[Mlir-commits] [mlir] [mlir][memref] Introduce `memref.distinct_objects` op (PR #156913)

[Ivan Butygin]

================
@@ -465,6 +465,50 @@ struct AssumeAlignmentOpLowering
   }
 };
 
+struct DistinctObjectsOpLowering
+    : public ConvertOpToLLVMPattern<memref::DistinctObjectsOp> {
+  using ConvertOpToLLVMPattern<
+      memref::DistinctObjectsOp>::ConvertOpToLLVMPattern;
+  explicit DistinctObjectsOpLowering(const LLVMTypeConverter &converter)
+      : ConvertOpToLLVMPattern<memref::DistinctObjectsOp>(converter) {}
+
+  LogicalResult
+  matchAndRewrite(memref::DistinctObjectsOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    ValueRange operands = adaptor.getOperands();
+    if (operands.size() <= 1) {
+      // Fast path.
+      rewriter.replaceOp(op, operands);
+      return success();
+    }
+
+    Location loc = op.getLoc();
+    SmallVector<Value> ptrs;
+    for (auto [origOperand, newOperand] :
+         llvm::zip_equal(op.getOperands(), operands)) {
+      auto memrefType = cast<MemRefType>(origOperand.getType());
+      Value ptr = getStridedElementPtr(rewriter, loc, memrefType, newOperand,
----------------
Hardcode84 wrote:

Copying comment from the other discussion
@krzysz00 

> I'm not sure this is the right thing? I think we want bufferPtr on MemRefDescriptor.

That's a good question, imagine we have code like
```
%0 = memref.alloc
%1 = memref.subview %0[%c0]
%2 = memref.subview %0[%c1024]
%3, %4 = memrefs.distinct_objects %1, %2
```

If we just use `bufferPtr` without the offset we will en up with `assume separate_storage(%ptr, %ptr)`, i.e. same pointer, which makes no sense.

https://github.com/llvm/llvm-project/pull/156913