Re: [RFC] Try vector<bool> as a new representation for vector masks

[Richard Biener <richard.guenther@gmail.com>]

On Thu, Sep 24, 2015 at 6:37 PM, Richard Henderson <rth@redhat.com> wrote:
> On 09/24/2015 01:09 AM, Richard Biener wrote:
>> Both are basically a (target) restriction on how we should expand a conditional
>> move (and its condition).  It's techincally convenient to tie both together by
>> having them in the same statement but it's also techincally very incovenient
>> in other places.  I'd say for targets where
>>
>> tem_1 = a_2 < b_3;
>> res_4 = tem_1 ? c_5 : d_6;
>> res_7 = tem_1 ? x_8 : z_9;
>>
>> presents a serious issue ("re-using" the flags register) out-of-SSA should
>> duplicate the conditionals so that TER can do its job (and RTL expansion
>> should use TER to get at the flags setter).
>
> Sure it's a target restriction, but it's an extremely common one.  Essentially
> all of our production platforms have it.  What do we gain by adding some sort
> of target hook for this?

A cleaner IL, no GENERIC expression tree building in GIMPLE (I guess that's
sth Andrew needs for his GIMPLE types project as well), less awkward
special-casing of comparisons based on context in code like genmatch.c
or in value-numbering.

>> I imagine that if we expand the above to adjacent statements the CPUs can
>> re-use the condition code.
>
> Sure, but IMO it should be the job of RTL CSE to make that decision, after all
> of the uses (and clobbers) of the flags register have been exposed.
>
>> To me where the condition is in GIMPLE is an implementation detail and the
>> inconveniences outweight the benefits.
>
> Why is a 3-operand gimple statement fine, but a 4-operand gimple statement
> inconvenient?

The inconvenience is not the number of operands but that we have two operation
codes and that we compute two values but only have an SSA name def for one
of them.  Oh, and did I mention that second operation is GENERIC?

So one way to clean things up would be to no longer use a GIMPLE_ASSIGN for
x = a < b ? c : d but instead use a GIMPLE_COND and give that a SSA def
for the result, using the true/false label operand places for 'c' and 'd'.

That still wouldn't get the compare a SSA def but at least it would get rid of
the 2nd operator code and the GENERIC expression operand.

From the GIMPLE side forcing out the comparison to a separate stmt looks
more obvious and if we're considering doing a different thing then we may as
well think of how to represent predicating arbitrary stmts or how to explicitely
model condition codes in GIMPLE.

It kind of looks like we want a GIMPLE PARALLEL ... (we already have
a GIMPLE stmt with multiple defs - GIMPLE_ASM)

Richard.

>
> r~