[Bug tree-optimization/66826] Unused result from dlsym in constructor results in a segfault

[pinskia at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66826

Andrew Pinski  changed:

   What|Removed |Added

URL|https://gist.github.com/dau |
   |rnimator/a468e01800752d11cd |
   |15  |
 Status|UNCONFIRMED |NEW
   Last reconfirmed||2021-09-02
 Ever confirmed|0   |1

--- Comment #7 from Andrew Pinski  ---
https://gist.github.com/daurnimator/a468e01800752d11cd15

Confirmed.

[Bug tree-optimization/66826] Unused result from dlsym in constructor results in a segfault

[tetra2005 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66826

--- Comment #6 from Yuri Gribov  ---
(In reply to Rich Felker from comment #5)
> maybe there are workarounds glibc could do to prevent tco without needing a
> new attribute...

X-posted to Glibc BZ: https://sourceware.org/bugzilla/show_bug.cgi?id=21050

[Bug tree-optimization/66826] Unused result from dlsym in constructor results in a segfault

[bugdal at aerifal dot cx]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66826

Rich Felker  changed:

   What|Removed |Added

 CC||bugdal at aerifal dot cx

--- Comment #5 from Rich Felker  ---
I think the issue is more complicated. Even if glibc were fixed not to crash,
code like the following:

return dlsym(RTLD_NEXT, "whatever");

would return the wrong result under tco when the caller's caller is in a
different dso. GCC probably needs a "notailcall" attribute to fix this, but
maybe there are workarounds glibc could do to prevent tco without needing a new
attribute...

[Bug tree-optimization/66826] Unused result from dlsym in constructor results in a segfault

[tetra2005 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66826

--- Comment #4 from Yuri Gribov  ---
As this is not a GCC bug I suggest you
* close this issue (as not-a-bug?)
* report to Glibc folks (perhaps they could do more checking of return address
or at least document their calling convention assumptions in manpages)

[Bug tree-optimization/66826] Unused result from dlsym in constructor results in a segfault

[tetra2005 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66826

Yuri Gribov  changed:

   What|Removed |Added

 CC||tetra2005 at gmail dot com

--- Comment #3 from Yuri Gribov  ---
This is a very funny bug but not related to GCC per se. Firstly, let's consider
a miminal repro:
__attribute__((constructor)) static void some_init() {
  dlsym(RTLD_DEFAULT, "anything");
}
(segfaults just as well). Under -O0 this produces a normal call:
calldlsym@PLT
...
ret
but with -O2 GCC is clever enough to tail-call-optimize it to a plain jump:
jmp dlsym@PLT

Now dlsym (and other dl-functions) secretly take shadow parameter - return
address on stack:
void *
__dlsym (void *handle, const char *name DL_CALLER_DECL)
{
...
  struct dlsym_args args;
  args.who = DL_CALLER;
  args.handle = handle;
  args.name = name;
(from dlsym.c). As in our case return address is missing, args.who argument is
missing which causes segfault during symbol resolution (dynamic linker is lame
on checks).