clover: Initialize Asm Parsers
[mesa.git] / src / gallium / state_trackers / clover / llvm / invocation.cpp
index 9c102d2428fc872a3ea3e03818803d95729f5ccb..95a9d036622a29117087221badb822da8a0059a9 100644 (file)
 // OTHER DEALINGS IN THE SOFTWARE.
 //
 
-#include "llvm/compat.hpp"
-#include "core/compiler.hpp"
-#include "util/algorithm.hpp"
-
-#include <clang/Frontend/CompilerInstance.h>
-#include <clang/Frontend/TextDiagnosticBuffer.h>
-#include <clang/Frontend/TextDiagnosticPrinter.h>
-#include <clang/CodeGen/CodeGenAction.h>
-#include <clang/Basic/TargetInfo.h>
-#include <llvm/Bitcode/BitstreamWriter.h>
-#include <llvm/Bitcode/ReaderWriter.h>
-#include <llvm/Linker/Linker.h>
-#include <llvm/IR/DiagnosticInfo.h>
 #include <llvm/IR/DiagnosticPrinter.h>
-#include <llvm/IR/DerivedTypes.h>
+#include <llvm/IR/DiagnosticInfo.h>
 #include <llvm/IR/LLVMContext.h>
-#include <llvm/IR/Module.h>
-#include <llvm/Support/SourceMgr.h>
-#include <llvm/IRReader/IRReader.h>
-#include <llvm/Support/CodeGen.h>
-#include <llvm/Support/TargetSelect.h>
-#include <llvm/Support/MemoryBuffer.h>
-#include <llvm/Support/FormattedStream.h>
-#include <llvm/Support/TargetRegistry.h>
-#include <llvm/Transforms/IPO.h>
+#include <llvm/Support/raw_ostream.h>
 #include <llvm/Transforms/IPO/PassManagerBuilder.h>
-#include <llvm/Transforms/Utils/Cloning.h>
-
+#include <llvm-c/Target.h>
+#ifdef HAVE_CLOVER_SPIRV
+#include <LLVMSPIRVLib/LLVMSPIRVLib.h>
+#endif
 
-#include <llvm/IR/DataLayout.h>
-#include <llvm/Target/TargetMachine.h>
-#include <llvm/Target/TargetOptions.h>
+#include <clang/CodeGen/CodeGenAction.h>
+#include <clang/Lex/PreprocessorOptions.h>
+#include <clang/Frontend/TextDiagnosticBuffer.h>
+#include <clang/Frontend/TextDiagnosticPrinter.h>
+#include <clang/Basic/TargetInfo.h>
 
-#include <llvm-c/Target.h>
-#include <llvm-c/TargetMachine.h>
-#include <llvm-c/Core.h>
+// We need to include internal headers last, because the internal headers
+// include CL headers which have #define's like:
+//
+//#define cl_khr_gl_sharing 1
+//#define cl_khr_icd 1
+//
+// Which will break the compilation of clang/Basic/OpenCLOptions.h
 
-#include "pipe/p_state.h"
-#include "util/u_memory.h"
-#include "util/u_math.h"
+#include "core/error.hpp"
+#include "llvm/codegen.hpp"
+#include "llvm/compat.hpp"
+#include "llvm/invocation.hpp"
+#include "llvm/metadata.hpp"
+#include "llvm/util.hpp"
+#ifdef HAVE_CLOVER_SPIRV
+#include "spirv/invocation.hpp"
+#endif
+#include "util/algorithm.hpp"
 
-#include <iostream>
-#include <iomanip>
-#include <fstream>
-#include <cstdio>
-#include <sstream>
-#include <libelf.h>
-#include <gelf.h>
 
 using namespace clover;
 using namespace clover::llvm;
@@ -82,42 +69,33 @@ using ::llvm::Module;
 using ::llvm::raw_string_ostream;
 
 namespace {
-   // XXX - Temporary hack to avoid breaking the build for the moment, will
-   //       get rid of this later.
-   namespace llvm {
-      using namespace ::llvm;
-   }
-
-   void debug_log(const std::string &msg, const std::string &suffix) {
-      const char *dbg_file = debug_get_option("CLOVER_DEBUG_FILE", "stderr");
-      if (!strcmp("stderr", dbg_file)) {
-         std::cerr << msg;
-       } else {
-        std::ofstream file(dbg_file + suffix, std::ios::app);
-        file << msg;
-       }
-   }
-
-   inline std::vector<std::string>
-   tokenize(const std::string &s) {
-      std::vector<std::string> ss;
-      std::istringstream iss(s);
-      std::string t;
-
-      while (getline(iss, t, ' '))
-         ss.push_back(t);
 
-      return ss;
-   }
+    struct cl_version {
+        std::string version_str; // CL Version
+        unsigned version_number; // Numeric CL Version
+    };
+
+   static const unsigned ANY_VERSION = 999;
+   const cl_version cl_versions[] = {
+      { "1.0", 100},
+      { "1.1", 110},
+      { "1.2", 120},
+      { "2.0", 200},
+      { "2.1", 210},
+      { "2.2", 220},
+   };
 
-   struct target {
-      target(const std::string &s) :
-         cpu(s.begin(), s.begin() + s.find_first_of("-")),
-         triple(s.begin() + s.find_first_of("-") + 1, s.end()) {}
+    struct clc_version_lang_std {
+        unsigned version_number; // CLC Version
+        clang::LangStandard::Kind clc_lang_standard;
+    };
 
-      std::string cpu;
-      std::string triple;
-   };
+    const clc_version_lang_std cl_version_lang_stds[] = {
+       { 100, compat::lang_opencl10},
+       { 110, clang::LangStandard::lang_opencl11},
+       { 120, clang::LangStandard::lang_opencl12},
+       { 200, clang::LangStandard::lang_opencl20},
+    };
 
    void
    init_targets() {
@@ -126,22 +104,19 @@ namespace {
          LLVMInitializeAllTargets();
          LLVMInitializeAllTargetInfos();
          LLVMInitializeAllTargetMCs();
+         LLVMInitializeAllAsmParsers();
          LLVMInitializeAllAsmPrinters();
          targets_initialized = true;
       }
    }
 
    void
-   diagnostic_handler(const llvm::DiagnosticInfo &di, void *data) {
-      if (di.getSeverity() == llvm::DS_Error) {
-         std::string message = *(std::string*)data;
-         llvm::raw_string_ostream stream(message);
-         llvm::DiagnosticPrinterRawOStream dp(stream);
-         di.print(dp);
-         stream.flush();
-         *(std::string*)data = message;
-
-         throw compile_error();
+   diagnostic_handler(const ::llvm::DiagnosticInfo &di, void *data) {
+      if (di.getSeverity() == ::llvm::DS_Error) {
+         raw_string_ostream os { *reinterpret_cast<std::string *>(data) };
+         ::llvm::DiagnosticPrinterRawOStream printer { os };
+         di.print(printer);
+         throw build_error();
       }
    }
 
@@ -149,17 +124,78 @@ namespace {
    create_context(std::string &r_log) {
       init_targets();
       std::unique_ptr<LLVMContext> ctx { new LLVMContext };
-      ctx->setDiagnosticHandler(diagnostic_handler, &r_log);
+      compat::set_diagnostic_handler(*ctx, diagnostic_handler, &r_log);
       return ctx;
    }
 
+   const struct clc_version_lang_std&
+   get_cl_lang_standard(unsigned requested, unsigned max = ANY_VERSION) {
+       for (const struct clc_version_lang_std &version : cl_version_lang_stds) {
+           if (version.version_number == max ||
+                   version.version_number == requested) {
+               return version;
+           }
+       }
+       throw build_error("Unknown/Unsupported language version");
+   }
+
+   const struct cl_version&
+   get_cl_version(const std::string &version_str,
+                  unsigned max = ANY_VERSION) {
+      for (const struct cl_version &version : cl_versions) {
+         if (version.version_number == max || version.version_str == version_str) {
+            return version;
+         }
+      }
+      throw build_error("Unknown/Unsupported language version");
+   }
+
+   clang::LangStandard::Kind
+   get_lang_standard_from_version_str(const std::string &version_str,
+                                      bool is_build_opt = false) {
+
+       //Per CL 2.0 spec, section 5.8.4.5:
+       //  If it's an option, use the value directly.
+       //  If it's a device version, clamp to max 1.x version, a.k.a. 1.2
+      const cl_version version =
+         get_cl_version(version_str, is_build_opt ? ANY_VERSION : 120);
+
+      const struct clc_version_lang_std standard =
+         get_cl_lang_standard(version.version_number);
+
+      return standard.clc_lang_standard;
+   }
+
+   clang::LangStandard::Kind
+   get_language_version(const std::vector<std::string> &opts,
+                        const std::string &device_version) {
+
+      const std::string search = "-cl-std=CL";
+
+      for (auto &opt: opts) {
+         auto pos = opt.find(search);
+         if (pos == 0){
+            const auto ver = opt.substr(pos + search.size());
+            const auto device_ver = get_cl_version(device_version);
+            const auto requested = get_cl_version(ver);
+            if (requested.version_number > device_ver.version_number) {
+               throw build_error();
+            }
+            return get_lang_standard_from_version_str(ver, true);
+         }
+      }
+
+      return get_lang_standard_from_version_str(device_version);
+   }
+
    std::unique_ptr<clang::CompilerInstance>
-   create_compiler_instance(const target &target,
+   create_compiler_instance(const device &dev, const std::string& ir_target,
                             const std::vector<std::string> &opts,
                             std::string &r_log) {
       std::unique_ptr<clang::CompilerInstance> c { new clang::CompilerInstance };
+      clang::TextDiagnosticBuffer *diag_buffer = new clang::TextDiagnosticBuffer;
       clang::DiagnosticsEngine diag { new clang::DiagnosticIDs,
-            new clang::DiagnosticOptions, new clang::TextDiagnosticBuffer };
+            new clang::DiagnosticOptions, diag_buffer };
 
       // Parse the compiler options.  A file name should be present at the end
       // and must have the .cl extension in order for the CompilerInvocation
@@ -167,9 +203,16 @@ namespace {
       const std::vector<const char *> copts =
          map(std::mem_fn(&std::string::c_str), opts);
 
-      if (!clang::CompilerInvocation::CreateFromArgs(
-             c->getInvocation(), copts.data(), copts.data() + copts.size(), diag))
-         throw error(CL_INVALID_COMPILER_OPTIONS);
+      const target &target = ir_target;
+      const std::string &device_clc_version = dev.device_clc_version();
+
+      if (!compat::create_compiler_invocation_from_args(
+             c->getInvocation(), copts, diag))
+         throw invalid_build_options_error();
+
+      diag_buffer->FlushDiagnostics(diag);
+      if (diag.hasErrorOccurred())
+         throw invalid_build_options_error();
 
       c->getTargetOpts().CPU = target.cpu;
       c->getTargetOpts().Triple = target.triple;
@@ -180,10 +223,10 @@ namespace {
       // http://www.llvm.org/bugs/show_bug.cgi?id=19735
       c->getDiagnosticOpts().ShowCarets = false;
 
-      compat::set_lang_defaults(c->getInvocation(), c->getLangOpts(),
-                                clang::IK_OpenCL, ::llvm::Triple(target.triple),
+      c->getInvocation().setLangDefaults(c->getLangOpts(),
+                                compat::ik_opencl, ::llvm::Triple(target.triple),
                                 c->getPreprocessorOpts(),
-                                clang::LangStandard::lang_opencl11);
+                                get_language_version(opts, device_clc_version));
 
       c->createDiagnostics(new clang::TextDiagnosticPrinter(
                               *new raw_string_ostream(r_log),
@@ -195,47 +238,55 @@ namespace {
       return c;
    }
 
-   llvm::Module *
-   compile_llvm(LLVMContext &ctx, clang::CompilerInstance &c,
-                const std::string &source, const header_map &headers,
-                const std::string &name, const std::string &target,
-                const std::string &opts, std::string &r_log) {
-      clang::EmitLLVMOnlyAction act(&ctx);
-
+   std::unique_ptr<Module>
+   compile(LLVMContext &ctx, clang::CompilerInstance &c,
+           const std::string &name, const std::string &source,
+           const header_map &headers, const device &dev,
+           const std::string &opts, bool use_libclc, std::string &r_log) {
       c.getFrontendOpts().ProgramAction = clang::frontend::EmitLLVMOnly;
       c.getHeaderSearchOpts().UseBuiltinIncludes = true;
       c.getHeaderSearchOpts().UseStandardSystemIncludes = true;
       c.getHeaderSearchOpts().ResourceDir = CLANG_RESOURCE_DIR;
 
-      // Add libclc generic search path
-      c.getHeaderSearchOpts().AddPath(LIBCLC_INCLUDEDIR,
-                                      clang::frontend::Angled,
-                                      false, false
-                                      );
+      if (use_libclc) {
+         // Add libclc generic search path
+         c.getHeaderSearchOpts().AddPath(LIBCLC_INCLUDEDIR,
+                                         clang::frontend::Angled,
+                                         false, false);
+
+         // Add libclc include
+         c.getPreprocessorOpts().Includes.push_back("clc/clc.h");
+      } else {
+         // Add opencl-c generic search path
+         c.getHeaderSearchOpts().AddPath(CLANG_RESOURCE_DIR,
+                                         clang::frontend::Angled,
+                                         false, false);
+
+         // Add opencl include
+         c.getPreprocessorOpts().Includes.push_back("opencl-c.h");
+      }
 
-      // Add libclc include
-      c.getPreprocessorOpts().Includes.push_back("clc/clc.h");
+      // Add definition for the OpenCL version
+      c.getPreprocessorOpts().addMacroDef("__OPENCL_VERSION__=" +
+              std::to_string(get_cl_version(
+                                  dev.device_version()).version_number));
 
       // clc.h requires that this macro be defined:
       c.getPreprocessorOpts().addMacroDef("cl_clang_storage_class_specifiers");
-
-      c.getPreprocessorOpts().addRemappedFile(name,
-                                              llvm::MemoryBuffer::getMemBuffer(source).release());
+      c.getPreprocessorOpts().addRemappedFile(
+              name, ::llvm::MemoryBuffer::getMemBuffer(source).release());
 
       if (headers.size()) {
          const std::string tmp_header_path = "/tmp/clover/";
 
          c.getHeaderSearchOpts().AddPath(tmp_header_path,
                                          clang::frontend::Angled,
-                                         false, false
-                                         );
-
-         for (header_map::const_iterator it = headers.begin();
-              it != headers.end(); ++it) {
-            const std::string path = tmp_header_path + std::string(it->first);
-            c.getPreprocessorOpts().addRemappedFile(path,
-                    llvm::MemoryBuffer::getMemBuffer(it->second.c_str()).release());
-         }
+                                         false, false);
+
+         for (const auto &header : headers)
+            c.getPreprocessorOpts().addRemappedFile(
+               tmp_header_path + header.first,
+               ::llvm::MemoryBuffer::getMemBuffer(header.second).release());
       }
 
       // Tell clang to link this file before performing any
@@ -245,49 +296,45 @@ namespace {
       // attribute will prevent Clang from creating illegal uses of
       // barrier() (e.g. Moving barrier() inside a conditional that is
       // no executed by all threads) during its optimizaton passes.
-      compat::add_link_bitcode_file(c.getCodeGenOpts(),
-                                    LIBCLC_LIBEXECDIR + target + ".bc");
+      if (use_libclc)
+         compat::add_link_bitcode_file(c.getCodeGenOpts(),
+                                       LIBCLC_LIBEXECDIR + dev.ir_target() + ".bc");
 
       // Compile the code
+      clang::EmitLLVMOnlyAction act(&ctx);
       if (!c.ExecuteAction(act))
-         throw compile_error();
+         throw build_error();
 
-      return act.takeModule().release();
+      return act.takeModule();
    }
+}
 
-   std::vector<llvm::Function *>
-   find_kernels(llvm::Module *mod) {
-      std::vector<llvm::Function *> kernels;
-#if HAVE_LLVM >= 0x0309
-      auto &list = mod->getFunctionList();
-      for_each(list.begin(), list.end(), [&](llvm::Function &f){
-         if (f.getMetadata("kernel_arg_type"))
-           kernels.push_back(&f);
-      });
-      return kernels;
-#endif
-      const llvm::NamedMDNode *kernel_node =
-                                 mod->getNamedMetadata("opencl.kernels");
-      // This means there are no kernels in the program.  The spec does not
-      // require that we return an error here, but there will be an error if
-      // the user tries to pass this program to a clCreateKernel() call.
-      if (!kernel_node) {
-         return std::vector<llvm::Function *>();
-      }
+module
+clover::llvm::compile_program(const std::string &source,
+                              const header_map &headers,
+                              const device &dev,
+                              const std::string &opts,
+                              std::string &r_log) {
+   if (has_flag(debug::clc))
+      debug::log(".cl", "// Options: " + opts + '\n' + source);
 
-      kernels.reserve(kernel_node->getNumOperands());
-      for (unsigned i = 0; i < kernel_node->getNumOperands(); ++i) {
-         kernels.push_back(llvm::mdconst::dyn_extract<llvm::Function>(
-                                    kernel_node->getOperand(i)->getOperand(0)));
-      }
-      return kernels;
-   }
+   auto ctx = create_context(r_log);
+   auto c = create_compiler_instance(dev, dev.ir_target(),
+                                     tokenize(opts + " input.cl"), r_log);
+   auto mod = compile(*ctx, *c, "input.cl", source, headers, dev, opts, true,
+                      r_log);
 
-   void
-   optimize(llvm::Module *mod, unsigned optimization_level) {
-      compat::pass_manager PM;
+   if (has_flag(debug::llvm))
+      debug::log(".ll", print_module_bitcode(*mod));
+
+   return build_module_library(*mod, module::section::text_intermediate);
+}
 
-      compat::add_data_layout_pass(PM);
+namespace {
+   void
+   optimize(Module &mod, unsigned optimization_level,
+            bool internalize_symbols) {
+      ::llvm::legacy::PassManager pm;
 
       // By default, the function internalizer pass will look for a function
       // called "main" and then mark all other functions as internal.  Marking
@@ -301,533 +348,118 @@ namespace {
       // list of kernel functions to the internalizer.  The internalizer will
       // treat the functions in the list as "main" functions and internalize
       // all of the other functions.
-      compat::add_internalize_pass(PM, map(std::mem_fn(&Function::getName),
-                                           find_kernels(mod)));
-
-      llvm::PassManagerBuilder PMB;
-      PMB.OptLevel = optimization_level;
-      PMB.LibraryInfo = new compat::target_library_info(
-         llvm::Triple(mod->getTargetTriple()));
-      PMB.populateModulePassManager(PM);
-      PM.run(*mod);
-   }
-
-   // Kernel metadata
-
-   struct kernel_arg_md {
-      llvm::StringRef type_name;
-      llvm::StringRef access_qual;
-      kernel_arg_md(llvm::StringRef type_name_, llvm::StringRef access_qual_):
-         type_name(type_name_), access_qual(access_qual_) {}
-   };
-#if HAVE_LLVM >= 0x0309
-   std::vector<kernel_arg_md>
-   get_kernel_arg_md(const llvm::Function *kernel_func) {
-
-      size_t num_args = kernel_func->getArgumentList().size();
-
-      auto aq = kernel_func->getMetadata("kernel_arg_access_qual");
-      auto ty = kernel_func->getMetadata("kernel_arg_type");
-
-      std::vector<kernel_arg_md> res;
-      res.reserve(num_args);
-      for (size_t i = 0; i < num_args; ++i) {
-         res.push_back(kernel_arg_md(
-            llvm::cast<llvm::MDString>(ty->getOperand(i))->getString(),
-            llvm::cast<llvm::MDString>(aq->getOperand(i))->getString()));
-      }
-
-      return res;
-   }
-
-#else
-
-   const llvm::MDNode *
-   get_kernel_metadata(const llvm::Function *kernel_func) {
-      auto mod = kernel_func->getParent();
-      auto kernels_node = mod->getNamedMetadata("opencl.kernels");
-      if (!kernels_node) {
-         return nullptr;
-      }
-
-      const llvm::MDNode *kernel_node = nullptr;
-      for (unsigned i = 0; i < kernels_node->getNumOperands(); ++i) {
-         auto func = llvm::mdconst::dyn_extract<llvm::Function>(
-               kernels_node->getOperand(i)->getOperand(0));
-         if (func == kernel_func) {
-            kernel_node = kernels_node->getOperand(i);
-            break;
-         }
-      }
-
-      return kernel_node;
-   }
-
-   llvm::MDNode*
-   node_from_op_checked(const llvm::MDOperand &md_operand,
-                        llvm::StringRef expect_name,
-                        unsigned expect_num_args)
-   {
-      auto node = llvm::cast<llvm::MDNode>(md_operand);
-      assert(node->getNumOperands() == expect_num_args &&
-             "Wrong number of operands.");
-
-      auto str_node = llvm::cast<llvm::MDString>(node->getOperand(0));
-      assert(str_node->getString() == expect_name &&
-             "Wrong metadata node name.");
-
-      return node;
-   }
-
-   std::vector<kernel_arg_md>
-   get_kernel_arg_md(const llvm::Function *kernel_func) {
-      auto num_args = kernel_func->getArgumentList().size();
-
-      auto kernel_node = get_kernel_metadata(kernel_func);
-      auto aq = node_from_op_checked(kernel_node->getOperand(2),
-                                     "kernel_arg_access_qual", num_args + 1);
-      auto ty = node_from_op_checked(kernel_node->getOperand(3),
-                                     "kernel_arg_type", num_args + 1);
-
-      std::vector<kernel_arg_md> res;
-      res.reserve(num_args);
-      for (unsigned i = 0; i < num_args; ++i) {
-         res.push_back(kernel_arg_md(
-            llvm::cast<llvm::MDString>(ty->getOperand(i+1))->getString(),
-            llvm::cast<llvm::MDString>(aq->getOperand(i+1))->getString()));
+      if (internalize_symbols) {
+         std::vector<std::string> names =
+            map(std::mem_fn(&Function::getName), get_kernels(mod));
+         pm.add(::llvm::createInternalizePass(
+                      [=](const ::llvm::GlobalValue &gv) {
+                         return std::find(names.begin(), names.end(),
+                                          gv.getName()) != names.end();
+                      }));
       }
 
-      return res;
+      ::llvm::PassManagerBuilder pmb;
+      pmb.OptLevel = optimization_level;
+      pmb.LibraryInfo = new ::llvm::TargetLibraryInfoImpl(
+         ::llvm::Triple(mod.getTargetTriple()));
+      pmb.populateModulePassManager(pm);
+      pm.run(mod);
    }
-#endif
-
-   std::vector<module::argument>
-   get_kernel_args(const llvm::Module *mod, const std::string &kernel_name,
-                   const clang::CompilerInstance &c) {
-      std::vector<module::argument> args;
-      const auto address_spaces = c.getTarget().getAddressSpaceMap();
-      llvm::Function *kernel_func = mod->getFunction(kernel_name);
-      assert(kernel_func && "Kernel name not found in module.");
-      auto arg_md = get_kernel_arg_md(kernel_func);
-
-      llvm::DataLayout TD(mod);
-      llvm::Type *size_type =
-         TD.getSmallestLegalIntType(mod->getContext(), sizeof(cl_uint) * 8);
-
-      for (const auto &arg: kernel_func->args()) {
-
-         llvm::Type *arg_type = arg.getType();
-         const unsigned arg_store_size = TD.getTypeStoreSize(arg_type);
-
-         // OpenCL 1.2 specification, Ch. 6.1.5: "A built-in data
-         // type that is not a power of two bytes in size must be
-         // aligned to the next larger power of two".  We need this
-         // alignment for three element vectors, which have
-         // non-power-of-2 store size.
-         const unsigned arg_api_size = util_next_power_of_two(arg_store_size);
-
-         llvm::Type *target_type = arg_type->isIntegerTy() ?
-               TD.getSmallestLegalIntType(mod->getContext(), arg_store_size * 8)
-               : arg_type;
-         unsigned target_size = TD.getTypeStoreSize(target_type);
-         unsigned target_align = TD.getABITypeAlignment(target_type);
-
-         llvm::StringRef type_name = arg_md[arg.getArgNo()].type_name;
-         llvm::StringRef access_qual = arg_md[arg.getArgNo()].access_qual;
-
-         // Image
-         const bool is_image2d = type_name == "image2d_t";
-         const bool is_image3d = type_name == "image3d_t";
-         if (is_image2d || is_image3d) {
-            const bool is_write_only = access_qual == "write_only";
-            const bool is_read_only = access_qual == "read_only";
-
-            enum module::argument::type marg_type;
-            if (is_image2d && is_read_only) {
-               marg_type = module::argument::image2d_rd;
-            } else if (is_image2d && is_write_only) {
-               marg_type = module::argument::image2d_wr;
-            } else if (is_image3d && is_read_only) {
-               marg_type = module::argument::image3d_rd;
-            } else if (is_image3d && is_write_only) {
-               marg_type = module::argument::image3d_wr;
-            } else {
-               assert(0 && "Wrong image access qualifier");
-            }
-
-            args.push_back(module::argument(marg_type,
-                                            arg_store_size, target_size,
-                                            target_align,
-                                            module::argument::zero_ext));
-            continue;
-         }
-
-         // Image size implicit argument
-         if (type_name == "__llvm_image_size") {
-            args.push_back(module::argument(module::argument::scalar,
-                                            sizeof(cl_uint),
-                                            TD.getTypeStoreSize(size_type),
-                                            TD.getABITypeAlignment(size_type),
-                                            module::argument::zero_ext,
-                                            module::argument::image_size));
-            continue;
-         }
-
-         // Image format implicit argument
-         if (type_name == "__llvm_image_format") {
-            args.push_back(module::argument(module::argument::scalar,
-                                            sizeof(cl_uint),
-                                            TD.getTypeStoreSize(size_type),
-                                            TD.getABITypeAlignment(size_type),
-                                            module::argument::zero_ext,
-                                            module::argument::image_format));
-            continue;
-         }
-
-         // Other types
-         if (llvm::isa<llvm::PointerType>(arg_type) && arg.hasByValAttr()) {
-            arg_type =
-                  llvm::dyn_cast<llvm::PointerType>(arg_type)->getElementType();
-         }
-
-         if (arg_type->isPointerTy()) {
-            unsigned address_space = llvm::cast<llvm::PointerType>(arg_type)->getAddressSpace();
-            if (address_space == address_spaces[clang::LangAS::opencl_local
-                                                     - clang::LangAS::Offset]) {
-               args.push_back(module::argument(module::argument::local,
-                                               arg_api_size, target_size,
-                                               target_align,
-                                               module::argument::zero_ext));
-            } else {
-               // XXX: Correctly handle constant address space.  There is no
-               // way for r600g to pass a handle for constant buffers back
-               // to clover like it can for global buffers, so
-               // creating constant arguments will break r600g.  For now,
-               // continue treating constant buffers as global buffers
-               // until we can come up with a way to create handles for
-               // constant buffers.
-               args.push_back(module::argument(module::argument::global,
-                                               arg_api_size, target_size,
-                                               target_align,
-                                               module::argument::zero_ext));
-           }
-
-         } else {
-            llvm::AttributeSet attrs = kernel_func->getAttributes();
-            enum module::argument::ext_type ext_type =
-                  (attrs.hasAttribute(arg.getArgNo() + 1,
-                                     llvm::Attribute::SExt) ?
-                   module::argument::sign_ext :
-                   module::argument::zero_ext);
-
-            args.push_back(
-               module::argument(module::argument::scalar, arg_api_size,
-                                target_size, target_align, ext_type));
-         }
-      }
-
-      // Append implicit arguments.  XXX - The types, ordering and
-      // vector size of the implicit arguments should depend on the
-      // target according to the selected calling convention.
-      args.push_back(
-         module::argument(module::argument::scalar, sizeof(cl_uint),
-                          TD.getTypeStoreSize(size_type),
-                          TD.getABITypeAlignment(size_type),
-                          module::argument::zero_ext,
-                          module::argument::grid_dimension));
-
-      args.push_back(
-         module::argument(module::argument::scalar, sizeof(cl_uint),
-                          TD.getTypeStoreSize(size_type),
-                          TD.getABITypeAlignment(size_type),
-                          module::argument::zero_ext,
-                          module::argument::grid_offset));
-
-      return args;
-   }
-
-   module
-   build_module_llvm(llvm::Module *mod,
-                     const clang::CompilerInstance &c) {
-
-      module m;
-      struct pipe_llvm_program_header header;
 
-      llvm::SmallVector<char, 1024> llvm_bitcode;
-      llvm::raw_svector_ostream bitcode_ostream(llvm_bitcode);
-      llvm::BitstreamWriter writer(llvm_bitcode);
-      llvm::WriteBitcodeToFile(mod, bitcode_ostream);
-#if HAVE_LLVM < 0x0308
-      bitcode_ostream.flush();
-#endif
-
-      const std::vector<llvm::Function *> kernels = find_kernels(mod);
-      for (unsigned i = 0; i < kernels.size(); ++i) {
-         std::string kernel_name = kernels[i]->getName();
-         std::vector<module::argument> args =
-               get_kernel_args(mod, kernel_name, c);
-
-         m.syms.push_back(module::symbol(kernel_name, 0, i, args ));
-      }
-
-      header.num_bytes = llvm_bitcode.size();
-      std::vector<char> data;
-      data.insert(data.end(), (char*)(&header),
-                              (char*)(&header) + sizeof(header));
-      data.insert(data.end(), llvm_bitcode.begin(),
-                                  llvm_bitcode.end());
-      m.secs.push_back(module::section(0, module::section::text,
-                                       header.num_bytes, data));
+   std::unique_ptr<Module>
+   link(LLVMContext &ctx, const clang::CompilerInstance &c,
+        const std::vector<module> &modules, std::string &r_log) {
+      std::unique_ptr<Module> mod { new Module("link", ctx) };
+      std::unique_ptr< ::llvm::Linker> linker { new ::llvm::Linker(*mod) };
 
-      return m;
-   }
-
-   void
-   emit_code(LLVMTargetMachineRef tm, LLVMModuleRef mod,
-             LLVMCodeGenFileType file_type,
-             LLVMMemoryBufferRef *out_buffer,
-             std::string &r_log) {
-      LLVMBool err;
-      char *err_message = NULL;
-
-      err = LLVMTargetMachineEmitToMemoryBuffer(tm, mod, file_type,
-                                                &err_message, out_buffer);
-
-      if (err) {
-         r_log = std::string(err_message);
+      for (auto &m : modules) {
+         if (linker->linkInModule(parse_module_library(m, ctx, r_log)))
+            throw build_error();
       }
 
-      LLVMDisposeMessage(err_message);
-
-      if (err) {
-         throw compile_error();
-      }
+      return mod;
    }
+}
 
-   std::vector<char>
-   compile_native(const llvm::Module *mod, const target &t, bool dump_asm,
-                  std::string &r_log) {
-
-      std::string log;
-      LLVMTargetRef target;
-      char *error_message;
-      LLVMMemoryBufferRef out_buffer;
-      unsigned buffer_size;
-      const char *buffer_data;
-      LLVMModuleRef mod_ref = wrap(mod);
-
-      if (LLVMGetTargetFromTriple(t.triple.c_str(), &target, &error_message)) {
-         r_log = std::string(error_message);
-         LLVMDisposeMessage(error_message);
-         throw compile_error();
-      }
+module
+clover::llvm::link_program(const std::vector<module> &modules,
+                           const device &dev, const std::string &opts,
+                           std::string &r_log) {
+   std::vector<std::string> options = tokenize(opts + " input.cl");
+   const bool create_library = count("-create-library", options);
+   erase_if(equals("-create-library"), options);
 
-      LLVMTargetMachineRef tm = LLVMCreateTargetMachine(
-            target, t.triple.c_str(), t.cpu.c_str(), "",
-            LLVMCodeGenLevelDefault, LLVMRelocDefault, LLVMCodeModelDefault);
+   auto ctx = create_context(r_log);
+   auto c = create_compiler_instance(dev, dev.ir_target(), options, r_log);
+   auto mod = link(*ctx, *c, modules, r_log);
 
-      if (!tm) {
-         r_log = "Could not create TargetMachine: " + t.triple;
-         throw compile_error();
-      }
+   optimize(*mod, c->getCodeGenOpts().OptimizationLevel, !create_library);
 
-      if (dump_asm) {
-         LLVMSetTargetMachineAsmVerbosity(tm, true);
-#if HAVE_LLVM >= 0x0308
-         LLVMModuleRef debug_mod = wrap(llvm::CloneModule(mod).release());
-#else
-         LLVMModuleRef debug_mod = wrap(llvm::CloneModule(mod));
-#endif
-         emit_code(tm, debug_mod, LLVMAssemblyFile, &out_buffer, r_log);
-         buffer_size = LLVMGetBufferSize(out_buffer);
-         buffer_data = LLVMGetBufferStart(out_buffer);
-         debug_log(std::string(buffer_data, buffer_size), ".asm");
-
-         LLVMSetTargetMachineAsmVerbosity(tm, false);
-         LLVMDisposeMemoryBuffer(out_buffer);
-         LLVMDisposeModule(debug_mod);
-      }
+   static std::atomic_uint seq(0);
+   const std::string id = "." + mod->getModuleIdentifier() + "-" +
+      std::to_string(seq++);
 
-      emit_code(tm, mod_ref, LLVMObjectFile, &out_buffer, r_log);
+   if (has_flag(debug::llvm))
+      debug::log(id + ".ll", print_module_bitcode(*mod));
 
-      buffer_size = LLVMGetBufferSize(out_buffer);
-      buffer_data = LLVMGetBufferStart(out_buffer);
+   if (create_library) {
+      return build_module_library(*mod, module::section::text_library);
 
-      std::vector<char> code(buffer_data, buffer_data + buffer_size);
+   } else if (dev.ir_format() == PIPE_SHADER_IR_NATIVE) {
+      if (has_flag(debug::native))
+         debug::log(id +  ".asm", print_module_native(*mod, dev.ir_target()));
 
-      LLVMDisposeMemoryBuffer(out_buffer);
-      LLVMDisposeTargetMachine(tm);
+      return build_module_native(*mod, dev.ir_target(), *c, r_log);
 
-      return code;
+   } else {
+      unreachable("Unsupported IR.");
    }
+}
 
-   std::map<std::string, unsigned>
-   get_kernel_offsets(std::vector<char> &code,
-                      const std::vector<llvm::Function *> &kernels,
-                      std::string &r_log) {
-
-      // One of the libelf implementations
-      // (http://www.mr511.de/software/english.htm) requires calling
-      // elf_version() before elf_memory().
-      //
-      elf_version(EV_CURRENT);
-
-      Elf *elf = elf_memory(&code[0], code.size());
-      size_t section_str_index;
-      elf_getshdrstrndx(elf, &section_str_index);
-      Elf_Scn *section = NULL;
-      Elf_Scn *symtab = NULL;
-      GElf_Shdr symtab_header;
-
-      // Find the symbol table
-      try {
-         while ((section = elf_nextscn(elf, section))) {
-            const char *name;
-            if (gelf_getshdr(section, &symtab_header) != &symtab_header) {
-               r_log = "Failed to read ELF section header.";
-               throw compile_error();
-            }
-            name = elf_strptr(elf, section_str_index, symtab_header.sh_name);
-           if (!strcmp(name, ".symtab")) {
-               symtab = section;
-               break;
-           }
-         }
-         if (!symtab) {
-            r_log = "Unable to find symbol table.";
-            throw compile_error();
-         }
-      } catch (compile_error &e) {
-         elf_end(elf);
-         throw e;
-      }
-
-
-      // Extract symbol information from the table
-      Elf_Data *symtab_data = NULL;
-      GElf_Sym *symbol;
-      GElf_Sym s;
-
-      std::map<std::string, unsigned> kernel_offsets;
-      symtab_data = elf_getdata(symtab, symtab_data);
+#ifdef HAVE_CLOVER_SPIRV
+module
+clover::llvm::compile_to_spirv(const std::string &source,
+                               const header_map &headers,
+                               const device &dev,
+                               const std::string &opts,
+                               std::string &r_log) {
+   if (has_flag(debug::clc))
+      debug::log(".cl", "// Options: " + opts + '\n' + source);
 
-      // Determine the offsets for each kernel
-      for (int i = 0; (symbol = gelf_getsym(symtab_data, i, &s)); i++) {
-         char *name = elf_strptr(elf, symtab_header.sh_link, symbol->st_name);
-         for (std::vector<llvm::Function*>::const_iterator it = kernels.begin(),
-              e = kernels.end(); it != e; ++it) {
-            llvm::Function *f = *it;
-            if (f->getName() == std::string(name))
-               kernel_offsets[f->getName()] = symbol->st_value;
-         }
-      }
-      elf_end(elf);
-      return kernel_offsets;
+   auto ctx = create_context(r_log);
+   const std::string target = dev.address_bits() == 32u ?
+      "-spir-unknown-unknown" :
+      "-spir64-unknown-unknown";
+   auto c = create_compiler_instance(dev, target,
+                                     tokenize(opts + " input.cl"), r_log);
+   auto mod = compile(*ctx, *c, "input.cl", source, headers, dev, opts, false,
+                      r_log);
+
+   if (has_flag(debug::llvm))
+      debug::log(".ll", print_module_bitcode(*mod));
+
+   std::string error_msg;
+   if (!::llvm::regularizeLlvmForSpirv(mod.get(), error_msg)) {
+      r_log += "Failed to regularize LLVM IR for SPIR-V: " + error_msg + ".\n";
+      throw error(CL_INVALID_VALUE);
    }
 
-   module
-   build_module_native(std::vector<char> &code,
-                       llvm::Module *mod,
-                       const clang::CompilerInstance &c,
-                       std::string &r_log) {
-
-      const std::vector<llvm::Function *> kernels = find_kernels(mod);
-
-      std::map<std::string, unsigned> kernel_offsets =
-            get_kernel_offsets(code, kernels, r_log);
-
-      // Begin building the clover module
-      module m;
-      struct pipe_llvm_program_header header;
-
-      // Store the generated ELF binary in the module's text section.
-      header.num_bytes = code.size();
-      std::vector<char> data;
-      data.insert(data.end(), (char*)(&header),
-                              (char*)(&header) + sizeof(header));
-      data.insert(data.end(), code.begin(), code.end());
-      m.secs.push_back(module::section(0, module::section::text,
-                                       header.num_bytes, data));
-
-      for (std::map<std::string, unsigned>::iterator i = kernel_offsets.begin(),
-           e = kernel_offsets.end(); i != e; ++i) {
-         std::vector<module::argument> args = get_kernel_args(mod, i->first, c);
-         m.syms.push_back(module::symbol(i->first, 0, i->second, args ));
-      }
-
-      return m;
+   std::ostringstream os;
+   if (!::llvm::writeSpirv(mod.get(), os, error_msg)) {
+      r_log += "Translation from LLVM IR to SPIR-V failed: " + error_msg + ".\n";
+      throw error(CL_INVALID_VALUE);
    }
 
-#define DBG_CLC  (1 << 0)
-#define DBG_LLVM (1 << 1)
-#define DBG_ASM  (1 << 2)
-
-   unsigned
-   get_debug_flags() {
-      static const struct debug_named_value debug_options[] = {
-         {"clc", DBG_CLC, "Dump the OpenCL C code for all kernels."},
-         {"llvm", DBG_LLVM, "Dump the generated LLVM IR for all kernels."},
-         {"asm", DBG_ASM, "Dump kernel assembly code for targets specifying "
-          "PIPE_SHADER_IR_NATIVE"},
-         DEBUG_NAMED_VALUE_END // must be last
-      };
-      static const unsigned debug_flags =
-         debug_get_flags_option("CLOVER_DEBUG", debug_options, 0);
-
-      return debug_flags;
+   const std::string osContent = os.str();
+   std::vector<char> binary(osContent.begin(), osContent.end());
+   if (binary.empty()) {
+      r_log += "Failed to retrieve SPIR-V binary.\n";
+      throw error(CL_INVALID_VALUE);
    }
 
-} // End anonymous namespace
-
-module
-clover::compile_program_llvm(const std::string &source,
-                             const header_map &headers,
-                             enum pipe_shader_ir ir,
-                             const std::string &target,
-                             const std::string &opts,
-                             std::string &r_log) {
-   if (get_debug_flags() & DBG_CLC)
-      debug_log("// Build options: " + opts + '\n' + source, ".cl");
+   if (has_flag(debug::spirv))
+      debug::log(".spvasm", spirv::print_module(binary, dev.device_version()));
 
-   auto ctx = create_context(r_log);
-   // The input file name must have the .cl extension in order for the
-   // CompilerInvocation class to recognize it as an OpenCL source file.
-   const auto c = create_compiler_instance(target, tokenize(opts + " input.cl"),
-                                           r_log);
-   Module *mod = compile_llvm(*ctx, *c, source, headers, "input.cl",
-                              target, opts, r_log);
-
-   optimize(mod, c->getCodeGenOpts().OptimizationLevel);
-
-   if (get_debug_flags() & DBG_LLVM) {
-      std::string log;
-      raw_string_ostream s_log(log);
-      mod->print(s_log, NULL);
-      s_log.flush();
-      debug_log(log, ".ll");
-    }
-
-   module m;
-   // Build the clover::module
-   switch (ir) {
-      case PIPE_SHADER_IR_NIR:
-      case PIPE_SHADER_IR_TGSI:
-         //XXX: Handle TGSI, NIR
-         assert(0);
-         m = module();
-         break;
-      case PIPE_SHADER_IR_LLVM:
-         m = build_module_llvm(mod, *c);
-         break;
-      case PIPE_SHADER_IR_NATIVE: {
-         std::vector<char> code = compile_native(mod, target,
-                                                 get_debug_flags() & DBG_ASM,
-                                                 r_log);
-         m = build_module_native(code, mod, *c, r_log);
-         break;
-      }
-   }
-   // The user takes ownership of the module.
-   delete mod;
-   return m;
+   return spirv::compile_program(binary, dev, r_log);
 }
+#endif