fix PR46029: reimplement if conversion of loads and stores

[Alan Lawrence]

Sebastian Pop wrote:
> On Thu, Jun 25, 2015 at 4:43 AM, Richard Biener
> <richard.guenther@gmail.com> wrote:
>> when the new scheme triggers vectorization cannot succeed on the
>> result as we get
>>
>>   if (cond)
>>     *p = val;
>>
>> if-converted to
>>
>>   tem = cond ? p : &scratch;
>>   *tem = val;
> 
> That's correct.
> 
>> and
>>
>>    if (cond)
>>      val = *p;
>>
>> if-converted to
>>
>>   scatch = val;
>>   tem = cond ? p : &scratch;
>>   val = *tem;
> 
> The patch does this slightly differently:
> 
>    tem = cond ? p : &scratch;
>    val = cond ? *tem : val;
> 
> I think I like your version better as it has only one cond_expr.

Another slight concern...by reusing scratchpad's, are we at risk of creating 
lots of false dependencies here, that restrict instruction scheduling / other 
opts later?

>> [...]
>> and thus the store and loads appear as scather/gather ones to
>> the vectorizer (if-conversion could directly generate masked
>> load/stores of course and not use a scratch-pad at all in that case).

Thank you Richard for much better expressing what I was thinking here too :)

> Abe also suggested to continue optimizing the other way in cases
> where we know to write or load from the same location on all branches:
> 
> if (c)
>   A[i] = ...
> else
>   A[i] = ...

The store here really should be commoned, yes.

(Related but different?: BZ 56625.)


I might add, I find this code much easier to follow than the old (removed) code 
about data references, memrefs_read_unconditionally, and 
write_memrefs_written_at_least-once...

Thanks, Alan