Re: PATCH: Add --alt-nops=short|long to x86/x86-64 assemblers

["H. J. Lu" <hjl@lucon.org>]

On Wed, Jun 14, 2006 at 04:14:23PM +0100, Nick Clifton wrote:
> Hi H. J.
> 
> >>I am not familiar with the x86 instruction set, so please can you 
> >>explain why it is necessary to have these two different versions of the 
> >>nop and why the user has to select one ?  ie if the space to be padded 
> >>is up to 10 bytes why can't the assembler just use the "short" version 
> >>automatically and if it is longer than 10 bytes use the "long" version ?
> >
> >x86/x86-64 assemblers need to fill the text section from 1 to
> >15 bytes for alignment. We have simple nop instructions for 1 to
> >10 bytes. For 11 to 15 bytes, we can add 0x66 prefix repeatedly
> >to 10 byte nop to get 11-15byte nops. However some processors
> >prefer simple nops. That is, 3 0x66 prefixes on the 10 byte nop
> >are slower on those processors than a 6 byte nop + a 7 byte nop.
> 
> Does the assembler not know the target processor variant ?  ie is the 
> assembler unable to make an informed choice of which type of nop to use ?

x86/x86-64 assemblers don't support -mtune/-march/-mcpu.

> 
> >User can use --alt-nops=short to avoid repeated 0x66 prefixes.
> >Maybe we can use something other than short|long. But I don't
> >have a better name.
> 
> How about "--allow-long-nop-sequences" and its inverse 
> "--no-allow-long-nop-sequences" ?  The default would presumably be to 
> allow the long sequences since this is faster on more modern processors, 
> yes ?

There are 2 problems. The new nop instructions are available only
on Pentium Pro or above. Without -mtune/-march/-mcpu, we don't
know if we can use them, short or long. We need a switch to tell
assembler if the new nop instructions can be used at all. Secondly,
not all modern processors prefer "long" versions. We need another
switch to tell which kind of the new nop instructions should be used,
short or long.

Maybe I should add -march= and -mtune= to assembler.



H.J.