[MLIR][GPU] Properly model step in parallel loop to gpu conversion.

[herhut]

Summary

The original patch had TODOs to add support for step computations,
which this commit addresses. The computations are expressed using
affine expressions so that the affine canonicalizers can simplify
the full bound and index computations.

Also cleans up the code a little and exposes the pass in the
header file.

[herhut]

PTAL.

[herhut]

Simplify afffine constant matching.

[bondhugula]

mlir/test/Conversion/LoopsToGPU/parallel_loop.mlir

235

Side question: where aren't we using affine.load/store instead of load/store and loop.parallel -> affine.parallel here? With the former, you'll get things like store to load fwd'ing, redundant load elimination, composition of ops supplying subscript values into the load/store itself, etc., infra for all of which exist and whenever you need them. All the mapping metadata should nicely fit into affine.parallel as well.

[ftynse]

mlir/lib/Conversion/LoopsToGPU/LoopsToGPU.cpp

542

Nit: could you early-return instead?

669

Nit "we use affine apply" ?

mlir/test/Conversion/LoopsToGPU/parallel_loop.mlir

18–19

Let's not match on attribute names here, i.e. prefer #[[MAP0:.*]] = affine_map<...

[herhut]

Review comments.

[herhut]

Thanks for the review!

mlir/test/Conversion/LoopsToGPU/parallel_loop.mlir

235

It is not pure coincidence that the mapping data fits :)

My hope is that this mapper will work equally well with affine.parallel. However, I do not want to restrict it to affine and currently the code we feed into this is not based on affine.parallel. I expect that we will generalize things in that direction eventually but would also be very happy if someone else looks into that.

[bondhugula]

mlir/test/Conversion/LoopsToGPU/parallel_loop.mlir

235

However, I do not want to restrict it to affine and currently the code
we feed into this is not based on affine.parallel. I expect that we will

"The code we feed into this": is the thing that's generating the loop.parallel's available somewhere or is it something that's planned for release in the future?

generalize things in that direction eventually but would also be very
happy if someone else looks into that.

But in order to do that one would also have to look at that converter that's generating the loop dialect ops and switch it to affine dialect ones. IMO, that'd prevent duplication of a lot of infrastructure in less powerful ways. All of these examples can be represented and transformed (whether or not you need any analysis) with the affine dialect.

mlir/include/mlir/Conversion/LoopsToGPU/LoopsToGPUPass.h

 
 #include "mlir/Support/LLVM.h"
 
 #include <memory>
 
 namespace mlir {
 class FuncOp;
 template <typename T> class OpPassBase;
+class Pass;
 
 /// Create a pass that converts loop nests into GPU kernels.  It considers
 /// top-level affine.for and linalg.for operations as roots of loop nests and
 /// converts them to the gpu.launch operations if possible.
 ///
 /// No check on the size of the block or grid, or on the validity of
 /// parallelization is performed, it is under the responsibility of the caller
 /// to strip-mine the loops and to perform the dependence analysis before
 /// calling the conversion.
 std::unique_ptr<OpPassBase<FuncOp>>
 createSimpleLoopsToGPUPass(unsigned numBlockDims, unsigned numThreadDims);
 
 /// Create a pass that converts every loop operation within the body of the
 /// FuncOp into a GPU launch. The number of workgroups and workgroup size for
 /// the implementation is controlled by SSA values passed into conversion
 /// method. For testing, the values are set as constants obtained from a command
 /// line flag. See convertLoopToGPULaunch for a description of the required
 /// semantics of the converted loop operation.
 std::unique_ptr<OpPassBase<FuncOp>>
 createLoopToGPUPass(ArrayRef<int64_t> numWorkGroups,
                     ArrayRef<int64_t> workGroupSize);
+
+/// Creates a pass that converts loop.parallel operations into a gpu.launch
+/// operation. The mapping of loop dimensions to launch dimensions is derived
+/// from mapping attributes. See ParallelToGpuLaunchLowering::matchAndRewrite
+/// for a description of the used attributes.
+std::unique_ptr<Pass> createParallelLoopToGpuPass();
+
 } // namespace mlir
 
 #endif // MLIR_CONVERSION_LOOPSTOGPU_LOOPSTOGPUPASS_H_

mlir/lib/Conversion/LoopsToGPU/LoopsToGPU.cpp

static MappingAnnotation extractMappingAnnotation(Attribute attribute) {
   AffineMap bound;
   if (boundAttr)
     bound = boundAttr.getValue();
   return {processor, map, bound};
 }
 
 /// Tries to derive a static upper bound from the defining operation of
 /// `upperBound`.
-static Value deriveStaticUpperBound(Value upperBound) {
+static Value deriveStaticUpperBound(Value upperBound,
+                                    PatternRewriter &rewriter) {
   Value constantBound = {};
+  // We cannot rely on canonicalization at the moment, so we have to
+  // dig throuhg the expression.
+  // See https://bugs.llvm.org/show_bug.cgi?id=45008.
   if (AffineMinOp minOp =
           dyn_cast_or_null<AffineMinOp>(upperBound.getDefiningOp())) {
     auto map = minOp.map();
     auto operands = minOp.operands();
     for (int sub = 0, e = map.getNumResults(); sub < e; ++sub) {
       AffineExpr expr = map.getResult(sub);
       if (AffineDimExpr dimExpr = expr.dyn_cast<AffineDimExpr>()) {
         auto dimOperand = operands[dimExpr.getPosition()];
         auto defOp = dimOperand.getDefiningOp();
         if (ConstantOp constOp = dyn_cast_or_null<ConstantOp>(defOp)) {
           constantBound = constOp;
           break;
         }
       }
+      if (AffineConstantExpr constExpr = expr.dyn_cast<AffineConstantExpr>()) {
+        constantBound = rewriter.create<ConstantIndexOp>(minOp.getLoc(),
+                                                         constExpr.getValue());
+        break;
+      }
     }
   }
   return constantBound;
 }
 
 /// Modifies the current transformation state to capture the effect of the given
 /// `loop.parallel` operation on index substitutions and the operations to be
 /// inserted.
for (auto config : llvm::zip(mapping, parallelOp.getInductionVars(),
     Attribute mappingAttribute;
     Value iv, lowerBound, upperBound, step;
     std::tie(mappingAttribute, iv, lowerBound, upperBound, step) = config;
     MappingAnnotation annotation = extractMappingAnnotation(mappingAttribute);
     Value newIndex;
 
     if (annotation.processor < gpu::LaunchOp::kNumConfigOperands) {
       // Use the corresponding thread/grid index as replacement for the loop iv.
-      // TODO(herhut): Make the iv calculation depend on lower & upper bound.
       Value operand = launchOp.body().front().getArgument(annotation.processor);
-      Value appliedMap =
-          rewriter.create<AffineApplyOp>(loc, annotation.indexMap, operand);
-      // Add the lower bound, as the maps are 0 based but the loop might not be.
-      // TODO(herhut): Maybe move this explicitly into the maps?
-      newIndex = rewriter.create<AddIOp>(
-          loc, appliedMap, cloningMap.lookupOrDefault(lowerBound));
+      // Take the indexmap and add the lower bound and step computations in.
+      // This computes operand * step + lowerBound.
+      // Use an affine map here so that it composes nicely with the provided
+      // annotation.
+      AffineMap lowerAndStep = AffineMap::get(
+          1, 2,
+          rewriter.getAffineDimExpr(0) * rewriter.getAffineSymbolExpr(0) +
+              rewriter.getAffineSymbolExpr(1));
+      newIndex = rewriter.create<AffineApplyOp>(
+          loc, annotation.indexMap.compose(lowerAndStep),
+          ValueRange{operand, step, lowerBound});
       // If there was also a bound, insert that, too.
       // TODO(herhut): Check that we do not assign bounds twice.
       if (annotation.boundMap) {
         // We pass as the single opererand to the bound-map the number of
-        // iterations, which is upperBound - lowerBound. To support inner loops
-        // with dynamic upper bounds (as generated by e.g. tiling), try to
-        // derive a max for the bounds. If the used bound for the hardware id is
-        // inprecise, wrap the contained code into a conditional.
-        // If the lower-bound is constant or defined before the launch, we can
-        // use it in the launch bounds. Otherwise fail.
+        // iterations, which is (upperBound - lowerBound) ceilDiv step. To
+        // support inner loops with dynamic upper bounds (as generated by e.g.
+        // tiling), try to derive a max for the bounds. If the used bound for
+        // the hardware id is imprecise, wrap the contained code into a
+        // conditional. If the lower-bound is constant or defined before the
+        // launch, we can use it in the launch bounds. Otherwise fail.
         if (!launchIndependent(lowerBound) &&
             !isa<ConstantOp>(lowerBound.getDefiningOp()))
           return failure();
+        // The step must also be constant or defined outside of the loop nest.
+        if (!launchIndependent(step) && !isa<ConstantOp>(step.getDefiningOp()))
+          return failure();
         // If the upper-bound is constant or defined before the launch, we can
         // use it in the launch bounds directly. Otherwise try derive a bound.
         bool boundIsPrecise = launchIndependent(upperBound) ||
                               isa<ConstantOp>(upperBound.getDefiningOp());
+        {
+          PatternRewriter::InsertionGuard guard(rewriter);
+          rewriter.setInsertionPoint(launchOp);
           if (!boundIsPrecise) {
-          upperBound = deriveStaticUpperBound(upperBound);
+            upperBound = deriveStaticUpperBound(upperBound, rewriter);
             if (!upperBound)
               return failure();
           }
-        {
-          PatternRewriter::InsertionGuard guard(rewriter);
-          rewriter.setInsertionPoint(launchOp);
-
-          Value iterations = rewriter.create<SubIOp>(
-              loc,
-              ensureLaunchIndependent(cloningMap.lookupOrDefault(upperBound)),
-              ensureLaunchIndependent(cloningMap.lookupOrDefault(lowerBound)));
+          // Compute the number of iterations needed. We compute this as an
+          // affine expression ceilDiv (upperBound - lowerBound) step. We use
+          // affine here so that it composes nicely with the provided map.
+          AffineMap stepMap =
+              AffineMap::get(0, 3,
+                             (rewriter.getAffineSymbolExpr(0) -
+                              rewriter.getAffineSymbolExpr(1).ceilDiv(
+                                  rewriter.getAffineSymbolExpr(2))));
           Value launchBound = rewriter.create<AffineApplyOp>(
-              loc, annotation.boundMap, iterations);
+              loc, annotation.boundMap.compose(stepMap),
+              ValueRange{
+                  ensureLaunchIndependent(
+                      cloningMap.lookupOrDefault(upperBound)),
+                  ensureLaunchIndependent(
+                      cloningMap.lookupOrDefault(lowerBound)),
+                  ensureLaunchIndependent(cloningMap.lookupOrDefault(step))});
           launchOp.setOperand(annotation.processor, launchBound);
         }
         if (!boundIsPrecise) {
           // We are using an approximation, create a surrounding conditional.
           Value originalBound = std::get<3>(config);
           CmpIOp pred = rewriter.create<CmpIOp>(
               loc, CmpIPredicate::slt, newIndex,
               cloningMap.lookupOrDefault(originalBound));
if (failed(processParallelLoop(parallelOp, launchOp, cloningMap, worklist,
     return matchFailure();
 
   // Whether we have seen any side-effects. Reset when leaving an inner scope.
   bool seenSideeffects = false;
   // Whether we have left a nesting scope (and hence are no longer innermost).
   bool leftNestingScope = false;
   while (!worklist.empty()) {
     Operation *op = worklist.pop_back_val();
-    launchOp.dump();
-
     // Now walk over the body and clone it.
     // TODO: This is only correct if there either is no further loop.parallel
     //       nested or this code is side-effect free. Otherwise we might need
     //       predication. We are overly consertaive for now and only allow
     //       side-effects in the innermost scope.
     if (auto nestedParallel = dyn_cast<ParallelOp>(op)) {
       // Before entering a nested scope, make sure there have been no
       // sideeffects until now.
if (auto nestedParallel = dyn_cast<ParallelOp>(op)) {
         return matchFailure();
     }
   }
 
   rewriter.eraseOp(parallelOp);
   return matchSuccess();
 }
 
-namespace {
-struct ParallelLoopToGpuPass : public OperationPass<ParallelLoopToGpuPass> {
-  void runOnOperation() override;
-};
-} // namespace
-
 void mlir::populateParallelLoopToGPUPatterns(OwningRewritePatternList &patterns,
                                              MLIRContext *ctx) {
   patterns.insert<ParallelToGpuLaunchLowering>(ctx);
 }
-
-void ParallelLoopToGpuPass::runOnOperation() {
-  OwningRewritePatternList patterns;
-  populateParallelLoopToGPUPatterns(patterns, &getContext());
-  ConversionTarget target(getContext());
-  target.addLegalDialect<StandardOpsDialect>();
-  target.addLegalDialect<AffineOpsDialect>();
-  target.addLegalDialect<gpu::GPUDialect>();
-  target.addLegalDialect<loop::LoopOpsDialect>();
-  target.addIllegalOp<loop::ParallelOp>();
-  if (failed(applyPartialConversion(getOperation(), target, patterns)))
-    signalPassFailure();
-}
-
-static PassRegistration<ParallelLoopToGpuPass>
-    pass("convert-parallel-loops-to-gpu", "Convert mapped loop.parallel ops"
-                                          " to gpu launch operations.");

mlir/lib/Conversion/LoopsToGPU/LoopsToGPUPass.cpp

 //===- LoopsToGPUPass.cpp - Convert a loop nest to a GPU kernel -----------===//
 //
 // 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 "mlir/Conversion/LoopsToGPU/LoopsToGPUPass.h"
 #include "mlir/Conversion/LoopsToGPU/LoopsToGPU.h"
 #include "mlir/Dialect/AffineOps/AffineOps.h"
+#include "mlir/Dialect/GPU/GPUDialect.h"
 #include "mlir/Dialect/LoopOps/LoopOps.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
 #include "mlir/Pass/Pass.h"
+#include "mlir/Transforms/DialectConversion.h"
 
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/Support/CommandLine.h"
 
 #define PASS_NAME "convert-loops-to-gpu"
 #define LOOPOP_TO_GPU_PASS_NAME "convert-loop-op-to-gpu"
 
 using namespace mlir;
for (Block &block : getFunction()) {
         }
       }
     }
   }
   SmallVector<int64_t, 3> numWorkGroups;
   SmallVector<int64_t, 3> workGroupSize;
 };
 
+struct ParallelLoopToGpuPass : public OperationPass<ParallelLoopToGpuPass> {
+  void runOnOperation() override {
+    OwningRewritePatternList patterns;
+    populateParallelLoopToGPUPatterns(patterns, &getContext());
+    ConversionTarget target(getContext());
+    target.addLegalDialect<StandardOpsDialect>();
+    target.addLegalDialect<AffineOpsDialect>();
+    target.addLegalDialect<gpu::GPUDialect>();
+    target.addLegalDialect<loop::LoopOpsDialect>();
+    target.addIllegalOp<loop::ParallelOp>();
+    if (failed(applyPartialConversion(getOperation(), target, patterns)))
+      signalPassFailure();
+  }
+};
+
 } // namespace
 
 std::unique_ptr<OpPassBase<FuncOp>>
 mlir::createSimpleLoopsToGPUPass(unsigned numBlockDims,
                                  unsigned numThreadDims) {
   return std::make_unique<ForLoopMapper>(numBlockDims, numThreadDims);
 }
 
 std::unique_ptr<OpPassBase<FuncOp>>
 mlir::createLoopToGPUPass(ArrayRef<int64_t> numWorkGroups,
                           ArrayRef<int64_t> workGroupSize) {
   return std::make_unique<ImperfectlyNestedForLoopMapper>(numWorkGroups,
                                                           workGroupSize);
 }
 
+std::unique_ptr<Pass> mlir::createParallelLoopToGpuPass() {
+  return std::make_unique<ParallelLoopToGpuPass>();
+}
+
 static PassRegistration<ForLoopMapper>
     registration(PASS_NAME, "Convert top-level loops to GPU kernels", [] {
       return std::make_unique<ForLoopMapper>(clNumBlockDims.getValue(),
                                              clNumThreadDims.getValue());
     });
 
 static PassRegistration<ImperfectlyNestedForLoopMapper> loopOpToGPU(
     LOOPOP_TO_GPU_PASS_NAME, "Convert top-level loop::ForOp to GPU kernels",
     [] {
       SmallVector<int64_t, 3> numWorkGroups, workGroupSize;
       numWorkGroups.assign(clNumWorkGroups.begin(), clNumWorkGroups.end());
       workGroupSize.assign(clWorkGroupSize.begin(), clWorkGroupSize.end());
       return std::make_unique<ImperfectlyNestedForLoopMapper>(numWorkGroups,
                                                               workGroupSize);
     });
+
+static PassRegistration<ParallelLoopToGpuPass>
+    pass("convert-parallel-loops-to-gpu", "Convert mapped loop.parallel ops"
+                                          " to gpu launch operations.");

mlir/test/Conversion/LoopsToGPU/parallel_loop.mlir

func @parallel_loop_bidy_bidx(%arg0 : index, %arg1 : index, %arg2 : index,
   loop.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
                                           step (%arg4, %step)  {
     %val = load %buf[%i0, %i1] : memref<?x?xf32>
     store %val, %res[%i1, %i0] : memref<?x?xf32>
   } { mapping = [{processor = 1, map = affine_map<(d0) -> (d0)>, bound = affine_map<(d0) -> (d0)>}, {processor = 0, map = affine_map<(d0) -> (d0)>, bound = affine_map<(d0) -> (d0)>}] }
   return
 }
 
-// CHECK:       #map0 = affine_map<(d0) -> (d0)>
-// CHECK:       module {
+// CHECK:       #map0 = affine_map<()[s0, s1, s2] -> (s0 - s1 ceildiv s2)>
+// CHECK:       #map1 = affine_map<(d0)[s0, s1] -> (d0 * s0 + s1)>
 
+// CHECK:       module {
 // CHECK-LABEL:   func @parallel_loop_bidy_bidx(
-// CHECK-SAME:                        [[VAL_0:%.*]]: index, [[VAL_1:%.*]]: index, [[VAL_2:%.*]]: index, [[VAL_3:%.*]]: index, [[VAL_4:%.*]]: index, [[VAL_5:%.*]]: memref<?x?xf32>, [[VAL_6:%.*]]: memref<?x?xf32>) {
+// CHECK-SAME:                                  [[VAL_0:%.*]]: index, [[VAL_1:%.*]]: index, [[VAL_2:%.*]]: index, [[VAL_3:%.*]]: index, [[VAL_4:%.*]]: index, [[VAL_5:%.*]]: memref<?x?xf32>, [[VAL_6:%.*]]: memref<?x?xf32>) {
 // CHECK:           [[VAL_7:%.*]] = constant 2 : index
 // CHECK:           [[VAL_8:%.*]] = constant 1 : index
-// CHECK:           [[VAL_9:%.*]] = subi [[VAL_2]], [[VAL_0]] : index
-// CHECK:           [[VAL_10:%.*]] = affine.apply #map0([[VAL_9]])
-// CHECK:           [[VAL_11:%.*]] = subi [[VAL_3]], [[VAL_1]] : index
-// CHECK:           [[VAL_12:%.*]] = affine.apply #map0([[VAL_11]])
-// CHECK:           gpu.launch blocks([[VAL_13:%.*]], [[VAL_14:%.*]], [[VAL_15:%.*]]) in ([[VAL_16:%.*]] = [[VAL_12]], [[VAL_17:%.*]] = [[VAL_10]], [[VAL_18:%.*]] = [[VAL_8]]) threads([[VAL_19:%.*]], [[VAL_20:%.*]], [[VAL_21:%.*]]) in ([[VAL_22:%.*]] = [[VAL_8]], [[VAL_23:%.*]] = [[VAL_8]], [[VAL_24:%.*]] = [[VAL_8]]) {
-// CHECK:             [[VAL_25:%.*]] = affine.apply #map0([[VAL_14]])
-// CHECK:             [[VAL_26:%.*]] = addi [[VAL_25]], [[VAL_0]] : index
-// CHECK:             [[VAL_27:%.*]] = affine.apply #map0([[VAL_13]])
-// CHECK:             [[VAL_28:%.*]] = addi [[VAL_27]], [[VAL_1]] : index
-// CHECK:             [[VAL_29:%.*]] = load [[VAL_5]]{{\[}}[[VAL_26]], [[VAL_28]]] : memref<?x?xf32>
-// CHECK:             store [[VAL_29]], [[VAL_6]]{{\[}}[[VAL_28]], [[VAL_26]]] : memref<?x?xf32>
+// CHECK:           [[VAL_9:%.*]] = affine.apply #map0(){{\[}}[[VAL_2]], [[VAL_0]], [[VAL_4]]]
+// CHECK:           [[VAL_10:%.*]] = affine.apply #map0(){{\[}}[[VAL_3]], [[VAL_1]], [[VAL_7]]]
+// CHECK:           gpu.launch blocks([[VAL_11:%.*]], [[VAL_12:%.*]], [[VAL_13:%.*]]) in ([[VAL_14:%.*]] = [[VAL_10]], [[VAL_15:%.*]] = [[VAL_9]], [[VAL_16:%.*]] = [[VAL_8]]) threads([[VAL_17:%.*]], [[VAL_18:%.*]], [[VAL_19:%.*]]) in ([[VAL_20:%.*]] = [[VAL_8]], [[VAL_21:%.*]] = [[VAL_8]], [[VAL_22:%.*]] = [[VAL_8]]) {
+// CHECK:             [[VAL_23:%.*]] = affine.apply #map1([[VAL_12]]){{\[}}[[VAL_4]], [[VAL_0]]]
+// CHECK:             [[VAL_24:%.*]] = affine.apply #map1([[VAL_11]]){{\[}}[[VAL_7]], [[VAL_1]]]
+// CHECK:             [[VAL_25:%.*]] = load [[VAL_5]]{{\[}}[[VAL_23]], [[VAL_24]]] : memref<?x?xf32>
+// CHECK:             store [[VAL_25]], [[VAL_6]]{{\[}}[[VAL_24]], [[VAL_23]]] : memref<?x?xf32>
 // CHECK:             gpu.terminator
 // CHECK:           }
 // CHECK:           return
 // CHECK:         }
 // CHECK:       }
 
 // -----
 
loop.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
      ] }
   } { mapping = [
       {processor = 1, map = affine_map<(d0) -> (d0)>, bound = affine_map<(d0) -> (d0)>},
       {processor = 0, map = affine_map<(d0) -> (d0)>, bound = affine_map<(d0) -> (d0)>}
     ] }
   return
 }
 
-// CHECK:       #map0 = affine_map<(d0) -> (d0)>
-// CHECK:       module {
+// CHECK:       #map0 = affine_map<()[s0, s1, s2] -> (s0 - s1 ceildiv s2)>
+// CHECK:       #map1 = affine_map<(d0)[s0, s1] -> (d0 * s0 + s1)>
 
+// CHECK:       module {
 // CHECK-LABEL:   func @parallel_loop_tiled(
-// CHECK-SAME:                              [[VAL_30:%.*]]: index, [[VAL_31:%.*]]: index, [[VAL_32:%.*]]: index, [[VAL_33:%.*]]: index, [[VAL_34:%.*]]: memref<?x?xf32>, [[VAL_35:%.*]]: memref<?x?xf32>) {
-// CHECK:           [[VAL_36:%.*]] = constant 0 : index
-// CHECK:           [[VAL_37:%.*]] = constant 1 : index
-// CHECK:           [[VAL_38:%.*]] = constant 4 : index
-// CHECK:           [[VAL_39:%.*]] = constant 1 : index
-// CHECK:           [[VAL_40:%.*]] = subi [[VAL_32]], [[VAL_30]] : index
-// CHECK:           [[VAL_41:%.*]] = affine.apply #map0([[VAL_40]])
-// CHECK:           [[VAL_42:%.*]] = subi [[VAL_33]], [[VAL_31]] : index
-// CHECK:           [[VAL_43:%.*]] = affine.apply #map0([[VAL_42]])
-// CHECK:           [[VAL_44:%.*]] = subi [[VAL_38]], [[VAL_36]] : index
-// CHECK:           [[VAL_45:%.*]] = affine.apply #map0([[VAL_44]])
-// CHECK:           [[VAL_46:%.*]] = subi [[VAL_38]], [[VAL_36]] : index
-// CHECK:           [[VAL_47:%.*]] = affine.apply #map0([[VAL_46]])
-// CHECK:           gpu.launch blocks([[VAL_48:%.*]], [[VAL_49:%.*]], [[VAL_50:%.*]]) in ([[VAL_51:%.*]] = [[VAL_43]], [[VAL_52:%.*]] = [[VAL_41]], [[VAL_53:%.*]] = [[VAL_39]]) threads([[VAL_54:%.*]], [[VAL_55:%.*]], [[VAL_56:%.*]]) in ([[VAL_57:%.*]] = [[VAL_47]], [[VAL_58:%.*]] = [[VAL_45]], [[VAL_59:%.*]] = [[VAL_39]]) {
-// CHECK:             [[VAL_60:%.*]] = affine.apply #map0([[VAL_49]])
-// CHECK:             [[VAL_61:%.*]] = addi [[VAL_60]], [[VAL_30]] : index
-// CHECK:             [[VAL_62:%.*]] = affine.apply #map0([[VAL_48]])
-// CHECK:             [[VAL_63:%.*]] = addi [[VAL_62]], [[VAL_31]] : index
-// CHECK:             [[VAL_64:%.*]] = affine.apply #map0([[VAL_55]])
-// CHECK:             [[VAL_65:%.*]] = addi [[VAL_64]], [[VAL_36]] : index
-// CHECK:             [[VAL_66:%.*]] = affine.apply #map0([[VAL_54]])
-// CHECK:             [[VAL_67:%.*]] = addi [[VAL_66]], [[VAL_36]] : index
-// CHECK:             [[VAL_68:%.*]] = addi [[VAL_61]], [[VAL_65]] : index
-// CHECK:             [[VAL_69:%.*]] = addi [[VAL_63]], [[VAL_67]] : index
-// CHECK:             [[VAL_70:%.*]] = load [[VAL_34]]{{\[}}[[VAL_68]], [[VAL_69]]] : memref<?x?xf32>
-// CHECK:             store [[VAL_70]], [[VAL_35]]{{\[}}[[VAL_69]], [[VAL_68]]] : memref<?x?xf32>
+// CHECK-SAME:                              [[VAL_26:%.*]]: index, [[VAL_27:%.*]]: index, [[VAL_28:%.*]]: index, [[VAL_29:%.*]]: index, [[VAL_30:%.*]]: memref<?x?xf32>, [[VAL_31:%.*]]: memref<?x?xf32>) {
+// CHECK:           [[VAL_32:%.*]] = constant 0 : index
+// CHECK:           [[VAL_33:%.*]] = constant 1 : index
+// CHECK:           [[VAL_34:%.*]] = constant 4 : index
+// CHECK:           [[VAL_35:%.*]] = constant 1 : index
+// CHECK:           [[VAL_36:%.*]] = affine.apply #map0(){{\[}}[[VAL_28]], [[VAL_26]], [[VAL_34]]]
+// CHECK:           [[VAL_37:%.*]] = affine.apply #map0(){{\[}}[[VAL_29]], [[VAL_27]], [[VAL_34]]]
+// CHECK:           [[VAL_38:%.*]] = affine.apply #map0(){{\[}}[[VAL_34]], [[VAL_32]], [[VAL_33]]]
+// CHECK:           [[VAL_39:%.*]] = affine.apply #map0(){{\[}}[[VAL_34]], [[VAL_32]], [[VAL_33]]]
+// CHECK:           gpu.launch blocks([[VAL_40:%.*]], [[VAL_41:%.*]], [[VAL_42:%.*]]) in ([[VAL_43:%.*]] = [[VAL_37]], [[VAL_44:%.*]] = [[VAL_36]], [[VAL_45:%.*]] = [[VAL_35]]) threads([[VAL_46:%.*]], [[VAL_47:%.*]], [[VAL_48:%.*]]) in ([[VAL_49:%.*]] = [[VAL_39]], [[VAL_50:%.*]] = [[VAL_38]], [[VAL_51:%.*]] = [[VAL_35]]) {
+// CHECK:             [[VAL_52:%.*]] = affine.apply #map1([[VAL_41]]){{\[}}[[VAL_34]], [[VAL_26]]]
+// CHECK:             [[VAL_53:%.*]] = affine.apply #map1([[VAL_40]]){{\[}}[[VAL_34]], [[VAL_27]]]
+// CHECK:             [[VAL_54:%.*]] = affine.apply #map1([[VAL_47]]){{\[}}[[VAL_33]], [[VAL_32]]]
+// CHECK:             [[VAL_55:%.*]] = affine.apply #map1([[VAL_46]]){{\[}}[[VAL_33]], [[VAL_32]]]
+// CHECK:             [[VAL_56:%.*]] = addi [[VAL_52]], [[VAL_54]] : index
+// CHECK:             [[VAL_57:%.*]] = addi [[VAL_53]], [[VAL_55]] : index
+// CHECK:             [[VAL_58:%.*]] = load [[VAL_30]]{{\[}}[[VAL_56]], [[VAL_57]]] : memref<?x?xf32>
+// CHECK:             store [[VAL_58]], [[VAL_31]]{{\[}}[[VAL_57]], [[VAL_56]]] : memref<?x?xf32>
 // CHECK:             gpu.terminator
 // CHECK:           }
 // CHECK:           return
 // CHECK:         }
 // CHECK:       }
 
 // -----
 
loop.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
     store %val, %res[%i1, %i0] : memref<?x?xf32>
   } { mapping = [
       {processor = 1, map = affine_map<(d0) -> (d0)>, bound = affine_map<(d0) -> (d0)>},
       {processor = 6, map = affine_map<(d0) -> (d0)>, bound = affine_map<(d0) -> (d0)>}
     ] }
   return
 }
 
-// CHECK:       #map0 = affine_map<(d0) -> (d0)>
-// CHECK:       module {
+// CHECK:       #map0 = affine_map<()[s0, s1, s2] -> (s0 - s1 ceildiv s2)>
+// CHECK:       #map1 = affine_map<(d0)[s0, s1] -> (d0 * s0 + s1)>
 
+// CHECK:       module {
 // CHECK-LABEL:   func @parallel_loop_bidy_seq(
-// CHECK-SAME:                        [[VAL_71:%.*]]: index, [[VAL_72:%.*]]: index, [[VAL_73:%.*]]: index, [[VAL_74:%.*]]: index, [[VAL_75:%.*]]: index, [[VAL_76:%.*]]: memref<?x?xf32>, [[VAL_77:%.*]]: memref<?x?xf32>) {
-// CHECK:           [[VAL_78:%.*]] = constant 2 : index
-// CHECK:           [[VAL_79:%.*]] = constant 1 : index
-// CHECK:           [[VAL_80:%.*]] = subi [[VAL_73]], [[VAL_71]] : index
-// CHECK:           [[VAL_81:%.*]] = affine.apply #map0([[VAL_80]])
-// CHECK:           gpu.launch blocks([[VAL_82:%.*]], [[VAL_83:%.*]], [[VAL_84:%.*]]) in ([[VAL_85:%.*]] = [[VAL_79]], [[VAL_86:%.*]] = [[VAL_81]], [[VAL_87:%.*]] = [[VAL_79]]) threads([[VAL_88:%.*]], [[VAL_89:%.*]], [[VAL_90:%.*]]) in ([[VAL_91:%.*]] = [[VAL_79]], [[VAL_92:%.*]] = [[VAL_79]], [[VAL_93:%.*]] = [[VAL_79]]) {
-// CHECK:             [[VAL_94:%.*]] = affine.apply #map0([[VAL_83]])
-// CHECK:             [[VAL_95:%.*]] = addi [[VAL_94]], [[VAL_71]] : index
-// CHECK:             loop.for [[VAL_96:%.*]] = [[VAL_72]] to [[VAL_74]] step [[VAL_78]] {
-// CHECK:               [[VAL_97:%.*]] = load [[VAL_76]]{{\[}}[[VAL_95]], [[VAL_96]]] : memref<?x?xf32>
-// CHECK:               store [[VAL_97]], [[VAL_77]]{{\[}}[[VAL_96]], [[VAL_95]]] : memref<?x?xf32>
+// CHECK-SAME:                                 [[VAL_59:%.*]]: index, [[VAL_60:%.*]]: index, [[VAL_61:%.*]]: index, [[VAL_62:%.*]]: index, [[VAL_63:%.*]]: index, [[VAL_64:%.*]]: memref<?x?xf32>, [[VAL_65:%.*]]: memref<?x?xf32>) {
+// CHECK:           [[VAL_66:%.*]] = constant 2 : index
+// CHECK:           [[VAL_67:%.*]] = constant 1 : index
+// CHECK:           [[VAL_68:%.*]] = affine.apply #map0(){{\[}}[[VAL_61]], [[VAL_59]], [[VAL_63]]]
+// CHECK:           gpu.launch blocks([[VAL_69:%.*]], [[VAL_70:%.*]], [[VAL_71:%.*]]) in ([[VAL_72:%.*]] = [[VAL_67]], [[VAL_73:%.*]] = [[VAL_68]], [[VAL_74:%.*]] = [[VAL_67]]) threads([[VAL_75:%.*]], [[VAL_76:%.*]], [[VAL_77:%.*]]) in ([[VAL_78:%.*]] = [[VAL_67]], [[VAL_79:%.*]] = [[VAL_67]], [[VAL_80:%.*]] = [[VAL_67]]) {
+// CHECK:             [[VAL_81:%.*]] = affine.apply #map1([[VAL_70]]){{\[}}[[VAL_63]], [[VAL_59]]]
+// CHECK:             loop.for [[VAL_82:%.*]] = [[VAL_60]] to [[VAL_62]] step [[VAL_66]] {
+// CHECK:               [[VAL_83:%.*]] = load [[VAL_64]]{{\[}}[[VAL_81]], [[VAL_82]]] : memref<?x?xf32>
+// CHECK:               store [[VAL_83]], [[VAL_65]]{{\[}}[[VAL_82]], [[VAL_81]]] : memref<?x?xf32>
 // CHECK:             }
 // CHECK:             gpu.terminator
 // CHECK:           }
 // CHECK:           return
 // CHECK:         }
 // CHECK:       }
 
 // -----
loop.parallel (%si0, %si1) = (%zero, %zero) to (%four, %four)
       ] }
   } { mapping = [
       {processor = 1, map = affine_map<(d0) -> (d0)>, bound = affine_map<(d0) -> (d0)>},
       {processor = 6, map = affine_map<(d0) -> (d0)>, bound = affine_map<(d0) -> (d0)>}
     ] }
   return
 }
 
-// CHECK:       #map0 = affine_map<(d0) -> (d0)>
-// CHECK:       module {
+// CHECK:       #map0 = affine_map<()[s0, s1, s2] -> (s0 - s1 ceildiv s2)>
+// CHECK:       #map1 = affine_map<(d0)[s0, s1] -> (d0 * s0 + s1)>
 
+// CHECK:       module {
 // CHECK-LABEL:   func @parallel_loop_tiled_seq(
-// CHECK-SAME:                        [[VAL_98:%.*]]: index, [[VAL_99:%.*]]: index, [[VAL_100:%.*]]: index, [[VAL_101:%.*]]: index, [[VAL_102:%.*]]: memref<?x?xf32>, [[VAL_103:%.*]]: memref<?x?xf32>) {
-// CHECK:           [[VAL_104:%.*]] = constant 0 : index
-// CHECK:           [[VAL_105:%.*]] = constant 1 : index
-// CHECK:           [[VAL_106:%.*]] = constant 4 : index
-// CHECK:           [[VAL_107:%.*]] = constant 1 : index
-// CHECK:           [[VAL_108:%.*]] = subi [[VAL_100]], [[VAL_98]] : index
-// CHECK:           [[VAL_109:%.*]] = affine.apply #map0([[VAL_108]])
-// CHECK:           [[VAL_110:%.*]] = subi [[VAL_106]], [[VAL_104]] : index
-// CHECK:           [[VAL_111:%.*]] = affine.apply #map0([[VAL_110]])
-// CHECK:           gpu.launch blocks([[VAL_112:%.*]], [[VAL_113:%.*]], [[VAL_114:%.*]]) in ([[VAL_115:%.*]] = [[VAL_107]], [[VAL_116:%.*]] = [[VAL_109]], [[VAL_117:%.*]] = [[VAL_107]]) threads([[VAL_118:%.*]], [[VAL_119:%.*]], [[VAL_120:%.*]]) in ([[VAL_121:%.*]] = [[VAL_107]], [[VAL_122:%.*]] = [[VAL_111]], [[VAL_123:%.*]] = [[VAL_107]]) {
-// CHECK:             [[VAL_124:%.*]] = affine.apply #map0([[VAL_113]])
-// CHECK:             [[VAL_125:%.*]] = addi [[VAL_124]], [[VAL_98]] : index
-// CHECK:             loop.for [[VAL_126:%.*]] = [[VAL_99]] to [[VAL_101]] step [[VAL_106]] {
-// CHECK:               [[VAL_127:%.*]] = affine.apply #map0([[VAL_119]])
-// CHECK:               [[VAL_128:%.*]] = addi [[VAL_127]], [[VAL_104]] : index
-// CHECK:               loop.for [[VAL_129:%.*]] = [[VAL_104]] to [[VAL_106]] step [[VAL_105]] {
-// CHECK:                 [[VAL_130:%.*]] = addi [[VAL_125]], [[VAL_128]] : index
-// CHECK:                 [[VAL_131:%.*]] = addi [[VAL_126]], [[VAL_129]] : index
-// CHECK:                 [[VAL_132:%.*]] = load [[VAL_102]]{{\[}}[[VAL_130]], [[VAL_131]]] : memref<?x?xf32>
-// CHECK:                 store [[VAL_132]], [[VAL_103]]{{\[}}[[VAL_131]], [[VAL_130]]] : memref<?x?xf32>
+// CHECK-SAME:                                  [[VAL_84:%.*]]: index, [[VAL_85:%.*]]: index, [[VAL_86:%.*]]: index, [[VAL_87:%.*]]: index, [[VAL_88:%.*]]: memref<?x?xf32>, [[VAL_89:%.*]]: memref<?x?xf32>) {
+// CHECK:           [[VAL_90:%.*]] = constant 0 : index
+// CHECK:           [[VAL_91:%.*]] = constant 1 : index
+// CHECK:           [[VAL_92:%.*]] = constant 4 : index
+// CHECK:           [[VAL_93:%.*]] = constant 1 : index
+// CHECK:           [[VAL_94:%.*]] = affine.apply #map0(){{\[}}[[VAL_86]], [[VAL_84]], [[VAL_92]]]
+// CHECK:           [[VAL_95:%.*]] = affine.apply #map0(){{\[}}[[VAL_92]], [[VAL_90]], [[VAL_91]]]
+// CHECK:           gpu.launch blocks([[VAL_96:%.*]], [[VAL_97:%.*]], [[VAL_98:%.*]]) in ([[VAL_99:%.*]] = [[VAL_93]], [[VAL_100:%.*]] = [[VAL_94]], [[VAL_101:%.*]] = [[VAL_93]]) threads([[VAL_102:%.*]], [[VAL_103:%.*]], [[VAL_104:%.*]]) in ([[VAL_105:%.*]] = [[VAL_93]], [[VAL_106:%.*]] = [[VAL_95]], [[VAL_107:%.*]] = [[VAL_93]]) {
+// CHECK:             [[VAL_108:%.*]] = affine.apply #map1([[VAL_97]]){{\[}}[[VAL_92]], [[VAL_84]]]
+// CHECK:             loop.for [[VAL_109:%.*]] = [[VAL_85]] to [[VAL_87]] step [[VAL_92]] {
+// CHECK:               [[VAL_110:%.*]] = affine.apply #map1([[VAL_103]]){{\[}}[[VAL_91]], [[VAL_90]]]
+// CHECK:               loop.for [[VAL_111:%.*]] = [[VAL_90]] to [[VAL_92]] step [[VAL_91]] {
+// CHECK:                 [[VAL_112:%.*]] = addi [[VAL_108]], [[VAL_110]] : index
+// CHECK:                 [[VAL_113:%.*]] = addi [[VAL_109]], [[VAL_111]] : index
+// CHECK:                 [[VAL_114:%.*]] = load [[VAL_88]]{{\[}}[[VAL_112]], [[VAL_113]]] : memref<?x?xf32>
+// CHECK:                 store [[VAL_114]], [[VAL_89]]{{\[}}[[VAL_113]], [[VAL_112]]] : memref<?x?xf32>
 // CHECK:               }
 // CHECK:             }
 // CHECK:             gpu.terminator
 // CHECK:           }
 // CHECK:           return
 // CHECK:         }
 // CHECK:       }
 
 // -----
 
+// This and the below are the same but with affine.min canonicalized (see https://bugs.llvm.org/show_bug.cgi?id=45008).
+// Once the bug is fixed, this version (and supporting code) can be removed.
+
 #map0 = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)>
 #map1 = affine_map<(d0, d1, d2) -> (d0, d1 - d2)>
 #map2 = affine_map<(d0, d1)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2)>
 #map3 = affine_map<(d0) -> (d0)>
 
 module {
   func @sum(%arg0: memref<?x?xf32, #map0>, %arg1: memref<?x?xf32, #map0>, %arg2: memref<?x?xf32, #map0>) {
     %c1 = constant 1 : index
loop.parallel (%arg3, %arg4) = (%c0, %c0) to (%0, %1) step (%c2, %c3) {
       %12 = dim %arg2, 0 : memref<?x?xf32, #map0>
       %13 = affine.min #map1(%c2, %12, %arg3)
       %14 = dim %arg2, 1 : memref<?x?xf32, #map0>
       %15 = affine.min #map1(%c3, %14, %arg4)
       %16 = std.subview %arg2[%arg3, %arg4][%13, %15][%c1, %c1] : memref<?x?xf32, #map0> to memref<?x?xf32, #map2>
       loop.parallel (%arg5, %arg6) = (%c0, %c0) to (%3, %5) step (%c1, %c1) {
         %17 = load %6[%arg5, %arg6] : memref<?x?xf32, #map2>
         %18 = load %11[%arg5, %arg6] : memref<?x?xf32, #map2>
         %19 = load %16[%arg5, %arg6] : memref<?x?xf32, #map2>
         %20 = addf %17, %18 : f32
         store %20, %16[%arg5, %arg6] : memref<?x?xf32, #map2>
         loop.yield
       } { mapping = [
           {processor = 3, map = #map3, bound = #map3},
           {processor = 4, map = #map3, bound = #map3}
         ] }
       loop.yield
     } { mapping = [
         {processor = 0, map = #map3, bound = #map3},
         {processor = 1, map = #map3, bound = #map3}
     ] }
     return
   }
 }
 
 // CHECK:       #map0 = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)>
-// CHECK:       #map1 = affine_map<(d0) -> (d0)>
-// CHECK:       #map2 = affine_map<(d0, d1, d2) -> (d0, d1 - d2)>
-// CHECK:       #map3 = affine_map<(d0, d1)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2)>
+// CHECK:       #map1 = affine_map<()[s0, s1, s2] -> (s0 - s1 ceildiv s2)>
+// CHECK:       #map2 = affine_map<(d0)[s0, s1] -> (d0 * s0 + s1)>
+// CHECK:       #map3 = affine_map<(d0, d1, d2) -> (d0, d1 - d2)>
+// CHECK:       #map4 = affine_map<(d0, d1)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2)>
 // CHECK:       module {
 
 // CHECK-LABEL:   func @sum(
-// CHECK-SAME:              [[VAL_133:%.*]]: memref<?x?xf32, #map0>, [[VAL_134:%.*]]: memref<?x?xf32, #map0>, [[VAL_135:%.*]]: memref<?x?xf32, #map0>) {
-// CHECK:           [[VAL_136:%.*]] = constant 1 : index
-// CHECK:           [[VAL_137:%.*]] = constant 0 : index
-// CHECK:           [[VAL_138:%.*]] = constant 3 : index
-// CHECK:           [[VAL_139:%.*]] = constant 2 : index
-// CHECK:           [[VAL_140:%.*]] = dim [[VAL_133]], 0 : memref<?x?xf32, #map0>
-// CHECK:           [[VAL_141:%.*]] = dim [[VAL_133]], 1 : memref<?x?xf32, #map0>
-// CHECK:           [[VAL_142:%.*]] = constant 1 : index
-// CHECK:           [[VAL_143:%.*]] = subi [[VAL_140]], [[VAL_137]] : index
-// CHECK:           [[VAL_144:%.*]] = affine.apply #map1([[VAL_143]])
-// CHECK:           [[VAL_145:%.*]] = subi [[VAL_141]], [[VAL_137]] : index
-// CHECK:           [[VAL_146:%.*]] = affine.apply #map1([[VAL_145]])
-// CHECK:           [[VAL_148:%.*]] = subi [[VAL_139]], [[VAL_137]] : index
-// CHECK:           [[VAL_149:%.*]] = affine.apply #map1([[VAL_148]])
-// CHECK:           [[VAL_151:%.*]] = subi [[VAL_138]], [[VAL_137]] : index
-// CHECK:           [[VAL_152:%.*]] = affine.apply #map1([[VAL_151]])
-// CHECK:           gpu.launch blocks([[VAL_153:%.*]], [[VAL_154:%.*]], [[VAL_155:%.*]]) in ([[VAL_156:%.*]] = [[VAL_144]], [[VAL_157:%.*]] = [[VAL_146]], [[VAL_158:%.*]] = [[VAL_142]]) threads([[VAL_159:%.*]], [[VAL_160:%.*]], [[VAL_161:%.*]]) in ([[VAL_162:%.*]] = [[VAL_149]], [[VAL_163:%.*]] = [[VAL_152]], [[VAL_164:%.*]] = [[VAL_142]]) {
-// CHECK:             [[VAL_165:%.*]] = affine.apply #map1([[VAL_153]])
-// CHECK:             [[VAL_166:%.*]] = addi [[VAL_165]], [[VAL_137]] : index
-// CHECK:             [[VAL_167:%.*]] = affine.apply #map1([[VAL_154]])
-// CHECK:             [[VAL_168:%.*]] = addi [[VAL_167]], [[VAL_137]] : index
-// CHECK:             [[VAL_169:%.*]] = dim [[VAL_133]], 0 : memref<?x?xf32, #map0>
-// CHECK:             [[VAL_170:%.*]] = affine.min #map2([[VAL_139]], [[VAL_169]], [[VAL_166]])
-// CHECK:             [[VAL_171:%.*]] = dim [[VAL_133]], 1 : memref<?x?xf32, #map0>
-// CHECK:             [[VAL_172:%.*]] = affine.min #map2([[VAL_138]], [[VAL_171]], [[VAL_168]])
-// CHECK:             [[VAL_173:%.*]] = std.subview [[VAL_133]]{{\[}}[[VAL_166]], [[VAL_168]]]{{\[}}[[VAL_170]], [[VAL_172]]]{{\[}}[[VAL_136]], [[VAL_136]]] : memref<?x?xf32, #map0> to memref<?x?xf32, #map3>
-// CHECK:             [[VAL_174:%.*]] = dim [[VAL_134]], 0 : memref<?x?xf32, #map0>
-// CHECK:             [[VAL_175:%.*]] = affine.min #map2([[VAL_139]], [[VAL_174]], [[VAL_166]])
-// CHECK:             [[VAL_176:%.*]] = dim [[VAL_134]], 1 : memref<?x?xf32, #map0>
-// CHECK:             [[VAL_177:%.*]] = affine.min #map2([[VAL_138]], [[VAL_176]], [[VAL_168]])
-// CHECK:             [[VAL_178:%.*]] = std.subview [[VAL_134]]{{\[}}[[VAL_166]], [[VAL_168]]]{{\[}}[[VAL_175]], [[VAL_177]]]{{\[}}[[VAL_136]], [[VAL_136]]] : memref<?x?xf32, #map0> to memref<?x?xf32, #map3>
-// CHECK:             [[VAL_179:%.*]] = dim [[VAL_135]], 0 : memref<?x?xf32, #map0>
-// CHECK:             [[VAL_180:%.*]] = affine.min #map2([[VAL_139]], [[VAL_179]], [[VAL_166]])
-// CHECK:             [[VAL_181:%.*]] = dim [[VAL_135]], 1 : memref<?x?xf32, #map0>
-// CHECK:             [[VAL_182:%.*]] = affine.min #map2([[VAL_138]], [[VAL_181]], [[VAL_168]])
-// CHECK:             [[VAL_183:%.*]] = std.subview [[VAL_135]]{{\[}}[[VAL_166]], [[VAL_168]]]{{\[}}[[VAL_180]], [[VAL_182]]]{{\[}}[[VAL_136]], [[VAL_136]]] : memref<?x?xf32, #map0> to memref<?x?xf32, #map3>
-// CHECK:             [[VAL_184:%.*]] = affine.apply #map1([[VAL_159]])
-// CHECK:             [[VAL_185:%.*]] = addi [[VAL_184]], [[VAL_137]] : index
-// CHECK:             [[VAL_186:%.*]] = cmpi "slt", [[VAL_185]], [[VAL_170]] : index
-// CHECK:             loop.if [[VAL_186]] {
-// CHECK:               [[VAL_187:%.*]] = affine.apply #map1([[VAL_160]])
-// CHECK:               [[VAL_188:%.*]] = addi [[VAL_187]], [[VAL_137]] : index
-// CHECK:               [[VAL_189:%.*]] = cmpi "slt", [[VAL_188]], [[VAL_172]] : index
-// CHECK:               loop.if [[VAL_189]] {
-// CHECK:                 [[VAL_190:%.*]] = load [[VAL_173]]{{\[}}[[VAL_185]], [[VAL_188]]] : memref<?x?xf32, #map3>
-// CHECK:                 [[VAL_191:%.*]] = load [[VAL_178]]{{\[}}[[VAL_185]], [[VAL_188]]] : memref<?x?xf32, #map3>
-// CHECK:                 [[VAL_192:%.*]] = load [[VAL_183]]{{\[}}[[VAL_185]], [[VAL_188]]] : memref<?x?xf32, #map3>
-// CHECK:                 [[VAL_193:%.*]] = addf [[VAL_190]], [[VAL_191]] : f32
-// CHECK:                 store [[VAL_193]], [[VAL_183]]{{\[}}[[VAL_185]], [[VAL_188]]] : memref<?x?xf32, #map3>
+// CHECK-SAME:              [[VAL_115:%.*]]: memref<?x?xf32, #map0>, [[VAL_116:%.*]]: memref<?x?xf32, #map0>, [[VAL_117:%.*]]: memref<?x?xf32, #map0>) {
+// CHECK:           [[VAL_118:%.*]] = constant 1 : index
+// CHECK:           [[VAL_119:%.*]] = constant 0 : index
+// CHECK:           [[VAL_120:%.*]] = constant 3 : index
+// CHECK:           [[VAL_121:%.*]] = constant 2 : index
+// CHECK:           [[VAL_122:%.*]] = dim [[VAL_115]], 0 : memref<?x?xf32, #map0>
+// CHECK:           [[VAL_123:%.*]] = dim [[VAL_115]], 1 : memref<?x?xf32, #map0>
+// CHECK:           [[VAL_124:%.*]] = constant 1 : index
+// CHECK:           [[VAL_125:%.*]] = affine.apply #map1(){{\[}}[[VAL_122]], [[VAL_119]], [[VAL_121]]]
+// CHECK:           [[VAL_126:%.*]] = affine.apply #map1(){{\[}}[[VAL_123]], [[VAL_119]], [[VAL_120]]]
+// CHECK:           [[VAL_127:%.*]] = affine.apply #map1(){{\[}}[[VAL_121]], [[VAL_119]], [[VAL_118]]]
+// CHECK:           [[VAL_128:%.*]] = affine.apply #map1(){{\[}}[[VAL_120]], [[VAL_119]], [[VAL_118]]]
+// CHECK:           gpu.launch blocks([[VAL_129:%.*]], [[VAL_130:%.*]], [[VAL_131:%.*]]) in ([[VAL_132:%.*]] = [[VAL_125]], [[VAL_133:%.*]] = [[VAL_126]], [[VAL_134:%.*]] = [[VAL_124]]) threads([[VAL_135:%.*]], [[VAL_136:%.*]], [[VAL_137:%.*]]) in ([[VAL_138:%.*]] = [[VAL_127]], [[VAL_139:%.*]] = [[VAL_128]], [[VAL_140:%.*]] = [[VAL_124]]) {
+// CHECK:             [[VAL_141:%.*]] = affine.apply #map2([[VAL_129]]){{\[}}[[VAL_121]], [[VAL_119]]]
+// CHECK:             [[VAL_142:%.*]] = affine.apply #map2([[VAL_130]]){{\[}}[[VAL_120]], [[VAL_119]]]
+// CHECK:             [[VAL_143:%.*]] = dim [[VAL_115]], 0 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_144:%.*]] = affine.min #map3([[VAL_121]], [[VAL_143]], [[VAL_141]])
+// CHECK:             [[VAL_145:%.*]] = dim [[VAL_115]], 1 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_146:%.*]] = affine.min #map3([[VAL_120]], [[VAL_145]], [[VAL_142]])
+// CHECK:             [[VAL_147:%.*]] = std.subview [[VAL_115]]{{\[}}[[VAL_141]], [[VAL_142]]]{{\[}}[[VAL_144]], [[VAL_146]]]{{\[}}[[VAL_118]], [[VAL_118]]] : memref<?x?xf32, #map0> to memref<?x?xf32, #map4>
+// CHECK:             [[VAL_148:%.*]] = dim [[VAL_116]], 0 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_149:%.*]] = affine.min #map3([[VAL_121]], [[VAL_148]], [[VAL_141]])
+// CHECK:             [[VAL_150:%.*]] = dim [[VAL_116]], 1 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_151:%.*]] = affine.min #map3([[VAL_120]], [[VAL_150]], [[VAL_142]])
+// CHECK:             [[VAL_152:%.*]] = std.subview [[VAL_116]]{{\[}}[[VAL_141]], [[VAL_142]]]{{\[}}[[VAL_149]], [[VAL_151]]]{{\[}}[[VAL_118]], [[VAL_118]]] : memref<?x?xf32, #map0> to memref<?x?xf32, #map4>
+// CHECK:             [[VAL_153:%.*]] = dim [[VAL_117]], 0 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_154:%.*]] = affine.min #map3([[VAL_121]], [[VAL_153]], [[VAL_141]])
+// CHECK:             [[VAL_155:%.*]] = dim [[VAL_117]], 1 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_156:%.*]] = affine.min #map3([[VAL_120]], [[VAL_155]], [[VAL_142]])
+// CHECK:             [[VAL_157:%.*]] = std.subview [[VAL_117]]{{\[}}[[VAL_141]], [[VAL_142]]]{{\[}}[[VAL_154]], [[VAL_156]]]{{\[}}[[VAL_118]], [[VAL_118]]] : memref<?x?xf32, #map0> to memref<?x?xf32, #map4>
+// CHECK:             [[VAL_158:%.*]] = affine.apply #map2([[VAL_135]]){{\[}}[[VAL_118]], [[VAL_119]]]
+// CHECK:             [[VAL_159:%.*]] = cmpi "slt", [[VAL_158]], [[VAL_144]] : index
+// CHECK:             loop.if [[VAL_159]] {
+// CHECK:               [[VAL_160:%.*]] = affine.apply #map2([[VAL_136]]){{\[}}[[VAL_118]], [[VAL_119]]]
+// CHECK:               [[VAL_161:%.*]] = cmpi "slt", [[VAL_160]], [[VAL_146]] : index
+// CHECK:               loop.if [[VAL_161]] {
+// CHECK:                 [[VAL_162:%.*]] = load [[VAL_147]]{{\[}}[[VAL_158]], [[VAL_160]]] : memref<?x?xf32, #map4>
+// CHECK:                 [[VAL_163:%.*]] = load [[VAL_152]]{{\[}}[[VAL_158]], [[VAL_160]]] : memref<?x?xf32, #map4>
+// CHECK:                 [[VAL_164:%.*]] = load [[VAL_157]]{{\[}}[[VAL_158]], [[VAL_160]]] : memref<?x?xf32, #map4>
+// CHECK:                 [[VAL_165:%.*]] = addf [[VAL_162]], [[VAL_163]] : f32
+// CHECK:                 store [[VAL_165]], [[VAL_157]]{{\[}}[[VAL_158]], [[VAL_160]]] : memref<?x?xf32, #map4>
 // CHECK:               }
 // CHECK:             }
 // CHECK:             gpu.terminator
 // CHECK:           }
 // CHECK:           return
 // CHECK:         }
 // CHECK:       }
 
+// -----
+
+#map0 = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)>
+#map1 = affine_map<(d0, d1, d2) -> (d0, d1 - d2)>
+#map2 = affine_map<(d0, d1)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2)>
+#map3 = affine_map<(d0) -> (d0)>
+#map4 = affine_map<(d1, d2) -> (2, d1 - d2)>
+#map5 = affine_map<(d1, d2) -> (3, d1 - d2)>
+
+
+module {
+  func @sum2(%arg0: memref<?x?xf32, #map0>, %arg1: memref<?x?xf32, #map0>, %arg2: memref<?x?xf32, #map0>) {
+    %c1 = constant 1 : index
+    %c0 = constant 0 : index
+    %c3 = constant 3 : index
+    %c2 = constant 2 : index
+    %0 = dim %arg0, 0 : memref<?x?xf32, #map0>
+    %1 = dim %arg0, 1 : memref<?x?xf32, #map0>
+    loop.parallel (%arg3, %arg4) = (%c0, %c0) to (%0, %1) step (%c2, %c3) {
+      %2 = dim %arg0, 0 : memref<?x?xf32, #map0>
+      %3 = affine.min #map4(%2, %arg3)
+      %4 = dim %arg0, 1 : memref<?x?xf32, #map0>
+      %5 = affine.min #map5(%4, %arg4)
+      %6 = std.subview %arg0[%arg3, %arg4][%3, %5][%c1, %c1] : memref<?x?xf32, #map0> to memref<?x?xf32, #map2>
+      %7 = dim %arg1, 0 : memref<?x?xf32, #map0>
+      %8 = affine.min #map4(%7, %arg3)
+      %9 = dim %arg1, 1 : memref<?x?xf32, #map0>
+      %10 = affine.min #map5(%9, %arg4)
+      %11 = std.subview %arg1[%arg3, %arg4][%8, %10][%c1, %c1] : memref<?x?xf32, #map0> to memref<?x?xf32, #map2>
+      %12 = dim %arg2, 0 : memref<?x?xf32, #map0>
+      %13 = affine.min #map4(%12, %arg3)
+      %14 = dim %arg2, 1 : memref<?x?xf32, #map0>
+      %15 = affine.min #map5(%14, %arg4)
+      %16 = std.subview %arg2[%arg3, %arg4][%13, %15][%c1, %c1] : memref<?x?xf32, #map0> to memref<?x?xf32, #map2>
+      loop.parallel (%arg5, %arg6) = (%c0, %c0) to (%3, %5) step (%c1, %c1) {
+        %17 = load %6[%arg5, %arg6] : memref<?x?xf32, #map2>
+        %18 = load %11[%arg5, %arg6] : memref<?x?xf32, #map2>
+        %19 = load %16[%arg5, %arg6] : memref<?x?xf32, #map2>
+        %20 = addf %17, %18 : f32
+        store %20, %16[%arg5, %arg6] : memref<?x?xf32, #map2>
+        loop.yield
+      } { mapping = [
+          {processor = 3, map = #map3, bound = #map3},
+          {processor = 4, map = #map3, bound = #map3}
+        ] }
+      loop.yield
+    } { mapping = [
+        {processor = 0, map = #map3, bound = #map3},
+        {processor = 1, map = #map3, bound = #map3}
+    ] }
+    return
+  }
+}
+
+// CHECK:       #map1 = affine_map<()[s0, s1, s2] -> (s0 - s1 ceildiv s2)>
+// CHECK:       #map2 = affine_map<(d0)[s0, s1] -> (d0 * s0 + s1)>
+// CHECK:       #map3 = affine_map<(d0, d1) -> (2, d0 - d1)>
+// CHECK:       #map4 = affine_map<(d0, d1) -> (3, d0 - d1)>
+// CHECK:       #map5 = affine_map<(d0, d1)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2)>
+
+// CHECK:       module {
+// CHECK-LABEL:   func @sum2(
+// CHECK-SAME:               [[VAL_166:%.*]]: memref<?x?xf32, #map0>, [[VAL_167:%.*]]: memref<?x?xf32, #map0>, [[VAL_168:%.*]]: memref<?x?xf32, #map0>) {
+// CHECK:           [[VAL_169:%.*]] = constant 1 : index
+// CHECK:           [[VAL_170:%.*]] = constant 0 : index
+// CHECK:           [[VAL_171:%.*]] = constant 3 : index
+// CHECK:           [[VAL_172:%.*]] = constant 2 : index
+// CHECK:           [[VAL_173:%.*]] = dim [[VAL_166]], 0 : memref<?x?xf32, #map0>
+// CHECK:           [[VAL_174:%.*]] = dim [[VAL_166]], 1 : memref<?x?xf32, #map0>
+// CHECK:           [[VAL_175:%.*]] = constant 1 : index
+// CHECK:           [[VAL_176:%.*]] = affine.apply #map1(){{\[}}[[VAL_173]], [[VAL_170]], [[VAL_172]]]
+// CHECK:           [[VAL_177:%.*]] = affine.apply #map1(){{\[}}[[VAL_174]], [[VAL_170]], [[VAL_171]]]
+// CHECK:           [[VAL_178:%.*]] = constant 2 : index
+// CHECK:           [[VAL_179:%.*]] = affine.apply #map1(){{\[}}[[VAL_178]], [[VAL_170]], [[VAL_169]]]
+// CHECK:           [[VAL_180:%.*]] = constant 3 : index
+// CHECK:           [[VAL_181:%.*]] = affine.apply #map1(){{\[}}[[VAL_180]], [[VAL_170]], [[VAL_169]]]
+// CHECK:           gpu.launch blocks([[VAL_182:%.*]], [[VAL_183:%.*]], [[VAL_184:%.*]]) in ([[VAL_185:%.*]] = [[VAL_176]], [[VAL_186:%.*]] = [[VAL_177]], [[VAL_187:%.*]] = [[VAL_175]]) threads([[VAL_188:%.*]], [[VAL_189:%.*]], [[VAL_190:%.*]]) in ([[VAL_191:%.*]] = [[VAL_179]], [[VAL_192:%.*]] = [[VAL_181]], [[VAL_193:%.*]] = [[VAL_175]]) {
+// CHECK:             [[VAL_194:%.*]] = affine.apply #map2([[VAL_182]]){{\[}}[[VAL_172]], [[VAL_170]]]
+// CHECK:             [[VAL_195:%.*]] = affine.apply #map2([[VAL_183]]){{\[}}[[VAL_171]], [[VAL_170]]]
+// CHECK:             [[VAL_196:%.*]] = dim [[VAL_166]], 0 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_197:%.*]] = affine.min #map3([[VAL_196]], [[VAL_194]])
+// CHECK:             [[VAL_198:%.*]] = dim [[VAL_166]], 1 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_199:%.*]] = affine.min #map4([[VAL_198]], [[VAL_195]])
+// CHECK:             [[VAL_200:%.*]] = std.subview [[VAL_166]]{{\[}}[[VAL_194]], [[VAL_195]]]{{\[}}[[VAL_197]], [[VAL_199]]]{{\[}}[[VAL_169]], [[VAL_169]]] : memref<?x?xf32, #map0> to memref<?x?xf32, #map5>
+// CHECK:             [[VAL_201:%.*]] = dim [[VAL_167]], 0 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_202:%.*]] = affine.min #map3([[VAL_201]], [[VAL_194]])
+// CHECK:             [[VAL_203:%.*]] = dim [[VAL_167]], 1 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_204:%.*]] = affine.min #map4([[VAL_203]], [[VAL_195]])
+// CHECK:             [[VAL_205:%.*]] = std.subview [[VAL_167]]{{\[}}[[VAL_194]], [[VAL_195]]]{{\[}}[[VAL_202]], [[VAL_204]]]{{\[}}[[VAL_169]], [[VAL_169]]] : memref<?x?xf32, #map0> to memref<?x?xf32, #map5>
+// CHECK:             [[VAL_206:%.*]] = dim [[VAL_168]], 0 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_207:%.*]] = affine.min #map3([[VAL_206]], [[VAL_194]])
+// CHECK:             [[VAL_208:%.*]] = dim [[VAL_168]], 1 : memref<?x?xf32, #map0>
+// CHECK:             [[VAL_209:%.*]] = affine.min #map4([[VAL_208]], [[VAL_195]])
+// CHECK:             [[VAL_210:%.*]] = std.subview [[VAL_168]]{{\[}}[[VAL_194]], [[VAL_195]]]{{\[}}[[VAL_207]], [[VAL_209]]]{{\[}}[[VAL_169]], [[VAL_169]]] : memref<?x?xf32, #map0> to memref<?x?xf32, #map5>
+// CHECK:             [[VAL_211:%.*]] = affine.apply #map2([[VAL_188]]){{\[}}[[VAL_169]], [[VAL_170]]]
+// CHECK:             [[VAL_212:%.*]] = cmpi "slt", [[VAL_211]], [[VAL_197]] : index
+// CHECK:             loop.if [[VAL_212]] {
+// CHECK:               [[VAL_213:%.*]] = affine.apply #map2([[VAL_189]]){{\[}}[[VAL_169]], [[VAL_170]]]
+// CHECK:               [[VAL_214:%.*]] = cmpi "slt", [[VAL_213]], [[VAL_199]] : index
+// CHECK:               loop.if [[VAL_214]] {
+// CHECK:                 [[VAL_215:%.*]] = load [[VAL_200]]{{\[}}[[VAL_211]], [[VAL_213]]] : memref<?x?xf32, #map5>
+// CHECK:                 [[VAL_216:%.*]] = load [[VAL_205]]{{\[}}[[VAL_211]], [[VAL_213]]] : memref<?x?xf32, #map5>
+// CHECK:                 [[VAL_217:%.*]] = load [[VAL_210]]{{\[}}[[VAL_211]], [[VAL_213]]] : memref<?x?xf32, #map5>
+// CHECK:                 [[VAL_218:%.*]] = addf [[VAL_215]], [[VAL_216]] : f32
+// CHECK:                 store [[VAL_218]], [[VAL_210]]{{\[}}[[VAL_211]], [[VAL_213]]] : memref<?x?xf32, #map5>
+// CHECK:               }
+// CHECK:             }
+// CHECK:             gpu.terminator
+// CHECK:           }
+// CHECK:           return
+// CHECK:         }
+// CHECK:       }