2006-12-08 12:59:25

by Adrian Bunk

[permalink] [raw]
Subject: UML and fastcall/FASTCALL

UML on i386 is now the only case where fastcall/FASTCALL is not a noop.

There are two use cases for fastcall/FASTCALL in UML on i386:


1. optimization for C code
A faster calling convention is used for the functions annotated this way.


2. interfacing with assembler code
But include/asm-um/linkage.h contains the following:

<-- snip -->

#ifndef __ASM_UM_LINKAGE_H
#define __ASM_UM_LINKAGE_H

#include "asm/arch/linkage.h"


/* <linux/linkage.h> will pick sane defaults */
#ifdef CONFIG_GPROF
#undef FASTCALL
#undef fastcall
#endif

#endif

<-- snip -->

E.g. if CONFIG_SMP was still available on UML, CONFIG_SMP=y,
CONFIG_GPROF=y would have some horrible effects when calling the
functions in arch/i386/lib/semaphore.S.


Are there any benchmark numbers that the existing fastcall/FASTCALL
annotations in the kernel really make a measurable difference for
C code?

Otherwise, we could use it only for assembler code using this calling
convention (if there is any used by UML) - and CONFIG_GPROF mustn't
change this.


cu
Adrian

--

"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon Seed


2006-12-08 16:39:39

by Jeff Dike

[permalink] [raw]
Subject: Re: UML and fastcall/FASTCALL

On Fri, Dec 08, 2006 at 01:59:28PM +0100, Adrian Bunk wrote:
> UML on i386 is now the only case where fastcall/FASTCALL is not a noop.

If i386 doesn't use it any more, then UML shouldn't either. The only reason
there's support for it in UML is that I was copying i386.

Jeff

--
Work email - jdike at linux dot intel dot com

2006-12-08 18:04:48

by Adrian Bunk

[permalink] [raw]
Subject: Re: UML and fastcall/FASTCALL

On Fri, Dec 08, 2006 at 11:33:04AM -0500, Jeff Dike wrote:
> On Fri, Dec 08, 2006 at 01:59:28PM +0100, Adrian Bunk wrote:
> > UML on i386 is now the only case where fastcall/FASTCALL is not a noop.
>
> If i386 doesn't use it any more, then UML shouldn't either. The only reason
> there's support for it in UML is that I was copying i386.

It's a noop on i386 due to -mregparm=3 ...

> Jeff

cu
Adrian

--

"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon Seed