[llvm-dev] Question about instruction selection

[Tim Northover via llvm-dev]

On Sat, 19 Jan 2019 at 02:03, Josh Sharp via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> In InstrInfo.td, MOVR has this dag pattern: [(set CPURegs:$ra, CPURegs:$rb)]. I don't understand how node t1's pattern matched MOVR's pattern.

A plain register class will match any node of the correct type, in
this case trivially. You'd expect that not to actually consume
anything and go into infinite recursion, but maybe the mere fact that
it has matched means it won't be considered again. I can't quite
picture what the DAG looks like after this match, but I suspect it's
actually cyclic and certainly not good.

You never really want to write patterns for plain moves like that;
MOVRs will get inserted later by calling the copyPhysReg function when
needed.

> Can someone point me to a location in the code where FrameIndex's pattern is defined?

A node you can use in more elaborate patterns is "frameindex" in
TargetSelectionDAG.td.cd, though that one is usually handled in C++
code rather than with TableGen patterns.

The idea is that you convert the FrameIndex into some kind of add
instruction that can operate on SP with a TargetFrameIndex as an
operand and usually dummy zeroes for the offset. That instruction is
usually a bit odd because what would normally be the LHS addition
register is this TargetFrameIndex node instead, and gets lowered into
actual stack arithmetic later. See for example
AArch64ISelDAGToDAG.cpp, around line 2888.

Cheers.

Tim.