[flang][openacc][openmp] Update stride computation for bounds

[clementval]

This patch updates the stride computation for the outer dimensions of multidimensional arrays where the stride is read from the descriptor.
For the inner dimension, the stride is the element size in bytes. Then it is multiplied by the n-1 extent for outer dimensions.

[llvmbot]

@llvm/pr-subscribers-openacc

@llvm/pr-subscribers-flang-fir-hlfir

Author: Valentin Clement (バレンタイン クレメン) (clementval)

Changes

This patch updates the stride computation for the outer dimensions of multidimensional arrays where the stride is read from the descriptor.
For the inner dimension, the stride is the element size in bytes. Then it is multiplied by the n-1 extent for outer dimensions.

---

Full diff: https://github.com/llvm/llvm-project/pull/72168.diff

2 Files Affected:

• (modified) flang/lib/Lower/DirectivesCommon.h (+7-3)
• (modified) flang/test/Lower/OpenACC/acc-bounds.f90 (+22)

diff --git a/flang/lib/Lower/DirectivesCommon.h b/flang/lib/Lower/DirectivesCommon.h
index f4903f607a2da8c..74f0c4b246b1cd1 100644
--- a/flang/lib/Lower/DirectivesCommon.h
+++ b/flang/lib/Lower/DirectivesCommon.h
@@ -597,6 +597,7 @@ genBoundsOpsFromBox(fir::FirOpBuilder &builder, mlir::Location loc,
   mlir::Value one = builder.createIntegerConstant(loc, idxTy, 1);
   assert(box.getType().isa<fir::BaseBoxType>() &&
          "expect fir.box or fir.class");
+  mlir::Value byteStride;
   for (unsigned dim = 0; dim < dataExv.rank(); ++dim) {
     mlir::Value d = builder.createIntegerConstant(loc, idxTy, dim);
     mlir::Value baseLb =
@@ -606,9 +607,12 @@ genBoundsOpsFromBox(fir::FirOpBuilder &builder, mlir::Location loc,
     mlir::Value lb = builder.createIntegerConstant(loc, idxTy, 0);
     mlir::Value ub =
         builder.create<mlir::arith::SubIOp>(loc, dimInfo.getExtent(), one);
-    mlir::Value bound =
-        builder.create<BoundsOp>(loc, boundTy, lb, ub, mlir::Value(),
-                                 dimInfo.getByteStride(), true, baseLb);
+    if (!byteStride) // First stride is the element size.
+      byteStride = dimInfo.getByteStride();
+    mlir::Value bound = builder.create<BoundsOp>(
+        loc, boundTy, lb, ub, mlir::Value(), byteStride, true, baseLb);
+    byteStride = builder.create<mlir::arith::MulIOp>(loc, byteStride,
+                                                     dimInfo.getExtent());
     bounds.push_back(bound);
   }
   return bounds;
diff --git a/flang/test/Lower/OpenACC/acc-bounds.f90 b/flang/test/Lower/OpenACC/acc-bounds.f90
index 519f9fa2bffd2a8..a1185292df22b55 100644
--- a/flang/test/Lower/OpenACC/acc-bounds.f90
+++ b/flang/test/Lower/OpenACC/acc-bounds.f90
@@ -102,4 +102,26 @@ subroutine acc_undefined_extent(a)
 ! HLFIR: %[[PRESENT:.*]] = acc.present varPtr(%[[DECL_ARG0]]#1 : !fir.ref<!fir.array<?xf32>>) bounds(%[[BOUND]]) -> !fir.ref<!fir.array<?xf32>> {name = "a"}
 ! CHECK: acc.kernels dataOperands(%[[PRESENT]] : !fir.ref<!fir.array<?xf32>>)
 
+  subroutine acc_multi_strides(a)
+    real, dimension(:,:,:) :: a
+
+    !$acc kernels present(a)
+    !$acc end kernels
+  end subroutine
+
+! CHECK-LABEL: func.func @_QMopenacc_boundsPacc_multi_strides(
+! CHECK-SAME: %[[ARG0:.*]]: !fir.box<!fir.array<?x?x?xf32>> {fir.bindc_name = "a"})
+! HLFIR: %[[DECL_ARG0:.*]]:2 = hlfir.declare %[[ARG0]] {uniq_name = "_QMopenacc_boundsFacc_multi_stridesEa"} : (!fir.box<!fir.array<?x?x?xf32>>) -> (!fir.box<!fir.array<?x?x?xf32>>, !fir.box<!fir.array<?x?x?xf32>>)
+! HLFIR: %[[BOX_DIMS0:.*]]:3 = fir.box_dims %[[DECL_ARG0]]#1, %c0{{.*}} : (!fir.box<!fir.array<?x?x?xf32>>, index) -> (index, index, index)
+! HLFIR: %[[BOUNDS0:.*]] = acc.bounds lowerbound(%{{.*}} : index) upperbound(%{{.*}} : index) stride(%[[BOX_DIMS0]]#2 : index) startIdx(%{{.*}} : index) {strideInBytes = true}
+! HLFIR: %[[STRIDE1:.*]] = arith.muli %27#2, %27#1 : index
+! HLFIR: %[[BOX_DIMS1:.*]]:3 = fir.box_dims %[[DECL_ARG0]]#1, %c1{{.*}} : (!fir.box<!fir.array<?x?x?xf32>>, index) -> (index, index, index)
+! HLFIR: %[[BOUNDS1:.*]] = acc.bounds lowerbound(%{{.*}} : index) upperbound(%{{.*}} : index) stride(%[[STRIDE1]] : index) startIdx(%{{.*}} : index) {strideInBytes = true}
+! HLFIR: %[[STRIDE2:.*]] = arith.muli %[[STRIDE1]], %[[BOX_DIMS1]]#1 : index
+! HLFIR: %[[BOX_DIMS2:.*]]:3 = fir.box_dims %[[DECL_ARG0]]#1, %c2{{.*}} : (!fir.box<!fir.array<?x?x?xf32>>, index) -> (index, index, index)
+! HLFIR: %[[BOUNDS2:.*]] = acc.bounds lowerbound(%{{.*}} : index) upperbound(%{{.*}} : index) stride(%[[STRIDE2]] : index) startIdx(%{{.*}} : index) {strideInBytes = true}
+! HLFIR: %[[BOX_ADDR:.*]] = fir.box_addr %[[DECL_ARG0]]#1 : (!fir.box<!fir.array<?x?x?xf32>>) -> !fir.ref<!fir.array<?x?x?xf32>>
+! HLFIR: %[[PRESENT:.*]] = acc.present varPtr(%[[BOX_ADDR]] : !fir.ref<!fir.array<?x?x?xf32>>) bounds(%29, %33, %37) -> !fir.ref<!fir.array<?x?x?xf32>> {name = "a"}
+! HLFIR: acc.kernels dataOperands(%[[PRESENT]] : !fir.ref<!fir.array<?x?x?xf32>>) {
+
 end module

[razvanlupusoru]

Should not use %27

[razvanlupusoru]

Looks great! Thank you for fixing this! Only a small concern about the test - but logic seems correct!