Re: [PATCH] gimple-match: Do not try UNCOND optimization with COND_LEN.

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

On Mon, Oct 16, 2023 at 11:59 PM Richard Sandiford
<richard.sandiford@arm.com> wrote:
>
> Robin Dapp <rdapp.gcc@gmail.com> writes:
> >> Why are the contents of this if statement wrong for COND_LEN?
> >> If the "else" value doesn't matter, then the masked form can use
> >> the "then" value for all elements.  I would have expected the same
> >> thing to be true of COND_LEN.
> >
> > Right, that one was overly pessimistic.  Removed.
> >
> >> But isn't the test whether res_op->code itself is an internal_function?
> >> In other words, shouldn't it just be:
> >>
> >>       if (internal_fn_p (res_op->code)
> >>        && internal_fn_len_index (as_internal_fn (res_op->code)) != -1)
> >>      return true;
> >>
> >> maybe_resimplify_conditional_op should already have converted to an
> >> internal function where possible, and if combined_fn (res_op->code)
> >> does any extra conversion on the fly, that conversion won't be reflected
> >> in res_op.
> >
> > I went through some of our test cases and believe most of the problems
> > are due to situations like the following:
> >
> > In vect-cond-arith-2.c we have (on riscv)
> >   vect_neg_xi_14.4_23 = -vect_xi_13.3_22;
> >   vect_res_2.5_24 = .COND_LEN_ADD ({ -1, ... }, vect_res_1.0_17, vect_neg_xi_14.4_23, vect_res_1.0_17, _29, 0);
> >
> > On aarch64 this is a situation that matches the VEC_COND_EXPR
> > simplification that I disabled with this patch.  We valueized
> > to _26 = vect_res_1.0_17 - vect_xi_13.3_22 and then create
> > vect_res_2.5_24 = VEC_COND_EXPR <loop_mask_22, _26, vect_res_1.0_19>;
> > This is later re-assembled into a COND_SUB.
> >
> > As we have two masks or COND_LEN we cannot use a VEC_COND_EXPR to
> > achieve the same thing.  Would it be possible to create a COND_OP
> > directly instead, though?  I tried the following (not very polished
> > obviously):
> >
> > -      new_op.set_op (VEC_COND_EXPR, res_op->type,
> > -                    res_op->cond.cond, res_op->ops[0],
> > -                    res_op->cond.else_value);
> > -      *res_op = new_op;
> > -      return gimple_resimplify3 (seq, res_op, valueize);
> > +      if (!res_op->cond.len)
> > +       {
> > +         new_op.set_op (VEC_COND_EXPR, res_op->type,
> > +                        res_op->cond.cond, res_op->ops[0],
> > +                        res_op->cond.else_value);
> > +         *res_op = new_op;
> > +         return gimple_resimplify3 (seq, res_op, valueize);
> > +       }
> > +      else if (seq && *seq && is_gimple_assign (*seq))
> > +       {
> > +         new_op.code = gimple_assign_rhs_code (*seq);
> > +         new_op.type = res_op->type;
> > +         new_op.num_ops = gimple_num_ops (*seq) - 1;
> > +         new_op.ops[0] = gimple_assign_rhs1 (*seq);
> > +         if (new_op.num_ops > 1)
> > +           new_op.ops[1] = gimple_assign_rhs2 (*seq);
> > +         if (new_op.num_ops > 2)
> > +           new_op.ops[2] = gimple_assign_rhs2 (*seq);
> > +
> > +         new_op.cond = res_op->cond;
> > +
> > +         gimple_match_op bla2;
> > +         if (convert_conditional_op (&new_op, &bla2))
> > +           {
> > +             *res_op = bla2;
> > +             // SEQ should now be dead.
> > +             return true;
> > +           }
> > +       }
> >
> > This would make the other hunk (check whether it was a LEN
> > and try to recreate it) redundant I hope.
> >
> > I don't know enough about valueization, whether it's always
> > safe to do that and other implications.  On riscv this seems
> > to work, though and the other backends never go through the LEN
> > path.  If, however, this is a feasible direction it could also
> > be done for the non-LEN targets?
>
> I don't know much about valueisation either :)  But it does feel
> like we're working around the lack of a LEN form of COND_EXPR.
> In other words, it seems odd that we can do:
>
>   IFN_COND_LEN_ADD (mask, a, 0, b, len, bias)
>
> but we can't do:
>
>   IFN_COND_LEN (mask, a, b, len, bias)
>
> There seems to be no way of applying a length without also finding an
> operation to perform.

Indeed .. maybe - _maybe_ we want to scrap VEC_COND_EXPR for
IFN_COND{,_LEN} to be more consistent here?

> Does IFN_COND_LEN make conceptual sense on RVV?  If so, would defining
> it solve some of these problems?
>
> I suppose in the worst case, IFN_COND_LEN is equivalent to IFN_COND_LEN_IOR
> with a zero input (and extended to floats).  So if the target can do
> IFN_COND_LEN_IOR, it could implement IFN_COND_LEN using the same instruction.

In principle one can construct a mask from the length via {0, 1, ... }
< len and then
AND that to the mask in a VEC_COND_EXPR but that's of course super ugly and
likely inefficient (or hard to match back on RTL land).

Richard.

> Thanks,
> Richard
>