This is the mail archive of the
mailing list for the GCC project.
Re: [PATCH][RFC] Come up with VEC_COND_OP_EXPRs.
- From: Richard Biener <richard dot guenther at gmail dot com>
- To: Richard Sandiford <richard dot sandiford at arm dot com>
- Cc: Martin Liška <mliska at suse dot cz>, GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Tue, 24 Sep 2019 13:29:38 +0200
- Subject: Re: [PATCH][RFC] Come up with VEC_COND_OP_EXPRs.
- References: <firstname.lastname@example.org> <email@example.com>
On Tue, Sep 24, 2019 at 1:11 PM Richard Sandiford
> Martin Liška <firstname.lastname@example.org> writes:
> > Hi.
> > The patch introduces couple of new TREE_CODEs that will help us to have
> > a proper GIMPLE representation of current VECT_COND_EXPR. Right now,
> > the first argument is typically a GENERIC tcc_expression tree with 2 operands
> > that are visited at various places in GIMPLE code. That said, based on the discussion
> > with Richi, I'm suggesting to come up with e.g.
> > VECT_COND_LT_EXPR<COND_LHS, COND_RHS, IF_CLAUSE, ELSE_CLAUSE>. Such a change logically
> > introduces new GIMPLE_QUATERNARY_RHS gassignments. For now, the VEC_COND_EXPR remains
> > and is only valid in GENERIC and gimplifier will take care of the corresponding transition.
> > The patch is a prototype and missing bits are:
> > - folding support addition for GIMPLE_QUATERNARY_RHS is missing
> > - fancy tcc_comparison expressions like LTGT_EXPR, UNORDERED_EXPR, ORDERED_EXPR,
> > UNLT_EXPR and others are not supported right now
> > - comments are missing for various functions added
> > Apart from that I was able to bootstrap and run tests with a quite small fallout.
> > Thoughts?
> > Martin
> I think this is going in the wrong direction. There are some targets
> that can only handle VEC_COND_EXPRs well if we know the associated
> condition, and others where a compare-and-VEC_COND_EXPR will always be
> two operations. In that situation, it seems like the native gimple
> representation should be the simpler representation rather than the
> more complex one. That way the comparisons can be optimised
> independently of any VEC_COND_EXPRs on targets that benefit from that.
> So IMO it would be better to use three-operand VEC_COND_EXPRs with
> no embedded conditions as the preferred gimple representation and
> have internal functions for the fused operations that some targets
> prefer. This means that using fused operations is "just" an instruction
> selection decision rather than hard-coded throughout gimple. (And that
> fits in well with the idea of doing more instruction selection in gimple.)
So I've been doing that before, but more generally also for COND_EXPR.
We cannot rely on TER and the existing RTL expansion "magic" for the
instruction selection issue you mention because TER isn't reliable. With
IFNs for optabs we could do actual [vector] condition instruction selection
before RTL expansion, ignoring "single-use" issues - is that what you are
hinting at? How should the vectorizer deal with this? Should it directly
use the optab IFNs then when facing "split" COND_EXPRs? IIRC the
most fallout of a simple patch (adjusting is_gimple_condexpr) is in the
Note I'm specifically looking for a solution that applies to both COND_EXPR
and VEC_COND_EXPR since both suffer from the same issues.
There was also recent work in putting back possibly trapping comparisons
into [VEC_]COND_EXPR because it doesn't interfere with EH and allows
better code. Also you SVE people had VN issues with cond-exprs and
VN runs into the exact same issue (but would handle separate comparisons
better - with the caveat of breaking TER).