[flang][hlfir] Array constructor lowering [part 1/4]

[jeanPerier]

Summary

This is the first and biggest chunk that introduces support for
array constructor to HLFIR.

This patch:

• adds a new ConvertArrayConstructor.cpp that centralizes the code dealing with array constructor lowering.
• introduces a framework to lower array constructor according to different strategies: A common analysis of the array constructor is done, and based on that, a lowering startegy is selected and driven through the ac-values of the array constructor. See ConvertArrayConstructor.cpp comments for more details.
• implements the first strategy that creates a temporary inlined and updates it with inlined code. This strategy can only be used if the temporary can be pre-allocated (i.e: the extents and length parameters can be pre-computed without evaluating any ac-values), and if all the ac-value expressions are scalars.

For the sake of simplicity, characters and derived type will be enabled
once all the strategies are added.

[jeanPerier]

Add empty namespaces around implementation classes and structs as per
LLVM developer guidelines.

[tblah]

flang/lib/Lower/ConvertArrayConstructor.cpp

88

I thought the plan was to centralize all array temporary generation in the hlfir bufferization pass? I guess there is not a way to write "give me an uninitialised hlfir.expr which will be given a buffer later"?

It would be nice if expression buffer allocation could be done in one place so that, eventually, the stack arrays pass (and all its messy data flow analysis) can go away.

Another option would be to always allocate temporaries on the stack (no matter if -fstack-arrays or -Ofast were specified) and rely on the memory-allocation-opt pass to move large allocations to the heap, if that's desired. As I understand it, non-temporary array variables will already always be allocated on the stack (no matter their size).

Feel free to ignore this comment if this was a deliberate design decision.

[clementval]

LGTM

flang/lib/Lower/ConvertArrayConstructor.cpp

38

107

115

393–394

[PeteSteinfeld]

Aside from some nits and questions, all builds and tests correctly and looks good.

I didn't understand Valentin's review. He says "LGTM", but it looks like he didn't approve. You might want to double check with him.

flang/lib/Lower/ConvertArrayConstructor.cpp

49

What about the case where there's an ac-implied-do-control, and the expression invokes an intrinsic that yields a constant? Also, are you handling array constructors for arrays of derived types?

70–71

This should either be "lower an array constructor" or "lower array constructors"

148

"become" should be "becomes"

152

"reduce" should be "reduces"

298

"require" should be "requires"

311

"are used" should be "is used"

430

"try gather" should be "try to gather"

[clementval]

In D144102#4130401, @PeteSteinfeld wrote:

I didn't understand Valentin's review. He says "LGTM", but it looks like he didn't approve. You might want to double check with him.

Just forgot to select the option.

[jeanPerier]

flang/lib/Lower/ConvertArrayConstructor.cpp

49

What about the case where there's an ac-implied-do-control, and the expression invokes an intrinsic that yields a constant?

By "the expression", do you mean and ac-implied-do control bound ([i, i=1, abs(-42)]), or the expression inside the ac-implied-do ([(acos(42.), i=1,n)])?
In both cases, the front-end would anyway fold the result and we would not see it, but even if we did we would use the third strategy: lower it as an hlfir.elemental.

> Also, are you handling array constructors for arrays of derived types?

The current code is blocking derived type so that I can add tests for this case in its own patch (see TODO in LengthAndTypeCollector<Fortran::evaluate::SomeDerived> line 277). Derived type will use the runtime strategy I think, mainly to ensure we do not call user defined assignment for to copy them inside the temporary for now (other than that, there is no deep reason the other two strategies would not work for them too).

88

I thought the plan was to centralize all array temporary generation in the hlfir bufferization pass?

It is as much as possible. But in some case, it is hard to implement the Fortran concept without using memory.
The plan is to "pay the abstraction price" of removing the early temp creation where it makes an impact, but lowering may still resort to creating temps in some cases.

I guess there is not a way to write "give me an uninitialised hlfir.expr which will be given a buffer later"?

Indeed, I think doing so and using chain of inserts would have a quite complex code generation to ensure a new temporary is not created at each insert.

It would be nice if expression buffer allocation could be done in one place so that, eventually, the stack arrays pass (and all its messy data flow analysis) can go away.

Agree. I will look into it. I think I first need to propagate the option here, so I will do it in a separate patch.

To be fair, I am not saying it is impossible/undesired to have a powerful abstraction to represent all kinds of array constructors in HLFIR. I think it will need some careful design to cope with all cases and ensure it "plugs" well with the optimization passes. I am thinking we could have something like:

%expr = hlfir.array_ctor [%optional_shape] [%optional_length_parameters] {
  fir.do_loop {
    hlfir.push_value %acValue ...
  }
  hlfir.push_value %acValue ....
}

The obvious benefit is that if %expr is then assigned, we could assume that the left hand side (if it is not a whole allocatable assignment) has the right storage size, and use that info to create the temp if needed, or if aliasing analysis proves the LHS is not used inside the array_ctor region.
Same if we have Array + array_ctor_hard_to_predict_size, we could use conformability requirement to deduce the array constructor size.

I am not going that way now because I would like to better think the implication of such a new operation with region.

[jeanPerier]

Fix typos in comments. Thanks for the review.

[tblah]

flang/lib/Lower/ConvertArrayConstructor.cpp

88

Thanks for explaining. I think it is fine to come back to this at a later time

flang/include/flang/Lower/ConvertArrayConstructor.h

+//===-- ConvertArrayConstructor.h -- Array constructor lowering -*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Coding style: https://mlir.llvm.org/getting_started/DeveloperGuide/
+//
+//===----------------------------------------------------------------------===//
+///
+/// Implements the conversion from evaluate::ArrayConstructor to HLFIR.
+///
+//===----------------------------------------------------------------------===//
+#ifndef FORTRAN_LOWER_CONVERTARRAYCONSTRUCTOR_H
+#define FORTRAN_LOWER_CONVERTARRAYCONSTRUCTOR_H
+
+#include "flang/Evaluate/type.h"
+#include "flang/Optimizer/Builder/HLFIRTools.h"
+
+namespace Fortran::evaluate {
+template <typename T>
+class ArrayConstructor;
+}
+
+namespace Fortran::lower {
+class AbstractConverter;
+class SymMap;
+class StatementContext;
+
+/// Class to lower evaluate::ArrayConstructor<T> to hlfir::EntityWithAttributes.
+template <typename T>
+class ArrayConstructorBuilder {
+public:
+  static hlfir::EntityWithAttributes
+  gen(mlir::Location loc, Fortran::lower::AbstractConverter &converter,
+      const Fortran::evaluate::ArrayConstructor<T> &expr,
+      Fortran::lower::SymMap &symMap,
+      Fortran::lower::StatementContext &stmtCtx);
+};
+using namespace evaluate;
+FOR_EACH_SPECIFIC_TYPE(extern template class ArrayConstructorBuilder, )
+} // namespace Fortran::lower
+
+#endif // FORTRAN_LOWER_CONVERTARRAYCONSTRUCTOR_H

flang/lib/Lower/CMakeLists.txt

 get_property(dialect_libs GLOBAL PROPERTY MLIR_DIALECT_LIBS)
 
 add_flang_library(FortranLower
   Allocatable.cpp
   Bridge.cpp
   CallInterface.cpp
   Coarray.cpp
+  ConvertArrayConstructor.cpp
   ConvertCall.cpp
   ConvertConstant.cpp
   ConvertExpr.cpp
   ConvertExprToHLFIR.cpp
   ConvertProcedureDesignator.cpp
   ConvertType.cpp
   ConvertVariable.cpp
   ComponentPath.cpp

flang/lib/Lower/ConvertArrayConstructor.cpp

+//===- ConvertArrayConstructor.cpp -- Array Constructor ---------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "flang/Lower/ConvertArrayConstructor.h"
+#include "flang/Evaluate/expression.h"
+#include "flang/Lower/AbstractConverter.h"
+#include "flang/Lower/ConvertExprToHLFIR.h"
+#include "flang/Lower/ConvertType.h"
+#include "flang/Lower/StatementContext.h"
+#include "flang/Lower/SymbolMap.h"
+#include "flang/Optimizer/Builder/HLFIRTools.h"
+#include "flang/Optimizer/Builder/Todo.h"
+#include "flang/Optimizer/HLFIR/HLFIROps.h"
+
+// Array constructors are lowered with three different strategies.
+// All strategies are not possible with all array constructors.
+//
+// - Strategy 1: runtime approach (RuntimeTempStrategy).
+//   This strategy works will all array constructors, but will create more
+//   complex code that is harder to optimize. An allocatable temp is created,
+//   it may be unallocated if the array constructor length parameters or extent
+//   could not be computed. Then, the runtime is called to push lowered
+//   ac-value (array constructor elements) into the allocatable. The runtime
+//   will allocate or reallocate as needed while values are being pushed.
+//   In the end, the allocatable contain a temporary with all the array
+//   constructor evaluated elements.
+//
+// - Strategy 2: inlined temporary approach (InlinedTempStrategyImpl)
+//   This strategy can only be used if the array constructor extent and length
+//   parameters can be pre-computed without evaluating any ac-value, and if all
+//   of the ac-value are scalars (at least for now).
+//   A temporary is allocated inline in one go, and an index pointing at the
+//   current ac-value position in the array constructor element sequence is
+//   maintained and used to store ac-value as they are being lowered.
+//
+// - Strategy 3: "function of the indices" approach (AsElementalStrategy)
+//   This strategy can only be used if the array constructor extent and length
+//   parameters can be pre-computed and, if the array constructor is of the
+//   form "[(scalar_expr, ac-implied-do-control)]". In this case, it is lowered
+//   into an hlfir.elemental without creating any temporary in lowering. This
+//   form should maximize the chance of array temporary elision when assigning
+//   the array constructor, potentially reshaped, to an array variable.
+//
+//   The array constructor lowering looks like:
+//   ```
+//     strategy = selectArrayCtorLoweringStrategy(array-ctor-expr);
+//     for (ac-value : array-ctor-expr)
+//       if (ac-value is expression) {
+//         strategy.pushValue(ac-value);
+//       } else if (ac-value is implied-do) {
+//         strategy.startImpliedDo(lower, upper, stride);
+//         // lower nested values
+//       }
+//     result = strategy.finishArrayCtorLowering();
+//   ```
+
+//===----------------------------------------------------------------------===//
+//   Definition of the lowering strategies. Each lowering strategy is defined
+//   as a class that implements "pushValue", "startImpliedDo", and
+//   "finishArrayCtorLowering".
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// Class that implements the "inlined temp strategy" to lower array
+/// constructors. It must be further provided a CounterType class to specify how
+/// the current ac-value insertion position is tracked.
+template <typename CounterType>
+class InlinedTempStrategyImpl {
+  /// Name that will be given to the temporary allocation and hlfir.declare in
+  /// the IR.
+  static constexpr char tempName[] = ".tmp.arrayctor";
+
+public:
+  /// Start lowering an array constructor according to the inline strategy.
+  /// The temporary is created right away.
+  InlinedTempStrategyImpl(mlir::Location loc, fir::FirOpBuilder &builder,
+                          fir::SequenceType declaredType, mlir::Value extent,
+                          llvm::ArrayRef<mlir::Value> lengths)
+      : one{builder.createIntegerConstant(loc, builder.getIndexType(), 1)},
+        counter{loc, builder, one} {
+    // Allocate the temporary storage.
+    llvm::SmallVector<mlir::Value, 1> extents{extent};
+    mlir::Value tempStorage = builder.createHeapTemporary(
+        loc, declaredType, tempName, extents, lengths);
+    mlir::Value shape = builder.genShape(loc, extents);
+    temp =
+        builder
+            .create<hlfir::DeclareOp>(loc, tempStorage, tempName, shape,
+                                      lengths, fir::FortranVariableFlagsAttr{})
+            .getBase();
+  }
+
+  /// Push a lowered ac-value into the current insertion point and
+  /// increment the insertion point.
+  void pushValue(mlir::Location loc, fir::FirOpBuilder &builder,
+                 hlfir::Entity value) {
+    assert(value.isScalar() && "cannot use inlined temp with array values");
+    mlir::Value indexValue = counter.getAndIncrementIndex(loc, builder, one);
+    hlfir::Entity tempElement = hlfir::getElementAt(
+        loc, builder, hlfir::Entity{temp}, mlir::ValueRange{indexValue});
+    // TODO: "copy" would probably be better than assign to ensure there are no
+    // side effects (user assignments, temp, lhs finalization)?
+    // This only makes a difference for derived types, so for now derived types
+    // will use the runtime strategy to avoid any bad behaviors.
+    builder.create<hlfir::AssignOp>(loc, value, tempElement);
+  }
+
+  /// Start a fir.do_loop with the control from an implied-do and return
+  /// the loop induction variable that is the ac-do-variable value.
+  /// Only usable if the counter is able to track the position through loops.
+  mlir::Value startImpliedDo(mlir::Location loc, fir::FirOpBuilder &builder,
+                             mlir::Value lower, mlir::Value upper,
+                             mlir::Value stride) {
+    if constexpr (!CounterType::canCountThroughLoops)
+      fir::emitFatalError(loc, "array constructor lowering is inconsistent");
+    auto loop = builder.create<fir::DoLoopOp>(loc, lower, upper, stride,
+                                              /*unordered=*/false,
+                                              /*finalCount=*/false);
+    builder.setInsertionPointToStart(loop.getBody());
+    return loop.getInductionVar();
+  }
+
+  /// Move the temporary to an hlfir.expr value (array constructors are not
+  /// variables and cannot be further modified).
+  hlfir::Entity finishArrayCtorLowering(mlir::Location loc,
+                                        fir::FirOpBuilder &builder) {
+    // Temp is created using createHeapTemporary.
+    mlir::Value mustFree = builder.createBool(loc, true);
+    auto hlfirExpr = builder.create<hlfir::AsExprOp>(loc, temp, mustFree);
+    return hlfir::Entity{hlfirExpr};
+  }
+
+private:
+  mlir::Value one;
+  CounterType counter;
+  mlir::Value temp;
+};
+
+/// A simple SSA value counter to lower array constructors without any
+/// implied-do in the "inlined temp strategy".
+/// The SSA value being tracked by the counter (hence, this
+/// cannot count through loops since the SSA value in the loop becomes
+/// inaccessible after the loop).
+/// Semantic analysis expression rewrites unroll implied do loop with
+/// compile time constant bounds (even if huge). So this minimalistic
+/// counter greatly reduces the generated IR for simple but big array
+/// constructors [(i,i=1,constant-expr)] that are expected to be quite
+/// common.
+class ValueCounter {
+public:
+  static constexpr bool canCountThroughLoops = false;
+  ValueCounter(mlir::Location loc, fir::FirOpBuilder &builder,
+               mlir::Value initialValue) {
+    indexValue = initialValue;
+  }
+
+  mlir::Value getAndIncrementIndex(mlir::Location loc,
+                                   fir::FirOpBuilder &builder,
+                                   mlir::Value increment) {
+    mlir::Value currentValue = indexValue;
+    indexValue =
+        builder.create<mlir::arith::AddIOp>(loc, indexValue, increment);
+    return currentValue;
+  }
+
+private:
+  mlir::Value indexValue;
+};
+using LooplessInlinedTempStrategy = InlinedTempStrategyImpl<ValueCounter>;
+
+/// A generic memory based counter that can deal with all cases of
+/// "inlined temp strategy". The counter value is stored in a temp
+/// from which it is loaded, incremented, and stored every time an
+/// ac-value is pushed.
+class InMemoryCounter {
+public:
+  static constexpr bool canCountThroughLoops = true;
+  InMemoryCounter(mlir::Location loc, fir::FirOpBuilder &builder,
+                  mlir::Value initialValue) {
+    indexVar = builder.createTemporary(loc, initialValue.getType());
+    builder.create<fir::StoreOp>(loc, initialValue, indexVar);
+  }
+
+  mlir::Value getAndIncrementIndex(mlir::Location loc,
+                                   fir::FirOpBuilder &builder,
+                                   mlir::Value increment) const {
+    mlir::Value indexValue = builder.create<fir::LoadOp>(loc, indexVar);
+    indexValue =
+        builder.create<mlir::arith::AddIOp>(loc, indexValue, increment);
+    builder.create<fir::StoreOp>(loc, indexValue, indexVar);
+    return indexValue;
+  }
+
+private:
+  mlir::Value indexVar;
+};
+using InlinedTempStrategy = InlinedTempStrategyImpl<InMemoryCounter>;
+
+// TODO: add and implement AsElementalStrategy.
+
+// TODO: add and implement RuntimeTempStrategy.
+
+/// Wrapper class that dispatch to the selected array constructor lowering
+/// strategy and does nothing else.
+class ArrayCtorLoweringStrategy {
+public:
+  template <typename A>
+  ArrayCtorLoweringStrategy(A &&impl) : implVariant{std::forward<A>(impl)} {}
+
+  void pushValue(mlir::Location loc, fir::FirOpBuilder &builder,
+                 hlfir::Entity value) {
+    return std::visit(
+        [&](auto &impl) { return impl.pushValue(loc, builder, value); },
+        implVariant);
+  }
+
+  mlir::Value startImpliedDo(mlir::Location loc, fir::FirOpBuilder &builder,
+                             mlir::Value lower, mlir::Value upper,
+                             mlir::Value stride) {
+    return std::visit(
+        [&](auto &impl) {
+          return impl.startImpliedDo(loc, builder, lower, upper, stride);
+        },
+        implVariant);
+  }
+
+  hlfir::Entity finishArrayCtorLowering(mlir::Location loc,
+                                        fir::FirOpBuilder &builder) {
+    return std::visit(
+        [&](auto &impl) { return impl.finishArrayCtorLowering(loc, builder); },
+        implVariant);
+  }
+
+private:
+  std::variant<InlinedTempStrategy, LooplessInlinedTempStrategy> implVariant;
+};
+} // namespace
+
+//===----------------------------------------------------------------------===//
+//   Definition of selectArrayCtorLoweringStrategy and its helpers.
+//   This is the code that analyses the evaluate::ArrayConstructor<T>,
+//   pre-lowers the array constructor extent and length parameters if it can,
+//   and chooses the lowering strategy.
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// Helper class to lower the array constructor type and its length parameters.
+/// The length parameters, if any, are only lowered if this does not require
+/// evaluating an ac-value.
+template <typename T>
+struct LengthAndTypeCollector {
+  static mlir::Type collect(mlir::Location,
+                            Fortran::lower::AbstractConverter &converter,
+                            const Fortran::evaluate::ArrayConstructor<T> &,
+                            Fortran::lower::SymMap &,
+                            Fortran::lower::StatementContext &,
+                            mlir::SmallVectorImpl<mlir::Value> &) {
+    // Numerical and Logical types.
+    return Fortran::lower::getFIRType(&converter.getMLIRContext(), T::category,
+                                      T::kind, /*lenParams*/ {});
+  }
+};
+
+template <>
+struct LengthAndTypeCollector<Fortran::evaluate::SomeDerived> {
+  static mlir::Type collect(
+      mlir::Location loc, Fortran::lower::AbstractConverter &converter,
+      const Fortran::evaluate::ArrayConstructor<Fortran::evaluate::SomeDerived>
+          &arrayCtorExpr,
+      Fortran::lower::SymMap &symMap, Fortran::lower::StatementContext &stmtCtx,
+      mlir::SmallVectorImpl<mlir::Value> &lengths) {
+    TODO(loc, "collect derived type and length");
+  }
+};
+
+template <int Kind>
+using Character =
+    Fortran::evaluate::Type<Fortran::common::TypeCategory::Character, Kind>;
+template <int Kind>
+struct LengthAndTypeCollector<Character<Kind>> {
+  static mlir::Type collect(
+      mlir::Location loc, Fortran::lower::AbstractConverter &converter,
+      const Fortran::evaluate::ArrayConstructor<Character<Kind>> &arrayCtorExpr,
+      Fortran::lower::SymMap &symMap, Fortran::lower::StatementContext &stmtCtx,
+      mlir::SmallVectorImpl<mlir::Value> &lengths) {
+    TODO(loc, "collect character type and length");
+  }
+};
+} // namespace
+
+/// Does the array constructor have length parameters that
+/// LengthAndTypeCollector::collect could not lower because this requires
+/// lowering an ac-value and must be delayed?
+static bool
+failedToGatherLengthParameters(mlir::Type elementType,
+                               llvm::ArrayRef<mlir::Value> lengths) {
+  return (elementType.isa<fir::CharacterType>() ||
+          fir::isRecordWithTypeParameters(elementType)) &&
+         lengths.empty();
+}
+
+namespace {
+/// Structure that analyses the ac-value and implied-do of
+/// evaluate::ArrayConstructor before they are lowered. It does not generate any
+/// IR. The result of this analysis pass is used to select the lowering
+/// strategy.
+struct ArrayCtorAnalysis {
+  template <typename T>
+  ArrayCtorAnalysis(
+      const Fortran::evaluate::ArrayConstructor<T> &arrayCtorExpr);
+
+  // Can the array constructor easily be rewritten into an hlfir.elemental ?
+  bool isSingleImpliedDoWithOneScalarExpr() const {
+    return !anyArrayExpr && isPerfectLoopNest &&
+           innerNumberOfExprIfPrefectNest == 1 && depthIfPerfectLoopNest == 1;
+  }
+
+  bool anyImpliedDo{false};
+  bool anyArrayExpr{false};
+  bool isPerfectLoopNest{true};
+  std::int64_t innerNumberOfExprIfPrefectNest = 0;
+  std::int64_t depthIfPerfectLoopNest = 0;
+};
+} // namespace
+
+template <typename T>
+ArrayCtorAnalysis::ArrayCtorAnalysis(
+    const Fortran::evaluate::ArrayConstructor<T> &arrayCtorExpr) {
+  llvm::SmallVector<const Fortran::evaluate::ArrayConstructorValues<T> *>
+      arrayValueListStack{&arrayCtorExpr};
+  // Loop through the ac-value-list(s) of the array constructor.
+  while (!arrayValueListStack.empty()) {
+    std::int64_t localNumberOfImpliedDo = 0;
+    std::int64_t localNumberOfExpr = 0;
+    // Loop though the ac-value of an ac-value list, and add any nested
+    // ac-value-list of ac-implied-do to the stack.
+    for (const Fortran::evaluate::ArrayConstructorValue<T> &acValue :
+         *arrayValueListStack.pop_back_val())
+      std::visit(Fortran::common::visitors{
+                     [&](const Fortran::evaluate::ImpliedDo<T> &impledDo) {
+                       arrayValueListStack.push_back(&impledDo.values());
+                       localNumberOfImpliedDo++;
+                     },
+                     [&](const Fortran::evaluate::Expr<T> &expr) {
+                       localNumberOfExpr++;
+                       anyArrayExpr = anyArrayExpr || expr.Rank() > 0;
+                     }},
+                 acValue.u);
+    anyImpliedDo = anyImpliedDo || localNumberOfImpliedDo > 0;
+
+    if (localNumberOfImpliedDo == 0) {
+      // Leaf ac-value-list in the array constructor ac-value tree.
+      if (isPerfectLoopNest)
+        // This this the only leaf of the array-constructor (the array
+        // constructor is a nest of single implied-do with a list of expression
+        // in the last deeper implied do). e.g: "[((i+j, i=1,n)j=1,m)]".
+        innerNumberOfExprIfPrefectNest = localNumberOfExpr;
+    } else if (localNumberOfImpliedDo == 1 && localNumberOfExpr == 0) {
+      // Perfect implied-do nest new level.
+      ++depthIfPerfectLoopNest;
+    } else {
+      // More than one implied-do, or at least one implied-do and an expr
+      // at that level. This will not form a perfect nest. Examples:
+      // "[a, (i, i=1,n)]" or "[(i, i=1,n), (j, j=1,m)]".
+      isPerfectLoopNest = false;
+    }
+  }
+}
+
+/// Helper to lower a scalar extent expression (like implied-do bounds).
+static mlir::Value lowerExtentExpr(mlir::Location loc,
+                                   Fortran::lower::AbstractConverter &converter,
+                                   Fortran::lower::SymMap &symMap,
+                                   Fortran::lower::StatementContext &stmtCtx,
+                                   const Fortran::evaluate::ExtentExpr &expr) {
+  fir::FirOpBuilder &builder = converter.getFirOpBuilder();
+  mlir::IndexType idxTy = builder.getIndexType();
+  hlfir::Entity value = Fortran::lower::convertExprToHLFIR(
+      loc, converter, toEvExpr(expr), symMap, stmtCtx);
+  value = hlfir::loadTrivialScalar(loc, builder, value);
+  return builder.createConvert(loc, idxTy, value);
+}
+
+/// Does \p expr contain no calls to user function?
+static bool isCallFreeExpr(const Fortran::evaluate::ExtentExpr &expr) {
+  for (const Fortran::semantics::Symbol &symbol :
+       Fortran::evaluate::CollectSymbols(expr))
+    if (Fortran::semantics::IsProcedure(symbol))
+      return false;
+  return true;
+}
+
+/// Core function that pre-lowers the extent and length parameters of
+/// array constructors if it can, runs the ac-value analysis and
+/// select the lowering strategy accordingly.
+template <typename T>
+static ArrayCtorLoweringStrategy selectArrayCtorLoweringStrategy(
+    mlir::Location loc, Fortran::lower::AbstractConverter &converter,
+    const Fortran::evaluate::ArrayConstructor<T> &arrayCtorExpr,
+    Fortran::lower::SymMap &symMap, Fortran::lower::StatementContext &stmtCtx) {
+  fir::FirOpBuilder &builder = converter.getFirOpBuilder();
+  mlir::Type idxType = builder.getIndexType();
+  // Try to gather the array constructor extent.
+  mlir::Value extent;
+  fir::SequenceType::Extent typeExtent = fir::SequenceType::getUnknownExtent();
+  auto shapeExpr =
+      Fortran::evaluate::GetShape(converter.getFoldingContext(), arrayCtorExpr);
+  if (shapeExpr && shapeExpr->size() == 1 && (*shapeExpr)[0]) {
+    const Fortran::evaluate::ExtentExpr &extentExpr = *(*shapeExpr)[0];
+    if (auto constantExtent = Fortran::evaluate::ToInt64(extentExpr)) {
+      typeExtent = *constantExtent;
+      extent = builder.createIntegerConstant(loc, idxType, typeExtent);
+    } else if (isCallFreeExpr(extentExpr)) {
+      // The expression built by expression analysis for the array constructor
+      // extent does not contain procedure symbols. It is side effect free.
+      // This could be relaxed to allow pure procedure, but some care must
+      // be taken to not bring in "unmapped" symbols from callee scopes.
+      extent = lowerExtentExpr(loc, converter, symMap, stmtCtx, extentExpr);
+    }
+    // Otherwise, the temporary will have to be built step by step with
+    // reallocation and the extent will only be known at the end of the array
+    // constructor evaluation.
+  }
+  // Convert the array constructor type and try to gather its length parameter
+  // values, if any.
+  mlir::SmallVector<mlir::Value> lengths;
+  mlir::Type elementType = LengthAndTypeCollector<T>::collect(
+      loc, converter, arrayCtorExpr, symMap, stmtCtx, lengths);
+  // Run an analysis of the array constructor ac-value.
+  ArrayCtorAnalysis analysis(arrayCtorExpr);
+  bool needToEvaluateOneExprToGetLengthParameters =
+      failedToGatherLengthParameters(elementType, lengths);
+
+  // Based on what was gathered and the result of the analysis, select and
+  // instantiate the right lowering strategy for the array constructor.
+  if (!extent || needToEvaluateOneExprToGetLengthParameters ||
+      analysis.anyArrayExpr)
+    TODO(loc, "Lowering of array constructor requiring the runtime");
+
+  auto declaredType = fir::SequenceType::get({typeExtent}, elementType);
+  if (analysis.isSingleImpliedDoWithOneScalarExpr())
+    TODO(loc, "Lowering of array constructor as hlfir.elemental");
+
+  if (analysis.anyImpliedDo)
+    return InlinedTempStrategy(loc, builder, declaredType, extent, lengths);
+
+  return LooplessInlinedTempStrategy(loc, builder, declaredType, extent,
+                                     lengths);
+}
+
+/// Lower an ac-value expression \p expr and forward it to the selected
+/// lowering strategy \p arrayBuilder,
+template <typename T>
+static void genAcValue(mlir::Location loc,
+                       Fortran::lower::AbstractConverter &converter,
+                       const Fortran::evaluate::Expr<T> &expr,
+                       Fortran::lower::SymMap &symMap,
+                       Fortran::lower::StatementContext &stmtCtx,
+                       ArrayCtorLoweringStrategy &arrayBuilder) {
+  if (expr.Rank() != 0)
+    TODO(loc, "array constructor with array ac-value in HLFIR");
+  // TODO: get rid of the toEvExpr indirection.
+  fir::FirOpBuilder &builder = converter.getFirOpBuilder();
+  hlfir::Entity value = Fortran::lower::convertExprToHLFIR(
+      loc, converter, toEvExpr(expr), symMap, stmtCtx);
+  value = hlfir::loadTrivialScalar(loc, builder, value);
+  arrayBuilder.pushValue(loc, builder, value);
+}
+
+/// Lowers an ac-value implied-do \p impledDo according to the selected
+/// lowering strategy \p arrayBuilder.
+template <typename T>
+static void genAcValue(mlir::Location loc,
+                       Fortran::lower::AbstractConverter &converter,
+                       const Fortran::evaluate::ImpliedDo<T> &impledDo,
+                       Fortran::lower::SymMap &symMap,
+                       Fortran::lower::StatementContext &stmtCtx,
+                       ArrayCtorLoweringStrategy &arrayBuilder) {
+  auto lowerIndex =
+      [&](const Fortran::evaluate::ExtentExpr expr) -> mlir::Value {
+    return lowerExtentExpr(loc, converter, symMap, stmtCtx, expr);
+  };
+  mlir::Value lower = lowerIndex(impledDo.lower());
+  mlir::Value upper = lowerIndex(impledDo.upper());
+  mlir::Value stride = lowerIndex(impledDo.stride());
+  fir::FirOpBuilder &builder = converter.getFirOpBuilder();
+  mlir::OpBuilder::InsertPoint insertPt = builder.saveInsertionPoint();
+  mlir::Value impliedDoIndexValue =
+      arrayBuilder.startImpliedDo(loc, builder, lower, upper, stride);
+  symMap.pushImpliedDoBinding(toStringRef(impledDo.name()),
+                              impliedDoIndexValue);
+  stmtCtx.pushScope();
+
+  for (const auto &acValue : impledDo.values())
+    std::visit(
+        [&](const auto &x) {
+          genAcValue(loc, converter, x, symMap, stmtCtx, arrayBuilder);
+        },
+        acValue.u);
+
+  stmtCtx.finalizeAndPop();
+  symMap.popImpliedDoBinding();
+  builder.restoreInsertionPoint(insertPt);
+}
+
+/// Entry point for evaluate::ArrayConstructor lowering.
+template <typename T>
+hlfir::EntityWithAttributes Fortran::lower::ArrayConstructorBuilder<T>::gen(
+    mlir::Location loc, Fortran::lower::AbstractConverter &converter,
+    const Fortran::evaluate::ArrayConstructor<T> &arrayCtorExpr,
+    Fortran::lower::SymMap &symMap, Fortran::lower::StatementContext &stmtCtx) {
+  fir::FirOpBuilder &builder = converter.getFirOpBuilder();
+  // Select the lowering strategy given the array constructor.
+  auto arrayBuilder = selectArrayCtorLoweringStrategy(
+      loc, converter, arrayCtorExpr, symMap, stmtCtx);
+  // Run the array lowering strategy through the ac-values.
+  for (const auto &acValue : arrayCtorExpr)
+    std::visit(
+        [&](const auto &x) {
+          genAcValue(loc, converter, x, symMap, stmtCtx, arrayBuilder);
+        },
+        acValue.u);
+  hlfir::Entity hlfirExpr = arrayBuilder.finishArrayCtorLowering(loc, builder);
+  // Insert the clean-up for the created hlfir.expr.
+  fir::FirOpBuilder *bldr = &builder;
+  stmtCtx.attachCleanup(
+      [=]() { bldr->create<hlfir::DestroyOp>(loc, hlfirExpr); });
+  return hlfir::EntityWithAttributes{hlfirExpr};
+}
+
+using namespace Fortran::evaluate;
+using namespace Fortran::common;
+FOR_EACH_SPECIFIC_TYPE(template class Fortran::lower::ArrayConstructorBuilder, )

flang/lib/Lower/ConvertExprToHLFIR.cpp

 //===-- ConvertExprToHLFIR.cpp --------------------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // Coding style: https://mlir.llvm.org/getting_started/DeveloperGuide/
 //
 //===----------------------------------------------------------------------===//
 
 #include "flang/Lower/ConvertExprToHLFIR.h"
 #include "flang/Evaluate/shape.h"
 #include "flang/Lower/AbstractConverter.h"
 #include "flang/Lower/CallInterface.h"
+#include "flang/Lower/ConvertArrayConstructor.h"
 #include "flang/Lower/ConvertCall.h"
 #include "flang/Lower/ConvertConstant.h"
 #include "flang/Lower/ConvertProcedureDesignator.h"
 #include "flang/Lower/ConvertType.h"
 #include "flang/Lower/ConvertVariable.h"
 #include "flang/Lower/StatementContext.h"
 #include "flang/Lower/SymbolMap.h"
 #include "flang/Optimizer/Builder/Complex.h"
if (auto addressOf = fir::getBase(exv).getDefiningOp<fir::AddrOfOp>()) {
           loc, builder, exv,
           addressOf.getSymbol().getRootReference().getValue(), flags);
     }
     fir::emitFatalError(loc, "Constant<T> was lowered to unexpected format");
   }
 
   template <typename T>
   hlfir::EntityWithAttributes
-  gen(const Fortran::evaluate::ArrayConstructor<T> &expr) {
-    TODO(getLoc(), "lowering ArrayCtor to HLFIR");
+  gen(const Fortran::evaluate::ArrayConstructor<T> &arrayCtor) {
+    return Fortran::lower::ArrayConstructorBuilder<T>::gen(
+        getLoc(), getConverter(), arrayCtor, getSymMap(), getStmtCtx());
   }
 
   template <typename D, typename R, typename O>
   hlfir::EntityWithAttributes
   gen(const Fortran::evaluate::Operation<D, R, O> &op) {
     auto &builder = getBuilder();
     mlir::Location loc = getLoc();
     const int rank = op.Rank();
case Fortran::evaluate::DescriptorInquiry::Field::Stride:
       // So far the front end does not generate this inquiry.
       TODO(loc, "stride inquiry");
     }
     llvm_unreachable("unknown descriptor inquiry");
   }
 
   hlfir::EntityWithAttributes
   gen(const Fortran::evaluate::ImpliedDoIndex &var) {
-    TODO(getLoc(), "lowering implied do index to HLFIR");
+    mlir::Value value = symMap.lookupImpliedDo(toStringRef(var.name));
+    assert(value && "impled do was not mapped");
+    return hlfir::EntityWithAttributes{value};
   }
 
   hlfir::EntityWithAttributes
   gen(const Fortran::evaluate::StructureConstructor &var) {
     TODO(getLoc(), "lowering structure constructor to HLFIR");
   }
 
   mlir::Location getLoc() const { return loc; }

flang/test/Lower/HLFIR/array-ctor-as-inlined-temp.f90

+! Test lowering of array constructors as inlined temporary.
+! RUN: bbc -emit-fir -hlfir -o - %s | FileCheck %s
+
+subroutine test_simple(i)
+  call takes_int([42, i])
+end subroutine
+! CHECK-LABEL: func.func @_QPtest_simple(
+! CHECK:  %[[VAL_1:.*]]:2 = hlfir.declare {{.*}}Ei
+! CHECK:  %[[VAL_2:.*]] = arith.constant 2 : index
+! CHECK:  %[[VAL_3:.*]] = arith.constant 1 : index
+! CHECK:  %[[VAL_4:.*]] = fir.allocmem !fir.array<2xi32> {bindc_name = ".tmp.arrayctor", uniq_name = ""}
+! CHECK:  %[[VAL_5:.*]] = fir.shape %[[VAL_2]] : (index) -> !fir.shape<1>
+! CHECK:  %[[VAL_6:.*]]:2 = hlfir.declare %[[VAL_4]](%[[VAL_5]]) {uniq_name = ".tmp.arrayctor"} : (!fir.heap<!fir.array<2xi32>>, !fir.shape<1>) -> (!fir.heap<!fir.array<2xi32>>, !fir.heap<!fir.array<2xi32>>)
+! CHECK:  %[[VAL_7:.*]] = arith.constant 42 : i32
+! CHECK:  %[[VAL_8:.*]] = arith.addi %[[VAL_3]], %[[VAL_3]] : index
+! CHECK:  %[[VAL_9:.*]] = hlfir.designate %[[VAL_6]]#0 (%[[VAL_3]])  : (!fir.heap<!fir.array<2xi32>>, index) -> !fir.ref<i32>
+! CHECK:  hlfir.assign %[[VAL_7]] to %[[VAL_9]] : i32, !fir.ref<i32>
+! CHECK:  %[[VAL_10:.*]] = fir.load %[[VAL_1]]#0 : !fir.ref<i32>
+! CHECK:  %[[VAL_11:.*]] = hlfir.designate %[[VAL_6]]#0 (%[[VAL_8]])  : (!fir.heap<!fir.array<2xi32>>, index) -> !fir.ref<i32>
+! CHECK:  hlfir.assign %[[VAL_10]] to %[[VAL_11]] : i32, !fir.ref<i32>
+! CHECK:  %[[VAL_12:.*]] = arith.constant true
+! CHECK:  %[[VAL_13:.*]] = hlfir.as_expr %[[VAL_6]]#0 move %[[VAL_12]] : (!fir.heap<!fir.array<2xi32>>, i1) -> !hlfir.expr<2xi32>
+! CHECK:  fir.call
+! CHECK:  hlfir.destroy %[[VAL_13]] : !hlfir.expr<2xi32>
+
+subroutine test_simple_real(x)
+  real(2) :: x
+  call takes_real_2([x, 0._2])
+end subroutine
+! CHECK-LABEL: func.func @_QPtest_simple_real(
+! CHECK:  %[[VAL_1:.*]]:2 = hlfir.declare {{.*}}Ex
+! CHECK:  %[[VAL_2:.*]] = arith.constant 2 : index
+! CHECK:  %[[VAL_3:.*]] = arith.constant 1 : index
+! CHECK:  %[[VAL_4:.*]] = fir.allocmem !fir.array<2xf16> {bindc_name = ".tmp.arrayctor", uniq_name = ""}
+! CHECK:  %[[VAL_5:.*]] = fir.shape %[[VAL_2]] : (index) -> !fir.shape<1>
+! CHECK:  %[[VAL_6:.*]]:2 = hlfir.declare %[[VAL_4]](%[[VAL_5]]) {uniq_name = ".tmp.arrayctor"} : (!fir.heap<!fir.array<2xf16>>, !fir.shape<1>) -> (!fir.heap<!fir.array<2xf16>>, !fir.heap<!fir.array<2xf16>>)
+! CHECK:  %[[VAL_7:.*]] = fir.load %[[VAL_1]]#0 : !fir.ref<f16>
+! CHECK:  %[[VAL_8:.*]] = arith.addi %[[VAL_3]], %[[VAL_3]] : index
+! CHECK:  %[[VAL_9:.*]] = hlfir.designate %[[VAL_6]]#0 (%[[VAL_3]])  : (!fir.heap<!fir.array<2xf16>>, index) -> !fir.ref<f16>
+! CHECK:  hlfir.assign %[[VAL_7]] to %[[VAL_9]] : f16, !fir.ref<f16>
+! CHECK:  %[[VAL_10:.*]] = arith.constant 0.000000e+00 : f16
+! CHECK:  %[[VAL_11:.*]] = hlfir.designate %[[VAL_6]]#0 (%[[VAL_8]])  : (!fir.heap<!fir.array<2xf16>>, index) -> !fir.ref<f16>
+! CHECK:  hlfir.assign %[[VAL_10]] to %[[VAL_11]] : f16, !fir.ref<f16>
+! CHECK:  %[[VAL_12:.*]] = arith.constant true
+! CHECK:  %[[VAL_13:.*]] = hlfir.as_expr %[[VAL_6]]#0 move %[[VAL_12]] : (!fir.heap<!fir.array<2xf16>>, i1) -> !hlfir.expr<2xf16>
+! CHECK:  fir.call
+! CHECK:  hlfir.destroy %[[VAL_13]] : !hlfir.expr<2xf16>
+
+subroutine test_simple_complex(z)
+  complex :: z
+  call takes_cmplx_8([complex(8):: 42, z])
+end subroutine
+! CHECK-LABEL: func.func @_QPtest_simple_complex(
+! CHECK:  %[[VAL_1:.*]]:2 = hlfir.declare {{.*}}Ez
+! CHECK:  %[[VAL_2:.*]] = arith.constant 2 : index
+! CHECK:  %[[VAL_3:.*]] = arith.constant 1 : index
+! CHECK:  %[[VAL_4:.*]] = fir.allocmem !fir.array<2x!fir.complex<8>> {bindc_name = ".tmp.arrayctor", uniq_name = ""}
+! CHECK:  %[[VAL_5:.*]] = fir.shape %[[VAL_2]] : (index) -> !fir.shape<1>
+! CHECK:  %[[VAL_6:.*]]:2 = hlfir.declare %[[VAL_4]](%[[VAL_5]]) {uniq_name = ".tmp.arrayctor"} : (!fir.heap<!fir.array<2x!fir.complex<8>>>, !fir.shape<1>) -> (!fir.heap<!fir.array<2x!fir.complex<8>>>, !fir.heap<!fir.array<2x!fir.complex<8>>>)
+! CHECK:  %[[VAL_7:.*]] = arith.constant 42 : i32
+! CHECK:  %[[VAL_8:.*]] = fir.convert %[[VAL_7]] : (i32) -> f64
+! CHECK:  %[[VAL_9:.*]] = arith.constant 0.000000e+00 : f64
+! CHECK:  %[[VAL_10:.*]] = fir.undefined !fir.complex<8>
+! CHECK:  %[[VAL_11:.*]] = fir.insert_value %[[VAL_10]], %[[VAL_8]], [0 : index] : (!fir.complex<8>, f64) -> !fir.complex<8>
+! CHECK:  %[[VAL_12:.*]] = fir.insert_value %[[VAL_11]], %[[VAL_9]], [1 : index] : (!fir.complex<8>, f64) -> !fir.complex<8>
+! CHECK:  %[[VAL_13:.*]] = arith.addi %[[VAL_3]], %[[VAL_3]] : index
+! CHECK:  %[[VAL_14:.*]] = hlfir.designate %[[VAL_6]]#0 (%[[VAL_3]])  : (!fir.heap<!fir.array<2x!fir.complex<8>>>, index) -> !fir.ref<!fir.complex<8>>
+! CHECK:  hlfir.assign %[[VAL_12]] to %[[VAL_14]] : !fir.complex<8>, !fir.ref<!fir.complex<8>>
+! CHECK:  %[[VAL_15:.*]] = fir.load %[[VAL_1]]#0 : !fir.ref<!fir.complex<4>>
+! CHECK:  %[[VAL_16:.*]] = fir.convert %[[VAL_15]] : (!fir.complex<4>) -> !fir.complex<8>
+! CHECK:  %[[VAL_17:.*]] = hlfir.designate %[[VAL_6]]#0 (%[[VAL_13]])  : (!fir.heap<!fir.array<2x!fir.complex<8>>>, index) -> !fir.ref<!fir.complex<8>>
+! CHECK:  hlfir.assign %[[VAL_16]] to %[[VAL_17]] : !fir.complex<8>, !fir.ref<!fir.complex<8>>
+! CHECK:  %[[VAL_18:.*]] = arith.constant true
+! CHECK:  %[[VAL_19:.*]] = hlfir.as_expr %[[VAL_6]]#0 move %[[VAL_18]] : (!fir.heap<!fir.array<2x!fir.complex<8>>>, i1) -> !hlfir.expr<2x!fir.complex<8>>
+! CHECK:  fir.call
+! CHECK:  hlfir.destroy %[[VAL_19]] : !hlfir.expr<2x!fir.complex<8>>
+
+subroutine test_simple_logical(a, b)
+  logical :: a, b
+  call takes_logical([a, a.and.b])
+end subroutine
+! CHECK-LABEL: func.func @_QPtest_simple_logical(
+! CHECK:  %[[VAL_2:.*]]:2 = hlfir.declare {{.*}}Ea
+! CHECK:  %[[VAL_3:.*]]:2 = hlfir.declare {{.*}}Eb
+! CHECK:  %[[VAL_4:.*]] = arith.constant 2 : index
+! CHECK:  %[[VAL_5:.*]] = arith.constant 1 : index
+! CHECK:  %[[VAL_6:.*]] = fir.allocmem !fir.array<2x!fir.logical<4>> {bindc_name = ".tmp.arrayctor", uniq_name = ""}
+! CHECK:  %[[VAL_7:.*]] = fir.shape %[[VAL_4]] : (index) -> !fir.shape<1>
+! CHECK:  %[[VAL_8:.*]]:2 = hlfir.declare %[[VAL_6]](%[[VAL_7]]) {uniq_name = ".tmp.arrayctor"} : (!fir.heap<!fir.array<2x!fir.logical<4>>>, !fir.shape<1>) -> (!fir.heap<!fir.array<2x!fir.logical<4>>>, !fir.heap<!fir.array<2x!fir.logical<4>>>)
+! CHECK:  %[[VAL_9:.*]] = fir.load %[[VAL_2]]#0 : !fir.ref<!fir.logical<4>>
+! CHECK:  %[[VAL_10:.*]] = arith.addi %[[VAL_5]], %[[VAL_5]] : index
+! CHECK:  %[[VAL_11:.*]] = hlfir.designate %[[VAL_8]]#0 (%[[VAL_5]])  : (!fir.heap<!fir.array<2x!fir.logical<4>>>, index) -> !fir.ref<!fir.logical<4>>
+! CHECK:  hlfir.assign %[[VAL_9]] to %[[VAL_11]] : !fir.logical<4>, !fir.ref<!fir.logical<4>>
+! CHECK:  %[[VAL_12:.*]] = fir.load %[[VAL_2]]#0 : !fir.ref<!fir.logical<4>>
+! CHECK:  %[[VAL_13:.*]] = fir.load %[[VAL_3]]#0 : !fir.ref<!fir.logical<4>>
+! CHECK:  %[[VAL_14:.*]] = fir.convert %[[VAL_12]] : (!fir.logical<4>) -> i1
+! CHECK:  %[[VAL_15:.*]] = fir.convert %[[VAL_13]] : (!fir.logical<4>) -> i1
+! CHECK:  %[[VAL_16:.*]] = arith.andi %[[VAL_14]], %[[VAL_15]] : i1
+! CHECK:  %[[VAL_17:.*]] = hlfir.designate %[[VAL_8]]#0 (%[[VAL_10]])  : (!fir.heap<!fir.array<2x!fir.logical<4>>>, index) -> !fir.ref<!fir.logical<4>>
+! CHECK:  hlfir.assign %[[VAL_16]] to %[[VAL_17]] : i1, !fir.ref<!fir.logical<4>>
+! CHECK:  %[[VAL_18:.*]] = arith.constant true
+! CHECK:  %[[VAL_19:.*]] = hlfir.as_expr %[[VAL_8]]#0 move %[[VAL_18]] : (!fir.heap<!fir.array<2x!fir.logical<4>>>, i1) -> !hlfir.expr<2x!fir.logical<4>>
+! CHECK:  fir.call
+! CHECK:  hlfir.destroy %[[VAL_19]] : !hlfir.expr<2x!fir.logical<4>>
+
+subroutine test_implied_do(n)
+  integer(8) :: n
+  ! This implied do cannot easily be promoted to hlfir.elemental because
+  ! the implied do contains more than one scalar ac-value.
+  call takes_int([(42, j, j=1,n)])
+end subroutine
+! CHECK-LABEL: func.func @_QPtest_implied_do(
+! CHECK:  %[[VAL_1:.*]] = fir.alloca index
+! CHECK:  %[[VAL_2:.*]]:2 = hlfir.declare {{.*}}En
+! CHECK:  %[[VAL_3:.*]] = arith.constant 0 : i64
+! CHECK:  %[[VAL_4:.*]] = arith.constant 2 : i64
+! CHECK:  %[[VAL_5:.*]] = fir.load %[[VAL_2]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_6:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_7:.*]] = arith.subi %[[VAL_5]], %[[VAL_6]] : i64
+! CHECK:  %[[VAL_8:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_9:.*]] = arith.addi %[[VAL_7]], %[[VAL_8]] : i64
+! CHECK:  %[[VAL_10:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_11:.*]] = arith.divsi %[[VAL_9]], %[[VAL_10]] : i64
+! CHECK:  %[[VAL_12:.*]] = arith.constant 0 : i64
+! CHECK:  %[[VAL_13:.*]] = arith.cmpi sgt, %[[VAL_11]], %[[VAL_12]] : i64
+! CHECK:  %[[VAL_14:.*]] = arith.select %[[VAL_13]], %[[VAL_11]], %[[VAL_12]] : i64
+! CHECK:  %[[VAL_15:.*]] = arith.muli %[[VAL_4]], %[[VAL_14]] : i64
+! CHECK:  %[[VAL_16:.*]] = arith.addi %[[VAL_3]], %[[VAL_15]] : i64
+! CHECK:  %[[VAL_17:.*]] = fir.convert %[[VAL_16]] : (i64) -> index
+! CHECK:  %[[VAL_18:.*]] = arith.constant 1 : index
+! CHECK:  fir.store %[[VAL_18]] to %[[VAL_1]] : !fir.ref<index>
+! CHECK:  %[[VAL_19:.*]] = fir.allocmem !fir.array<?xi32>, %[[VAL_17]] {bindc_name = ".tmp.arrayctor", uniq_name = ""}
+! CHECK:  %[[VAL_20:.*]] = fir.shape %[[VAL_17]] : (index) -> !fir.shape<1>
+! CHECK:  %[[VAL_21:.*]]:2 = hlfir.declare %[[VAL_19]](%[[VAL_20]]) {uniq_name = ".tmp.arrayctor"} : (!fir.heap<!fir.array<?xi32>>, !fir.shape<1>) -> (!fir.box<!fir.array<?xi32>>, !fir.heap<!fir.array<?xi32>>)
+! CHECK:  %[[VAL_22:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_23:.*]] = fir.convert %[[VAL_22]] : (i64) -> index
+! CHECK:  %[[VAL_24:.*]] = fir.load %[[VAL_2]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_25:.*]] = fir.convert %[[VAL_24]] : (i64) -> index
+! CHECK:  %[[VAL_26:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_27:.*]] = fir.convert %[[VAL_26]] : (i64) -> index
+! CHECK:  fir.do_loop %[[VAL_28:.*]] = %[[VAL_23]] to %[[VAL_25]] step %[[VAL_27]] {
+! CHECK:    %[[VAL_29:.*]] = arith.constant 42 : i32
+! CHECK:    %[[VAL_30:.*]] = fir.load %[[VAL_1]] : !fir.ref<index>
+! CHECK:    %[[VAL_31:.*]] = arith.addi %[[VAL_30]], %[[VAL_18]] : index
+! CHECK:    fir.store %[[VAL_31]] to %[[VAL_1]] : !fir.ref<index>
+! CHECK:    %[[VAL_32:.*]] = hlfir.designate %[[VAL_21]]#0 (%[[VAL_31]])  : (!fir.box<!fir.array<?xi32>>, index) -> !fir.ref<i32>
+! CHECK:    hlfir.assign %[[VAL_29]] to %[[VAL_32]] : i32, !fir.ref<i32>
+! CHECK:    %[[VAL_33:.*]] = fir.convert %[[VAL_28]] : (index) -> i32
+! CHECK:    %[[VAL_34:.*]] = fir.load %[[VAL_1]] : !fir.ref<index>
+! CHECK:    %[[VAL_35:.*]] = arith.addi %[[VAL_34]], %[[VAL_18]] : index
+! CHECK:    fir.store %[[VAL_35]] to %[[VAL_1]] : !fir.ref<index>
+! CHECK:    %[[VAL_36:.*]] = hlfir.designate %[[VAL_21]]#0 (%[[VAL_35]])  : (!fir.box<!fir.array<?xi32>>, index) -> !fir.ref<i32>
+! CHECK:    hlfir.assign %[[VAL_33]] to %[[VAL_36]] : i32, !fir.ref<i32>
+! CHECK:  }
+! CHECK:  %[[VAL_37:.*]] = arith.constant true
+! CHECK:  %[[VAL_38:.*]] = hlfir.as_expr %[[VAL_21]]#0 move %[[VAL_37]] : (!fir.box<!fir.array<?xi32>>, i1) -> !hlfir.expr<?xi32>
+! CHECK:  fir.call
+! CHECK:  hlfir.destroy %[[VAL_38]] : !hlfir.expr<?xi32>
+
+subroutine test_strided_implied_do(lb, ub, stride)
+  integer(8) :: lb, ub, stride
+  call takes_int([(42, j, j=lb,ub,stride)])
+end subroutine
+! CHECK-LABEL: func.func @_QPtest_strided_implied_do(
+! CHECK:  %[[VAL_3:.*]] = fir.alloca index
+! CHECK:  %[[VAL_4:.*]]:2 = hlfir.declare {{.*}}Elb
+! CHECK:  %[[VAL_5:.*]]:2 = hlfir.declare {{.*}}Estride
+! CHECK:  %[[VAL_6:.*]]:2 = hlfir.declare {{.*}}Eub
+! CHECK:  %[[VAL_7:.*]] = arith.constant 0 : i64
+! CHECK:  %[[VAL_8:.*]] = arith.constant 2 : i64
+! CHECK:  %[[VAL_9:.*]] = fir.load %[[VAL_6]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_10:.*]] = fir.load %[[VAL_4]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_11:.*]] = arith.subi %[[VAL_9]], %[[VAL_10]] : i64
+! CHECK:  %[[VAL_12:.*]] = fir.load %[[VAL_5]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_13:.*]] = arith.addi %[[VAL_11]], %[[VAL_12]] : i64
+! CHECK:  %[[VAL_14:.*]] = fir.load %[[VAL_5]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_15:.*]] = arith.divsi %[[VAL_13]], %[[VAL_14]] : i64
+! CHECK:  %[[VAL_16:.*]] = arith.constant 0 : i64
+! CHECK:  %[[VAL_17:.*]] = arith.cmpi sgt, %[[VAL_15]], %[[VAL_16]] : i64
+! CHECK:  %[[VAL_18:.*]] = arith.select %[[VAL_17]], %[[VAL_15]], %[[VAL_16]] : i64
+! CHECK:  %[[VAL_19:.*]] = arith.muli %[[VAL_8]], %[[VAL_18]] : i64
+! CHECK:  %[[VAL_20:.*]] = arith.addi %[[VAL_7]], %[[VAL_19]] : i64
+! CHECK:  %[[VAL_21:.*]] = fir.convert %[[VAL_20]] : (i64) -> index
+! CHECK:  %[[VAL_22:.*]] = arith.constant 1 : index
+! CHECK:  fir.store %[[VAL_22]] to %[[VAL_3]] : !fir.ref<index>
+! CHECK:  %[[VAL_23:.*]] = fir.allocmem !fir.array<?xi32>, %[[VAL_21]] {bindc_name = ".tmp.arrayctor", uniq_name = ""}
+! CHECK:  %[[VAL_24:.*]] = fir.shape %[[VAL_21]] : (index) -> !fir.shape<1>
+! CHECK:  %[[VAL_25:.*]]:2 = hlfir.declare %[[VAL_23]](%[[VAL_24]]) {uniq_name = ".tmp.arrayctor"} : (!fir.heap<!fir.array<?xi32>>, !fir.shape<1>) -> (!fir.box<!fir.array<?xi32>>, !fir.heap<!fir.array<?xi32>>)
+! CHECK:  %[[VAL_26:.*]] = fir.load %[[VAL_4]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_27:.*]] = fir.convert %[[VAL_26]] : (i64) -> index
+! CHECK:  %[[VAL_28:.*]] = fir.load %[[VAL_6]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_29:.*]] = fir.convert %[[VAL_28]] : (i64) -> index
+! CHECK:  %[[VAL_30:.*]] = fir.load %[[VAL_5]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_31:.*]] = fir.convert %[[VAL_30]] : (i64) -> index
+! CHECK:  fir.do_loop %[[VAL_32:.*]] = %[[VAL_27]] to %[[VAL_29]] step %[[VAL_31]] {
+! CHECK:    %[[VAL_33:.*]] = arith.constant 42 : i32
+! CHECK:    %[[VAL_34:.*]] = fir.load %[[VAL_3]] : !fir.ref<index>
+! CHECK:    %[[VAL_35:.*]] = arith.addi %[[VAL_34]], %[[VAL_22]] : index
+! CHECK:    fir.store %[[VAL_35]] to %[[VAL_3]] : !fir.ref<index>
+! CHECK:    %[[VAL_36:.*]] = hlfir.designate %[[VAL_25]]#0 (%[[VAL_35]])  : (!fir.box<!fir.array<?xi32>>, index) -> !fir.ref<i32>
+! CHECK:    hlfir.assign %[[VAL_33]] to %[[VAL_36]] : i32, !fir.ref<i32>
+! CHECK:    %[[VAL_37:.*]] = fir.convert %[[VAL_32]] : (index) -> i32
+! CHECK:    %[[VAL_38:.*]] = fir.load %[[VAL_3]] : !fir.ref<index>
+! CHECK:    %[[VAL_39:.*]] = arith.addi %[[VAL_38]], %[[VAL_22]] : index
+! CHECK:    fir.store %[[VAL_39]] to %[[VAL_3]] : !fir.ref<index>
+! CHECK:    %[[VAL_40:.*]] = hlfir.designate %[[VAL_25]]#0 (%[[VAL_39]])  : (!fir.box<!fir.array<?xi32>>, index) -> !fir.ref<i32>
+! CHECK:    hlfir.assign %[[VAL_37]] to %[[VAL_40]] : i32, !fir.ref<i32>
+! CHECK:  }
+! CHECK:  %[[VAL_41:.*]] = arith.constant true
+! CHECK:  %[[VAL_42:.*]] = hlfir.as_expr %[[VAL_25]]#0 move %[[VAL_41]] : (!fir.box<!fir.array<?xi32>>, i1) -> !hlfir.expr<?xi32>
+! CHECK:  fir.call
+! CHECK:  hlfir.destroy %[[VAL_42]] : !hlfir.expr<?xi32>
+
+subroutine test_nested_implied_do(n, m)
+  integer(8) :: n, m
+  call takes_int([((i+j, i=1,m), j=1,n)])
+end subroutine
+! CHECK-LABEL: func.func @_QPtest_nested_implied_do(
+! CHECK:  %[[VAL_2:.*]] = fir.alloca index
+! CHECK:  %[[VAL_3:.*]]:2 = hlfir.declare {{.*}}Em
+! CHECK:  %[[VAL_4:.*]]:2 = hlfir.declare {{.*}}En
+! CHECK:  %[[VAL_5:.*]] = arith.constant 0 : i64
+! CHECK:  %[[VAL_6:.*]] = arith.constant 0 : i64
+! CHECK:  %[[VAL_7:.*]] = fir.load %[[VAL_3]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_8:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_9:.*]] = arith.subi %[[VAL_7]], %[[VAL_8]] : i64
+! CHECK:  %[[VAL_10:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_11:.*]] = arith.addi %[[VAL_9]], %[[VAL_10]] : i64
+! CHECK:  %[[VAL_12:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_13:.*]] = arith.divsi %[[VAL_11]], %[[VAL_12]] : i64
+! CHECK:  %[[VAL_14:.*]] = arith.constant 0 : i64
+! CHECK:  %[[VAL_15:.*]] = arith.cmpi sgt, %[[VAL_13]], %[[VAL_14]] : i64
+! CHECK:  %[[VAL_16:.*]] = arith.select %[[VAL_15]], %[[VAL_13]], %[[VAL_14]] : i64
+! CHECK:  %[[VAL_17:.*]] = arith.addi %[[VAL_6]], %[[VAL_16]] : i64
+! CHECK:  %[[VAL_18:.*]] = fir.load %[[VAL_4]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_19:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_20:.*]] = arith.subi %[[VAL_18]], %[[VAL_19]] : i64
+! CHECK:  %[[VAL_21:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_22:.*]] = arith.addi %[[VAL_20]], %[[VAL_21]] : i64
+! CHECK:  %[[VAL_23:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_24:.*]] = arith.divsi %[[VAL_22]], %[[VAL_23]] : i64
+! CHECK:  %[[VAL_25:.*]] = arith.constant 0 : i64
+! CHECK:  %[[VAL_26:.*]] = arith.cmpi sgt, %[[VAL_24]], %[[VAL_25]] : i64
+! CHECK:  %[[VAL_27:.*]] = arith.select %[[VAL_26]], %[[VAL_24]], %[[VAL_25]] : i64
+! CHECK:  %[[VAL_28:.*]] = arith.muli %[[VAL_17]], %[[VAL_27]] : i64
+! CHECK:  %[[VAL_29:.*]] = arith.addi %[[VAL_5]], %[[VAL_28]] : i64
+! CHECK:  %[[VAL_30:.*]] = fir.convert %[[VAL_29]] : (i64) -> index
+! CHECK:  %[[VAL_31:.*]] = arith.constant 1 : index
+! CHECK:  fir.store %[[VAL_31]] to %[[VAL_2]] : !fir.ref<index>
+! CHECK:  %[[VAL_32:.*]] = fir.allocmem !fir.array<?xi32>, %[[VAL_30]] {bindc_name = ".tmp.arrayctor", uniq_name = ""}
+! CHECK:  %[[VAL_33:.*]] = fir.shape %[[VAL_30]] : (index) -> !fir.shape<1>
+! CHECK:  %[[VAL_34:.*]]:2 = hlfir.declare %[[VAL_32]](%[[VAL_33]]) {uniq_name = ".tmp.arrayctor"} : (!fir.heap<!fir.array<?xi32>>, !fir.shape<1>) -> (!fir.box<!fir.array<?xi32>>, !fir.heap<!fir.array<?xi32>>)
+! CHECK:  %[[VAL_35:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_36:.*]] = fir.convert %[[VAL_35]] : (i64) -> index
+! CHECK:  %[[VAL_37:.*]] = fir.load %[[VAL_4]]#0 : !fir.ref<i64>
+! CHECK:  %[[VAL_38:.*]] = fir.convert %[[VAL_37]] : (i64) -> index
+! CHECK:  %[[VAL_39:.*]] = arith.constant 1 : i64
+! CHECK:  %[[VAL_40:.*]] = fir.convert %[[VAL_39]] : (i64) -> index
+! CHECK:  fir.do_loop %[[VAL_41:.*]] = %[[VAL_36]] to %[[VAL_38]] step %[[VAL_40]] {
+! CHECK:    %[[VAL_42:.*]] = arith.constant 1 : i64
+! CHECK:    %[[VAL_43:.*]] = fir.convert %[[VAL_42]] : (i64) -> index
+! CHECK:    %[[VAL_44:.*]] = fir.load %[[VAL_3]]#0 : !fir.ref<i64>
+! CHECK:    %[[VAL_45:.*]] = fir.convert %[[VAL_44]] : (i64) -> index
+! CHECK:    %[[VAL_46:.*]] = arith.constant 1 : i64
+! CHECK:    %[[VAL_47:.*]] = fir.convert %[[VAL_46]] : (i64) -> index
+! CHECK:    fir.do_loop %[[VAL_48:.*]] = %[[VAL_43]] to %[[VAL_45]] step %[[VAL_47]] {
+! CHECK:      %[[VAL_49:.*]] = fir.convert %[[VAL_48]] : (index) -> i32
+! CHECK:      %[[VAL_50:.*]] = fir.convert %[[VAL_41]] : (index) -> i32
+! CHECK:      %[[VAL_51:.*]] = arith.addi %[[VAL_49]], %[[VAL_50]] : i32
+! CHECK:      %[[VAL_52:.*]] = fir.load %[[VAL_2]] : !fir.ref<index>
+! CHECK:      %[[VAL_53:.*]] = arith.addi %[[VAL_52]], %[[VAL_31]] : index
+! CHECK:      fir.store %[[VAL_53]] to %[[VAL_2]] : !fir.ref<index>
+! CHECK:      %[[VAL_54:.*]] = hlfir.designate %[[VAL_34]]#0 (%[[VAL_53]])  : (!fir.box<!fir.array<?xi32>>, index) -> !fir.ref<i32>
+! CHECK:      hlfir.assign %[[VAL_51]] to %[[VAL_54]] : i32, !fir.ref<i32>
+! CHECK:    }
+! CHECK:  }
+! CHECK:  %[[VAL_55:.*]] = arith.constant true
+! CHECK:  %[[VAL_56:.*]] = hlfir.as_expr %[[VAL_34]]#0 move %[[VAL_55]] : (!fir.box<!fir.array<?xi32>>, i1) -> !hlfir.expr<?xi32>
+! CHECK:  fir.call
+! CHECK:  hlfir.destroy %[[VAL_56]] : !hlfir.expr<?xi32>