code review 10036044: runtime: add lr, ctxt, ret to Gobuf

runtime: add lr, ctxt, ret to Gobuf

Add gostartcall and gostartcallfn.
The old gogocall = gostartcall + gogo.
The old gogocallfn = gostartcallfn + gogo.

[dvyukov, 2013-06-06 09:36:50]

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c
File src/pkg/runtime/stack.c (right):

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c#new...
src/pkg/runtime/stack.c:281: runtime·gostartcall(&label, m->morepc, nil);
the context argument here must be m->cret
after changing it to:

	label.ret = 0;
	label.ctxt = nil;
	if(reflectcall)
		runtime·gostartcallfn(&label, (FuncVal*)m->morepc);
	else
		runtime·gostartcall(&label, m->morepc, (void*)m->cret);
	m->morepc = nil;
	m->cret = 0;
	runtime·gogo(&label);

it seems to work fine

[dvyukov, 2013-06-06 09:57:02]

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/asm_386.s
File src/pkg/runtime/asm_386.s (right):

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/asm_386.s#n...
src/pkg/runtime/asm_386.s:385: TEXT runtime·atomicstorep(SB), 7, $0
I am getting a weird compilation failure here:
./asm_386.s:385 syntax error, last name: runtime
If I insert /**/ before this line, it works
If I change runtime·atomicstorep to ·atomicstorep, it works
I am not sure it is related

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/mgc0.c
File src/pkg/runtime/mgc0.c (right):

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/mgc0.c#newc...
src/pkg/runtime/mgc0.c:1405: if(pc == (byte*)f->entry) {
The idea behind my CL that adds 'started' parameter to addframeroots
(https://codereview.appspot.com/9873045)
is to not let all this platform-dependent mess into GC code.
I want it to be situated in traceback code, and provide slightly higher-level
interface to GC code.

Preempted functions will have *not* adjusted SP, but PC != f->entry (PC points
to the point after morestack call).
So this condition will need to be updated in 2 places -- traceback and GC.

You wrote:
>This CL will need to be revised if Dmitriy cleans up my CL tweaking the
> goroutine representation. In that CL there is no longer a concept of 
> a non-started > goroutine. There are just functions on the stack that
> haven't yet bumped the SP to allocate a frame (those are identified by pc ==
f->entry).

My 'started' parameter means exactly that -- whether SP is adjusted or not. If
it does not adjust SP yet, it's not started in some sense. Probably it's a bad
name.
A better option may be to pass properly adjusted SP and FP to addframeroots.

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c
File src/pkg/runtime/stack.c (right):

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c#new...
src/pkg/runtime/stack.c:138: if(0) {
drop {} ;)

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/sys_x86.c
File src/pkg/runtime/sys_x86.c (right):

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/sys_x86.c#n...
src/pkg/runtime/sys_x86.c:28: runtime·gostartcall(gobuf, fv->fn, fv);
Can we just do this call inplace?

[rsc, 2013-06-10 19:03:46]

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/mgc0.c
File src/pkg/runtime/mgc0.c (right):

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/mgc0.c#newc...
src/pkg/runtime/mgc0.c:1405: if(pc == (byte*)f->entry) {
On 2013/06/06 09:57:02, dvyukov wrote:
> The idea behind my CL that adds 'started' parameter to addframeroots
> (https://codereview.appspot.com/9873045)
> is to not let all this platform-dependent mess into GC code.
> I want it to be situated in traceback code, and provide slightly higher-level
> interface to GC code.
> 
> Preempted functions will have *not* adjusted SP, but PC != f->entry (PC points
> to the point after morestack call).
> So this condition will need to be updated in 2 places -- traceback and GC.
> 
> You wrote:
> >This CL will need to be revised if Dmitriy cleans up my CL tweaking the
> > goroutine representation. In that CL there is no longer a concept of 
> > a non-started > goroutine. There are just functions on the stack that
> > haven't yet bumped the SP to allocate a frame (those are identified by pc ==
> f->entry).
> 
> My 'started' parameter means exactly that -- whether SP is adjusted or not. If
> it does not adjust SP yet, it's not started in some sense. Probably it's a bad
> name.
> A better option may be to pass properly adjusted SP and FP to addframeroots.

I am in favor of a cleaner fix, but I think it should be done in a separate CL.
I agree that perhaps passing in separate sp fp is best.

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c
File src/pkg/runtime/stack.c (right):

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c#new...
src/pkg/runtime/stack.c:138: if(0) {
On 2013/06/06 09:57:02, dvyukov wrote:
> drop {} ;)

Done.

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c#new...
src/pkg/runtime/stack.c:281: runtime·gostartcall(&label, m->morepc, nil);
On 2013/06/06 09:36:50, dvyukov wrote:
> the context argument here must be m->cret
> after changing it to:
> 
> 	label.ret = 0;
> 	label.ctxt = nil;
> 	if(reflectcall)
> 		runtime·gostartcallfn(&label, (FuncVal*)m->morepc);
> 	else
> 		runtime·gostartcall(&label, m->morepc, (void*)m->cret);
> 	m->morepc = nil;
> 	m->cret = 0;
> 	runtime·gogo(&label);
> 
> it seems to work fine

Ugh. Thanks. I added a comment too.

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/sys_x86.c
File src/pkg/runtime/sys_x86.c (right):

https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/sys_x86.c#n...
src/pkg/runtime/sys_x86.c:28: runtime·gostartcall(gobuf, fv->fn, fv);
On 2013/06/06 09:57:02, dvyukov wrote:
> Can we just do this call inplace?

I guess, but then there are two copies. Does it matter? This doesn't run very
often.

[dvyukov, 2013-06-10 19:21:41]

On 2013/06/10 19:03:46, rsc wrote:
> https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/mgc0.c
> File src/pkg/runtime/mgc0.c (right):
> 
>
https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/mgc0.c#newc...
> src/pkg/runtime/mgc0.c:1405: if(pc == (byte*)f->entry) {
> On 2013/06/06 09:57:02, dvyukov wrote:
> > The idea behind my CL that adds 'started' parameter to addframeroots
> > (https://codereview.appspot.com/9873045)
> > is to not let all this platform-dependent mess into GC code.
> > I want it to be situated in traceback code, and provide slightly
higher-level
> > interface to GC code.
> > 
> > Preempted functions will have *not* adjusted SP, but PC != f->entry (PC
points
> > to the point after morestack call).
> > So this condition will need to be updated in 2 places -- traceback and GC.
> > 
> > You wrote:
> > >This CL will need to be revised if Dmitriy cleans up my CL tweaking the
> > > goroutine representation. In that CL there is no longer a concept of 
> > > a non-started > goroutine. There are just functions on the stack that
> > > haven't yet bumped the SP to allocate a frame (those are identified by pc
==
> > f->entry).
> > 
> > My 'started' parameter means exactly that -- whether SP is adjusted or not.
If
> > it does not adjust SP yet, it's not started in some sense. Probably it's a
bad
> > name.
> > A better option may be to pass properly adjusted SP and FP to addframeroots.
> 
> I am in favor of a cleaner fix, but I think it should be done in a separate
CL.
> I agree that perhaps passing in separate sp fp is best.
> 
> https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c
> File src/pkg/runtime/stack.c (right):
> 
>
https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c#new...
> src/pkg/runtime/stack.c:138: if(0) {
> On 2013/06/06 09:57:02, dvyukov wrote:
> > drop {} ;)
> 
> Done.
> 
>
https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/stack.c#new...
> src/pkg/runtime/stack.c:281: runtime·gostartcall(&label, m->morepc, nil);
> On 2013/06/06 09:36:50, dvyukov wrote:
> > the context argument here must be m->cret
> > after changing it to:
> > 
> > 	label.ret = 0;
> > 	label.ctxt = nil;
> > 	if(reflectcall)
> > 		runtime·gostartcallfn(&label, (FuncVal*)m->morepc);
> > 	else
> > 		runtime·gostartcall(&label, m->morepc, (void*)m->cret);
> > 	m->morepc = nil;
> > 	m->cret = 0;
> > 	runtime·gogo(&label);
> > 
> > it seems to work fine
> 
> Ugh. Thanks. I added a comment too.
> 
> https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/sys_x86.c
> File src/pkg/runtime/sys_x86.c (right):
> 
>
https://codereview.appspot.com/10036044/diff/5001/src/pkg/runtime/sys_x86.c#n...
> src/pkg/runtime/sys_x86.c:28: runtime·gostartcall(gobuf, fv->fn, fv);
> On 2013/06/06 09:57:02, dvyukov wrote:
> > Can we just do this call inplace?
> 
> I guess, but then there are two copies. Does it matter? This doesn't run very
> often.

It does not matter too much.

The changes are not uploaded yet, right?

I think Ian must take a look as well.

[rsc, 2013-06-10 20:16:06]

PTAL

On 386, all.bash compiles and passes (thanks to CL 10167043).

On amd64, all.bash almost passes. It fails at fixedbugs/issue5493.go with not
all finalizers called (8 of 10). I can't tell if this is the CL's fault or the
test is just flaky.

I have not tested ARM.

[minux1, 2013-06-10 21:09:46]

Hi Russ,
after fixing these superficial errors on ARM, i got this from go_bootstrap:
releasep: m=0x329490 m->p=0x10612000 p->m=0x329490 m->mcache=0x400c9000
p->mcache=0x400c9000 p->status=3
fatal error: releasep: invalid p state

goroutine 1 [runnable]:

goroutine 2 [runnable]:

if you don't have time, i will take a closer look after i get some sleep.

https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s
File src/pkg/runtime/asm_arm.s (right):

https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s#...
src/pkg/runtime/asm_arm.s:104: MOVW	$0, gobuf_lr(R0)
please use a temporary register to assign 0 to memory (I used R11).

https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s#...
src/pkg/runtime/asm_arm.s:122: MOVQ	gobuf_ret(R0), R0
s/MOVQ/MOVW/

https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/sys_arm.c
File src/pkg/runtime/sys_arm.c (right):

https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/sys_arm.c#...
src/pkg/runtime/sys_arm.c:16: gobuf->pc = (uintptr)fn;
s/(uintptr)/(byte *)/

[dvyukov, 2013-06-10 21:16:54]

On 2013/06/10 20:16:06, rsc wrote:
> PTAL
> 
> On 386, all.bash compiles and passes (thanks to CL 10167043).
> 
> On amd64, all.bash almost passes. It fails at fixedbugs/issue5493.go with not
> all finalizers called (8 of 10). I can't tell if this is the CL's fault or the
> test is just flaky.

I think you need to zeroize gp->sched.ctxt in mcall or in gogo (or probably
runtime.goexit will do for now).
ctxt is void*, so when GC scans runtime.allg, it sees that pointer. Even if you
make it uintptr, G can be reached from other places and be scanned
conservatively, and then the closure will be leaked again.

[dvyukov, 2013-06-10 21:19:26]

On Tue, Jun 11, 2013 at 1:09 AM,  <minux.ma@gmail.com> wrote:
> Hi Russ,
> after fixing these superficial errors on ARM, i got this from
> go_bootstrap:
> releasep: m=0x329490 m->p=0x10612000 p->m=0x329490 m->mcache=0x400c9000
> p->mcache=0x400c9000 p->status=3
> fatal error: releasep: invalid p state

This is Pgcstop.
Can you get a stacktrace somehow? I usually traceback from runtime.startpanic.

>
> goroutine 1 [runnable]:
>
> goroutine 2 [runnable]:
>
> if you don't have time, i will take a closer look after i get some
> sleep.
>
>
> https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s
> File src/pkg/runtime/asm_arm.s (right):
>
>
https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s#...
> src/pkg/runtime/asm_arm.s:104: MOVW     $0, gobuf_lr(R0)
> please use a temporary register to assign 0 to memory (I used R11).
>
>
https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s#...
> src/pkg/runtime/asm_arm.s:122: MOVQ     gobuf_ret(R0), R0
> s/MOVQ/MOVW/
>
> https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/sys_arm.c
> File src/pkg/runtime/sys_arm.c (right):
>
>
https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/sys_arm.c#...
> src/pkg/runtime/sys_arm.c:16: gobuf->pc = (uintptr)fn;
> s/(uintptr)/(byte *)/
>
> https://codereview.appspot.com/10036044/

[rsc, 2013-06-11 01:41:36]

https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s
File src/pkg/runtime/asm_arm.s (right):

https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s#...
src/pkg/runtime/asm_arm.s:104: MOVW	$0, gobuf_lr(R0)
On 2013/06/10 21:09:46, minux wrote:
> please use a temporary register to assign 0 to memory (I used R11).

Done, but doesn't 5l do that for me?

https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s#...
src/pkg/runtime/asm_arm.s:122: MOVQ	gobuf_ret(R0), R0
On 2013/06/10 21:09:46, minux wrote:
> s/MOVQ/MOVW/

Done.

[rsc, 2013-06-11 04:14:24]

PTAL

amd64 and 386 are happy. arm is happy with the stack switches but seems to
collect some non-garbage, causing crashes.

[minux1, 2013-06-11 18:16:15]

On Tue, Jun 11, 2013 at 9:41 AM, <rsc@golang.org> wrote:

> https://codereview.appspot.**com/10036044/diff/16001/src/**
>
pkg/runtime/asm_arm.s#**newcode104<https://codereview.appspot.com/10036044/diff/16001/src/pkg/runtime/asm_arm.s#newcode104>
> src/pkg/runtime/asm_arm.s:104: MOVW     $0, gobuf_lr(R0)
> On 2013/06/10 21:09:46, minux wrote:
>
>> please use a temporary register to assign 0 to memory (I used R11).
>>
>
> Done, but doesn't 5l do that for me?

at least 5l doesn't do that now. Do you want it to be implemented?

[minux1, 2013-06-11 18:37:16]

On Tue, Jun 11, 2013 at 12:14 PM, Russ Cox <rsc@golang.org> wrote:

> amd64 and 386 are happy. arm is happy with the stack switches but seems to
> collect some non-garbage, causing crashes.
>
with Patch Set 7:

$ GOTRACEBACK=2 ../pkg/tool/linux_arm/go_bootstrap
fatal error: malloc/free - deadlock
[signal 0xb code=0x1 addr=0x24 pc=0x74144]

goroutine 1 [running]:
[fp=0x4029ab1c] runtime.throw(0x323470)
/home/minux/go.hg/src/pkg/runtime/panic.c:475 +0x50
[fp=0x4029ab50] runtime.mallocgc(0x8, 0x0, 0x1, 0x1)
/home/minux/go.hg/src/pkg/runtime/zmalloc_linux_arm.c:34 +0x54
[fp=0x4029ab64] runtime.mal(0x8)
 /home/minux/go.hg/src/pkg/runtime/zmalloc_linux_arm.c:658 +0x3c
[fp=0x4029ab80] copyin(0x1bbc78, 0x4029aba4, 0x4029abb0)
/home/minux/go.hg/src/pkg/runtime/iface.c:152 +0x60
[fp=0x4029ab9c] runtime.convT2E()
/home/minux/go.hg/src/pkg/runtime/iface.c:220 +0x44
[fp=0x4029abb4] runtime.newErrorString(0x322b08, 0x31, 0x4029abcc)
 /home/minux/go.hg/src/pkg/runtime/error.go:74 +0x38
[fp=0x4029abd4] runtime.panicstring(0x322b08)
/home/minux/go.hg/src/pkg/runtime/panic.c:488 +0x78
[fp=0x4029abe0] runtime.sigpanic()
/home/minux/go.hg/src/pkg/runtime/os_linux.c:237 +0x114
[fp=0x4029ac18] runtime.mallocgc(0x20, 0x0, 0x1, 0x1)
 /home/minux/go.hg/src/pkg/runtime/zmalloc_linux_arm.c:54 +0xd8
[fp=0x4029ac30] runtime.new()
/home/minux/go.hg/src/pkg/runtime/zmalloc_linux_arm.c:675 +0x50
[fp=0x4029ac84] text/template/parse.(*Tree).textOrAction(0x10612930, 0x12,
0x0)
/home/minux/go.hg/src/pkg/text/template/parse/parse.go:346 +0xd8
[fp=0x4029ad7c] text/template/parse.(*Tree).parse(0x10612930, 0x10622fe0,
0x0, 0x0)
/home/minux/go.hg/src/pkg/text/template/parse/parse.go:294 +0x330
[fp=0x4029adc4] text/template/parse.(*Tree).Parse(0x10612930, 0x2269f8,
0x8fc, 0x0, 0x0, ...)
/home/minux/go.hg/src/pkg/text/template/parse/parse.go:233 +0x198
[fp=0x4029ae00] text/template/parse.Parse(0x1dbe00, 0x4, 0x2269f8, 0x8fc,
0x0, ...)
/home/minux/go.hg/src/pkg/text/template/parse/parse.go:42 +0xe0
[fp=0x4029ae94] text/template.(*Template).Parse(0x10622f60, 0x2269f8,
0x8fc, 0x106234b0, 0x1d66a0, ...)
/home/minux/go.hg/src/pkg/text/template/template.go:171 +0x1a8
[fp=0x4029afac] main.init()
/home/minux/go.hg/src/cmd/go/vet.go:-4958 +0xfa8
[fp=0x4029afc4] runtime.main()
/home/minux/go.hg/src/pkg/runtime/proc.c:180 +0x70
[fp=0x4029afc8] runtime.goexit()
/home/minux/go.hg/src/pkg/runtime/proc.c:1237

goroutine 2 [syscall]:
runtime.entersyscallblock()
 /home/minux/go.hg/src/pkg/runtime/proc.c:1347 +0xc0
runtime.MHeap_Scavenger()
/home/minux/go.hg/src/pkg/runtime/mheap.c:470 +0xf8
runtime.goexit()
/home/minux/go.hg/src/pkg/runtime/proc.c:1237
created by runtime.main
/home/minux/go.hg/src/pkg/runtime/proc.c:166

goroutine 3 [syscall]:
runtime.entersyscallblock()
/home/minux/go.hg/src/pkg/runtime/proc.c:1347 +0xc0
created by os/signal.init·1
 /home/minux/go.hg/src/pkg/os/signal/signal_unix.go:27 +0x44

goroutine 4 [chan send]:
runtime.park(0x5f2d8, 0x1062353c, 0x3228a0)
 /home/minux/go.hg/src/pkg/runtime/proc.c:1185 +0x40
runtime.chansend(0x18fac0, 0x10623510, 0x402a7f6c, 0x0, 0x16e65c, ...)
/home/minux/go.hg/src/pkg/runtime/chan.c:227 +0x224
runtime.chansend1()
/home/minux/go.hg/src/pkg/runtime/chan.c:444 +0x44
text/template/parse.lexLeftDelim(0x10641600, 0x220a7c)
 /home/minux/go.hg/src/pkg/text/template/parse/lex.go:238 +0x1d8
text/template/parse.(*lexer).run(0x10641600)
/home/minux/go.hg/src/pkg/text/template/parse/lex.go:198 +0x44
runtime.goexit()
/home/minux/go.hg/src/pkg/runtime/proc.c:1237
created by text/template/parse.lex
/home/minux/go.hg/src/pkg/text/template/parse/lex.go:191 +0x118

The signal received is a SIGSEGV at pkg/runtime/malloc.goc:63,
and v = 0x24

Dmitry, do you have any ideas?

[rsc, 2013-06-11 18:52:41]

I'm pretty sure that the new code and the stack frame walker are not
getting along: if I turn the stack frame-based gc off, all is well. I'm
working on simplifying the stack frame walker a bit, which should help.

Russ

[rsc, 2013-06-12 18:00:05]

PTAL

all.bash passes on arm, amd64, and 386, so I think this is ready to go in.

[minux1, 2013-06-12 19:10:47]

LGTM.

https://codereview.appspot.com/10036044/diff/69001/src/pkg/runtime/runtime.h
File src/pkg/runtime/runtime.h (right):

https://codereview.appspot.com/10036044/diff/69001/src/pkg/runtime/runtime.h#...
src/pkg/runtime/runtime.h:212: // The offsets of these fields are known to
(hard-coded in) libmach.
update this comment (only offsets of the first three is hardcoded in libmach)?

btw, after cmd/prof and cmd/cov is removed, i think much of libmach is not
used anymore, should we remove them?

https://codereview.appspot.com/10036044/diff/69001/src/pkg/runtime/sys_x86.c
File src/pkg/runtime/sys_x86.c (right):

https://codereview.appspot.com/10036044/diff/69001/src/pkg/runtime/sys_x86.c#...
src/pkg/runtime/sys_x86.c:26: runtime·gostartcallfn(Gobuf *gobuf, FuncVal *fv)
should we move this function to a portable file?

[rsc, 2013-06-12 19:15:07]

Thanks for the review.

https://codereview.appspot.com/10036044/diff/69001/src/pkg/runtime/runtime.h
File src/pkg/runtime/runtime.h (right):

https://codereview.appspot.com/10036044/diff/69001/src/pkg/runtime/runtime.h#...
src/pkg/runtime/runtime.h:212: // The offsets of these fields are known to
(hard-coded in) libmach.
On 2013/06/12 19:10:48, minux wrote:
> update this comment (only offsets of the first three is hardcoded in libmach)?

Done.

> btw, after cmd/prof and cmd/cov is removed, i think much of libmach is not
> used anymore, should we remove them?

I would rather keep it for now. It still has functionality that we don't have a
replacement for, such as disassembly. Also we will always need it for 'go tool
nm' and the like.

https://codereview.appspot.com/10036044/diff/69001/src/pkg/runtime/sys_x86.c
File src/pkg/runtime/sys_x86.c (right):

https://codereview.appspot.com/10036044/diff/69001/src/pkg/runtime/sys_x86.c#...
src/pkg/runtime/sys_x86.c:26: runtime·gostartcallfn(Gobuf *gobuf, FuncVal *fv)
On 2013/06/12 19:10:48, minux wrote:
> should we move this function to a portable file?

Done. I've probably broken the arm now, but we'll see...

[rsc, 2013-06-12 19:22:22]

Hello dvyukov@google.com, minux.ma@gmail.com (cc: golang-dev@googlegroups.com),

I'd like you to review this change to
https://code.google.com/p/go/

[rsc, 2013-06-12 19:22:29]

*** Submitted as https://code.google.com/p/go/source/detail?r=916cb45c0d5f ***

runtime: add lr, ctxt, ret to Gobuf

Add gostartcall and gostartcallfn.
The old gogocall = gostartcall + gogo.
The old gogocallfn = gostartcallfn + gogo.

R=dvyukov, minux.ma
CC=golang-dev
https://codereview.appspot.com/10036044