[MLIR][DISC] Revise ParallelLoopTilingPass with inbound_check mode

[• linearhit]

Summary

Expand ParallelLoopTilingPass with an inbound_check mode.

In default mode, the upper bound of the inner loop is from the min op; in inbound_check mode, the upper bound of the inner loop is the step of the outer loop and an additional inbound check will be emitted inside of the inner loop. This was 'FIXME' in the original codes and a typical usage is for GPU backends, thus the outer loop and inner loop can be mapped to blocks/threads in seperate.

[mehdi_amini]

mlir/include/mlir/Dialect/SCF/Passes.h

37

Please document the new param

mlir/lib/Dialect/SCF/Transforms/ParallelLoopTiling.cpp

42

(Nit: Indent this to match the first dimension)

I think this is the crux of the change here: the tiles will be all the same size instead of being dynamic. I was wondering if there is a better naming here for the option: right now it is pointing at the inbound check inside the body instead of pointing at the change in the tile size.
Adding Uday to have another opinion here.

[bondhugula]

mlir/include/mlir/Dialect/SCF/Passes.td

44–47

The name of this option doesn't really convey what it's doing - in fact, it seems to imply exactly the opposite: it is putting the check in the body instead of in the bound! (i.e., in-body-check instead of in-bound-check).

Please update the various instances of upperbound, etc. - I'd recommend space in between upper and bound.

[herhut]

I think a better name here might be something like use-static-loop-bounds or ensure-static-loop-bounds.

We currently get rid of the affine.min when mapping to gpu. Producing static bounded loops like done here would work, too.

I believe the original comment was more about enabling easy vectorization, which this might not achieve. We are using the loop specialization pass for this, which transforms a loop with an upper bound defined via affine.min into a conditional and two loops. That makes the LLVM vectorizer understand the pattern. That is more for CPU use, though.

Just to be clear, I am not arguing against this change.

[rriddle]

nit on the tag in the commit title: What is disc? How does that relate to MLIR upstream (more specifically SCF)?

[• linearhit]

mlir/include/mlir/Dialect/SCF/Passes.h

37

done.

mlir/include/mlir/Dialect/SCF/Passes.td

44–47

done, I rename withInboundCheck into useStaticLoopBounds as Herhut suggests.
"upperbound" is changed to "upper bound"

mlir/lib/Dialect/SCF/Transforms/ParallelLoopTiling.cpp

42

Indent fixed.

[• linearhit]

In D105455#2871426, @herhut wrote:

I think a better name here might be something like use-static-loop-bounds or ensure-static-loop-bounds.

We currently get rid of the affine.min when mapping to gpu. Producing static bounded loops like done here would work, too.

I believe the original comment was more about enabling easy vectorization, which this might not achieve. We are using the loop specialization pass for this, which transforms a loop with an upper bound defined via affine.min into a conditional and two loops. That makes the LLVM vectorizer understand the pattern. That is more for CPU use, though.

Just to be clear, I am not arguing against this change.

I adopt Herhut's suggenstion and rename it into use-static-loop-bounds

[• linearhit]

In D105455#2872566, @rriddle wrote:

nit on the tag in the commit title: What is disc? How does that relate to MLIR upstream (more specifically SCF)?

DISC stands for an end2end dynamic shape compiler, which is doing upstreaming now. Most of the codes will be into mlir-hlo repo, with only a few necessary changes will be submitted to LLVM repo.
Please refer to https://drive.google.com/file/d/1t6Q_VhZVWBhi--fYTxTOLGklIemlKQmV/view?usp=sharing

[mehdi_amini]

In D105455#2878999, @linearhit wrote:

In D105455#2872566, @rriddle wrote:

nit on the tag in the commit title: What is disc? How does that relate to MLIR upstream (more specifically SCF)?

DISC stands for an end2end dynamic shape compiler, which is doing upstreaming now.

We just don't need to tag anything "DISC" in the commit message, I don't think it brings much value to the reader. Downstream projects in general send patches as they make sense in the context of the MLIR project.

[bondhugula]

I adopt Herhut's suggenstion and rename it into use-static-loop-bounds

This is a terminology issue but static would be inaccurate here. Throughout the compiler literature pervasively, bounds where you have symbols that do not depend on data are treated as static. They don't have to be constant nor need to be free of min/max. min(0, N) or max(M, N) is also considered statically predictable the same way as 32*N or M if M and N are treated as symbols. Are you specifically avoiding a min/max? In that case, you can make this more descriptive - no-min-max-bounds? "fixed upper bound" would also be confusing in the pass description.

I also notice multiple instances of "inbound" in the code comments which aren't meaningful.

[• linearhit]

In D105455#2879257, @bondhugula wrote:

I adopt Herhut's suggenstion and rename it into use-static-loop-bounds

This is a terminology issue but static would be inaccurate here. Throughout the compiler literature pervasively, bounds where you have symbols that do not depend on data are treated as static. They don't have to be constant nor need to be free of min/max. min(0, N) or max(M, N) is also considered statically predictable the same way as 32*N or M if M and N are treated as symbols. Are you specifically avoiding a min/max? In that case, you can make this more descriptive - no-min-max-bounds? "fixed upper bound" would also be confusing in the pass description.

I also notice multiple instances of "inbound" in the code comments which aren't meaningful.

I'm also okay on this.
Just to confirm it, is everybody agree with the name no-min-max-bounds? @herhut @mehdi_amini

In my

In D105455#2879257, @bondhugula wrote:

I adopt Herhut's suggenstion and rename it into use-static-loop-bounds

This is a terminology issue but static would be inaccurate here. Throughout the compiler literature pervasively, bounds where you have symbols that do not depend on data are treated as static. They don't have to be constant nor need to be free of min/max. min(0, N) or max(M, N) is also considered statically predictable the same way as 32*N or M if M and N are treated as symbols. Are you specifically avoiding a min/max? In that case, you can make this more descriptive - no-min-max-bounds? "fixed upper bound" would also be confusing in the pass description.

I also notice multiple instances of "inbound" in the code comments which aren't meaningful.

[mehdi_amini]

no-min-max-bounds is fine with me. Thanks!

[• linearhit]

In D105455#2895208, @mehdi_amini wrote:

no-min-max-bounds is fine with me. Thanks!

done!

[• linearhit]

In D105455#2879001, @mehdi_amini wrote:

In D105455#2878999, @linearhit wrote:

In D105455#2872566, @rriddle wrote:

nit on the tag in the commit title: What is disc? How does that relate to MLIR upstream (more specifically SCF)?

DISC stands for an end2end dynamic shape compiler, which is doing upstreaming now.

We just don't need to tag anything "DISC" in the commit message, I don't think it brings much value to the reader. Downstream projects in general send patches as they make sense in the context of the MLIR project.

done!

[herhut]

In D105455#2879257, @bondhugula wrote:

I adopt Herhut's suggenstion and rename it into use-static-loop-bounds

This is a terminology issue but static would be inaccurate here. Throughout the compiler literature pervasively, bounds where you have symbols that do not depend on data are treated as static. They don't have to be constant nor need to be free of min/max. min(0, N) or max(M, N) is also considered statically predictable the same way as 32*N or M if M and N are treated as symbols. Are you specifically avoiding a min/max? In that case, you can make this more descriptive - no-min-max-bounds? "fixed upper bound" would also be confusing in the pass description.

This is also the requirement here. You can have min and max and really any computation as long as such computation is not loop dependent. So min(0, N) would be fine. The current code we have introduces a min between the loop index of the outer loop and a symbol, which I would not consider static.

That said, I am also fine with the current name of the flag.

[• linearhit]

In D105455#2919385, @herhut wrote:

In D105455#2879257, @bondhugula wrote:

I adopt Herhut's suggenstion and rename it into use-static-loop-bounds

This is a terminology issue but static would be inaccurate here. Throughout the compiler literature pervasively, bounds where you have symbols that do not depend on data are treated as static. They don't have to be constant nor need to be free of min/max. min(0, N) or max(M, N) is also considered statically predictable the same way as 32*N or M if M and N are treated as symbols. Are you specifically avoiding a min/max? In that case, you can make this more descriptive - no-min-max-bounds? "fixed upper bound" would also be confusing in the pass description.

This is also the requirement here. You can have min and max and really any computation as long as such computation is not loop dependent. So min(0, N) would be fine. The current code we have introduces a min between the loop index of the outer loop and a symbol, which I would not consider static.

Thanks @herhut
It seems to me that I don't have the permission to merge , could you help with that?

That said, I am also fine with the current name of the flag.

[herhut]

In D105455#2924400, @linearhit wrote:

Thanks @herhut
It seems to me that I don't have the permission to merge , could you help with that?

Sure. Do you have a preferred name/email for me to use as the author of the patch? Essentially what git log would show on your end. I could not find that information.

[• linearhit]

In D105455#2928431, @herhut wrote:

In D105455#2924400, @linearhit wrote:

Thanks @herhut
It seems to me that I don't have the permission to merge , could you help with that?

Sure. Do you have a preferred name/email for me to use as the author of the patch? Essentially what git log would show on your end. I could not find that information.

Sorry for a late reply, please use "tashuang.zk" <[e-mail]>
Thanks.

mlir/include/mlir/Dialect/SCF/Passes.h

 std::unique_ptr<Pass> createParallelLoopFusionPass();
 
 /// Creates a pass that specializes parallel loop for unrolling and
 /// vectorization.
 std::unique_ptr<Pass> createParallelLoopSpecializationPass();
 
 /// Creates a pass which tiles innermost parallel loops.
 std::unique_ptr<Pass>
-createParallelLoopTilingPass(llvm::ArrayRef<int64_t> tileSize = {});
+createParallelLoopTilingPass(llvm::ArrayRef<int64_t> tileSize = {},
+                             bool withInboundCheck = false);
 
 /// Creates a pass which folds arith ops on induction variable into
 /// loop range.
 std::unique_ptr<Pass> createForLoopRangeFoldingPass();
 
 //===----------------------------------------------------------------------===//
 // Registration
 //===----------------------------------------------------------------------===//
 
 /// Generate the code for registering passes.
 #define GEN_PASS_REGISTRATION
 #include "mlir/Dialect/SCF/Passes.h.inc"
 
 } // namespace mlir
 
 #endif // MLIR_DIALECT_SCF_PASSES_H_

mlir/include/mlir/Dialect/SCF/Passes.td

 }
 
 def SCFParallelLoopTiling : FunctionPass<"parallel-loop-tiling"> {
   let summary = "Tile parallel loops";
   let constructor = "mlir::createParallelLoopTilingPass()";
   let options = [
     ListOption<"tileSizes", "parallel-loop-tile-sizes", "int64_t",
                "Factors to tile parallel loops by",
-               "llvm::cl::ZeroOrMore, llvm::cl::MiscFlags::CommaSeparated">
+               "llvm::cl::ZeroOrMore, llvm::cl::MiscFlags::CommaSeparated">,
+    Option<"withInboundCheck", "with-inbound-check", "bool",
+           /*default=*/"false",
+           "Perform tiling with fixed upperbound with inbound check "
+           "for internal loops">
   ];
   let dependentDialects = ["AffineDialect"];
 }
 
 def SCFForLoopRangeFolding
     : Pass<"for-loop-range-folding"> {
   let summary = "Fold add/mul ops into loop range";
   let constructor = "mlir::createForLoopRangeFoldingPass()";
 }
 
 #endif // MLIR_DIALECT_SCF_PASSES

mlir/include/mlir/Dialect/SCF/Transforms.h

 ///     scf.parallel (%j0, %j1) = (0, 0) to (min(tileSize[0], %arg2-%j0)
 ///                                           min(tileSize[1], %arg3-%j1))
 ///                                        step (%arg4, %arg5)
 /// The old loop is replaced with the new one.
 ///
 /// The function returns the resulting ParallelOps, i.e. {outer_loop_op,
 /// inner_loop_op}.
 std::pair<ParallelOp, ParallelOp>
-tileParallelLoop(ParallelOp op, llvm::ArrayRef<int64_t> tileSizes);
+tileParallelLoop(ParallelOp op, llvm::ArrayRef<int64_t> tileSizes,
+                 bool withInboundCheck);
 
 /// Populates patterns for SCF structural type conversions and sets up the
 /// provided ConversionTarget with the appropriate legality configuration for
 /// the ops to get converted properly.
 ///
 /// A "structural" type conversion is one where the underlying ops are
 /// completely agnostic to the actual types involved and simply need to update
 /// their types. An example of this is scf.if -- the scf.if op and the

mlir/lib/Dialect/SCF/Transforms/ParallelLoopTiling.cpp

 
 #include "PassDetail.h"
 #include "mlir/Dialect/Affine/IR/AffineOps.h"
 #include "mlir/Dialect/SCF/Passes.h"
 #include "mlir/Dialect/SCF/SCF.h"
 #include "mlir/Dialect/SCF/Transforms.h"
 #include "mlir/Dialect/SCF/Utils.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
+#include "llvm/ADT/Sequence.h"
 
 using namespace mlir;
 using namespace mlir::scf;
 
 /// Tile a parallel loop of the form
 ///   scf.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
 ///                                            step (%arg4, %arg5)
 ///
 /// into
 ///   scf.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
 ///                                            step (%arg4*tileSize[0],
 ///                                                  %arg5*tileSize[1])
 ///     scf.parallel (%j0, %j1) = (0, 0) to (min(%arg4*tileSize[0], %arg2-%i0)
 ///                                          min(%arg5*tileSize[1], %arg3-%i1))
 ///                                      step (%arg4, %arg5)
 ///
+/// or, when with-inbound-check is true, into
+///   scf.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
+///                                            step (%arg4*tileSize[0],
+///                                                  %arg5*tileSize[1])
+///     scf.parallel (%j0, %j1) = (0, 0) to (%arg4*tileSize[0],
+///     %arg5*tileSize[1])
+///                                      step (%arg4, %arg5)
+///        %inbound = (%j0 * %arg4 + %i0 < %arg2) &&
+///                   (%j1 * %arg5 + %i1 < %arg3)
+///        scf.if (%inbound)
+///          ....
+///
 /// where the uses of %i0 and %i1 in the loop body are replaced by
 /// %i0 + j0 and %i1 + %j1.
 //
 /// The old loop is replaced with the new one.
 std::pair<ParallelOp, ParallelOp>
-mlir::scf::tileParallelLoop(ParallelOp op, ArrayRef<int64_t> tileSizes) {
+mlir::scf::tileParallelLoop(ParallelOp op, ArrayRef<int64_t> tileSizes,
+                            bool withInboundCheck) {
   OpBuilder b(op);
   auto zero = b.create<ConstantIndexOp>(op.getLoc(), 0);
   SmallVector<Value, 2> tileSizeConstants;
   tileSizeConstants.reserve(op.upperBound().size());
   for (size_t i = 0, end = op.upperBound().size(); i != end; ++i) {
     if (i < tileSizes.size())
       tileSizeConstants.push_back(
           b.create<ConstantIndexOp>(op.getLoc(), tileSizes[i]));
auto minMap = AffineMap::get(
       {getAffineDimExpr(/*position=*/0, b.getContext()),
        getAffineDimExpr(/*position=*/1, b.getContext()) -
            getAffineDimExpr(/*position=*/2, b.getContext())},
       b.getContext());
 
   // Create the inner loop with adjusted bounds.
   SmallVector<Value, 2> newBounds;
   newBounds.reserve(op.upperBound().size());
+  bool needInboundCheck = false;
   for (auto dim : llvm::zip(outerLoop.lowerBound(), outerLoop.upperBound(),
                             outerLoop.step(), outerLoop.getInductionVars(),
                             op.step(), tileSizeConstants)) {
     Value lowerBound, upperBound, newStep, iv, step, tileSizeConstant;
     std::tie(lowerBound, upperBound, newStep, iv, step, tileSizeConstant) = dim;
     // Collect the statically known loop bounds
     auto lowerBoundConstant =
         dyn_cast_or_null<ConstantIndexOp>(lowerBound.getDefiningOp());
if (lowerBoundConstant && upperBoundConstant && stepConstant) {
       auto numIterations = llvm::divideCeil(upperBoundConstant.getValue() -
                                                 lowerBoundConstant.getValue(),
                                             stepConstant.getValue());
       if (numIterations % tileSize == 0) {
         newBounds.push_back(newStep);
         continue;
       }
     }
+
+    // For InboundCheck mode, just use the variable outer step
+    if (withInboundCheck) {
+      newBounds.push_back(newStep);
+      needInboundCheck = true;
+      continue;
+    }
+
     // Otherwise, we dynamically compute the bound for
     // each iteration of the outer loop.
     newBounds.push_back(
         b.create<AffineMinOp>(op.getLoc(), b.getIndexType(), minMap,
                               ValueRange{newStep, upperBound, iv}));
   }
   auto innerLoop = b.create<ParallelOp>(
       op.getLoc(), SmallVector<Value, 2>(newBounds.size(), zero), newBounds,
       op.step());
 
-  // Steal the body of the old parallel loop and erase it.
+  if (withInboundCheck && needInboundCheck) {
+    b.setInsertionPointToStart(innerLoop.getBody());
+    // Insert in-bound check
+    Value inbound =
+        b.create<ConstantOp>(op.getLoc(), b.getIntegerType(1),
+                             b.getIntegerAttr(b.getIntegerType(1), 1));
+    for (auto dim :
+         llvm::zip(outerLoop.upperBound(), outerLoop.getInductionVars(),
+                   innerLoop.getInductionVars(), innerLoop.step())) {
+      Value outerUpperBound, outerIV, innerIV, innerStep;
+      std::tie(outerUpperBound, outerIV, innerIV, innerStep) = dim;
+      // %in_bound = %in_bound &&
+      //             (%inner_iv * %inner_step + %outer_iv < %outer_upper_bound)
+      Value index = b.create<AddIOp>(
+          op.getLoc(), b.create<MulIOp>(op.getLoc(), innerIV, innerStep),
+          outerIV);
+      Value dimInbound = b.create<CmpIOp>(op.getLoc(), CmpIPredicate::ult,
+                                          index, outerUpperBound);
+      inbound = b.create<AndOp>(op.getLoc(), inbound, dimInbound);
+    }
+    auto ifInbound = b.create<IfOp>(op.getLoc(),
+                                    /*resultTypes*/ ArrayRef<Type>{}, inbound,
+                                    /*hasElseRegion*/ false);
+    ifInbound.thenRegion().takeBody(op.region());
+    Block &thenBlock = ifInbound.thenRegion().front();
+    b.setInsertionPointToStart(innerLoop.getBody());
+    for (auto ivs : llvm::enumerate(llvm::zip(innerLoop.getInductionVars(),
+                                              outerLoop.getInductionVars()))) {
+      AddIOp newIndex = b.create<AddIOp>(op.getLoc(), std::get<0>(ivs.value()),
+                                         std::get<1>(ivs.value()));
+      thenBlock.getArgument(ivs.index())
+          .replaceAllUsesExcept(newIndex, newIndex);
+    }
+    thenBlock.eraseArguments(llvm::to_vector<4>(
+        llvm::seq((unsigned)0, thenBlock.getNumArguments())));
+  } else {
     innerLoop.region().takeBody(op.region());
-
-  // Insert computation for new index vectors and replace uses.
     b.setInsertionPointToStart(innerLoop.getBody());
-  for (auto ivs :
-       llvm::zip(innerLoop.getInductionVars(), outerLoop.getInductionVars())) {
+    for (auto ivs : llvm::zip(innerLoop.getInductionVars(),
+                              outerLoop.getInductionVars())) {
       Value inner_index = std::get<0>(ivs);
       AddIOp newIndex =
           b.create<AddIOp>(op.getLoc(), std::get<0>(ivs), std::get<1>(ivs));
       inner_index.replaceAllUsesExcept(newIndex, newIndex);
     }
+  }
 
   op.erase();
   return std::make_pair(outerLoop, innerLoop);
 }
 
 namespace {
 struct ParallelLoopTiling
     : public SCFParallelLoopTilingBase<ParallelLoopTiling> {
   ParallelLoopTiling() = default;
-  explicit ParallelLoopTiling(ArrayRef<int64_t> tileSizes) {
+  explicit ParallelLoopTiling(ArrayRef<int64_t> tileSizes,
+                              bool withInboundCheck = false) {
     this->tileSizes = tileSizes;
+    this->withInboundCheck = withInboundCheck;
   }
 
   void runOnFunction() override {
     SmallVector<ParallelOp, 2> innermostPloops;
     getInnermostParallelLoops(getFunction().getOperation(), innermostPloops);
     for (ParallelOp ploop : innermostPloops) {
       // FIXME: Add reduction support.
       if (ploop.getNumReductions() == 0)
-        tileParallelLoop(ploop, tileSizes);
+        tileParallelLoop(ploop, tileSizes, withInboundCheck);
     }
   }
 };
 } // namespace
 
 std::unique_ptr<Pass>
-mlir::createParallelLoopTilingPass(ArrayRef<int64_t> tileSizes) {
-  return std::make_unique<ParallelLoopTiling>(tileSizes);
+mlir::createParallelLoopTilingPass(ArrayRef<int64_t> tileSizes,
+                                   bool withInboundCheck) {
+  return std::make_unique<ParallelLoopTiling>(tileSizes, withInboundCheck);
 }

mlir/test/Dialect/SCF/parallel-loop-tiling-inbound-check.mlir

+// RUN: mlir-opt %s -pass-pipeline='func(parallel-loop-tiling{parallel-loop-tile-sizes=1,4 with-inbound-check=true})' -split-input-file | FileCheck %s
+
+func @parallel_loop(%arg0 : index, %arg1 : index, %arg2 : index,
+                    %arg3 : index, %arg4 : index, %arg5 : index,
+        %A: memref<?x?xf32>, %B: memref<?x?xf32>,
+                    %C: memref<?x?xf32>, %result: memref<?x?xf32>) {
+  scf.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3) step (%arg4, %arg5) {
+    %B_elem = memref.load %B[%i0, %i1] : memref<?x?xf32>
+    %C_elem = memref.load %C[%i0, %i1] : memref<?x?xf32>
+    %sum_elem = addf %B_elem, %C_elem : f32
+    memref.store %sum_elem, %result[%i0, %i1] : memref<?x?xf32>
+  }
+  return
+}
+
+// CHECK-LABEL:   func @parallel_loop(
+// CHECK-SAME:                        [[ARG1:%.*]]: index, [[ARG2:%.*]]: index, [[ARG3:%.*]]: index, [[ARG4:%.*]]: index, [[ARG5:%.*]]: index, [[ARG6:%.*]]: index, [[ARG7:%.*]]: memref<?x?xf32>, [[ARG8:%.*]]: memref<?x?xf32>, [[ARG9:%.*]]: memref<?x?xf32>, [[ARG10:%.*]]: memref<?x?xf32>) {
+// CHECK:           [[C0:%.*]] = constant 0 : index
+// CHECK:           [[C1:%.*]] = constant 1 : index
+// CHECK:           [[C4:%.*]] = constant 4 : index
+// CHECK:           [[V1:%.*]] = muli [[ARG5]], [[C1]] : index
+// CHECK:           [[V2:%.*]] = muli [[ARG6]], [[C4]] : index
+// CHECK:           scf.parallel ([[V3:%.*]], [[V4:%.*]]) = ([[ARG1]], [[ARG2]]) to ([[ARG3]], [[ARG4]]) step ([[V1]], [[V2]]) {
+// CHECK:             scf.parallel ([[V7:%.*]], [[V8:%.*]]) = ([[C0]], [[C0]]) to ([[V1]], [[V2]]) step ([[ARG5]], [[ARG6]]) {
+// CHECK:               [[V9:%.*]] = addi [[V7]], [[V3]] : index
+// CHECK:               [[V10:%.*]] = addi [[V8]], [[V4]] : index
+// CHECK:               %true = constant true
+// CHECK:               [[V11:%.*]] = muli [[V7]], [[ARG5]] : index
+// CHECK:               [[V12:%.*]] = addi [[V11]], [[V3]] : index
+// CHECK:               [[V13:%.*]] = cmpi ult, [[V12]], [[ARG3]] : index
+// CHECK:               [[V14:%.*]] = and %true, [[V13]] : i1
+// CHECK:               [[V15:%.*]] = muli [[V8]], [[ARG6]] : index
+// CHECK:               [[V16:%.*]] = addi [[V15]], [[V4]] : index
+// CHECK:               [[V17:%.*]] = cmpi ult, [[V16]], [[ARG4]] : index
+// CHECK:               [[V18:%.*]] = and [[V14]], [[V17]] : i1
+// CHECK:               scf.if [[V18]] {
+// CHECK:                 [[V19:%.*]] = memref.load [[ARG8]]{{\[}}[[V9]], [[V10]]] : memref<?x?xf32>
+// CHECK:                 [[V20:%.*]] = memref.load [[ARG9]]{{\[}}[[V9]], [[V10]]] : memref<?x?xf32>
+// CHECK:                 [[V21:%.*]] = addf [[V19]], [[V20]] : f32
+// CHECK:                 memref.store [[V21]], [[ARG10]]{{\[}}[[V9]], [[V10]]] : memref<?x?xf32>
+// CHECK:               }
+// CHECK:             }
+// CHECK:           }
+// CHECK:           return
+
+// -----
+
+func @static_loop_with_step() {
+  %c0 = constant 0 : index
+  %c3 = constant 3 : index
+  %c22 = constant 22 : index
+  %c24 = constant 24 : index
+  scf.parallel (%i0, %i1) = (%c0, %c0) to (%c22, %c24) step (%c3, %c3) {
+  }
+  return
+}
+
+// CHECK-LABEL:   func @static_loop_with_step() {
+// CHECK:           [[C0:%.*]] = constant 0 : index
+// CHECK:           [[C3:%.*]] = constant 3 : index
+// CHECK:           [[C22:%.*]] = constant 22 : index
+// CHECK:           [[C24:%.*]] = constant 24 : index
+// CHECK:           [[C0_1:%.*]] = constant 0 : index
+// CHECK:           [[C1:%.*]] = constant 1 : index
+// CHECK:           [[C4:%.*]] = constant 4 : index
+// CHECK:           [[V1:%.*]] = muli [[C3]], [[C1]] : index
+// CHECK:           [[V2:%.*]] = muli [[C3]], [[C4]] : index
+// CHECK:           scf.parallel ([[V3:%.*]], [[V4:%.*]]) = ([[C0]], [[C0]]) to ([[C22]], [[C24]]) step ([[V1]], [[V2]]) {
+// CHECK:             scf.parallel ([[V5:%.*]], [[V6:%.*]]) = ([[C0_1]], [[C0_1]]) to ([[V1]], [[V2]]) step ([[C3]], [[C3]]) {
+// CHECK-NOT:           scf.if
+// CHECK:               = addi [[V5]], [[V3]] : index
+// CHECK:               = addi [[V6]], [[V4]] : index
+// CHECK:             }
+// CHECK:           }
+// CHECK:           return
+
+// -----
+
+func @tile_nested_innermost() {
+  %c2 = constant 2 : index
+  %c0 = constant 0 : index
+  %c1 = constant 1 : index
+  scf.parallel (%i, %j) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
+    scf.parallel (%k, %l) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
+    }
+  }
+  scf.parallel (%i, %j) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
+  }
+  return
+}
+
+// CHECK-LABEL:   func @tile_nested_innermost() {
+// CHECK:           [[C2:%.*]] = constant 2 : index
+// CHECK:           [[C0:%.*]] = constant 0 : index
+// CHECK:           [[C1:%.*]] = constant 1 : index
+// CHECK:           scf.parallel ([[V1:%.*]], [[V2:%.*]]) = ([[C0]], [[C0]]) to ([[C2]], [[C2]]) step ([[C1]], [[C1]]) {
+// CHECK:             [[C0_1:%.*]] = constant 0 : index
+// CHECK:             [[C1_1:%.*]] = constant 1 : index
+// CHECK:             [[C4:%.*]] = constant 4 : index
+// CHECK:             [[V3:%.*]] = muli [[C1]], [[C1_1]] : index
+// CHECK:             [[V4:%.*]] = muli [[C1]], [[C4]] : index
+// CHECK:             scf.parallel ([[V5:%.*]], [[V6:%.*]]) = ([[C0]], [[C0]]) to ([[C2]], [[C2]]) step ([[V3]], [[V4]]) {
+// CHECK:               scf.parallel ([[V8:%.*]], [[V9:%.*]]) = ([[C0_1]], [[C0_1]]) to ([[V3]], [[V4]]) step ([[C1]], [[C1]]) {
+// CHECK:                 = addi [[V8]], [[V5]] : index
+// CHECK:                 = addi [[V9]], [[V6]] : index
+// CHECK:                 scf.if
+// CHECK:               }
+// CHECK:             }
+// CHECK:           }
+// CHECK:           [[C0_2:%.*]] = constant 0 : index
+// CHECK:           [[C1_2:%.*]] = constant 1 : index
+// CHECK:           [[C4_1:%.*]] = constant 4 : index
+// CHECK:           [[V10:%.*]] = muli [[C1]], [[C1_2]] : index
+// CHECK:           [[V11:%.*]] = muli [[C1]], [[C4_1]] : index
+// CHECK:           scf.parallel ([[V12:%.*]], [[V13:%.*]]) = ([[C0]], [[C0]]) to ([[C2]], [[C2]]) step ([[V10]], [[V11]]) {
+// CHECK:             scf.parallel ([[V15:%.*]], [[V16:%.*]]) = ([[C0_2]], [[C0_2]]) to ([[V10]], [[V11]]) step ([[C1]], [[C1]]) {
+// CHECK:               = addi [[V15]], [[V12]] : index
+// CHECK:               = addi [[V16]], [[V13]] : index
+// CHECK:               scf.if
+// CHECK:             }
+// CHECK:           }
+// CHECK:           return
+// CHECK:         }
+
+// -----
+
+func @tile_nested_in_non_ploop() {
+  %c0 = constant 0 : index
+  %c1 = constant 1 : index
+  %c2 = constant 2 : index
+  scf.for %i = %c0 to %c2 step %c1 {
+    scf.for %j = %c0 to %c2 step %c1 {
+      scf.parallel (%k, %l) = (%c0, %c0) to (%c2, %c2) step (%c1, %c1) {
+      }
+    }
+  }
+  return
+}
+
+// CHECK-LABEL: func @tile_nested_in_non_ploop
+// CHECK:         scf.for
+// CHECK:           scf.for
+// CHECK:             scf.parallel
+// CHECK:               scf.parallel
+// CHECK:               }
+// CHECK:             }
+// CHECK:           }
+// CHECK:         }
+// CHECK:       }