# Chapter 5 - Extending the language: Control flow from collections import namedtuple from ctypes import CFUNCTYPE, c_double from enum import Enum import llvmlite.ir as ir import llvmlite.binding as llvm # Each token is a tuple of kind and value. kind is one of the enumeration values # in TokenKind. value is the textual value of the token in the input. class TokenKind(Enum): EOF = -1 DEF = -2 EXTERN = -3 IDENTIFIER = -4 NUMBER = -5 OPERATOR = -6 IF = -7 THEN = -8 ELSE = -9 FOR = -10 IN = -11 Token = namedtuple('Token', 'kind value') class Lexer(object): """Lexer for Kaleidoscope. Initialize the lexer with a string buffer. tokens() returns a generator that can be queried for tokens. The generator will emit an EOF token before stopping. """ def __init__(self, buf): assert len(buf) >= 1 self.buf = buf self.pos = 0 self.lastchar = self.buf[0] self._keyword_map = { 'def': TokenKind.DEF, 'extern': TokenKind.EXTERN, 'if': TokenKind.IF, 'then': TokenKind.THEN, 'else': TokenKind.ELSE, 'for': TokenKind.FOR, 'in': TokenKind.IN, } def tokens(self): while self.lastchar: # Skip whitespace while self.lastchar.isspace(): self._advance() # Identifier or keyword if self.lastchar.isalpha(): id_str = '' while self.lastchar.isalnum(): id_str += self.lastchar self._advance() if id_str in self._keyword_map: yield Token(kind=self._keyword_map[id_str], value=id_str) else: yield Token(kind=TokenKind.IDENTIFIER, value=id_str) # Number elif self.lastchar.isdigit() or self.lastchar == '.': num_str = '' while self.lastchar.isdigit() or self.lastchar == '.': num_str += self.lastchar self._advance() yield Token(kind=TokenKind.NUMBER, value=num_str) # Comment elif self.lastchar == '#': self._advance() while self.lastchar and self.lastchar not in '\r\n': self._advance() elif self.lastchar: # Some other char yield Token(kind=TokenKind.OPERATOR, value=self.lastchar) self._advance() yield Token(kind=TokenKind.EOF, value='') def _advance(self): try: self.pos += 1 self.lastchar = self.buf[self.pos] except IndexError: self.lastchar = '' # AST hierarchy class ASTNode(object): def dump(self, indent=0): raise NotImplementedError class ExprAST(ASTNode): pass class NumberExprAST(ExprAST): def __init__(self, val): self.val = val def dump(self, indent=0): return '{0}{1}[{2}]'.format( ' ' * indent, self.__class__.__name__, self.val) class VariableExprAST(ExprAST): def __init__(self, name): self.name = name def dump(self, indent=0): return '{0}{1}[{2}]'.format( ' ' * indent, self.__class__.__name__, self.name) class BinaryExprAST(ExprAST): def __init__(self, op, lhs, rhs): self.op = op self.lhs = lhs self.rhs = rhs def dump(self, indent=0): s = '{0}{1}[{2}]\n'.format( ' ' * indent, self.__class__.__name__, self.op) s += self.lhs.dump(indent + 2) + '\n' s += self.rhs.dump(indent + 2) return s class IfExprAST(ExprAST): def __init__(self, cond_expr, then_expr, else_expr): self.cond_expr = cond_expr self.then_expr = then_expr self.else_expr = else_expr def dump(self, indent=0): prefix = ' ' * indent s = '{0}{1}\n'.format(prefix, self.__class__.__name__) s += '{0} Condition:\n{1}\n'.format( prefix, self.cond_expr.dump(indent + 2)) s += '{0} Then:\n{1}\n'.format( prefix, self.then_expr.dump(indent + 2)) s += '{0} Else:\n{1}'.format( prefix, self.else_expr.dump(indent + 2)) return s class ForExprAST(ExprAST): def __init__(self, id_name, start_expr, end_expr, step_expr, body): self.id_name = id_name self.start_expr = start_expr self.end_expr = end_expr self.step_expr = step_expr self.body = body def dump(self, indent=0): prefix = ' ' * indent s = '{0}{1}\n'.format(prefix, self.__class__.__name__) s += '{0} Start [{1}]:\n{2}\n'.format( prefix, self.id_name, self.start_expr.dump(indent + 2)) s += '{0} End:\n{1}\n'.format( prefix, self.end_expr.dump(indent + 2)) s += '{0} Step:\n{1}\n'.format( prefix, self.step_expr.dump(indent + 2)) s += '{0} Body:\n{1}\n'.format( prefix, self.body.dump(indent + 2)) return s class CallExprAST(ExprAST): def __init__(self, callee, args): self.callee = callee self.args = args def dump(self, indent=0): s = '{0}{1}[{2}]\n'.format( ' ' * indent, self.__class__.__name__, self.callee) for arg in self.args: s += arg.dump(indent + 2) + '\n' return s[:-1] # snip out trailing '\n' class PrototypeAST(ASTNode): def __init__(self, name, argnames): self.name = name self.argnames = argnames def dump(self, indent=0): return '{0}{1}[{2}]'.format( ' ' * indent, self.__class__.__name__, ', '.join(self.argnames)) class FunctionAST(ASTNode): def __init__(self, proto, body): self.proto = proto self.body = body _anonymous_function_counter = 0 @classmethod def create_anonymous(klass, expr): """Create an anonymous function to hold an expression.""" klass._anonymous_function_counter += 1 return klass( PrototypeAST('_anon{0}'.format(klass._anonymous_function_counter), []), expr) def is_anonymous(self): return self.proto.name.startswith('_anon') def dump(self, indent=0): s = '{0}{1}[{2}]\n'.format( ' ' * indent, self.__class__.__name__, self.proto.dump()) s += self.body.dump(indent + 2) + '\n' return s class ParseError(Exception): pass class Parser(object): """Parser for the Kaleidoscope language. After the parser is created, invoke parse_toplevel multiple times to parse Kaleidoscope source into an AST. """ def __init__(self): self.token_generator = None self.cur_tok = None # toplevel ::= definition | external | expression | ';' def parse_toplevel(self, buf): """Given a string, returns an AST node representing it.""" self.token_generator = Lexer(buf).tokens() self.cur_tok = None self._get_next_token() if self.cur_tok.kind == TokenKind.EXTERN: return self._parse_external() elif self.cur_tok.kind == TokenKind.DEF: return self._parse_definition() elif self._cur_tok_is_operator(';'): self._get_next_token() return None else: return self._parse_toplevel_expression() def _get_next_token(self): self.cur_tok = next(self.token_generator) def _match(self, expected_kind, expected_value=None): """Consume the current token; verify that it's of the expected kind. If expected_kind == TokenKind.OPERATOR, verify the operator's value. """ if (expected_kind == TokenKind.OPERATOR and not self._cur_tok_is_operator(expected_value)): raise ParseError('Expected "{0}"'.format(expected_value)) elif expected_kind != self.cur_tok.kind: raise ParseError('Expected "{0}"'.format(expected_kind)) self._get_next_token() _precedence_map = {'<': 10, '+': 20, '-': 20, '*': 40} def _cur_tok_precedence(self): """Get the operator precedence of the current token.""" try: return Parser._precedence_map[self.cur_tok.value] except KeyError: return -1 def _cur_tok_is_operator(self, op): """Query whether the current token is the operator op""" return (self.cur_tok.kind == TokenKind.OPERATOR and self.cur_tok.value == op) # identifierexpr # ::= identifier # ::= identifier '(' expression* ')' def _parse_identifier_expr(self): id_name = self.cur_tok.value self._get_next_token() # If followed by a '(' it's a call; otherwise, a simple variable ref. if not self._cur_tok_is_operator('('): return VariableExprAST(id_name) self._get_next_token() args = [] if not self._cur_tok_is_operator(')'): while True: args.append(self._parse_expression()) if self._cur_tok_is_operator(')'): break self._match(TokenKind.OPERATOR, ',') self._get_next_token() # consume the ')' return CallExprAST(id_name, args) # numberexpr ::= number def _parse_number_expr(self): result = NumberExprAST(self.cur_tok.value) self._get_next_token() # consume the number return result # parenexpr ::= '(' expression ')' def _parse_paren_expr(self): self._get_next_token() # consume the '(' expr = self._parse_expression() self._match(TokenKind.OPERATOR, ')') return expr # primary # ::= identifierexpr # ::= numberexpr # ::= parenexpr # ::= ifexpr # ::= forexpr def _parse_primary(self): if self.cur_tok.kind == TokenKind.IDENTIFIER: return self._parse_identifier_expr() elif self.cur_tok.kind == TokenKind.NUMBER: return self._parse_number_expr() elif self._cur_tok_is_operator('('): return self._parse_paren_expr() elif self.cur_tok.kind == TokenKind.IF: return self._parse_if_expr() elif self.cur_tok.kind == TokenKind.FOR: return self._parse_for_expr() else: raise ParseError('Unknown token when expecting an expression') # ifexpr ::= 'if' expression 'then' expression 'else' expression def _parse_if_expr(self): self._get_next_token() # consume the 'if' cond_expr = self._parse_expression() self._match(TokenKind.THEN) then_expr = self._parse_expression() self._match(TokenKind.ELSE) else_expr = self._parse_expression() return IfExprAST(cond_expr, then_expr, else_expr) # forexpr ::= 'for' identifier '=' expr ',' expr (',' expr)? 'in' expr def _parse_for_expr(self): self._get_next_token() # consume the 'for' id_name = self.cur_tok.value self._match(TokenKind.IDENTIFIER) self._match(TokenKind.OPERATOR, '=') start_expr = self._parse_expression() self._match(TokenKind.OPERATOR, ',') end_expr = self._parse_expression() # The step part is optional if self._cur_tok_is_operator(','): self._get_next_token() step_expr = self._parse_expression() else: step_expr = None self._match(TokenKind.IN) body = self._parse_expression() return ForExprAST(id_name, start_expr, end_expr, step_expr, body) # binoprhs ::= ( primary)* def _parse_binop_rhs(self, expr_prec, lhs): """Parse the right-hand-side of a binary expression. expr_prec: minimal precedence to keep going (precedence climbing). lhs: AST of the left-hand-side. """ while True: cur_prec = self._cur_tok_precedence() # If this is a binary operator with precedence lower than the # currently parsed sub-expression, bail out. If it binds at least # as tightly, keep going. # Note that the precedence of non-operators is defined to be -1, # so this condition handles cases when the expression ended. if cur_prec < expr_prec: return lhs op = self.cur_tok.value self._get_next_token() # consume the operator rhs = self._parse_primary() next_prec = self._cur_tok_precedence() # There are three options: # 1. next_prec > cur_prec: we need to make a recursive call # 2. next_prec == cur_prec: no need for a recursive call, the next # iteration of this loop will handle it. # 3. next_prec < cur_prec: no need for a recursive call, combine # lhs and the next iteration will immediately bail out. if cur_prec < next_prec: rhs = self._parse_binop_rhs(cur_prec + 1, rhs) # Merge lhs/rhs lhs = BinaryExprAST(op, lhs, rhs) # expression ::= primary binoprhs def _parse_expression(self): lhs = self._parse_primary() # Start with precedence 0 because we want to bind any operator to the # expression at this point. return self._parse_binop_rhs(0, lhs) # prototype ::= id '(' id* ')' def _parse_prototype(self): name = self.cur_tok.value self._match(TokenKind.IDENTIFIER) self._match(TokenKind.OPERATOR, '(') argnames = [] while self.cur_tok.kind == TokenKind.IDENTIFIER: argnames.append(self.cur_tok.value) self._get_next_token() self._match(TokenKind.OPERATOR, ')') return PrototypeAST(name, argnames) # external ::= 'extern' prototype def _parse_external(self): self._get_next_token() # consume 'extern' return self._parse_prototype() # definition ::= 'def' prototype expression def _parse_definition(self): self._get_next_token() # consume 'def' proto = self._parse_prototype() expr = self._parse_expression() return FunctionAST(proto, expr) # toplevel ::= expression def _parse_toplevel_expression(self): expr = self._parse_expression() return FunctionAST.create_anonymous(expr) class CodegenError(Exception): pass class LLVMCodeGenerator(object): def __init__(self): """Initialize the code generator. This creates a new LLVM module into which code is generated. The generate_code() method can be called multiple times. It adds the code generated for this node into the module, and returns the IR value for the node. At any time, the current LLVM module being constructed can be obtained from the module attribute. """ self.module = ir.Module() # Current IR builder. self.builder = None # Manages a symbol table while a function is being codegen'd. Maps var # names to ir.Value. self.func_symtab = {} def generate_code(self, node): assert isinstance(node, (PrototypeAST, FunctionAST)) return self._codegen(node) def _codegen(self, node): """Node visitor. Dispathces upon node type. For AST node of class Foo, calls self._codegen_Foo. Each visitor is expected to return a llvmlite.ir.Value. """ method = '_codegen_' + node.__class__.__name__ return getattr(self, method)(node) def _codegen_NumberExprAST(self, node): return ir.Constant(ir.DoubleType(), float(node.val)) def _codegen_VariableExprAST(self, node): return self.func_symtab[node.name] def _codegen_BinaryExprAST(self, node): lhs = self._codegen(node.lhs) rhs = self._codegen(node.rhs) if node.op == '+': return self.builder.fadd(lhs, rhs, 'addtmp') elif node.op == '-': return self.builder.fsub(lhs, rhs, 'subtmp') elif node.op == '*': return self.builder.fmul(lhs, rhs, 'multmp') elif node.op == '<': cmp = self.builder.fcmp_unordered('<', lhs, rhs, 'cmptmp') return self.builder.uitofp(cmp, ir.DoubleType(), 'booltmp') else: raise CodegenError('Unknown binary operator', node.op) def _codegen_IfExprAST(self, node): # Emit comparison value cond_val = self._codegen(node.cond_expr) cmp = self.builder.fcmp_ordered( '!=', cond_val, ir.Constant(ir.DoubleType(), 0.0)) # Create basic blocks to express the control flow, with a conditional # branch to either then_bb or else_bb depending on cmp. else_bb and # merge_bb are not yet attached to the function's list of BBs because # if a nested IfExpr is generated we want to have a reasonably nested # order of BBs generated into the function. then_bb = self.builder.function.append_basic_block('then') else_bb = ir.Block(self.builder.function, 'else') merge_bb = ir.Block(self.builder.function, 'ifcont') self.builder.cbranch(cmp, then_bb, else_bb) # Emit the 'then' part self.builder.position_at_start(then_bb) then_val = self._codegen(node.then_expr) self.builder.branch(merge_bb) # Emission of then_val could have modified the current basic block. To # properly set up the PHI, remember which block the 'then' part ends in. then_bb = self.builder.block # Emit the 'else' part self.builder.function.basic_blocks.append(else_bb) self.builder.position_at_start(else_bb) else_val = self._codegen(node.else_expr) self.builder.branch(merge_bb) # Emit the merge ('ifcnt') block self.builder.function.basic_blocks.append(merge_bb) self.builder.position_at_start(merge_bb) phi = self.builder.phi(ir.DoubleType(), 'iftmp') phi.add_incoming(then_val, then_bb) phi.add_incoming(else_val, else_bb) return phi def _codegen_ForExprAST(self, node): # Output this as: # ... # start = startexpr # goto loop # loop: # variable = phi [start, loopheader], [nextvariable, loopend] # ... # bodyexpr # ... # loopend: # step = stepexpr # nextvariable = variable + step # endcond = endexpr # br endcond, loop, endloop # outloop: # Emit the start expr first, without the variable in scope. start_val = self._codegen(node.start_expr) preheader_bb = self.builder.block loop_bb = self.builder.function.append_basic_block('loop') # Insert an explicit fall through from the current block to loop_bb self.builder.branch(loop_bb) self.builder.position_at_start(loop_bb) # Start the PHI node with an entry for start phi = self.builder.phi(ir.DoubleType(), node.id_name) phi.add_incoming(start_val, preheader_bb) # Within the loop, the variable is defined equal to the PHI node. If it # shadows an existing variable, we have to restore it, so save it now. oldval = self.func_symtab.get(node.id_name) self.func_symtab[node.id_name] = phi # Emit the body of the loop. This, like any other expr, can change the # current BB. Note that we ignore the value computed by the body. body_val = self._codegen(node.body) if node.step_expr is None: stepval = ir.Constant(ir.DoubleType(), 1.0) else: stepval = self._codegen(node.step_expr) nextvar = self.builder.fadd(phi, stepval, 'nextvar') # Compute the end condition endcond = self._codegen(node.end_expr) cmp = self.builder.fcmp_ordered( '!=', endcond, ir.Constant(ir.DoubleType(), 0.0), 'loopcond') # Create the 'after loop' block and insert it loop_end_bb = self.builder.block after_bb = self.builder.function.append_basic_block('afterloop') # Insert the conditional branch into the end of loop_end_bb self.builder.cbranch(cmp, loop_bb, after_bb) # New code will be inserted into after_bb self.builder.position_at_start(after_bb) # Add a new entry to the PHI node for the backedge phi.add_incoming(nextvar, loop_end_bb) # Remove the loop variable from the symbol table; if it shadowed an # existing variable, restore that. if oldval is None: del self.func_symtab[node.id_name] else: self.func_symtab[node.id_name] = oldval # The 'for' expression always returns 0 return ir.Constant(ir.DoubleType(), 0.0) def _codegen_CallExprAST(self, node): callee_func = self.module.globals.get(node.callee, None) if callee_func is None or not isinstance(callee_func, ir.Function): raise CodegenError('Call to unknown function', node.callee) if len(callee_func.args) != len(node.args): raise CodegenError('Call argument length mismatch', node.callee) call_args = [self._codegen(arg) for arg in node.args] return self.builder.call(callee_func, call_args, 'calltmp') def _codegen_PrototypeAST(self, node): funcname = node.name # Create a function type func_ty = ir.FunctionType(ir.DoubleType(), [ir.DoubleType()] * len(node.argnames)) # If a function with this name already exists in the module... if funcname in self.module.globals: # We only allow the case in which a declaration exists and now the # function is defined (or redeclared) with the same number of args. existing_func = self.module[funcname] if not isinstance(existing_func, ir.Function): raise CodegenError('Function/Global name collision', funcname) if not existing_func.is_declaration(): raise CodegenError('Redifinition of {0}'.format(funcname)) if len(existing_func.function_type.args) != len(func_ty.args): raise CodegenError( 'Redifinition with different number of arguments') func = self.module.globals[funcname] else: # Otherwise create a new function func = ir.Function(self.module, func_ty, funcname) # Set function argument names from AST for i, arg in enumerate(func.args): arg.name = node.argnames[i] self.func_symtab[arg.name] = arg return func def _codegen_FunctionAST(self, node): # Reset the symbol table. Prototype generation will pre-populate it with # function arguments. self.func_symtab = {} # Create the function skeleton from the prototype. func = self._codegen(node.proto) # Create the entry BB in the function and set the builder to it. bb_entry = func.append_basic_block('entry') self.builder = ir.IRBuilder(bb_entry) retval = self._codegen(node.body) self.builder.ret(retval) return func class KaleidoscopeEvaluator(object): """Evaluator for Kaleidoscope expressions. Once an object is created, calls to evaluate() add new expressions to the module. Definitions (including externs) are only added into the IR - no JIT compilation occurs. When a toplevel expression is evaluated, the whole module is JITed and the result of the expression is returned. """ def __init__(self): llvm.initialize() llvm.initialize_native_target() llvm.initialize_native_asmprinter() self.codegen = LLVMCodeGenerator() self._add_builtins(self.codegen.module) self.target = llvm.Target.from_default_triple() def evaluate(self, codestr, optimize=True, llvmdump=False): """Evaluate code in codestr. Returns None for definitions and externs, and the evaluated expression value for toplevel expressions. """ # Parse the given code and generate code from it ast = Parser().parse_toplevel(codestr) self.codegen.generate_code(ast) if llvmdump: print('======== Unoptimized LLVM IR') print(str(self.codegen.module)) # If we're evaluating a definition or extern declaration, don't do # anything else. If we're evaluating an anonymous wrapper for a toplevel # expression, JIT-compile the module and run the function to get its # result. if not (isinstance(ast, FunctionAST) and ast.is_anonymous()): return None # Convert LLVM IR into in-memory representation llvmmod = llvm.parse_assembly(str(self.codegen.module)) # Optimize the module if optimize: pmb = llvm.create_pass_manager_builder() pmb.opt_level = 2 pm = llvm.create_module_pass_manager() pmb.populate(pm) pm.run(llvmmod) if llvmdump: print('======== Optimized LLVM IR') print(str(llvmmod)) # Create a MCJIT execution engine to JIT-compile the module. Note that # ee takes ownership of target_machine, so it has to be recreated anew # each time we call create_mcjit_compiler. target_machine = self.target.create_target_machine() with llvm.create_mcjit_compiler(llvmmod, target_machine) as ee: ee.finalize_object() if llvmdump: print('======== Machine code') print(target_machine.emit_assembly(llvmmod)) func = llvmmod.get_function(ast.proto.name) fptr = CFUNCTYPE(c_double)(ee.get_pointer_to_function(func)) result = fptr() return result def _add_builtins(self, module): # The C++ tutorial adds putchard() simply by defining it in the host C++ # code, which is then accessible to the JIT. It doesn't work as simply # for us; but luckily it's very easy to define new "C level" functions # for our JITed code to use - just emit them as LLVM IR. This is what # this method does. # Add the declaration of putchar putchar_ty = ir.FunctionType(ir.IntType(32), [ir.IntType(32)]) putchar = ir.Function(module, putchar_ty, 'putchar') # Add putchard putchard_ty = ir.FunctionType(ir.DoubleType(), [ir.DoubleType()]) putchard = ir.Function(module, putchard_ty, 'putchard') irbuilder = ir.IRBuilder(putchard.append_basic_block('entry')) ival = irbuilder.fptoui(putchard.args[0], ir.IntType(32), 'intcast') irbuilder.call(putchar, [ival]) irbuilder.ret(ir.Constant(ir.DoubleType(), 0)) #---- Some unit tests ----# import unittest class TestEvaluator(unittest.TestCase): def test_basic_if(self): e = KaleidoscopeEvaluator() e.evaluate('def foo(a b) a * if a < b then a + 1 else b + 1') self.assertEqual(e.evaluate('foo(3, 4)'), 12) self.assertEqual(e.evaluate('foo(5, 4)'), 25) def test_nested_if(self): e = KaleidoscopeEvaluator() e.evaluate(''' def foo(a b c) if a < b then if a < c then a * 2 else c * 2 else b * 2''') self.assertEqual(e.evaluate('foo(1, 20, 300)'), 2) self.assertEqual(e.evaluate('foo(10, 2, 300)'), 4) self.assertEqual(e.evaluate('foo(100, 2000, 30)'), 60) def test_for(self): # For doesn't return anything, so just make sure evaluating it doesn't # crash. e = KaleidoscopeEvaluator() e.evaluate(''' def foo(a b c) if a < b then for x = 1.0, x < b, c in x+a+c*b else c * 2''') self.assertEqual(e.evaluate('foo(1, 2, 3)'), 0) self.assertEqual(e.evaluate('foo(3, 2, 30)'), 60) if __name__ == '__main__': # Evaluate some code. kalei = KaleidoscopeEvaluator() kalei.evaluate('def foo(a b) for x = 65, x < a, b in putchard(x)') # We call these for their side effect of printing to stdout, so don't care # about the return value. kalei.evaluate('foo(79, 1)') kalei.evaluate('putchard(10)')