[Mlir-commits] [mlir] [mlir][Vector] Fix vector.extract lowering to llvm for 0-d vectors (PR #117731)

[Andrzej Warzyński]

================
@@ -1096,43 +1096,56 @@ class VectorExtractOpConversion
     SmallVector<OpFoldResult> positionVec = getMixedValues(
         adaptor.getStaticPosition(), adaptor.getDynamicPosition(), rewriter);
 
-    // Extract entire vector. Should be handled by folder, but just to be safe.
-    ArrayRef<OpFoldResult> position(positionVec);
-    if (position.empty()) {
-      rewriter.replaceOp(extractOp, adaptor.getVector());
-      return success();
-    }
-
-    // One-shot extraction of vector from array (only requires extractvalue).
-    // Except for extracting 1-element vectors.
-    if (isa<VectorType>(resultType) &&
-        position.size() !=
-            static_cast<size_t>(extractOp.getSourceVectorType().getRank())) {
-      if (extractOp.hasDynamicPosition())
-        return failure();
-
-      Value extracted = rewriter.create<LLVM::ExtractValueOp>(
-          loc, adaptor.getVector(), getAsIntegers(position));
-      rewriter.replaceOp(extractOp, extracted);
-      return success();
-    }
+    // The LLVM lowering models multi dimension vectors as stacked 1-d vectors.
----------------
banach-space wrote:

It's not really the LLVM lowering, but the Vector dialect:
* https://mlir.llvm.org/docs/Dialects/Vector/#llvm-lowering-tradeoffs

Note that over there it's referred to as "nested aggregates" rather than "stacking".

https://github.com/llvm/llvm-project/pull/117731