[mlir] Fix correct memset range in OwningMemRef zero-init (llvm#158200)

`OwningMemref` allocates with overprovision + manual alignment.
This is fixing the zero-initialization of the data, the existing code
was potentially overrunning the allocation:

```cpp
memset(descriptor.data, 0, size + desiredAlignment); // ❌ may overrun
```

This is invalid because `descriptor.data` (the aligned pointer) **does
not point to the full allocated block** (`size + desiredAlignment`).
Zeroing that much from the aligned start can write past the end of the
allocation.

Instead we only initialize the data from the aligned pointer for the expected
buffer size. The padding from [allocatedPtr, alignedDataPtr] is left untouched.

Author: chokobole

mlir/include/mlir/ExecutionEngine/MemRefUtils.h

@@ -164,19 +164,17 @@ class OwningMemRef {
     int64_t nElements = 1;
     for (int64_t s : shapeAlloc)
       nElements *= s;
-    auto [data, alignedData] =
+    auto [allocatedPtr, alignedData] =
         detail::allocAligned<T>(nElements, allocFun, alignment);
-    descriptor = detail::makeStridedMemRefDescriptor<Rank>(data, alignedData,
-                                                           shape, shapeAlloc);
+    descriptor = detail::makeStridedMemRefDescriptor<Rank>(
+        allocatedPtr, alignedData, shape, shapeAlloc);
     if (init) {
       for (StridedMemrefIterator<T, Rank> it = descriptor.begin(),
                                           end = descriptor.end();
            it != end; ++it)
         init(*it, it.getIndices());
     } else {
-      memset(descriptor.data, 0,
-             nElements * sizeof(T) +
-                 alignment.value_or(detail::nextPowerOf2(sizeof(T))));
+      memset(alignedData, 0, nElements * sizeof(T));
     }
   }
   /// Take ownership of an existing descriptor with a custom deleter.

mlir/unittests/ExecutionEngine/Invoke.cpp

@@ -251,6 +251,24 @@ TEST(NativeMemRefJit, SKIP_WITHOUT_JIT(BasicMemref)) {
   EXPECT_EQ((a[{2, 1}]), 42.);
 }
 
+TEST(NativeMemRefJit, SKIP_WITHOUT_JIT(OwningMemrefZeroInit)) {
+  constexpr int k = 3;
+  constexpr int m = 7;
+  int64_t shape[] = {k, m};
+  // Use a large alignment to stress the case where the memref data/basePtr are
+  // disjoint.
+  int alignment = 8192;
+  OwningMemRef<float, 2> a(shape, {}, {}, alignment);
+  ASSERT_EQ(
+      (void *)(((uintptr_t)a->basePtr + alignment - 1) & ~(alignment - 1)),
+      a->data);
+  for (int i = 0; i < k; ++i) {
+    for (int j = 0; j < m; ++j) {
+      EXPECT_EQ((a[{i, j}]), 0.);
+    }
+  }
+}
+
 // A helper function that will be called from the JIT
 static void memrefMultiply(::StridedMemRefType<float, 2> *memref,
                            int32_t coefficient) {