[LLVMdev] Simplifing the handling of pre-legalize vector nodes
sabre at nondot.org
Wed May 23 00:03:30 PDT 2007
Sorry for the delay, things are more insane than usual, with the pending
release, the dev mtg, etc :)
On Mon, 21 May 2007, Dan Gohman wrote:
> Right now there are special SelectionDAG node kinds for operations on
> "abstract" vector types (VLOAD, VADD, and all the rest), and a special
> MVT::Vector ValueType for them. These nodes carry two additional operands,
> constants which specify the vector length and element type. All of this is
> only used before legalize; then they are replaced with regular node kinds
> and value types.
> It seems that a number of things would be considerably simpler if the
> pre-legalize nodes could use the same node kinds as post-legalize; the only
> thing preventing that is that the MVT::ValueType enum isn't able to describe
> vector types with arbitrarily long vector lengths.
Yep, I'm sure you know this, but this is to support generic vectors. For
example, if you had an input .ll that operated on 128-wide vectors, we
want to be able to split that up to use 4-wide vectors if your target
Going forward, we will also likely have to do something similar to this
for integer types, in order to handle stuff like i47 that gets legalized
into 2 x i32 or something. I am not looking forward to duplicating all of
the arithmetic nodes for IADD variants etc (urk, imagine vectors of
> I'm currently considering ways to make this possible. One idea is to rename
> the MVT::ValueType enum to something else and make MVT::ValueType a plain
> integer type. Values beyond the range of the original enum are interpreted
> as indices into a UniqueVector which holds pairs of vector lengths and
> element types.
That is a very interesting idea. It handles the generality of arbitrary
value types, but is nice and fast for the common code that doesn't use the
craziness :). I like it!
> This would require most of the ValueType utility routines
> to be aware of the new kinds of ValueTypes so that they could handle them
> properly, but once that's done, it would allow vector nodes to be handled
> consistently between pre-legalize and post-legalize, without obscure
> constructs like
> ... blahblah->getNumOperands()-2 ...
> ... *(blahblah->op_end()-1) ...
> to get the vector length and/or element type in the pre-legalize form.
That would be amazingly wonderful. The other nice thing about this is
that 'normal' code that uses (e.g.) sse intrinsics wouldn't have to have
all the nodes rewritten multiple times from VADD->ADD->... Anything in a
legal type would just be created and stay legal.
> It would make -view-legalize-dags and earlier DAG drawings a lot easier to
> read in DAGS with vector nodes, because they wouldn't need to be edges from
> each vector node up to the common element type node and vector length node.
> Finally, and of particular interest for me, it may help with the next steps
> beyond PR400, as it would allow vector loads to use plain LoadSDNode nodes
> instead of VLOAD, and thus have an alignment, not to mention a volatile flag
> and alias information, in a consistent manner with regular loads.
Yep. For this specific one, it would be possible to enhance vload, but I
agree that this is not a very appealing solution :)
> Before I do much more experimentation with this, I wanted to run the idea by
> the list to see what feedback it might get. Has anyone thought about doing
> this before? Are there other approaches that might be better?
This approach sounds very sensible. Make sure the SelectionDAG owns the
> Another idea is to get rid of the ValueType concept altogether and just use
> Type* for everything. I don't know how this would work with various
> tablegen'd pieces of code though.
I'd suggest sticking with ValueTypes. Among other things, they can be
indexed into arrays and Type*'s can't. This is important for the various
tables in targetlowering that the legalizer consults to find out if a
type/operation is valid.
More information about the llvm-dev