[lno] [RFC] if-conversion and auto vectorizer

[Daniel Berlin]

On Mar 14, 2004, at 2:22 PM, Andrew Pinski wrote:

>
> On Mar 14, 2004, at 03:07, Dorit Naishlos wrote:
>
>>> I will also be doing the following soon (after MIN/MAX becomes gimple
>>> which has already be discussed and approved):
>>
>> I seem to have missed the MIN/MAX discussion thread. Do you have a 
>> link to
>> that, or was it off the mailing list? just curious when is it planned 
>> to
>> add the MIN/MAX tree codes to gimple and who is responsible for that?
>
> It was discussed off list (on IRC).  Really MIN/MAX is already gimple, 
> just that it gets
> lowered when gimplifing happens.  Also right now DOM does not 
> understand MIN/MAX.
> The current plan is remove the lowering when the tree-ssa gets merged 
> into the mainline.
> I think Daniel Berlin said he would be the one to remove the lowering 
> as the linear loop
> transformation generates better code with MIN/MAX instead of different 
> basic blocks.
It's not better code so much as much less complex code.  This is in the 
case of integer min/max.
Otherwise, performing a min/max requires 1 cond_expr, two new basic 
blocks, two new labels, and two new gotos, per MAX/MIN operation.
In the case of floating point min/max, RTH pointed out that it is very 
hard to put the lowered form back together into MAX/MIN.

> I think he also has a patch to teach DOM about it too but I am not too 
> sure.
>

There isn't really much to teach DOM about here.

Previously, DOM saw

if (a < b)
	c = a;
else
	c = b;

It knew that in each arm, a was either < or > b, respectively.

With
c = MAX (a, b);

we know nothing about each arm (because they are implicit), but it 
doesn't actually matter, because you couldn't have done anything with 
the info anyway here.

So i'm not sure what you want me to teach DOM.