[RFC] optabs and tree-codes for vector operations

[James E Wilson]

Dorit Naishlos wrote:
> The attached html document contains the first major set of proposed
> optabs/tree-codes for vectorization.

I read through this.  It seems like a reasonable start.  I do have some 
comments, but the way you have posted this makes it hard for me to quote 
you.  Hopefully my comments will make sense.

First a general comment, that we need better definitions of what the 
existing vector operations do.  They are ambiguous as to endianness, and 
some targets defining them differently for different endians, and some 
don't.  Also, there appears to be confusion about what vec_merge means, 
as it means something different in the altivec.md file than it does in 
all other md files.  We need well defined definitions used consistently 
by all ports before we can have good optimization support for vectors.

Trees have type info, RTL does not.  So we have two kinds of compare 
optabs to represent signed/unsigned compares, and different rtl 
operations for signed/unsigned compares, but we only need one tree 
operator to represent both.

I don't see any reason why we can't use the existing 
cmp_optab/ucmp_optab for vectors.

The current RTL vec_select operator takes a parallel containing a list 
of integers.  You are defining the tree operator VSELECT_EXPR as taking 
a vector of condition codes, produced by a vector compare.  Also, 
vec_select takes one vector operand, and VSELECT_EXPR takes two.  These 
are very different operations.  This could be confusing. vec_select is 
really more like a shuffle, whereas vec_merge is more like your 
VSELECT_EXPR.

The current RTL vec_select operator is really not much different than 
the vec_merge operator.  It just rearranges elements in a different way. 
  I have found that almost anything that can be expressed with one can 
be expressed with the other.  Maybe redefining the RTL vec_select 
operator would make sense to eliminate the duplication.

This duplication at the RTL level could make optimization difficult. 
All RTL needs to have a canonical representation for combine and other 
optimizations to work, but with the current situation there is no 
obvious way to pick whether vec_select or vec_merge is more canonical 
than the other.  The result is that targets may get stuck defining 
multiple patterns to match every way to express an operation, which is 
undesirable.

There is a general principle here, that you should avoid defining new 
RTL operations that can be easily represented by others, and that 
prevents us from easily determining the canonical representations. 
There can also be similar problems at the tree level, since optimizers 
need to know how to simplify something, and this can be a problem if 
there are multiple ways to represent something.

In MIPS, a vector compare does not return a vector mask of 0/1.  It 
returns a vector of condition code results.  We have 8 condition code 
registers, and if you compare an order 8 vector, it sets every condition 
code register.  Similarly, the conditional move instructions check 
multiple condition code registers.  The current RTL vec_merge/vec_select 
can not express this, because neither one can take a vector of condition 
code registers as input.  This should be considered in the design of the 
vector compare support.

Your compare proposals don't seem to take the needs of FP into account. 
  We need additional comparison operators for FP, because of unordered 
results.  This also matters when inverting FP compares, as the opposite 
of < is not >=.

The VPERM_EXPR you define is similar to the vec_select rtx operation 
that already exists.  It is confusing to have closely related tree and 
rtl operators with different names.

Your VMERGE_EXPR proprosal is confusing.  You defined it with 
permutation, but the existing vec_merge RTL operator does not do 
permutation.  Also, you didn't define how you can do the permutation if 
you only have one index mask.  However, I see that you defined the range 
of a 4-element vector mask to be 0-7 which implies it can index any 
element of either vector.  That seems confusing to me.  Why not just 
have a separate permute and merge operations?  That is what we do now at 
the RTL level with vec_select and vec_merge.

If a 4 element vector mask can go from 0-7, then what happens for 
VPERM_EXPR which has only one input vector?  You need a different  kind 
of vector mask here that only goes from 0 to 3.

misalignment can already be handled with shifts or rotate.  It isn't 
clear that we need a special operator for this.

I think VSEL_MULT_EXPR is an unnecessary complication.  Similarly the 
lo/hi/odd/even stuff.  This can all be represented with a permutation 
and a multiply.

It isn't clear that we need MULT_HI_EXPR.

I suppose there is some trade off here about having tree operations that 
can represent VMX instructions, so they can be directly generated, 
versus have an optimization pass like combine create some of them after 
the fact.  It isn't obvious without implementation experience where to 
draw the line.

These complicated tree operations will all need to be expanded into rtl 
operations for targets that don't have appropriate instructions.  So 
there is also a trade off here about how much middle end code we need 
for expanding tree operations.

The saturation, packing, unpacking stuff seems pretty reasonable.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com