[flang] Replace !fir.ref<BoxType> -> BoxType conversion with a load prior to calling fir.dispatch

[Renaud-K]

Summary

Converting a !fir.ref<BoxType> -> BoxType is sort of ok because all boxes are bufferized during codegen and replaced with a memory reference. Then CodeGen of fir.dispatch does the load. But as we are working on moving polymorphic transformation into the fir-polymorphic-op pass (https://reviews.llvm.org/D144921), the load needs to be made explicit so the lowering of fir.dispatch can be done in FIR instead of LLVM dialect.
I thought it would be better to make this change into its own patch in case it introduces issues.

[PeteSteinfeld]

All builds and tests correctly, but I don't understand this code very well. Someone who does should approve.

[clementval]

As you pointed out this is ok to do the conversion but I think it is looks better to do the load. Thanks for the patch Renaud! LGTM

[clementval]

There is a failure on the pre commit CI but I doesn't look related to your change at all.

Driver/code-gen-rv64.f90

flang/lib/Lower/ConvertCall.cpp

if (std::optional<unsigned> passArg = caller.getPassArgIndex()) {
       // NOPASS
       const Fortran::evaluate::Component *component =
           caller.getCallDescription().proc().GetComponent();
       assert(component && "expect component for type-bound procedure call.");
       fir::ExtendedValue pass = converter.getSymbolExtendedValue(
           component->GetFirstSymbol(), &symMap);
       mlir::Value passObject = fir::getBase(pass);
       if (fir::isa_ref_type(passObject.getType()))
-        passObject = builder.create<fir::ConvertOp>(
-            loc, passObject.getType().dyn_cast<fir::ReferenceType>().getEleTy(),
-            passObject);
+        passObject = builder.create<fir::LoadOp>(loc, passObject);
       dispatch = builder.create<fir::DispatchOp>(
           loc, funcType.getResults(), builder.getStringAttr(procName),
           passObject, operands, nullptr);
     }
     callResult = dispatch.getResult(0);
     callNumResults = dispatch.getNumResults();
   } else {
     // Standard procedure call with fir.call.

flang/test/Lower/allocatable-polymorphic.f90

 ! CHECK: %[[TYPE_DESC_P1_CAST:.*]] = fir.convert %[[TYPE_DESC_P1]] : (!fir.tdesc<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) -> !fir.ref<none>
 ! CHECK: %[[RANK:.*]] = arith.constant 0 : i32
 ! CHECK: %[[CORANK:.*]] = arith.constant 0 : i32
 ! CHECK: %{{.*}} = fir.call @_FortranAPointerNullifyDerived(%[[P_DESC_CAST]], %[[TYPE_DESC_P1_CAST]], %[[RANK]], %[[CORANK]]) {{.*}}: (!fir.ref<!fir.box<none>>, !fir.ref<none>, i32, i32) -> none
 ! CHECK: %[[P_DESC_CAST:.*]] = fir.convert %[[P_DESC]] : (!fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.ref<!fir.box<none>>
 ! CHECK: %{{.*}} = fir.call @_FortranAPointerAllocate(%[[P_DESC_CAST]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}}: (!fir.ref<!fir.box<none>>, i1, !fir.box<none>, !fir.ref<i8>, i32) -> i32
 
 ! call p%proc1()
-! CHECK: %[[P_CAST:.*]] = fir.convert %[[P_DESC:.*]] : (!fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>
-! CHECK: fir.dispatch "proc1"(%[[P_CAST]] : !fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
+! CHECK: %[[P_LOAD:.*]] = fir.load %[[P_DESC:.*]] : !fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
+! CHECK: fir.dispatch "proc1"(%[[P_LOAD]] : !fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
 
 ! CHECK: %[[TYPE_DESC_P1:.*]] = fir.type_desc !fir.type<_QMpolyTp1{a:i32,b:i32}>
 ! CHECK: %[[C1_DESC_CAST:.*]] = fir.convert %[[C1_DESC:.*]] : (!fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.ref<!fir.box<none>>
 ! CHECK: %[[TYPE_DESC_P1_CAST:.*]] = fir.convert %[[TYPE_DESC_P1]] : (!fir.tdesc<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) -> !fir.ref<none>
 ! CHECK: %[[RANK:.*]] = arith.constant 0 : i32
 ! CHECK: %[[CORANK:.*]] = arith.constant 0 : i32
 ! CHECK: %{{.*}} = fir.call @_FortranAPointerNullifyDerived(%[[C1_DESC_CAST]], %[[TYPE_DESC_P1_CAST]], %[[RANK]], %[[CORANK]]) {{.*}}: (!fir.ref<!fir.box<none>>, !fir.ref<none>, i32, i32) -> none
 ! CHECK: %[[C1_DESC_CAST:.*]] = fir.convert %[[C1_DESC]] : (!fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.ref<!fir.box<none>>
 ! CHECK: %{{.*}} = fir.call @_FortranAPointerAllocate(%[[C1_DESC_CAST]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}}: (!fir.ref<!fir.box<none>>, i1, !fir.box<none>, !fir.ref<i8>, i32) -> i32
 
 ! CHECK: %[[TYPE_DESC_P2:.*]] = fir.type_desc !fir.type<_QMpolyTp2{a:i32,b:i32,c:i32}>
 ! CHECK: %[[C2_DESC_CAST:.*]] = fir.convert %[[C2_DESC]] : (!fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.ref<!fir.box<none>>
 ! CHECK: %[[TYPE_DESC_P2_CAST:.*]] = fir.convert %[[TYPE_DESC_P2]] : (!fir.tdesc<!fir.type<_QMpolyTp2{a:i32,b:i32,c:i32}>>) -> !fir.ref<none>
 ! CHECK: %[[RANK:.*]] = arith.constant 0 : i32
 ! CHECK: %[[CORANK:.*]] = arith.constant 0 : i32
 ! CHECK: %{{.*}} = fir.call @_FortranAPointerNullifyDerived(%[[C2_DESC_CAST]], %[[TYPE_DESC_P2_CAST]], %[[RANK]], %[[CORANK]]) {{.*}}: (!fir.ref<!fir.box<none>>, !fir.ref<none>, i32, i32) -> none
 ! CHECK: %[[C2_DESC_CAST:.*]] = fir.convert %[[C2_DESC]] : (!fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.ref<!fir.box<none>>
 ! CHECK: %{{.*}} = fir.call @_FortranAPointerAllocate(%[[C2_DESC_CAST]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}}: (!fir.ref<!fir.box<none>>, i1, !fir.box<none>, !fir.ref<i8>, i32) -> i32
 
 ! call c1%proc1()
-! CHECK: %[[C1_DESC_CAST:.*]] = fir.convert %[[C1_DESC]] : (!fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>
-! CHECK: fir.dispatch "proc1"(%[[C1_DESC_CAST]] : !fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
+! CHECK: %[[C1_DESC_LOAD:.*]] = fir.load %[[C1_DESC]] : !fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
+! CHECK: fir.dispatch "proc1"(%[[C1_DESC_LOAD]] : !fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
 
 ! call c2%proc1()
-! CHECK: %[[C2_DESC_CAST:.*]] = fir.convert %[[C2_DESC]] : (!fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>
-! CHECK: fir.dispatch "proc1"(%[[C2_DESC_CAST]] : !fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
+! CHECK: %[[C2_DESC_LOAD:.*]] = fir.load %[[C2_DESC]] : !fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
+! CHECK: fir.dispatch "proc1"(%[[C2_DESC_LOAD]] : !fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
 
 ! call c1%proc2()
 ! CHECK: %[[C1_LOAD:.*]] = fir.load %[[C1_DESC]] : !fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
 ! CHECK: %[[C1_REBOX:.*]] = fir.rebox %[[C1_LOAD]] : (!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>) -> !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>
 ! CHECK: fir.dispatch "proc2"(%[[C1_REBOX]] : !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) (%[[C1_REBOX]] : !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) {pass_arg_pos = 0 : i32}
 
 ! call c2%proc2()
 ! CHECK: %[[C2_LOAD:.*]] = fir.load %[[C2_DESC]] : !fir.ref<!fir.class<!fir.ptr<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>

 ! CHECK: %[[C0:.*]] = arith.constant 0 : i32
 ! CHECK: %[[C4_CAST:.*]] = fir.convert %[[C4]] : (!fir.ref<!fir.class<!fir.heap<!fir.array<?x!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>>) -> !fir.ref<!fir.box<none>>
 ! CHECK: %[[C1_I64:.*]] = fir.convert %[[CST1]] : (index) -> i64
 ! CHECK: %[[C20_I64:.*]] = fir.convert %[[C20]] : (i32) -> i64
 ! CHECK: %{{.*}} = fir.call @_FortranAAllocatableSetBounds(%[[C4_CAST]], %[[C0]], %[[C1_I64]], %[[C20_I64]]) {{.*}}: (!fir.ref<!fir.box<none>>, i32, i64, i64) -> none
 ! CHECK: %[[C4_CAST:.*]] = fir.convert %[[C4]] : (!fir.ref<!fir.class<!fir.heap<!fir.array<?x!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>>) -> !fir.ref<!fir.box<none>>
 ! CHECK: %{{.*}} = fir.call @_FortranAAllocatableAllocate(%[[C4_CAST]], %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}}: (!fir.ref<!fir.box<none>>, i1, !fir.box<none>, !fir.ref<i8>, i32) -> i32
 
-! CHECK: %[[C1_CAST:.*]] = fir.convert %[[C1_DESC]] : (!fir.ref<!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>
-! CHECK: fir.dispatch "proc1"(%[[C1_CAST]] : !fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
+! CHECK: %[[C1_LOAD1:.*]] = fir.load %[[C1_DESC]] : !fir.ref<!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
+! CHECK: fir.dispatch "proc1"(%[[C1_LOAD1]] : !fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
 
-! CHECK: %[[C2_CAST:.*]] = fir.convert %[[C2_DESC]] : (!fir.ref<!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>
-! CHECK: fir.dispatch "proc1"(%[[C2_CAST]] : !fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
+! CHECK: %[[C2_LOAD1:.*]] = fir.load %[[C2_DESC]] : !fir.ref<!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
+! CHECK: fir.dispatch "proc1"(%[[C2_LOAD1]] : !fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>)
 
-! CHECK: %[[C1_LOAD:.*]] = fir.load %[[C1_DESC]] : !fir.ref<!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
-! CHECK: %[[C1_REBOX:.*]] = fir.rebox %[[C1_LOAD]] : (!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>) -> !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>
+! CHECK: %[[C1_LOAD2:.*]] = fir.load %[[C1_DESC]] : !fir.ref<!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
+! CHECK: %[[C1_REBOX:.*]] = fir.rebox %[[C1_LOAD2]] : (!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>) -> !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>
 ! CHECK: fir.dispatch "proc2"(%[[C1_REBOX]] : !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) (%[[C1_REBOX]] : !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) {pass_arg_pos = 0 : i32}
 
-! CHECK: %[[C2_LOAD:.*]] = fir.load %[[C2_DESC]] : !fir.ref<!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
-! CHECK: %[[C2_REBOX:.*]] = fir.rebox %[[C2_LOAD]] : (!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>) -> !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>
+! CHECK: %[[C2_LOAD2:.*]] = fir.load %[[C2_DESC]] : !fir.ref<!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>
+! CHECK: %[[C2_REBOX:.*]] = fir.rebox %[[C2_LOAD2]] : (!fir.class<!fir.heap<!fir.type<_QMpolyTp1{a:i32,b:i32}>>>) -> !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>
 ! CHECK: fir.dispatch "proc2"(%[[C2_REBOX]] : !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) (%[[C2_REBOX]] : !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) {pass_arg_pos = 0 : i32}
 
 ! CHECK-LABEL: %{{.*}} = fir.do_loop
 ! CHECK: %[[C3_LOAD:.*]] = fir.load %[[C3_DESC]] : !fir.ref<!fir.class<!fir.heap<!fir.array<?x!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>>
 ! CHECK: %[[C3_COORD:.*]] = fir.coordinate_of %[[C3_LOAD]], %{{.*}} : (!fir.class<!fir.heap<!fir.array<?x!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>, i64) -> !fir.ref<!fir.type<_QMpolyTp1{a:i32,b:i32}>>
 ! CHECK: %[[C3_EMBOX:.*]] = fir.embox %[[C3_COORD]] source_box %[[C3_LOAD]] : (!fir.ref<!fir.type<_QMpolyTp1{a:i32,b:i32}>>, !fir.class<!fir.heap<!fir.array<?x!fir.type<_QMpolyTp1{a:i32,b:i32}>>>>) -> !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>
 ! CHECK: fir.dispatch "proc2"(%[[C3_EMBOX]] : !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) (%[[C3_EMBOX]] : !fir.class<!fir.type<_QMpolyTp1{a:i32,b:i32}>>) {pass_arg_pos = 0 : i32}