[flang][hlfir] relax the strictness of intrinsic verifiers

[tblah]

Summary

The verifiers for hlfir.matmul and hlfir.transpose try to ensure that
the shape of the result value makes sense given the shapes of the input
argument(s).

It there are some cases in the gfortran tests where lowering knows a bit
more about shape information than (HL)FIR. I think the cases here will be
solved when hlfir.shape_meet is implemented.

But in the meantime, and to improve robustness, I've relaxed the
verifier to allow the return type to have more precise shape information
than can be deduced from the argument type(s).

[vzakhari]

Thank you!

[kiranchandramohan]

flang/test/Lower/HLFIR/matmul.f90

23

Should we test the generated IR? If not, at least a check for the subroutine label?

[tblah]

Thanks for taking a look. The intention of that test was just to make sure the
verifier didn't return an error (which would be reflected in flang's exit code).

I have expanded it to check some of the generated IR.

[tblah]

Check label in transpose test

flang/lib/Optimizer/HLFIR/IR/HLFIROps.cpp

if (lhsRank == 2) {
     }
   } else {
     // lhsRank == 1
     // rhsRank == 2
     expectedResultShape.push_back(rhsShape[1]);
   }
   if (resultShape.size() != expectedResultShape.size())
     return emitOpError("incorrect result shape");
-  if (resultShape[0] != expectedResultShape[0])
+  if (resultShape[0] != expectedResultShape[0] &&
+      expectedResultShape[0] != unknownExtent)
     return emitOpError("incorrect result shape");
-  if (resultShape.size() == 2 && resultShape[1] != expectedResultShape[1])
+  if (resultShape.size() == 2 && resultShape[1] != expectedResultShape[1] &&
+      expectedResultShape[1] != unknownExtent)
     return emitOpError("incorrect result shape");
 
   return mlir::success();
 }
 
 mlir::LogicalResult
 hlfir::MatmulOp::canonicalize(MatmulOp matmulOp,
                               mlir::PatternRewriter &rewriter) {
mlir::LogicalResult hlfir::TransposeOp::verify() {
   hlfir::ExprType resultTy = getResult().getType().cast<hlfir::ExprType>();
   llvm::ArrayRef<int64_t> resultShape = resultTy.getShape();
   std::size_t resultRank = resultShape.size();
   mlir::Type resultEleTy = resultTy.getEleTy();
 
   if (rank != 2 || resultRank != 2)
     return emitOpError("input and output arrays should have rank 2");
 
-  if (inShape[0] != resultShape[1] || inShape[1] != resultShape[0])
+  constexpr int64_t unknownExtent = fir::SequenceType::getUnknownExtent();
+  if ((inShape[0] != resultShape[1]) && (inShape[0] != unknownExtent))
+    return emitOpError("output shape does not match input array");
+  if ((inShape[1] != resultShape[0]) && (inShape[1] != unknownExtent))
     return emitOpError("output shape does not match input array");
 
   if (eleTy != resultEleTy)
     return emitOpError(
         "input and output arrays should have the same element type");
 
   return mlir::success();
 }

flang/test/Lower/HLFIR/matmul.f90

 ! CHECK-DAG:     %[[LHS_VAR:.*]]:2 = hlfir.declare %[[LHS]]
 ! CHECK-DAG:     %[[RHS_VAR:.*]]:2 = hlfir.declare %[[RHS]]
 ! CHECK-DAG:     %[[RES_VAR:.*]]:2 = hlfir.declare %[[RES]]
 ! CHECK-NEXT:    %[[EXPR:.*]] = hlfir.matmul %[[LHS_VAR]]#0 %[[RHS_VAR]]#0 {fastmath = #arith.fastmath<contract>} : (!fir.box<!fir.array<?x?xi32>>, !fir.box<!fir.array<?x?xi32>>) -> !hlfir.expr<?x?xi32>
 ! CHECK-NEXT:    hlfir.assign %[[EXPR]] to %[[RES_VAR]]#0 : !hlfir.expr<?x?xi32>, !fir.box<!fir.array<?x?xi32>>
 ! CHECK-NEXT:    hlfir.destroy %[[EXPR]]
 ! CHECK-NEXT:    return
 ! CHECK-NEXT:   }
+
+! regression test for a case where the AST and FIR have different amounts of
+! shape inference
+subroutine matmul2(c)
+  integer, parameter :: N = 4
+  integer, dimension(:,:), allocatable :: a, b, c
+  integer, dimension(N,N) :: x
+
+  allocate(a(3*N, N), b(N, N), c(3*N, N))
+
+  call fill(a)
+  call fill(b)
+  call fill(x)
+
+  c = matmul(a, b - x)
+endsubroutine
+! CHECK-LABEL: func.func @_QPmatmul2
+! CHECK:           %[[C_ARG:.*]]: !fir.ref<!fir.box<!fir.heap<!fir.array<?x?xi32>>>>
+! CHECK:         %[[B_BOX_ALLOC:.*]] = fir.alloca !fir.box<!fir.heap<!fir.array<?x?xi32>>> {bindc_name = "b"
+! CHECK:         %[[B_BOX_DECL:.*]]:2 = hlfir.declare %[[B_BOX_ALLOC]] {{.*}} uniq_name = "_QFmatmul2Eb"
+
+
+! CHECK:         fir.call @_QPfill
+! CHECK:         fir.call @_QPfill
+! CHECK:         fir.call @_QPfill
+! CHECK-NEXT:    %[[B_BOX:.*]] = fir.load %[[B_BOX_DECL]]#0 : !fir.ref<!fir.box<!fir.heap<!fir.array<?x?xi32>>>>
+! CHECK-NEXT:    %[[C0:.*]] = arith.constant 0 : index
+! CHECK-NEXT:    %[[B_DIMS_0:.*]]:3 = fir.box_dims %[[B_BOX]], %[[C0]]
+! CHECK-NEXT:    %[[C1:.*]] = arith.constant 1 : index
+! CHECK-NEXT:    %[[B_DIMS_1:.*]]:3 = fir.box_dims %[[B_BOX]], %[[C1]]
+! CHECK-NEXT:    %[[B_SHAPE:.*]] = fir.shape %[[B_DIMS_0]]#1, %[[B_DIMS_1]]#1
+! CHECK-NEXT:    %[[ELEMENTAL:.*]] = hlfir.elemental %[[B_SHAPE]] : (!fir.shape<2>) -> !hlfir.expr<?x?xi32> {
+
+! CHECK:         }
+! CHECK-NEXT:    %[[A_BOX:.*]] = fir.load %{{.*}} : !fir.ref<!fir.box<!fir.heap<!fir.array<?x?xi32>>>>
+
+! The shapes in these types are what is being tested:
+! CHECK-NEXT:    %[[MATMUL:.*]] = hlfir.matmul %[[A_BOX]] %[[ELEMENTAL]] {{.*}} : (!fir.box<!fir.heap<!fir.array<?x?xi32>>>, !hlfir.expr<?x?xi32>) -> !hlfir.expr<?x4xi32>

flang/test/Lower/HLFIR/transpose.f90

 ! CHECK:           %[[RES_ARG:.*]]: !fir.ref<!fir.array<2x1xi32>>
 ! CHECK-DAG:     %[[ARG:.*]]:2 = hlfir.declare %[[M_ARG]](%[[M_SHAPE:.*]]) {[[NAME:.*]]} : (!fir.ref<!fir.array<1x2xi32>>, !fir.shape<2>) -> (!fir.ref<!fir.array<1x2xi32>>, !fir.ref<!fir.array<1x2xi32>>)
 ! CHECK-DAG:     %[[RES:.*]]:2 = hlfir.declare %[[RES_ARG]](%[[RES_SHAPE:.*]]) {[[NAME2:.*]]} : (!fir.ref<!fir.array<2x1xi32>>, !fir.shape<2>) -> (!fir.ref<!fir.array<2x1xi32>>, !fir.ref<!fir.array<2x1xi32>>)
 ! CHECK:         %[[EXPR:.*]] = hlfir.transpose %[[ARG]]#0 : (!fir.ref<!fir.array<1x2xi32>>) -> !hlfir.expr<2x1xi32>
 ! CHECK-NEXT:    hlfir.assign %[[EXPR]] to %[[RES]]#0
 ! CHECK-NEXT:    hlfir.destroy %[[EXPR]]
 ! CHECK-NEXT:    return
 ! CHECK-NEXT:  }
+
+! test the case where lowering has more exact information about the output
+! shape than is available from the argument
+subroutine transpose2(a, out)
+  real, allocatable, dimension(:) :: a
+  real, dimension(:,:) :: out
+  integer, parameter :: N = 3
+  integer, parameter :: M = 4
+
+  allocate(a(N*M))
+  out = transpose(reshape(a, (/N, M/)))
+end subroutine