Re: [PATCH v6 5/9] vsprintf: Factor out %pV handler as va_format()

[Rasmus Villemoes]

On 12/02/2019 18.58, Joe Perches wrote:
> On Tue, 2019-02-12 at 14:00 +0100, Petr Mladek wrote:
>> On Fri 2019-02-08 09:11:17, Joe Perches wrote:
>>> On Fri, 2019-02-08 at 16:23 +0100, Petr Mladek wrote:
>>>> Move the code from the long pointer() function. We are going to improve
>>>> error handling that will make it more complicated.
>>>>
>>>> This patch does not change the existing behavior.
>>>
>>> But doesn't this increase stack use?
>>> %pV is recursive and increasing the stack is undesired
>>> for this use.
>>
>> %pV handler is stack sensitive because the entire vsnprintf()
>> machinery is called recursively. This one extra call does
>> not make it much worse.
> 
> That's an argument?.
> 
> Refactoring is good, but you need to add
> __always_inline here.
> 
> 

No. No no no. Please at least try to send the code through a compiler.
Petr's patch actually makes things uniformly better, since gcc ends up
doing a tail call from pointer().

    2e65:       4c 89 cf                mov    %r9,%rdi
    2e68:       e9 23 05 00 00          jmpq   3390 <va_format>

$ scripts/stackusage -o before.su lib/vsprintf.o
$ cp lib/vsprintf.o lib/vsprintf.o.0
$ # apply patch
$ scripts/stackusage -o after.su lib/vsprintf.o
$ cp lib/vsprintf.o lib/vsprintf.o.1
$ scripts/stackdelta before.su after.su
./lib/vsprintf.c        pointer 40      8       -32
$ grep va_format after.su
after.su:./lib/vsprintf.c:1437     va_format       40      static
# Icing
$ scripts/bloat-o-meter lib/vsprintf.o.0 lib/vsprintf.o.1
add/remove: 1/0 grow/shrink: 0/1 up/down: 77/-186 (-109)
Function                                     old     new   delta
va_format                                      -      77     +77
pointer                                      634     448    -186

Of course, all of this depends very much on compiler version,
architecture (if the calling convention says to pass arguments on stack,
tail calls are a lot harder to come by), CONFIG_STACKPROTECTOR etc.
etc., but nothing suggests that forcing the va_format logic into
pointer() would be a good thing.

Rasmus