Il 19/09/2013 10:11, Stefan Hajnoczi ha scritto:
> On Thu, Sep 19, 2013 at 8:36 AM, Taimoor Mirza <taimoor.m...@gmail.com> wrote:
>
> Please use Reply-All to keep the CC list on your reply email. This
> way others on the mailing list can follow the discussion and
> participate.
>
>> O1 disassembly is at http://pastebin.com/yxWAmmmf
>>
>> O2 disassembly is at http://pastebin.com/v0uqHuYY
>
> I diffed them. The differences look reasonable, nothing obviously broken:
> * The broken -O2 version inlines qemu_coroutine_switch() into
> coroutine_trampoline().
> * The broken -O2 version changes the ordering of basic blocks and
> switches a few instructions for shorter/cheaper equivalents.
>
> Taimoor: Can you try "make check" with the -O2 version? In
> particular, the test-coroutine tests may fail and that could help us
> figure out which operation is broken.
>
> Paolo: CCed you in case you want to diff the working -O1 against the
> broken -O2 disassembly. Perhaps Taimoor's gcc 4.6.2 -O2 subtly breaks
> around the Windows Fiber API calls.
Taimoor, can you please try compiling coroutine-win32.c at
-O2 but without the -fstack-protector-all compiler option? If it works,
please try with -fstack-protector instead of -fstack-protector-all.
Also (separately from the above experiment), can you try this patch and
see if it breaks/works at -O1 and -O2? Then, independent of the result,
please post -O1 and -O2 assembly outputs with the patch.
diff --git a/coroutine-win32.c b/coroutine-win32.c
index edc1f72..3b10189 100644
--- a/coroutine-win32.c
+++ b/coroutine-win32.c
@@ -55,7 +55,13 @@ static void CALLBACK coroutine_trampoline(void *co_)
while (true) {
co->entry(co->entry_arg);
- qemu_coroutine_switch(co, co->caller, COROUTINE_TERMINATE);
+ Coroutine *to_ = co->caller;
+ CoroutineWin32 *from = DO_UPCAST(CoroutineWin32, base, co);
+ CoroutineWin32 *to = DO_UPCAST(CoroutineWin32, base, to_);
+
+ current = to_;
+ to->action = COROUTINE_TERMINATE;
+ SwitchToFiber(to->fiber);
}
}
Thanks,
Paolo
