsuites for x86_64-linux-gnu,
mipsisa64-elf and mips64-linux-gnu.
Because several changes are needed, I've tried to split things up for
ease of review. I'll post each patch as a follow-up.
More information about the Binutils