Re: [PATCH] powerpc: New feature - HWCAP/HWCAP2 bits in the TCB

[Adhemerval Zanella <adhemerval dot zanella at linaro dot org>]

On 09-06-2015 15:51, Steven Munroe wrote:
> On Tue, 2015-06-09 at 11:33 -0700, Roland McGrath wrote:
>>> I believe the idea is to provide a fast way to emulate a functionality
>>> similar to __builtin_cpu_supports for powerpc.  For x86, this builtin
>>> will create 'cpuid' instruction, but since powerpc lacks a similar one
>>> it should rely on hardware capability information provided by kernel.
>>
>> On x86 using cpuid is quite slow as instruction-level overheads go.
>> It's certainly nowhere near as fast as doing a direct load from memory.
>> So this analogue does not suggest anything like justification for the
>> kind of microoptimization being discussed.
> 
> In the X86 implementation the cpuid is cached by __builtin_cpu_init(). I
> suspect the result is saved in static or TLS. 
> 
> That said the x86/x86_64 ISA and micro arch are different from POWER
> with different tradeoffs.
> 
> It would inappropriate to impose these assumptions on other platforms
> 
> Our proposal is appropriate for the reality of POWER and using the
> HWCAP.
> 

In fact the __builtin_cpu_supports generate for x86_64 a read from
a static struct defined in libgcc: 

* libgcc/config/i386/cpuinfo.c:

struct __processor_model
{
  unsigned int __cpu_vendor;
  unsigned int __cpu_type;
  unsigned int __cpu_subtype;
  unsigned int __cpu_features[1];
} __cpu_model = { };

And it is initialized in constructor (__cpu_indicator_init) using the
cpuid. Either way, for powerpc even using the same mechanism will
incur in a static GOT relocation as it is defined in a dynamic library
(with the different it won't have a dynamic relocation).