[CIR][Dialect] Make addrspace in pointer types to model LangAS

clang/include/clang/CIR/Dialect/Builder/CIRBaseBuilder.h

@@ -75,14 +75,18 @@ class CIRBaseBuilderTy : public mlir::OpBuilder {
     return mlir::cir::IntType::get(getContext(), N, true);
   }
 
-  mlir::cir::PointerType getPointerTo(mlir::Type ty,
-                                      unsigned addressSpace = 0) {
-    assert(!addressSpace && "address space is NYI");
-    return mlir::cir::PointerType::get(getContext(), ty);
+  mlir::cir::PointerType
+  getPointerTo(mlir::Type ty, clang::LangAS langAS = clang::LangAS::Default) {
+    mlir::cir::AddressSpaceAttr addrSpaceAttr;
+    if (langAS != clang::LangAS::Default)
+      addrSpaceAttr = mlir::cir::AddressSpaceAttr::get(getContext(), langAS);
+
+    return mlir::cir::PointerType::get(getContext(), ty, addrSpaceAttr);
   }
 
-  mlir::cir::PointerType getVoidPtrTy(unsigned addressSpace = 0) {
-    return getPointerTo(::mlir::cir::VoidType::get(getContext()), addressSpace);
+  mlir::cir::PointerType
+  getVoidPtrTy(clang::LangAS langAS = clang::LangAS::Default) {
+    return getPointerTo(::mlir::cir::VoidType::get(getContext()), langAS);
   }
 
   mlir::Value createLoad(mlir::Location loc, mlir::Value ptr,

clang/include/clang/CIR/Dialect/IR/CIRAttrs.td

@@ -617,6 +617,143 @@ def DynamicCastInfoAttr
   }];
 }
 
+//===----------------------------------------------------------------------===//
+// AddressSpaceAttr
+//===----------------------------------------------------------------------===//
+
+// TODO: other CIR AS cases
+def AS_Target : I32EnumAttrCase<"target", 21>;
+
+def AddressSpaceAttr : CIR_Attr<"AddressSpace", "addrspace"> {
+
+  let summary = "Address space attribute for pointer types";
+  let description = [{
+    The address space attribute models `clang::LangAS` rather than the LLVM
+    address space, which means it's not yet converted by the address space map
+    to carry target-specific semantics.
+
+    The representation is one-to-one except for `LangAS::Default`, which
+    corresponds to a null attribute instead.
+  }];
+
+  let parameters = (ins "int32_t":$value);
+
+  let assemblyFormat = [{
+    `<` $value `>`
+  }];
+
+  let builders = [
+    AttrBuilder<(ins "clang::LangAS":$langAS), [{
+      assert(langAS != clang::LangAS::Default &&
+        "Default address space is encoded as null attribute");
+      return $_get($_ctxt, getValueFromLangAS(langAS).value());
+    }]>
+  ];
+
+  let cppNamespace = "::mlir::cir";
+
+  // The following codes implement these conversions:
+  // clang::LangAS -> int32_t <-> text-form CIR
+
+  // CIR_PointerType manipulates the parse- and stringify- methods to provide
+  // simplified assembly format `custom<PointerAddrSpace>`.
+
+  list<I32EnumAttrCase> langASCases = [
+    // TODO: includes all non-target CIR AS cases here
+  ];
+
+  I32EnumAttrCase targetASCase = AS_Target;
+
+  let extraClassDeclaration = [{
+    static constexpr char kTargetKeyword[] = "}]#targetASCase.symbol#[{";
+    static constexpr int32_t kFirstTargetASValue = }]#targetASCase.value#[{;
+
+    bool isLang() const;
+    bool isTarget() const;
+    unsigned getTargetValue() const;
+
+    static std::optional<int32_t> parseValueFromString(llvm::StringRef s);
+    static std::optional<int32_t> getValueFromLangAS(clang::LangAS v);
+    static std::optional<llvm::StringRef> stringifyValue(int32_t v);
+  }];
+
+  let extraClassDefinition = [{
+    bool $cppClass::isLang() const {
+      return !isTarget();
+    }
+
+    bool $cppClass::isTarget() const {
+      return getValue() >= kFirstTargetASValue;
+    }
+
+    unsigned $cppClass::getTargetValue() const {
+      assert(isTarget() && "Not a target address space");
+      return getValue() - kFirstTargetASValue;
+    }
+
+    std::optional<int32_t>
+    $cppClass::parseValueFromString(llvm::StringRef str) {
+      return llvm::StringSwitch<::std::optional<int32_t>>(str)
+        }]
+        #
+        !interleave(
+          !foreach(case, langASCases,
+            ".Case(\""#case.symbol# "\", "#case.value # ")\n"
+          ),
+          "\n"
+        )
+        #
+        [{
+        // Target address spaces are not parsed here
+        .Default(std::nullopt);
+    }
+
+    std::optional<llvm::StringRef>
+    $cppClass::stringifyValue(int32_t value) {
+      switch (value) {
+      }]
+      #
+      !interleave(
+        !foreach(case, langASCases,
+          "case "#case.value
+          # ": return \""#case.symbol # "\";" ),
+        "\n"
+      )
+      #
+      [{
+      default:
+        // Target address spaces are not processed here
+        return std::nullopt;
+      }
+    }
+
+    std::optional<int32_t>
+    $cppClass::getValueFromLangAS(clang::LangAS langAS) {
+      assert((langAS == clang::LangAS::Default ||
+             clang::isTargetAddressSpace(langAS)) &&
+             "Language-specific address spaces are not supported");
+      switch (langAS) {
+      }]
+      #
+      !interleave(
+        !foreach(case, langASCases,
+          "case clang::LangAS::"#case.symbol
+          # ": return "#case.value # ";" ),
+        "\n"
+      )
+      #
+      [{
+      case clang::LangAS::Default:
+        // Default address space should be encoded as a null attribute.
+        return std::nullopt;
+      default:
+        // Target address space offset arithmetics
+        return clang::toTargetAddressSpace(langAS) + kFirstTargetASValue;
+      }
+    }
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // AST Wrappers
 //===----------------------------------------------------------------------===//

clang/include/clang/CIR/Dialect/IR/CIRTypes.h

@@ -20,6 +20,8 @@
 
 #include "clang/CIR/Interfaces/ASTAttrInterfaces.h"
 
+#include "clang/CIR/Dialect/IR/CIROpsEnums.h"
+
 //===----------------------------------------------------------------------===//
 // CIR StructType
 //

clang/include/clang/CIR/Dialect/IR/CIRTypes.td

@@ -208,24 +208,34 @@ def CIR_PointerType : CIR_Type<"Pointer", "ptr",
     `CIR.ptr` is a type returned by any op generating a pointer in C++.
   }];
 
-  let parameters = (ins "mlir::Type":$pointee,
-                        DefaultValuedParameter<"unsigned", "0">:$addrSpace);
+  let parameters = (ins
+    "mlir::Type":$pointee,
+    // FIXME(cir): Currently unable to directly use AddressSpaceAttr because of
+    // cyclic dep. Workaround with the top type and verifier.
+    OptionalParameter<"mlir::Attribute">:$addrSpace
+  );
 
   let builders = [
     TypeBuilderWithInferredContext<(ins
-      "mlir::Type":$pointee, CArg<"unsigned", "0">:$addrSpace), [{
-      return Base::get(pointee.getContext(), pointee, addrSpace);
+      "mlir::Type":$pointee,
+      CArg<"mlir::Attribute", "{}">:$addrSpace), [{
+      return $_get(pointee.getContext(), pointee, addrSpace);
     }]>,
     TypeBuilder<(ins
-      "mlir::Type":$pointee, CArg<"unsigned", "0">:$addrSpace), [{
-      return Base::get($_ctxt, pointee, addrSpace);
-    }]>,
+      "mlir::Type":$pointee,
+      CArg<"mlir::Attribute", "{}">:$addrSpace), [{
+      return $_get($_ctxt, pointee, addrSpace);
+    }]>
   ];
 
   let assemblyFormat = [{
-    `<` $pointee ( `,` `addrspace` `(` $addrSpace^ `)` )? `>`
+    `<` $pointee ( `,` `addrspace` `(`
+      custom<PointerAddrSpace>($addrSpace)^
+    `)` )? `>`
   }];
 
+  let genVerifyDecl = 1;
+
   let skipDefaultBuilders = 1;
 
   let extraClassDeclaration = [{

clang/include/clang/CIR/MissingFeatures.h

@@ -156,6 +156,7 @@ struct MissingFeatures {
   static bool constantFoldsToSimpleInteger() { return false; }
   static bool checkFunctionCallABI() { return false; }
   static bool zeroInitializer() { return false; }
+  static bool targetLoweringInfoAddressSpaceMap() { return false; }
   static bool targetCodeGenInfoIsProtoCallVariadic() { return false; }
   static bool targetCodeGenInfoGetNullPointer() { return false; }
   static bool operandBundles() { return false; }

clang/lib/CIR/CodeGen/CIRGenTypes.cpp

@@ -605,9 +605,7 @@ mlir::Type CIRGenTypes::ConvertType(QualType T) {
     const ReferenceType *RTy = cast<ReferenceType>(Ty);
     QualType ETy = RTy->getPointeeType();
     auto PointeeType = convertTypeForMem(ETy);
-    ResultType = ::mlir::cir::PointerType::get(
-        Builder.getContext(), PointeeType,
-        Context.getTargetAddressSpace(ETy.getAddressSpace()));
+    ResultType = Builder.getPointerTo(PointeeType, ETy.getAddressSpace());
     assert(ResultType && "Cannot get pointer type?");
     break;
   }
@@ -622,9 +620,7 @@ mlir::Type CIRGenTypes::ConvertType(QualType T) {
     // if (PointeeType->isVoidTy())
     //  PointeeType = Builder.getI8Type();
 
-    ResultType = ::mlir::cir::PointerType::get(
-        Builder.getContext(), PointeeType,
-        Context.getTargetAddressSpace(ETy.getAddressSpace()));
+    ResultType = Builder.getPointerTo(PointeeType, ETy.getAddressSpace());
     assert(ResultType && "Cannot get pointer type?");
     break;
   }

clang/lib/CIR/Dialect/IR/CIRTypes.cpp

@@ -45,6 +45,11 @@ parseFuncTypeArgs(mlir::AsmParser &p, llvm::SmallVector<mlir::Type> &params,
 static void printFuncTypeArgs(mlir::AsmPrinter &p,
                               mlir::ArrayRef<mlir::Type> params, bool isVarArg);
 
+static mlir::ParseResult parsePointerAddrSpace(mlir::AsmParser &p,
+                                               mlir::Attribute &addrSpaceAttr);
+static void printPointerAddrSpace(mlir::AsmPrinter &p,
+                                  mlir::Attribute addrSpaceAttr);
+
 //===----------------------------------------------------------------------===//
 // Get autogenerated stuff
 //===----------------------------------------------------------------------===//
@@ -873,6 +878,65 @@ llvm::ArrayRef<mlir::Type> FuncType::getReturnTypes() const {
 
 bool FuncType::isVoid() const { return mlir::isa<VoidType>(getReturnType()); }
 
+//===----------------------------------------------------------------------===//
+// PointerType Definitions
+//===----------------------------------------------------------------------===//
+
+mlir::LogicalResult
+PointerType::verify(llvm::function_ref<mlir::InFlightDiagnostic()> emitError,
+                    mlir::Type pointee, mlir::Attribute addrSpace) {
+  if (addrSpace && !mlir::isa<mlir::cir::AddressSpaceAttr>(addrSpace)) {
+    emitError() << "unexpected addrspace attribute type";
+    return mlir::failure();
+  }
+  return mlir::success();
+}
+
+mlir::ParseResult parsePointerAddrSpace(mlir::AsmParser &p,
+                                        mlir::Attribute &addrSpaceAttr) {
+  using mlir::cir::AddressSpaceAttr;
+  auto attrLoc = p.getCurrentLocation();
+
+  llvm::StringRef addrSpaceKind;
+  if (mlir::failed(p.parseOptionalKeyword(&addrSpaceKind))) {
+    p.emitError(attrLoc, "expected keyword for addrspace kind");
+    return mlir::failure();
+  }
+
+  if (addrSpaceKind == AddressSpaceAttr::kTargetKeyword) {
+    int64_t targetValue = -1;
+    if (p.parseLess() || p.parseInteger(targetValue) || p.parseGreater()) {
+      return mlir::failure();
+    }
+    addrSpaceAttr = AddressSpaceAttr::get(
+        p.getContext(), AddressSpaceAttr::kFirstTargetASValue + targetValue);
+  } else {
+    std::optional<int64_t> value =
+        AddressSpaceAttr::parseValueFromString(addrSpaceKind);
+    // not target AS, must be wrong keyword if no value
+    if (!value.has_value()) {
+      p.emitError(attrLoc, "invalid addrspace kind keyword: " + addrSpaceKind);
+      return mlir::failure();
+    }
+
+    addrSpaceAttr = AddressSpaceAttr::get(p.getContext(), *value);
+  }
+
+  return mlir::success();
+}
+
+void printPointerAddrSpace(mlir::AsmPrinter &p,
+                           mlir::Attribute rawAddrSpaceAttr) {
+  using mlir::cir::AddressSpaceAttr;
+  auto addrSpaceAttr = mlir::cast<AddressSpaceAttr>(rawAddrSpaceAttr);
+  if (addrSpaceAttr.isTarget()) {
+    p << AddressSpaceAttr::kTargetKeyword << "<"
+      << addrSpaceAttr.getTargetValue() << ">";
+  } else {
+    p << AddressSpaceAttr::stringifyValue(addrSpaceAttr.getValue());
+  }
+}
+
 //===----------------------------------------------------------------------===//
 // CIR Dialect
 //===----------------------------------------------------------------------===//

clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp

@@ -3335,8 +3335,20 @@ void prepareTypeConverter(mlir::LLVMTypeConverter &converter,
                           mlir::DataLayout &dataLayout) {
   converter.addConversion([&](mlir::cir::PointerType type) -> mlir::Type {
     // Drop pointee type since LLVM dialect only allows opaque pointers.
+
+    auto addrSpace =
+        mlir::cast_if_present<mlir::cir::AddressSpaceAttr>(type.getAddrSpace());
+    // null addrspace attribute indicates the default addrspace
+    if (!addrSpace)
+      return mlir::LLVM::LLVMPointerType::get(type.getContext());
+
+    // TODO(cir): Query the target-specific address space map to lower other ASs
+    // like `opencl_private`.
+    assert(!MissingFeatures::targetLoweringInfoAddressSpaceMap());
+    assert(addrSpace.isTarget() && "NYI");
+
     return mlir::LLVM::LLVMPointerType::get(type.getContext(),
-                                            type.getAddrSpace());
+                                            addrSpace.getTargetValue());
   });
   converter.addConversion([&](mlir::cir::DataMemberType type) -> mlir::Type {
     return mlir::IntegerType::get(type.getContext(),

clang/test/CIR/CodeGen/OpenCL/addrspace-alloca.cl

@@ -3,22 +3,25 @@
 // RUN: %clang_cc1 -cl-std=CL3.0 -O0 -fclangir -emit-llvm -triple spirv64-unknown-unknown %s -o %t.ll
 // RUN: FileCheck --input-file=%t.ll %s --check-prefix=LLVM
 
-// CIR: cir.func @func(%arg0: !cir.ptr<!s32i, addrspace(3)>
+// Lowering of language-specific AS not supported
+// XFAIL: *
+
+// CIR: cir.func @func(%arg0: !cir.ptr<!s32i, addrspace(target<3>)>
 // LLVM: @func(ptr addrspace(3)
 kernel void func(local int *p) {
-  // CIR-NEXT: %[[#ALLOCA_P:]] = cir.alloca !cir.ptr<!s32i, addrspace(3)>, !cir.ptr<!cir.ptr<!s32i, addrspace(3)>>, ["p", init] {alignment = 8 : i64}
+  // CIR-NEXT: %[[#ALLOCA_P:]] = cir.alloca !cir.ptr<!s32i, addrspace(target<3>)>, !cir.ptr<!cir.ptr<!s32i, addrspace(target<3>)>>, ["p", init] {alignment = 8 : i64}
   // LLVM-NEXT: %[[#ALLOCA_P:]] = alloca ptr addrspace(3), i64 1, align 8
 
   int x;
   // CIR-NEXT: %[[#ALLOCA_X:]] = cir.alloca !s32i, !cir.ptr<!s32i>, ["x"] {alignment = 4 : i64}
   // LLVM-NEXT: %[[#ALLOCA_X:]] = alloca i32, i64 1, align 4
 
   global char *b;
-  // CIR-NEXT: %[[#ALLOCA_B:]] = cir.alloca !cir.ptr<!s8i, addrspace(1)>, !cir.ptr<!cir.ptr<!s8i, addrspace(1)>>, ["b"] {alignment = 8 : i64}
+  // CIR-NEXT: %[[#ALLOCA_B:]] = cir.alloca !cir.ptr<!s8i, addrspace(target<1>)>, !cir.ptr<!cir.ptr<!s8i, addrspace(target<1>)>>, ["b"] {alignment = 8 : i64}
   // LLVM-NEXT: %[[#ALLOCA_B:]] = alloca ptr addrspace(1), i64 1, align 8
 
   // Store of the argument `p`
-  // CIR-NEXT: cir.store %arg0, %[[#ALLOCA_P]] : !cir.ptr<!s32i, addrspace(3)>, !cir.ptr<!cir.ptr<!s32i, addrspace(3)>>
+  // CIR-NEXT: cir.store %arg0, %[[#ALLOCA_P]] : !cir.ptr<!s32i, addrspace(target<3>)>, !cir.ptr<!cir.ptr<!s32i, addrspace(target<3>)>>
   // LLVM-NEXT: store ptr addrspace(3) %{{[0-9]+}}, ptr %[[#ALLOCA_P]], align 8
 
   return;

clang/test/CIR/CodeGen/OpenCL/spirv-target.cl

@@ -4,6 +4,9 @@
 // RUN: %clang_cc1 -cl-std=CL3.0 -fclangir -emit-llvm -triple spirv64-unknown-unknown %s -o %t_64.ll
 // RUN: FileCheck --input-file=%t_64.ll %s --check-prefix=LLVM-SPIRV64
 
+// Lowering of language-specific AS not supported
+// XFAIL: *
+
 // CIR-SPIRV64: cir.triple = "spirv64-unknown-unknown"
 // LLVM-SPIRV64: target triple = "spirv64-unknown-unknown"