Re: [PATCH] S/390: Fix two issues with the IFUNC optimized mem* routines

[Andreas Krebbel <krebbel at linux dot vnet dot ibm dot com>]

On 30/08/12 18:32, Richard Henderson wrote:
> On 08/30/2012 08:13 AM, Andreas Krebbel wrote:
>>> There's no reason to *ever* drop a hidden function pointer
>>> into the constant pool, since LARL is always available.
>>
>> The problem was that I unfortunately missed to mark the references
>> hidden in a way visible to GCC :(
> 
> Ah, well, that'll do it every time.  ;-)
> 
> All that said, it's probably reasonable have the linker transform
> 
> 	lg   r1,x@got(b2)
> or
> 	lgrl r1,x@got
> into
> 	larl r1,x
> 
> when it finds that X has appropriate alignment and has been made
> hidden via linker maps or other mechanisms not visible to the compiler.

Yes. I'll work on this. Unfortunately it will not cover the currently most common case -march <z10
and -fPIC. In that case we do:

	larl	%r1,foo@GOTENT
	lg	%r1,0(%r1)

But starting with z10 that can easily be optimized as you say. While looking at it I noticed that we
currently don't use load relative for GOT slots. Fixed with this patch:
http://gcc.gnu.org/ml/gcc-patches/2012-09/msg00088.html

> The use of lgrl would be an enhancement to gcc for use on z10+ to
> completely eliminate the got register.  Freeing up another register
> definitely seems like a useful optimization.

I tried this when implementing the z10 support. Unfortunately in most cases using a GOT pointer is
faster since all the instructions support the <disp>(r12) address format and therefore can directly
deal with the GOT slot. Not having a GOT pointer requires more additional load relatives to be put
into the code.

Bye,

-Andreas-