[Mlir-commits] [mlir] [memref] Simplify loads from reinterpret_cast of 1D contiguous memrefs (PR #188459)

[Andrzej Warzyński]

================
@@ -195,6 +196,237 @@ struct CopyToScalarLoadAndStore : public OpRewritePattern<memref::CopyOp> {
   }
 };
 
+static bool isConstZero(Value v) { return matchPattern(v, m_Zero()); }
+
+static bool isPureRankReshape(memref::ReinterpretCastOp rc, memref::LoadOp op) {
+  auto inputTy = cast<MemRefType>(rc.getSource().getType());
+  auto outputTy = cast<MemRefType>(rc.getResult().getType());
+
+  // This fold only handles reinterpret_casts that behave like pure rank
+  // reshapes of a single logical dimension:
+  //
+  //   - all metadata is static
+  //   - offset is 0
+  //   - source/result each have at most one non-unit dim
+  //   - if a non-unit dim exists, it is at the left or right boundary
+  //
+  // Examples accepted by this shape restriction:
+  //   memref<999xf32>       <-> memref<1x1x999xf32>
+  //   memref<1x108xf32>     <-> memref<1x1x1x108xf32>
+  //   memref<100x1xf32>     <-> memref<100x1x1xf32>
+  //
+  // General reinterpret_casts are intentionally rejected.
+
+  auto offsets = rc.getStaticOffsets();
+  assert(offsets.size() == 1 && "Expecting single offset");
+
+  // The rewrite drops the reinterpret_cast and remaps indices directly to the
+  // source memref. That is only correct if there is no storage shift.
----------------
banach-space wrote:

"The rewrite" -> which rewrite? You are making an assumption that this method will only be called by `FoldReinterpretCastLoad`, but we can't guarantee that.

> That is only correct if there is no storage shift.

Correct? Or are you trying to say that the rewrite only supports "index re-use" (as opposed to "index re-write/adjustment")

https://github.com/llvm/llvm-project/pull/188459