VEC_COND_EXPR optimizations v2

[Marc Glisse]

On Fri, 7 Aug 2020, Richard Biener wrote:

> On Thu, Aug 6, 2020 at 8:07 PM Marc Glisse <marc.glisse@inria.fr> wrote:
>>
>> On Thu, 6 Aug 2020, Christophe Lyon wrote:
>>
>>>> Was I on the right track configuring with
>>>> --target=arm-none-linux-gnueabihf --with-cpu=cortex-a9
>>>> --with-fpu=neon-fp16
>>>> then compiling without any special option?
>>>
>>> Maybe you also need --with-float=hard, I don't remember if it's
>>> implied by the 'hf' target suffix
>>
>> Thanks! That's what I was missing to reproduce the issue. Now I can
>> reproduce it with just
>>
>> typedef unsigned int vec __attribute__((vector_size(16)));
>> typedef int vi __attribute__((vector_size(16)));
>> vi f(vec a,vec b){
>>      return a==5 | b==7;
>> }
>>
>> with -fdisable-tree-forwprop1 -fdisable-tree-forwprop2 -fdisable-tree-forwprop3 -O1
>>
>>    _1 = a_5(D) == { 5, 5, 5, 5 };
>>    _3 = b_6(D) == { 7, 7, 7, 7 };
>>    _9 = _1 | _3;
>>    _7 = .VCOND (_9, { 0, 0, 0, 0 }, { -1, -1, -1, -1 }, { 0, 0, 0, 0 }, 107);
>>
>> we fail to expand the equality comparison (expand_vec_cmp_expr_p returns
>> false), while with -fdisable-tree-forwprop4 we do manage to expand
>>
>>    _2 = .VCONDU (a_5(D), { 5, 5, 5, 5 }, { -1, -1, -1, -1 }, { 0, 0, 0, 0 }, 112);
>>
>> It doesn't make much sense to me that we can expand the more complicated
>> form and not the simpler form of the same operation (both compare a to 5
>> and produce a vector of -1 or 0 of the same size), especially when the
>> target has an instruction (vceq) that does just what we want.
>>
>> Introducing boolean vectors was fine, but I think they should be real
>> types, that we can operate on, not be forced to appear only as the first
>> argument of a vcond.
>>
>> I can think of 2 natural ways to improve things: either implement vector
>> comparisons in the ARM backend (possibly by forwarding to their existing
>> code for vcond), or in the generic expansion code try using vcond if the
>> direct comparison opcode is not provided.
>>
>> We can temporarily revert my patch, but I would like it to be temporary.
>> Since aarch64 seems to handle the same code just fine, maybe someone who
>> knows arm could copy the relevant code over?
>>
>> Does my message make sense, do people have comments?
>
> So what complicates things now (and to some extent pre-existed when you
> used AVX512 which _could_ operate on boolean vectors) is that we
> have split out the condition from VEC_COND_EXPR to separate stmts
> but we do not expect backends to be able to code-generate the separate
> form - instead we rely on the ISEL pass to trasform VEC_COND_EXPRs
> to .VCOND[U] "merging" the compares again.  Now that process breaks
> down once we have things like _9 = _1 | _3;  -  at some point I argued
> that we should handle vector compares [and operations on boolean vectors]
> as well in ISEL but then when it came up again for some reason I
> disregarded that again.
>
> Thus - we don't want to go back to fixing up the generic expansion code
> (which looks at one instruction at a time and is restricted by TER single-use
> restrictions).

Here, it would only be handling comparisons left over by ISEL because they 
do not feed a VEC_COND_EXPR, maybe not so bad. Handling it in ISEL would 
be more consistent, but then we may have to teach the vector lowering pass 
about this.

> Instead we want to deal with this in ISEL which should
> behave more intelligently.  In the above case it might involve turning
> the _1 and _3 defs into .VCOND [with different result type], doing
> _9 in that type and then somehow dealing with _7 ... but this eventually
> means undoing the match simplification that introduced the code?

For targets that do not have compact boolean vectors, 
VEC_COND_EXPR<*,-1,0> does nothing, it is just there as a technicality. 
Removing it to allow more simplifications makes sense, reintroducing it 
for expansion can make sense as well, I think it is ok if the second one 
reverses the first, but very locally, without having to propagate a change 
of type to the uses. So on ARM we could turn _1 and _3 into .VCOND 
producing bool:32 vectors, and replace _7 with a VIEW_CONVERT_EXPR. Or 
does using bool vectors like that seem bad? (maybe they aren't guaranteed 
to be equivalent to signed integer vectors with values 0 and -1 and we 
need to query the target to know if that is the case, or introduce an 
extra .VCOND)

Also, we only want to replace comparisons with .VCOND if the target 
doesn't handle the comparison directly. For AVX512, we do want to produce 
SImode bool vectors (for k* registers) and operate on them directly, 
that's the whole point of introducing the bool vectors (if bool vectors 
were only used to feed VEC_COND_EXPR and were all turned into .VCOND 
before expansion, I don't see the point of specifying different types for 
bool vectors for AVX512 vs non-AVX512, as it would make no difference on 
what is passed to the backend).

I think targets should provide some number of operations, including for 
instance bit_and and bit_ior on bool vectors, and be a bit consistent 
about what they provide, it becomes unmanageable in the middle-end 
otherwise...

-- 
Marc Glisse