On Tue, Jan 23, 2018 at 02:53:33PM +0000, David Woodhouse wrote:
> We were doing a fresh CPUID for every single bit in every single output
> register. Do it once and then harvest *all* the bits we want.
>
> We were also doing the max_level check with a new CPUID invocation for
> every single bit. Stop that too.
>
> Signed-off-by: David Woodhouse <d...@amazon.co.uk>
> ---
> Spotted this in my travels; it offended me.
Ok, I see that it's itching so let's scratch it properly :-)
If we're going to optimize scattered.c, let's do something like this:
* do CPUID for each function once.
* for each set bit in there, set feature flag
which means we'd have to change the data structure.
struct cpuid_leaf {
u32 level;
u32 sub_leaf;
struct cpuid_bit bits[];
};
and that last thing is:
struct cpuid_bit {
u16 feature;
u8 reg;
u8 bit;
};
So that you have something like (for example with leaf 0x10):
struct cpuid_leaf leafs[] = {
...
{
.level = 0x00000010,
.sub_leaf = 0,
.bits = {
{ X86_FEATURE_CAT_L3, CPUID_EBX, 1 },
{ X86_FEATURE_CAT_L2, CPUID_EBX, 2 },
{ X86_FEATURE_MBA , CPUID_EBX, 3 },
{ 0 }
}
}
...
}
This way you get the CPUID only once and then iterate over bits[] and
you can do the cleaner max level computation
cpuid_eax(level & 0xffff0000);
without having to do the extended level checks.
Anyway, something like that.
It's probably not even worth doing anything though - I doubt the speedup
is visible at all.
But I certainly understand the intent to fix an annoying thing like that. :-))
Thx.
--
Regards/Gruss,
Boris.
Good mailing practices for 400: avoid top-posting and trim the reply.
