Re: [PATCH] IPC driver for Intel Mobile Internet Device (MID) platforms

[Andrew Morton <akpm@linux-foundation.org>]

On Thu, 22 Apr 2010 14:16:11 +0100 Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:

> > > +static inline int busy_loop(void) /* Wait till scu status is busy */
> > > +{
> > > +	u32 status = 0;
> > > +	u32 loop_count = 0;
> > > +
> > > +	status = __raw_readl(IPC_STATUS_REG);
> > > +	while (status & 1) {
> > > +		udelay(1); /* scu processing time is in few u secods */
> > > +		status = __raw_readl(IPC_STATUS_REG);
> > > +		loop_count++;
> > > +		/* break if scu doesn't reset busy bit after huge retry */
> > > +		if (loop_count > 100000)
> > > +			return -ETIMEDOUT;
> 
> > This function has seven-odd callsites and is waaaaaaaay to fat and slow
> > to be inlined.
> 
> Looking at the asm I'm not convinced.

I tried two version of gcc and both of them just refuse to inline
the function anyway.

If the function is changed to __always_inline, gcc will then inline it.

                  text    data     bss     dec     hex filename
inline            6182     960    1680    8822    2276 arch/x86/kernel/intel_scu_ipc.o
__always_inline   6582     928    1760    9270    2436 arch/x86/kernel/intel_scu_ipc.o

I agree with gcc ;)