[MIPS] Fix constraint modifier for syscall clone

[MIPS] Fix constraint modifier for syscall clone

__v0 has to be marked as read-write rather than early-clobbered.
This fixes an issue in which return value from the system call is
unexpectedly overridden.

When v0 is marked as early-clobbered,
(asm) "li    %0,%2\n"

will be coded as:
(objdump) "58:    2410ffea    li    s0,-22"

and at label '1' the value is restored to v0:
(objdump) "b0:    02001021    move  v0,s0"

However, this clobbers the return value from the syscall.

When v0 is marked as read-write, (-EINVAL) is correctly loaded in v0:

(objdump) "58:    2402ffea    li    v0,-22"

and the value is not overridden later.

BUG = described above
TEST= used by breakpad for MIPS
Committed: https://code.google.com/p/linux-syscall-support/source/detail?r=21

[Markus (顧孟勤), 2013-04-09 00:20:51]

This looks fine, I think.

But I really would appreciate a more verbose checkin comment explaining what the
observed buggy behavior was, and how this change fixes it.

[Mark Seaborn, 2013-04-09 00:29:55]

LGTM.  Can you list the faulty assembly output that GCC produced in the commit
message?

https://codereview.chromium.org/13846002/diff/1/lss/linux_syscall_support.h
File lss/linux_syscall_support.h (left):

https://codereview.chromium.org/13846002/diff/1/lss/linux_syscall_support.h#o...
lss/linux_syscall_support.h:2445: register unsigned long __v0 __asm__("$2");
Background question (since I'm not very familiar with the MIPS code here):  Why
does this code use reg numbers ($2) rather than names ($v0)?

https://codereview.chromium.org/13846002/diff/1/lss/linux_syscall_support.h#o...
lss/linux_syscall_support.h:2447: {
Background question: Why is there a block here?

[petarj, 2013-04-09 12:17:40]

I have added additional description about the issue in the commit message.

https://codereview.chromium.org/13846002/diff/1/lss/linux_syscall_support.h
File lss/linux_syscall_support.h (left):

https://codereview.chromium.org/13846002/diff/1/lss/linux_syscall_support.h#o...
lss/linux_syscall_support.h:2445: register unsigned long __v0 __asm__("$2");
On 2013/04/09 00:29:55, Mark Seaborn wrote:
> Background question (since I'm not very familiar with the MIPS code here): 
Why
> does this code use reg numbers ($2) rather than names ($v0)?

Older binutils could not recognize symbolic register names, so I assume
whoever wrote the original code had an older version of binutils or wanted
to support older versions.

https://codereview.chromium.org/13846002/diff/1/lss/linux_syscall_support.h#o...
lss/linux_syscall_support.h:2447: {
On 2013/04/09 00:29:55, Mark Seaborn wrote:
> Background question: Why is there a block here?

I do not know. It does not seem like it is needed.

[petarj, 2013-04-09 12:30:20]

For anyone interested in the faulty code for sys_clone (that I refer to in the
commit message), here it is:

00000000 <sys_clone>:
   0:   3c1c0000        lui     gp,0x0
   4:   279c0000        addiu   gp,gp,0
   8:   0399e021        addu    gp,gp,t9
   c:   27bdffd0        addiu   sp,sp,-48
  10:   afbf002c        sw      ra,44(sp)
  14:   afbe0028        sw      s8,40(sp)
  18:   afb00024        sw      s0,36(sp)
  1c:   03a0f021        move    s8,sp
  20:   afbc0010        sw      gp,16(sp)
  24:   afc40030        sw      a0,48(s8)
  28:   afc50034        sw      a1,52(s8)
  2c:   afc60038        sw      a2,56(s8)
  30:   afc7003c        sw      a3,60(s8)
  34:   8fc20044        lw      v0,68(s8)
  38:   00403821        move    a3,v0
  3c:   8fc40038        lw      a0,56(s8)
  40:   8fc50034        lw      a1,52(s8)
  44:   8fc60040        lw      a2,64(s8)
  48:   8fc80048        lw      t0,72(s8)
  4c:   8fc20030        lw      v0,48(s8)
  50:   8fc3003c        lw      v1,60(s8)
  54:   27bdffe8        addiu   sp,sp,-24
  58:   2410ffea        li      s0,-22
  5c:   10400013        beqz    v0,ac <sys_clone+0xac>
  60:   00000000        nop
  64:   10a00011        beqz    a1,ac <sys_clone+0xac>
  68:   00000000        nop
  6c:   24a5ffe0        addiu   a1,a1,-32
  70:   aca20000        sw      v0,0(a1)
  74:   aca30004        sw      v1,4(a1)
  78:   24021018        li      v0,4120
  7c:   0000000c        syscall
  80:   14e0000a        bnez    a3,ac <sys_clone+0xac>
  84:   00000000        nop
  88:   14400008        bnez    v0,ac <sys_clone+0xac>
  8c:   00000000        nop
  90:   8fb90000        lw      t9,0(sp)
  94:   8fa40004        lw      a0,4(sp)
  98:   0320f809        jalr    t9
  9c:   00000000        nop
  a0:   00402021        move    a0,v0
  a4:   24020fa1        li      v0,4001
  a8:   0000000c        syscall
  ac:   27bd0018        addiu   sp,sp,24
  b0:   02001021        move    v0,s0
  b4:   00e01821        move    v1,a3
  b8:   1060000a        beqz    v1,e4 <sys_clone+0xe4>
  bc:   00000000        nop
  c0:   afc20018        sw      v0,24(s8)
  c4:   8f820000        lw      v0,0(gp)
  c8:   0040c821        move    t9,v0
  cc:   0320f809        jalr    t9
  d0:   00000000        nop
  d4:   8fdc0010        lw      gp,16(s8)
  d8:   8fc30018        lw      v1,24(s8)
  dc:   ac430000        sw      v1,0(v0)
  e0:   2402ffff        li      v0,-1
  e4:   03c0e821        move    sp,s8
  e8:   8fbf002c        lw      ra,44(sp)
  ec:   8fbe0028        lw      s8,40(sp)
  f0:   8fb00024        lw      s0,36(sp)
  f4:   27bd0030        addiu   sp,sp,48
  f8:   03e00008        jr      ra
  fc:   00000000        nop

[Markus (顧孟勤), 2013-04-09 18:33:17]

lgtm

[petarj, 2013-04-09 23:25:22]

On 2013/04/09 18:33:17, Markus (顧孟勤) wrote:

Mark, Markus,

can any of you two submit this change?
CQ does not work for this repository.

Thanks.

Petar

[Mark Seaborn, 2013-04-09 23:32:10]

On 9 April 2013 16:25, <petarj@mips.com> wrote:

> can any of you two submit this change?
> CQ does not work for this repository.
>

I've added you as a committer, so you can commit directly using gcl or
git-cl.

Mark

[petarj, 2013-04-10 00:35:57]

Committed patchset #1 manually as r21 (presubmit successful).

[petarj, 2013-04-10 00:38:15]

On 2013/04/09 23:32:10, Mark Seaborn wrote:
> On 9 April 2013 16:25, <mailto:petarj@mips.com> wrote:
> 
> > can any of you two submit this change?
> > CQ does not work for this repository.
> >
> 
> I've added you as a committer, so you can commit directly using gcl or
> git-cl.
> 
> Mark

Done. Thanks, Mark.