[mlir][omp] Add omp.tile operation (llvm#160292)

Add the `omp.tile` loop transformations for the OpenMP dialect. Used for
lowering a standalone `!$omp tile` in Flang.

Author: Meinersbur

mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td

@@ -995,6 +995,35 @@ class OpenMP_NumTeamsClauseSkip<
 
 def OpenMP_NumTeamsClause : OpenMP_NumTeamsClauseSkip<>;
 
+//===----------------------------------------------------------------------===//
+//  V5.1: [10.1.2] `sizes` clause
+//===----------------------------------------------------------------------===//
+
+class OpenMP_SizesClauseSkip<
+    bit traits = false, bit arguments = false, bit assemblyFormat = false,
+    bit description = false, bit extraClassDeclaration = false
+  > : OpenMP_Clause<traits, arguments, assemblyFormat, description,
+                    extraClassDeclaration> {
+  let arguments = (ins
+    Variadic<IntLikeType>:$sizes
+  );
+
+  let optAssemblyFormat = [{
+    `sizes` `(` $sizes `:` type($sizes) `)`
+  }];
+
+  let description = [{
+    The `sizes` clauses defines the size of a grid over a multi-dimensional
+    logical iteration space. This grid is used for loop transformations such as
+    `tile` and `strip`. The size per dimension can be a variable, but only
+    values that are not at least 2 make sense. It is not specified what happens
+    when smaller values are used, but should still result in a loop nest that
+    executes each logical iteration once.
+  }];
+}
+
+def OpenMP_SizesClause : OpenMP_SizesClauseSkip<>;
+
 //===----------------------------------------------------------------------===//
 // V5.2: [10.1.2] `num_threads` clause
 //===----------------------------------------------------------------------===//

mlir/include/mlir/Dialect/OpenMP/OpenMPOpBase.td

@@ -38,6 +38,44 @@ def OpenMP_MapBoundsType : OpenMP_Type<"MapBounds", "map_bounds_ty"> {
   let summary = "Type for representing omp map clause bounds information";
 }
 
+//===---------------------------------------------------------------------===//
+// OpenMP Canonical Loop Info Type
+//===---------------------------------------------------------------------===//
+
+def CanonicalLoopInfoType : OpenMP_Type<"CanonicalLoopInfo", "cli"> {
+  let summary = "Type for representing a reference to a canonical loop";
+  let description = [{
+    A variable of type CanonicalLoopInfo refers to an OpenMP-compatible
+    canonical loop in the same function. Values of this type are not
+    available at runtime and therefore cannot be used by the program itself,
+    i.e. an opaque type. It is similar to the transform dialect's
+    `!transform.interface` type, but instead of implementing an interface
+    for each transformation, the OpenMP dialect itself defines possible
+    operations on this type.
+
+    A value of type CanonicalLoopInfoType (in the following: CLI) value can be
+
+    1. created by omp.new_cli.
+    2. passed to omp.canonical_loop to associate the loop to that CLI. A CLI
+       can only be associated once.
+    3. passed to an omp loop transformation operation that modifies the loop
+       associated with the CLI. The CLI is the "applyee" and the operation is
+       the consumer. A CLI can only be consumed once.
+    4. passed to an omp loop transformation operation to associate the cli with
+       a result of that transformation. The CLI is the "generatee" and the
+       operation is the generator.
+
+    A CLI cannot
+
+    1. be returned from a function.
+    2. be passed to operations that are not specifically designed to take a
+       CanonicalLoopInfoType, including AnyType.
+
+    A CLI directly corresponds to an object of
+    OpenMPIRBuilder's CanonicalLoopInfo struct when lowering to LLVM-IR.
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // Base classes for OpenMP dialect operations.
 //===----------------------------------------------------------------------===//
@@ -211,8 +249,35 @@ class OpenMP_Op<string mnemonic, list<Trait> traits = [],
 // Doesn't actually create a C++ base class (only defines default values for
 // tablegen classes that derive from this). Use LoopTransformationInterface
 // instead for common operations.
-class OpenMPTransform_Op<string mnemonic, list<Trait> traits = []> :
-      OpenMP_Op<mnemonic, !listconcat([DeclareOpInterfaceMethods<LoopTransformationInterface>], traits)  > {
+class OpenMPTransform_Op<string mnemonic,
+                         list<Trait> traits = [],
+                         list<OpenMP_Clause> clauses = []> :
+      OpenMP_Op<mnemonic,
+                traits = !listconcat([DeclareOpInterfaceMethods<LoopTransformationInterface>], traits),
+                clauses = clauses> {
+}
+
+// Base clause for loop transformations using the standard syntax.
+//
+//     omp.opname ($generatees) <- ($applyees) clause(...) clause(...) ... <attr-dicr>
+//     omp.opname                  ($applyees) clause(...) clause(...) ... <attr-dict>
+//
+// $generatees is optional and is assumed to be empty if omitted
+class OpenMPTransformBase_Op<string mnemonic,
+                         list<Trait> traits = [],
+                         list<OpenMP_Clause> clauses = []> :
+      OpenMPTransform_Op<mnemonic,
+                         traits = !listconcat(traits, [AttrSizedOperandSegments]),
+                         clauses = clauses> {
+
+  let arguments = !con(
+                       (ins Variadic<CanonicalLoopInfoType>:$generatees,
+                            Variadic<CanonicalLoopInfoType>:$applyees
+                      ), clausesArgs);
+
+  let assemblyFormat = [{ custom<LoopTransformClis>($generatees, $applyees) }]
+                         # clausesAssemblyFormat
+                         # [{ attr-dict }];
 }
 
 #endif  // OPENMP_OP_BASE

mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td

@@ -357,44 +357,6 @@ def SingleOp : OpenMP_Op<"single", traits = [
   let hasVerifier = 1;
 }
 
-//===---------------------------------------------------------------------===//
-// OpenMP Canonical Loop Info Type
-//===---------------------------------------------------------------------===//
-
-def CanonicalLoopInfoType : OpenMP_Type<"CanonicalLoopInfo", "cli"> {
-  let summary = "Type for representing a reference to a canonical loop";
-  let description = [{
-    A variable of type CanonicalLoopInfo refers to an OpenMP-compatible
-    canonical loop in the same function. Values of this type are not
-    available at runtime and therefore cannot be used by the program itself,
-    i.e. an opaque type. It is similar to the transform dialect's
-    `!transform.interface` type, but instead of implementing an interface
-    for each transformation, the OpenMP dialect itself defines possible
-    operations on this type.
-
-    A value of type CanonicalLoopInfoType (in the following: CLI) value can be
-
-    1. created by omp.new_cli.
-    2. passed to omp.canonical_loop to associate the loop to that CLI. A CLI
-       can only be associated once.
-    3. passed to an omp loop transformation operation that modifies the loop
-       associated with the CLI. The CLI is the "applyee" and the operation is
-       the consumer. A CLI can only be consumed once.
-    4. passed to an omp loop transformation operation to associate the cli with
-       a result of that transformation. The CLI is the "generatee" and the
-       operation is the generator.
-
-    A CLI cannot
-
-    1. be returned from a function.
-    2. be passed to operations that are not specifically designed to take a
-       CanonicalLoopInfoType, including AnyType.
-
-    A CLI directly corresponds to an object of
-    OpenMPIRBuilder's CanonicalLoopInfo struct when lowering to LLVM-IR.
-  }];
-}
-
 //===---------------------------------------------------------------------===//
 // OpenMP Canonical Loop Info Creation
 //===---------------------------------------------------------------------===//
@@ -563,6 +525,31 @@ def UnrollHeuristicOp : OpenMPTransform_Op<"unroll_heuristic", []> {
   let hasCustomAssemblyFormat = 1;
 }
 
+//===----------------------------------------------------------------------===//
+// OpenMP tile operation
+//===----------------------------------------------------------------------===//
+
+def TileOp : OpenMPTransformBase_Op<"tile",
+                                clauses = [OpenMP_SizesClause]> {
+  let summary = "OpenMP tile operation";
+  let description = [{
+    Represents the OpenMP tile directive introduced in OpenMP 5.1.
+
+    The construct partitions the logical iteration space of the affected loops
+    into equally-sized tiles, then creates two sets of nested loops. The outer
+    loops, called the grid loops, iterate over all tiles. The inner loops,
+    called the intratile loops, iterate over the logical iterations of a tile.
+    The sizes clause determines the size of a tile.
+
+    Currently, the affected loops must be rectangular (the tripcount of the
+    inner loop must not depend on any iv of an surrounding affected loop) and
+    perfectly nested (except for the innermost affected loop, no operations
+    other than the nested loop and the terminator in the loop body).
+  }] # clausesDescription;
+
+  let hasVerifier = 1;
+}
+
 //===----------------------------------------------------------------------===//
 // 2.8.3 Workshare Construct
 //===----------------------------------------------------------------------===//

mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp

@@ -33,6 +33,7 @@
 #include "llvm/ADT/TypeSwitch.h"
 #include "llvm/ADT/bit.h"
 #include "llvm/Frontend/OpenMP/OMPConstants.h"
+#include "llvm/Support/InterleavedRange.h"
 #include <cstddef>
 #include <iterator>
 #include <optional>
@@ -3385,6 +3386,9 @@ void NewCliOp::getAsmResultNames(OpAsmSetValueNameFn setNameFn) {
   Value result = getResult();
   auto [newCli, gen, cons] = decodeCli(result);
 
+  // Structured binding `gen` cannot be captured in lambdas before C++20
+  OpOperand *generator = gen;
+
   // Derive the CLI variable name from its generator:
   //  * "canonloop" for omp.canonical_loop
   //  * custom name for loop transformation generatees
@@ -3403,6 +3407,24 @@ void NewCliOp::getAsmResultNames(OpAsmSetValueNameFn setNameFn) {
             .Case([&](UnrollHeuristicOp op) -> std::string {
               llvm_unreachable("heuristic unrolling does not generate a loop");
             })
+            .Case([&](TileOp op) -> std::string {
+              auto [generateesFirst, generateesCount] =
+                  op.getGenerateesODSOperandIndexAndLength();
+              unsigned firstGrid = generateesFirst;
+              unsigned firstIntratile = generateesFirst + generateesCount / 2;
+              unsigned end = generateesFirst + generateesCount;
+              unsigned opnum = generator->getOperandNumber();
+              // In the OpenMP apply and looprange clauses, indices are 1-based
+              if (firstGrid <= opnum && opnum < firstIntratile) {
+                unsigned gridnum = opnum - firstGrid + 1;
+                return ("grid" + Twine(gridnum)).str();
+              }
+              if (firstIntratile <= opnum && opnum < end) {
+                unsigned intratilenum = opnum - firstIntratile + 1;
+                return ("intratile" + Twine(intratilenum)).str();
+              }
+              llvm_unreachable("Unexpected generatee argument");
+            })
             .Default([&](Operation *op) {
               assert(false && "TODO: Custom name for this operation");
               return "transformed";
@@ -3631,6 +3653,138 @@ UnrollHeuristicOp::getGenerateesODSOperandIndexAndLength() {
   return {0, 0};
 }
 
+//===----------------------------------------------------------------------===//
+// TileOp
+//===----------------------------------------------------------------------===//
+
+static void printLoopTransformClis(OpAsmPrinter &p, TileOp op,
+                                   OperandRange generatees,
+                                   OperandRange applyees) {
+  if (!generatees.empty())
+    p << '(' << llvm::interleaved(generatees) << ')';
+
+  if (!applyees.empty())
+    p << " <- (" << llvm::interleaved(applyees) << ')';
+}
+
+static ParseResult parseLoopTransformClis(
+    OpAsmParser &parser,
+    SmallVectorImpl<OpAsmParser::UnresolvedOperand> &generateesOperands,
+    SmallVectorImpl<OpAsmParser::UnresolvedOperand> &applyeesOperands) {
+  if (parser.parseOptionalLess()) {
+    // Syntax 1: generatees present
+
+    if (parser.parseOperandList(generateesOperands,
+                                mlir::OpAsmParser::Delimiter::Paren))
+      return failure();
+
+    if (parser.parseLess())
+      return failure();
+  } else {
+    // Syntax 2: generatees omitted
+  }
+
+  // Parse `<-` (`<` has already been parsed)
+  if (parser.parseMinus())
+    return failure();
+
+  if (parser.parseOperandList(applyeesOperands,
+                              mlir::OpAsmParser::Delimiter::Paren))
+    return failure();
+
+  return success();
+}
+
+LogicalResult TileOp::verify() {
+  if (getApplyees().empty())
+    return emitOpError() << "must apply to at least one loop";
+
+  if (getSizes().size() != getApplyees().size())
+    return emitOpError() << "there must be one tile size for each applyee";
+
+  if (!getGeneratees().empty() &&
+      2 * getSizes().size() != getGeneratees().size())
+    return emitOpError()
+           << "expecting two times the number of generatees than applyees";
+
+  DenseSet<Value> parentIVs;
+
+  Value parent = getApplyees().front();
+  for (auto &&applyee : llvm::drop_begin(getApplyees())) {
+    auto [parentCreate, parentGen, parentCons] = decodeCli(parent);
+    auto [create, gen, cons] = decodeCli(applyee);
+
+    if (!parentGen)
+      return emitOpError() << "applyee CLI has no generator";
+
+    auto parentLoop = dyn_cast_or_null<CanonicalLoopOp>(parentGen->getOwner());
+    if (!parentGen)
+      return emitOpError()
+             << "currently only supports omp.canonical_loop as applyee";
+
+    parentIVs.insert(parentLoop.getInductionVar());
+
+    if (!gen)
+      return emitOpError() << "applyee CLI has no generator";
+    auto loop = dyn_cast_or_null<CanonicalLoopOp>(gen->getOwner());
+    if (!loop)
+      return emitOpError()
+             << "currently only supports omp.canonical_loop as applyee";
+
+    // Canonical loop must be perfectly nested, i.e. the body of the parent must
+    // only contain the omp.canonical_loop of the nested loops, and
+    // omp.terminator
+    bool isPerfectlyNested = [&]() {
+      auto &parentBody = parentLoop.getRegion();
+      if (!parentBody.hasOneBlock())
+        return false;
+      auto &parentBlock = parentBody.getBlocks().front();
+
+      auto nestedLoopIt = parentBlock.begin();
+      if (nestedLoopIt == parentBlock.end() ||
+          (&*nestedLoopIt != loop.getOperation()))
+        return false;
+
+      auto termIt = std::next(nestedLoopIt);
+      if (termIt == parentBlock.end() || !isa<TerminatorOp>(termIt))
+        return false;
+
+      if (std::next(termIt) != parentBlock.end())
+        return false;
+
+      return true;
+    }();
+    if (!isPerfectlyNested)
+      return emitOpError() << "tiled loop nest must be perfectly nested";
+
+    if (parentIVs.contains(loop.getTripCount()))
+      return emitOpError() << "tiled loop nest must be rectangular";
+
+    parent = applyee;
+  }
+
+  // TODO: The tile sizes must be computed before the loop, but checking this
+  // requires dominance analysis. For instance:
+  //
+  //      %canonloop = omp.new_cli
+  //      omp.canonical_loop(%canonloop) %iv : i32 in range(%tc) {
+  //        // write to %x
+  //        omp.terminator
+  //      }
+  //      %ts = llvm.load %x
+  //      omp.tile <- (%canonloop) sizes(%ts : i32)
+
+  return success();
+}
+
+std::pair<unsigned, unsigned> TileOp ::getApplyeesODSOperandIndexAndLength() {
+  return getODSOperandIndexAndLength(odsIndex_applyees);
+}
+
+std::pair<unsigned, unsigned> TileOp::getGenerateesODSOperandIndexAndLength() {
+  return getODSOperandIndexAndLength(odsIndex_generatees);
+}
+
 //===----------------------------------------------------------------------===//
 // Critical construct (2.17.1)
 //===----------------------------------------------------------------------===//