Re: [PATCH 5/5] lib: Add error string support to printks

[Daniel Santos]

On 09/19/2013 08:07 PM, Joe Perches wrote:
> On Wed, 2013-09-18 at 20:27 -0500, Daniel Santos wrote:
> > if I use ERR_PTR() on a signed int on a x86_64 where pointer
> > is 64 bits and int is 32, wouldn't that mean a signed conversion
> > instruction where the sign bit has to be moved from bit 31 to 63?
> No.  It's cast to long
>
> static inline void * __must_check ERR_PTR(long error)
> {
> 	return (void *) error;
> }

Yes, but it is that cast from int to long that costs us a signed extend 
instruction on platforms where sizeof(int) != sizeof(long).  This 
example should demonstrate the issue:

        extern void funca(void *ptr);

        static inline void * ERR_PTR(long error)
        {
                return (void *) error;
        }

        void funcb(int i)
        {
                funca(ERR_PTR(i));
        }

        void funcc(long l)
        {
                funca(ERR_PTR(l));
        }

And here is the generated code on x86_64 with -O2:

        0000000000000000 <funcb>:
                return (void *) error;
        }

        void funcb(int i)
        {
                funca(ERR_PTR(i));
           0:   48 63 ff                movslq %edi,%rdi
           3:   e9 00 00 00 00          jmpq   8 <funcb+0x8>
                                4: R_X86_64_PC32 funca-0x4
           8:   0f 1f 84 00 00 00 00    nopl 0x0(%rax,%rax,1)
           f:   00

        0000000000000010 <funcc>:
        }

        void funcc(long l)
        {
                funca(ERR_PTR(l));
          10:   e9 00 00 00 00          jmpq   15 <funcc+0x5>
                                11: R_X86_64_PC32 funca-0x4


So on x86_64 this movslq is 3 bytes of text, plus register pollution for 
each time you convert an int to a pointer. I don't know the precise 
number of cases where error numbers are passed to printk as ints, but I 
presume it is enough that this could add several kilobytes of text.  
Either way, for a popular function like vsnprintf, it's better to take a 
moderate bloat in the function than a little bloat at many call sites, 
especially when it's not performance critical.

> > Either way, %pE does seem to make a lot of sense for conditions where we
> > already have a pointer that we would otherwise use PTR_ERR() to convert,
> > but it just seems klunky to use it on an int, to have it treated like a
> > pointer and then re-interpreted as an int.  Maybe we can add %pE as well
> > as %dE and leave [ioxXu] out of it?
> I think having just one way to format is better.
Yeah, I do agree, I just don't see how to do it without introducing 
unnecessary bloat.

Daniel
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/