[flang][cuda] Support shape shift in data transfer op. (#115929)

When an array is declared with a non default lower bound, the declare op
`getShape` will return a `ShapeShiftOp`. This result is used in data
transfer operation to compute the number of bytes to transfer. Update
the op to support `ShapeShiftOp`.

Author: clementval

flang/include/flang/Optimizer/Dialect/CUF/CUFOps.td

@@ -162,7 +162,7 @@ def cuf_DataTransferOp : cuf_Op<"data_transfer", []> {
 
   let arguments = (ins Arg<AnyType, "", [MemRead]>:$src,
                        Arg<AnyRefOrBoxType, "", [MemWrite]>:$dst,
-                       Optional<fir_ShapeType>:$shape,
+                       Optional<AnyShapeOrShiftType>:$shape,
                        cuf_DataTransferKindAttr:$transfer_kind);
 
   let assemblyFormat = [{

flang/lib/Optimizer/Dialect/CUF/CUFOps.cpp

@@ -102,7 +102,7 @@ llvm::LogicalResult cuf::DataTransferOp::verify() {
   mlir::Type srcTy = getSrc().getType();
   mlir::Type dstTy = getDst().getType();
   if (getShape()) {
-    if (!fir::isa_ref_type(srcTy) || !fir::isa_ref_type(dstTy))
+    if (!fir::isa_ref_type(srcTy) && !fir::isa_ref_type(dstTy))
       return emitOpError()
              << "shape can only be specified on data transfer with references";
   }

flang/lib/Optimizer/Transforms/CUFOpConversion.cpp

@@ -570,13 +570,21 @@ struct CUFDataTransferOpConversion
       mlir::Type i64Ty = builder.getI64Type();
       mlir::Value nbElement;
       if (op.getShape()) {
-        auto shapeOp =
-            mlir::dyn_cast<fir::ShapeOp>(op.getShape().getDefiningOp());
-        nbElement = rewriter.create<fir::ConvertOp>(loc, i64Ty,
-                                                    shapeOp.getExtents()[0]);
-        for (unsigned i = 1; i < shapeOp.getExtents().size(); ++i) {
-          auto operand = rewriter.create<fir::ConvertOp>(
-              loc, i64Ty, shapeOp.getExtents()[i]);
+        llvm::SmallVector<mlir::Value> extents;
+        if (auto shapeOp =
+                mlir::dyn_cast<fir::ShapeOp>(op.getShape().getDefiningOp())) {
+          extents = shapeOp.getExtents();
+        } else if (auto shapeShiftOp = mlir::dyn_cast<fir::ShapeShiftOp>(
+                       op.getShape().getDefiningOp())) {
+          for (auto i : llvm::enumerate(shapeShiftOp.getPairs()))
+            if (i.index() & 1)
+              extents.push_back(i.value());
+        }
+
+        nbElement = rewriter.create<fir::ConvertOp>(loc, i64Ty, extents[0]);
+        for (unsigned i = 1; i < extents.size(); ++i) {
+          auto operand =
+              rewriter.create<fir::ConvertOp>(loc, i64Ty, extents[i]);
           nbElement =
               rewriter.create<mlir::arith::MulIOp>(loc, nbElement, operand);
         }

flang/test/Fir/CUDA/cuda-data-transfer.fir

@@ -327,4 +327,39 @@ func.func @_QPtest_array_type() {
 // CHECK: %[[BYTES:.*]] = arith.muli %c10{{.*}}, %c12{{.*}} : i64
 // CHECK: fir.call @_FortranACUFDataTransferPtrPtr(%{{.*}}, %{{.*}}, %[[BYTES]], %c0{{.*}}, %{{.*}}, %{{.*}}) : (!fir.llvm_ptr<i8>, !fir.llvm_ptr<i8>, i64, i32, !fir.ref<i8>, i32) -> none
 
+func.func @_QPshape_shift() {
+  %c0_i32 = arith.constant 0 : i32
+  %c11 = arith.constant 11 : index
+  %c10 = arith.constant 10 : index
+  %0 = cuf.alloc !fir.array<10xi32> {bindc_name = "cdev", data_attr = #cuf.cuda<device>, uniq_name = "_QFshape_shiftEcdev"} -> !fir.ref<!fir.array<10xi32>>
+  %1 = fir.shape_shift %c11, %c10 : (index, index) -> !fir.shapeshift<1>
+  %2:2 = hlfir.declare %0(%1) {data_attr = #cuf.cuda<device>, uniq_name = "_QFshape_shiftEcdev"} : (!fir.ref<!fir.array<10xi32>>, !fir.shapeshift<1>) -> (!fir.box<!fir.array<10xi32>>, !fir.ref<!fir.array<10xi32>>)
+  cuf.data_transfer %c0_i32 to %2#1, %1 : !fir.shapeshift<1> {transfer_kind = #cuf.cuda_transfer<host_device>} : i32, !fir.ref<!fir.array<10xi32>>
+  cuf.free %2#1 : !fir.ref<!fir.array<10xi32>> {data_attr = #cuf.cuda<device>}
+  return
+}
+
+// CHECK-LABEL: func.func @_QPshape_shift()
+// CHECK: fir.call @_FortranACUFDataTransferDescDescNoRealloc
+
+func.func @_QPshape_shift2() {
+  %c11 = arith.constant 11 : index
+  %c10 = arith.constant 10 : index
+  %0 = fir.alloca !fir.array<10xi32> {bindc_name = "ahost", uniq_name = "_QFshape_shift2Eahost"}
+  %1 = fir.shape %c10 : (index) -> !fir.shape<1>
+  %2:2 = hlfir.declare %0(%1) {uniq_name = "_QFshape_shift2Eahost"} : (!fir.ref<!fir.array<10xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<10xi32>>, !fir.ref<!fir.array<10xi32>>)
+  %3 = cuf.alloc !fir.array<10xi32> {bindc_name = "cdev", data_attr = #cuf.cuda<device>, uniq_name = "_QFshape_shift2Ecdev"} -> !fir.ref<!fir.array<10xi32>>
+  %4 = fir.shape_shift %c11, %c10 : (index, index) -> !fir.shapeshift<1>
+  %5:2 = hlfir.declare %3(%4) {data_attr = #cuf.cuda<device>, uniq_name = "_QFshape_shift2Ecdev"} : (!fir.ref<!fir.array<10xi32>>, !fir.shapeshift<1>) -> (!fir.box<!fir.array<10xi32>>, !fir.ref<!fir.array<10xi32>>)
+  cuf.data_transfer %2#0 to %5#1, %4 : !fir.shapeshift<1> {transfer_kind = #cuf.cuda_transfer<host_device>} : !fir.ref<!fir.array<10xi32>>, !fir.ref<!fir.array<10xi32>>
+  cuf.free %5#1 : !fir.ref<!fir.array<10xi32>> {data_attr = #cuf.cuda<device>}
+  return
+}
+
+// CHECK-LABEL: func.func @_QPshape_shift2()
+// CHECK: %[[C10:.*]] = fir.convert %c10{{.*}} : (index) -> i64
+// CHECK: %[[BYTES:.*]] = arith.muli %[[C10]], %c4{{.*}} : i64
+// CHECK: fir.call @_FortranACUFDataTransferPtrPtr(%{{.*}}, %{{.*}}, %[[BYTES]], %{{.*}}, %{{.*}}, %{{.*}}) : (!fir.llvm_ptr<i8>, !fir.llvm_ptr<i8>, i64, i32, !fir.ref<i8>, i32) -> none
+
+
 } // end of module

flang/test/Lower/CUDA/cuda-data-transfer.cuf

@@ -7,6 +7,8 @@ module mod1
     integer :: i
   end type
 
+  integer, device, dimension(11:20) :: cdev
+
 contains
   function dev1(a)
     integer, device :: a(:)
@@ -16,6 +18,7 @@ contains
 end
 
 subroutine sub1()
+  use mod1
   integer, device :: m
   integer, device :: adev(10)
   integer :: i, ahost(10), bhost(10)
@@ -34,6 +37,8 @@ subroutine sub1()
 
   adev = 10
 
+  cdev = 0
+
 end
 
 ! CHECK-LABEL: func.func @_QPsub1()
@@ -70,6 +75,8 @@ end
 
 ! CHECK: cuf.data_transfer %c10{{.*}} to %[[ADEV]]#0 {transfer_kind = #cuf.cuda_transfer<host_device>} : i32, !fir.ref<!fir.array<10xi32>>
 
+! CHECK: cuf.data_transfer %c0{{.*}} to %{{.*}}#1, %{{.*}} : !fir.shapeshift<1> {transfer_kind = #cuf.cuda_transfer<host_device>} : i32, !fir.ref<!fir.array<10xi32>>
+
 subroutine sub2()
   integer, device :: m
   integer, device :: adev(10), bdev(10)
@@ -124,7 +131,7 @@ end
 ! CHECK: %[[TMP:.*]] = fir.alloca !fir.type<_QMmod1Tt1{i:i32}> {bindc_name = ".tmp"}
 ! CHECK: %[[AHOST:.*]]:2 = hlfir.declare %{{.*}}(%{{.*}}) {uniq_name = "_QFsub3Eahost"} : (!fir.ref<!fir.array<10xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<10xi32>>, !fir.ref<!fir.array<10xi32>>)
 ! CHECK: %[[BHOST:.*]]:2 = hlfir.declare %{{.*}}(%{{.*}}) {uniq_name = "_QFsub3Ebhost"} : (!fir.ref<!fir.array<10xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<10xi32>>, !fir.ref<!fir.array<10xi32>>)
-! CHECK: %[[T:.*]]:2 = hlfir.declare %7 {data_attr = #cuf.cuda<device>, uniq_name = "_QFsub3Et"} : (!fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>) -> (!fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>, !fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>)
+! CHECK: %[[T:.*]]:2 = hlfir.declare %{{.*}} {data_attr = #cuf.cuda<device>, uniq_name = "_QFsub3Et"} : (!fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>) -> (!fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>, !fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>)
 ! CHECK: %[[TMP_DECL:.*]]:2 = hlfir.declare %0 {uniq_name = ".tmp"} : (!fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>) -> (!fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>, !fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>)
 ! CHECK: cuf.data_transfer %[[T]]#1 to %[[TMP_DECL]]#0 {transfer_kind = #cuf.cuda_transfer<device_host>} : !fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>, !fir.ref<!fir.type<_QMmod1Tt1{i:i32}>>