[vector][mlir] Restrict vector.shuffle to fixed-width vectors #88733
Conversation
At the moment there is no support for vector.shuffle for scalable vectors - various hooks/helpers related to `vector.shuffle` simply ignore the scalable flags (e.g. ` ShuffleOp::inferReturnTypes`). This is unlikely to change any time soon (vector shuffles are known to be tricky for scalable vectors), hence this patch restricts `vector.shuffle` to fixed width vectors.
|
@llvm/pr-subscribers-mlir-vector @llvm/pr-subscribers-mlir Author: Andrzej Warzyński (banach-space) ChangesAt the moment there is no support for vector.shuffle for scalable This is unlikely to change any time soon (vector shuffles are known to Full diff: https://github.com/llvm/llvm-project/pull/88733.diff 2 Files Affected:
diff --git a/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td b/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
index 147bc2354977d7..de6617c2062b0c 100644
--- a/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
+++ b/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
@@ -420,7 +420,7 @@ def Vector_ShuffleOp :
PredOpTrait<"second operand v2 and result have same element type",
TCresVTEtIsSameAsOpBase<0, 1>>,
InferTypeOpAdaptor]>,
- Arguments<(ins AnyVectorOfAnyRank:$v1, AnyVectorOfAnyRank:$v2,
+ Arguments<(ins AnyFixedVector:$v1, AnyFixedVector:$v2,
I64ArrayAttr:$mask)>,
Results<(outs AnyVector:$vector)> {
let summary = "shuffle operation";
@@ -444,6 +444,8 @@ def Vector_ShuffleOp :
mask values must be within range, viz. given two k-D operands v1 and v2
above, all mask values are in the range [0,s_1+t_1)
+ Note, at the moment only fixed-width vectors are supported.
+
Example:
```mlir
diff --git a/mlir/test/Dialect/Vector/invalid.mlir b/mlir/test/Dialect/Vector/invalid.mlir
index c16f1cb2876dbd..c9f7e9c6e2fb0b 100644
--- a/mlir/test/Dialect/Vector/invalid.mlir
+++ b/mlir/test/Dialect/Vector/invalid.mlir
@@ -84,6 +84,13 @@ func.func @shuffle_index_out_of_range(%arg0: vector<2xf32>, %arg1: vector<2xf32>
// -----
+func.func @shuffle_scalable_vec(%arg0: vector<[2]xf32>, %arg1: vector<[2]xf32>) {
+ // expected-error@+1 {{'vector.shuffle' op operand #0 must be fixed-length vector of any type values}}
+ %1 = vector.shuffle %arg0, %arg1 [0, 1, 2, 3] : vector<[2]xf32>, vector<[2]xf32>
+}
+
+// -----
+
func.func @shuffle_empty_mask(%arg0: vector<2xf32>, %arg1: vector<2xf32>) {
// expected-error@+1 {{'vector.shuffle' op invalid mask length}}
%1 = vector.shuffle %arg0, %arg1 [] : vector<2xf32>, vector<2xf32>
|
| @@ -444,6 +444,8 @@ def Vector_ShuffleOp : | |||
| mask values must be within range, viz. given two k-D operands v1 and v2 | |||
| above, all mask values are in the range [0,s_1+t_1) | |||
|
|
|||
| Note, at the moment only fixed-width vectors are supported. | |||
There was a problem hiding this comment.
nit: I'd drop "at the moment" as it suggests support is missing but as you mention it seems unlikely this will change given we have dedicated ops like splat / broadcast and interleave.
…ectors Remove a test that's no longer valid
|
I'm about to land this. Windows CI is failing, but that's when building Flang (i.e. unrelated to this change):
|
At the moment there is no support for vector.shuffle for scalable
vectors - various hooks/helpers related to
vector.shufflesimplyignore the scalable flags (e.g.
ShuffleOp::inferReturnTypes).This is unlikely to change any time soon (vector shuffles are known to
be tricky for scalable vectors), hence this patch restricts
vector.shuffleto fixed width vectors.