[PATCH] Optimize MIPS memcpy

[Maxim Kuvyrkov]

On 28/09/2012, at 3:47 PM, Maxim Kuvyrkov wrote:

> On 21/09/2012, at 6:38 AM, Steve Ellcey wrote:
> 
>> On Thu, 2012-09-20 at 21:05 +1200, Maxim Kuvyrkov wrote:
>> 
>>> What testing was done for this patch, does it pass glibc testsuite?
>>> 
>>> I have a benchmark that exercises various string and mem* routines failing with it.
>>> 
>>> Thank you,
>>> 
>>> --
>>> Maxim Kuvyrkov
>>> CodeSourcery / Mentor Graphics
>> 
>> Is the benchmark anything you can share?  I ran the glibc testsuite and
>> got some failures but I don't think they are due to the new memcpy.  I
>> am going back now and running the glibc testsuite with no changes to get
>> a baseline so I can verify that.  Hopefully I will have an answer later
>> today.
>> 
>> I ran some other tests like the gcc testsuite using a glibc with this
>> change in it and that didn't have any problems and there is the one I
>> sent to the list
>> http://sourceware.org/ml/libc-ports/2012-09/msg00007.html that also ran
>> with no problems.
> 
> As I mentioned in a different email, I can't share the benchmark, but I think I've got a testcase of sorts for you to investigate.  It appears your memcpy clobbers a couple of bytes just before DEST in certain cases.
> 
> In particular when ABI=N32, DEST=0x10060008, SRC=0x1002c088, N=0x172 it clobbers DEST[-1] and DEST[-2] bytes.

Steve and I have debugged these failures and they now seem to be resolved.  I'll let Steve to followup with analysis and a new patch.

Meanwhile, I've benchmarked Steve's patch against mine.  On the benchmark that I use both implementations provide equal performance for N64 ABI, but on N32 ABI Steve's patch is only half as fast.  This is, probably, due to using 4-byte operations instead of 8-byte operations for N32 ABI:

#if _MIPS_SIM == _ABI64
#define USE_DOUBLE
#endif

It should be easy to improve Steve's patch for N32 ABI.  Steve, will you look into that?

I would also appreciate if you look into making your version of memcpy memmove-safe, if it is not already.

Thank you,

--
Maxim Kuvyrkov
CodeSourcery / Mentor Graphics