LICM pass zero trip count loop handling; zero trip count loop removal in

-simplify-affine-structures - addresses tensorflow#194 - change name of tablegen auto-generated internal helper for op interfaces to avoid potential conflicts with methods of same name in dialect namespaces. (addresses tensorflow#197) Signed-off-by: Uday Bondhugula <[email protected]>
bondhugula · Oct 23, 2019 · e3c2e48 · e3c2e48
1 parent 705a743
commit e3c2e48
Show file tree

Hide file tree

Showing 11 changed files with 144 additions and 33 deletions.
diff --git a/include/mlir/Dialect/LoopOps/LoopOps.h b/include/mlir/Dialect/LoopOps/LoopOps.h
@@ -52,6 +52,8 @@ void ensureLoopTerminator(Region &region, Builder &builder, Location loc);
 /// not an induction variable, then return nullptr.
 ForOp getForInductionVarOwner(Value *val);
 
+/// Returns the trip count of the loop if it's a constant, None otherwise.
+Optional<uint64_t> getConstantTripCount(ForOp forOp);
 } // end namespace loop
 } // end namespace mlir
 #endif // MLIR_LOOPOPS_OPS_H_
diff --git a/include/mlir/Transforms/LoopLikeInterface.h b/include/mlir/Transforms/LoopLikeInterface.h
@@ -22,6 +22,7 @@
 #ifndef MLIR_TRANSFORMS_LOOPLIKEINTERFACE_H_
 #define MLIR_TRANSFORMS_LOOPLIKEINTERFACE_H_
 
+#include "mlir/Analysis/LoopAnalysis.h"
 #include "mlir/IR/OpDefinition.h"
 #include "mlir/Support/LogicalResult.h"
 #include "llvm/ADT/ArrayRef.h"

diff --git a/include/mlir/Transforms/LoopLikeInterface.td b/include/mlir/Transforms/LoopLikeInterface.td
@@ -56,6 +56,10 @@ def LoopLikeOpInterface : OpInterface<"LoopLikeOpInterface"> {
       }],
       "LogicalResult", "moveOutOfLoop", (ins "ArrayRef<Operation *>":$ops)
     >,
+    InterfaceMethod<[{ Get the trip count if it is a constant. }],
+      "llvm::Optional<uint64_t>", "getConstantTripCount", (ins), [{
+        return getConstantTripCount(op);
+    }]>,
   ];
 }
 

diff --git a/lib/Dialect/LoopOps/LoopOps.cpp b/lib/Dialect/LoopOps/LoopOps.cpp
@@ -140,7 +140,7 @@ bool ForOp::isDefinedOutsideOfLoop(Value *value) {
 
 LogicalResult ForOp::moveOutOfLoop(ArrayRef<Operation *> ops) {
   for (auto *op : ops)
-    op->moveBefore(this->getOperation());
+    op->moveBefore(*this);
   return success();
 }
 
@@ -153,6 +153,31 @@ ForOp mlir::loop::getForInductionVarOwner(Value *val) {
   return dyn_cast_or_null<ForOp>(containingInst);
 }
 
+Optional<uint64_t> mlir::loop::getConstantTripCount(ForOp forOp) {
+  Value *lb = forOp.lowerBound();
+  Value *ub = forOp.upperBound();
+
+  if (lb == ub)
+    return 0;
+
+  IntegerAttr lbCst, ubCst, step;
+  if (!matchPattern(lb, m_Constant(&lbCst)) ||
+      !matchPattern(ub, m_Constant(&ubCst)))
+    return llvm::None;
+
+  int64_t lbConst = lbCst.getValue().getSExtValue();
+  int64_t ubConst = ubCst.getValue().getSExtValue();
+  if (ubConst - lbConst <= 0)
+    return 0;
+
+  if (!matchPattern(forOp.step(), m_Constant(&step)))
+    return llvm::None;
+
+  // Step is guaranteed to be positive.
+  int64_t stepConst = step.getValue().getSExtValue();
+  return llvm::divideCeil(ubConst - lbConst, stepConst);
+}
+
 //===----------------------------------------------------------------------===//
 // IfOp
 //===----------------------------------------------------------------------===//

diff --git a/lib/Transforms/AffineLoopInvariantCodeMotion.cpp b/lib/Transforms/AffineLoopInvariantCodeMotion.cpp
@@ -40,17 +40,17 @@
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/raw_ostream.h"
 
-#define DEBUG_TYPE "licm"
+#define DEBUG_TYPE "affine-licm"
 
 using namespace mlir;
 
 namespace {
 
-/// Loop invariant code motion (LICM) pass.
+/// Affine loop invariant code motion (LICM) pass.
+/// TODO: This pass should be removed once the new LICM pass can handle its
+///       uses.
 /// TODO(asabne) : The pass is missing zero-trip tests.
 /// TODO(asabne) : Check for the presence of side effects before hoisting.
-/// TODO: This code should be removed once the new LICM pass can handle its
-///       uses.
 struct LoopInvariantCodeMotion : public FunctionPass<LoopInvariantCodeMotion> {
   void runOnFunction() override;
   void runOnAffineForOp(AffineForOp forOp);
@@ -245,4 +245,4 @@ mlir::createAffineLoopInvariantCodeMotionPass() {
 
 static PassRegistration<LoopInvariantCodeMotion>
     pass("affine-loop-invariant-code-motion",
-         "Hoist loop invariant instructions outside of the loop");
+         "Hoist loop invariant operations outside of the loop");
diff --git a/lib/Transforms/LoopInvariantCodeMotion.cpp b/lib/Transforms/LoopInvariantCodeMotion.cpp
@@ -62,17 +62,15 @@ static bool canBeHoisted(Operation *op,
   auto thisOpIsSideEffecting = sideEffecting;
   if (thisOpIsSideEffecting != SideEffecting::Never) {
     thisOpIsSideEffecting = interface.isSideEffecting(op);
-    // If the op always has sideeffects, we cannot hoist.
+    // If the op always has side effects, we cannot hoist.
     if (thisOpIsSideEffecting == SideEffecting::Always)
       return false;
   }
   // Recurse into the regions for this op and check whether the contained ops
   // can be hoisted.
   for (auto &region : op->getRegions()) {
     for (auto &block : region.getBlocks()) {
-      for (auto &innerOp : block) {
-        if (innerOp.isKnownTerminator())
-          continue;
+      for (auto &innerOp : block.without_terminator()) {
         if (!canBeHoisted(&innerOp, definedOutside, thisOpIsSideEffecting,
                           interface))
           return false;
@@ -112,7 +110,7 @@ static LogicalResult moveLoopInvariantCode(LoopLikeOpInterface looplike,
     }
   }
 
-  // For all instructions that we found to be invariant, move outside of the
+  // For all operations that we found to be invariant, move outside of the
   // loop.
   auto result = looplike.moveOutOfLoop(opsToMove);
   LLVM_DEBUG(looplike.print(llvm::dbgs() << "Modified loop\n"));
@@ -128,6 +126,13 @@ void LoopInvariantCodeMotion::runOnOperation() {
   // the outer loop, which in turn can be further LICM'ed.
   getOperation()->walk([&](Operation *op) {
     if (auto looplike = dyn_cast<LoopLikeOpInterface>(op)) {
+      // Skip zero trip count loops. For unknown trip counts, we still move
+      // invariant code since it is side-effect free, and in general profitable.
+      // TODO: when necessary, we could only move when the trip count is
+      // guaranteed to be at least one.
+      auto tripCount = looplike.getConstantTripCount();
+      if (tripCount == 0UL)
+        return;
       LLVM_DEBUG(op->print(llvm::dbgs() << "\nOriginal loop\n"));
       if (failed(moveLoopInvariantCode(looplike, interface)))
         signalPassFailure();
@@ -146,4 +151,4 @@ std::unique_ptr<Pass> mlir::createLoopInvariantCodeMotionPass() {
 
 static PassRegistration<LoopInvariantCodeMotion>
     pass("loop-invariant-code-motion",
-         "Hoist loop invariant instructions outside of the loop");
+         "Hoist loop invariant operations outside of the loop");
diff --git a/lib/Transforms/SimplifyAffineStructures.cpp b/lib/Transforms/SimplifyAffineStructures.cpp
@@ -22,6 +22,7 @@
 #include "mlir/Analysis/AffineStructures.h"
 #include "mlir/IR/IntegerSet.h"
 #include "mlir/Pass/Pass.h"
+#include "mlir/Transforms/LoopLikeInterface.h"
 #include "mlir/Transforms/Passes.h"
 #include "mlir/Transforms/Utils.h"
 
@@ -93,12 +94,12 @@ std::unique_ptr<OpPassBase<FuncOp>> mlir::createSimplifyAffineStructuresPass() {
 void SimplifyAffineStructures::runOnFunction() {
   auto func = getFunction();
   simplifiedAttributes.clear();
-  func.walk([&](Operation *opInst) {
-    for (auto attr : opInst->getAttrs()) {
+  func.walk([&](Operation *op) {
+    for (auto attr : op->getAttrs()) {
       if (auto mapAttr = attr.second.dyn_cast<AffineMapAttr>())
-        simplifyAndUpdateAttribute(opInst, attr.first, mapAttr);
+        simplifyAndUpdateAttribute(op, attr.first, mapAttr);
       else if (auto setAttr = attr.second.dyn_cast<IntegerSetAttr>())
-        simplifyAndUpdateAttribute(opInst, attr.first, setAttr);
+        simplifyAndUpdateAttribute(op, attr.first, setAttr);
     }
   });
 
@@ -110,8 +111,17 @@ void SimplifyAffineStructures::runOnFunction() {
   for (auto allocOp : allocOps) {
     normalizeMemRef(allocOp);
   }
+
+  // Remove zero trip count loops.
+  // TODO: this could be moved to a more appropriate place.
+  func.walk([&](LoopLikeOpInterface loopOp) {
+    auto tripCount = loopOp.getConstantTripCount();
+    if (tripCount == 0UL)
+      loopOp.erase();
+  });
 }
 
 static PassRegistration<SimplifyAffineStructures>
-    pass("simplify-affine-structures",
-         "Simplify affine expressions in maps/sets and normalize memrefs");
+    pass("simplify-affine-structures", "Simplify expressions in afine map/set "
+                                       "attributes, normalize memrefs, remove "
+                                       "zero trip-count loops");
diff --git a/test/Transforms/affine-loop-invariant-code-motion.mlir b/test/Transforms/affine-loop-invariant-code-motion.mlir
@@ -17,27 +17,23 @@ func @nested_loops_both_having_invariant_code() {
   // CHECK-NEXT: %cst_0 = constant 8.000000e+00 : f32
   // CHECK-NEXT: %1 = addf %cst, %cst_0 : f32
   // CHECK-NEXT: affine.for %arg0 = 0 to 10 {
-  // CHECK-NEXT: affine.store %1, %0[%arg0] : memref<10xf32>
+  // CHECK-NEXT:   affine.store %1, %0[%arg0] : memref<10xf32>
 
   return
 }
 
-// The store-load forwarding can see through affine apply's since it relies on
-// dependence information.
-// CHECK-LABEL: func @store_affine_apply
-func @store_affine_apply() -> memref<10xf32> {
+// CHECK-LABEL: func @store_affine_for
+func @store_affine_for() -> memref<10xf32> {
   %cf7 = constant 7.0 : f32
   %m = alloc() : memref<10xf32>
   affine.for %arg0 = 0 to 10 {
-      %t0 = affine.apply (d1) -> (d1 + 1)(%arg0)
-      affine.store %cf7, %m[%t0] : memref<10xf32>
+    affine.store %cf7, %m[%arg0 + 1] : memref<10xf32>
   }
   return %m : memref<10xf32>
 // CHECK:       %cst = constant 7.000000e+00 : f32
 // CHECK-NEXT:  %0 = alloc() : memref<10xf32>
 // CHECK-NEXT:  affine.for %arg0 = 0 to 10 {
-// CHECK-NEXT:      %1 = affine.apply #map3(%arg0)
-// CHECK-NEXT:      affine.store %cst, %0[%1] : memref<10xf32>
+// CHECK-NEXT:      affine.store %cst, %0[%arg0 + 1] : memref<10xf32>
 // CHECK-NEXT:  }
 // CHECK-NEXT:  return %0 : memref<10xf32>
 }

diff --git a/test/Transforms/loop-invariant-code-motion.mlir b/test/Transforms/loop-invariant-code-motion.mlir
@@ -199,10 +199,54 @@ func @invariant_affine_nested_if_else() {
   // CHECK-NEXT: }
   // CHECK-NEXT: }
 
+  return
+}
+
+// CHECK-LABEL: func @zero_trip_count_affine
+func @zero_trip_count_affine() {
+  %m = alloc() : memref<10xf32>
+  %cf7 = constant 7.0 : f32
+  %N = constant 0 : index
+
+  affine.for %arg0 = 0 to %N {
+    affine.for %arg1 = 0 to 10 {
+      %v0 = addf %cf7, %cf7 : f32
+    }
+  }
+  // CHECK:      alloc() : memref<10xf32>
+  // CHECK-NEXT: %cst = constant 7.000000e+00 : f32
+  // CHECK-NEXT: %c0 = constant 0 : index
+  // CHECK-NEXT: affine.for
+  // CHECK-NEXT:   addf
+  // CHECK-NEXT:   affine.for
+
+  return
+}
+
+// CHECK-LABEL: func @zero_trip_count_loop
+func @zero_trip_count_loop(%N : index) {
+  %m = alloc() : memref<10xf32>
+  %cf7 = constant 7.0 : f32
+  %c1 = constant 1 : index
+  %c5 = constant 5 : index
+
+  loop.for %i = %N to %N step %c1 {
+    loop.for %j = %c5 to %c5 step %c1 {
+      addf %cf7, %cf7 : f32
+    }
+  }
+  // CHECK:      alloc() : memref<10xf32>
+  // CHECK-NEXT: %cst = constant 7.000000e+00 : f32
+  // CHECK-NEXT: %c1 = constant 1 : index
+  // CHECK-NEXT: %c5 = constant 5 : index
+  // CHECK-NEXT: loop.for
+  // CHECK-NEXT:   loop.for
+  // CHECK-NEXT:     addf
 
   return
 }
 
+// CHECK-LABEL: func @invariant_loop_dialect
 func @invariant_loop_dialect() {
   %ci0 = constant 0 : index
   %ci10 = constant 10 : index
@@ -211,7 +255,7 @@ func @invariant_loop_dialect() {
   %cf7 = constant 7.0 : f32
   %cf8 = constant 8.0 : f32
   loop.for %arg0 = %ci0 to %ci10 step %ci1 {
-    loop.for %arg1 = %ci0 to %ci10 step %ci1 {
+    loop.for %arg1 = %ci0 to %ci1 step %ci10 {
       %v0 = addf %cf7, %cf8 : f32
     }
   }
@@ -237,8 +281,8 @@ func @variant_loop_dialect() {
 
   // CHECK: %0 = alloc() : memref<10xf32>
   // CHECK-NEXT: loop.for
-  // CHECK-NEXT: loop.for
-  // CHECK-NEXT: addi
+  // CHECK-NEXT:   loop.for
+  // CHECK-NEXT:     addi
 
   return
 }
diff --git a/test/Transforms/simplify-affine-structures.mlir b/test/Transforms/simplify-affine-structures.mlir
@@ -235,3 +235,22 @@ func @test_empty_set(%N : index) {
 
   return
 }
+
+// CHECK-LABEL: func @zero_trip_count_loops
+func @zero_trip_count_loops(%N : index) {
+  %c0 = constant 0 : index
+  %c1 = constant 1 : index
+  %c-1 = constant -1 : index
+  %M = affine.apply (d0) -> ((2*d0 + 4) mod 2)(%N)
+  affine.for %i = 0 to %M {
+  }
+  affine.for %i = 0 to -1 {
+  }
+  loop.for %i = %M to %M step %c1 {
+  }
+  loop.for %i = %c0 to %c-1 step %N {
+  }
+  // All loops above should disappear.
+  // CHECK-NOT: loop.for
+  return
+}
diff --git a/tools/mlir-tblgen/OpInterfacesGen.cpp b/tools/mlir-tblgen/OpInterfacesGen.cpp
@@ -41,7 +41,10 @@ using mlir::tblgen::OpInterfaceMethod;
 // beginning of the argument list.
 static void emitMethodNameAndArgs(const OpInterfaceMethod &method,
                                   raw_ostream &os, bool addOperationArg) {
-  os << method.getName() << '(';
+  // Whenever an operation argument is added, suffix helper method name with an
+  // underscore to avoid conflicts with free functions of same name on the
+  // concrete ops using this interface.
+  os << method.getName() << (addOperationArg ? "_(" : "(");
   if (addOperationArg)
     os << "Operation *tablegen_opaque_op" << (method.arg_empty() ? "" : ", ");
   interleaveComma(method.getArguments(), os,
@@ -64,9 +67,11 @@ static void emitInterfaceDef(OpInterface &interface, raw_ostream &os) {
     emitMethodNameAndArgs(method, os, /*addOperationArg=*/false);
 
     // Forward to the method on the concrete operation type.
-    os << " {\n      return getImpl()->" << method.getName() << '(';
+    os << " {\n      return getImpl()->" << method.getName();
     if (!method.isStatic())
-      os << "getOperation()" << (method.arg_empty() ? "" : ", ");
+      os << "_(getOperation()" << (method.arg_empty() ? "" : ", ");
+    else
+      os << "(";
     interleaveComma(
         method.getArguments(), os,
         [&](const OpInterfaceMethod::Argument &arg) { os << arg.name; });