[Bug tree-optimization/53726] [4.8 Regression] aes test performance drop for eembc_2_0_peak_32

[hubicka at ucw dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53726

--- Comment #22 from Jan Hubicka <hubicka at ucw dot cz> 2012-06-22 22:45:35 UTC ---
> Yes.  The question is what is "very small" and how can we possibly

As what is very small is defined in the i386.c in the cost tables.
I simply run a small benchmark testing library&GCC implementations to
fill it in.  With new glibcs these tables may need upating.  I updated them
on some to make glibc in SUSE 11.x.

PR  43052 is about memcmp. Memcpy/memset should behave more or less sanely.
(that also reminds me that I should look again at the SSE memcpy/memset
implementation for 4.8)

> detect "very small".  For this testcase we can derive an upper bound
> of the size, which is 8, but the size is not constant.  I think unless
> we know we can expand the variable-size memcpy with, say, three
> CPU instructions inline there is no reason to not call memcpy.
> 
> Thus if the CPU could do
> 
>   tem = unaligned-load-8-bytes-from-src-and-ignore-faults;
>   mask = generate mask from size
>   store-unaligned-8-bytes-with-maxk
> 
> then expanding the memcpy call inline would be a win I suppose.
> AVX has VMASKMOV, but I'm not sure using that for sizes <= 16
> bytes is profitable?  Note that from the specs
> of VMASKMOV it seems the memory operands need to be aligned and
> the mask does not support byte-granularity.
> 
> Which would leave us to inline expanding the case of at most 2 byte
> memcpy.  Of course currently there is no way to record an upper
> bound for the size (we do not retain value-range information - but
> we of course should).

My secret plan was to make VRP produce value profiling histogram
when value is known to be with small range.  Should be quite easy
to implement.

Honza