Add OP_SELF_TAIL_CALL + cache len(instrs): 37-50% faster loops
- OP_SELF_TAIL_CALL: named-let loops reuse env instead of allocating new child env each iteration. Eliminates 50,000 dict allocations per sum-to(50000). - Cache n_instrs = len(instrs): avoid per-dispatch len() call. Updated in all code-switching paths (CALL, TAIL_CALL, RETURN, CALL_CC, continuation resume). - Combined with LOOK_ADD1/LOOK_SUB1 from previous commit: sum-to: 99x → 50x, ackermann: 79x → 50x, fib-rec: 85x → 63x. - Serialization updated for new opcode operand format. - 571 tests green.
This commit is contained in:
parent
e700273136
commit
b31e03216e
1 changed files with 38 additions and 12 deletions
|
|
@ -1116,6 +1116,7 @@ OP_LOOK_ADD1 = 81 # lookup + increment: arg = sym
|
|||
OP_LOOK_SUB1 = 82 # lookup + decrement: arg = sym
|
||||
OP_CONST_EQ_JF = 83 # push const, compare TOS, branch: arg = (const, jump_addr)
|
||||
OP_LOOK_CONST_CALL2 = 84 # lookup func, push const, call(2): arg = (sym, const)
|
||||
OP_SELF_TAIL_CALL = 85 # self-recursive tail call (reuse env): arg = (n_args, params_tuple)
|
||||
|
||||
# Specialization table: {symbol: {arity: opcode}}
|
||||
_BC_SPECIALIZE = {
|
||||
|
|
@ -1169,7 +1170,7 @@ _BC_FOLDABLE = {
|
|||
|
||||
class CodeObj:
|
||||
"""Compiled bytecode chunk."""
|
||||
__slots__ = ('instrs', 'name', 'source_map', 'ic', '_cur_line')
|
||||
__slots__ = ('instrs', 'name', 'source_map', 'ic', '_cur_line', '_self_name', '_self_params')
|
||||
def __init__(self, name=None):
|
||||
self.instrs = []; self.name = name
|
||||
self.source_map = [] # parallel to instrs: line number or None
|
||||
|
|
@ -1363,7 +1364,8 @@ def _bc(expr, code, env, tail=False):
|
|||
# Named let: (let loop ((v init)...) body...)
|
||||
name = a[0]; binds = _L(a[1]); body = a[2:]
|
||||
bps = [_L(b)[0] for b in binds]
|
||||
inner = _bc_lambda(body, bps, None, env, name=str(name))
|
||||
inner = _bc_lambda(body, bps, None, env, name=str(name),
|
||||
self_name=str(name), self_params=bps)
|
||||
code.emit(OP_PUSH_ENV)
|
||||
code.emit(OP_MAKE_CLOSURE, (inner, bps, None))
|
||||
code.emit(OP_DUP)
|
||||
|
|
@ -1490,6 +1492,12 @@ def _bc(expr, code, env, tail=False):
|
|||
for arg in call_args: _bc(arg, code, env)
|
||||
code.emit(spec[n]); return
|
||||
|
||||
# --- Self tail call optimization ---
|
||||
if tail and isinstance(head, Symbol) and hasattr(code, '_self_name') and str(head) == code._self_name:
|
||||
params = code._self_params
|
||||
for arg in call_args: _bc(arg, code, env)
|
||||
code.emit(OP_SELF_TAIL_CALL, (len(call_args), tuple(params))); return
|
||||
|
||||
# --- Function call ---
|
||||
_bc(head, code, env)
|
||||
for arg in call_args: _bc(arg, code, env)
|
||||
|
|
@ -1503,9 +1511,12 @@ def _bc_body(body, code, env, tail=False):
|
|||
_bc(body[-1], code, env, tail=tail)
|
||||
|
||||
|
||||
def _bc_lambda(body, params, rest, env, name=None):
|
||||
def _bc_lambda(body, params, rest, env, name=None, self_name=None, self_params=None):
|
||||
"""Compile a lambda body into a CodeObj."""
|
||||
inner = CodeObj(name=name)
|
||||
if self_name:
|
||||
inner._self_name = self_name
|
||||
inner._self_params = self_params
|
||||
# Handle internal defines (letrec* semantics)
|
||||
def_names = []
|
||||
body_list = list(body)
|
||||
|
|
@ -1622,7 +1633,8 @@ def vm_exec(code, env):
|
|||
c = ci.cont
|
||||
frames = [(i, p, _deep_copy_env(e), list(s)) for i, p, e, s in c.frames]
|
||||
stack = list(c.stack); stack.append(ci.val)
|
||||
ip = c.ip; instrs = c.instrs; env = _deep_copy_env(c.env)
|
||||
ip = c.ip; instrs = c.instrs; n_instrs = len(instrs)
|
||||
env = _deep_copy_env(c.env)
|
||||
smap = None
|
||||
except LispErr as e:
|
||||
if e.source_line is None and smap and ip > 0 and ip - 1 < len(smap):
|
||||
|
|
@ -1636,7 +1648,8 @@ def _vm_loop(instrs, ip, stack, env, frames, vm_id):
|
|||
_ap = stack.append; _po = stack.pop
|
||||
_isinstance = isinstance; _CP = CompiledProc; _Pr = Proc
|
||||
_ic = {} # inline cache: {instr_idx: (cached_env, cached_val)}
|
||||
while ip < len(instrs):
|
||||
n_instrs = len(instrs)
|
||||
while ip < n_instrs:
|
||||
op, arg = instrs[ip]; ip += 1
|
||||
if op == OP_CONST: _ap(arg)
|
||||
elif op == OP_LOOKUP:
|
||||
|
|
@ -1677,7 +1690,7 @@ def _vm_loop(instrs, ip, stack, env, frames, vm_id):
|
|||
if _isinstance(func, _CP):
|
||||
frames.append((instrs, ip, env, stack))
|
||||
env = func.env.child(func.params, func.rest, args_)
|
||||
instrs = func.code.instrs; ip = 0
|
||||
instrs = func.code.instrs; ip = 0; n_instrs = len(instrs)
|
||||
stack = []; _ap = stack.append; _po = stack.pop
|
||||
continue
|
||||
elif _isinstance(func, _Pr):
|
||||
|
|
@ -1692,7 +1705,7 @@ def _vm_loop(instrs, ip, stack, env, frames, vm_id):
|
|||
func = _po()
|
||||
if _isinstance(func, _CP):
|
||||
env = func.env.child(func.params, func.rest, args_)
|
||||
instrs = func.code.instrs; ip = 0
|
||||
instrs = func.code.instrs; ip = 0; n_instrs = len(instrs)
|
||||
stack.clear()
|
||||
if _portal_checkpoint[0] is not None:
|
||||
_check_portal_checkpoint(instrs, ip, stack, env, frames, vm_id)
|
||||
|
|
@ -1703,20 +1716,20 @@ def _vm_loop(instrs, ip, stack, env, frames, vm_id):
|
|||
for e in body[:-1]: leval(e, c)
|
||||
ret = leval(body[-1], c)
|
||||
if not frames: return ret
|
||||
instrs, ip, env, stack = frames.pop()
|
||||
instrs, ip, env, stack = frames.pop(); n_instrs = len(instrs)
|
||||
_ap = stack.append; _po = stack.pop
|
||||
_ap(ret); continue
|
||||
elif callable(func):
|
||||
ret = func(args_, env)
|
||||
if not frames: return ret
|
||||
instrs, ip, env, stack = frames.pop()
|
||||
instrs, ip, env, stack = frames.pop(); n_instrs = len(instrs)
|
||||
_ap = stack.append; _po = stack.pop
|
||||
_ap(ret); continue
|
||||
else: raise LispErr(f'not callable: {show(func)}')
|
||||
elif op == OP_RETURN:
|
||||
ret = _po() if stack else VOID
|
||||
if not frames: return ret
|
||||
instrs, ip, env, stack = frames.pop()
|
||||
instrs, ip, env, stack = frames.pop(); n_instrs = len(instrs)
|
||||
_ap = stack.append; _po = stack.pop
|
||||
_ap(ret); continue
|
||||
elif op == OP_MAKE_CLOSURE:
|
||||
|
|
@ -1732,7 +1745,7 @@ def _vm_loop(instrs, ip, stack, env, frames, vm_id):
|
|||
if _isinstance(proc, _CP):
|
||||
frames.append((instrs, ip, env, stack))
|
||||
env = proc.env.child(proc.params, proc.rest, [cont])
|
||||
instrs = proc.code.instrs; ip = 0
|
||||
instrs = proc.code.instrs; ip = 0; n_instrs = len(instrs)
|
||||
stack = []; _ap = stack.append; _po = stack.pop
|
||||
continue
|
||||
elif _isinstance(proc, _Pr):
|
||||
|
|
@ -1772,6 +1785,13 @@ def _vm_loop(instrs, ip, stack, env, frames, vm_id):
|
|||
elif op == OP_CONST_EQ_JF:
|
||||
c, addr = arg
|
||||
if _po() != c: ip = addr
|
||||
elif op == OP_SELF_TAIL_CALL:
|
||||
n, params = arg
|
||||
if n: args_ = stack[-n:]; del stack[-n:]
|
||||
else: args_ = []
|
||||
b = env.b
|
||||
for p, a in zip(params, args_): b[p] = a
|
||||
ip = 0; stack.clear(); continue
|
||||
elif op == OP_LOOK_CONST_CALL2:
|
||||
sym, c = arg
|
||||
func = env.lookup(sym)
|
||||
|
|
@ -1779,7 +1799,7 @@ def _vm_loop(instrs, ip, stack, env, frames, vm_id):
|
|||
frames.append((instrs, ip, env, stack))
|
||||
env = func.env.child(func.params, func.rest, [stack[-1], c])
|
||||
del stack[-1:]
|
||||
instrs = func.code.instrs; ip = 0
|
||||
instrs = func.code.instrs; ip = 0; n_instrs = len(instrs)
|
||||
stack = []; _ap = stack.append; _po = stack.pop
|
||||
continue
|
||||
elif callable(func):
|
||||
|
|
@ -1822,6 +1842,9 @@ def _serialize_operand(val):
|
|||
return {'t': 'closure', 'code': _serialize_code(code),
|
||||
'params': [str(p) for p in params],
|
||||
'rest': str(rest) if rest else None}
|
||||
# OP_SELF_TAIL_CALL: (n_args, params_tuple)
|
||||
if len(val) == 2 and isinstance(val[0], int) and isinstance(val[1], tuple):
|
||||
return {'t': 'stc', 'n': val[0], 'p': [str(p) for p in val[1]]}
|
||||
return {'t': 'repr', 'v': repr(val)}
|
||||
|
||||
def _deserialize_operand(data):
|
||||
|
|
@ -1851,6 +1874,8 @@ def _deserialize_operand(data):
|
|||
params = [S(p) for p in data['params']]
|
||||
rest = S(data['rest']) if data['rest'] else None
|
||||
return (code, params, rest)
|
||||
if t == 'stc':
|
||||
return (data['n'], tuple(S(p) for p in data['p']))
|
||||
return data
|
||||
|
||||
def _serialize_code(code):
|
||||
|
|
@ -2228,6 +2253,7 @@ def _disassemble(proc):
|
|||
OP_VEC_REF: 'VEC_REF', OP_VEC_SET: 'VEC_SET',
|
||||
OP_LOOK_LOOK: 'LOOK²', OP_LOOK_ADD1: 'LOOK+1', OP_LOOK_SUB1: 'LOOK-1',
|
||||
OP_CONST_EQ_JF: 'CONST=JF', OP_LOOK_CONST_CALL2: 'LOOK_C_CALL2',
|
||||
OP_SELF_TAIL_CALL: 'SELF_TCALL',
|
||||
}
|
||||
lines = [f'--- {proc.name or "λ"} '
|
||||
f'({" ".join(str(p) for p in proc.params)}'
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue