[OpenACC][CIR] Implement 'gang' lowering for 'loop'

This clause requires an entire additional collection to keep track of
the gang 'kind' or 'type'. That work is maintained in the OpenACC
dialect functions. Otherwise, this is effectively the same as the
worker/vectors.

Author: erichkeane

clang/lib/CIR/CodeGen/CIRGenOpenACCClause.h

@@ -107,6 +107,18 @@ class OpenACCClauseCIREmitter final
         .CaseLower("radeon", mlir::acc::DeviceType::Radeon);
   }
 
+  mlir::acc::GangArgType decodeGangType(OpenACCGangKind GK) {
+    switch (GK) {
+    case OpenACCGangKind::Num:
+      return mlir::acc::GangArgType::Num;
+    case OpenACCGangKind::Dim:
+      return mlir::acc::GangArgType::Dim;
+    case OpenACCGangKind::Static:
+      return mlir::acc::GangArgType::Static;
+    }
+    llvm_unreachable("unknown gang kind");
+  }
+
 public:
   OpenACCClauseCIREmitter(OpTy &operation, CIRGen::CIRGenFunction &cgf,
                           CIRGen::CIRGenBuilderTy &builder,
@@ -424,6 +436,42 @@ class OpenACCClauseCIREmitter final
       return clauseNotImplemented(clause);
     }
   }
+
+  void VisitGangClause(const OpenACCGangClause &clause) {
+    if constexpr (isOneOfTypes<OpTy, mlir::acc::LoopOp>) {
+      if (clause.getNumExprs() == 0) {
+        operation.addEmptyGang(builder.getContext(), lastDeviceTypeValues);
+      } else {
+        llvm::SmallVector<mlir::Value> values;
+        llvm::SmallVector<mlir::acc::GangArgType> argTypes;
+        for (unsigned I = 0; I < clause.getNumExprs(); ++I) {
+          auto [kind, expr] = clause.getExpr(I);
+          mlir::Location exprLoc = cgf.cgm.getLoc(expr->getBeginLoc());
+          argTypes.push_back(decodeGangType(kind));
+          if (kind == OpenACCGangKind::Dim) {
+            llvm::APInt curValue =
+                expr->EvaluateKnownConstInt(cgf.cgm.getASTContext());
+            // The value is 1, 2, or 3, but the type isn't necessarily smaller
+            // than 64.
+            curValue = curValue.sextOrTrunc(64);
+            values.push_back(
+                createConstantInt(exprLoc, 64, curValue.getSExtValue()));
+          } else if (isa<OpenACCAsteriskSizeExpr>(expr)) {
+            values.push_back(createConstantInt(exprLoc, 64, -1));
+          } else {
+            values.push_back(createIntExpr(expr));
+          }
+        }
+
+        operation.addGangOperands(builder.getContext(), lastDeviceTypeValues,
+                                  argTypes, values);
+      }
+    } else {
+      // TODO: When we've implemented this for everything, switch this to an
+      // unreachable. Combined constructs remain.
+      return clauseNotImplemented(clause);
+    }
+  }
 };
 
 template <typename OpTy>

clang/test/CIR/CodeGenOpenACC/loop.cpp

@@ -323,4 +323,73 @@ extern "C" void acc_loop(int *A, int *B, int *C, int N) {
   // CHECK: acc.yield
   // CHECK-NEXT: } loc
   }
+
+#pragma acc parallel
+  // CHECK: acc.parallel {
+  {
+#pragma acc loop gang
+  for(unsigned I = 0; I < N; ++I);
+  // CHECK-NEXT: acc.loop gang {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+#pragma acc loop gang device_type(nvidia) gang
+  for(unsigned I = 0; I < N; ++I);
+  // CHECK-NEXT: acc.loop gang([#acc.device_type<none>, #acc.device_type<nvidia>]) {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+#pragma acc loop gang(dim:1) device_type(nvidia) gang(dim:2)
+  for(unsigned I = 0; I < N; ++I);
+  // CHECK-NEXT: %[[ONE_CONST:.*]] = arith.constant 1 : i64
+  // CHECK-NEXT: %[[TWO_CONST:.*]] = arith.constant 2 : i64
+  // CHECK-NEXT: acc.loop gang({dim=%[[ONE_CONST]] : i64}, {dim=%[[TWO_CONST]] : i64} [#acc.device_type<nvidia>]) {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+#pragma acc loop gang(static:N, dim: 1) device_type(nvidia, radeon) gang(static:*, dim : 2)
+  for(unsigned I = 0; I < N; ++I);
+  // CHECK-NEXT: %[[N_LOAD:.*]] = cir.load %[[ALLOCA_N]] : !cir.ptr<!s32i>, !s32i
+  // CHECK-NEXT: %[[N_CONV:.*]] = builtin.unrealized_conversion_cast %[[N_LOAD]] : !s32i to si32
+  // CHECK-NEXT: %[[ONE_CONST:.*]] = arith.constant 1 : i64
+  // CHECK-NEXT: %[[STAR_CONST:.*]] = arith.constant -1 : i64
+  // CHECK-NEXT: %[[TWO_CONST:.*]] = arith.constant 2 : i64
+  // CHECK-NEXT: acc.loop gang({static=%[[N_CONV]] : si32, dim=%[[ONE_CONST]] : i64}, {static=%[[STAR_CONST]] : i64, dim=%[[TWO_CONST]] : i64} [#acc.device_type<nvidia>], {static=%[[STAR_CONST]] : i64, dim=%[[TWO_CONST]] : i64} [#acc.device_type<radeon>]) {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+  }
+#pragma acc kernels
+  // CHECK: acc.kernels {
+  {
+#pragma acc loop gang(num:N) device_type(nvidia, radeon) gang(num:N)
+  for(unsigned I = 0; I < N; ++I);
+  // CHECK-NEXT: %[[N_LOAD:.*]] = cir.load %[[ALLOCA_N]] : !cir.ptr<!s32i>, !s32i
+  // CHECK-NEXT: %[[N_CONV:.*]] = builtin.unrealized_conversion_cast %[[N_LOAD]] : !s32i to si32
+  // CHECK-NEXT: %[[N_LOAD2:.*]] = cir.load %[[ALLOCA_N]] : !cir.ptr<!s32i>, !s32i
+  // CHECK-NEXT: %[[N_CONV2:.*]] = builtin.unrealized_conversion_cast %[[N_LOAD2]] : !s32i to si32
+  // CHECK-NEXT: acc.loop gang({num=%[[N_CONV]] : si32}, {num=%[[N_CONV2]] : si32} [#acc.device_type<nvidia>], {num=%[[N_CONV2]] : si32} [#acc.device_type<radeon>]) {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+#pragma acc loop gang(static:N) device_type(nvidia) gang(static:*)
+  for(unsigned I = 0; I < N; ++I);
+  // CHECK-NEXT: %[[N_LOAD:.*]] = cir.load %[[ALLOCA_N]] : !cir.ptr<!s32i>, !s32i
+  // CHECK-NEXT: %[[N_CONV:.*]] = builtin.unrealized_conversion_cast %[[N_LOAD]] : !s32i to si32
+  // CHECK-NEXT: %[[STAR_CONST:.*]] = arith.constant -1 : i64
+  // CHECK-NEXT: acc.loop gang({static=%[[N_CONV]] : si32}, {static=%[[STAR_CONST]] : i64} [#acc.device_type<nvidia>]) {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+#pragma acc loop gang(static:N, num: N + 1) device_type(nvidia) gang(static:*, num : N + 2)
+  for(unsigned I = 0; I < N; ++I);
+  // CHECK-NEXT: %[[N_LOAD:.*]] = cir.load %[[ALLOCA_N]] : !cir.ptr<!s32i>, !s32i
+  // CHECK-NEXT: %[[N_CONV:.*]] = builtin.unrealized_conversion_cast %[[N_LOAD]] : !s32i to si32
+  // CHECK-NEXT: %[[N_LOAD2:.*]] = cir.load %[[ALLOCA_N]] : !cir.ptr<!s32i>, !s32i
+  // CHECK-NEXT: %[[CIR_ONE_CONST:.*]] = cir.const #cir.int<1> : !s32i
+  // CHECK-NEXT: %[[N_PLUS_ONE:.*]] = cir.binop(add, %[[N_LOAD2]], %[[CIR_ONE_CONST]]) nsw : !s32i
+  // CHECK-NEXT: %[[N_PLUS_ONE_CONV:.*]] = builtin.unrealized_conversion_cast %[[N_PLUS_ONE]] : !s32i to si32
+  // CHECK-NEXT: %[[STAR_CONST:.*]] = arith.constant -1 : i64
+  // CHECK-NEXT: %[[N_LOAD3:.*]] = cir.load %[[ALLOCA_N]] : !cir.ptr<!s32i>, !s32i
+  // CHECK-NEXT: %[[CIR_TWO_CONST:.*]] = cir.const #cir.int<2> : !s32i
+  // CHECK-NEXT: %[[N_PLUS_TWO:.*]] = cir.binop(add, %[[N_LOAD3]], %[[CIR_TWO_CONST]]) nsw : !s32i
+  // CHECK-NEXT: %[[N_PLUS_TWO_CONV:.*]] = builtin.unrealized_conversion_cast %[[N_PLUS_TWO]] : !s32i to si32
+  // CHECK-NEXT: acc.loop gang({static=%[[N_CONV]] : si32, num=%[[N_PLUS_ONE_CONV]] : si32}, {static=%[[STAR_CONST]] : i64, num=%[[N_PLUS_TWO_CONV]] : si32} [#acc.device_type<nvidia>]) {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+  }
 }

mlir/include/mlir/Dialect/OpenACC/OpenACCOps.td

@@ -2231,6 +2231,16 @@ def OpenACC_LoopOp : OpenACC_Op<"loop",
     // device_types. This is for the case where there is no expression specified
     // in a 'worker'.
     void addEmptyWorker(MLIRContext *, llvm::ArrayRef<DeviceType>);
+
+    // Adds a collection of operands for a 'gang' clause that has various types
+    // corresponding to each operand.
+    void addGangOperands(MLIRContext *, llvm::ArrayRef<DeviceType>,
+                         llvm::ArrayRef<GangArgType>, mlir::ValueRange);
+
+    // Add an empty value to the 'gang' list with a current list of
+    // device_types. This is for the case where there is no expression specified
+    // in a 'gang'.
+    void addEmptyGang(MLIRContext *, llvm::ArrayRef<DeviceType>);
   }];
 
   let hasCustomAssemblyFormat = 1;

mlir/lib/Dialect/OpenACC/IR/OpenACC.cpp

@@ -2748,6 +2748,49 @@ void acc::LoopOp::addEmptyWorker(
                                                    effectiveDeviceTypes));
 }
 
+void acc::LoopOp::addEmptyGang(
+    MLIRContext *context, llvm::ArrayRef<DeviceType> effectiveDeviceTypes) {
+  setGangAttr(addDeviceTypeAffectedOperandHelper(context, getGangAttr(),
+                                                 effectiveDeviceTypes));
+}
+
+void acc::LoopOp::addGangOperands(
+    MLIRContext *context, llvm::ArrayRef<DeviceType> effectiveDeviceTypes,
+    llvm::ArrayRef<GangArgType> argTypes, mlir::ValueRange values) {
+  llvm::SmallVector<int32_t> segments;
+  if (getGangOperandsSegments())
+    llvm::copy(*getGangOperandsSegments(), std::back_inserter(segments));
+
+  unsigned beforeCount = segments.size();
+
+  setGangOperandsDeviceTypeAttr(addDeviceTypeAffectedOperandHelper(
+      context, getGangOperandsDeviceTypeAttr(), effectiveDeviceTypes, values,
+      getGangOperandsMutable(), segments));
+
+  setGangOperandsSegments(segments);
+
+  // This is a bit of extra work to make sure we update the 'types' correctly by
+  // adding to the types collection the correct number of times. We could
+  // potentially add something similar to the
+  // addDeviceTypeAffectedOperandHelper, but it seems that would be pretty
+  // excessive for a one-off case.
+  unsigned numAdded = segments.size() - beforeCount;
+
+  if (numAdded > 0) {
+    llvm::SmallVector<mlir::Attribute> gangTypes;
+    if (getGangOperandsArgTypeAttr())
+      llvm::copy(getGangOperandsArgTypeAttr(), std::back_inserter(gangTypes));
+
+    for (unsigned I = 0; I < numAdded; ++I)
+      llvm::transform(argTypes, std::back_inserter(gangTypes),
+                      [=](mlir::acc::GangArgType gangTy) {
+                        return mlir::acc::GangArgTypeAttr::get(context, gangTy);
+                      });
+
+    setGangOperandsArgTypeAttr(mlir::ArrayAttr::get(context, gangTypes));
+  }
+}
+
 //===----------------------------------------------------------------------===//
 // DataOp
 //===----------------------------------------------------------------------===//