[Bug ld/32387] New: ppc64 TLS local-dynamic optimization bug when built with -fno-plt

[bergner at linux dot ibm.com]

https://sourceware.org/bugzilla/show_bug.cgi?id=32387

            Bug ID: 32387
           Summary: ppc64 TLS local-dynamic optimization bug when built
                    with -fno-plt
           Product: binutils
           Version: unspecified
            Status: NEW
          Severity: normal
          Priority: P2
         Component: ld
          Assignee: unassigned at sourceware dot org
          Reporter: bergner at linux dot ibm.com
  Target Milestone: ---

We SEGV on the following simple test case when our library function is compiled
with -fno-plt.

bergner@c643n10lp1:bug$ cat libtlsdso.c
unsigned long *
dsofunc_ld (void)
{
  static _Thread_local __attribute ((tls_model("local-dynamic"))) unsigned long
tlsvar = 0xdeadbeef;
  return &tlsvar;
}

bergner@c643n10lp1:bug$ cat prog.c
#include <stdio.h>
extern unsigned long *dsofunc_ld (void);
int
main (void)
{
  unsigned long *var = dsofunc_ld ();
  printf ("0x%08x\n", *var);
  return 0;
}

gcc -O2 -fPIC -fplt -c -o libtlsdso.o libtlsdso.c
gcc -O2 -fPIC -shared -Wl,-soname=libtlsdso.so -o libtlsdso.so libtlsdso.o
gcc -O2 -R./ -L./ prog.c -o prog -ltlsdso
bergner@c643n10lp1:bug$ ./prog
0xdeadbeef


gcc -O2 -fPIC -fno-plt -c -o libtlsdso.o libtlsdso.c
gcc -O2 -fPIC -shared -Wl,-soname=libtlsdso.so -o libtlsdso.so libtlsdso.o
gcc -O2 -R./ -L./ prog.c -o prog -ltlsdso
bergner@c643n10lp1:bug$ ./prog
Segmentation fault (core dumped)

In the working case (-fplt), the call to __tls_get_addr calls through the PLT
call stub with tld module id = 0 and offset = 4096 (ie, this is a local-dynamic
optimization).  The PLT call stub for __tls_get_addr is "special" in that is
has special code to look for module id == 0 and it just returns the offset in
that case.

In the failing case, the inline PLT call skips the PLT stub and goes directly
to __tls_get_addr which assumes the module id == 0 case has been handled, so
things go badly.

I see a couple of solutions, 1) The inline plt call stub could be expanded to
handle the module id == 0 case itself or 2) Change GCC so that calls to
__tls_get_addr don't use inline PLT call stubs and instead use the normal PLT
mechanism or 3) The linker could emit the special __tls_get_addr PLT stub and
force the inline PLT call to branch to the special PLT stub rather than
__tls_get_addr itself. 

Did I miss any other options?

Option 1 would give the highest performing code, but 1 and 2 seem to go against
Alan's original design decision of not needing any GCC changes for the
local-dynamic optimization.

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