[PATCH] powerpc64le: add optimized strlen for P9

[Paul E Murphy]

On 6/3/20 3:44 PM, Paul A. Clarke wrote:
>> +	vctzlsbb  r3,v6
>> +	blr
>> +
>> +	/* Test 64B 16B at a time.  The 64B vector loop is optimized for
>> +	   longer strings.  Likewise, we check a multiple of 64B to avoid
>> +	   breaking the alignment calculation below.  */
>> +L(aligned):
>> +	add	  r4,r3,r9
>> +	rldicl.	  r5,r4,60,62  /* Determine the number of 48B loops needed for
>> +                                  alignment to 64B.  And test for zero.  */
> 
> Would it be bad to move the "rldicl." down...
> 
>> +
>> +	lxv	  v0+32,0(r4)
>> +	vcmpequb. v6,v0,v18
>> +	bne	  cr6,L(tail1)
>> +
>> +	lxv	  v0+32,16(r4)
>> +	vcmpequb. v6,v0,v18
>> +	bne 	  cr6,L(tail2)
>> +
>> +	lxv	  v0+32,32(r4)
>> +	vcmpequb. v6,v0,v18
>> +	bne 	  cr6,L(tail3)
>> +
>> +	lxv	  v0+32,48(r4)
>> +	vcmpequb. v6,v0,v18
>> +	bne 	  cr6,L(tail4)
> 
> ...to here, to avoid needlessly penalizing the cases above?

My thinking is the integer op is basically free while we're busy 
executing vector ops (similarly for li).  Likewise, this may put enough 
space between the conditional branch below to ease prediction.  I don't 
claim to have run this against a cycle accurate simulator, so it may not 
really matter.

Similarly, binutils 2.26 should support all P9 instructions here and in 
Anton's other patches.  I did check before pushing the rawmemchr/stpcpy.