Merge pull request #7598 from FireFox317/more-llvm-stage2

stage2: More improvements to self-hosted LLVM backend

build.zig

@@ -91,6 +91,7 @@ pub fn build(b: *Builder) !void {
     exe.addBuildOption(bool, "have_llvm", enable_llvm);
     if (enable_llvm) {
         const cmake_cfg = if (static_llvm) null else findAndParseConfigH(b, config_h_path_option);
+
         if (is_stage1) {
             exe.addIncludeDir("src");
             exe.addIncludeDir("deps/SoftFloat-3e/source/include");
@@ -109,28 +110,8 @@ pub fn build(b: *Builder) !void {
             softfloat.addCSourceFiles(&softfloat_sources, &[_][]const u8{ "-std=c99", "-O3" });
             exe.linkLibrary(softfloat);
 
-            const exe_cflags = [_][]const u8{
-                "-std=c++14",
-                "-D__STDC_CONSTANT_MACROS",
-                "-D__STDC_FORMAT_MACROS",
-                "-D__STDC_LIMIT_MACROS",
-                "-D_GNU_SOURCE",
-                "-fvisibility-inlines-hidden",
-                "-fno-exceptions",
-                "-fno-rtti",
-                "-Werror=type-limits",
-                "-Wno-missing-braces",
-                "-Wno-comment",
-            };
             exe.addCSourceFiles(&stage1_sources, &exe_cflags);
             exe.addCSourceFiles(&optimized_c_sources, &[_][]const u8{ "-std=c99", "-O3" });
-            if (cmake_cfg == null) {
-                // We need this because otherwise zig_clang_cc1_main.cpp ends up pulling
-                // in a dependency on llvm::cfg::Update<llvm::BasicBlock*>::dump() which is
-                // unavailable when LLVM is compiled in Release mode.
-                const zig_cpp_cflags = exe_cflags ++ [_][]const u8{"-DNDEBUG=1"};
-                exe.addCSourceFiles(&zig_cpp_sources, &zig_cpp_cflags);
-            }
         }
         if (cmake_cfg) |cfg| {
             // Inside this code path, we have to coordinate with system packaged LLVM, Clang, and LLD.
@@ -139,79 +120,14 @@ pub fn build(b: *Builder) !void {
             if (cfg.cmake_prefix_path.len > 0) {
                 b.addSearchPrefix(cfg.cmake_prefix_path);
             }
-            exe.addObjectFile(fs.path.join(b.allocator, &[_][]const u8{
-                cfg.cmake_binary_dir,
-                "zigcpp",
-                b.fmt("{s}{s}{s}", .{ exe.target.libPrefix(), "zigcpp", exe.target.staticLibSuffix() }),
-            }) catch unreachable);
-            assert(cfg.lld_include_dir.len != 0);
-            exe.addIncludeDir(cfg.lld_include_dir);
-            addCMakeLibraryList(exe, cfg.clang_libraries);
-            addCMakeLibraryList(exe, cfg.lld_libraries);
-            addCMakeLibraryList(exe, cfg.llvm_libraries);
 
-            const need_cpp_includes = tracy != null;
-
-            // System -lc++ must be used because in this code path we are attempting to link
-            // against system-provided LLVM, Clang, LLD.
-            if (exe.target.getOsTag() == .linux) {
-                // First we try to static link against gcc libstdc++. If that doesn't work,
-                // we fall back to -lc++ and cross our fingers.
-                addCxxKnownPath(b, cfg, exe, "libstdc++.a", "", need_cpp_includes) catch |err| switch (err) {
-                    error.RequiredLibraryNotFound => {
-                        exe.linkSystemLibrary("c++");
-                    },
-                    else => |e| return e,
-                };
-
-                exe.linkSystemLibrary("pthread");
-            } else if (exe.target.isFreeBSD()) {
-                try addCxxKnownPath(b, cfg, exe, "libc++.a", null, need_cpp_includes);
-                exe.linkSystemLibrary("pthread");
-            } else if (exe.target.getOsTag() == .openbsd) {
-                try addCxxKnownPath(b, cfg, exe, "libc++.a", null, need_cpp_includes);
-                try addCxxKnownPath(b, cfg, exe, "libc++abi.a", null, need_cpp_includes);
-            } else if (exe.target.isDarwin()) {
-                if (addCxxKnownPath(b, cfg, exe, "libgcc_eh.a", "", need_cpp_includes)) {
-                    // Compiler is GCC.
-                    try addCxxKnownPath(b, cfg, exe, "libstdc++.a", null, need_cpp_includes);
-                    exe.linkSystemLibrary("pthread");
-                    // TODO LLD cannot perform this link.
-                    // Set ZIG_SYSTEM_LINKER_HACK env var to use system linker ld instead.
-                    // See https://github.com/ziglang/zig/issues/1535
-                } else |err| switch (err) {
-                    error.RequiredLibraryNotFound => {
-                        // System compiler, not gcc.
-                        exe.linkSystemLibrary("c++");
-                    },
-                    else => |e| return e,
-                }
-            }
-
-            if (cfg.dia_guids_lib.len != 0) {
-                exe.addObjectFile(cfg.dia_guids_lib);
-            }
+            try addCmakeCfgOptionsToExe(b, cfg, tracy, exe);
+            try addCmakeCfgOptionsToExe(b, cfg, tracy, test_stage2);
         } else {
             // Here we are -Denable-llvm but no cmake integration.
-            for (clang_libs) |lib_name| {
-                exe.linkSystemLibrary(lib_name);
-            }
 
-            for (lld_libs) |lib_name| {
-                exe.linkSystemLibrary(lib_name);
-            }
-
-            for (llvm_libs) |lib_name| {
-                exe.linkSystemLibrary(lib_name);
-            }
-
-            // This means we rely on clang-or-zig-built LLVM, Clang, LLD libraries.
-            exe.linkSystemLibrary("c++");
-
-            if (target.getOs().tag == .windows) {
-                exe.linkSystemLibrary("version");
-                exe.linkSystemLibrary("uuid");
-            }
+            try addStaticLlvmOptionsToExe(exe);
+            try addStaticLlvmOptionsToExe(test_stage2);
         }
     }
     if (link_libc) {
@@ -357,6 +273,112 @@ pub fn build(b: *Builder) !void {
     test_step.dependOn(docs_step);
 }
 
+const exe_cflags = [_][]const u8{
+    "-std=c++14",
+    "-D__STDC_CONSTANT_MACROS",
+    "-D__STDC_FORMAT_MACROS",
+    "-D__STDC_LIMIT_MACROS",
+    "-D_GNU_SOURCE",
+    "-fvisibility-inlines-hidden",
+    "-fno-exceptions",
+    "-fno-rtti",
+    "-Werror=type-limits",
+    "-Wno-missing-braces",
+    "-Wno-comment",
+};
+
+fn addCmakeCfgOptionsToExe(
+    b: *Builder,
+    cfg: CMakeConfig,
+    tracy: ?[]const u8,
+    exe: *std.build.LibExeObjStep,
+) !void {
+    exe.addObjectFile(fs.path.join(b.allocator, &[_][]const u8{
+        cfg.cmake_binary_dir,
+        "zigcpp",
+        b.fmt("{s}{s}{s}", .{ exe.target.libPrefix(), "zigcpp", exe.target.staticLibSuffix() }),
+    }) catch unreachable);
+    assert(cfg.lld_include_dir.len != 0);
+    exe.addIncludeDir(cfg.lld_include_dir);
+    addCMakeLibraryList(exe, cfg.clang_libraries);
+    addCMakeLibraryList(exe, cfg.lld_libraries);
+    addCMakeLibraryList(exe, cfg.llvm_libraries);
+
+    const need_cpp_includes = tracy != null;
+
+    // System -lc++ must be used because in this code path we are attempting to link
+    // against system-provided LLVM, Clang, LLD.
+    if (exe.target.getOsTag() == .linux) {
+        // First we try to static link against gcc libstdc++. If that doesn't work,
+        // we fall back to -lc++ and cross our fingers.
+        addCxxKnownPath(b, cfg, exe, "libstdc++.a", "", need_cpp_includes) catch |err| switch (err) {
+            error.RequiredLibraryNotFound => {
+                exe.linkSystemLibrary("c++");
+            },
+            else => |e| return e,
+        };
+
+        exe.linkSystemLibrary("pthread");
+    } else if (exe.target.isFreeBSD()) {
+        try addCxxKnownPath(b, cfg, exe, "libc++.a", null, need_cpp_includes);
+        exe.linkSystemLibrary("pthread");
+    } else if (exe.target.getOsTag() == .openbsd) {
+        try addCxxKnownPath(b, cfg, exe, "libc++.a", null, need_cpp_includes);
+        try addCxxKnownPath(b, cfg, exe, "libc++abi.a", null, need_cpp_includes);
+    } else if (exe.target.isDarwin()) {
+        if (addCxxKnownPath(b, cfg, exe, "libgcc_eh.a", "", need_cpp_includes)) {
+            // Compiler is GCC.
+            try addCxxKnownPath(b, cfg, exe, "libstdc++.a", null, need_cpp_includes);
+            exe.linkSystemLibrary("pthread");
+            // TODO LLD cannot perform this link.
+            // Set ZIG_SYSTEM_LINKER_HACK env var to use system linker ld instead.
+            // See https://github.com/ziglang/zig/issues/1535
+        } else |err| switch (err) {
+            error.RequiredLibraryNotFound => {
+                // System compiler, not gcc.
+                exe.linkSystemLibrary("c++");
+            },
+            else => |e| return e,
+        }
+    }
+
+    if (cfg.dia_guids_lib.len != 0) {
+        exe.addObjectFile(cfg.dia_guids_lib);
+    }
+}
+
+fn addStaticLlvmOptionsToExe(
+    exe: *std.build.LibExeObjStep,
+) !void {
+    // Adds the Zig C++ sources which both stage1 and stage2 need.
+    //
+    // We need this because otherwise zig_clang_cc1_main.cpp ends up pulling
+    // in a dependency on llvm::cfg::Update<llvm::BasicBlock*>::dump() which is
+    // unavailable when LLVM is compiled in Release mode.
+    const zig_cpp_cflags = exe_cflags ++ [_][]const u8{"-DNDEBUG=1"};
+    exe.addCSourceFiles(&zig_cpp_sources, &zig_cpp_cflags);
+
+    for (clang_libs) |lib_name| {
+        exe.linkSystemLibrary(lib_name);
+    }
+
+    for (lld_libs) |lib_name| {
+        exe.linkSystemLibrary(lib_name);
+    }
+
+    for (llvm_libs) |lib_name| {
+        exe.linkSystemLibrary(lib_name);
+    }
+
+    // This means we rely on clang-or-zig-built LLVM, Clang, LLD libraries.
+    exe.linkSystemLibrary("c++");
+
+    if (exe.target.getOs().tag == .windows) {
+        exe.linkSystemLibrary("version");
+        exe.linkSystemLibrary("uuid");
+    }
+}
+
 fn addCxxKnownPath(
     b: *Builder,
     ctx: CMakeConfig,

src/Cache.zig

@@ -16,7 +16,8 @@ const Allocator = std.mem.Allocator;
 /// This protection is conditionally compiled depending on `want_debug_deadlock`.
 var all_cache_digest_set: std.AutoHashMapUnmanaged(BinDigest, void) = .{};
 var all_cache_digest_lock: std.Mutex = .{};
-const want_debug_deadlock = std.debug.runtime_safety;
+// TODO: Figure out how to make sure that `all_cache_digest_set` does not leak memory!
+pub const want_debug_deadlock = false;
 const DebugBinDigest = if (want_debug_deadlock) BinDigest else void;
 const null_debug_bin_digest = if (want_debug_deadlock) ([1]u8{0} ** bin_digest_len) else {};
 

src/Compilation.zig

@@ -1173,6 +1173,7 @@ pub fn destroy(self: *Compilation) void {
 
     const gpa = self.gpa;
     self.work_queue.deinit();
+    self.c_object_work_queue.deinit();
 
     {
         var it = self.crt_files.iterator();
@@ -1202,6 +1203,10 @@ pub fn destroy(self: *Compilation) void {
         crt_file.deinit(gpa);
     }
 
+    if (self.glibc_so_files) |*glibc_file| {
+        glibc_file.deinit(gpa);
+    }
+
     for (self.c_object_table.items()) |entry| {
         entry.key.destroy(gpa);
     }

src/DepTokenizer.zig

@@ -927,7 +927,7 @@ fn depTokenizer(input: []const u8, expect: []const u8) !void {
     try out.writeAll("\n");
     try printSection(out, "<<<< input", input);
     try printSection(out, "==== expect", expect);
-    try printSection(out, ">>>> got", got);
+    try printSection(out, ">>>> got", buffer.items);
     try printRuler(out);
 
     testing.expect(false);

src/link/Coff.zig

@@ -811,8 +811,11 @@ fn linkWithLLD(self: *Coff, comp: *Compilation) !void {
     // If there is no Zig code to compile, then we should skip flushing the output file because it
     // will not be part of the linker line anyway.
     const module_obj_path: ?[]const u8 = if (self.base.options.module) |module| blk: {
-        const use_stage1 = build_options.is_stage1 and self.base.options.use_llvm;
-        if (use_stage1) {
+        // Both stage1 and stage2 LLVM backend put the object file in the cache directory.
+        if (self.base.options.use_llvm) {
+            // Stage2 has to call flushModule since that outputs the LLVM object file.
+            if (!build_options.is_stage1) try self.flushModule(comp);
+
             const obj_basename = try std.zig.binNameAlloc(arena, .{
                 .root_name = self.base.options.root_name,
                 .target = self.base.options.target,

src/link/Elf.zig

@@ -1251,8 +1251,11 @@ fn linkWithLLD(self: *Elf, comp: *Compilation) !void {
     // If there is no Zig code to compile, then we should skip flushing the output file because it
     // will not be part of the linker line anyway.
     const module_obj_path: ?[]const u8 = if (self.base.options.module) |module| blk: {
-        const use_stage1 = build_options.is_stage1 and self.base.options.use_llvm;
-        if (use_stage1) {
+        // Both stage1 and stage2 LLVM backend put the object file in the cache directory.
+        if (self.base.options.use_llvm) {
+            // Stage2 has to call flushModule since that outputs the LLVM object file.
+            if (!build_options.is_stage1) try self.flushModule(comp);
+
             const obj_basename = try std.zig.binNameAlloc(arena, .{
                 .root_name = self.base.options.root_name,
                 .target = self.base.options.target,

src/llvm_backend.zig

@@ -1,4 +1,5 @@
 const std = @import("std");
+const assert = std.debug.assert;
 const Allocator = std.mem.Allocator;
 const Compilation = @import("Compilation.zig");
 const llvm = @import("llvm_bindings.zig");
@@ -141,17 +142,36 @@ pub const LLVMIRModule = struct {
     target_machine: *const llvm.TargetMachineRef,
     builder: *const llvm.BuilderRef,
 
-    output_path: []const u8,
+    object_path: []const u8,
 
     gpa: *Allocator,
     err_msg: ?*Compilation.ErrorMsg = null,
 
+    /// This stores the LLVM values used in a function, such that they can be
+    /// referred to in other instructions. This table is cleared before every function is generated.
+    func_inst_table: std.AutoHashMapUnmanaged(*Inst, *const llvm.ValueRef) = .{},
+
+    /// These fields are used to refer to the LLVM value of the function paramaters in an Arg instruction.
+    args: []*const llvm.ValueRef = &[_]*const llvm.ValueRef{},
+    arg_index: usize = 0,
+
     pub fn create(allocator: *Allocator, sub_path: []const u8, options: link.Options) !*LLVMIRModule {
         const self = try allocator.create(LLVMIRModule);
         errdefer allocator.destroy(self);
 
         const gpa = options.module.?.gpa;
 
+        const obj_basename = try std.zig.binNameAlloc(gpa, .{
+            .root_name = options.root_name,
+            .target = options.target,
+            .output_mode = .Obj,
+        });
+        defer gpa.free(obj_basename);
+
+        const o_directory = options.module.?.zig_cache_artifact_directory;
+        const object_path = try o_directory.join(gpa, &[_][]const u8{obj_basename});
+        errdefer gpa.free(object_path);
+
         initializeLLVMTargets();
 
         const root_nameZ = try gpa.dupeZ(u8, options.root_name);
@@ -203,7 +223,7 @@ pub const LLVMIRModule = struct {
             .llvm_module = llvm_module,
             .target_machine = target_machine,
             .builder = builder,
-            .output_path = sub_path,
+            .object_path = object_path,
             .gpa = gpa,
         };
         return self;
@@ -213,6 +233,10 @@ pub const LLVMIRModule = struct {
         self.builder.disposeBuilder();
         self.target_machine.disposeTargetMachine();
         self.llvm_module.disposeModule();
+
+        self.func_inst_table.deinit(self.gpa);
+        self.gpa.free(self.object_path);
+
         allocator.destroy(self);
     }
 
@@ -245,15 +269,13 @@ pub const LLVMIRModule = struct {
             }
         }
 
-        const output_pathZ = try self.gpa.dupeZ(u8, self.output_path);
-        defer self.gpa.free(output_pathZ);
+        const object_pathZ = try self.gpa.dupeZ(u8, self.object_path);
+        defer self.gpa.free(object_pathZ);
 
         var error_message: [*:0]const u8 = undefined;
-        // TODO: where to put the output object, zig-cache something?
-        // TODO: caching?
         if (self.target_machine.emitToFile(
             self.llvm_module,
-            output_pathZ.ptr,
+            object_pathZ.ptr,
             .ObjectFile,
             &error_message,
         )) {
@@ -271,6 +293,7 @@ pub const LLVMIRModule = struct {
             error.CodegenFail => {
                 decl.analysis = .codegen_failure;
                 try module.failed_decls.put(module.gpa, decl, self.err_msg.?);
+                self.err_msg = null;
                 return;
             },
             else => |e| return e,
@@ -278,47 +301,71 @@ pub const LLVMIRModule = struct {
     }
 
     fn gen(self: *LLVMIRModule, module: *Module, typed_value: TypedValue, src: usize) !void {
-        switch (typed_value.ty.zigTypeTag()) {
-            .Fn => {
-                const func = typed_value.val.castTag(.function).?.data;
+        if (typed_value.val.castTag(.function)) |func_inst| {
+            const func = func_inst.data;
 
-                const llvm_func = try self.resolveLLVMFunction(func);
+            const llvm_func = try self.resolveLLVMFunction(func, src);
 
-                // We remove all the basic blocks of a function to support incremental
-                // compilation!
-                // TODO: remove all basic blocks if functions can have more than one
-                if (llvm_func.getFirstBasicBlock()) |bb| {
-                    bb.deleteBasicBlock();
-                }
+            // This gets the LLVM values from the function and stores them in `self.args`.
+            const fn_param_len = func.owner_decl.typed_value.most_recent.typed_value.ty.fnParamLen();
+            var args = try self.gpa.alloc(*const llvm.ValueRef, fn_param_len);
+            defer self.gpa.free(args);
 
-                const entry_block = llvm_func.appendBasicBlock("Entry");
-                self.builder.positionBuilderAtEnd(entry_block);
+            for (args) |*arg, i| {
+                arg.* = llvm.getParam(llvm_func, @intCast(c_uint, i));
+            }
+            self.args = args;
+            self.arg_index = 0;
 
-                const instructions = func.body.instructions;
-                for (instructions) |inst| {
-                    switch (inst.tag) {
-                        .breakpoint => try self.genBreakpoint(inst.castTag(.breakpoint).?),
-                        .call => try self.genCall(inst.castTag(.call).?),
-                        .unreach => self.genUnreach(inst.castTag(.unreach).?),
-                        .retvoid => self.genRetVoid(inst.castTag(.retvoid).?),
-                        .arg => self.genArg(inst.castTag(.arg).?),
-                        .dbg_stmt => {
-                            // TODO: implement debug info
-                        },
-                        else => |tag| return self.fail(src, "TODO implement LLVM codegen for Zir instruction: {}", .{tag}),
-                    }
-                }
-            },
-            else => |ty| return self.fail(src, "TODO implement LLVM codegen for top-level decl type: {}", .{ty}),
+            // Make sure no other LLVM values from other functions can be referenced
+            self.func_inst_table.clearRetainingCapacity();
+
+            // We remove all the basic blocks of a function to support incremental
+            // compilation!
+            // TODO: remove all basic blocks if functions can have more than one
+            if (llvm_func.getFirstBasicBlock()) |bb| {
+                bb.deleteBasicBlock();
+            }
+
+            const entry_block = llvm_func.appendBasicBlock("Entry");
+            self.builder.positionBuilderAtEnd(entry_block);
+
+            const instructions = func.body.instructions;
+            for (instructions) |inst| {
+                const opt_llvm_val: ?*const llvm.ValueRef = switch (inst.tag) {
+                    .add => try self.genAdd(inst.castTag(.add).?),
+                    .alloc => try self.genAlloc(inst.castTag(.alloc).?),
+                    .arg => try self.genArg(inst.castTag(.arg).?),
+                    .bitcast => try self.genBitCast(inst.castTag(.bitcast).?),
+                    .breakpoint => try self.genBreakpoint(inst.castTag(.breakpoint).?),
+                    .call => try self.genCall(inst.castTag(.call).?),
+                    .intcast => try self.genIntCast(inst.castTag(.intcast).?),
+                    .load => try self.genLoad(inst.castTag(.load).?),
+                    .not => try self.genNot(inst.castTag(.not).?),
+                    .ret => try self.genRet(inst.castTag(.ret).?),
+                    .retvoid => self.genRetVoid(inst.castTag(.retvoid).?),
+                    .store => try self.genStore(inst.castTag(.store).?),
+                    .sub => try self.genSub(inst.castTag(.sub).?),
+                    .unreach => self.genUnreach(inst.castTag(.unreach).?),
+                    .dbg_stmt => blk: {
+                        // TODO: implement debug info
+                        break :blk null;
+                    },
+                    else => |tag| return self.fail(src, "TODO implement LLVM codegen for Zir instruction: {}", .{tag}),
+                };
+                if (opt_llvm_val) |llvm_val| try self.func_inst_table.putNoClobber(self.gpa, inst, llvm_val);
+            }
+        } else {
+            return self.fail(src, "TODO implement LLVM codegen for top-level decl type: {}", .{typed_value.ty});
         }
     }
 
-    fn genCall(self: *LLVMIRModule, inst: *Inst.Call) !void {
+    fn genCall(self: *LLVMIRModule, inst: *Inst.Call) !?*const llvm.ValueRef {
         if (inst.func.value()) |func_value| {
             if (func_value.castTag(.function)) |func_payload| {
                 const func = func_payload.data;
                 const zig_fn_type = func.owner_decl.typed_value.most_recent.typed_value.ty;
-                const llvm_fn = try self.resolveLLVMFunction(func);
+                const llvm_fn = try self.resolveLLVMFunction(func, inst.base.src);
 
                 const num_args = inst.args.len;
 
@@ -338,54 +385,165 @@ pub const LLVMIRModule = struct {
                     "",
                 );
 
-                if (zig_fn_type.fnReturnType().zigTypeTag() == .NoReturn) {
+                const return_type = zig_fn_type.fnReturnType();
+                if (return_type.tag() == .noreturn) {
                     _ = self.builder.buildUnreachable();
                 }
+
+                // No need to store the LLVM value if the return type is void or noreturn
+                if (!return_type.hasCodeGenBits()) return null;
+
+                return call;
             }
         }
+        return self.fail(inst.base.src, "TODO implement calling runtime known function pointer LLVM backend", .{});
     }
 
-    fn genRetVoid(self: *LLVMIRModule, inst: *Inst.NoOp) void {
+    fn genRetVoid(self: *LLVMIRModule, inst: *Inst.NoOp) ?*const llvm.ValueRef {
         _ = self.builder.buildRetVoid();
+        return null;
     }
 
-    fn genUnreach(self: *LLVMIRModule, inst: *Inst.NoOp) void {
+    fn genRet(self: *LLVMIRModule, inst: *Inst.UnOp) !?*const llvm.ValueRef {
+        _ = self.builder.buildRet(try self.resolveInst(inst.operand));
+        return null;
+    }
+
+    fn genNot(self: *LLVMIRModule, inst: *Inst.UnOp) !?*const llvm.ValueRef {
+        return self.builder.buildNot(try self.resolveInst(inst.operand), "");
+    }
+
+    fn genUnreach(self: *LLVMIRModule, inst: *Inst.NoOp) ?*const llvm.ValueRef {
         _ = self.builder.buildUnreachable();
+        return null;
     }
 
-    fn genArg(self: *LLVMIRModule, inst: *Inst.Arg) void {
-        // TODO: implement this
+    fn genAdd(self: *LLVMIRModule, inst: *Inst.BinOp) !?*const llvm.ValueRef {
+        const lhs = try self.resolveInst(inst.lhs);
+        const rhs = try self.resolveInst(inst.rhs);
+
+        if (!inst.base.ty.isInt())
+            return self.fail(inst.base.src, "TODO implement 'genAdd' for type {}", .{inst.base.ty});
+
+        return if (inst.base.ty.isSignedInt())
+            self.builder.buildNSWAdd(lhs, rhs, "")
+        else
+            self.builder.buildNUWAdd(lhs, rhs, "");
     }
 
-    fn genBreakpoint(self: *LLVMIRModule, inst: *Inst.NoOp) !void {
-        // TODO: Store this function somewhere such that we dont have to add it again
-        const fn_type = llvm.TypeRef.functionType(llvm.voidType(), null, 0, false);
-        const func = self.llvm_module.addFunction("llvm.debugtrap", fn_type);
-        // TODO: add assertion: LLVMGetIntrinsicID
-        _ = self.builder.buildCall(func, null, 0, "");
+    fn genSub(self: *LLVMIRModule, inst: *Inst.BinOp) !?*const llvm.ValueRef {
+        const lhs = try self.resolveInst(inst.lhs);
+        const rhs = try self.resolveInst(inst.rhs);
+
+        if (!inst.base.ty.isInt())
+            return self.fail(inst.base.src, "TODO implement 'genSub' for type {}", .{inst.base.ty});
+
+        return if (inst.base.ty.isSignedInt())
+            self.builder.buildNSWSub(lhs, rhs, "")
+        else
+            self.builder.buildNUWSub(lhs, rhs, "");
+    }
+
+    fn genIntCast(self: *LLVMIRModule, inst: *Inst.UnOp) !?*const llvm.ValueRef {
+        const val = try self.resolveInst(inst.operand);
+
+        const signed = inst.base.ty.isSignedInt();
+        // TODO: Should we use intcast here or just a simple bitcast?
+        //       LLVM does truncation vs bitcast (+signed extension) in the intcast depending on the sizes
+        return self.builder.buildIntCast2(val, try self.getLLVMType(inst.base.ty, inst.base.src), signed, "");
+    }
+
+    fn genBitCast(self: *LLVMIRModule, inst: *Inst.UnOp) !?*const llvm.ValueRef {
+        const val = try self.resolveInst(inst.operand);
+        const dest_type = try self.getLLVMType(inst.base.ty, inst.base.src);
+
+        return self.builder.buildBitCast(val, dest_type, "");
+    }
+
+    fn genArg(self: *LLVMIRModule, inst: *Inst.Arg) !?*const llvm.ValueRef {
+        const arg_val = self.args[self.arg_index];
+        self.arg_index += 1;
+
+        const ptr_val = self.builder.buildAlloca(try self.getLLVMType(inst.base.ty, inst.base.src), "");
+        _ = self.builder.buildStore(arg_val, ptr_val);
+        return self.builder.buildLoad(ptr_val, "");
+    }
+
+    fn genAlloc(self: *LLVMIRModule, inst: *Inst.NoOp) !?*const llvm.ValueRef {
+        // buildAlloca expects the pointee type, not the pointer type, so assert that
+        // a Payload.PointerSimple is passed to the alloc instruction.
+        const pointee_type = inst.base.ty.castPointer().?.data;
+
+        // TODO: figure out a way to get the name of the var decl.
+        // TODO: set alignment and volatile
+        return self.builder.buildAlloca(try self.getLLVMType(pointee_type, inst.base.src), "");
+    }
+
+    fn genStore(self: *LLVMIRModule, inst: *Inst.BinOp) !?*const llvm.ValueRef {
+        const val = try self.resolveInst(inst.rhs);
+        const ptr = try self.resolveInst(inst.lhs);
+        _ = self.builder.buildStore(val, ptr);
+        return null;
+    }
+
+    fn genLoad(self: *LLVMIRModule, inst: *Inst.UnOp) !?*const llvm.ValueRef {
+        const ptr_val = try self.resolveInst(inst.operand);
+        return self.builder.buildLoad(ptr_val, "");
+    }
+
+    fn genBreakpoint(self: *LLVMIRModule, inst: *Inst.NoOp) !?*const llvm.ValueRef {
+        const llvn_fn = self.getIntrinsic("llvm.debugtrap");
+        _ = self.builder.buildCall(llvn_fn, null, 0, "");
+        return null;
+    }
+
+    fn getIntrinsic(self: *LLVMIRModule, name: []const u8) *const llvm.ValueRef {
+        const id = llvm.lookupIntrinsicID(name.ptr, name.len);
+        assert(id != 0);
+        // TODO: add support for overload intrinsics by passing the prefix of the intrinsic
+        //       to `lookupIntrinsicID` and then passing the correct types to
+        //       `getIntrinsicDeclaration`
+        return self.llvm_module.getIntrinsicDeclaration(id, null, 0);
     }
 
     fn resolveInst(self: *LLVMIRModule, inst: *ir.Inst) !*const llvm.ValueRef {
-        if (inst.castTag(.constant)) |const_inst| {
-            return self.genTypedValue(inst.src, .{ .ty = inst.ty, .val = const_inst.val });
+        if (inst.value()) |val| {
+            return self.genTypedValue(inst.src, .{ .ty = inst.ty, .val = val });
         }
-        return self.fail(inst.src, "TODO implement resolveInst", .{});
+        if (self.func_inst_table.get(inst)) |value| return value;
+
+        return self.fail(inst.src, "TODO implement global llvm values (or the value is not in the func_inst_table table)", .{});
     }
 
-    fn genTypedValue(self: *LLVMIRModule, src: usize, typed_value: TypedValue) !*const llvm.ValueRef {
-        const llvm_type = self.getLLVMType(typed_value.ty);
+    fn genTypedValue(self: *LLVMIRModule, src: usize, tv: TypedValue) !*const llvm.ValueRef {
+        const llvm_type = try self.getLLVMType(tv.ty, src);
 
-        if (typed_value.val.isUndef())
+        if (tv.val.isUndef())
             return llvm_type.getUndef();
 
-        switch (typed_value.ty.zigTypeTag()) {
-            .Bool => return if (typed_value.val.toBool()) llvm_type.constAllOnes() else llvm_type.constNull(),
-            else => return self.fail(src, "TODO implement const of type '{}'", .{typed_value.ty}),
+        switch (tv.ty.zigTypeTag()) {
+            .Bool => return if (tv.val.toBool()) llvm_type.constAllOnes() else llvm_type.constNull(),
+            .Int => {
+                var bigint_space: Value.BigIntSpace = undefined;
+                const bigint = tv.val.toBigInt(&bigint_space);
+
+                if (bigint.eqZero()) return llvm_type.constNull();
+
+                if (bigint.limbs.len != 1) {
+                    return self.fail(src, "TODO implement bigger bigint", .{});
+                }
+                const llvm_int = llvm_type.constInt(bigint.limbs[0], false);
+                if (!bigint.positive) {
+                    return llvm.constNeg(llvm_int);
+                }
+                return llvm_int;
+            },
+            else => return self.fail(src, "TODO implement const of type '{}'", .{tv.ty}),
         }
     }
 
     /// If the llvm function does not exist, create it
-    fn resolveLLVMFunction(self: *LLVMIRModule, func: *Module.Fn) !*const llvm.ValueRef {
+    fn resolveLLVMFunction(self: *LLVMIRModule, func: *Module.Fn, src: usize) !*const llvm.ValueRef {
         // TODO: do we want to store this in our own datastructure?
         if (self.llvm_module.getNamedFunction(func.owner_decl.name)) |llvm_fn| return llvm_fn;
 
@@ -402,25 +560,25 @@ pub const LLVMIRModule = struct {
         defer self.gpa.free(llvm_param);
 
         for (fn_param_types) |fn_param, i| {
-            llvm_param[i] = self.getLLVMType(fn_param);
+            llvm_param[i] = try self.getLLVMType(fn_param, src);
         }
 
         const fn_type = llvm.TypeRef.functionType(
-            self.getLLVMType(return_type),
+            try self.getLLVMType(return_type, src),
             if (fn_param_len == 0) null else llvm_param.ptr,
             @intCast(c_uint, fn_param_len),
             false,
         );
         const llvm_fn = self.llvm_module.addFunction(func.owner_decl.name, fn_type);
 
-        if (return_type.zigTypeTag() == .NoReturn) {
+        if (return_type.tag() == .noreturn) {
             llvm_fn.addFnAttr("noreturn");
         }
 
         return llvm_fn;
     }
 
-    fn getLLVMType(self: *LLVMIRModule, t: Type) *const llvm.TypeRef {
+    fn getLLVMType(self: *LLVMIRModule, t: Type, src: usize) error{ OutOfMemory, CodegenFail }!*const llvm.TypeRef {
         switch (t.zigTypeTag()) {
             .Void => return llvm.voidType(),
             .NoReturn => return llvm.voidType(),
@@ -429,13 +587,18 @@ pub const LLVMIRModule = struct {
                 return llvm.intType(info.bits);
             },
             .Bool => return llvm.intType(1),
-            else => unreachable,
+            .Pointer => {
+                const pointer = t.castPointer().?;
+                const elem_type = try self.getLLVMType(pointer.data, src);
+                return elem_type.pointerType(0);
+            },
+            else => return self.fail(src, "TODO implement getLLVMType for type '{}'", .{t}),
         }
     }
 
     pub fn fail(self: *LLVMIRModule, src: usize, comptime format: []const u8, args: anytype) error{ OutOfMemory, CodegenFail } {
         @setCold(true);
-        std.debug.assert(self.err_msg == null);
+        assert(self.err_msg == null);
         self.err_msg = try Compilation.ErrorMsg.create(self.gpa, src, format, args);
         return error.CodegenFail;
     }

src/llvm_bindings.zig

@@ -43,8 +43,14 @@ pub const TypeRef = opaque {
     pub const constAllOnes = LLVMConstAllOnes;
     extern fn LLVMConstAllOnes(Ty: *const TypeRef) *const ValueRef;
 
+    pub const constInt = LLVMConstInt;
+    extern fn LLVMConstInt(IntTy: *const TypeRef, N: c_ulonglong, SignExtend: LLVMBool) *const ValueRef;
+
     pub const getUndef = LLVMGetUndef;
     extern fn LLVMGetUndef(Ty: *const TypeRef) *const ValueRef;
+
+    pub const pointerType = LLVMPointerType;
+    extern fn LLVMPointerType(ElementType: *const TypeRef, AddressSpace: c_uint) *const TypeRef;
 };
 
 pub const ModuleRef = opaque {
@@ -63,10 +69,16 @@ pub const ModuleRef = opaque {
     pub const getNamedFunction = LLVMGetNamedFunction;
     extern fn LLVMGetNamedFunction(*const ModuleRef, Name: [*:0]const u8) ?*const ValueRef;
 
+    pub const getIntrinsicDeclaration = LLVMGetIntrinsicDeclaration;
+    extern fn LLVMGetIntrinsicDeclaration(Mod: *const ModuleRef, ID: c_uint, ParamTypes: ?[*]*const TypeRef, ParamCount: usize) *const ValueRef;
+
     pub const printToString = LLVMPrintModuleToString;
     extern fn LLVMPrintModuleToString(*const ModuleRef) [*:0]const u8;
 };
 
+pub const lookupIntrinsicID = LLVMLookupIntrinsicID;
+extern fn LLVMLookupIntrinsicID(Name: [*]const u8, NameLen: usize) c_uint;
+
 pub const disposeMessage = LLVMDisposeMessage;
 extern fn LLVMDisposeMessage(Message: [*:0]const u8) void;
 
@@ -76,9 +88,15 @@ pub const VerifierFailureAction = extern enum {
     ReturnStatus,
 };
 
+pub const constNeg = LLVMConstNeg;
+extern fn LLVMConstNeg(ConstantVal: *const ValueRef) *const ValueRef;
+
 pub const voidType = LLVMVoidType;
 extern fn LLVMVoidType() *const TypeRef;
 
+pub const getParam = LLVMGetParam;
+extern fn LLVMGetParam(Fn: *const ValueRef, Index: c_uint) *const ValueRef;
+
 pub const getEnumAttributeKindForName = LLVMGetEnumAttributeKindForName;
 extern fn LLVMGetEnumAttributeKindForName(Name: [*]const u8, SLen: usize) c_uint;
 
@@ -117,11 +135,41 @@ pub const BuilderRef = opaque {
     pub const buildRetVoid = LLVMBuildRetVoid;
     extern fn LLVMBuildRetVoid(*const BuilderRef) *const ValueRef;
 
+    pub const buildRet = LLVMBuildRet;
+    extern fn LLVMBuildRet(*const BuilderRef, V: *const ValueRef) *const ValueRef;
+
     pub const buildUnreachable = LLVMBuildUnreachable;
     extern fn LLVMBuildUnreachable(*const BuilderRef) *const ValueRef;
 
     pub const buildAlloca = LLVMBuildAlloca;
     extern fn LLVMBuildAlloca(*const BuilderRef, Ty: *const TypeRef, Name: [*:0]const u8) *const ValueRef;
+
+    pub const buildStore = LLVMBuildStore;
+    extern fn LLVMBuildStore(*const BuilderRef, Val: *const ValueRef, Ptr: *const ValueRef) *const ValueRef;
+
+    pub const buildLoad = LLVMBuildLoad;
+    extern fn LLVMBuildLoad(*const BuilderRef, PointerVal: *const ValueRef, Name: [*:0]const u8) *const ValueRef;
+
+    pub const buildNot = LLVMBuildNot;
+    extern fn LLVMBuildNot(*const BuilderRef, V: *const ValueRef, Name: [*:0]const u8) *const ValueRef;
+
+    pub const buildNSWAdd = LLVMBuildNSWAdd;
+    extern fn LLVMBuildNSWAdd(*const BuilderRef, LHS: *const ValueRef, RHS: *const ValueRef, Name: [*:0]const u8) *const ValueRef;
+
+    pub const buildNUWAdd = LLVMBuildNUWAdd;
+    extern fn LLVMBuildNUWAdd(*const BuilderRef, LHS: *const ValueRef, RHS: *const ValueRef, Name: [*:0]const u8) *const ValueRef;
+
+    pub const buildNSWSub = LLVMBuildNSWSub;
+    extern fn LLVMBuildNSWSub(*const BuilderRef, LHS: *const ValueRef, RHS: *const ValueRef, Name: [*:0]const u8) *const ValueRef;
+
+    pub const buildNUWSub = LLVMBuildNUWSub;
+    extern fn LLVMBuildNUWSub(*const BuilderRef, LHS: *const ValueRef, RHS: *const ValueRef, Name: [*:0]const u8) *const ValueRef;
+
+    pub const buildIntCast2 = LLVMBuildIntCast2;
+    extern fn LLVMBuildIntCast2(*const BuilderRef, Val: *const ValueRef, DestTy: *const TypeRef, IsSigned: LLVMBool, Name: [*:0]const u8) *const ValueRef;
+
+    pub const buildBitCast = LLVMBuildBitCast;
+    extern fn LLVMBuildBitCast(*const BuilderRef, Val: *const ValueRef, DestTy: *const TypeRef, Name: [*:0]const u8) *const ValueRef;
 };
 
 pub const BasicBlockRef = opaque {

src/test.zig

@@ -135,6 +135,7 @@ pub const TestContext = struct {
         extension: Extension,
         object_format: ?std.builtin.ObjectFormat = null,
         emit_h: bool = false,
+        llvm_backend: bool = false,
 
         files: std.ArrayList(File),
 
@@ -266,6 +267,21 @@ pub const TestContext = struct {
         return &ctx.cases.items[ctx.cases.items.len - 1];
     }
 
+    /// Adds a test case that uses the LLVM backend to emit an executable.
+    /// Currently this implies linking libc, because only then we can generate a testable executable.
+    pub fn exeUsingLlvmBackend(ctx: *TestContext, name: []const u8, target: CrossTarget) *Case {
+        ctx.cases.append(Case{
+            .name = name,
+            .target = target,
+            .updates = std.ArrayList(Update).init(ctx.cases.allocator),
+            .output_mode = .Exe,
+            .extension = .Zig,
+            .files = std.ArrayList(File).init(ctx.cases.allocator),
+            .llvm_backend = true,
+        }) catch unreachable;
+        return &ctx.cases.items[ctx.cases.items.len - 1];
+    }
+
     pub fn addObj(
         ctx: *TestContext,
         name: []const u8,
@@ -518,10 +534,30 @@ pub const TestContext = struct {
         try thread_pool.init(std.testing.allocator);
         defer thread_pool.deinit();
 
+        // Use the same global cache dir for all the tests, such that we for example don't have to
+        // rebuild musl libc for every case (when LLVM backend is enabled).
+        var global_tmp = std.testing.tmpDir(.{});
+        defer global_tmp.cleanup();
+
+        var cache_dir = try global_tmp.dir.makeOpenPath("zig-cache", .{});
+        defer cache_dir.close();
+        const tmp_dir_path = try std.fs.path.join(std.testing.allocator, &[_][]const u8{ ".", "zig-cache", "tmp", &global_tmp.sub_path });
+        defer std.testing.allocator.free(tmp_dir_path);
+
+        const global_cache_directory: Compilation.Directory = .{
+            .handle = cache_dir,
+            .path = try std.fs.path.join(std.testing.allocator, &[_][]const u8{ tmp_dir_path, "zig-cache" }),
+        };
+        defer std.testing.allocator.free(global_cache_directory.path.?);
+
         for (self.cases.items) |case| {
             if (build_options.skip_non_native and case.target.getCpuArch() != std.Target.current.cpu.arch)
                 continue;
 
+            // Skip tests that require LLVM backend when it is not available
+            if (!build_options.have_llvm and case.llvm_backend)
+                continue;
+
             var prg_node = root_node.start(case.name, case.updates.items.len);
             prg_node.activate();
             defer prg_node.end();
@@ -537,6 +573,7 @@ pub const TestContext = struct {
                 case,
                 zig_lib_directory,
                 &thread_pool,
+                global_cache_directory,
             );
         }
     }
@@ -548,6 +585,7 @@ pub const TestContext = struct {
         case: Case,
         zig_lib_directory: Compilation.Directory,
         thread_pool: *ThreadPool,
+        global_cache_directory: Compilation.Directory,
     ) !void {
         const target_info = try std.zig.system.NativeTargetInfo.detect(allocator, case.target);
         const target = target_info.target;
@@ -601,7 +639,7 @@ pub const TestContext = struct {
             null;
         const comp = try Compilation.create(allocator, .{
             .local_cache_directory = zig_cache_directory,
-            .global_cache_directory = zig_cache_directory,
+            .global_cache_directory = global_cache_directory,
             .zig_lib_directory = zig_lib_directory,
             .thread_pool = thread_pool,
             .root_name = "test_case",
@@ -619,6 +657,9 @@ pub const TestContext = struct {
             .object_format = case.object_format,
             .is_native_os = case.target.isNativeOs(),
             .is_native_abi = case.target.isNativeAbi(),
+            .link_libc = case.llvm_backend,
+            .use_llvm = case.llvm_backend,
+            .self_exe_path = std.testing.zig_exe_path,
         });
         defer comp.destroy();
 

src/zig_clang.h

@@ -8,14 +8,32 @@
 #ifndef ZIG_ZIG_CLANG_H
 #define ZIG_ZIG_CLANG_H
 
-#include "stage1/stage2.h"
 #include <inttypes.h>
 #include <stdbool.h>
+#include <stddef.h>
+
+#ifdef __cplusplus
+#define ZIG_EXTERN_C extern "C"
+#else
+#define ZIG_EXTERN_C
+#endif
 
 // ATTENTION: If you modify this file, be sure to update the corresponding
 // extern function declarations in the self-hosted compiler file
 // src/clang.zig.
 
+// ABI warning
+struct Stage2ErrorMsg {
+    const char *filename_ptr; // can be null
+    size_t filename_len;
+    const char *msg_ptr;
+    size_t msg_len;
+    const char *source; // valid until the ASTUnit is freed. can be null
+    unsigned line; // 0 based
+    unsigned column; // 0 based
+    unsigned offset; // byte offset into source
+};
+
 struct ZigClangSourceLocation {
     unsigned ID;
 };

test/stage2/llvm_backend.zig

@@ -0,0 +1,30 @@
+const std = @import("std");
+const TestContext = @import("../../src/test.zig").TestContext;
+const build_options = @import("build_options");
+
+// These tests should work with all platforms, but we're using linux_x64 for
+// now for consistency. Will be expanded eventually.
+const linux_x64 = std.zig.CrossTarget{
+    .cpu_arch = .x86_64,
+    .os_tag = .linux,
+};
+
+pub fn addCases(ctx: *TestContext) !void {
+    {
+        var case = ctx.exeUsingLlvmBackend("simple addition and subtraction", linux_x64);
+
+        case.addCompareOutput(
+            \\fn add(a: i32, b: i32) i32 {
+            \\    return a + b;
+            \\}
+            \\
+            \\export fn main() c_int {
+            \\    var a: i32 = -5;
+            \\    const x = add(a, 7);
+            \\    var y = add(2, 0);
+            \\    y -= x;
+            \\    return y;
+            \\}
+        , "");
+    }
+}

test/stage2/test.zig

@@ -31,6 +31,7 @@ pub fn addCases(ctx: *TestContext) !void {
     try @import("spu-ii.zig").addCases(ctx);
     try @import("arm.zig").addCases(ctx);
     try @import("aarch64.zig").addCases(ctx);
+    try @import("llvm_backend.zig").addCases(ctx);
 
     {
         var case = ctx.exe("hello world with updates", linux_x64);