[Mlir-commits] [mlir] [MLIR] Implement emulation of static indexing subbyte type vector stores (PR #115922)

[llvmlistbot at llvm.org]

================
@@ -292,6 +296,100 @@ emulatedVectorLoad(OpBuilder &rewriter, Location loc, Value base,
       newLoad);
 }
 
+/// Downcast two values to `downcastType`, then select values
+/// based on `mask`, and casts the result to `upcastType`.
+static Value downcastSelectAndUpcast(OpBuilder &builder, Location loc,
+                                     VectorType downcastType,
+                                     VectorType upcastType, Value mask,
+                                     Value trueValue, Value falseValue) {
+  assert(
+      downcastType.getNumElements() * downcastType.getElementTypeBitWidth() ==
+          upcastType.getNumElements() * upcastType.getElementTypeBitWidth() &&
+      "expected upcastType size to be twice the size of downcastType");
+  if (trueValue.getType() != downcastType)
+    trueValue = builder.create<vector::BitCastOp>(loc, downcastType, trueValue);
+  if (falseValue.getType() != downcastType)
+    falseValue =
+        builder.create<vector::BitCastOp>(loc, downcastType, falseValue);
----------------
lialan wrote:

:-) You are right. But here I actually prefer having those checks if you don't mind, because generating something like:
```mlir
288:  %8 = vector.bitcast %1 : vector<4xi2> to vector<4xi2>
```

actually adds more IR lines, so I personally love to make it a little bit cleaner :-)


https://github.com/llvm/llvm-project/pull/115922