[mlir][scf]: Expose emitNormalizedLoopBounds/denormalizeInductionVariable util functions

* Also updated normarlize/denormalize loop bounds to be folded if
possible.

Author: Aviad Cohen

mlir/include/mlir/Dialect/SCF/Utils/Utils.h

@@ -120,6 +120,31 @@ LogicalResult loopUnrollByFactor(
     scf::ForOp forOp, uint64_t unrollFactor,
     function_ref<void(unsigned, Operation *, OpBuilder)> annotateFn = nullptr);
 
+/// This structure is to pass and return sets of loop parameters without
+/// confusing the order.
+struct LoopParams {
+  OpFoldResult lowerBound;
+  OpFoldResult upperBound;
+  OpFoldResult step;
+};
+
+/// Transform a loop with a strictly positive step
+///   for %i = %lb to %ub step %s
+/// into a 0-based loop with step 1
+///   for %ii = 0 to ceildiv(%ub - %lb, %s) step 1 {
+///     %i = %ii * %s + %lb
+/// Insert the induction variable remapping in the body of `inner`, which is
+/// expected to be either `loop` or another loop perfectly nested under `loop`.
+/// Insert the definition of new bounds immediate before `outer`, which is
+/// expected to be either `loop` or its parent in the loop nest.
+LoopParams emitNormalizedLoopBounds(RewriterBase &rewriter, Location loc,
+                                    Value lb, Value ub, Value step);
+
+/// Get back the original induction variable values after loop normalization.
+void denormalizeInductionVariable(RewriterBase &rewriter, Location loc,
+                                  Value normalizedIv, Value origLb,
+                                  Value origStep);
+
 /// Tile a nest of standard for loops rooted at `rootForOp` by finding such
 /// parametric tile sizes that the outer loops have a fixed number of iterations
 /// as defined in `sizes`.

mlir/lib/Dialect/SCF/Utils/Utils.cpp

@@ -18,6 +18,7 @@
 #include "mlir/Dialect/SCF/IR/SCF.h"
 #include "mlir/IR/BuiltinOps.h"
 #include "mlir/IR/IRMapping.h"
+#include "mlir/IR/OpDefinition.h"
 #include "mlir/IR/PatternMatch.h"
 #include "mlir/Interfaces/SideEffectInterfaces.h"
 #include "mlir/Support/MathExtras.h"
@@ -29,16 +30,6 @@
 
 using namespace mlir;
 
-namespace {
-// This structure is to pass and return sets of loop parameters without
-// confusing the order.
-struct LoopParams {
-  Value lowerBound;
-  Value upperBound;
-  Value step;
-};
-} // namespace
-
 SmallVector<scf::ForOp> mlir::replaceLoopNestWithNewYields(
     RewriterBase &rewriter, MutableArrayRef<scf::ForOp> loopNest,
     ValueRange newIterOperands, const NewYieldValuesFn &newYieldValuesFn,
@@ -473,17 +464,8 @@ LogicalResult mlir::loopUnrollByFactor(
   return success();
 }
 
-/// Transform a loop with a strictly positive step
-///   for %i = %lb to %ub step %s
-/// into a 0-based loop with step 1
-///   for %ii = 0 to ceildiv(%ub - %lb, %s) step 1 {
-///     %i = %ii * %s + %lb
-/// Insert the induction variable remapping in the body of `inner`, which is
-/// expected to be either `loop` or another loop perfectly nested under `loop`.
-/// Insert the definition of new bounds immediate before `outer`, which is
-/// expected to be either `loop` or its parent in the loop nest.
-static LoopParams emitNormalizedLoopBounds(RewriterBase &rewriter, Location loc,
-                                           Value lb, Value ub, Value step) {
+LoopParams mlir::emitNormalizedLoopBounds(RewriterBase &rewriter, Location loc,
+                                          Value lb, Value ub, Value step) {
   // For non-index types, generate `arith` instructions
   // Check if the loop is already known to have a constant zero lower bound or
   // a constant one step.
@@ -501,26 +483,27 @@ static LoopParams emitNormalizedLoopBounds(RewriterBase &rewriter, Location loc,
   if (isZeroBased && isStepOne)
     return {lb, ub, step};
 
-  Value diff = isZeroBased ? ub : rewriter.create<arith::SubIOp>(loc, ub, lb);
-  Value newUpperBound =
-      isStepOne ? diff : rewriter.create<arith::CeilDivSIOp>(loc, diff, step);
-
-  Value newLowerBound = isZeroBased
-                            ? lb
-                            : rewriter.create<arith::ConstantOp>(
-                                  loc, rewriter.getZeroAttr(lb.getType()));
-  Value newStep = isStepOne
-                      ? step
-                      : rewriter.create<arith::ConstantOp>(
-                            loc, rewriter.getIntegerAttr(step.getType(), 1));
+  auto diff =
+      isZeroBased ? ub : rewriter.createOrFold<arith::SubIOp>(loc, ub, lb);
+  auto newUpperBound =
+      isStepOne ? diff
+                : rewriter.createOrFold<arith::CeilDivSIOp>(loc, diff, step);
+
+  auto newLowerBound = isZeroBased
+                           ? lb
+                           : rewriter.createOrFold<arith::ConstantOp>(
+                                 loc, rewriter.getZeroAttr(lb.getType()));
+  auto newStep = isStepOne
+                     ? step
+                     : rewriter.createOrFold<arith::ConstantOp>(
+                           loc, rewriter.getIntegerAttr(step.getType(), 1));
 
   return {newLowerBound, newUpperBound, newStep};
 }
 
-/// Get back the original induction variable values after loop normalization
-static void denormalizeInductionVariable(RewriterBase &rewriter, Location loc,
-                                         Value normalizedIv, Value origLb,
-                                         Value origStep) {
+void mlir::denormalizeInductionVariable(RewriterBase &rewriter, Location loc,
+                                        Value normalizedIv, Value origLb,
+                                        Value origStep) {
   Value denormalizedIv;
   SmallPtrSet<Operation *, 2> preserve;
   bool isStepOne = isConstantIntValue(origStep, 1);
@@ -638,9 +621,9 @@ LogicalResult mlir::coalesceLoops(RewriterBase &rewriter,
         emitNormalizedLoopBounds(rewriter, loop.getLoc(), lb, ub, step);
 
     rewriter.modifyOpInPlace(loop, [&]() {
-      loop.setLowerBound(newLoopParams.lowerBound);
-      loop.setUpperBound(newLoopParams.upperBound);
-      loop.setStep(newLoopParams.step);
+      loop.setLowerBound(cast<Value>(newLoopParams.lowerBound));
+      loop.setUpperBound(cast<Value>(newLoopParams.upperBound));
+      loop.setStep(cast<Value>(newLoopParams.step));
     });
 
     rewriter.setInsertionPointToStart(innermost.getBody());
@@ -778,18 +761,15 @@ void mlir::collapseParallelLoops(
     llvm::sort(dims);
 
   // Normalize ParallelOp's iteration pattern.
-  SmallVector<Value, 3> normalizedLowerBounds, normalizedSteps,
-      normalizedUpperBounds;
+  SmallVector<Value, 3> normalizedUpperBounds;
   for (unsigned i = 0, e = loops.getNumLoops(); i < e; ++i) {
     OpBuilder::InsertionGuard g2(rewriter);
     rewriter.setInsertionPoint(loops);
     Value lb = loops.getLowerBound()[i];
     Value ub = loops.getUpperBound()[i];
     Value step = loops.getStep()[i];
     auto newLoopParams = emitNormalizedLoopBounds(rewriter, loc, lb, ub, step);
-    normalizedLowerBounds.push_back(newLoopParams.lowerBound);
-    normalizedUpperBounds.push_back(newLoopParams.upperBound);
-    normalizedSteps.push_back(newLoopParams.step);
+    normalizedUpperBounds.push_back(cast<Value>(newLoopParams.upperBound));
 
     rewriter.setInsertionPointToStart(loops.getBody());
     denormalizeInductionVariable(rewriter, loc, loops.getInductionVars()[i], lb,

mlir/test/Dialect/Affine/loop-coalescing.mlir

@@ -74,32 +74,27 @@ func.func @multi_use() {
 
 func.func @unnormalized_loops() {
   // CHECK: %[[orig_step_i:.*]] = arith.constant 2
-  // CHECK: %[[orig_step_j:.*]] = arith.constant 3
+
+  // CHECK: %[[orig_step_j_and_numiter_i:.*]] = arith.constant 3
   // CHECK: %[[orig_lb_i:.*]] = arith.constant 5
   // CHECK: %[[orig_lb_j:.*]] = arith.constant 7
-  // CHECK: %[[orig_ub_i:.*]] = arith.constant 10
-  // CHECK: %[[orig_ub_j:.*]] = arith.constant 17
   %c2 = arith.constant 2 : index
   %c3 = arith.constant 3 : index
   %c5 = arith.constant 5 : index
   %c7 = arith.constant 7 : index
   %c10 = arith.constant 10 : index
   %c17 = arith.constant 17 : index
 
-  // Number of iterations in the outer scf.
-  // CHECK: %[[diff_i:.*]] = arith.subi %[[orig_ub_i]], %[[orig_lb_i]]
-  // CHECK: %[[numiter_i:.*]] = arith.ceildivsi %[[diff_i]], %[[orig_step_i]]
-
   // Normalized lower bound and step for the outer scf.
   // CHECK: %[[lb_i:.*]] = arith.constant 0
   // CHECK: %[[step_i:.*]] = arith.constant 1
 
   // Number of iterations in the inner loop, the pattern is the same as above,
   // only capture the final result.
-  // CHECK: %[[numiter_j:.*]] = arith.ceildivsi {{.*}}, %[[orig_step_j]]
+  // CHECK: %[[numiter_j:.*]] = arith.constant 4
 
   // New bounds of the outer scf.
-  // CHECK: %[[range:.*]] = arith.muli %[[numiter_i]], %[[numiter_j]]
+  // CHECK: %[[range:.*]] = arith.muli %[[orig_step_j_and_numiter_i:.*]], %[[numiter_j]]
   // CHECK: scf.for %[[i:.*]] = %[[lb_i]] to %[[range]] step %[[step_i]]
   scf.for %i = %c5 to %c10 step %c2 {
     // The inner loop has been removed.
@@ -108,7 +103,7 @@ func.func @unnormalized_loops() {
       // The IVs are rewritten.
       // CHECK: %[[normalized_j:.*]] = arith.remsi %[[i]], %[[numiter_j]]
       // CHECK: %[[normalized_i:.*]] = arith.divsi %[[i]], %[[numiter_j]]
-      // CHECK: %[[scaled_j:.*]] = arith.muli %[[normalized_j]], %[[orig_step_j]]
+      // CHECK: %[[scaled_j:.*]] = arith.muli %[[normalized_j]], %[[orig_step_j_and_numiter_i]]
       // CHECK: %[[orig_j:.*]] = arith.addi %[[scaled_j]], %[[orig_lb_j]]
       // CHECK: %[[scaled_i:.*]] = arith.muli %[[normalized_i]], %[[orig_step_i]]
       // CHECK: %[[orig_i:.*]] = arith.addi %[[scaled_i]], %[[orig_lb_i]]