[flang-commits] [flang] [flang] Canonicalize fir.array_coor by pulling in embox/rebox. (PR #92858)

[Slava Zakharin via flang-commits]

================
@@ -417,6 +423,216 @@ mlir::LogicalResult fir::ArrayCoorOp::verify() {
   return mlir::success();
 }
 
+// Pull in fir.embox and fir.rebox into fir.array_coor when possible.
+struct SimplifyArrayCoorOp : public mlir::OpRewritePattern<fir::ArrayCoorOp> {
+  using mlir::OpRewritePattern<fir::ArrayCoorOp>::OpRewritePattern;
+  mlir::LogicalResult
+  matchAndRewrite(fir::ArrayCoorOp op,
+                  mlir::PatternRewriter &rewriter) const override {
+    mlir::Value memref = op.getMemref();
+    if (!mlir::isa<fir::BaseBoxType>(memref.getType()))
+      return mlir::failure();
+
+    mlir::Value boxedMemref, boxedShape, boxedSlice;
+    if (auto emboxOp =
+            mlir::dyn_cast_or_null<fir::EmboxOp>(memref.getDefiningOp())) {
+      boxedMemref = emboxOp.getMemref();
+      boxedShape = emboxOp.getShape();
+      boxedSlice = emboxOp.getSlice();
+      // If any of operands, that are not currently supported for migration
+      // to ArrayCoorOp, is present, don't rewrite.
+      if (!emboxOp.getTypeparams().empty() || emboxOp.getSourceBox() ||
+          emboxOp.getAccessMap())
+        return mlir::failure();
+    } else if (auto reboxOp = mlir::dyn_cast_or_null<fir::ReboxOp>(
+                   memref.getDefiningOp())) {
+      boxedMemref = reboxOp.getBox();
+      boxedShape = reboxOp.getShape();
+      boxedSlice = reboxOp.getSlice();
+    } else {
+      return mlir::failure();
+    }
+
+    // Slices changing the number of dimensions are not supported
+    // for array_coor yet.
+    unsigned origBoxRank;
+    if (mlir::isa<fir::BaseBoxType>(boxedMemref.getType()))
+      origBoxRank = fir::getBoxRank(boxedMemref.getType());
+    else if (auto arrTy = mlir::dyn_cast<fir::SequenceType>(
+                 fir::unwrapRefType(boxedMemref.getType())))
+      origBoxRank = arrTy.getDimension();
+    else
+      return mlir::failure();
+
+    if (fir::getBoxRank(memref.getType()) != origBoxRank)
+      return mlir::failure();
+
+    // Slices with substring are not supported by array_coor.
+    if (boxedSlice)
+      if (auto sliceOp =
+              mlir::dyn_cast_or_null<fir::SliceOp>(boxedSlice.getDefiningOp()))
+        if (!sliceOp.getSubstr().empty())
+          return mlir::failure();
+
+    // If embox/rebox and array_coor have conflicting shapes or slices,
+    // do nothing.
+    if (op.getShape() && boxedShape && boxedShape != op.getShape())
+      return mlir::failure();
+    if (op.getSlice() && boxedSlice && boxedSlice != op.getSlice())
+      return mlir::failure();
+
+    // The embox/rebox and array_coor either have compatible
+    // shape/slice at this point or shape/slice is null
+    // in one of them but not in the other.
+    // The compatibility means they are equal or both null.
+    if (!op.getShape()) {
+      if (boxedShape) {
+        if (op.getSlice()) {
+          if (!boxedSlice) {
+            // %0 = fir.rebox %arg(%shape)
+            // %1 = fir.array_coor %0 [%slice] %idx
+            // Both the slice indices and %idx are 1-based, so the rebox
+            // may be pulled in as:
+            // %1 = fir.array_coor %arg [%slice] %idx
+            boxedShape = nullptr;
----------------
vzakhari wrote:

Thank you for the great examples again!  Please check if I got them right in the new version.

I am a bit confused by the examples, though.  I am trying to understand if the constants are just off or I am missing something.

Here (and in the second example, as well):
```
%1 = fir.shape_shift %c10, %c19, %c10, %c19 : (index, index, index, index) -> !fir.shapeshift<2>
%2 = fir.embox %arg0(%1) : (!fir.ref<!fir.array<?x?xi32>>, !fir.shapeshift<2>) -> !fir.box<!fir.array<?x?xi32>>
%s = fir.slice %c10, %c19, %c1, %c10, %c19, %c1 : (index, index, index, index, index, index) -> !fir.slice<2>
%3 = fir.array_coor %2 [%s] %c12, %c12 : (!fir.box<!fir.array<?x?xi32>>, !fir.slice<2>, index, index) -> !fir.ref<i32>
```
Since there is no shape in the array_coor, the slice and the indices should be 1-based, so the current indices in array_coor overrun the slice, aren't they?

https://github.com/llvm/llvm-project/pull/92858