Pattern matching in a SelectionDAG

Project Infrastructure LLVM Dev List Archives
1

Hi,

I'm trying to write an LLVM backend using LLVM's retargetable code
generator, and I'm confused about how the SelectionDAG works. Let me give
you an example from the SPARC backend (as this is what is often mentioned
in the documentation). This is how the "branch always" instruction is
defined:

def BA : BranchSP<0b1000, (ins brtarget:$dst),
                      "ba $dst",
                      [(br bb:$dst)]>;

The pattern that is to be matched is simply (br bb: $dst). Based on this, I
would have expected the pattern for an add instruction to look somehow like
this: (add IntRegs:$b, IntRegs:$c). But in reality, it looks like this:
(set IntRegs:$dst, (add IntRegs:$b, IntRegs:$c))
What is this "set" thing all about? It doesn't seem to be a node in the
graph, as the pattern is also matched in a graph in which the add node's
ancestor is, say, a mul node. So what is it?

Cheers
Matthias

2

The graph of your second instruction looks like this:

$dst <----- +
/
/
$b $c

No set in BA instruction pattern because no dest register should be updated.
in ADD instruction pattern, you should update your dest register $dst with the value (add IntRegs:$b, IntRegs:$c).