Re: [PATCH] parisc: futex: Use same lock set as lws calls
On 10/21/2011 10:49 AM, Carlos O'Donell wrote:
On Wed, Oct 19, 2011 at 9:47 PM, John David Anglin<dave.anglin@bell.net> wrote:
I looked at this a bit. The problem is legitimize_tls_address doesn't work
properly. First, GCC doesn't know that the libcall needs r26 and ret0 when the
__thread variable is an argument to a call. Secondly, the implementation of __tls_get_addr
clobbers some other call clobbered registers. I'm thinking the glibc implementation
might need to be in assembly language so that the clobbered registers are limited (i.e.,
it needs to save registers). This all seems really ugly...
The issues occur in generating PIC code.
Yes, __tls_get_addr is a normal C function and follows normal function
register usage.
Yes. The issue here is expand_call doesn't appear to be aware that
loading the
address of a __thread variable may require a call. Thus, it just emits
an insn to
load r26. This load is dead because r26 is reloaded for the call to
__tls_get_addr.
The first argument for the call needs to be saved to a temp, and then
r26 needs
to be loaded just before the actual call. This is works fine for normal
calls.
Why does __tls_get_addr need to be a special function?
Almost all the targets in glibc-ports have C versions of __tls_get_addr.
For example I notice that on Alpha the call to __tls_get_addr is *not*
done via a emit_library_call_value, instead they use some emit_insn,
emit_libcall_block and use_reg.
use_reg won't help. This is attached to the call insn and only comes
into play
during register allocation. The problem occurs during rtl expansion.
So it looks like they tamper with the register usage via use_reg
before the call to __tls_get_addr?
Another example is Sparc which also uses a custom sequence and
manipulates the used registers.
sparc64 tls support is also broken... See,
http://www.loria.fr/~zimmerma/software/compilerbugs.html
I think our legitimize_tls_address needs to be rewritten to match
something like what alpha or sparc has, otherwise we are going to run
into trouble trying to get emit_library_call_value to work correctly.
Dave
--
John David Anglin dave.anglin@bell.net
Reply to: