Tue, Dec 27, 2005 at 08:11:47 +0000, Nick Hudson wrote:

not sure I've got sh3 right?

sh3 also needs r12 to be set up before doing plt calls. You can leave
sh3 out and I will take care of it.

Index:

RCS file: /,v
retrieving revision 1.10
diff -u -p -u -r1.10 pthread_switch.S
23 Apr 2004 02:58:27 -00001.10
27 Dec 2005 07:12:17 -0000
@@ -136,7 +136,11 @@ ENTRY(pthread__switch)
* Edit the context so that it continues as if returning from
* the _setcontext_u below.
*/
+#ifdef PIC
+movqPIC_GT(pthread__switch_return_point), %r15
+#else
leaqpthread__switch_return_point(%rip), %r15
+#endif
movq%r15, UC_RIP(%r14)
movq%r14, PT_UC(%r12)

@@ -157,7 +161,6 @@ NENTRY(pthread__switch_return_point)
* Helper switch code used by pthread__locked_switch() and
* pthread__upcall_switch() when they discover spinlock preemption.
*/
-.globl pthread__switch_away
pthread__switch_away:
STACK_SWITCH

@@ -204,7 +207,11 @@ ENTRY(pthread__locked_switch)
* Edit the context so that it continues as if returning from
* the _setcontext_u below.
*/
+#ifdef PIC
+movqPIC_GT(locked_return_point), %r15
+#else
leaqlocked_return_point(%rip), %r15
+#endif
movq%r15, UC_RIP(%r14)
movq%r14, PT_UC(%r12)

I think these *_return_point symbols should be made .hidden, then
these diffs will not be necessary. As with cerror, I don't see a
reason why we would want to make them globally visible/overridable.

SY, Uwe

Tue, Dec 27, 2005 at 11:38:05 +0300, Valeriy E. Ushakov wrote:

Index:

RCS file: /,v
retrieving revision 1.10
diff -u -p -u -r1.10 pthread_switch.S
23 Apr 2004 02:58:27 -00001.10
27 Dec 2005 07:12:17 -0000
@@ -136,7 +136,11 @@ ENTRY(pthread__switch)
* Edit the context so that it continues as if returning from
* the _setcontext_u below.
*/
+#ifdef PIC
+movqPIC_GT(pthread__switch_return_point), %r15
+#else
leaqpthread__switch_return_point(%rip), %r15
+#endif
movq%r15, UC_RIP(%r14)
movq%r14, PT_UC(%r12)

I think these *_return_point symbols should be made .hidden, then
these diffs will not be necessary. As with cerror, I don't see a
reason why we would want to make them globally visible/overridable.

I did a quick review of the arch/*/pthread_switch.S and the situation
seems to be quite messy.

pthread__switch_return_point is mentioned in pthread_sa.c as:

extern void pthread__switch_return_point(void);

In pthread__upcall() the code checks if:

psrp = pthread__switch_return_point;
pc = (fptr_t)((intptr_t)
pthread__uc_pc(victim->pt_uc));
if ((victim->pt_spinlocks == 0) &&
((victim->pt_switchto != NULL) ||
(pc == psrp))) {

}

Now, my reading of the pthread_switch.S files is that most of the
architectures set the pc (pthread__uc_pc(victim->pt_uc)) to the
absolute address of the pthread__switch_return_point label during the
pthread__switch. But psrp = pthread__switch_return_point assignment
will set psrp to the PLT entry instead, b/c pthread__switch_return_point
is not .hidden, so (pc == psrp) will never be true.

PS: Then there's VAX on which pthread__switch_return_point is *not*
the label that pthread__switch use, so on VAX that (pc == psrp) will
never be true even for static linking, or pthread__switch_return_point
marked as .hidden.

SY, Uwe