Re: Do not spill variables/registers on the stack

[David Brown <david at westcontrol dot com>]

On 04/02/2011 13:19, Drasko DRASKOVIC wrote:
> On Fri, Feb 4, 2011 at 12:29 PM, David Brown<david@westcontrol.com>
> wrote:
> > On 03/02/2011 23:36, Drasko DRASKOVIC wrote:
> > > Back to the topic, few things crossed my mind :
> > >
> > > 1) Register keyword is a "hint" to the compiler, and thus not
> > > obligatory. We saw that for -O1 and above it is "useless hint",
> > > as it is always ignored. Is it however always respected by GCC
> > > when -O0 ? If yes - why is it a hint ? If no, in which cases it
> > > is not respected ? Only when you assign more vars than you have
> > > reg, or GCC might ignore it some other times even with -O0 ?
> > >
> > > 2) Is this behavior documented anywhere ? Can we find docs that
> > > note : a) that "regiser" keyword is not ignored only with -O0 b)
> > > when -O0 in which cases it is taken into account and in which
> > > cases not.
> >
> > <http://gcc.gnu.org/onlinedocs/gcc/Hints-implementation.html>
> >
> >
> > To quote:
> >
> > 4.8 Hints
> >
> > * The extent to which suggestions made by using the register
> > storage-class specifier are effective (C90 6.5.1, C99 6.7.1).
> >
> > The register specifier affects code generation only in these ways:
> > o When used as part of the register variable extension, see
> > Explicit Reg Vars. o When -O0 is in use, the compiler allocates
> > distinct stack memory for all variables that do not have the
> > register storage-class specifier; if register is specified, the
> > variable may have a shorter lifespan than the code would indicate
> > and may never be placed in memory. o On some rare x86 targets,
> > setjmp doesn't save the registers in all circumstances. In those
> > cases, GCC doesn't allocate any variables in registers unless they
> > are marked register.
>
> Hi David, thanks for the llink.
>
> From what I see, for -O0 "register" does not look like a "hint" but
> an obligation to the compiler (i.e. by this document it is promised
> that variable will finish into the register).

No, it's a hint - but one that gcc will take seriously and do its best 
to honour (at -O0).  The compiler cannot guarantee that it will be able 
to put the data into a register - there are many things that could keep 
it out of the register, including register pressure and the way the data 
is used.  The text says that with -O0 gcc will put all non-register 
variables on the stack - that does /not/ imply that it will put all 
register variables in registers.

> Also, no mention what happens for -O1 and above - there is no
> promise that variable will finish into the reg, but also no mention
> that it will be ignored. Maybe in this case it is really a hint ?

No, it is ignored for -O1 and above, except in odd combinations of 
particular x86 targets when using setjmp - as detailed on that page. 
The page says "The register specifier affects code generation /only/ in 
these ways..." - if these don't apply, the hint is ignored.

> What about global register variables ? Is it the same, ignored for
> -O1 and above ?

Variables allocated to a specific named register are a different matter 
entirely - that is a gcc extension that is related to its inline 
assembler, and is not related to the "register" storage specifier.  It's 
an example of C's bad habit of overloading keywords - albeit a fairly 
natural one.

> I am asking this because I know that for example U-Boot uses
> register variable to hold some global information (pointer to global
> structure), as you can see at the bottom of this file :
> http://git.denx.de/?p=u-boot.git;a=blob;f=arch/arm/include/asm/global_data.h;h=2a84d27a4e15e10a86cc5842eae25e80d21981e1;hb=42d44f631c4e8e5359775bdc098f2fffde4e5c05
>
>  Then the code takes care not to clobber this reg in the .S files,
> and in .c files it counts on GCC to to always allocate this reg to
> this global variable, and keep it protected from clobbering during C
> compilation. And U-Boot is surely compiled with highest possible
> optimizations. But not only that it works, it is specified in the
> official U-Boot documentation that it should always be done like
> this.

That's correct - you can put data into specified dedicated named 
registers like this, if you are /very/ careful.  But it is unrelated to 
the register storage specifier, and unaffected by any optimisation levels.