Re: [PATCH] powerpc/vdso32: Add missing _restgpr_31_x to fix build failure

[Segher Boessenkool <segher@kernel.crashing.org>]

On Mon, Mar 15, 2021 at 04:38:52PM +0000, David Laight wrote:
> From: Rasmus Villemoes
> > Sent: 15 March 2021 16:24
> > On 12/03/2021 03.29, Segher Boessenkool wrote:
> > > On Tue, Mar 09, 2021 at 06:19:30AM +0000, Christophe Leroy wrote:
> > >> With some defconfig including CONFIG_CC_OPTIMIZE_FOR_SIZE,
> > >> (for instance mvme5100_defconfig and ps3_defconfig), gcc 5
> > >> generates a call to _restgpr_31_x.
> > >
> > >> I don't know if there is a way to tell GCC not to emit that call, because at the end we get more
> > instructions than needed.
> > >
> > > The function is required by the ABI, you need to have it.
> > >
> > > You get *fewer* insns statically, and that is what -Os is about: reduce
> > > the size of the binaries.
> > 
> > Is there any reason to not just always build the vdso with -O2? It's one
> > page/one VMA either way, and the vdso is about making certain system
> > calls cheaper, so if unconditional -O2 could save a few cycles compared
> > to -Os, why not? (And if, as it seems, there's only one user within the
> > DSO of _restgpr_31_x, yes, the overall size of the .text segment
> > probably increases slightly).
> 
> Sometimes -Os generates such horrid code you really never want to use it.
> A classic is on x86 where it replaces 'load register with byte constant'
> with 'push byte' 'pop register'.
> The code is actually smaller but the execution time is horrid.
> 
> There are also cases where -O2 actually generates smaller code.

Yes, as with all heuristics it doesn't always work out.  But usually -Os
is smaller.

> Although you may need to disable loop unrolling (often dubious at best)
> and either force or disable some function inlining.

The cases where GCC does loop unrolling at -O2 always help quite a lot.
Or, do you have a counter-example?  We'd love to see one.

And yup, inlining is hard.  GCC's heuristics there are very good
nowadays, but any single decision has big effects.  Doing the important
spots manually (always_inline or noinline) has good payoff.


Segher