[patch] i386: switch_to(): misplaced parentheses

[patch] i386: switch_to(): misplaced parentheses

[Chuck Ebbert]

Recent changes in i386 __switch_to() have a misplaced closing
parenthesis causing an unlikely() to terminate early.

Signed-off-by: Chuck Ebbert <76306.1226@compuserve.com>

--- 2.6.18-rc2-32.orig/arch/i386/kernel/process.c
+++ 2.6.18-rc2-32/arch/i386/kernel/process.c
@@ -690,8 +690,8 @@ struct task_struct fastcall * __switch_t
 	/*
 	 * Now maybe handle debug registers and/or IO bitmaps
 	 */
-	if (unlikely((task_thread_info(next_p)->flags & _TIF_WORK_CTXSW))
-	    || test_tsk_thread_flag(prev_p, TIF_IO_BITMAP))
+	if (unlikely((task_thread_info(next_p)->flags & _TIF_WORK_CTXSW)
+	    || test_tsk_thread_flag(prev_p, TIF_IO_BITMAP)))
 		__switch_to_xtra(next_p, tss);
 
 	disable_tsc(prev_p, next_p);
-- 
Chuck

Re: [patch] i386: switch_to(): misplaced parentheses

[Steven Rostedt]

On Tue, 2006-07-25 at 16:15 -0400, Chuck Ebbert wrote:
> Recent changes in i386 __switch_to() have a misplaced closing
> parenthesis causing an unlikely() to terminate early.
> 
> Signed-off-by: Chuck Ebbert <76306.1226@compuserve.com>
> 
> --- 2.6.18-rc2-32.orig/arch/i386/kernel/process.c
> +++ 2.6.18-rc2-32/arch/i386/kernel/process.c
> @@ -690,8 +690,8 @@ struct task_struct fastcall * __switch_t
>  	/*
>  	 * Now maybe handle debug registers and/or IO bitmaps
>  	 */
> -	if (unlikely((task_thread_info(next_p)->flags & _TIF_WORK_CTXSW))
> -	    || test_tsk_thread_flag(prev_p, TIF_IO_BITMAP))
> +	if (unlikely((task_thread_info(next_p)->flags & _TIF_WORK_CTXSW)
> +	    || test_tsk_thread_flag(prev_p, TIF_IO_BITMAP)))
>  		__switch_to_xtra(next_p, tss);
>  
>  	disable_tsc(prev_p, next_p);

Unlikely's with or's are kind of ambiguous.  An 'and' makes sense but
or's don't.  Because a branch is going to happen anyway.  Just to test
this out, I made a little function and tried out different types of
parenthesis placements.

Here: (the original placement)

	#define unlikely(x) __builtin_expect(!!(x), 0)

	int switch_to(int x, int y)
	{
		int a = 0;
		if (unlikely(x==24) || (y==83))
			a=10;
		return a;
	}


and now yours:

	#define unlikely(x) __builtin_expect(!!(x), 0)

	int switch_to(int x, int y)
	{
		int a = 0;
		if (unlikely(x==24 || y==83))
			a=10;
		return a;
	}

The original gave me this: gcc -O2 -c switch.c

objdump -Dr switch.o

00000000 <switch_to>:
   0:   55                      push   %ebp
   1:   89 e5                   mov    %esp,%ebp
   3:   83 7d 08 18             cmpl   $0x18,0x8(%ebp)
   7:   74 0a                   je     13 <switch_to+0x13>
   9:   31 c0                   xor    %eax,%eax
   b:   83 7d 0c 53             cmpl   $0x53,0xc(%ebp)
   f:   74 02                   je     13 <switch_to+0x13>
  11:   5d                      pop    %ebp
  12:   c3                      ret
  13:   5d                      pop    %ebp
  14:   b8 0a 00 00 00          mov    $0xa,%eax
  19:   c3                      ret


and then yours:  gcc -O2 -c switch_alt.c

objdump -Dr switch_alt.o

00000000 <switch_to>:
   0:   55                      push   %ebp
   1:   89 e5                   mov    %esp,%ebp
   3:   83 7d 08 18             cmpl   $0x18,0x8(%ebp)
   7:   74 0a                   je     13 <switch_to+0x13>
   9:   31 c0                   xor    %eax,%eax
   b:   83 7d 0c 53             cmpl   $0x53,0xc(%ebp)
   f:   74 02                   je     13 <switch_to+0x13>
  11:   5d                      pop    %ebp
  12:   c3                      ret
  13:   5d                      pop    %ebp
  14:   b8 0a 00 00 00          mov    $0xa,%eax
  19:   c3                      ret


They are identical!!!

This is not surprising since you still need to test the other OR case if
it fails, which you are saying it will.  So saying it is unlikely is
meaningless.  Likely's are good for 'or' and unlikely's are good for
'and'.  But not the other way around.

|| and && short circuit when they can.

So the statement of A || B is:

TEST A
JUMP_IF_TRUE true
TEST B
JUMP_IF_FALSE false
true:
 do true statement
JUMP out:
false:
 do else statement
out:

Saying TEST A is likely to fail really doesn't help the situation.
Saying it will likely pass will. Then we could do this:

TEST A
JUMP_IF_FALSE test2
true:
 do true statement
JUMP out:
false:
  do false statement
jump out:
test2:
TEST B
JUMP_IF_TRUE true
JUMP false:
out:

Where here we can move the true after the first compare and prevent the
conditional branch.

So what I'm saying is that unlikely(a || b) has really no good meaning.

Out of curiosity, I removed the unlikely altogether and here's what I
got.

gcc -O2 -c switch_none.c

00000000 <switch_to>:
   0:   55                      push   %ebp
   1:   89 e5                   mov    %esp,%ebp
   3:   83 7d 08 18             cmpl   $0x18,0x8(%ebp)
   7:   74 0a                   je     13 <switch_to+0x13>
   9:   31 c0                   xor    %eax,%eax
   b:   83 7d 0c 53             cmpl   $0x53,0xc(%ebp)
   f:   74 02                   je     13 <switch_to+0x13>
  11:   5d                      pop    %ebp
  12:   c3                      ret
  13:   5d                      pop    %ebp
  14:   b8 0a 00 00 00          mov    $0xa,%eax
  19:   c3                      ret


Ha! still the same.

Is that unlikely there really do any good?

-- Steve

Re: [patch] i386: switch_to(): misplaced parentheses

[Steven Rostedt]

On Fri, 2006-07-28 at 21:39 -0400, Steven Rostedt wrote:

> 
> Unlikely's with or's are kind of ambiguous.  An 'and' makes sense but
> or's don't.  Because a branch is going to happen anyway.  Just to test
> this out, I made a little function and tried out different types of
> parenthesis placements.

OK, I take this back. Thinking that it might make a difference to the
second compare, I added a little more to my test program:

---
#define unlikely(x) __builtin_expect(!!(x), 0)

int do_func(void);
int a;

int switch_to(int x, int y)
{
        a = 2;
        if (unlikely(x==24 || y==83))
                a=10;
        do_func();
        return a;
}
---


This way I now have a global 'a' and a function that just might use that
'a'.  So this does make a difference:  Kind of funky though. It wastes
space to make it avoid branching when we don't have your total unlikely
(see below):

The above gave:

00000000 <switch_to>:
   0:   55                      push   %ebp
   1:   ba 02 00 00 00          mov    $0x2,%edx
   6:   89 e5                   mov    %esp,%ebp
   8:   83 ec 08                sub    $0x8,%esp
   b:   83 7d 08 18             cmpl   $0x18,0x8(%ebp)
   f:   89 15 00 00 00 00       mov    %edx,0x0
                        11: R_386_32    a
  15:   74 12                   je     29 <switch_to+0x29>
  17:   83 7d 0c 53             cmpl   $0x53,0xc(%ebp)
  1b:   74 0c                   je     29 <switch_to+0x29>
  1d:   e8 fc ff ff ff          call   1e <switch_to+0x1e>
                        1e: R_386_PC32  do_func
  22:   a1 00 00 00 00          mov    0x0,%eax
                        23: R_386_32    a
  27:   c9                      leave
  28:   c3                      ret
  29:   b8 0a 00 00 00          mov    $0xa,%eax
  2e:   a3 00 00 00 00          mov    %eax,0x0
                        2f: R_386_32    a
  33:   eb e8                   jmp    1d <switch_to+0x1d>


Which looks the best, so your patch may be good after all :-)

The other tests looked pretty much the same:

  Between (unlikely(x==24) || (y==83)) and
      ((x==24) || unlikely(y==83)) and
      (x==24 || y==83)

which was this: (done with the unlikely(y==83))

00000000 <switch_to>:
   0:   55                      push   %ebp
   1:   ba 02 00 00 00          mov    $0x2,%edx
   6:   89 e5                   mov    %esp,%ebp
   8:   83 ec 08                sub    $0x8,%esp
   b:   83 7d 08 18             cmpl   $0x18,0x8(%ebp)
   f:   89 15 00 00 00 00       mov    %edx,0x0
                        11: R_386_32    a
  15:   74 19                   je     30 <switch_to+0x30>
  17:   83 7d 0c 53             cmpl   $0x53,0xc(%ebp)
  1b:   74 13                   je     30 <switch_to+0x30>
  1d:   e8 fc ff ff ff          call   1e <switch_to+0x1e>
                        1e: R_386_PC32  do_func
  22:   a1 00 00 00 00          mov    0x0,%eax
                        23: R_386_32    a
  27:   c9                      leave
  28:   c3                      ret
  29:   8d b4 26 00 00 00 00    lea    0x0(%esi),%esi
  30:   b8 0a 00 00 00          mov    $0xa,%eax
  35:   a3 00 00 00 00          mov    %eax,0x0
                        36: R_386_32    a
  3a:   e8 fc ff ff ff          call   3b <switch_to+0x3b>
                        3b: R_386_PC32  do_func
  3f:   a1 00 00 00 00          mov    0x0,%eax
                        40: R_386_32    a
  44:   c9                      leave
  45:   c3                      ret


So the big savings isn't the branching, but the entire branch is
expected to fail, so it doesn't bother with the duplicate code to speed
things up (see the two do_func calls).  So your patch really just saves
space, and not really speed (but you can argue that this space savings
increases speed by not wasting cache).

So I do give credence to your patch.

Acked-by: Steven Rostedt <rostedt@goodmis.org>

-- Steve