[PATCH] D93312: [RISCV] Add ISel support for RVV .vx and .vi forms

[Fraser Cormack via Phabricator via llvm-commits]

frasercrmck added inline comments.


================
Comment at: llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td:636
+let Complexity = 1 in
+def SplatPat       : ComplexPattern<vAny, 1, "SelectVSplat", [splat_vector, rv32_splat_i64]>;
+def SplatPat_simm5 : ComplexPattern<vAny, 1, "SelectVSplat_simm5", [splat_vector, rv32_splat_i64]>;
----------------
craig.topper wrote:
> frasercrmck wrote:
> > craig.topper wrote:
> > > I don't think you want anything in the square brackets. That autogenerates a pattern for splat_vector/rv32_splat_i64 as the root node for the isel match. I guess your vmv.v.v pattern is getting priority so its not causing a problem?
> > I'm not sure I follow, sorry. I think I'll need to read up on `ComplexPattern`s as I've managed to avoid them thus far.
> The list in square brackets is not used when the ComplexPattern is used as an operand of another node. That's why you have to check for the opcode in the C++.
> 
> The bracket list is only used if you write something like this
> 
> ```
> def : Pat<(SplatPat GPR:$rs)>;
> ```
> 
> Where the ComplexPattern is the first node of the pattern. You have patterns that where splat_vector, rv32_splat_i64, and rv32_splat_i64 appear as the first node and don't use SplatPat this way.
Right, thanks! I just looked through the TableGen source and I guess the clue's in the name: "root" nodes. I'll get rid of them as you suggest, since I'm not using them.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D93312/new/

https://reviews.llvm.org/D93312