Re: [PATCH, i386]: Emit cld instruction when stringops are used

[Uros Bizjak <ubizjak at gmail dot com>]

Chris Lattner wrote:

> > This patch adds back generation of cld instructions when stringops are
> > used in the function. cld insn is emitted as unspec volatile at the
> > end of function prologue. Currently, patched gcc emits cld when
> > "-mcld" is added to compile flags.
> >
> > Currently, there is no configure-time support, I think we should first
> > reach a consensus on how to configure this. We can
>
> If I understand the issue correctly, this patch is insufficient.  You 
> should also have GCC emit a cld instruction before any inline asm.  Do 
> you think it's a good idea to add a 52 cycle delay (on the P4) to any 
> function that uses inline asm?

No, this patch in effect restores previous functionality for what is 
considered buggy code (in the sense that doesn't follow psABI w.r.t. to 
direction flag). This patch should be treated as a band-aid for corner 
cases, nothing less and nothing more. In previous versions, inline asm 
was not handled at all and all [asm] code was written with this fact in 
mind.

Due to this, IMO there is no point to implement cld inserting via LCM 
(this would require major work, since we will have to restore REG_DIR 
fake register in order to make _all_ stringop patterns consumers of this 
register). There is also no point to trigger cld insertion with inline 
asm, as this will fix a non-bug.

IMO, -mcld should be off by default. This flag should provide a way to 
still compile the code that is now _known_ to be buggy. But it is up to 
distributions to decide if they want to penalize applications with 
unnecessary cld instructions. I expect that the need for -mcld will 
phase out relatively quickly.


Uros.