bzero optimization rarely does

[Kaveh R. Ghazi]

 > From: Roger Sayle <roger@eyesopen.com>
 > 
 > Hi Kaveh,
 > >  > 2002-07-11  Roger Sayle  <roger@eyesopen.com>
 > >  >
 > >  > 	* expr.c [CLEAR_RATIO]: New macro defining the maximum number
 > >  >         of move instructions to use when clearing memory,
 > >  >         c.f. MOVE_RATIO.  [CLEAR_BY_PIECES_P]: New macro, using
 > >  >         CLEAR_RATIO, to determine whether clear_by_pieces should be
 > >  >         used to clear storage.  (clear_storage): Use CLEAR_BY_PIECES_P
 > >  >         instead of MOVE_BY_PIECES_P.
 > >
 > > How about a testcase patterned after execute/string-opt-17.c?
 > > (Must len really be less than 2 before we abort in the static memset?)
 > 
 > The problem is that the backend has control over the decision of
 > whether to call memset or to use a sequence of clear instructions.
 > So for example, in ns32k.h the "#define MOVE_RATIO 0" means that
 > this platform always calls memset!

I thought the point of your patch was to divorce memset from
MOVE_RATIO?


 > When I wrote execute/string-opt-17.c and its corresponding memset
 > optimization, I always transformed writing a single byte with
 > memset into a byte store instruction.  This allows the "abort"
 > in the testcase to check whether the code fires at all.
 > 
 > Alas there is no simple check whether the clear_storage heuristics
 > are firing correctly.  All the sequences that GCC generates should
 > have identical behaviour, and the selection between them is target
 > and compiler option (-Os) specific.


You can always check __OPTIMIZE_SIZE__ and bypass the abort for -Os
just like we do for -O0 and __OPTIMIZE__.

Once your patch goes in (except with -Os) everybody should have a
CLEAR_RATIO of at least 15 which is good for more than len == 2,
right?  And some of the platforms can do any length if clrstr is
there.

If you're motivated you can always write something like what's in
gcc.dg/20020312-2.c to set the upper bound on how many max bytes
you'll test memset with for each cpu.  And for those cpus with clrstr
you add a test with a huge upper bound.

I know it seems like a pain, but in the end IMHO it's worth it because
then the regression checker ensures it's never broken ever again.

		--Kaveh
--
Kaveh R. Ghazi			Director of Systems Architecture
ghazi@caip.rutgers.edu		Qwest Solutions