[Mlir-commits] [mlir] [mlir][vector] VectorEmulateNarrowType uses deinterleave (PR #94946)

[Cullen Rhodes]

================
@@ -922,39 +922,23 @@ static Value rewriteI8ToI4Trunc(PatternRewriter &rewriter, Location loc,
          "Expected i8 type");
 
   // 1. De-interleave low and high i8 elements.
-  int64_t vecDimSize = srcVecType.getShape().back();
-  SmallVector<int64_t> deinterleaveLowMaskValues;
-  SmallVector<int64_t> deinterleaveHighMaskValues;
-  assert((vecDimSize % 2) == 0 && "Odd number of i4 elements");
-  deinterleaveLowMaskValues.reserve(vecDimSize / 2);
-  deinterleaveHighMaskValues.reserve(vecDimSize / 2);
-  for (int i = 0, end = vecDimSize; i < end; i += 2) {
-    deinterleaveLowMaskValues.push_back(i);
-    deinterleaveHighMaskValues.push_back(i + 1);
-  }
-
-  auto lowShuffleOp = rewriter.create<vector::ShuffleOp>(
-      loc, srcValue, srcValue,
-      rewriter.getI64ArrayAttr(deinterleaveLowMaskValues));
-  auto highShuffleOp = rewriter.create<vector::ShuffleOp>(
-      loc, srcValue, srcValue,
-      rewriter.getI64ArrayAttr(deinterleaveHighMaskValues));
+  auto deinterleaveOp = rewriter.create<vector::DeinterleaveOp>(loc, srcValue);
----------------
c-rhodes wrote:

the doc at the top of this function (`rewriteI8ToI4Trunc`) needs updating to reflect this now uses deinterleave rather than shuffles?

https://github.com/llvm/llvm-project/pull/94946