diff options
| author | Timon Kruiper <timonkruiper@gmail.com> | 2021-03-08 00:09:03 +0100 |
|---|---|---|
| committer | Andrew Kelley <andrew@ziglang.org> | 2021-03-20 15:10:44 -0700 |
| commit | a710368054096889385562addaed2d16f0705332 (patch) | |
| tree | 66cc9673d9bd3ddd914aafeee964d50e55cf4d50 /src/codegen/llvm.zig | |
| parent | 56677f2f2da41af5999b84b7f740d7bc463d1032 (diff) | |
| download | zig-a710368054096889385562addaed2d16f0705332.tar.gz zig-a710368054096889385562addaed2d16f0705332.zip | |
stage2: restructure LLVM backend
The LLVM backend is now structured into 3 different structs, namely
Object, DeclGen and FuncGen. Object represents an object that is
generated by the LLVM backend. DeclGen is responsible for generating
a decl and FuncGen is responsible for generating llvm instructions
from tzir in a function.
Diffstat (limited to 'src/codegen/llvm.zig')
| -rw-r--r-- | src/codegen/llvm.zig | 743 |
1 files changed, 387 insertions, 356 deletions
diff --git a/src/codegen/llvm.zig b/src/codegen/llvm.zig index d0ba4b8986..06f70db665 100644 --- a/src/codegen/llvm.zig +++ b/src/codegen/llvm.zig @@ -16,7 +16,6 @@ const Value = @import("../value.zig").Value; const Type = @import("../type.zig").Type; const LazySrcLoc = Module.LazySrcLoc; -const SrcLoc = Module.SrcLoc; pub fn targetTriple(allocator: *Allocator, target: std.Target) ![:0]u8 { const llvm_arch = switch (target.cpu.arch) { @@ -146,83 +145,42 @@ pub fn targetTriple(allocator: *Allocator, target: std.Target) ![:0]u8 { return std.fmt.allocPrintZ(allocator, "{s}-unknown-{s}-{s}", .{ llvm_arch, llvm_os, llvm_abi }); } -pub const LLVMIRModule = struct { - module: *Module, +pub const Object = struct { llvm_module: *const llvm.Module, context: *const llvm.Context, target_machine: *const llvm.TargetMachine, - builder: *const llvm.Builder, - - object_path: []const u8, - - gpa: *Allocator, - err_msg: ?*Module.ErrorMsg = null, - - // TODO: The fields below should really move into a different struct, - // because they are only valid when generating a function - - /// TODO: this should not be undefined since it should be in another per-decl struct - /// Curent decl we are analysing. Stored to get source locations from relative info - decl: *Module.Decl = undefined, - - /// 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. - /// TODO: Change this to a stack of Branch. Currently we store all the values from all the blocks - /// in here, however if a block ends, the instructions can be thrown away. - func_inst_table: std.AutoHashMapUnmanaged(*Inst, *const llvm.Value) = .{}, - - /// These fields are used to refer to the LLVM value of the function paramaters in an Arg instruction. - args: []*const llvm.Value = &[_]*const llvm.Value{}, - arg_index: usize = 0, - - entry_block: *const llvm.BasicBlock = undefined, - /// This fields stores the last alloca instruction, such that we can append more alloca instructions - /// to the top of the function. - latest_alloca_inst: ?*const llvm.Value = null, - - llvm_func: *const llvm.Value = undefined, - - /// This data structure is used to implement breaking to blocks. - blocks: std.AutoHashMapUnmanaged(*Inst.Block, struct { - parent_bb: *const llvm.BasicBlock, - break_bbs: *BreakBasicBlocks, - break_vals: *BreakValues, - }) = .{}, - - src_loc: Module.SrcLoc, - - const BreakBasicBlocks = std.ArrayListUnmanaged(*const llvm.BasicBlock); - const BreakValues = std.ArrayListUnmanaged(*const llvm.Value); + object_pathZ: [:0]const u8, - pub fn create(allocator: *Allocator, sub_path: []const u8, options: link.Options) !*LLVMIRModule { - const self = try allocator.create(LLVMIRModule); + pub fn create(allocator: *Allocator, sub_path: []const u8, options: link.Options) !*Object { + const self = try allocator.create(Object); errdefer allocator.destroy(self); - const gpa = options.module.?.gpa; - - const obj_basename = try std.zig.binNameAlloc(gpa, .{ + const obj_basename = try std.zig.binNameAlloc(allocator, .{ .root_name = options.root_name, .target = options.target, .output_mode = .Obj, }); - defer gpa.free(obj_basename); + defer allocator.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); + const object_path = try o_directory.join(allocator, &[_][]const u8{obj_basename}); + defer allocator.free(object_path); + + const object_pathZ = try allocator.dupeZ(u8, object_path); + errdefer allocator.free(object_pathZ); const context = llvm.Context.create(); errdefer context.dispose(); initializeLLVMTargets(); - const root_nameZ = try gpa.dupeZ(u8, options.root_name); - defer gpa.free(root_nameZ); + const root_nameZ = try allocator.dupeZ(u8, options.root_name); + defer allocator.free(root_nameZ); const llvm_module = llvm.Module.createWithName(root_nameZ.ptr, context); errdefer llvm_module.dispose(); - const llvm_target_triple = try targetTriple(gpa, options.target); - defer gpa.free(llvm_target_triple); + const llvm_target_triple = try targetTriple(allocator, options.target); + defer allocator.free(llvm_target_triple); var error_message: [*:0]const u8 = undefined; var target: *const llvm.Target = undefined; @@ -257,34 +215,21 @@ pub const LLVMIRModule = struct { ); errdefer target_machine.dispose(); - const builder = context.createBuilder(); - errdefer builder.dispose(); - self.* = .{ - .module = options.module.?, .llvm_module = llvm_module, .context = context, .target_machine = target_machine, - .builder = builder, - .object_path = object_path, - .gpa = gpa, - // TODO move this field into a struct that is only instantiated per gen() call - .src_loc = undefined, + .object_pathZ = object_pathZ, }; return self; } - pub fn deinit(self: *LLVMIRModule, allocator: *Allocator) void { - self.builder.dispose(); + pub fn deinit(self: *Object, allocator: *Allocator) void { self.target_machine.dispose(); self.llvm_module.dispose(); self.context.dispose(); - self.func_inst_table.deinit(self.gpa); - self.gpa.free(self.object_path); - - self.blocks.deinit(self.gpa); - + allocator.free(self.object_pathZ); allocator.destroy(self); } @@ -296,7 +241,7 @@ pub const LLVMIRModule = struct { llvm.initializeAllAsmParsers(); } - pub fn flushModule(self: *LLVMIRModule, comp: *Compilation) !void { + pub fn flushModule(self: *Object, comp: *Compilation) !void { if (comp.verbose_llvm_ir) { const dump = self.llvm_module.printToString(); defer llvm.disposeMessage(dump); @@ -317,13 +262,10 @@ pub const LLVMIRModule = struct { } } - const object_pathZ = try self.gpa.dupeZ(u8, self.object_path); - defer self.gpa.free(object_pathZ); - var error_message: [*:0]const u8 = undefined; if (self.target_machine.emitToFile( self.llvm_module, - object_pathZ.ptr, + self.object_pathZ.ptr, .ObjectFile, &error_message, ).toBool()) { @@ -335,23 +277,55 @@ pub const LLVMIRModule = struct { } } - pub fn updateDecl(self: *LLVMIRModule, module: *Module, decl: *Module.Decl) !void { - self.gen(module, decl) catch |err| switch (err) { + pub fn updateDecl(self: *Object, module: *Module, decl: *Module.Decl) !void { + var dg: DeclGen = .{ + .object = self, + .module = module, + .decl = decl, + .err_msg = null, + .gpa = module.gpa, + }; + dg.genDecl() catch |err| switch (err) { error.CodegenFail => { decl.analysis = .codegen_failure; - try module.failed_decls.put(module.gpa, decl, self.err_msg.?); - self.err_msg = null; + try module.failed_decls.put(module.gpa, decl, dg.err_msg.?); + dg.err_msg = null; return; }, else => |e| return e, }; } +}; + +pub const DeclGen = struct { + object: *Object, + module: *Module, + decl: *Module.Decl, + err_msg: ?*Module.ErrorMsg, + + gpa: *Allocator, + + fn fail(self: *DeclGen, src: LazySrcLoc, comptime format: []const u8, args: anytype) error{ OutOfMemory, CodegenFail } { + @setCold(true); + assert(self.err_msg == null); + const src_loc = src.toSrcLocWithDecl(self.decl); + self.err_msg = try Module.ErrorMsg.create(self.gpa, src_loc, format, args); + return error.CodegenFail; + } + + fn llvmModule(self: *DeclGen) *const llvm.Module { + return self.object.llvm_module; + } + + fn context(self: *DeclGen) *const llvm.Context { + return self.object.context; + } - fn gen(self: *LLVMIRModule, module: *Module, decl: *Module.Decl) !void { + fn genDecl(self: *DeclGen) !void { + const decl = self.decl; const typed_value = decl.typed_value.most_recent.typed_value; - self.src_loc = decl.srcLoc(); - self.decl = decl; - const src = self.src_loc.lazy; + + const src = decl.srcLoc().lazy; log.debug("gen: {s} type: {}, value: {}", .{ decl.name, typed_value.ty, typed_value.val }); @@ -363,16 +337,10 @@ pub const LLVMIRModule = struct { // 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.Value, fn_param_len); - defer self.gpa.free(args); for (args) |*arg, i| { arg.* = llvm.getParam(llvm_func, @intCast(c_uint, i)); } - self.args = args; - self.arg_index = 0; - - // 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! @@ -381,12 +349,25 @@ pub const LLVMIRModule = struct { bb.deleteBasicBlock(); } - self.entry_block = self.context.appendBasicBlock(llvm_func, "Entry"); - self.builder.positionBuilderAtEnd(self.entry_block); - self.latest_alloca_inst = null; - self.llvm_func = llvm_func; + const builder = self.context().createBuilder(); + + const entry_block = self.context().appendBasicBlock(llvm_func, "Entry"); + builder.positionBuilderAtEnd(entry_block); + + var fg: FuncGen = .{ + .dg = self, + .builder = builder, + .args = args, + .arg_index = 0, + .func_inst_table = .{}, + .entry_block = entry_block, + .latest_alloca_inst = null, + .llvm_func = llvm_func, + .blocks = .{}, + }; + defer fg.deinit(); - try self.genBody(func.body); + try fg.genBody(func.body); } else if (typed_value.val.castTag(.extern_fn)) |extern_fn| { _ = try self.resolveLLVMFunction(extern_fn.data, src); } else { @@ -394,7 +375,267 @@ pub const LLVMIRModule = struct { } } - fn genBody(self: *LLVMIRModule, body: ir.Body) error{ OutOfMemory, CodegenFail }!void { + /// If the llvm function does not exist, create it + fn resolveLLVMFunction(self: *DeclGen, func: *Module.Decl, src: LazySrcLoc) !*const llvm.Value { + // TODO: do we want to store this in our own datastructure? + if (self.llvmModule().getNamedFunction(func.name)) |llvm_fn| return llvm_fn; + + const zig_fn_type = func.typed_value.most_recent.typed_value.ty; + const return_type = zig_fn_type.fnReturnType(); + + const fn_param_len = zig_fn_type.fnParamLen(); + + const fn_param_types = try self.gpa.alloc(Type, fn_param_len); + defer self.gpa.free(fn_param_types); + zig_fn_type.fnParamTypes(fn_param_types); + + const llvm_param = try self.gpa.alloc(*const llvm.Type, fn_param_len); + defer self.gpa.free(llvm_param); + + for (fn_param_types) |fn_param, i| { + llvm_param[i] = try self.getLLVMType(fn_param, src); + } + + const fn_type = llvm.Type.functionType( + 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.llvmModule().addFunction(func.name, fn_type); + + if (return_type.tag() == .noreturn) { + self.addFnAttr(llvm_fn, "noreturn"); + } + + return llvm_fn; + } + + fn resolveGlobalDecl(self: *DeclGen, decl: *Module.Decl, src: LazySrcLoc) error{ OutOfMemory, CodegenFail }!*const llvm.Value { + // TODO: do we want to store this in our own datastructure? + if (self.llvmModule().getNamedGlobal(decl.name)) |val| return val; + + const typed_value = decl.typed_value.most_recent.typed_value; + + // TODO: remove this redundant `getLLVMType`, it is also called in `genTypedValue`. + const llvm_type = try self.getLLVMType(typed_value.ty, src); + const val = try self.genTypedValue(src, typed_value, null); + const global = self.llvmModule().addGlobal(llvm_type, decl.name); + llvm.setInitializer(global, val); + + // TODO ask the Decl if it is const + // https://github.com/ziglang/zig/issues/7582 + + return global; + } + + fn getLLVMType(self: *DeclGen, t: Type, src: LazySrcLoc) error{ OutOfMemory, CodegenFail }!*const llvm.Type { + switch (t.zigTypeTag()) { + .Void => return self.context().voidType(), + .NoReturn => return self.context().voidType(), + .Int => { + const info = t.intInfo(self.module.getTarget()); + return self.context().intType(info.bits); + }, + .Bool => return self.context().intType(1), + .Pointer => { + if (t.isSlice()) { + return self.fail(src, "TODO: LLVM backend: implement slices", .{}); + } else { + const elem_type = try self.getLLVMType(t.elemType(), src); + return elem_type.pointerType(0); + } + }, + .Array => { + const elem_type = try self.getLLVMType(t.elemType(), src); + return elem_type.arrayType(@intCast(c_uint, t.abiSize(self.module.getTarget()))); + }, + .Optional => { + if (!t.isPtrLikeOptional()) { + var buf: Type.Payload.ElemType = undefined; + const child_type = t.optionalChild(&buf); + + var optional_types: [2]*const llvm.Type = .{ + try self.getLLVMType(child_type, src), + self.context().intType(1), + }; + return self.context().structType(&optional_types, 2, .False); + } else { + return self.fail(src, "TODO implement optional pointers as actual pointers", .{}); + } + }, + else => return self.fail(src, "TODO implement getLLVMType for type '{}'", .{t}), + } + } + + // TODO: figure out a way to remove the FuncGen argument + fn genTypedValue(self: *DeclGen, src: LazySrcLoc, tv: TypedValue, fg: ?*FuncGen) error{ OutOfMemory, CodegenFail }!*const llvm.Value { + const llvm_type = try self.getLLVMType(tv.ty, src); + + if (tv.val.isUndef()) + return llvm_type.getUndef(); + + 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; + }, + .Pointer => switch (tv.val.tag()) { + .decl_ref => { + const decl = tv.val.castTag(.decl_ref).?.data; + const val = try self.resolveGlobalDecl(decl, src); + + const usize_type = try self.getLLVMType(Type.initTag(.usize), src); + + // TODO: second index should be the index into the memory! + var indices: [2]*const llvm.Value = .{ + usize_type.constNull(), + usize_type.constNull(), + }; + + // TODO: consider using buildInBoundsGEP2 for opaque pointers + return fg.?.builder.buildInBoundsGEP(val, &indices, 2, ""); + }, + .ref_val => { + const elem_value = tv.val.castTag(.ref_val).?.data; + const elem_type = tv.ty.castPointer().?.data; + const alloca = fg.?.buildAlloca(try self.getLLVMType(elem_type, src)); + _ = fg.?.builder.buildStore(try self.genTypedValue(src, .{ .ty = elem_type, .val = elem_value }, fg), alloca); + return alloca; + }, + else => return self.fail(src, "TODO implement const of pointer type '{}'", .{tv.ty}), + }, + .Array => { + if (tv.val.castTag(.bytes)) |payload| { + const zero_sentinel = if (tv.ty.sentinel()) |sentinel| blk: { + if (sentinel.tag() == .zero) break :blk true; + return self.fail(src, "TODO handle other sentinel values", .{}); + } else false; + + return self.context().constString(payload.data.ptr, @intCast(c_uint, payload.data.len), llvm.Bool.fromBool(!zero_sentinel)); + } else { + return self.fail(src, "TODO handle more array values", .{}); + } + }, + .Optional => { + if (!tv.ty.isPtrLikeOptional()) { + var buf: Type.Payload.ElemType = undefined; + const child_type = tv.ty.optionalChild(&buf); + const llvm_child_type = try self.getLLVMType(child_type, src); + + if (tv.val.tag() == .null_value) { + var optional_values: [2]*const llvm.Value = .{ + llvm_child_type.constNull(), + self.context().intType(1).constNull(), + }; + return self.context().constStruct(&optional_values, 2, .False); + } else { + var optional_values: [2]*const llvm.Value = .{ + try self.genTypedValue(src, .{ .ty = child_type, .val = tv.val }, fg), + self.context().intType(1).constAllOnes(), + }; + return self.context().constStruct(&optional_values, 2, .False); + } + } else { + return self.fail(src, "TODO implement const of optional pointer", .{}); + } + }, + else => return self.fail(src, "TODO implement const of type '{}'", .{tv.ty}), + } + } + + // Helper functions + fn addAttr(self: *DeclGen, val: *const llvm.Value, index: llvm.AttributeIndex, name: []const u8) void { + const kind_id = llvm.getEnumAttributeKindForName(name.ptr, name.len); + assert(kind_id != 0); + const llvm_attr = self.context().createEnumAttribute(kind_id, 0); + val.addAttributeAtIndex(index, llvm_attr); + } + + fn addFnAttr(self: *DeclGen, val: *const llvm.Value, attr_name: []const u8) void { + // TODO: improve this API, `addAttr(-1, attr_name)` + self.addAttr(val, std.math.maxInt(llvm.AttributeIndex), attr_name); + } +}; + +pub const FuncGen = struct { + dg: *DeclGen, + + builder: *const llvm.Builder, + + /// 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. + /// TODO: Change this to a stack of Branch. Currently we store all the values from all the blocks + /// in here, however if a block ends, the instructions can be thrown away. + func_inst_table: std.AutoHashMapUnmanaged(*Inst, *const llvm.Value), + + /// These fields are used to refer to the LLVM value of the function paramaters in an Arg instruction. + args: []*const llvm.Value, + arg_index: usize, + + entry_block: *const llvm.BasicBlock, + /// This fields stores the last alloca instruction, such that we can append more alloca instructions + /// to the top of the function. + latest_alloca_inst: ?*const llvm.Value, + + llvm_func: *const llvm.Value, + + /// This data structure is used to implement breaking to blocks. + blocks: std.AutoHashMapUnmanaged(*Inst.Block, struct { + parent_bb: *const llvm.BasicBlock, + break_bbs: *BreakBasicBlocks, + break_vals: *BreakValues, + }), + + const BreakBasicBlocks = std.ArrayListUnmanaged(*const llvm.BasicBlock); + const BreakValues = std.ArrayListUnmanaged(*const llvm.Value); + + fn deinit(self: *FuncGen) void { + self.builder.dispose(); + self.func_inst_table.deinit(self.gpa()); + self.gpa().free(self.args); + self.blocks.deinit(self.gpa()); + } + + fn fail(self: *FuncGen, src: LazySrcLoc, comptime format: []const u8, args: anytype) error{ OutOfMemory, CodegenFail } { + @setCold(true); + return self.dg.fail(src, format, args); + } + + fn llvmModule(self: *FuncGen) *const llvm.Module { + return self.dg.object.llvm_module; + } + + fn context(self: *FuncGen) *const llvm.Context { + return self.dg.object.context; + } + + fn gpa(self: *FuncGen) *Allocator { + return self.dg.gpa; + } + + fn resolveInst(self: *FuncGen, inst: *ir.Inst) !*const llvm.Value { + if (inst.value()) |val| { + return self.dg.genTypedValue(inst.src, .{ .ty = inst.ty, .val = val }, self); + } + 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 genBody(self: *FuncGen, body: ir.Body) error{ OutOfMemory, CodegenFail }!void { for (body.instructions) |inst| { const opt_value = switch (inst.tag) { .add => try self.genAdd(inst.castTag(.add).?), @@ -434,11 +675,11 @@ pub const LLVMIRModule = struct { }, else => |tag| return self.fail(inst.src, "TODO implement LLVM codegen for Zir instruction: {}", .{tag}), }; - if (opt_value) |val| try self.func_inst_table.putNoClobber(self.gpa, inst, val); + if (opt_value) |val| try self.func_inst_table.putNoClobber(self.gpa(), inst, val); } } - fn genCall(self: *LLVMIRModule, inst: *Inst.Call) !?*const llvm.Value { + fn genCall(self: *FuncGen, inst: *Inst.Call) !?*const llvm.Value { if (inst.func.value()) |func_value| { const fn_decl = if (func_value.castTag(.extern_fn)) |extern_fn| extern_fn.data @@ -448,12 +689,12 @@ pub const LLVMIRModule = struct { unreachable; const zig_fn_type = fn_decl.typed_value.most_recent.typed_value.ty; - const llvm_fn = try self.resolveLLVMFunction(fn_decl, inst.base.src); + const llvm_fn = try self.dg.resolveLLVMFunction(fn_decl, inst.base.src); const num_args = inst.args.len; - const llvm_param_vals = try self.gpa.alloc(*const llvm.Value, num_args); - defer self.gpa.free(llvm_param_vals); + const llvm_param_vals = try self.gpa().alloc(*const llvm.Value, num_args); + defer self.gpa().free(llvm_param_vals); for (inst.args) |arg, i| { llvm_param_vals[i] = try self.resolveInst(arg); @@ -482,17 +723,17 @@ pub const LLVMIRModule = struct { } } - fn genRetVoid(self: *LLVMIRModule, inst: *Inst.NoOp) ?*const llvm.Value { + fn genRetVoid(self: *FuncGen, inst: *Inst.NoOp) ?*const llvm.Value { _ = self.builder.buildRetVoid(); return null; } - fn genRet(self: *LLVMIRModule, inst: *Inst.UnOp) !?*const llvm.Value { + fn genRet(self: *FuncGen, inst: *Inst.UnOp) !?*const llvm.Value { _ = self.builder.buildRet(try self.resolveInst(inst.operand)); return null; } - fn genCmp(self: *LLVMIRModule, inst: *Inst.BinOp, op: math.CompareOperator) !?*const llvm.Value { + fn genCmp(self: *FuncGen, inst: *Inst.BinOp, op: math.CompareOperator) !?*const llvm.Value { const lhs = try self.resolveInst(inst.lhs); const rhs = try self.resolveInst(inst.rhs); @@ -513,21 +754,21 @@ pub const LLVMIRModule = struct { return self.builder.buildICmp(operation, lhs, rhs, ""); } - fn genBlock(self: *LLVMIRModule, inst: *Inst.Block) !?*const llvm.Value { - const parent_bb = self.context.createBasicBlock("Block"); + fn genBlock(self: *FuncGen, inst: *Inst.Block) !?*const llvm.Value { + const parent_bb = self.context().createBasicBlock("Block"); // 5 breaks to a block seems like a reasonable default. - var break_bbs = try BreakBasicBlocks.initCapacity(self.gpa, 5); - var break_vals = try BreakValues.initCapacity(self.gpa, 5); - try self.blocks.putNoClobber(self.gpa, inst, .{ + var break_bbs = try BreakBasicBlocks.initCapacity(self.gpa(), 5); + var break_vals = try BreakValues.initCapacity(self.gpa(), 5); + try self.blocks.putNoClobber(self.gpa(), inst, .{ .parent_bb = parent_bb, .break_bbs = &break_bbs, .break_vals = &break_vals, }); defer { self.blocks.removeAssertDiscard(inst); - break_bbs.deinit(self.gpa); - break_vals.deinit(self.gpa); + break_bbs.deinit(self.gpa()); + break_vals.deinit(self.gpa()); } try self.genBody(inst.body); @@ -538,7 +779,7 @@ pub const LLVMIRModule = struct { // If the block does not return a value, we dont have to create a phi node. if (!inst.base.ty.hasCodeGenBits()) return null; - const phi_node = self.builder.buildPhi(try self.getLLVMType(inst.base.ty, inst.base.src), ""); + const phi_node = self.builder.buildPhi(try self.dg.getLLVMType(inst.base.ty, inst.base.src), ""); phi_node.addIncoming( break_vals.items.ptr, break_bbs.items.ptr, @@ -547,7 +788,7 @@ pub const LLVMIRModule = struct { return phi_node; } - fn genBr(self: *LLVMIRModule, inst: *Inst.Br) !?*const llvm.Value { + fn genBr(self: *FuncGen, inst: *Inst.Br) !?*const llvm.Value { var block = self.blocks.get(inst.block).?; // If the break doesn't break a value, then we don't have to add @@ -560,25 +801,25 @@ pub const LLVMIRModule = struct { // For the phi node, we need the basic blocks and the values of the // break instructions. - try block.break_bbs.append(self.gpa, self.builder.getInsertBlock()); - try block.break_vals.append(self.gpa, val); + try block.break_bbs.append(self.gpa(), self.builder.getInsertBlock()); + try block.break_vals.append(self.gpa(), val); _ = self.builder.buildBr(block.parent_bb); } return null; } - fn genBrVoid(self: *LLVMIRModule, inst: *Inst.BrVoid) !?*const llvm.Value { + fn genBrVoid(self: *FuncGen, inst: *Inst.BrVoid) !?*const llvm.Value { var block = self.blocks.get(inst.block).?; _ = self.builder.buildBr(block.parent_bb); return null; } - fn genCondBr(self: *LLVMIRModule, inst: *Inst.CondBr) !?*const llvm.Value { + fn genCondBr(self: *FuncGen, inst: *Inst.CondBr) !?*const llvm.Value { const condition_value = try self.resolveInst(inst.condition); - const then_block = self.context.appendBasicBlock(self.llvm_func, "Then"); - const else_block = self.context.appendBasicBlock(self.llvm_func, "Else"); + const then_block = self.context().appendBasicBlock(self.llvm_func, "Then"); + const else_block = self.context().appendBasicBlock(self.llvm_func, "Else"); { const prev_block = self.builder.getInsertBlock(); defer self.builder.positionBuilderAtEnd(prev_block); @@ -593,8 +834,8 @@ pub const LLVMIRModule = struct { return null; } - fn genLoop(self: *LLVMIRModule, inst: *Inst.Loop) !?*const llvm.Value { - const loop_block = self.context.appendBasicBlock(self.llvm_func, "Loop"); + fn genLoop(self: *FuncGen, inst: *Inst.Loop) !?*const llvm.Value { + const loop_block = self.context().appendBasicBlock(self.llvm_func, "Loop"); _ = self.builder.buildBr(loop_block); self.builder.positionBuilderAtEnd(loop_block); @@ -604,20 +845,20 @@ pub const LLVMIRModule = struct { return null; } - fn genNot(self: *LLVMIRModule, inst: *Inst.UnOp) !?*const llvm.Value { + fn genNot(self: *FuncGen, inst: *Inst.UnOp) !?*const llvm.Value { return self.builder.buildNot(try self.resolveInst(inst.operand), ""); } - fn genUnreach(self: *LLVMIRModule, inst: *Inst.NoOp) ?*const llvm.Value { + fn genUnreach(self: *FuncGen, inst: *Inst.NoOp) ?*const llvm.Value { _ = self.builder.buildUnreachable(); return null; } - fn genIsNonNull(self: *LLVMIRModule, inst: *Inst.UnOp, operand_is_ptr: bool) !?*const llvm.Value { + fn genIsNonNull(self: *FuncGen, inst: *Inst.UnOp, operand_is_ptr: bool) !?*const llvm.Value { const operand = try self.resolveInst(inst.operand); if (operand_is_ptr) { - const index_type = self.context.intType(32); + const index_type = self.context().intType(32); var indices: [2]*const llvm.Value = .{ index_type.constNull(), @@ -630,15 +871,15 @@ pub const LLVMIRModule = struct { } } - fn genIsNull(self: *LLVMIRModule, inst: *Inst.UnOp, operand_is_ptr: bool) !?*const llvm.Value { + fn genIsNull(self: *FuncGen, inst: *Inst.UnOp, operand_is_ptr: bool) !?*const llvm.Value { return self.builder.buildNot((try self.genIsNonNull(inst, operand_is_ptr)).?, ""); } - fn genOptionalPayload(self: *LLVMIRModule, inst: *Inst.UnOp, operand_is_ptr: bool) !?*const llvm.Value { + fn genOptionalPayload(self: *FuncGen, inst: *Inst.UnOp, operand_is_ptr: bool) !?*const llvm.Value { const operand = try self.resolveInst(inst.operand); if (operand_is_ptr) { - const index_type = self.context.intType(32); + const index_type = self.context().intType(32); var indices: [2]*const llvm.Value = .{ index_type.constNull(), @@ -651,7 +892,7 @@ pub const LLVMIRModule = struct { } } - fn genAdd(self: *LLVMIRModule, inst: *Inst.BinOp) !?*const llvm.Value { + fn genAdd(self: *FuncGen, inst: *Inst.BinOp) !?*const llvm.Value { const lhs = try self.resolveInst(inst.lhs); const rhs = try self.resolveInst(inst.rhs); @@ -664,7 +905,7 @@ pub const LLVMIRModule = struct { self.builder.buildNUWAdd(lhs, rhs, ""); } - fn genSub(self: *LLVMIRModule, inst: *Inst.BinOp) !?*const llvm.Value { + fn genSub(self: *FuncGen, inst: *Inst.BinOp) !?*const llvm.Value { const lhs = try self.resolveInst(inst.lhs); const rhs = try self.resolveInst(inst.rhs); @@ -677,44 +918,44 @@ pub const LLVMIRModule = struct { self.builder.buildNUWSub(lhs, rhs, ""); } - fn genIntCast(self: *LLVMIRModule, inst: *Inst.UnOp) !?*const llvm.Value { + fn genIntCast(self: *FuncGen, inst: *Inst.UnOp) !?*const llvm.Value { 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), llvm.Bool.fromBool(signed), ""); + return self.builder.buildIntCast2(val, try self.dg.getLLVMType(inst.base.ty, inst.base.src), llvm.Bool.fromBool(signed), ""); } - fn genBitCast(self: *LLVMIRModule, inst: *Inst.UnOp) !?*const llvm.Value { + fn genBitCast(self: *FuncGen, inst: *Inst.UnOp) !?*const llvm.Value { const val = try self.resolveInst(inst.operand); - const dest_type = try self.getLLVMType(inst.base.ty, inst.base.src); + const dest_type = try self.dg.getLLVMType(inst.base.ty, inst.base.src); return self.builder.buildBitCast(val, dest_type, ""); } - fn genArg(self: *LLVMIRModule, inst: *Inst.Arg) !?*const llvm.Value { + fn genArg(self: *FuncGen, inst: *Inst.Arg) !?*const llvm.Value { const arg_val = self.args[self.arg_index]; self.arg_index += 1; - const ptr_val = self.buildAlloca(try self.getLLVMType(inst.base.ty, inst.base.src)); + const ptr_val = self.buildAlloca(try self.dg.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.Value { + fn genAlloc(self: *FuncGen, inst: *Inst.NoOp) !?*const llvm.Value { // 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.buildAlloca(try self.getLLVMType(pointee_type, inst.base.src)); + return self.buildAlloca(try self.dg.getLLVMType(pointee_type, inst.base.src)); } /// Use this instead of builder.buildAlloca, because this function makes sure to /// put the alloca instruction at the top of the function! - fn buildAlloca(self: *LLVMIRModule, t: *const llvm.Type) *const llvm.Value { + fn buildAlloca(self: *FuncGen, t: *const llvm.Type) *const llvm.Value { const prev_block = self.builder.getInsertBlock(); defer self.builder.positionBuilderAtEnd(prev_block); @@ -736,240 +977,30 @@ pub const LLVMIRModule = struct { return val; } - fn genStore(self: *LLVMIRModule, inst: *Inst.BinOp) !?*const llvm.Value { + fn genStore(self: *FuncGen, inst: *Inst.BinOp) !?*const llvm.Value { 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.Value { + fn genLoad(self: *FuncGen, inst: *Inst.UnOp) !?*const llvm.Value { const ptr_val = try self.resolveInst(inst.operand); return self.builder.buildLoad(ptr_val, ""); } - fn genBreakpoint(self: *LLVMIRModule, inst: *Inst.NoOp) !?*const llvm.Value { + fn genBreakpoint(self: *FuncGen, inst: *Inst.NoOp) !?*const llvm.Value { 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.Value { + fn getIntrinsic(self: *FuncGen, name: []const u8) *const llvm.Value { 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.Value { - if (inst.value()) |val| { - return self.genTypedValue(inst.src, .{ .ty = inst.ty, .val = val }); - } - 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: LazySrcLoc, tv: TypedValue) error{ OutOfMemory, CodegenFail }!*const llvm.Value { - const llvm_type = try self.getLLVMType(tv.ty, src); - - if (tv.val.isUndef()) - return llvm_type.getUndef(); - - 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; - }, - .Pointer => switch (tv.val.tag()) { - .decl_ref => { - const decl = tv.val.castTag(.decl_ref).?.data; - const val = try self.resolveGlobalDecl(decl, src); - - const usize_type = try self.getLLVMType(Type.initTag(.usize), src); - - // TODO: second index should be the index into the memory! - var indices: [2]*const llvm.Value = .{ - usize_type.constNull(), - usize_type.constNull(), - }; - - // TODO: consider using buildInBoundsGEP2 for opaque pointers - return self.builder.buildInBoundsGEP(val, &indices, 2, ""); - }, - .ref_val => { - const elem_value = tv.val.castTag(.ref_val).?.data; - const elem_type = tv.ty.castPointer().?.data; - const alloca = self.buildAlloca(try self.getLLVMType(elem_type, src)); - _ = self.builder.buildStore(try self.genTypedValue(src, .{ .ty = elem_type, .val = elem_value }), alloca); - return alloca; - }, - else => return self.fail(src, "TODO implement const of pointer type '{}'", .{tv.ty}), - }, - .Array => { - if (tv.val.castTag(.bytes)) |payload| { - const zero_sentinel = if (tv.ty.sentinel()) |sentinel| blk: { - if (sentinel.tag() == .zero) break :blk true; - return self.fail(src, "TODO handle other sentinel values", .{}); - } else false; - - return self.context.constString(payload.data.ptr, @intCast(c_uint, payload.data.len), llvm.Bool.fromBool(!zero_sentinel)); - } else { - return self.fail(src, "TODO handle more array values", .{}); - } - }, - .Optional => { - if (!tv.ty.isPtrLikeOptional()) { - var buf: Type.Payload.ElemType = undefined; - const child_type = tv.ty.optionalChild(&buf); - const llvm_child_type = try self.getLLVMType(child_type, src); - - if (tv.val.tag() == .null_value) { - var optional_values: [2]*const llvm.Value = .{ - llvm_child_type.constNull(), - self.context.intType(1).constNull(), - }; - return self.context.constStruct(&optional_values, 2, .False); - } else { - var optional_values: [2]*const llvm.Value = .{ - try self.genTypedValue(src, .{ .ty = child_type, .val = tv.val }), - self.context.intType(1).constAllOnes(), - }; - return self.context.constStruct(&optional_values, 2, .False); - } - } else { - return self.fail(src, "TODO implement const of optional pointer", .{}); - } - }, - else => return self.fail(src, "TODO implement const of type '{}'", .{tv.ty}), - } - } - - fn getLLVMType(self: *LLVMIRModule, t: Type, src: LazySrcLoc) error{ OutOfMemory, CodegenFail }!*const llvm.Type { - switch (t.zigTypeTag()) { - .Void => return self.context.voidType(), - .NoReturn => return self.context.voidType(), - .Int => { - const info = t.intInfo(self.module.getTarget()); - return self.context.intType(info.bits); - }, - .Bool => return self.context.intType(1), - .Pointer => { - if (t.isSlice()) { - return self.fail(src, "TODO: LLVM backend: implement slices", .{}); - } else { - const elem_type = try self.getLLVMType(t.elemType(), src); - return elem_type.pointerType(0); - } - }, - .Array => { - const elem_type = try self.getLLVMType(t.elemType(), src); - return elem_type.arrayType(@intCast(c_uint, t.abiSize(self.module.getTarget()))); - }, - .Optional => { - if (!t.isPtrLikeOptional()) { - var buf: Type.Payload.ElemType = undefined; - const child_type = t.optionalChild(&buf); - - var optional_types: [2]*const llvm.Type = .{ - try self.getLLVMType(child_type, src), - self.context.intType(1), - }; - return self.context.structType(&optional_types, 2, .False); - } else { - return self.fail(src, "TODO implement optional pointers as actual pointers", .{}); - } - }, - else => return self.fail(src, "TODO implement getLLVMType for type '{}'", .{t}), - } - } - - fn resolveGlobalDecl(self: *LLVMIRModule, decl: *Module.Decl, src: LazySrcLoc) error{ OutOfMemory, CodegenFail }!*const llvm.Value { - // TODO: do we want to store this in our own datastructure? - if (self.llvm_module.getNamedGlobal(decl.name)) |val| return val; - - const typed_value = decl.typed_value.most_recent.typed_value; - - // TODO: remove this redundant `getLLVMType`, it is also called in `genTypedValue`. - const llvm_type = try self.getLLVMType(typed_value.ty, src); - const val = try self.genTypedValue(src, typed_value); - const global = self.llvm_module.addGlobal(llvm_type, decl.name); - llvm.setInitializer(global, val); - - // TODO ask the Decl if it is const - // https://github.com/ziglang/zig/issues/7582 - - return global; - } - - /// If the llvm function does not exist, create it - fn resolveLLVMFunction(self: *LLVMIRModule, func: *Module.Decl, src: LazySrcLoc) !*const llvm.Value { - // TODO: do we want to store this in our own datastructure? - if (self.llvm_module.getNamedFunction(func.name)) |llvm_fn| return llvm_fn; - - const zig_fn_type = func.typed_value.most_recent.typed_value.ty; - const return_type = zig_fn_type.fnReturnType(); - - const fn_param_len = zig_fn_type.fnParamLen(); - - const fn_param_types = try self.gpa.alloc(Type, fn_param_len); - defer self.gpa.free(fn_param_types); - zig_fn_type.fnParamTypes(fn_param_types); - - const llvm_param = try self.gpa.alloc(*const llvm.Type, fn_param_len); - defer self.gpa.free(llvm_param); - - for (fn_param_types) |fn_param, i| { - llvm_param[i] = try self.getLLVMType(fn_param, src); - } - - const fn_type = llvm.Type.functionType( - 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.name, fn_type); - - if (return_type.tag() == .noreturn) { - self.addFnAttr(llvm_fn, "noreturn"); - } - - return llvm_fn; - } - - // Helper functions - fn addAttr(self: LLVMIRModule, val: *const llvm.Value, index: llvm.AttributeIndex, name: []const u8) void { - const kind_id = llvm.getEnumAttributeKindForName(name.ptr, name.len); - assert(kind_id != 0); - const llvm_attr = self.context.createEnumAttribute(kind_id, 0); - val.addAttributeAtIndex(index, llvm_attr); - } - - fn addFnAttr(self: *LLVMIRModule, val: *const llvm.Value, attr_name: []const u8) void { - // TODO: improve this API, `addAttr(-1, attr_name)` - self.addAttr(val, std.math.maxInt(llvm.AttributeIndex), attr_name); - } - - pub fn fail(self: *LLVMIRModule, src: LazySrcLoc, comptime format: []const u8, args: anytype) error{ OutOfMemory, CodegenFail } { - @setCold(true); - assert(self.err_msg == null); - const src_loc = src.toSrcLocWithDecl(self.decl); - self.err_msg = try Module.ErrorMsg.create(self.gpa, src_loc, format, args); - return error.CodegenFail; + return self.llvmModule().getIntrinsicDeclaration(id, null, 0); } }; |
