[PATCH 2/2] rs6000: tune loop size for cunroll at O2

[Jiufu Guo]

"Kewen.Lin" <linkw@linux.ibm.com> writes:

> Hi Jeff,
>
> on 2020/5/20 上午11:58, Jiufu Guo via Gcc-patches wrote:
>> Hi,
>> 
>> This patch check the size of a loop to be unrolled/peeled completely,
>> and set the limits to a number (24).  This prevents large loop from
>> being unrolled, then avoid binary size increasing, and this limit keeps
>> performance.
>> 
>> Bootstrap&regtest pass on powerpc64le, ok for trunk?
>> 
>> Jiufu
>> 
>> ---
>>  gcc/config/rs6000/rs6000.c | 19 ++++++++++++++++++-
>>  1 file changed, 18 insertions(+), 1 deletion(-)
>> 
>> diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
>> index a1a3f9cb583..f3abb92d046 100644
>> --- a/gcc/config/rs6000/rs6000.c
>> +++ b/gcc/config/rs6000/rs6000.c
>> @@ -5142,6 +5142,22 @@ rs6000_loop_unroll_adjust (unsigned nunroll, struct loop *loop)
>>    return nunroll;
>>  }
>>  
>> +/* Count the number of statements in LOOP.  */
>> +
>> +static int
>> +num_stmt_in_loop (class loop *loop)
>> +{
>> +  int res = 0;
>> +  basic_block *bbs = get_loop_body (loop);
>> +  for (unsigned i = 0; i < loop->num_nodes; i++)
>> +    for (gimple_stmt_iterator bsi = gsi_start_bb (bbs[i]); !gsi_end_p (bsi);
>> +	 gsi_next (&bsi))
>> +      if (!is_gimple_debug (gsi_stmt (bsi)))
>> +	res++;
>> +
>> +  return res;
>> +}
>> +
>
> Would it be possible to use tree_num_loop_insns here?
> If no, this can be part of tree-ssa-loop.c, and we need to free(bbs).

free (bbs) is missed here. 
tree_num_loop_insns may be used, I would have a try. 

Thanks!
Jiufu(Jeff)

>
> BR,
> Kewen
>
>>  /* Implement targetm.loop_allow_unroll_completely_peel.  */
>>  
>>  static bool
>> @@ -5151,7 +5167,8 @@ rs6000_loop_allow_unroll_completely_peel (HOST_WIDE_INT maxiter, tree niter,
>>    if (unroll_only_small_loops && optimize == 2)
>>      {
>>        if (maxiter >= 4
>> -	  && !(TREE_CODE (niter) == INTEGER_CST && single_exit (loop)))
>> +	  && !(TREE_CODE (niter) == INTEGER_CST && num_stmt_in_loop (loop) < 24
>> +	       && single_exit (loop)))
>>  	return false;
>>      }
>>  
>>