http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55354
--- Comment #9 from Konstantin Serebryany <konstantin.s.serebryany at gmail dot
com> 2012-11-18 19:35:43 UTC ---
As dvyuokv@ pointed out,
-ftls-model=initial-exec improves the situation, but does not fully help.
Experiment:
% cat x.c
__thread int a;
int foo() {
return a;
}
HORRIBLE: -fPIC -shared
% gcc x.c -O2 -fPIC -shared -o x.so ; objdump -d x.so | grep foo.: -A 5
00000000000006e0 <foo>:
6e0: 66 48 8d 3d f0 08 20 lea 0x2008f0(%rip),%rdi # 200fd8
<_DYNAMIC+0x1b8>
6e7: 00
6e8: 66 66 48 e8 10 ff ff callq 600 <__tls_get_addr@plt>
6ef: ff
6f0: 8b 00 mov (%rax),%eax
NOT-SO-BAD: -fPIC -shared -ftls-model=initial-exec
% gcc x.c -O2 -fPIC -shared -o x.so -ftls-model=initial-exec ; objdump -d x.so
| grep foo.: -A 5
0000000000000630 <foo>:
630: 48 8b 05 a9 09 20 00 mov 0x2009a9(%rip),%rax # 200fe0
<_DYNAMIC+0x1b8>
637: 64 8b 00 mov %fs:(%rax),%eax
63a: c3 retq
GOOD: -fPIE
% gcc -c x.c -O2 -fPIE -o x.o ; objdump -d x.o | grep foo.: -A 5
0000000000000000 <foo>:
0: 64 8b 04 25 00 00 00 mov %fs:0x0,%eax
7: 00
8: c3 retq
So, while -ftls-model=initial-exec improves the TLS performance, it is still
2x slower than -fPIE.
For tsan, which does this for *every* memory access in the original program,
this will cost 5%-10% slowdown.
For our users this is a big deal, so they will link the static library whenever
possible. Which default is used in gcc -- I don't care that much.
