Initial Metal backend support to dcompute

driver/dcomputecodegenerator.cpp

@@ -17,7 +17,7 @@
 #include <string>
 #include <algorithm>
 
-#if !(LDC_LLVM_SUPPORTED_TARGET_SPIRV || LDC_LLVM_SUPPORTED_TARGET_NVPTX)
+#if !(LDC_LLVM_SUPPORTED_TARGET_SPIRV || LDC_LLVM_SUPPORTED_TARGET_NVPTX || LDC_LLVM_SUPPORTED_TARGET_AArch64)
 
 DComputeCodeGenManager::DComputeCodeGenManager(llvm::LLVMContext &c) : ctx(c) {}
 void DComputeCodeGenManager::emit(Module *) {}
@@ -43,6 +43,20 @@ DComputeCodeGenManager::createComputeTarget(const std::string &s) {
 #endif
   }
 
+  if (s.substr(0, 6) == "metal-") {
+#if  LDC_LLVM_SUPPORTED_TARGET_AArch64 //&& LDC_LLVM_VER >= 2100
+#define METAL_VALID_VER_INIT 400
+    const std::array<int, 6> valid_metal_versions = {{METAL_VALID_VER_INIT}};
+   const int v = atoi(s.c_str() + 6);
+   if (std::find(valid_metal_versions.begin(), valid_metal_versions.end(), v) !=
+        valid_metal_versions.end()) {
+      return createMetalTarget(ctx, v);
+    } 
+#else
+    error(Loc(), "LDC was not built with Apple Metal Dcompute support!");
+#endif
+  }
+
   if (s.substr(0, 5) == "cuda-") {
 #if LDC_LLVM_SUPPORTED_TARGET_NVPTX
 #define CUDA_VALID_VER_INIT 100, 110, 120, 130, 200, 210, 300, 350, 370,\
@@ -64,14 +78,18 @@ DComputeCodeGenManager::createComputeTarget(const std::string &s) {
 
   error(Loc(),
         "Unrecognised or invalid DCompute targets: the format is ocl-xy0 "
-        "for OpenCl x.y and cuda-xy0 for CUDA CC x.y."
+        "for OpenCl x.y and cuda-xy0 for CUDA CC x.y and metal-xy0 for Metal x.y."
 #if LDC_LLVM_SUPPORTED_TARGET_SPIRV
         " Valid version strings for OpenCl are ocl-{" XSTR(OCL_VALID_VER_INIT) "}."
 #endif
 #if LDC_LLVM_SUPPORTED_TARGET_NVPTX
         " Valid version strings for CUDA are cuda-{" XSTR(CUDA_VALID_VER_INIT) "}."
+#endif
+#if LDC_LLVM_SUPPORTED_TARGET_AArch64
+        "Valid version strings for Metal are metal-{" XSTR(METAL_VALID_VER_INIT) "}"
 #endif
   );
+
 
 #undef XSTR
 #undef STR
@@ -119,4 +137,4 @@ DComputeCodeGenManager::~DComputeCodeGenManager() {
   gABI = oldGABI;
 }
 
-#endif // LDC_LLVM_SUPPORTED_TARGET_SPIRV || LDC_LLVM_SUPPORTED_TARGET_NVPTX
+#endif // LDC_LLVM_SUPPORTED_TARGET_SPIRV || LDC_LLVM_SUPPORTED_TARGET_NVPTX || LDC_LLVM_SUPPORTED_TARGET_AArch64

driver/dcomputecodegenerator.h

@@ -24,7 +24,7 @@ struct TargetABI;
 class DComputeCodeGenManager {
 
   llvm::LLVMContext &ctx;
-  llvm::SmallVector<DComputeTarget *, 2> targets;
+  llvm::SmallVector<DComputeTarget *, 3> targets;
   DComputeTarget *createComputeTarget(const std::string &s);
   IRState *oldGIR = nullptr;
   llvm::TargetMachine *oldGTargetMachine = nullptr;

driver/targetmachine.cpp

@@ -666,6 +666,16 @@ ComputeBackend::Type getComputeTargetType(llvm::Module* m) {
     return ComputeBackend::SPIRV;
   else if (a == llvm::Triple::nvptx || a == llvm::Triple::nvptx64)
     return ComputeBackend::NVPTX;
-  else
-    return ComputeBackend::None;
+
+
+#if LDC_LLVM_SUPPORTED_TARGET_AArch64 && LLVM_VERSION_MAJOR >= 21
+    llvm::StringRef tripleString = m->getTargetTriple().str();
+#else
+    llvm::StringRef tripleString = m->getTargetTriple();
+#endif
+
+
+  if (tripleString.starts_with("air64"))
+    return ComputeBackend::METAL; 
+  return ComputeBackend::None;
 }

driver/targetmachine.h

@@ -41,7 +41,7 @@ class Module;
 }
 
 namespace ComputeBackend {
-enum Type { None, SPIRV, NVPTX };
+enum Type { None, SPIRV, NVPTX, METAL };
 }
 
 ComputeBackend::Type getComputeTargetType(llvm::Module*);

driver/toobj.cpp

@@ -35,10 +35,10 @@
 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/CodeGen/TargetSubtargetInfo.h"
+#include "llvm/Transforms/IPO/AlwaysInliner.h"
 #include "llvm/Transforms/Utils/Cloning.h"
 #include "llvm/IR/Module.h"
 #include <cstddef>
-#include <fstream>
 
 namespace llvm {
 namespace codegen {
@@ -58,6 +58,40 @@ void runDLLImportRelocationPass(llvm::TargetMachine &Target, llvm::Module &m) {
   pm.run(m);
 }
 
+void inlineDComputeKernelFunctions(llvm::Module *m) {
+  // Create a PassManager to hold and optimize the collection of passes we are
+  // about to build.
+  llvm::legacy::PassManager Passes;
+
+  llvm::SmallPtrSet<llvm::Function*, 8> kernelFunctions;
+
+  // Extract all the kernel functions
+  if (auto *kernelMetadata = m->getNamedMetadata("air.kernel")) {
+    for(auto *op: kernelMetadata->operands()) {
+      if (auto *F = llvm::mdconst::dyn_extract<llvm::Function>(op->getOperand(0))) {
+        kernelFunctions.insert(F);
+      }
+    }
+  }
+
+  // Prepare non-kernel functions to be inlined
+  for(auto& F: *m) {
+    if (!F.isDeclaration() && !kernelFunctions.contains(&F)) {
+     F.addFnAttr(llvm::Attribute::AlwaysInline);
+    }
+  }
+
+  Passes.add(llvm::createAlwaysInlinerLegacyPass());
+
+  Passes.run(*m);
+
+  // Terminate upon errors during the LLVM passes.
+  if (global.errors || global.warnings) {
+  error(Loc(), "Aborting because of errors/warnings during LLVM passes");
+    fatal();
+  }
+}
+
 // based on llc code, University of Illinois Open Source License
 void codegenModule(llvm::TargetMachine &Target, llvm::Module &m,
                    const char *filename,
@@ -73,6 +107,63 @@ void codegenModule(llvm::TargetMachine &Target, llvm::Module &m,
 #endif
   }
 
+#ifdef LDC_LLVM_SUPPORTED_TARGET_AArch64
+  if (cb == ComputeBackend::METAL) {
+#ifdef __APPLE__
+    {
+      // Inline non-kernel functions for Metal dcompute target
+      inlineDComputeKernelFunctions(&m);
+
+      std::error_code errinfo;
+      llvm::ToolOutputFile out(filename, errinfo, llvm::sys::fs::OF_None);
+      if (errinfo) {
+        error(Loc(), "cannot write file '%s': %s", filename,
+              errinfo.message().c_str());
+        fatal();
+      }
+
+      llvm::WriteBitcodeToFile(m, out.os());
+
+      out.keep();
+
+      // Terminate upon errors during the LLVM passes.
+      if (global.errors || global.warnings) {
+        error(Loc(), "Aborting because of errors/warnings during LLVM passes!");
+        fatal();
+      }
+    }
+
+    auto xcrunpath = llvm::sys::findProgramByName("xcrun");
+    if (!xcrunpath) {
+      error(Loc(), "xcrun not found - XCode should be installed first!");
+      fatal();
+    }
+
+    llvm::SmallString<256> metallibOutPath;
+    llvm::sys::fs::current_path(metallibOutPath);
+    llvm::sys::path::append(metallibOutPath, llvm::sys::path::filename(filename));
+    llvm::sys::path::replace_extension(metallibOutPath, "metallib");
+
+    std::vector<std::string> args = {
+      xcrunpath.get(), "-sdk", "macosx", "metallib", filename, "-o", metallibOutPath.c_str()
+    };
+
+    std::string errorMsg;
+
+    int status =  executeToolAndWait(Loc(), args[0], args);
+
+    if (status < 0) {
+      error(Loc(), "program received signal %d (%s)", -status,
+          strsignal(-status));
+      fatal();
+    }
+
+#endif
+
+    return;
+  }
+#endif
+
   std::error_code errinfo;
   llvm::ToolOutputFile out(filename, errinfo, llvm::sys::fs::OF_None);
   if (errinfo) {
@@ -303,6 +394,15 @@ std::string replaceExtensionWith(const DArray<const char> &ext,
 }
 
 void writeModule(llvm::Module *m, const char *filename) {
+// Inline non-kernel functions for Metal dcompute target
+#ifdef LDC_LLVM_SUPPORTED_TARGET_AArch64
+  const ComputeBackend::Type cb = getComputeTargetType(m);
+
+  if (cb == ComputeBackend::METAL) {
+    inlineDComputeKernelFunctions(m);
+  }
+#endif
+
   const bool doLTO = opts::isUsingLTO();
   const bool outputObj = shouldOutputObjectFile();
   const bool assembleExternally = shouldAssembleExternally();

gen/abi/abi.cpp

@@ -11,8 +11,6 @@
 
 #include "dmd/argtypes.h"
 #include "dmd/expression.h"
-#include "dmd/id.h"
-#include "dmd/identifier.h"
 #include "dmd/target.h"
 #include "gen/abi/targets.h"
 #include "gen/abi/generic.h"
@@ -24,7 +22,6 @@
 #include "gen/tollvm.h"
 #include "ir/irfunction.h"
 #include "ir/irfuncty.h"
-#include <algorithm>
 
 using namespace dmd;
 
@@ -302,6 +299,13 @@ TargetABI *TargetABI::getTarget() {
   case llvm::Triple::wasm32:
   case llvm::Triple::wasm64:
     return getWasmTargetABI();
+
+  case llvm::Triple::UnknownArch:
+    if (global.params.targetTriple->getArchName() == "air64") {
+      return createMetalABI();
+    }
+    // fallthrough
+
   default:
     warning(Loc(),
             "unknown target ABI, falling back to generic implementation. C/C++ "

gen/abi/metal.cpp

@@ -0,0 +1,60 @@
+//===-- gen/abi-metal.cpp ---------------------------------------*- C++ -*-===//
+//
+//                         LDC – the LLVM D compiler
+//
+// This file is distributed under the BSD-style LDC license. See the LICENSE
+// file for details.
+//
+//===----------------------------------------------------------------------===//
+
+#include "gen/abi/abi.h"
+#include "gen/dcompute/druntime.h"
+#include "gen/dcompute/abi-rewrites.h"
+#include "ir/irfuncty.h"
+#include "dmd/mtype.h"
+#include <optional>
+
+using namespace dmd;
+
+struct MetalABI : TargetABI {
+    DComputePointerRewrite pointerRewite;
+    DcomputeMetalScalarRewrite metalScalarRewrite;
+
+    auto returnInArg(TypeFunction *tf, bool needsThis) -> bool override {
+        return false;
+    }
+
+    auto passByVal(TypeFunction *tf, Type*t) -> bool override {
+        return false;
+    }
+
+    void rewriteFunctionType(IrFuncTy &fty) override {
+      for (auto arg : fty.args) {
+        if (!arg->byref) {
+          rewriteArgument(fty, *arg);
+        }
+      }
+    }
+
+    void rewriteArgument(IrFuncTy &fty, IrFuncTyArg &arg) override {
+        TargetABI::rewriteArgument(fty, arg);
+
+        if (arg.rewrite) {
+            return;
+        }
+
+        Type *ty = arg.type->toBasetype();
+        std::optional<DcomputePointer> ptr;
+
+        if (ty->ty == TY::Tstruct &&
+            (ptr = toDcomputePointer(static_cast<TypeStruct *>(ty)->sym))) {
+                pointerRewite.applyTo(arg);
+            }
+
+        if (dmd::isScalar(ty)) {
+          metalScalarRewrite.applyTo(arg);
+        }
+    }
+};
+
+TargetABI* createMetalABI() { return new MetalABI(); }

gen/abi/targets.h

@@ -40,3 +40,5 @@ TargetABI *getX86TargetABI();
 TargetABI *getLoongArch64TargetABI();
 
 TargetABI *getWasmTargetABI();
+
+TargetABI* createMetalABI();

gen/dcompute/abi-rewrites.h

@@ -32,3 +32,19 @@ struct DComputePointerRewrite : ABIRewrite {
     return ptr->toLLVMType(true);
   }
 };
+
+struct DcomputeMetalScalarRewrite : ABIRewrite {
+  LLType *type(Type* t) override {
+    // XXX: Scalar variables are stored in the constant memory space for Metal GPU
+    return llvm::PointerType::get(gIR->context(), 2/*Constant Memory space*/);
+  }
+
+  LLValue *getLVal(Type *dty, LLValue *v) override {
+    return v;
+  }
+
+  LLValue *put(DValue *v, bool isLValueExp, bool) override {
+    auto value = DtoRVal(v);
+    return value;
+  }
+};

gen/dcompute/target.cpp

@@ -7,7 +7,7 @@
 //
 //===----------------------------------------------------------------------===//
 
-#if LDC_LLVM_SUPPORTED_TARGET_SPIRV || LDC_LLVM_SUPPORTED_TARGET_NVPTX
+#if LDC_LLVM_SUPPORTED_TARGET_SPIRV || LDC_LLVM_SUPPORTED_TARGET_NVPTX || LDC_LLVM_SUPPORTED_TARGET_AArch64
 
 #include "dmd/dsymbol.h"
 #include "dmd/errors.h"

gen/dcompute/target.h

@@ -28,7 +28,7 @@ class DComputeTarget {
 public:
   llvm::LLVMContext &ctx;
   int tversion; // OpenCL or CUDA CC version:major*100 + minor*10
-  enum class ID { Host = 0, OpenCL = 1, CUDA = 2 };
+  enum class ID { Host = 0, OpenCL = 1, CUDA = 2, Metal = 3 };
   ID target;    // ID for codegen time conditional compilation.
   const char *short_name;
   const char *binSuffix;
@@ -64,6 +64,10 @@ class DComputeTarget {
 DComputeTarget *createCUDATarget(llvm::LLVMContext &c, int sm);
 #endif
 
+#if LDC_LLVM_SUPPORTED_TARGET_AArch64
+DComputeTarget* createMetalTarget(llvm::LLVMContext &c, int version);
+#endif
+
 #if LDC_LLVM_SUPPORTED_TARGET_SPIRV
 DComputeTarget *createOCLTarget(llvm::LLVMContext &c, int oclver);
 #endif