Re: Status of CONFIG_FORCED_INLINING?

From: Rob Landley
Date: Thu May 24 2007 - 13:43:19 EST


On Thursday 24 May 2007 12:29 pm, Jan Engelhardt wrote:
>
> On May 23 2007 23:22, Adrian Bunk wrote:
> >
> >And we need only two different inline levels (__always_inline and
> >"let the compiler decide"), not three (__always_inline, inline and
> >"let the compiler decide").
>
> "inline" is "let the compiler decide".

The compiler decides anyway. That's why we need a noinline to tell it _not_
to spontaneously inline things it shouldn't.

> If it is not, then it is "let the
> compiler decide, based on my bias that I think it should be inlined".

Things like unlikely() actually produce different code (and are, technically
speaking, workarounds for the high branch misprediction penalty in modern
processor designs).

Things like "register" are a hint to the compiler that it's free to ignore.
How often do we use the "register" keyword in Linux? Register allocation is
totally different on different processors, and deprived of context the hint
is almomst meaningless. If it doesn't guarantee anything and has far less of
a performance impact than the difference between -O2 and -Os (or between gcc
3.4 and 4.1), then it's probably an unnecessary complication.

Inlining is even worse because whether or not it's a performance win depends
on the L1 cache size, cache line size and alignment, L2 cache, DRAM fetch
latency, and so on. This varies all over the map within a given processor
line, let alone between processors.

If we actually need something inlined, we can tell it to do so reliably with
__always_inline (and change __always_inline to be MORE explicit each time gcc
breaks the previous way of saying "yes really inline it dammit"). If we can
live with it not being inlined, why not leave it to the compiler whether or
not to inline it?

Rob
-
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/