[Bug ld/27491] ld: relocation R_X86_64_PC32 against undefined protected symbol `__start_xx' can not be used when making a shared object

[i at maskray dot me]

https://sourceware.org/bugzilla/show_bug.cgi?id=27491

--- Comment #3 from Fangrui Song <i at maskray dot me> ---
(In reply to H.J. Lu from comment #2)
> -z start-stop-gc is a bad idea.  LLVM should use __gc_start/__gc_stop.

I think the history is:

* In 2006-10, [BZ3400](https://sourceware.org/bugzilla/show_bug.cgi?id=3400)
reported that stdio flushing did not work with static linking &&
`--gc-sections`.
* In 2010-01, the problem was re-raised and gold got the rule
(<https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=f1ec9ded5c740c22735843025e5d3a8ff4c4079e>).
* [PR11133#c13](https://sourceware.org/bugzilla/show_bug.cgi?id=11133#c13)
installed a rule for GNU ld, but it did not appear to work or did not work with
other translation units.
* In 2015-10, the rule was properly installed
([PR19161](https://sourceware.org/bugzilla/show_bug.cgi?id=19161)
[PR19167](https://sourceware.org/bugzilla/show_bug.cgi?id=19167)).

I don't know what happened to the 2010-01 patch. It likely did not let __start_
reference retain magic sections in other translation unit (in metadata section
usage __start_/__stop_ references are always in a separate runtime file), so
"__start_ retaining all magic sections" was not effective before 2015-10.  Why
do we need new start/stop symbols? The orphan __start_/__stop_ usage worked
before 2015-10 and will work with -z start-stop-gc.

(I think of the possibility enabling -z start-stop-gc by default because the
rule was essentially invalid before 2015-10.)




As is, all clang -fprofile-generate and -fprofile-instr-generate instrumented
text sections cannot be GCed. (Fuchsia folks measured that GC can reduce binary
size by 28%.). Below is the story:

A function references its `__llvm_prf_cnts` and may reference its
`__llvm_prf_data` if value profiling applies.
The `__llvm_prf_data` references the text section, the associated
`__llvm_prf_cnts` and the associated `__llvm_prf_vals`.

If `__start_` references retain the `__llvm_prf_data` and the
`__llvm_prf_cnts`, the text section will be retained as well.
With GNU ld and gold all instrumented text sections cannot be discarded.

Another example is https://clang.llvm.org/docs/SanitizerCoverage.html#pc-table:
"Note: this instrumentation might be incompatible with dead code stripping
(-Wl,-gc-sections) for linkers other than LLD, thus resulting in a significant
binary size overhead. For more information, see Bug 34636."

I envision new metadata section usage will appear and those sections will need
a way for GC.

If people are interested a longer analysis, please check out
https://maskray.me/blog/2021-01-31-metadata-sections-comdat-and-shf-link-order

-- 
You are receiving this mail because:
You are on the CC list for the bug.