[CIR][CodeGen] Basic skeleton of SPIRV64 target support (#671)

* SPIRV64 TargetInfo
* Calling convention `SPIR_KERNEL`
* Minimal ABI with Direct/Extend for arguments and Ignore for return

---------

Co-authored-by: Vinicius Couto Espindola <[email protected]>

Author: 2 people

clang/lib/CIR/CodeGen/CIRGenDecl.cpp

@@ -36,10 +36,10 @@ CIRGenFunction::AutoVarEmission
 CIRGenFunction::buildAutoVarAlloca(const VarDecl &D,
                                    mlir::OpBuilder::InsertPoint ip) {
   QualType Ty = D.getType();
-  // TODO: (|| Ty.getAddressSpace() == LangAS::opencl_private &&
-  //        getLangOpts().OpenCL))
   assert(!MissingFeatures::openCL());
-  assert(Ty.getAddressSpace() == LangAS::Default);
+  assert(
+      Ty.getAddressSpace() == LangAS::Default ||
+      (Ty.getAddressSpace() == LangAS::opencl_private && getLangOpts().OpenCL));
   assert(!D.hasAttr<AnnotateAttr>() && "not implemented");
 
   auto loc = getLoc(D.getSourceRange());

clang/lib/CIR/CodeGen/CIRGenFunction.cpp

@@ -993,7 +993,8 @@ void CIRGenFunction::StartFunction(GlobalDecl GD, QualType RetTy,
     llvm_unreachable("NYI");
 
   if (FD && getLangOpts().OpenCL) {
-    llvm_unreachable("NYI");
+    // TODO(cir): Emit OpenCL kernel metadata
+    assert(!MissingFeatures::openCL());
   }
 
   // If we are checking function types, emit a function type signature as

clang/lib/CIR/CodeGen/CIRGenTypes.cpp

@@ -25,8 +25,14 @@ using namespace clang;
 using namespace cir;
 
 unsigned CIRGenTypes::ClangCallConvToCIRCallConv(clang::CallingConv CC) {
-  assert(CC == CC_C && "No other calling conventions implemented.");
-  return cir::CallingConv::C;
+  switch (CC) {
+  case CC_C:
+    return cir::CallingConv::C;
+  case CC_OpenCLKernel:
+    return CGM.getTargetCIRGenInfo().getOpenCLKernelCallingConv();
+  default:
+    llvm_unreachable("No other calling conventions implemented.");
+  }
 }
 
 CIRGenTypes::CIRGenTypes(CIRGenModule &cgm)

clang/lib/CIR/CodeGen/CallingConv.h

@@ -34,6 +34,12 @@ enum {
   /// with typical C calling conventions, the callee/caller have to tolerate
   /// certain amounts of prototype mismatch.
   C = 0,
+
+  /// Used for SPIR kernel functions. Inherits the restrictions of SPIR_FUNC,
+  /// except it cannot have non-void return values, it cannot have variable
+  /// arguments, it can also be called by the host or it is externally
+  /// visible.
+  SPIR_KERNEL = 76,
 };
 
 } // namespace CallingConv

clang/lib/CIR/CodeGen/TargetInfo.cpp

@@ -20,6 +20,68 @@ static bool testIfIsVoidTy(QualType Ty) {
   return k == BuiltinType::Void;
 }
 
+static bool isAggregateTypeForABI(QualType T) {
+  return !CIRGenFunction::hasScalarEvaluationKind(T) ||
+         T->isMemberFunctionPointerType();
+}
+
+/// Pass transparent unions as if they were the type of the first element. Sema
+/// should ensure that all elements of the union have the same "machine type".
+static QualType useFirstFieldIfTransparentUnion(QualType Ty) {
+  assert(!Ty->getAsUnionType() && "NYI");
+  return Ty;
+}
+
+namespace {
+
+/// The default implementation for ABI specific
+/// details. This implementation provides information which results in
+/// self-consistent and sensible LLVM IR generation, but does not
+/// conform to any particular ABI.
+class DefaultABIInfo : public ABIInfo {
+public:
+  DefaultABIInfo(CIRGenTypes &CGT) : ABIInfo(CGT) {}
+
+  virtual ~DefaultABIInfo() = default;
+
+  ABIArgInfo classifyReturnType(QualType RetTy) const {
+    if (RetTy->isVoidType())
+      return ABIArgInfo::getIgnore();
+
+    llvm_unreachable("Non-void return type NYI");
+  }
+
+  ABIArgInfo classifyArgumentType(QualType Ty) const {
+    Ty = useFirstFieldIfTransparentUnion(Ty);
+
+    if (isAggregateTypeForABI(Ty)) {
+      llvm_unreachable("NYI");
+    }
+
+    // Treat an enum type as its underlying type.
+    if (const EnumType *EnumTy = Ty->getAs<EnumType>())
+      llvm_unreachable("NYI");
+
+    ASTContext &Context = getContext();
+    if (const auto *EIT = Ty->getAs<BitIntType>())
+      llvm_unreachable("NYI");
+
+    if (isPromotableIntegerTypeForABI(Ty)) {
+      llvm_unreachable("ArgInfo integer extend NYI");
+    } else {
+      return ABIArgInfo::getDirect();
+    }
+  }
+
+  void computeInfo(CIRGenFunctionInfo &FI) const override {
+    if (!getCXXABI().classifyReturnType(FI))
+      FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
+    for (auto &I : FI.arguments())
+      I.info = classifyArgumentType(I.type);
+  }
+};
+} // namespace
+
 //===----------------------------------------------------------------------===//
 // AArch64 ABI Implementation
 //===----------------------------------------------------------------------===//
@@ -151,6 +213,66 @@ class X86_64TargetCIRGenInfo : public TargetCIRGenInfo {
 };
 } // namespace
 
+//===----------------------------------------------------------------------===//
+// Base ABI and target codegen info implementation common between SPIR and
+// SPIR-V.
+//===----------------------------------------------------------------------===//
+
+namespace {
+class CommonSPIRABIInfo : public DefaultABIInfo {
+public:
+  CommonSPIRABIInfo(CIRGenTypes &CGT) : DefaultABIInfo(CGT) {}
+};
+
+class SPIRVABIInfo : public CommonSPIRABIInfo {
+public:
+  SPIRVABIInfo(CIRGenTypes &CGT) : CommonSPIRABIInfo(CGT) {}
+  void computeInfo(CIRGenFunctionInfo &FI) const override {
+    // The logic is same as in DefaultABIInfo with an exception on the kernel
+    // arguments handling.
+    llvm::CallingConv::ID CC = FI.getCallingConvention();
+
+    bool cxxabiHit = getCXXABI().classifyReturnType(FI);
+    assert(!cxxabiHit && "C++ ABI not considered");
+
+    FI.getReturnInfo() = classifyReturnType(FI.getReturnType());
+
+    for (auto &I : FI.arguments()) {
+      if (CC == llvm::CallingConv::SPIR_KERNEL) {
+        I.info = classifyKernelArgumentType(I.type);
+      } else {
+        I.info = classifyArgumentType(I.type);
+      }
+    }
+  }
+
+private:
+  ABIArgInfo classifyKernelArgumentType(QualType Ty) const {
+    assert(!getContext().getLangOpts().CUDAIsDevice && "NYI");
+    return classifyArgumentType(Ty);
+  }
+};
+} // namespace
+namespace {
+
+class CommonSPIRTargetCIRGenInfo : public TargetCIRGenInfo {
+public:
+  CommonSPIRTargetCIRGenInfo(std::unique_ptr<ABIInfo> ABIInfo)
+      : TargetCIRGenInfo(std::move(ABIInfo)) {}
+
+  unsigned getOpenCLKernelCallingConv() const override {
+    return llvm::CallingConv::SPIR_KERNEL;
+  }
+};
+
+class SPIRVTargetCIRGenInfo : public CommonSPIRTargetCIRGenInfo {
+public:
+  SPIRVTargetCIRGenInfo(CIRGenTypes &CGT)
+      : CommonSPIRTargetCIRGenInfo(std::make_unique<SPIRVABIInfo>(CGT)) {}
+};
+
+} // namespace
+
 // TODO(cir): remove the attribute once this gets used.
 LLVM_ATTRIBUTE_UNUSED
 static bool classifyReturnType(const CIRGenCXXABI &CXXABI,
@@ -189,13 +311,6 @@ void X86_64ABIInfo::computeInfo(CIRGenFunctionInfo &FI) const {
     FI.getReturnInfo() = ABIArgInfo::getDirect(CGT.ConvertType(RetTy));
 }
 
-/// Pass transparent unions as if they were the type of the first element. Sema
-/// should ensure that all elements of the union have the same "machine type".
-static QualType useFirstFieldIfTransparentUnion(QualType Ty) {
-  assert(!Ty->getAsUnionType() && "NYI");
-  return Ty;
-}
-
 /// GetINTEGERTypeAtOffset - The ABI specifies that a value should be passed in
 /// an 8-byte GPR. This means that we either have a scalar or we are talking
 /// about the high or low part of an up-to-16-byte struct. This routine picks
@@ -459,5 +574,9 @@ const TargetCIRGenInfo &CIRGenModule::getTargetCIRGenInfo() {
       return SetCIRGenInfo(new X86_64TargetCIRGenInfo(genTypes, AVXLevel));
     }
   }
+
+  case llvm::Triple::spirv64: {
+    return SetCIRGenInfo(new SPIRVTargetCIRGenInfo(genTypes));
+  }
   }
 }

clang/lib/CIR/CodeGen/TargetInfo.h

@@ -78,6 +78,21 @@ class TargetCIRGenInfo {
                                            mlir::Type DestTy,
                                            bool IsNonNull = false) const;
 
+  /// Get LLVM calling convention for OpenCL kernel.
+  virtual unsigned getOpenCLKernelCallingConv() const {
+    // OpenCL kernels are called via an explicit runtime API with arguments
+    // set with clSetKernelArg(), not as normal sub-functions.
+    // Return SPIR_KERNEL by default as the kernel calling convention to
+    // ensure the fingerprint is fixed such way that each OpenCL argument
+    // gets one matching argument in the produced kernel function argument
+    // list to enable feasible implementation of clSetKernelArg() with
+    // aggregates etc. In case we would use the default C calling conv here,
+    // clSetKernelArg() might break depending on the target-specific
+    // conventions; different targets might split structs passed as values
+    // to multiple function arguments etc.
+    return llvm::CallingConv::SPIR_KERNEL;
+  }
+
   virtual ~TargetCIRGenInfo() {}
 };
 

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

@@ -0,0 +1,25 @@
+// RUN: %clang_cc1 -cl-std=CL3.0 -O0 -fclangir -emit-cir -triple spirv64-unknown-unknown %s -o %t.cir
+// RUN: FileCheck --input-file=%t.cir %s --check-prefix=CIR
+// 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)>
+// 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}
+  // 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}
+  // 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)>>
+  // LLVM-NEXT: store ptr addrspace(3) %{{[0-9]+}}, ptr %[[#ALLOCA_P]], align 8
+
+  return;
+}

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

@@ -0,0 +1,29 @@
+// See also: clang/test/CodeGenOpenCL/spirv_target.cl
+// RUN: %clang_cc1 -cl-std=CL3.0 -fclangir -emit-cir -triple spirv64-unknown-unknown %s -o %t_64.cir
+// RUN: FileCheck --input-file=%t_64.cir %s --check-prefix=CIR-SPIRV64
+// 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
+
+// CIR-SPIRV64: cir.triple = "spirv64-unknown-unknown"
+// LLVM-SPIRV64: target triple = "spirv64-unknown-unknown"
+
+typedef struct {
+  char c;
+  void *v;
+  void *v2;
+} my_st;
+
+// CIR-SPIRV64: cir.func @func(
+// LLVM-SPIRV64: @func(
+kernel void func(global long *arg) {
+  int res1[sizeof(my_st)  == 24 ? 1 : -1]; // expected-no-diagnostics
+  int res2[sizeof(void *) ==  8 ? 1 : -1]; // expected-no-diagnostics
+  int res3[sizeof(arg)    ==  8 ? 1 : -1]; // expected-no-diagnostics
+
+  my_st *tmp = 0;
+
+  // LLVM-SPIRV64: store i64 8, ptr addrspace(1)
+  arg[0] = (long)(&tmp->v);
+  // LLVM-SPIRV64: store i64 16, ptr addrspace(1)
+  arg[1] = (long)(&tmp->v2);
+}