[flang] Keep original data type for do-variable value.

[vzakhari]

Summary

Keep the original data type of integer do-variables
for structured loops. When do-variable's data type
is an integer type shorter than IndexType, processing
the do-variable separately from the DoLoop's iteration index
allows getting rid of type casts, which can make backend
optimizations easier.

For example,

do i = 2, n-1
  do j = 2, n-1
    ... = a(j-1, i)
  end do
end do

If value of 'j' is computed by casting the DoLoop's iteration
index to 'i32', then Flang will produce the following LLVM IR:

%1 = trunc i64 %iter_index to i32
%2 = sub i32 %1, 1
%3 = sext i32 %2 to i64

LLVM's InstCombine may try to get rid of the sign extension,
and may transform this into:

%1 = shl i64 %iter_index, 32
%2 = add i64 %1, -4294967296
%3 = ashr exact i64 %2, 32

The extra computations for the element address applied on top
of this awkward pattern confuse LLVM vectorizer so that
it does not recognize the unit-strided access of 'a'.

Measured performance improvements on SPEC CPU2000@IceLake:

168.wupwise:    11.96%
171.swim:       11.22%
172.mrgid:      56.38%
178.galgel:      7.29%
301.apsi:        8.32%

[vzakhari]

This should fix the Windows build failure.

[clementval]

flang/lib/Lower/Bridge.cpp

173

Any good reason to have 64 inlined elements here?

1248

IndexType's bit width is meant to be target specific so I guess it is normal you cannot get the information at this stage.

[vzakhari]

flang/lib/Lower/Bridge.cpp

173

No particular reason except to keep the default vector size of 64. Without specifying 64 the compilation asserts that IncrementLoopInfo is smaller than 256 bytes, otherwise, it is considered to be too big. When the vector size is explicitly specified the assertion is ignored regardless of the size.

1248

Yep, this sounds right.

[clementval]

flang/lib/Lower/Bridge.cpp

173

This is not specifying the size of the vector but the number of inlined elements.

[vzakhari]

flang/lib/Lower/Bridge.cpp

173

Right, I did not use the right words :)

[clementval]

LGTM

flang/lib/Lower/Bridge.cpp

173

Ok then fine for me.

[vdonaldson]

Please wait on this until Jean has had a chance to review it.

This approach will only affect user loops. It won't affect compiler generated loops.

If a motivation for using an index type is to accommodate affine code optimization, there might be some alternatives. It might be possible to generate the loop with the original type, convert it to an index type before an affine pass, and then convert it back. Or some such. If something like that works, it might be cleaner.

flang/lib/Lower/Bridge.cpp

99

// Do loop block argument holding the current value

1246–1248

The width of an integer type is intTy.getWidth(). Doesn't that apply here?

[vdonaldson]

flang/lib/Lower/Bridge.cpp

173

Isn't a loop nest depth of 64 massive overkill? 4 would cover the vast majority of use cases.

The code originally had a lot of explicit llvm::SmallVector sizes, and someone (mlir folks?) suggested it was better to not do that. Why is this case different than other cases?

[vzakhari]

In D132176#3735961, @vdonaldson wrote:

Please wait on this until Jean has had a chance to review it.

This approach will only affect user loops. It won't affect compiler generated loops.

This is correct. At the same time, the compiler generated loops mean to traverse all elements of some array, and since the extents have IndexType we have to use IndexType loops in many cases. These loops may use shorter IV's when the extents are known to fit a shorter integer type, but I am not sure how easy it is to recognize such cases in the current lowering scheme.
This change only tries to preserve what is written in the user code explicitly, i.e. if user uses INTEGER(4) IV, then it will stay in i32.

If a motivation for using an index type is to accommodate affine code optimization, there might be some alternatives. It might be possible to generate the loop with the original type, convert it to an index type before an affine pass, and then convert it back. Or some such. If something like that works, it might be cleaner.

This may require changing DoLoop definition that only supports IndexType index currently (just like scf::ForOp), but this is possible. I suppose the LLVM IR dialect lowering of such a DoLoop will have to compute the iteration count in i64 anyway to handle loops with more iterations than signed i32 can represent.

I am not sure how easy it will be to "remember" a loop's original indexing value type between affine promotion/demotion since affine transformations can delete loops (e.g. loop fusion) and create new ones (e.g. loop distribution).

I will wait for Jean's comments.

Thanks for review!

flang/lib/Lower/Bridge.cpp

99

Will change. Thanks!

1246–1248

IndexType is not an IntegerType and it does not provide means for getting its width, so there is no way to compare widths of IntegerType and IndexType.

[vzakhari]

flang/lib/Lower/Bridge.cpp

173

I guess it is not different from other cases. To select the right constant we need to know the average number of elements for all programs. I do not know it in this case, so 64, 4 and even 1 seem the same to me :) In the end we are just balancing between stack usage and heap allocation overhead.

[vzakhari]

Fixed comment, and changed pre-allocated vector size to 8.

[vdonaldson]

flang/lib/Lower/Bridge.cpp

1250

A loop that hasRealControl is unstructured, so this half of the condition is a nop.

[jeanPerier]

There is one alternative to your fix at the flang level that I see, but I am not sure it is legal/ok to go that route:

From a Fortran point of view, would it be allowed to use the index (i64) induction variable directly when the Fortran loop variable appears inside the loops (in places where its value is needed, and it does not need to be passed by address) If it were allowed, it would be possible to avoid generating the casts and to add an extra induction variable. But I am not sure this is legal, so your approach might just be the best workaround to this LLVM optimization issue.

To illustrate, I am wondering if it would it be legal to lower:

 do j = 2, n-1
    ... = a(j-42)
end do

into:

 fir.do_loop %induct = %lb to %ub step %step-> index {
  %induct_cast = fir.convert %induct : (index) -> i32
  fir.store %induct_cast to %induct_cast : !fir.ref<i32>
  ! Using %induct directly inside the loop
  %add = arith.subi %induct, %c42 : index
  %addr = fir.coordinate_of %a_base, %add
}

That way, the vectorizer would probably work and the generated code would be even cleaner. However, this code would behave differently in case of integer overflow, and while overflow in indexing is probably always bad (I am not sure what Fortran or flang guarantees here), I am not sure about the cases where the induction variable would appear in a general integer expression.

[vzakhari]

In D132176#3740653, @jeanPerier wrote:

There is one alternative to your fix at the flang level that I see, but I am not sure it is legal/ok to go that route:

From a Fortran point of view, would it be allowed to use the index (i64) induction variable directly when the Fortran loop variable appears inside the loops (in places where its value is needed, and it does not need to be passed by address) If it were allowed, it would be possible to avoid generating the casts and to add an extra induction variable. But I am not sure this is legal, so your approach might just be the best workaround to this LLVM optimization issue.

To illustrate, I am wondering if it would it be legal to lower:

 do j = 2, n-1
    ... = a(j-42)
end do

into:

 fir.do_loop %induct = %lb to %ub step %step-> index {
  %induct_cast = fir.convert %induct : (index) -> i32
  fir.store %induct_cast to %induct_cast : !fir.ref<i32>
  ! Using %induct directly inside the loop
  %add = arith.subi %induct, %c42 : index
  %addr = fir.coordinate_of %a_base, %add
}

That way, the vectorizer would probably work and the generated code would be even cleaner. However, this code would behave differently in case of integer overflow, and while overflow in indexing is probably always bad (I am not sure what Fortran or flang guarantees here), I am not sure about the cases where the induction variable would appear in a general integer expression.

I think we do not care about the integer overflow, because a Fortran program relying on particular overflow behavior is non-conformant. At the same time, I am not sure how easy it is to extend the whole computation tree(s) dependent on the induction variable from their original data type to index. It does not sound trivial to do during the lowering.

[jeanPerier]

In D132176#3740735, @vzakhari wrote:

I think we do not care about the integer overflow, because a Fortran program relying on particular overflow behavior is non-conformant. At the same time, I am not sure how easy it is to extend the whole computation tree(s) dependent on the induction variable from their original data type to index. It does not sound trivial to do during the lowering.

Fair point, this would complexify lowering of expressions that expects getting the same operand types. Your fix is more localized, so I think it is best to go that way for now.

[vzakhari]

Removed the check for hasRealControl. Thanks, Val!

[vzakhari]

Did some code restructuring after discussing it with Val.

[vdonaldson]

Thanks Slava

flang/lib/Lower/Bridge.cpp

struct IncrementLoopInfo {
   bool isUnordered; // do concurrent, forall
   llvm::SmallVector<const Fortran::semantics::Symbol *> localInitSymList;
   llvm::SmallVector<const Fortran::semantics::Symbol *> sharedSymList;
   mlir::Value loopVariable = nullptr;
   mlir::Value stepValue = nullptr; // possible uses in multiple blocks
 
   // Data members for structured loops.
   fir::DoLoopOp doLoop = nullptr;
+  // Argument block of the DoLoop holding the current value
+  // of the do-variable. It has the same data type as the original
+  // do-variable.
+  mlir::Value doVarValue = nullptr;
 
   // Data members for unstructured loops.
   bool hasRealControl = false;
   mlir::Value tripVariable = nullptr;
   mlir::Block *headerBlock = nullptr; // loop entry and test block
   mlir::Block *maskBlock = nullptr;   // concurrent loop mask block
   mlir::Block *bodyBlock = nullptr;   // first loop body block
   mlir::Block *exitBlock = nullptr;   // loop exit target block
private:
   /// symbol in registeredTypeInfoSymbols.
   bool skipRegistration = false;
   /// Track symbols symbols processed during and after the registration
   /// to avoid infinite loops between type conversions and global variable
   /// creation.
   llvm::SmallSetVector<Fortran::semantics::SymbolRef, 64> seen;
 };
 
 using IncrementLoopNestInfo = llvm::SmallVector<IncrementLoopInfo>;
 } // namespace
 
 //===----------------------------------------------------------------------===//
 // FirConverter
 //===----------------------------------------------------------------------===//
 
 namespace {
 
for (IncrementLoopInfo &info : incrementLoopNestInfo) {
       info.loopVariable =
           genLoopVariableAddress(loc, info.loopVariableSym, info.isUnordered);
       mlir::Value lowerValue = genControlValue(info.lowerExpr, info);
       mlir::Value upperValue = genControlValue(info.upperExpr, info);
       info.stepValue = genControlValue(info.stepExpr, info);
 
       // Structured loop - generate fir.do_loop.
       if (info.isStructured()) {
+        // Keep the original data type of integer do-variables
+        // for structured loops. When do-variable's data type
+        // is an integer type shorter than IndexType, processing
+        // the do-variable separately from the DoLoop's iteration index
+        // allows getting rid of type casts, which can make backend
+        // optimizations easier.
+        //
+        // In order to make backends' life easier, we can add an extra
+        // iteration variable in the DoLoop's iter_args() list. This extra
+        // variable has the same data type as the original do-variable.
+        //
+        // We could have disabled an extra iteration variable usage
+        // for cases when its data type is not shorter than IndexType,
+        // but there seem to be no way to get IndexType's bit width.
+        mlir::Value doVarInit = nullptr;
+        if (!info.isUnordered && !info.hasRealControl)
+          doVarInit = builder->createConvert(loc, info.getLoopVariableType(),
+                                             lowerValue);
+
         info.doLoop = builder->create<fir::DoLoopOp>(
             loc, lowerValue, upperValue, info.stepValue, info.isUnordered,
-            /*finalCountValue=*/!info.isUnordered);
+            /*finalCountValue=*/!info.isUnordered,
+            doVarInit ? mlir::ValueRange{doVarInit} : mlir::ValueRange{});
         builder->setInsertionPointToStart(info.doLoop.getBody());
-        // Update the loop variable value, as it may have non-index references.
-        mlir::Value value = builder->createConvert(
-            loc, info.getLoopVariableType(), info.doLoop.getInductionVar());
+        mlir::Value value;
+        if (!doVarInit) {
+          // Update the loop variable value, as it may have non-index
+          // references.
+          value = builder->createConvert(loc, info.getLoopVariableType(),
+                                         info.doLoop.getInductionVar());
+        } else {
+          // The loop variable value is the region's argument rather
+          // than the DoLoop's index value.
+          value = info.doLoop.getRegionIterArgs()[0];
+          info.doVarValue = value;
+        }
         builder->create<fir::StoreOp>(loc, value, info.loopVariable);
         if (info.maskExpr) {
           Fortran::lower::StatementContext stmtCtx;
           mlir::Value maskCond = createFIRExpr(loc, info.maskExpr, stmtCtx);
           stmtCtx.finalize();
           mlir::Value maskCondCast =
               builder->createConvert(loc, builder->getI1Type(), maskCond);
           auto ifOp = builder->create<fir::IfOp>(loc, maskCondCast,
void genFIRIncrementLoopEnd(IncrementLoopNestInfo &incrementLoopNestInfo) {
     for (auto it = incrementLoopNestInfo.rbegin(),
               rend = incrementLoopNestInfo.rend();
          it != rend; ++it) {
       IncrementLoopInfo &info = *it;
       if (info.isStructured()) {
         // End fir.do_loop.
         if (!info.isUnordered) {
           builder->setInsertionPointToEnd(info.doLoop.getBody());
-          mlir::Value result = builder->create<mlir::arith::AddIOp>(
-              loc, info.doLoop.getInductionVar(), info.doLoop.getStep());
-          builder->create<fir::ResultOp>(loc, result);
+          llvm::SmallVector<mlir::Value, 2> results;
+          results.push_back(builder->create<mlir::arith::AddIOp>(
+              loc, info.doLoop.getInductionVar(), info.doLoop.getStep()));
+          if (info.doVarValue) {
+            // If we use an extra iteration variable of the same data
+            // type as the original do-variable, we have to increment
+            // it by the step value. Note that the step has 'index'
+            // type, so we need to cast it, first.
+            mlir::Value stepCast = builder->createConvert(
+                loc, info.getLoopVariableType(), info.doLoop.getStep());
+            results.push_back(builder->create<mlir::arith::AddIOp>(
+                loc, info.doVarValue, stepCast));
+          }
+          builder->create<fir::ResultOp>(loc, results);
         }
         builder->setInsertionPointAfter(info.doLoop);
         if (info.isUnordered)
           continue;
         // The loop control variable may be used after loop execution.
-        mlir::Value lcv = builder->createConvert(
-            loc, info.getLoopVariableType(), info.doLoop.getResult(0));
+        mlir::Value lcv = nullptr;
+        if (info.doVarValue) {
+          // Final do-variable value is the second result of the DoLoop.
+          assert(info.doLoop.getResults().size() == 2 &&
+                 "invalid do-variable handling");
+          lcv = info.doLoop.getResult(1);
+        } else {
+          lcv = builder->createConvert(loc, info.getLoopVariableType(),
+                                       info.doLoop.getResult(0));
+        }
         builder->create<fir::StoreOp>(loc, lcv, info.loopVariable);
         continue;
       }
 
       // Unstructured loop - decrement tripVariable and step loopVariable.
       mlir::Value tripCount =
           builder->create<fir::LoadOp>(loc, info.tripVariable);
       mlir::Value one =

flang/test/Lower/OpenMP/omp-parallel-private-clause-fixes.f90

 ! CHECK:           %[[VAL_5:.*]] = arith.constant 1 : i32
 ! CHECK:           omp.wsloop for (%[[VAL_6:.*]]) : i32 = (%[[ONE]]) to (%[[VAL_3]]) inclusive step (%[[VAL_5]]) {
 ! CHECK:             fir.store %[[VAL_6]] to %[[PRIV_I]] : !fir.ref<i32>
 ! CHECK:             %[[VAL_7:.*]] = arith.constant 1 : i32
 ! CHECK:             %[[VAL_8:.*]] = fir.convert %[[VAL_7]] : (i32) -> index
 ! CHECK:             %[[VAL_9:.*]] = fir.load %[[VAL_4]] : !fir.ref<i32>
 ! CHECK:             %[[VAL_10:.*]] = fir.convert %[[VAL_9]] : (i32) -> index
 ! CHECK:             %[[VAL_11:.*]] = arith.constant 1 : index
-! CHECK:             %[[VAL_12:.*]] = fir.do_loop %[[VAL_13:.*]] = %[[VAL_8]] to %[[VAL_10]] step %[[VAL_11]] -> index {
-! CHECK:               %[[VAL_14:.*]] = fir.convert %[[VAL_13]] : (index) -> i32
-! CHECK:               fir.store %[[VAL_14]] to %[[PRIV_J]] : !fir.ref<i32>
+! CHECK:             %[[LB:.*]] = fir.convert %[[VAL_8]] : (index) -> i32
+! CHECK:             %[[VAL_12:.*]]:2 = fir.do_loop %[[VAL_13:[^ ]*]] =
+! CHECK-SAME:            %[[VAL_8]] to %[[VAL_10]] step %[[VAL_11]]
+! CHECK-SAME:            iter_args(%[[IV:.*]] = %[[LB]]) -> (index, i32) {
+! CHECK:               fir.store %[[IV]] to %[[PRIV_J]] : !fir.ref<i32>
 ! CHECK:               %[[LOAD:.*]] = fir.load %[[PRIV_I]] : !fir.ref<i32>
 ! CHECK:               %[[VAL_15:.*]] = fir.load %[[PRIV_J]] : !fir.ref<i32>
 ! CHECK:               %[[VAL_16:.*]] = arith.addi %[[LOAD]], %[[VAL_15]] : i32
 ! CHECK:               fir.store %[[VAL_16]] to %[[PRIV_X]] : !fir.ref<i32>
 ! CHECK:               %[[VAL_17:.*]] = arith.addi %[[VAL_13]], %[[VAL_11]] : index
-! CHECK:               fir.result %[[VAL_17]] : index
+! CHECK:               %[[STEPCAST:.*]] = fir.convert %[[VAL_11]] : (index) -> i32
+! CHECK:               %[[IVINC:.*]] = arith.addi %[[IV]], %[[STEPCAST]]
+! CHECK:               fir.result %[[VAL_17]], %[[IVINC]] : index, i32
 ! CHECK:             }
-! CHECK:             %[[VAL_18:.*]] = fir.convert %[[VAL_19:.*]] : (index) -> i32
-! CHECK:             fir.store %[[VAL_18]] to %[[PRIV_J]] : !fir.ref<i32>
+! CHECK:             fir.store %[[VAL_12]]#1 to %[[PRIV_J]] : !fir.ref<i32>
 ! CHECK:             omp.yield
 ! CHECK:           }
 ! CHECK:           omp.terminator
 ! CHECK:         }
 ! CHECK:         return
 subroutine multiple_private_fix(gama)
         integer :: i, j, x, gama
 !$OMP PARALLEL DO PRIVATE(j,x)

flang/test/Lower/OpenMP/omp-wsloop-variable.f90

 !CHECK:         omp.wsloop   for  (%[[ARG0:.*]]) : i32 = (%[[VAL_7]]) to (%[[VAL_10]]) inclusive step (%[[VAL_11]]) {
 !CHECK:           fir.store %[[ARG0]] to %[[STORE_IV:.*]] : !fir.ref<i32>
 !CHECK:           %[[VAL_13:.*]] = fir.load %[[VAL_0]] : !fir.ref<i128>
 !CHECK:           %[[VAL_14:.*]] = fir.convert %[[VAL_13]] : (i128) -> index
 !CHECK:           %[[VAL_15:.*]] = arith.constant 100 : i32
 !CHECK:           %[[VAL_16:.*]] = fir.convert %[[VAL_15]] : (i32) -> index
 !CHECK:           %[[VAL_17:.*]] = fir.load %[[VAL_4]] : !fir.ref<i32>
 !CHECK:           %[[VAL_18:.*]] = fir.convert %[[VAL_17]] : (i32) -> index
-!CHECK:           %[[VAL_19:.*]] = fir.do_loop %[[VAL_20:.*]] = %[[VAL_14]] to %[[VAL_16]] step %[[VAL_18]] -> index {
-!CHECK:             %[[VAL_21:.*]] = fir.convert %[[VAL_20]] : (index) -> i64
-!CHECK:             fir.store %[[VAL_21]] to %[[VAL_5]] : !fir.ref<i64>
+!CHECK:           %[[LB:.*]] = fir.convert %[[VAL_14]] : (index) -> i64
+!CHECK:           %[[VAL_19:.*]]:2 = fir.do_loop %[[VAL_20:[^ ]*]] =
+!CHECK-SAME:          %[[VAL_14]] to %[[VAL_16]] step %[[VAL_18]]
+!CHECK-SAME:          iter_args(%[[IV:.*]] = %[[LB]]) -> (index, i64) {
+!CHECK:             fir.store %[[IV]] to %[[VAL_5]] : !fir.ref<i64>
 !CHECK:             %[[LOAD_IV:.*]] = fir.load %[[STORE_IV]] : !fir.ref<i32>
 !CHECK:             %[[VAL_22:.*]] = fir.convert %[[LOAD_IV]] : (i32) -> i64
 !CHECK:             %[[VAL_23:.*]] = fir.load %[[VAL_5]] : !fir.ref<i64>
 !CHECK:             %[[VAL_24:.*]] = arith.addi %[[VAL_22]], %[[VAL_23]] : i64
 !CHECK:             %[[VAL_25:.*]] = fir.convert %[[VAL_24]] : (i64) -> f32
 !CHECK:             fir.store %[[VAL_25]] to %[[VAL_6]] : !fir.ref<f32>
 !CHECK:             %[[VAL_26:.*]] = arith.addi %[[VAL_20]], %[[VAL_18]] : index
-!CHECK:             fir.result %[[VAL_26]] : index
+!CHECK:             %[[STEPCAST:.*]] = fir.convert %[[VAL_18]] : (index) -> i64
+!CHECK:             %[[IVINC:.*]] = arith.addi %[[IV]], %[[STEPCAST]]
+!CHECK:             fir.result %[[VAL_26]], %[[IVINC]] : index, i64
 !CHECK:           }
-!CHECK:           %[[VAL_27:.*]] = fir.convert %[[VAL_28:.*]] : (index) -> i64
-!CHECK:           fir.store %[[VAL_27]] to %[[VAL_5]] : !fir.ref<i64>
+!CHECK:           fir.store %[[VAL_19]]#1 to %[[VAL_5]] : !fir.ref<i64>
 !CHECK:           omp.yield
 !CHECK:         }
 !CHECK:         return
 !CHECK:       }
 
 subroutine wsloop_variable_sub
   integer(kind=1) :: i1_ub
   integer(kind=2) :: i2, i2_s

flang/test/Lower/array-expression-slice-1.f90

 ! CHECK-DAG:         %[[VAL_25:.*]] = fir.address_of(@_QFEa1) : !fir.ref<!fir.array<10x10xf32>>
 ! CHECK-DAG:         %[[VAL_26:.*]] = fir.address_of(@_QFEa2) : !fir.ref<!fir.array<3xf32>>
 ! CHECK-DAG:         %[[VAL_27:.*]] = fir.address_of(@_QFEa3) : !fir.ref<!fir.array<10xf32>>
 ! CHECK-DAG:         %[[VAL_28:.*]] = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFEi"}
 ! CHECK-DAG:         %[[VAL_29:.*]] = fir.address_of(@_QFEiv) : !fir.ref<!fir.array<3xi32>>
 ! CHECK-DAG:         %[[VAL_30:.*]] = fir.alloca i32 {bindc_name = "j", uniq_name = "_QFEj"}
 ! CHECK-DAG:         %[[VAL_31:.*]] = fir.alloca i32 {bindc_name = "k", uniq_name = "_QFEk"}
 ! CHECK:         fir.store %[[VAL_24]] to %[[VAL_31]] : !fir.ref<i32>
-! CHECK:         br ^bb1(%[[VAL_5]], %[[VAL_0]] : index, index)
-! CHECK:       ^bb1(%[[VAL_32:.*]]: index, %[[VAL_33:.*]]: index):
+! CHECK:         %[[LB:.*]] = fir.convert %[[VAL_5]] : (index) -> i32
+! CHECK:         br ^bb1(%[[LB]], %[[VAL_0]] : i32, index)
+! CHECK:       ^bb1(%[[VAL_32:.*]]: i32, %[[VAL_33:.*]]: index):
 ! CHECK:         %[[VAL_34:.*]] = arith.cmpi sgt, %[[VAL_33]], %[[VAL_6]] : index
 ! CHECK:         cond_br %[[VAL_34]], ^bb2, ^bb6
 ! CHECK:       ^bb2:
-! CHECK:         %[[VAL_35:.*]] = fir.convert %[[VAL_32]] : (index) -> i32
-! CHECK:         fir.store %[[VAL_35]] to %[[VAL_30]] : !fir.ref<i32>
-! CHECK:         br ^bb3(%[[VAL_5]], %[[VAL_0]] : index, index)
-! CHECK:       ^bb3(%[[VAL_36:.*]]: index, %[[VAL_37:.*]]: index):
+! CHECK:         fir.store %[[VAL_32]] to %[[VAL_30]] : !fir.ref<i32>
+! CHECK:         br ^bb3(%[[LB]], %[[VAL_0]] : i32, index)
+! CHECK:       ^bb3(%[[VAL_36:.*]]: i32, %[[VAL_37:.*]]: index):
 ! CHECK:         %[[VAL_38:.*]] = arith.cmpi sgt, %[[VAL_37]], %[[VAL_6]] : index
 ! CHECK:         cond_br %[[VAL_38]], ^bb4, ^bb5
 ! CHECK:       ^bb4:
-! CHECK:         %[[VAL_39:.*]] = fir.convert %[[VAL_36]] : (index) -> i32
-! CHECK:         fir.store %[[VAL_39]] to %[[VAL_28]] : !fir.ref<i32>
+! CHECK:         fir.store %[[VAL_36]] to %[[VAL_28]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_40:.*]] = fir.load %[[VAL_31]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_41:.*]] = arith.addi %[[VAL_40]], %[[VAL_23]] : i32
 ! CHECK:         fir.store %[[VAL_41]] to %[[VAL_31]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_42:.*]] = fir.load %[[VAL_31]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_43:.*]] = fir.convert %[[VAL_42]] : (i32) -> f32
 ! CHECK:         %[[VAL_44:.*]] = fir.call @fir.cos.f32.f32(%[[VAL_43]]) : (f32) -> f32
 ! CHECK:         %[[VAL_45:.*]] = fir.load %[[VAL_28]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_46:.*]] = fir.convert %[[VAL_45]] : (i32) -> i64
 ! CHECK:         %[[VAL_47:.*]] = arith.subi %[[VAL_46]], %[[VAL_20]] : i64
 ! CHECK:         %[[VAL_48:.*]] = fir.load %[[VAL_30]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_49:.*]] = fir.convert %[[VAL_48]] : (i32) -> i64
 ! CHECK:         %[[VAL_50:.*]] = arith.subi %[[VAL_49]], %[[VAL_20]] : i64
 ! CHECK:         %[[VAL_51:.*]] = fir.coordinate_of %[[VAL_25]], %[[VAL_47]], %[[VAL_50]] : (!fir.ref<!fir.array<10x10xf32>>, i64, i64) -> !fir.ref<f32>
 ! CHECK:         fir.store %[[VAL_44]] to %[[VAL_51]] : !fir.ref<f32>
-! CHECK:         %[[VAL_52:.*]] = arith.addi %[[VAL_36]], %[[VAL_5]] : index
+! CHECK:         %[[VAL_52:.*]] = arith.addi %[[VAL_36]], %[[LB]] : i32
 ! CHECK:         %[[VAL_53:.*]] = arith.subi %[[VAL_37]], %[[VAL_5]] : index
-! CHECK:         br ^bb3(%[[VAL_52]], %[[VAL_53]] : index, index)
+! CHECK:         br ^bb3(%[[VAL_52]], %[[VAL_53]] : i32, index)
 ! CHECK:       ^bb5:
-! CHECK:         %[[VAL_54:.*]] = fir.convert %[[VAL_36]] : (index) -> i32
-! CHECK:         fir.store %[[VAL_54]] to %[[VAL_28]] : !fir.ref<i32>
+! CHECK:         fir.store %[[VAL_36]] to %[[VAL_28]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_55:.*]] = fir.load %[[VAL_31]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_56:.*]] = fir.convert %[[VAL_55]] : (i32) -> f32
 ! CHECK:         %[[VAL_57:.*]] = fir.call @fir.sin.f32.f32(%[[VAL_56]]) : (f32) -> f32
 ! CHECK:         %[[VAL_58:.*]] = fir.load %[[VAL_30]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_59:.*]] = fir.convert %[[VAL_58]] : (i32) -> i64
 ! CHECK:         %[[VAL_60:.*]] = arith.subi %[[VAL_59]], %[[VAL_20]] : i64
 ! CHECK:         %[[VAL_61:.*]] = fir.coordinate_of %[[VAL_27]], %[[VAL_60]] : (!fir.ref<!fir.array<10xf32>>, i64) -> !fir.ref<f32>
 ! CHECK:         fir.store %[[VAL_57]] to %[[VAL_61]] : !fir.ref<f32>
-! CHECK:         %[[VAL_62:.*]] = arith.addi %[[VAL_32]], %[[VAL_5]] : index
+! CHECK:         %[[VAL_62:.*]] = arith.addi %[[VAL_32]], %[[LB]] : i32
 ! CHECK:         %[[VAL_63:.*]] = arith.subi %[[VAL_33]], %[[VAL_5]] : index
-! CHECK:         br ^bb1(%[[VAL_62]], %[[VAL_63]] : index, index)
+! CHECK:         br ^bb1(%[[VAL_62]], %[[VAL_63]] : i32, index)
 ! CHECK:       ^bb6:
-! CHECK:         %[[VAL_64:.*]] = fir.convert %[[VAL_32]] : (index) -> i32
-! CHECK:         fir.store %[[VAL_64]] to %[[VAL_30]] : !fir.ref<i32>
+! CHECK:         fir.store %[[VAL_32]] to %[[VAL_30]] : !fir.ref<i32>
 ! CHECK:         %[[VAL_65:.*]] = fir.shape %[[VAL_11]] : (index) -> !fir.shape<1>
 ! CHECK:         %[[VAL_66:.*]] = fir.undefined index
 ! CHECK:         %[[VAL_67:.*]] = fir.shape %[[VAL_0]], %[[VAL_0]] : (index, index) -> !fir.shape<2>
 ! CHECK:         %[[VAL_68:.*]] = fir.slice %[[VAL_16]], %[[VAL_66]], %[[VAL_66]], %[[VAL_4]], %[[VAL_0]], %[[VAL_11]] : (i64, index, index, index, index, index) -> !fir.slice<2>
 ! CHECK:         br ^bb7(%[[VAL_6]], %[[VAL_11]] : index, index)
 ! CHECK:       ^bb7(%[[VAL_69:.*]]: index, %[[VAL_70:.*]]: index):
 ! CHECK:         %[[VAL_71:.*]] = arith.cmpi sgt, %[[VAL_70]], %[[VAL_6]] : index
 ! CHECK:         cond_br %[[VAL_71]], ^bb8, ^bb9

flang/test/Lower/do_loop.f90

subroutine simple_loop
   ! CHECK: %[[I_REF:.*]] = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFsimple_loopEi"}
   integer :: i
 
   ! CHECK: %[[C1:.*]] = arith.constant 1 : i32
   ! CHECK: %[[C1_CVT:.*]] = fir.convert %c1_i32 : (i32) -> index
   ! CHECK: %[[C5:.*]] = arith.constant 5 : i32
   ! CHECK: %[[C5_CVT:.*]] = fir.convert %c5_i32 : (i32) -> index
   ! CHECK: %[[C1:.*]] = arith.constant 1 : index
-  ! CHECK: %[[LI_RES:.*]] = fir.do_loop %[[LI:.*]] = %[[C1_CVT]] to %[[C5_CVT]] step %[[C1]] -> index {
+  ! CHECK: %[[LB:.*]] = fir.convert %[[C1_CVT]] : (index) -> i32
+  ! CHECK: %[[LI_RES:.*]]:2 = fir.do_loop %[[LI:[^ ]*]] =
+  ! CHECK-SAME: %[[C1_CVT]] to %[[C5_CVT]] step %[[C1]]
+  ! CHECK-SAME: iter_args(%[[IV:.*]] = %[[LB]]) -> (index, i32) {
   do i=1,5
-  ! CHECK:   %[[LI_CVT:.*]] = fir.convert %[[LI]] : (index) -> i32
-  ! CHECK:   fir.store %[[LI_CVT]] to %[[I_REF]] : !fir.ref<i32>
+  ! CHECK:   fir.store %[[IV]] to %[[I_REF]] : !fir.ref<i32>
   ! CHECK:   %[[LI_NEXT:.*]] = arith.addi %[[LI]], %[[C1]] : index
-  ! CHECK:  fir.result %[[LI_NEXT]] : index
+  ! CHECK:   %[[STEPCAST:.*]] = fir.convert %[[C1]] : (index) -> i32
+  ! CHECK:   %[[IVINC:.*]] = arith.addi %[[IV]], %[[STEPCAST]] : i32
+  ! CHECK:  fir.result %[[LI_NEXT]], %[[IVINC]] : index, i32
   ! CHECK: }
   end do
-  ! CHECK: %[[LI_RES_CVT:.*]] = fir.convert %[[LI_RES]] : (index) -> i32
-  ! CHECK: fir.store %[[LI_RES_CVT]] to %[[I_REF]] : !fir.ref<i32>
+  ! CHECK: fir.store %[[LI_RES]]#1 to %[[I_REF]] : !fir.ref<i32>
   ! CHECK: %[[I:.*]] = fir.load %[[I_REF]] : !fir.ref<i32>
   ! CHECK: %{{.*}} = fir.call @_FortranAioOutputInteger32(%{{.*}}, %[[I]]) : (!fir.ref<i8>, i32) -> i1
   print *, i
 end subroutine
 
 ! Test a 2-nested loop with a body composed of a reduction. Values are read from a 2d array.
 ! CHECK-LABEL: nested_loop
 subroutine nested_loop
   ! CHECK: %[[ARR_REF:.*]] = fir.alloca !fir.array<5x5xi32> {bindc_name = "arr", uniq_name = "_QFnested_loopEarr"}
   ! CHECK: %[[ASUM_REF:.*]] = fir.alloca i32 {bindc_name = "asum", uniq_name = "_QFnested_loopEasum"}
   ! CHECK: %[[I_REF:.*]] = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFnested_loopEi"}
   ! CHECK: %[[J_REF:.*]] = fir.alloca i32 {bindc_name = "j", uniq_name = "_QFnested_loopEj"}
   integer :: asum, arr(5,5)
   integer :: i, j
   asum = 0
   ! CHECK: %[[S_I:.*]] = arith.constant 1 : i32
   ! CHECK: %[[S_I_CVT:.*]] = fir.convert %[[S_I]] : (i32) -> index
   ! CHECK: %[[E_I:.*]] = arith.constant 5 : i32
   ! CHECK: %[[E_I_CVT:.*]] = fir.convert %[[E_I]] : (i32) -> index
   ! CHECK: %[[ST_I:.*]] = arith.constant 1 : index
-  ! CHECK: %[[I_RES:.*]] = fir.do_loop %[[LI:.*]] = %[[S_I_CVT]] to %[[E_I_CVT]] step %[[ST_I]] -> index {
+  ! CHECK: %[[I_LB:.*]] = fir.convert %[[S_I_CVT]] : (index) -> i32
+  ! CHECK: %[[I_RES:.*]]:2 = fir.do_loop %[[LI:[^ ]*]] =
+  ! CHECK-SAME: %[[S_I_CVT]] to %[[E_I_CVT]] step %[[ST_I]]
+  ! CHECK-SAME: iter_args(%[[I_IV:.*]] = %[[I_LB]]) -> (index, i32) {
   do i=1,5
-    ! CHECK: %[[LI_CVT:.*]] = fir.convert %[[LI]] : (index) -> i32
-    ! CHECK: fir.store %[[LI_CVT]] to %[[I_REF]] : !fir.ref<i32>
+    ! CHECK: fir.store %[[I_IV]] to %[[I_REF]] : !fir.ref<i32>
     ! CHECK: %[[S_J:.*]] = arith.constant 1 : i32
     ! CHECK: %[[S_J_CVT:.*]] = fir.convert %[[S_J]] : (i32) -> index
     ! CHECK: %[[E_J:.*]] = arith.constant 5 : i32
     ! CHECK: %[[E_J_CVT:.*]] = fir.convert %[[E_J]] : (i32) -> index
     ! CHECK: %[[ST_J:.*]] = arith.constant 1 : index
-    ! CHECK: %[[J_RES:.*]] = fir.do_loop %[[LJ:.*]] = %[[S_J_CVT]] to %[[E_J_CVT]] step %[[ST_J]] -> index {
+    ! CHECK: %[[J_LB:.*]] = fir.convert %[[S_J_CVT]] : (index) -> i32
+    ! CHECK: %[[J_RES:.*]]:2 = fir.do_loop %[[LJ:[^ ]*]] =
+    ! CHECK-SAME: %[[S_J_CVT]] to %[[E_J_CVT]] step %[[ST_J]]
+    ! CHECK-SAME: iter_args(%[[J_IV:.*]] = %[[J_LB]]) -> (index, i32) {
     do j=1,5
-      ! CHECK: %[[LJ_CVT:.*]] = fir.convert %[[LJ]] : (index) -> i32
-      ! CHECK: fir.store %[[LJ_CVT]] to %[[J_REF]] : !fir.ref<i32>
+      ! CHECK: fir.store %[[J_IV]] to %[[J_REF]] : !fir.ref<i32>
       ! CHECK: %[[ASUM:.*]] = fir.load %[[ASUM_REF]] : !fir.ref<i32>
       ! CHECK: %[[I:.*]] = fir.load %[[I_REF]] : !fir.ref<i32>
       ! CHECK: %[[I_CVT:.*]] = fir.convert %[[I]] : (i32) -> i64
       ! CHECK: %[[C1_I:.*]] = arith.constant 1 : i64
       ! CHECK: %[[I_INDX:.*]] = arith.subi %[[I_CVT]], %[[C1_I]] : i64
       ! CHECK: %[[J:.*]] = fir.load %[[J_REF]] : !fir.ref<i32>
       ! CHECK: %[[J_CVT:.*]] = fir.convert %[[J]] : (i32) -> i64
       ! CHECK: %[[C1_J:.*]] = arith.constant 1 : i64
       ! CHECK: %[[J_INDX:.*]] = arith.subi %[[J_CVT]], %[[C1_J]] : i64
       ! CHECK: %[[ARR_IJ_REF:.*]] = fir.coordinate_of %[[ARR_REF]], %[[I_INDX]], %[[J_INDX]] : (!fir.ref<!fir.array<5x5xi32>>, i64, i64) -> !fir.ref<i32>
       ! CHECK: %[[ARR_VAL:.*]] = fir.load %[[ARR_IJ_REF]] : !fir.ref<i32>
       ! CHECK: %[[ASUM_NEW:.*]] = arith.addi %[[ASUM]], %[[ARR_VAL]] : i32
       ! CHECK: fir.store %[[ASUM_NEW]] to %[[ASUM_REF]] : !fir.ref<i32>
       asum = asum + arr(i,j)
       ! CHECK: %[[LJ_NEXT:.*]] = arith.addi %[[LJ]], %[[ST_J]] : index
-      ! CHECK: fir.result %[[LJ_NEXT]] : index
+      ! CHECK: %[[J_STEPCAST:.*]] = fir.convert %[[ST_J]] : (index) -> i32
+      ! CHECK: %[[J_IVINC:.*]] = arith.addi %[[J_IV]], %[[J_STEPCAST]] : i32
+      ! CHECK: fir.result %[[LJ_NEXT]], %[[J_IVINC]] : index, i32
     ! CHECK: }
     end do
-    ! CHECK: %[[J_RES_CVT:.*]] = fir.convert %[[J_RES]] : (index) -> i32
-    ! CHECK: fir.store %[[J_RES_CVT]] to %[[J_REF]] : !fir.ref<i32>
+    ! CHECK: fir.store %[[J_RES]]#1 to %[[J_REF]] : !fir.ref<i32>
     ! CHECK: %[[LI_NEXT:.*]] = arith.addi %[[LI]], %[[ST_I]] : index
-    ! CHECK: fir.result %[[LI_NEXT]] : index
+    ! CHECK: %[[I_STEPCAST:.*]] = fir.convert %[[ST_I]] : (index) -> i32
+    ! CHECK: %[[I_IVINC:.*]] = arith.addi %[[I_IV]], %[[I_STEPCAST]] : i32
+    ! CHECK: fir.result %[[LI_NEXT]], %[[I_IVINC]] : index, i32
   ! CHECK: }
   end do
-  ! CHECK: %[[I_RES_CVT:.*]] = fir.convert %[[I_RES]] : (index) -> i32
-  ! CHECK: fir.store %[[I_RES_CVT]] to %[[I_REF]] : !fir.ref<i32>
+  ! CHECK: fir.store %[[I_RES]]#1 to %[[I_REF]] : !fir.ref<i32>
 end subroutine
 
 ! Test a downcounting loop
 ! CHECK-LABEL: down_counting_loop
 subroutine down_counting_loop()
   integer :: i
   ! CHECK: %[[I_REF:.*]] = fir.alloca i32 {bindc_name = "i", uniq_name = "_QFdown_counting_loopEi"}
 
   ! CHECK: %[[C5:.*]] = arith.constant 5 : i32
   ! CHECK: %[[C5_CVT:.*]] = fir.convert %[[C5]] : (i32) -> index
   ! CHECK: %[[C1:.*]] = arith.constant 1 : i32
   ! CHECK: %[[C1_CVT:.*]] = fir.convert %[[C1]] : (i32) -> index
   ! CHECK: %[[CMINUS1:.*]] = arith.constant -1 : i32
   ! CHECK: %[[CMINUS1_STEP_CVT:.*]] = fir.convert %[[CMINUS1]] : (i32) -> index
-  ! CHECK: %[[I_RES:.*]] = fir.do_loop %[[LI:.*]] = %[[C5_CVT]] to %[[C1_CVT]] step %[[CMINUS1_STEP_CVT]] -> index {
+  ! CHECK: %[[I_LB:.*]] = fir.convert %[[C5_CVT]] : (index) -> i32
+  ! CHECK: %[[I_RES:.*]]:2 = fir.do_loop %[[LI:[^ ]*]] =
+  ! CHECK-SAME: %[[C5_CVT]] to %[[C1_CVT]] step %[[CMINUS1_STEP_CVT]]
+  ! CHECK-SAME: iter_args(%[[I_IV:.*]] = %[[I_LB]]) -> (index, i32) {
   do i=5,1,-1
-  ! CHECK: %[[LI_CVT:.*]] = fir.convert %[[LI]] : (index) -> i32
-  ! CHECK: fir.store %[[LI_CVT]] to %[[I_REF]] : !fir.ref<i32>
+  ! CHECK: fir.store %[[I_IV]] to %[[I_REF]] : !fir.ref<i32>
   ! CHECK: %[[LI_NEXT:.*]] = arith.addi %[[LI]], %[[CMINUS1_STEP_CVT]] : index
-  ! CHECK: fir.result %[[LI_NEXT]] : index
+  ! CHECK: %[[I_STEPCAST:.*]] = fir.convert %[[CMINUS1_STEP_CVT]] : (index) -> i32
+  ! CHECK: %[[I_IVINC:.*]] = arith.addi %[[I_IV]], %[[I_STEPCAST]] : i32
+  ! CHECK: fir.result %[[LI_NEXT]], %[[I_IVINC]] : index, i32
   ! CHECK: }
   end do
-  ! CHECK: %[[I_RES_CVT:.*]] = fir.convert %[[I_RES]] : (index) -> i32
-  ! CHECK: fir.store %[[I_RES_CVT]] to %[[I_REF]] : !fir.ref<i32>
+  ! CHECK: fir.store %[[I_RES]]#1 to %[[I_REF]] : !fir.ref<i32>
 end subroutine
 
 ! Test a general loop with a variable step
 ! CHECK-LABEL: loop_with_variable_step
 ! CHECK-SAME: (%[[S_REF:.*]]: !fir.ref<i32> {fir.bindc_name = "s"}, %[[E_REF:.*]]: !fir.ref<i32> {fir.bindc_name = "e"}, %[[ST_REF:.*]]: !fir.ref<i32> {fir.bindc_name = "st"}) {
 subroutine loop_with_variable_step(s,e,st)
   integer :: s, e, st
   ! CHECK: %[[S:.*]] = fir.load %[[S_REF]] : !fir.ref<i32>
   ! CHECK: %[[S_CVT:.*]] = fir.convert %[[S]] : (i32) -> index
   ! CHECK: %[[E:.*]] = fir.load %[[E_REF]] : !fir.ref<i32>
   ! CHECK: %[[E_CVT:.*]] = fir.convert %[[E]] : (i32) -> index
   ! CHECK: %[[ST:.*]] = fir.load %[[ST_REF]] : !fir.ref<i32>
   ! CHECK: %[[ST_CVT:.*]] = fir.convert %[[ST]] : (i32) -> index
-  ! CHECK: %[[I_RES:.*]] = fir.do_loop %[[LI:.*]] = %[[S_CVT]] to %[[E_CVT]] step %[[ST_CVT]] -> index {
+  ! CHECK: %[[I_LB:.*]] = fir.convert %[[S_CVT]] : (index) -> i32
+  ! CHECK: %[[I_RES:.*]]:2 = fir.do_loop %[[LI:[^ ]*]] =
+  ! CHECK-SAME: %[[S_CVT]] to %[[E_CVT]] step %[[ST_CVT]]
+  ! CHECK-SAME: iter_args(%[[I_IV:.*]] = %[[I_LB]]) -> (index, i32) {
   do i=s,e,st
-  ! CHECK:  %[[LI_CVT:.*]] = fir.convert %[[LI]] : (index) -> i32
-  ! CHECK:  fir.store %[[LI_CVT]] to %[[I_REF]] : !fir.ref<i32>
+  ! CHECK:  fir.store %[[I_IV]] to %[[I_REF]] : !fir.ref<i32>
   ! CHECK:  %[[LI_NEXT:.*]] = arith.addi %[[LI]], %[[ST_CVT]] : index
-  ! CHECK:  fir.result %[[LI_NEXT]] : index
+  ! CHECK: %[[I_STEPCAST:.*]] = fir.convert %[[ST_CVT]] : (index) -> i32
+  ! CHECK: %[[I_IVINC:.*]] = arith.addi %[[I_IV]], %[[I_STEPCAST]] : i32
+  ! CHECK:  fir.result %[[LI_NEXT]], %[[I_IVINC]] : index, i32
   ! CHECK: }
   end do
-  ! CHECK: %[[I_RES_CVT:.*]] = fir.convert %[[I_RES]] : (index) -> i32
-  ! CHECK: fir.store %[[I_RES_CVT]] to %[[I_REF]] : !fir.ref<i32>
+  ! CHECK: fir.store %[[I_RES]]#1 to %[[I_REF]] : !fir.ref<i32>
 end subroutine
 
 ! Test usage of pointer variables as index, start, end and step variables
 ! CHECK-LABEL: loop_with_pointer_variables
 ! CHECK-SAME: (%[[S_REF:.*]]: !fir.ref<i32> {fir.bindc_name = "s", fir.target}, %[[E_REF:.*]]: !fir.ref<i32> {fir.bindc_name = "e", fir.target}, %[[ST_REF:.*]]: !fir.ref<i32> {fir.bindc_name = "st", fir.target}) {
 subroutine loop_with_pointer_variables(s,e,st)
 ! CHECK:  %[[E_PTR_REF:.*]] = fir.alloca !fir.ptr<i32> {uniq_name = "_QFloop_with_pointer_variablesEeptr.addr"}
 ! CHECK:  %[[I_REF:.*]] = fir.alloca i32 {bindc_name = "i", fir.target, uniq_name = "_QFloop_with_pointer_variablesEi"}

 ! CHECK:  %[[S:.*]] = fir.load %[[S_PTR]] : !fir.ptr<i32>
 ! CHECK:  %[[S_CVT:.*]] = fir.convert %[[S]] : (i32) -> index
 ! CHECK:  %[[E_PTR:.*]] = fir.load %[[E_PTR_REF]] : !fir.ref<!fir.ptr<i32>>
 ! CHECK:  %[[E:.*]] = fir.load %[[E_PTR]] : !fir.ptr<i32>
 ! CHECK:  %[[E_CVT:.*]] = fir.convert %[[E]] : (i32) -> index
 ! CHECK:  %[[ST_PTR:.*]] = fir.load %[[ST_PTR_REF]] : !fir.ref<!fir.ptr<i32>>
 ! CHECK:  %[[ST:.*]] = fir.load %[[ST_PTR]] : !fir.ptr<i32>
 ! CHECK:  %[[ST_CVT:.*]] = fir.convert %[[ST]] : (i32) -> index
-! CHECK:  %[[I_RES:.*]] = fir.do_loop %[[LI:.*]] = %[[S_CVT]] to %[[E_CVT]] step %[[ST_CVT]] -> index {
+! CHECK:  %[[I_LB:.*]] = fir.convert %[[S_CVT]] : (index) -> i32
+! CHECK:  %[[I_RES:.*]]:2 = fir.do_loop %[[LI:[^ ]*]] =
+! CHECK-SAME: %[[S_CVT]] to %[[E_CVT]] step %[[ST_CVT]]
+! CHECK-SAME: iter_args(%[[I_IV:.*]] = %[[I_LB]]) -> (index, i32) {
   do iptr=sptr,eptr,stptr
-! CHECK:    %[[LI_CVT:.*]] = fir.convert %[[LI]] : (index) -> i32
-! CHECK:    fir.store %[[LI_CVT]] to %[[I_PTR]] : !fir.ptr<i32>
+! CHECK:    fir.store %[[I_IV]] to %[[I_PTR]] : !fir.ptr<i32>
 ! CHECK:    %[[LI_NEXT:.*]] = arith.addi %[[LI]], %[[ST_CVT]] : index
-! CHECK:    fir.result %[[LI_NEXT]] : index
+! CHECK:    %[[I_STEPCAST:.*]] = fir.convert %[[ST_CVT]] : (index) -> i32
+! CHECK:    %[[I_IVINC:.*]] = arith.addi %[[I_IV]], %[[I_STEPCAST]] : i32
+! CHECK:    fir.result %[[LI_NEXT]], %[[I_IVINC]] : index, i32
   end do
 ! CHECK:  }
-! CHECK:  %[[I_RES_CVT:.*]] = fir.convert %[[I_RES]] : (index) -> i32
-! CHECK:  fir.store %[[I_RES_CVT:.*]] to %[[I_PTR]] : !fir.ptr<i32>
+! CHECK:  fir.store %[[I_RES]]#1 to %[[I_PTR]] : !fir.ptr<i32>
 end subroutine
 
 ! Test usage of non-default integer kind for loop control and loop index variable
 ! CHECK-LABEL: loop_with_non_default_integer
 ! CHECK-SAME: (%[[S_REF:.*]]: !fir.ref<i64> {fir.bindc_name = "s"}, %[[E_REF:.*]]: !fir.ref<i64> {fir.bindc_name = "e"}, %[[ST_REF:.*]]: !fir.ref<i64> {fir.bindc_name = "st"}) {
 subroutine loop_with_non_default_integer(s,e,st)
   ! CHECK: %[[I_REF:.*]] = fir.alloca i64 {bindc_name = "i", uniq_name = "_QFloop_with_non_default_integerEi"}
   integer(kind=8):: i
   ! CHECK: %[[S:.*]] = fir.load %[[S_REF]] : !fir.ref<i64>
   ! CHECK: %[[S_CVT:.*]] = fir.convert %[[S]] : (i64) -> index
   ! CHECK: %[[E:.*]] = fir.load %[[E_REF]] : !fir.ref<i64>
   ! CHECK: %[[E_CVT:.*]] = fir.convert %[[E]] : (i64) -> index
   ! CHECK: %[[ST:.*]] = fir.load %[[ST_REF]] : !fir.ref<i64>
   ! CHECK: %[[ST_CVT:.*]] = fir.convert %[[ST]] : (i64) -> index
   integer(kind=8) :: s, e, st
 
-  ! CHECK: %[[I_RES:.*]] = fir.do_loop %[[LI:.*]] = %[[S_CVT]] to %[[E_CVT]] step %[[ST_CVT]] -> index {
+  ! CHECK: %[[I_LB:.*]] = fir.convert %[[S_CVT]] : (index) -> i64
+  ! CHECK: %[[I_RES:.*]]:2 = fir.do_loop %[[LI:[^ ]*]] =
+  ! CHECK-SAME: %[[S_CVT]] to %[[E_CVT]] step %[[ST_CVT]]
+  ! CHECK-SAME: iter_args(%[[I_IV:.*]] = %[[I_LB]]) -> (index, i64) {
   do i=s,e,st
-    ! CHECK: %[[LI_CVT:.*]] = fir.convert %[[LI]] : (index) -> i64
-    ! CHECK: fir.store %[[LI_CVT]] to %[[I_REF]] : !fir.ref<i64>
+    ! CHECK: fir.store %[[I_IV]] to %[[I_REF]] : !fir.ref<i64>
     ! CHECK: %[[LI_NEXT:.*]] = arith.addi %[[LI]], %[[ST_CVT]] : index
-    ! CHECK: fir.result %[[LI_NEXT]] : index
+    ! CHECK: %[[I_STEPCAST:.*]] = fir.convert %[[ST_CVT]] : (index) -> i64
+    ! CHECK: %[[I_IVINC:.*]] = arith.addi %[[I_IV]], %[[I_STEPCAST]] : i64
+    ! CHECK: fir.result %[[LI_NEXT]], %[[I_IVINC]] : index, i64
   end do
   ! CHECK: }
-  ! CHECK: %[[I_RES_CVT:.*]] = fir.convert %[[I_RES]] : (index) -> i64
-  ! CHECK: fir.store %[[I_RES_CVT]] to %[[I_REF]] : !fir.ref<i64>
+  ! CHECK: fir.store %[[I_RES]]#1 to %[[I_REF]] : !fir.ref<i64>
 end subroutine
 
 ! Test real loop control.
 ! CHECK-LABEL: loop_with_real_control
 ! CHECK-SAME: (%[[S_REF:.*]]: !fir.ref<f32> {fir.bindc_name = "s"}, %[[E_REF:.*]]: !fir.ref<f32> {fir.bindc_name = "e"}, %[[ST_REF:.*]]: !fir.ref<f32> {fir.bindc_name = "st"}) {
 subroutine loop_with_real_control(s,e,st)
   ! CHECK-DAG: %[[INDEX_REF:.*]] = fir.alloca index
   ! CHECK-DAG: %[[X_REF:.*]] = fir.alloca f32 {bindc_name = "x", uniq_name = "_QFloop_with_real_controlEx"}

flang/test/Lower/do_loop_unstructured.f90

 ! CHECK:   fir.store %[[I_START]] to %[[LOOP_VAR_I_REF]] : !fir.ref<i32>
 ! CHECK:   cf.br ^[[HEADER:.*]]
 ! CHECK: ^[[HEADER]]:
 ! CHECK:   %[[TRIP_VAR:.*]] = fir.load %[[TRIP_VAR_I_REF]] : !fir.ref<i32>
 ! CHECK:   %[[ZERO:.*]] = arith.constant 0 : i32
 ! CHECK:   %[[COND:.*]] = arith.cmpi sgt, %[[TRIP_VAR]], %[[ZERO]] : i32
 ! CHECK:   cf.cond_br %[[COND]], ^[[BODY:.*]], ^[[EXIT:.*]]
 ! CHECK: ^[[BODY]]:
-! CHECK:   %{{.*}} = fir.do_loop %[[J_INDEX:.*]] = %{{.*}} to %{{.*}} step %{{.*}} -> index {
-! CHECK:     %[[J_INDEX_CVT:.*]] = fir.convert %[[J_INDEX]] : (index) -> i32
-! CHECK:     fir.store %[[J_INDEX_CVT]] to %[[LOOP_VAR_J_REF]] : !fir.ref<i32>
+! CHECK:   %{{.*}} = fir.do_loop %[[J_INDEX:[^ ]*]] =
+! CHECK-SAME: %{{.*}} to %{{.*}} step %{{[^ ]*}}
+! CHECK-SAME: iter_args(%[[J_IV:.*]] = %{{.*}}) -> (index, i32) {
+! CHECK:     fir.store %[[J_IV]] to %[[LOOP_VAR_J_REF]] : !fir.ref<i32>
 ! CHECK:   }
 ! CHECK:   %[[TRIP_VAR_I:.*]] = fir.load %[[TRIP_VAR_I_REF]] : !fir.ref<i32>
 ! CHECK:   %[[C1_3:.*]] = arith.constant 1 : i32
 ! CHECK:   %[[TRIP_VAR_I_NEXT:.*]] = arith.subi %[[TRIP_VAR_I]], %[[C1_3]] : i32
 ! CHECK:   fir.store %[[TRIP_VAR_I_NEXT]] to %[[TRIP_VAR_I_REF]] : !fir.ref<i32>
 ! CHECK:   %[[LOOP_VAR_I:.*]] = fir.load %[[LOOP_VAR_I_REF]] : !fir.ref<i32>
 ! CHECK:   %[[LOOP_VAR_I_NEXT:.*]] = arith.addi %[[LOOP_VAR_I]], %c1_i32_0 : i32
 ! CHECK:   fir.store %[[LOOP_VAR_I_NEXT]] to %[[LOOP_VAR_I_REF]] : !fir.ref<i32>
 ! CHECK:   cf.br ^[[HEADER]]
 ! CHECK: ^[[EXIT]]:
 ! CHECK:   return

flang/test/Lower/infinite_loop.f90

 ! CHECK: ^[[EXIT]]:
 ! CHECK:  cf.br ^[[RETURN:.*]]
 ! CHECK: ^[[BODY2:.*]]:
 ! CHECK:  %[[C1:.*]] = arith.constant 1 : i32
 ! CHECK:  %[[C1_INDEX:.*]] = fir.convert %[[C1]] : (i32) -> index
 ! CHECK:  %[[C10:.*]] = arith.constant 10 : i32
 ! CHECK:  %[[C10_INDEX:.*]] = fir.convert %[[C10]] : (i32) -> index
 ! CHECK:  %[[C1_1:.*]] = arith.constant 1 : index
-! CHECK:  %[[J_FINAL:.*]] = fir.do_loop %[[J:.*]] = %[[C1_INDEX]] to %[[C10_INDEX]] step %[[C1_1]] -> index {
-! CHECK:    %[[J_I32:.*]] = fir.convert %[[J]] : (index) -> i32
-! CHECK:    fir.store %[[J_I32]] to %[[J_REF]] : !fir.ref<i32>
+! CHECK:  %[[J_LB:.*]] = fir.convert %[[C1_INDEX]] : (index) -> i32
+! CHECK:  %[[J_FINAL:.*]]:2 = fir.do_loop %[[J:[^ ]*]] =
+! CHECK-SAME: %[[C1_INDEX]] to %[[C10_INDEX]] step %[[C1_1]]
+! CHECK-SAME: iter_args(%[[J_IV:.*]] = %[[J_LB]]) -> (index, i32) {
+! CHECK:    fir.store %[[J_IV]] to %[[J_REF]] : !fir.ref<i32>
 ! CHECK:    %[[J_NEXT:.*]] = arith.addi %[[J]], %[[C1_1]] : index
-! CHECK:    fir.result %[[J_NEXT]] : index
+! CHECK:    %[[J_STEPCAST:.*]] = fir.convert %[[C1_1]] : (index) -> i32
+! CHECK:    %[[J_IVINC:.*]] = arith.addi %[[J_IV]], %[[J_STEPCAST]] : i32
+! CHECK:    fir.result %[[J_NEXT]], %[[J_IVINC]] : index, i32
 ! CHECK:  }
-! CHECK:  %[[J_I32:.*]] = fir.convert %[[J_FINAL]] : (index) -> i32
-! CHECK:  fir.store %[[J_I32]] to %[[J_REF]] : !fir.ref<i32>
+! CHECK:  fir.store %[[J_FINAL]]#1 to %[[J_REF]] : !fir.ref<i32>
 ! CHECK:  cf.br ^[[BODY1]]
 ! CHECK: ^[[RETURN]]:
 ! CHECK:   return
 
 subroutine empty_infinite_in_while(i)
   integer :: i
   do while (i .gt. 50)
     do

flang/test/Lower/loops.f90

subroutine loop_test
 
   ! CHECK-COUNT-3: fir.do_loop {{.*}} unordered
   ! CHECK: fir.if
   do concurrent (integer(1)::i=1:5, j=1:5, k=1:5, i.ne.j .and. k.ne.3) shared(a)
     ! CHECK-COUNT-2: fir.coordinate_of
     a(i,j,k) = a(i,j,k) + 1
   enddo
 
-  ! CHECK-COUNT-3: fir.do_loop {{[^un]*}} -> index
+  ! CHECK-COUNT-3: fir.do_loop {{[^un]*}} -> (index, i32)
   asum = 0
   do i=1,5
     do j=1,5
       do k=1,5
         ! CHECK: fir.coordinate_of
         asum = asum + a(i,j,k)
       enddo
     enddo

flang/test/Lower/loops2.f90

 contains
 ! CHECK-LABEL: func @_QMtest_loop_varPtest_pointer
   subroutine test_pointer()
     do i_pointer=1,10
     enddo
 ! CHECK:         %[[VAL_0:.*]] = fir.address_of(@_QMtest_loop_varEi_pointer) : !fir.ref<!fir.box<!fir.ptr<i32>>>
 ! CHECK:         %[[VAL_1:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.ptr<i32>>>
 ! CHECK:         %[[VAL_2:.*]] = fir.box_addr %[[VAL_1]] : (!fir.box<!fir.ptr<i32>>) -> !fir.ptr<i32>
-! CHECK:         fir.do_loop %[[VAL_9:.*]] =
-! CHECK:           %[[VAL_10:.*]] = fir.convert %[[VAL_9]] : (index) -> i32
-! CHECK:           fir.store %[[VAL_10]] to %[[VAL_2]] : !fir.ptr<i32>
+! CHECK:         %[[VAL_9:.*]]:2 = fir.do_loop{{.*}}iter_args(%[[IV:.*]] = {{.*}})
+! CHECK:           fir.store %[[IV]] to %[[VAL_2]] : !fir.ptr<i32>
 ! CHECK:         }
-! CHECK:         %[[VAL_12:.*]] = fir.convert %[[VAL_13:.*]] : (index) -> i32
-! CHECK:         fir.store %[[VAL_12]] to %[[VAL_2]] : !fir.ptr<i32>
+! CHECK:         fir.store %[[VAL_9]]#1 to %[[VAL_2]] : !fir.ptr<i32>
   end subroutine
 
 ! CHECK-LABEL: func @_QMtest_loop_varPtest_allocatable
   subroutine test_allocatable()
     do i_allocatable=1,10
     enddo
 ! CHECK:         %[[VAL_0:.*]] = fir.address_of(@_QMtest_loop_varEi_allocatable) : !fir.ref<!fir.box<!fir.heap<i32>>>
 ! CHECK:         %[[VAL_1:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.heap<i32>>>
 ! CHECK:         %[[VAL_2:.*]] = fir.box_addr %[[VAL_1]] : (!fir.box<!fir.heap<i32>>) -> !fir.heap<i32>
-! CHECK:         fir.do_loop %[[VAL_9:.*]] =
-! CHECK:           %[[VAL_10:.*]] = fir.convert %[[VAL_9]] : (index) -> i32
-! CHECK:           fir.store %[[VAL_10]] to %[[VAL_2]] : !fir.heap<i32>
+! CHECK:         %[[VAL_9:.*]]:2 = fir.do_loop{{.*}}iter_args(%[[IV:.*]] = {{.*}})
+! CHECK:           fir.store %[[IV]] to %[[VAL_2]] : !fir.heap<i32>
 ! CHECK:         }
-! CHECK:         %[[VAL_12:.*]] = fir.convert %[[VAL_13:.*]] : (index) -> i32
-! CHECK:         fir.store %[[VAL_12]] to %[[VAL_2]] : !fir.heap<i32>
+! CHECK:         fir.store %[[VAL_9]]#1 to %[[VAL_2]] : !fir.heap<i32>
   end subroutine
 
 ! CHECK-LABEL: func @_QMtest_loop_varPtest_real_pointer
   subroutine test_real_pointer()
     do x_pointer=1,10
     enddo
 ! CHECK:         %[[VAL_0:.*]] = fir.alloca index
 ! CHECK:         %[[VAL_1:.*]] = fir.address_of(@_QMtest_loop_varEx_pointer) : !fir.ref<!fir.box<!fir.ptr<f32>>>

flang/test/Lower/mixed_loops.f90

integer :: i, j
     do while (j .lt. 21)
       ! CHECK: ^[[BODY1]]:  // pred: ^[[HDR1]]
 
       ! CHECK-DAG: %[[C8_I32:.*]] = arith.constant 8 : i32
       ! CHECK-DAG: %[[C8:.*]] = fir.convert %[[C8_I32]] : (i32) -> index
       ! CHECK-DAG: %[[C13_I32:.*]] = arith.constant 13 : i32
       ! CHECK-DAG: %[[C13:.*]] = fir.convert %[[C13_I32]] : (i32) -> index
       ! CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index
-      ! CHECK: %[[RESULT:.*]] = fir.do_loop %[[IDX:.*]] = %[[C8]] to %[[C13]] step %[[C1]] -> index {
-        ! CHECK: %[[I32:.*]] = fir.convert %[[IDX]] : (index) -> i32
-        ! CHECK: fir.store %[[I32]] to %[[I_REF]] : !fir.ref<i32>
+      ! CHECK: %[[I_LB:.*]] = fir.convert %[[C8]] : (index) -> i32
+      ! CHECK: %[[RESULT:.*]]:2 = fir.do_loop %[[IDX:[^ ]*]] =
+      ! CHECK-SAME: %[[C8]] to %[[C13]] step %[[C1]]
+      ! CHECK-SAME: iter_args(%[[I_IV:.*]] = %[[I_LB]]) -> (index, i32) {
+        ! CHECK: fir.store %[[I_IV]] to %[[I_REF]] : !fir.ref<i32>
         ! CHECK-DAG: %[[J2:.*]] = fir.load %[[J_REF]] : !fir.ref<i32>
         ! CHECK-DAG: %[[C2:.*]] = arith.constant 2 : i32
         ! CHECK: %[[JINC:.*]] = arith.muli %[[C2]], %[[J2]] : i32
         ! CHECK: fir.store %[[JINC]] to %[[J_REF]] : !fir.ref<i32>
         ! CHECK: %[[IINC:.*]] = arith.addi %[[IDX]], %[[C1]] : index
-        ! CHECK: fir.result %[[IINC]] : index
+        ! CHECK: %[[I_STEPCAST:.*]] = fir.convert %[[C1]] : (index) -> i32
+        ! CHECK: %[[I_IVINC:.*]] = arith.addi %[[I_IV]], %[[I_STEPCAST]] : i32
+        ! CHECK: fir.result %[[IINC]], %[[I_IVINC]] : index, i32
       do i=8,13
         j=j*2
 
-      ! CHECK: %[[IFINAL:.*]] = fir.convert %[[RESULT]] : (index) -> i32
-      ! CHECK: fir.store %[[IFINAL]] to %[[I_REF]] : !fir.ref<i32>
+      ! CHECK: fir.store %[[RESULT]]#1 to %[[I_REF]] : !fir.ref<i32>
       end do
 
     ! CHECK: br ^[[HDR1]]
     end do
 
   ! CHECK: ^[[EXIT1]]:  // pred: ^[[HDR1]]
   ! CHECK: %[[IPRINT:.*]] = fir.load %[[I_REF]] : !fir.ref<i32>
   ! CHECK: fir.call @_FortranAioOutputInteger32(%{{.*}}, %[[IPRINT]]) : (!fir.ref<i8>, i32) -> i1
   ! CHECK: %[[JPRINT:.*]] = fir.load %[[J_REF]] : !fir.ref<i32>
   ! CHECK: fir.call @_FortranAioOutputInteger32(%{{.*}}, %[[JPRINT]]) : (!fir.ref<i8>, i32) -> i1
   print *, i, j
 end subroutine