Incorrect code from EGCS-1.1 on PowerPC

Incorrect code from EGCS-1.1 on PowerPC

[Gary Thomas]

[-- Attachment #1: Type: text/plain, Size: 721 bytes --]

The attached code snippet (from the Linux kernel) generates incorrect
code when compiled on Linux/PPC with "-O2" options.

The problem stems from inlined functions and a pointer (skb) being
dereferenced (skb->sb) before it is tested for NULL as a result of 
the loop being reorganized.

------------------------------------------------------------------------
Gary Thomas                              |
email: gdt@linuxppc.org                  | "Fine wine is a necessity of
   ... opinions expressed here are mine  |        life for me"
       and no one else would claim them! |
                                         |      Thomas Jefferson
------------------------------------------------------------------------



[-- Attachment #2: bad_ppc.c --]
[-- Type: text/x-c, Size: 1204 bytes --]

struct sk_buff {
  struct sk_buff	* next;			 
  struct sk_buff	* prev;			 
  struct sk_buff_head * list;		 
  struct sock		* sk;
};

struct sk_buff_head {
  struct sk_buff	* next;
  struct sk_buff	* prev;
  int			qlen;		 

};

struct sock {
  struct sock 	* receive_queue;
};

extern struct sk_buff *		skb_dequeue(struct sk_buff_head *list);

 
extern __inline__ struct sk_buff *__skb_dequeue(struct sk_buff_head *list)
{
	struct sk_buff *next, *prev, *result;
	prev = (struct sk_buff *) list;
	next = prev->next;
	result = ((void *) 0) ;
	if (next != prev) {
		result = next;
		next = next->next;
		list->qlen--;
		next->prev = prev;
		prev->next = next;
		result->next = ((void *) 0) ;
		result->prev = ((void *) 0) ;
		result->list = ((void *) 0) ;
	}
	return result;
}

extern __inline__ struct sk_buff *skb_dequeue(struct sk_buff_head *list)
{
	long flags;
	struct sk_buff *result;
	__save_flags(&( flags )) ;
	cli();
	result = __skb_dequeue(list);
	__restore_flags( flags ) ;
	return result;
}
 
static void tcp_close_pending (struct sock *sk)
{
	struct sk_buff *skb;
	while ((skb = skb_dequeue(&sk->receive_queue)) != ((void *) 0) )
	{
		tcp_close(skb->sk, 0);
		kfree_skb(skb, 1 );
	}
	return;
}

[-- Attachment #3: bad_ppc.s --]
[-- Type: text/x-asm, Size: 867 bytes --]

	.file	"xx1.c"
gcc2_compiled.:
	.section	".text"
	.align 2
	.type	 tcp_close_pending,@function
tcp_close_pending:
	mr 12,1
	la 1,-48(1)
	stw 12,0(1)
	mflr 0
	stw 29,36(1)
	stw 30,40(1)
	stw 31,44(1)
	stw 0,52(1)
	mr 30,3
	addi 29,1,8
	b .L7
.L9:
	crxor 6,6,6
	bl tcp_close
	li 4,1
	mr 3,31
	crxor 6,6,6
	bl kfree_skb
.L7:
	mr 3,29
	crxor 6,6,6
	bl __save_flags
	li 31,0
	bl cli
	lwz 11,0(30)
	lwz 3,8(1)
	cmpw 0,11,30
	li 0,0
	bc 12,2,.L10
	lwz 9,8(30)
	mr 31,11
	lwz 11,0(31)
	addi 9,9,-1
	stw 9,8(30)
	stw 30,4(11)
	stw 11,0(30)
	stw 0,8(31)
	stw 0,0(31)
	stw 0,4(31)
.L10:
	crxor 6,6,6
	bl __restore_flags
	cmpwi 0,31,0
	li 4,0
	lwz 3,12(31)
	bc 4,2,.L9
	lwz 0,52(1)
	mtlr 0
	lwz 29,36(1)
	lwz 30,40(1)
	lwz 31,44(1)
	la 1,48(1)
	blr
.Lfe1:
	.size	 tcp_close_pending,.Lfe1-tcp_close_pending
	.ident	"GCC: (GNU) egcs-2.91.53 19980803 (gcc2 ss-980609 experimental)"

Re: Incorrect code from EGCS-1.1 on PowerPC

[Franz Sirl]

Am Tue, 18 Aug 1998 schrieb Gary Thomas:
>>The attached code snippet (from the Linux kernel) generates incorrect
>code when compiled on Linux/PPC with "-O2" options.
>
>The problem stems from inlined functions and a pointer (skb) being
>dereferenced (skb->sb) before it is tested for NULL as a result of 
>the loop being reorganized.

I verified the bug with a cvs checkout 1845UTC 19980818. -fno-schedule-insns
makes the bug go away.

Franz.

Re: Incorrect code from EGCS-1.1 on PowerPC

[Gary Thomas]

On 18-Aug-98 Franz Sirl wrote:
> Am Tue, 18 Aug 1998 schrieb Gary Thomas:
>>>The attached code snippet (from the Linux kernel) generates incorrect
>>code when compiled on Linux/PPC with "-O2" options.
>>
>>The problem stems from inlined functions and a pointer (skb) being
>>dereferenced (skb->sb) before it is tested for NULL as a result of 
>>the loop being reorganized.
> 
> I verified the bug with a cvs checkout 1845UTC 19980818. -fno-schedule-insns
> makes the bug go away.
> 

You reported on the 'egcs' list that the current snapshot is "OK"
for building the latest 2.1.XX kernel.  Don't you see this code generation
error when you build it?  Or is it that pattern sensitive that changes
between 2.0.32 and 2.1.115 keep it from happening?

------------------------------------------------------------------------
Gary Thomas                              |
email: gdt@linuxppc.org                  | "Fine wine is a necessity of
   ... opinions expressed here are mine  |        life for me"
       and no one else would claim them! |
                                         |      Thomas Jefferson
------------------------------------------------------------------------

Re: Incorrect code from EGCS-1.1 on PowerPC

[Franz Sirl]

At 09:17 19.08.98 , Gary Thomas wrote:
>
>On 18-Aug-98 Franz Sirl wrote:
>> Am Tue, 18 Aug 1998 schrieb Gary Thomas:
>>>>The attached code snippet (from the Linux kernel) generates incorrect
>>>code when compiled on Linux/PPC with "-O2" options.
>>>
>>>The problem stems from inlined functions and a pointer (skb) being
>>>dereferenced (skb->sb) before it is tested for NULL as a result of 
>>>the loop being reorganized.
>> 
>> I verified the bug with a cvs checkout 1845UTC 19980818.
-fno-schedule-insns
>> makes the bug go away.
>> 
>
>You reported on the 'egcs' list that the current snapshot is "OK"
>for building the latest 2.1.XX kernel.  Don't you see this code generation
>error when you build it?  Or is it that pattern sensitive that changes
>between 2.0.32 and 2.1.115 keep it from happening?

Hmm, I said "no known bugs due to egcs" and that is still true, cause I
haven't seen any bug reports yet with egcs-1.1 and linux-2.1.115. And I
(and others) use 2.1.115vger since its out.
Probably you are right and something has changed that hides this bug for
2.1.115 (I think networking code was changed heavily in 2.1).

Franz.

Re: Incorrect code from EGCS-1.1 on PowerPC

[Franz Sirl]

Am Wed, 19 Aug 1998 schrieb Gary Thomas:
>On 18-Aug-98 Franz Sirl wrote:
>> Am Tue, 18 Aug 1998 schrieb Gary Thomas:
>>>>The attached code snippet (from the Linux kernel) generates incorrect
>>>code when compiled on Linux/PPC with "-O2" options.
>>>
>>>The problem stems from inlined functions and a pointer (skb) being
>>>dereferenced (skb->sb) before it is tested for NULL as a result of 
>>>the loop being reorganized.
>> 
>> I verified the bug with a cvs checkout 1845UTC 19980818. -fno-schedule-insns
>> makes the bug go away.
>> 
>
>You reported on the 'egcs' list that the current snapshot is "OK"
>for building the latest 2.1.XX kernel.  Don't you see this code generation
>error when you build it?  Or is it that pattern sensitive that changes
>between 2.0.32 and 2.1.115 keep it from happening?

Gary,

the latest CVS-egcs seems to fix this problem here. I assume it was the change
to haifa-sched.c.
Can you verify this?

Franz.

Re: Incorrect code from EGCS-1.1 on PowerPC

[Jeffrey A Law]

  In message < XFMail.980818151402.gdt@linuxppc.org >you write:
  > This message is in MIME format
  > --_=XFMail.1.3.p0.Linux:980818151357:722=_
  > Content-Type: text/plain; charset=us-ascii
  > 
  > The attached code snippet (from the Linux kernel) generates incorrect
  > code when compiled on Linux/PPC with "-O2" options.
  > 
  > The problem stems from inlined functions and a pointer (skb) being
  > dereferenced (skb->sb) before it is tested for NULL as a result of 
  > the loop being reorganized.
I just looked at the debugging dumps for this one.  It was definitely
the haifa-sched.c change that fixed this problem.  Amazing.  I wonder
if this is the same problem that Michal has been having (he's seeing
problems in tcp_close_pending, but on the Alpha).

If I had known this particular bug in haifa was causing these problems
I would have encouraged Catherine to nail it down quicker.  The good
news is we do have a testcase for this bug in the testsuite (980506-1.c)

jeff

Re: Incorrect code from EGCS-1.1 on PowerPC

[Gary Thomas]

On 20-Aug-98 Jeffrey A Law wrote:
> 
>   In message < XFMail.980818151402.gdt@linuxppc.org >you write:
>   > This message is in MIME format
>   > --_=XFMail.1.3.p0.Linux:980818151357:722=_
>   > Content-Type: text/plain; charset=us-ascii
>   > 
>   > The attached code snippet (from the Linux kernel) generates incorrect
>   > code when compiled on Linux/PPC with "-O2" options.
>   > 
>   > The problem stems from inlined functions and a pointer (skb) being
>   > dereferenced (skb->sb) before it is tested for NULL as a result of 
>   > the loop being reorganized.
> I just looked at the debugging dumps for this one.  It was definitely
> the haifa-sched.c change that fixed this problem.  Amazing.  I wonder
> if this is the same problem that Michal has been having (he's seeing
> problems in tcp_close_pending, but on the Alpha).
> 
> If I had known this particular bug in haifa was causing these problems
> I would have encouraged Catherine to nail it down quicker.  The good
> news is we do have a testcase for this bug in the testsuite (980506-1.c)
> 

This is good news.  I'll be testing with the latest CVS sources today
and let you know if the problem is truly fixed.

------------------------------------------------------------------------
Gary Thomas                              |
email: gdt@linuxppc.org                  | "Fine wine is a necessity of
   ... opinions expressed here are mine  |        life for me"
       and no one else would claim them! |
                                         |      Thomas Jefferson
------------------------------------------------------------------------

Re: Incorrect code from EGCS-1.1 on PowerPC

[Gary Thomas]

On 20-Aug-98 Jeffrey A Law wrote:
> 
>   In message < XFMail.980818151402.gdt@linuxppc.org >you write:
>   > This message is in MIME format
>   > --_=XFMail.1.3.p0.Linux:980818151357:722=_
>   > Content-Type: text/plain; charset=us-ascii
>   > 
>   > The attached code snippet (from the Linux kernel) generates incorrect
>   > code when compiled on Linux/PPC with "-O2" options.
>   > 
>   > The problem stems from inlined functions and a pointer (skb) being
>   > dereferenced (skb->sb) before it is tested for NULL as a result of 
>   > the loop being reorganized.
> I just looked at the debugging dumps for this one.  It was definitely
> the haifa-sched.c change that fixed this problem.  Amazing.  I wonder
> if this is the same problem that Michal has been having (he's seeing
> problems in tcp_close_pending, but on the Alpha).
> 
> If I had known this particular bug in haifa was causing these problems
> I would have encouraged Catherine to nail it down quicker.  The good
> news is we do have a testcase for this bug in the testsuite (980506-1.c)
> 

I built from CVS yesterday (20.08.98) and used this compiler to build
the Linux kernel 2.0.32 which seems to work fine.  The problem I reorted
is now fixed.

------------------------------------------------------------------------
Gary Thomas                              |
email: gdt@linuxppc.org                  | "Fine wine is a necessity of
   ... opinions expressed here are mine  |        life for me"
       and no one else would claim them! |
                                         |      Thomas Jefferson
------------------------------------------------------------------------