1339 lines
55 KiB
Plaintext
1339 lines
55 KiB
Plaintext
import t, c
|
||
from stdint import *
|
||
import string
|
||
import viperlib
|
||
import memhub
|
||
import stdio
|
||
import stdlib
|
||
from linkedlist import GSListNode, GSList
|
||
from .__types import LLVMType, ParamNode, TypePrint
|
||
|
||
|
||
# dump 哨兵：IR 链表遍历步数上限（成环检测，fail-fast，与 __module.vp 对齐）。
|
||
# 函数/基本块/指令行链的 Next 被越界写破坏成环时，遍历无限循环挂起；
|
||
# 超过上限即打印链表名并 exit(1)，把"卡死"变成"确定性报错定位"。
|
||
IR_DUMP_CHAIN_LIMIT: t.CDefine = 100000
|
||
|
||
|
||
# ============================================================
|
||
# Line: 指令文本行
|
||
#
|
||
# 每条 LLVM IR 指令序列化为一行文本，通过链表组织。
|
||
# 继承 GSListNode[Line] 获取强类型 Next: Line|CPtr
|
||
#
|
||
# 注意: 必须加 @t.NoVTable。否则 TransPyC 为子类添加 vtable 指针(T*)，
|
||
# 但完全丢失父类 GSListNode[T] 的 Next 字段，导致 struct 布局错位，
|
||
# 所有字段访问读取垃圾值（见续 XV 会话根因分析）。
|
||
# ============================================================
|
||
@t.NoVTable
|
||
class Line(GSListNode[Line]):
|
||
Buf: str # 行文本（NUL 结尾）
|
||
Len: t.CSizeT # 文本长度（不含 NUL）
|
||
|
||
|
||
# ============================================================
|
||
# BasicBlock: 基本块
|
||
#
|
||
# 持有名称 + 指令行链表 + 终止指令标志。
|
||
# 继承 GSListNode[BasicBlock] 获取强类型 Next: BasicBlock|CPtr
|
||
# Lines 为 GSList[Line] 指针，提供 O(1) 追加。
|
||
#
|
||
# 注意: 必须加 @t.NoVTable（理由同 Line）。
|
||
# ============================================================
|
||
@t.NoVTable
|
||
class BasicBlock(GSListNode[BasicBlock]):
|
||
Name: str # 块名（如 "entry"/"if.then"）
|
||
Lines: rowptr[GSList[Line]] # 指令行链表容器（Head/Tail/Count）
|
||
IsTerminated: int # 1=已有终止指令（br/ret）
|
||
|
||
|
||
# ============================================================
|
||
# Param: 函数参数
|
||
#
|
||
# 继承 GSListNode[Param] 获取强类型 Next: Param|CPtr
|
||
# Attrs 存储参数属性文本（如 "noundef"/"nocapture"/"readonly"），None=无属性
|
||
#
|
||
# 注意: 必须加 @t.NoVTable（理由同 Line）。
|
||
# ============================================================
|
||
@t.NoVTable
|
||
class Param(GSListNode[Param]):
|
||
Ty: rowptr[LLVMType] # 参数类型
|
||
Name: str # 参数名（如 "%0"/"%a"）
|
||
Attrs: str # 参数属性文本（None=无属性）
|
||
|
||
|
||
# ============================================================
|
||
# Function: 函数
|
||
#
|
||
# 持有名称 + 返回类型 + 参数链表 + 基本块链表。
|
||
# 继承 GSListNode[Function] 获取强类型 Next: Function|CPtr
|
||
# Params/Blocks 为 GSList 指针，提供 O(1) 追加。
|
||
# Attrs 存储函数级属性文本（如 "nounwind"/"noinline"），None=无属性
|
||
#
|
||
# 注意: 必须加 @t.NoVTable（理由同 Line）。
|
||
# ============================================================
|
||
@t.NoVTable
|
||
class Function(GSListNode[Function]):
|
||
Name: str # 函数名
|
||
RetTy: rowptr[LLVMType] # 返回类型
|
||
Params: rowptr[GSList[Param]] # 参数链表容器
|
||
Blocks: rowptr[GSList[BasicBlock]] # 基本块链表容器
|
||
IsDeclared: int # 1=仅声明(declare), 0=定义(define)
|
||
IsVarArg: int # 1=变参函数(...)
|
||
Attrs: str # 函数属性文本（None=无属性）
|
||
AttrGroup: str # 字符串属性组内容（如 '"section"=".text.startup"'，None=无）
|
||
AttrGroupId: int # 属性组编号（>0 时函数行输出 #N，模块尾输出 attributes #N）
|
||
|
||
|
||
# ============================================================
|
||
# 工厂函数
|
||
# ============================================================
|
||
# Line 硬编码大小: Next(8) + Buf(8) + Len(8) = 24 字节
|
||
LINE_SIZE: t.CDefine = 24
|
||
|
||
def new_line(pool: rowptr[memhub.MemBuddy], text: str) -> rowptr[Line]:
|
||
"""创建一个指令行，复制 text 到新分配的缓冲区"""
|
||
ptr: rowptr[Line] = pool.alloc(LINE_SIZE)
|
||
if ptr is None: return None
|
||
string.memset(ptr, 0, LINE_SIZE)
|
||
if text is not None:
|
||
slen: t.CSizeT = string.strlen(text)
|
||
buf: str = pool.alloc(slen + 1)
|
||
if buf is not None:
|
||
string.strcpy(buf, text)
|
||
ptr.Buf = buf
|
||
ptr.Len = slen
|
||
return ptr
|
||
|
||
|
||
def _new_gslist(pool: rowptr[memhub.MemBuddy], list_size: t.CSizeT) -> t.CPtr:
|
||
"""从 pool 分配并零初始化一个 GSList 容器（Head/Tail/Count 全零）"""
|
||
gsl: t.CPtr = pool.alloc(list_size)
|
||
if gsl is None: return None
|
||
string.memset(gsl, 0, list_size)
|
||
return gsl
|
||
|
||
|
||
# BasicBlock 硬编码大小: Next(8) + Name(8) + Lines(8) + IsTerminated(4) = 32 字节
|
||
BLOCK_SIZE: t.CDefine = 32
|
||
|
||
def new_basic_block(pool: rowptr[memhub.MemBuddy], name: str) -> rowptr[BasicBlock]:
|
||
"""创建一个基本块"""
|
||
ptr: rowptr[BasicBlock] = pool.alloc(BLOCK_SIZE)
|
||
if ptr is None: return None
|
||
string.memset(ptr, 0, BLOCK_SIZE)
|
||
ptr.Name = name
|
||
ptr.Lines = _new_gslist(pool, 24) # GSList: Head+Tail+Count
|
||
ptr.IsTerminated = 0
|
||
return ptr
|
||
|
||
|
||
# Param 硬编码大小: Next(8) + Ty(8) + Name(8) + Attrs(8) = 32 字节
|
||
PARAM_SIZE: t.CDefine = 32
|
||
|
||
def new_param(pool: rowptr[memhub.MemBuddy], ty: rowptr[LLVMType],
|
||
name: str) -> rowptr[Param]:
|
||
"""创建一个函数参数"""
|
||
ptr: rowptr[Param] = pool.alloc(PARAM_SIZE)
|
||
if ptr is None: return None
|
||
string.memset(ptr, 0, PARAM_SIZE)
|
||
ptr.Ty = ty
|
||
ptr.Name = name
|
||
ptr.Attrs = None
|
||
return ptr
|
||
|
||
|
||
# Function 硬编码大小: Next(8) + Name(8) + RetTy(8) + Params(8) + Blocks(8) + IsDeclared(4) + IsVarArg(4) + Attrs(8) + AttrGroup(8) + AttrGroupId(8) = 72 字节
|
||
FUNC_SIZE: t.CDefine = 72
|
||
|
||
def new_function(pool: rowptr[memhub.MemBuddy], name: str,
|
||
ret_ty: rowptr[LLVMType]) -> rowptr[Function]:
|
||
"""创建一个函数"""
|
||
ptr: rowptr[Function] = pool.alloc(FUNC_SIZE)
|
||
if ptr is None: return None
|
||
string.memset(ptr, 0, FUNC_SIZE)
|
||
ptr.Name = name
|
||
ptr.RetTy = ret_ty
|
||
ptr.Params = _new_gslist(pool, 24) # GSList: Head+Tail+Count
|
||
ptr.Blocks = _new_gslist(pool, 24)
|
||
ptr.IsDeclared = 0
|
||
ptr.Attrs = None
|
||
ptr.AttrGroup = None
|
||
ptr.AttrGroupId = 0
|
||
return ptr
|
||
|
||
|
||
# ============================================================
|
||
# 链表操作（委托 GSList.append）
|
||
# ============================================================
|
||
def function_add_param(func: rowptr[Function], param: rowptr[Param]):
|
||
"""将参数追加到函数参数链表"""
|
||
if func is None or param is None: return
|
||
# 注意：必须先用局部变量 + 类型参数（对齐 function_move_block_to_end 注释），
|
||
# 链式 func.Params.append() 会被编译器解析成 LinkedNode.append 且 this 前移
|
||
# sizeof(LinkedNode)（兄弟 Mixin 兜底误匹配），写穿容器前 48 字节破坏数据结构。
|
||
params: rowptr[GSList[Param]] = func.Params
|
||
if params is None: return
|
||
params.append(param)
|
||
|
||
|
||
def function_add_block(func: rowptr[Function], block: rowptr[BasicBlock]):
|
||
"""将基本块追加到函数块链表"""
|
||
if func is None or block is None: return
|
||
blocks: rowptr[GSList[BasicBlock]] = func.Blocks
|
||
if blocks is None: return
|
||
blocks.append(block)
|
||
|
||
|
||
def function_move_block_to_end(func: rowptr[Function], block: rowptr[BasicBlock]):
|
||
"""将 block 移动到函数块链表末尾（保证 BB 文本顺序与控制流顺序一致）。
|
||
|
||
用于解决 SSA 名非单调问题：BB 按创建顺序排列，但指令按控制流顺序生成，
|
||
导致 SSA 编号在文本输出时非单调递增，llc 报错。
|
||
在 position_at_end 时调用此函数，使 BB 顺序与 position 顺序一致。
|
||
|
||
注意：必须使用 GSList[BasicBlock] 带类型参数，否则 TransPyC 创建不完整
|
||
struct 类型，导致 blocks.Head/Tail/Count 字段访问读取垃圾值。
|
||
"""
|
||
if func is None or block is None: return
|
||
blocks: rowptr[GSList[BasicBlock]] = func.Blocks
|
||
if blocks is None: return
|
||
|
||
# 已经是 Tail，无需移动
|
||
if blocks.Tail is block:
|
||
return
|
||
|
||
# 从链表中移除 block
|
||
if blocks.Head is block:
|
||
# block 是 Head，直接后移 Head
|
||
blocks.Head = block.Next
|
||
else:
|
||
# 遍历找前驱
|
||
prev: rowptr[BasicBlock] = blocks.Head
|
||
while prev is not None and prev.Next is not block:
|
||
prev = prev.Next
|
||
if prev is None:
|
||
# block 不在链表中，直接 append
|
||
blocks.append(block)
|
||
return
|
||
prev.Next = block.Next
|
||
|
||
# block.Next 由移除逻辑处理，重置为 None
|
||
block.Next = None
|
||
|
||
# append 到末尾（不调用 blocks.append 以避免 Count 重复递增）
|
||
if blocks.Tail is not None:
|
||
blocks.Tail.Next = block
|
||
blocks.Tail = block
|
||
|
||
|
||
def block_append_line(block: rowptr[BasicBlock], line: rowptr[Line]):
|
||
"""将指令行追加到基本块"""
|
||
if block is None or line is None: return
|
||
lines: rowptr[GSList[Line]] = block.Lines
|
||
if lines is None: return
|
||
lines.append(line)
|
||
|
||
|
||
def block_set_terminated(block: rowptr[BasicBlock], val: int):
|
||
"""设置块的终止标志（供跨模块调用，避免链式属性访问的 opaque 类型问题）"""
|
||
if block is None: return
|
||
block.IsTerminated = val
|
||
|
||
|
||
def block_get_terminated(block: rowptr[BasicBlock]) -> int:
|
||
"""获取块的终止标志（供跨模块调用，避免链式属性访问的 opaque 类型问题）"""
|
||
if block is None:
|
||
return 1
|
||
return block.IsTerminated
|
||
|
||
|
||
def block_append_text(pool: rowptr[memhub.MemBuddy], block: rowptr[BasicBlock],
|
||
text: str):
|
||
"""创建指令行并追加到基本块"""
|
||
if block is None or text is None: return
|
||
line: rowptr[Line] = new_line(pool, text)
|
||
if line is not None:
|
||
block_append_line(block, line)
|
||
|
||
|
||
def block_insert_text_before_terminator(pool: rowptr[memhub.MemBuddy],
|
||
block: rowptr[BasicBlock],
|
||
text: str):
|
||
"""创建指令行并插入到基本块终止指令之前
|
||
|
||
如果块未终止，追加到末尾。
|
||
如果块已终止（有 br/ret），在终止指令之前插入，确保 alloca 等指令
|
||
不出现在终止指令之后。
|
||
"""
|
||
if block is None or text is None: return
|
||
if block.Lines is None: return
|
||
line: rowptr[Line] = new_line(pool, text)
|
||
if line is None: return
|
||
# 块未终止：直接追加
|
||
if block.IsTerminated == 0:
|
||
block_append_line(block, line)
|
||
return
|
||
# 块已终止：在 Tail（终止指令）之前插入
|
||
lines: rowptr[GSList[Line]] = block.Lines
|
||
if lines.Head is None:
|
||
block_append_line(block, line)
|
||
return
|
||
# 只有一个节点：在 Head 之前插入
|
||
if lines.Head is lines.Tail:
|
||
line.Next = lines.Head
|
||
lines.Head = line
|
||
lines.Count += 1
|
||
return
|
||
# 找到 Tail 的前驱节点
|
||
prev: rowptr[Line] = lines.Head
|
||
while prev is not None and prev.Next is not lines.Tail:
|
||
prev = prev.Next
|
||
if prev is None:
|
||
# 找不到前驱，回退到追加
|
||
block_append_line(block, line)
|
||
return
|
||
# 在 prev 和 Tail 之间插入 line
|
||
line.Next = lines.Tail
|
||
prev.Next = line
|
||
lines.Count += 1
|
||
|
||
|
||
# ============================================================
|
||
# 属性设置
|
||
# ============================================================
|
||
def function_set_attrs(func: rowptr[Function], attrs: str):
|
||
"""设置函数级属性文本（如 "nounwind"/"noinline"）"""
|
||
if func is None: return
|
||
func.Attrs = attrs
|
||
|
||
|
||
_g_attr_group_counter: int = 0
|
||
|
||
def function_set_attr_group(func: rowptr[Function], group: str):
|
||
"""设置函数字符串属性组（如 '"section"=".text.startup"'）并分配组编号 #N。
|
||
|
||
LLVM 字符串属性（带值的，如 section）不能内联在函数签名行，
|
||
必须用属性组：函数行输出 #N，组定义输出在函数尾。
|
||
"""
|
||
global _g_attr_group_counter
|
||
if func is None or group is None:
|
||
return
|
||
func.AttrGroup = group
|
||
_g_attr_group_counter += 1
|
||
func.AttrGroupId = _g_attr_group_counter
|
||
|
||
|
||
def param_set_attrs(param: rowptr[Param], attrs: str):
|
||
"""设置参数属性文本（如 "noundef"/"nocapture"/"readonly"）"""
|
||
if param is None: return
|
||
param.Attrs = attrs
|
||
|
||
|
||
# ============================================================
|
||
# _ll_name_needs_quote - 检查 LLVM 标识符是否需要加引号
|
||
#
|
||
# LLVM IR 规范：标识符合法字符集为 [-a-zA-Z$._][-a-zA-Z$._0-9]*
|
||
# 以数字开头或含非合法字符（如 '['、']'、' '、'@' 等）时必须用双引号包围。
|
||
# '.' 是合法字符，不需要加引号（如 @.str、@llvm.memcpy 都不加引号）。
|
||
# SHA1 命名空间的函数名（如 "4dad2ba72ed7ba09.GSListNode[Param]"）含 '['、']'，
|
||
# 需要加引号；但纯 hex+点 的名字（如 "abc.def"）不需要。
|
||
# ============================================================
|
||
def _ll_name_needs_quote(name: str) -> int:
|
||
"""检查 LLVM 标识符是否需要加引号（以数字开头或含非合法字符）
|
||
|
||
LLVM IR 合法字符：a-z, A-Z, 0-9, -, $, ., _
|
||
首字符是数字时需要加引号（LLVM 要求局部/全局标识符首字符非数字，除非加引号）。
|
||
含其他字符（如 '['、']'、空格）时需要加引号。
|
||
"""
|
||
if name is None:
|
||
return 0
|
||
c0: t.CChar = name[0]
|
||
if c0 >= '0' and c0 <= '9':
|
||
return 1
|
||
name_len: t.CSizeT = string.strlen(name)
|
||
i: t.CSizeT = 0
|
||
while i < name_len:
|
||
ch: t.CChar = name[i]
|
||
# 合法字符：a-z, A-Z, 0-9, -, $, ., _
|
||
if not ((ch >= 'a' and ch <= 'z') or (ch >= 'A' and ch <= 'Z') or
|
||
(ch >= '0' and ch <= '9') or ch == '-' or ch == '$' or
|
||
ch == '.' or ch == '_'):
|
||
return 1
|
||
i += 1
|
||
return 0
|
||
|
||
|
||
# ============================================================
|
||
# RenumberSSA: 重编号函数内数字 SSA 名
|
||
#
|
||
# 解决 build_alloca_at_entry 在入口块终止指令前插入 alloca 导致
|
||
# SSA 名在文本输出中非单调递增的问题（llc 报错:
|
||
# "instruction expected to be numbered '%N' or greater"）。
|
||
#
|
||
# 现象: entry 块原指令 %0-%6 + br，alloca_at_entry 插入 %28/%41 到 br 之前，
|
||
# 文本输出顺序为 entry(%0..%6, %28, %41) → if.then(%7..%27)，
|
||
# 导致 %28 出现在 %7 之前，llc 拒绝编译。
|
||
#
|
||
# 算法:
|
||
# 1. 扫描所有 BB 指令行，按文本输出顺序收集数字 SSA 名 %N
|
||
# 2. 按首次出现顺序建立 old→new 映射（从 0 递增）
|
||
# 3. 替换所有指令行文本中的 %N 为新编号
|
||
#
|
||
# 注意:
|
||
# - 仅重编号纯数字 SSA 名（%0, %1, ...）
|
||
# - 命名 SSA 名（%entry, %.tmp, %self）、块标签、全局符号 @N 不重编号
|
||
# - 仅对 define（有函数体）调用，declare 无函数体无需处理
|
||
# - 当前 TransPyV 参数多为命名（%self/%env/%n），非数字，无需重编号参数
|
||
# ============================================================
|
||
|
||
# 判定字符是否为十进制数字
|
||
def _is_digit_ch(ch: t.CChar) -> int:
|
||
"""判断字符是否是 '0'-'9'
|
||
|
||
ch 是值类型 (t.CChar = i8)，TPC 已修复值类型 is None 处理：
|
||
值类型 is None 恒为 false（生成 i1 0 常量），不会解引用空指针。
|
||
"""
|
||
if ch is None:
|
||
return 0
|
||
if ch >= '0' and ch <= '9':
|
||
return 1
|
||
return 0
|
||
|
||
|
||
# ============================================================
|
||
# i32 数组的安全访问（显式字节偏移）
|
||
#
|
||
# TransPyC 的 ptr[i] 指针索引可能被当作字节索引（i8*）而非 i32* 索引，
|
||
# 导致内存损坏。必须显式计算字节偏移：addr = base + idx * 4，再转型为 i32*。
|
||
# 参见 HandlesTranslator.py:358 的 t.CUInt64T(ptr) + offset * sizeof 用法。
|
||
# ============================================================
|
||
def _map_get_i32(map_arr: str, idx: int) -> int:
|
||
"""读取 i32 数组 map_arr[idx]（显式字节偏移 idx*4）
|
||
|
||
使用 (rowptr[int])(uint64) 强转将字节偏移转为 i32*，再用 c.Deref 解引用。
|
||
TPC 已修复类型联合强转（_HandleTypeUnionCastLlvm 添加 fallback），此写法可用。
|
||
"""
|
||
if map_arr is None:
|
||
return -1
|
||
byte_off: t.CUInt64T = t.CUInt64T(map_arr) + idx * 4
|
||
elem_ptr: rowptr[int] = (rowptr[int])(byte_off)
|
||
return c.Deref(elem_ptr)
|
||
|
||
|
||
def _map_set_i32(map_arr: str, idx: int, val: int):
|
||
"""写入 map_arr[idx] = val（i32 数组，显式字节偏移 idx*4）
|
||
|
||
使用 (rowptr[int])(uint64) 强转 + c.DerefAs 解引用赋值。
|
||
TPC 已修复类型联合强转，此写法可用。
|
||
"""
|
||
if map_arr is None:
|
||
return
|
||
byte_off: t.CUInt64T = t.CUInt64T(map_arr) + idx * 4
|
||
elem_ptr: rowptr[int] = (rowptr[int])(byte_off)
|
||
c.DerefAs(elem_ptr, val)
|
||
|
||
|
||
def _scan_max_ssa_in_text(text: str, cur_max: int) -> int:
|
||
"""扫描文本，找出 %N（N 为数字）中最大的 N，与 cur_max 比较返回更大值
|
||
|
||
仅匹配 '%' 后紧跟数字的情况，跳过命名 SSA 名（%entry 等）。
|
||
"""
|
||
result: int = cur_max
|
||
p: str = text
|
||
while p is not None and p[0] != '\0':
|
||
if p[0] == '%':
|
||
nxt: str = p + 1
|
||
if nxt is not None and _is_digit_ch(nxt[0]) != 0:
|
||
num: int = string.atoi(nxt)
|
||
if num > result:
|
||
result = num
|
||
p = p + 1
|
||
return result
|
||
|
||
|
||
def _find_max_ssa_number(func: rowptr[Function]) -> int:
|
||
"""遍历所有 BB 指令行，找出最大的 %N 数字；无数字 SSA 名返回 -1"""
|
||
max_n: int = -1
|
||
if func is None or func.Blocks is None:
|
||
return max_n
|
||
blocks: rowptr[GSList[BasicBlock]] = func.Blocks
|
||
blk: rowptr[BasicBlock] = blocks.Head
|
||
blk_steps: t.CSizeT = 0
|
||
while blk is not None:
|
||
blk_steps += 1
|
||
if blk_steps > IR_DUMP_CHAIN_LIMIT:
|
||
stdio.printf("[IR-DUMP] 链表成环: func blocks（函数名=%s，ssa 扫描终止）\n", func.Name)
|
||
stdlib.exit(1)
|
||
if blk.Lines is not None:
|
||
lines: rowptr[GSList[Line]] = blk.Lines
|
||
line: rowptr[Line] = lines.Head
|
||
line_steps: t.CSizeT = 0
|
||
while line is not None:
|
||
line_steps += 1
|
||
if line_steps > IR_DUMP_CHAIN_LIMIT:
|
||
stdio.printf("[IR-DUMP] 链表成环: block lines（函数名=%s，块=%s，ssa 扫描终止）\n",
|
||
func.Name, blk.Name)
|
||
stdlib.exit(1)
|
||
if line.Buf is not None:
|
||
max_n = _scan_max_ssa_in_text(line.Buf, max_n)
|
||
line = line.Next
|
||
blk = blk.Next
|
||
return max_n
|
||
|
||
|
||
def _assign_new_numbers(text: str,
|
||
map_arr: str,
|
||
max_n: int, next_new: int) -> int:
|
||
"""扫描文本，对每个 %N（数字），若 map_arr[N]==-1 则分配 next_new
|
||
|
||
返回更新后的 next_new。按文本首次出现顺序建立映射。
|
||
"""
|
||
result: int = next_new
|
||
p: str = text
|
||
while p is not None and p[0] != '\0':
|
||
if p[0] == '%':
|
||
nxt: str = p + 1
|
||
if nxt is not None and _is_digit_ch(nxt[0]) != 0:
|
||
num: int = string.atoi(nxt)
|
||
if num >= 0 and num <= max_n:
|
||
if _map_get_i32(map_arr, num) == -1:
|
||
_map_set_i32(map_arr, num, result)
|
||
result += 1
|
||
p = p + 1
|
||
return result
|
||
|
||
|
||
def _replace_ssa_numbers(text: str,
|
||
pool: rowptr[memhub.MemBuddy],
|
||
map_arr: str,
|
||
max_n: int) -> str:
|
||
"""构建新文本，将 %N 替换为映射后的新编号
|
||
|
||
逐字符扫描：遇到 %N（数字）查映射表替换，其他字符原样复制。
|
||
返回新分配的缓冲区（调用方无需释放旧 Buf，pool 统一管理）。
|
||
"""
|
||
src_len: t.CSizeT = string.strlen(text)
|
||
# 新缓冲区：最坏情况每个数字变长（位数增加），分配 src_len*2 + 64
|
||
new_size: t.CSizeT = src_len * 2 + 64
|
||
new_buf: str = pool.alloc(new_size)
|
||
if new_buf is None:
|
||
return None
|
||
# 一次性分配数字缓冲区（循环中复用）
|
||
num_buf: str = pool.alloc(16)
|
||
if num_buf is None:
|
||
new_buf[0] = '\0'
|
||
return new_buf
|
||
|
||
dst_idx: t.CSizeT = 0
|
||
src_p: str = text
|
||
while src_p is not None and src_p[0] != '\0':
|
||
if dst_idx + 16 >= new_size:
|
||
# 缓冲区不足，截断
|
||
break
|
||
if src_p[0] == '%':
|
||
nxt: str = src_p + 1
|
||
if nxt is not None and _is_digit_ch(nxt[0]) != 0:
|
||
num: int = string.atoi(nxt)
|
||
# 先做范围检查再读 map_arr：num 超 max_n 时 map_arr 只有
|
||
# (max_n+1)*4 字节，先读后查会越界读远端内存
|
||
if num >= 0 and num <= max_n:
|
||
new_num: int = _map_get_i32(map_arr, num)
|
||
if new_num != -1:
|
||
# 写入 '%'
|
||
new_buf[dst_idx] = '%'
|
||
dst_idx += 1
|
||
# 写入新编号
|
||
num_buf[0] = '\0'
|
||
viperlib.snprintf(num_buf, 16, "%d", new_num)
|
||
k: t.CSizeT = 0
|
||
while num_buf[k] != '\0' and dst_idx + 1 < new_size:
|
||
new_buf[dst_idx] = num_buf[k]
|
||
dst_idx += 1
|
||
k += 1
|
||
# 跳过旧数字字符
|
||
src_p = src_p + 1
|
||
while src_p is not None and _is_digit_ch(src_p[0]) != 0:
|
||
src_p = src_p + 1
|
||
continue
|
||
# 普通字符：直接复制
|
||
new_buf[dst_idx] = src_p[0]
|
||
dst_idx += 1
|
||
src_p = src_p + 1
|
||
|
||
new_buf[dst_idx] = '\0'
|
||
return new_buf
|
||
|
||
|
||
def RenumberSSA(func: rowptr[Function], pool: rowptr[memhub.MemBuddy]):
|
||
"""重编号函数内所有数字 SSA 名，确保按文本输出顺序单调递增
|
||
|
||
在 FunctionPrint 序列化函数体之前调用。
|
||
仅对 define（IsDeclared==0）且 Blocks 非空 的函数生效。
|
||
"""
|
||
if func is None or pool is None:
|
||
return
|
||
if func.IsDeclared == 1:
|
||
return # declare 无函数体
|
||
if func.Blocks is None:
|
||
return
|
||
|
||
# 1. 第一遍扫描：找出最大 SSA 编号
|
||
max_n: int = _find_max_ssa_number(func)
|
||
if max_n < 0:
|
||
return # 无数字 SSA 名
|
||
|
||
# 2. 分配并初始化映射数组 map_arr[0..max_n] = -1
|
||
# 用 i8* 存储（pool.alloc 返回 i8*），通过 _map_get_i32/_map_set_i32 访问。
|
||
# 初始化为 -1：每个 i32 元素 4 字节，0xFFFFFFFF = -1（有符号 i32）。
|
||
# 用 memset 一次性填充，避免循环内 ptr[i] 指针索引问题。
|
||
map_bytes: t.CSizeT = (max_n + 1) * 4 # int = 4 字节
|
||
map_arr: str = pool.alloc(map_bytes)
|
||
if map_arr is None:
|
||
return
|
||
string.memset(map_arr, 0xFF, map_bytes) # 0xFF * 4 = 0xFFFFFFFF = -1
|
||
|
||
# 3. 第二遍扫描：按文本顺序分配新编号
|
||
next_new: int = 0
|
||
blocks: rowptr[GSList[BasicBlock]] = func.Blocks
|
||
blk: rowptr[BasicBlock] = blocks.Head
|
||
blk_steps: t.CSizeT = 0
|
||
while blk is not None:
|
||
blk_steps += 1
|
||
if blk_steps > IR_DUMP_CHAIN_LIMIT:
|
||
stdio.printf("[IR-DUMP] 链表成环: func blocks（函数名=%s，ssa 重编号终止）\n", func.Name)
|
||
stdlib.exit(1)
|
||
if blk.Lines is not None:
|
||
lines: rowptr[GSList[Line]] = blk.Lines
|
||
line: rowptr[Line] = lines.Head
|
||
line_steps: t.CSizeT = 0
|
||
while line is not None:
|
||
line_steps += 1
|
||
if line_steps > IR_DUMP_CHAIN_LIMIT:
|
||
stdio.printf("[IR-DUMP] 链表成环: block lines（函数名=%s，块=%s，ssa 重编号终止）\n",
|
||
func.Name, blk.Name)
|
||
stdlib.exit(1)
|
||
if line.Buf is not None:
|
||
next_new = _assign_new_numbers(line.Buf, map_arr, max_n, next_new)
|
||
line = line.Next
|
||
blk = blk.Next
|
||
|
||
# 4. 检查是否有需要重编号的（old != new）
|
||
need_renumber: int = 0
|
||
i: int = 0
|
||
while i <= max_n:
|
||
v: int = _map_get_i32(map_arr, i)
|
||
if v != -1 and v != i:
|
||
need_renumber = 1
|
||
break
|
||
i += 1
|
||
if need_renumber == 0:
|
||
return # 编号已单调，无需替换
|
||
|
||
# 5. 第三遍扫描：替换所有指令行文本
|
||
blk = blocks.Head
|
||
blk_steps = 0
|
||
while blk is not None:
|
||
blk_steps += 1
|
||
if blk_steps > IR_DUMP_CHAIN_LIMIT:
|
||
stdio.printf("[IR-DUMP] 链表成环: func blocks（函数名=%s，ssa 替换终止）\n", func.Name)
|
||
stdlib.exit(1)
|
||
if blk.Lines is not None:
|
||
lines: rowptr[GSList[Line]] = blk.Lines
|
||
line: rowptr[Line] = lines.Head
|
||
line_steps = 0
|
||
while line is not None:
|
||
line_steps += 1
|
||
if line_steps > IR_DUMP_CHAIN_LIMIT:
|
||
stdio.printf("[IR-DUMP] 链表成环: block lines（函数名=%s，块=%s，ssa 替换终止）\n",
|
||
func.Name, blk.Name)
|
||
stdlib.exit(1)
|
||
if line.Buf is not None:
|
||
new_buf: str = _replace_ssa_numbers(
|
||
line.Buf, pool, map_arr, max_n)
|
||
if new_buf is not None:
|
||
line.Buf = new_buf
|
||
line.Len = string.strlen(new_buf)
|
||
line = line.Next
|
||
blk = blk.Next
|
||
|
||
|
||
# ============================================================
|
||
# 本地窥孔优化（文本层，输出前执行，保持语义）
|
||
# T1: ret <iN> zeroinitializer → ret <iN> 0；ret <ptr*> zeroinitializer → null
|
||
# T2: 入口 alloca 槽：唯一一次 store 且值无副作用（形参/%SSA/整数/浮点字面量）
|
||
# 时，把 store→load 转发为直接使用该值，删除 alloca/store/load 冗余行
|
||
# （"虚拟寄存器优化"，如 printf(i8* %msg) 前不再 alloca+store+load）。
|
||
# 安全性：仅当 alloca 与唯一 store 都在 entry、loads 任意、无其它取址/别名用途。
|
||
# ============================================================
|
||
def _peep_is_startwith(s: str, pfx: str) -> int:
|
||
if s is None or pfx is None:
|
||
return 0
|
||
pl: t.CSizeT = string.strlen(pfx)
|
||
if pl == 0:
|
||
return 1
|
||
if string.strlen(s) < pl:
|
||
return 0
|
||
return 1 if string.strncmp(s, pfx, pl) == 0 else 0
|
||
|
||
|
||
def _peep_replace_token(pool: rowptr[memhub.MemBuddy],
|
||
line_buf: str, tok: str, repl: str) -> str:
|
||
"""将 line_buf 中独立记号 %<tok> 替换为 repl，返回新缓冲（无变化返回原指针）"""
|
||
if line_buf is None or tok is None or repl is None:
|
||
return line_buf
|
||
tlen: t.CSizeT = string.strlen(tok)
|
||
if tlen == 0:
|
||
return line_buf
|
||
rlen: t.CSizeT = string.strlen(repl)
|
||
ln: t.CSizeT = string.strlen(line_buf)
|
||
# 预扫是否命中
|
||
hit: int = 0
|
||
i: t.CSizeT = 0
|
||
while i + tlen <= ln:
|
||
if line_buf[i] == '%' and string.strncmp(line_buf + i + 1, tok, tlen) == 0:
|
||
nxt: int = 0
|
||
if i + 1 + tlen < ln:
|
||
cc: int = line_buf[i + 1 + tlen]
|
||
if (cc >= 48 and cc <= 57) or (cc >= 65 and cc <= 90) or (cc >= 97 and cc <= 122) or cc == 95:
|
||
nxt = 1
|
||
if nxt == 0:
|
||
hit = 1
|
||
break
|
||
i += 1
|
||
if hit == 0:
|
||
return line_buf
|
||
# 缓冲大小上界：单行最多 ⌊ln/2⌋ 次替换，每次去掉 ≥2 字节（%+名）换入 rlen 字节；
|
||
# 最坏输出 ≤ ln + (ln/2)·rlen + NUL。ln + 1 + rlen·(ln+1) 恒覆盖该上界，
|
||
# 避免 K≥2 且替换串更长时把新缓冲写穿（随机内存破坏）。
|
||
out: str = pool.alloc(ln + 1 + rlen * (ln + 1))
|
||
if out is None:
|
||
return line_buf
|
||
op: t.CSizeT = 0
|
||
oq: t.CSizeT = 0
|
||
i = 0
|
||
while i < ln:
|
||
if line_buf[i] == '%' and i + 1 + tlen <= ln \
|
||
and string.strncmp(line_buf + i + 1, tok, tlen) == 0:
|
||
nxt2: int = 0
|
||
if i + 1 + tlen < ln:
|
||
cc2: int = line_buf[i + 1 + tlen]
|
||
if (cc2 >= 48 and cc2 <= 57) or (cc2 >= 65 and cc2 <= 90) or (cc2 >= 97 and cc2 <= 122) or cc2 == 95:
|
||
nxt2 = 1
|
||
if nxt2 == 0:
|
||
j: t.CSizeT = 0
|
||
while j < rlen:
|
||
out[oq] = repl[j]
|
||
oq += 1
|
||
j += 1
|
||
i += 1 + tlen
|
||
continue
|
||
out[oq] = line_buf[i]
|
||
oq += 1
|
||
i += 1
|
||
out[oq] = '\0'
|
||
return out
|
||
|
||
|
||
def _peep_last_ref_eq(sb: str, a_name: str) -> int:
|
||
"""行内最后一个 '%name' 记号是否等于 a_name（store/load 的指针操作数在行尾）"""
|
||
if sb is None or a_name is None:
|
||
return 0
|
||
r: str = None
|
||
probe: str = sb
|
||
while True:
|
||
hit: str = string.strchr(probe, '%')
|
||
if hit is None:
|
||
break
|
||
r = hit
|
||
probe = hit + 1
|
||
if r is None:
|
||
return 0
|
||
nl: t.CSizeT = string.strlen(a_name)
|
||
if string.strncmp(r + 1, a_name, nl) == 0:
|
||
nxtc: int = r[1 + nl]
|
||
if nxtc == 0 or nxtc == 32 or nxtc == 44 or nxtc == 41:
|
||
return 1
|
||
return 0
|
||
|
||
|
||
def _peep_alloc_name(eb: str, pool: rowptr[memhub.MemBuddy]) -> str:
|
||
"""从 '%name = alloca ...' 提取 name（不含 %），失败返回 None"""
|
||
if eb is None or eb[0] != '%':
|
||
return None
|
||
eqpos: str = string.strstr(eb, " = alloca ")
|
||
if eqpos is None:
|
||
return None
|
||
nlen: t.CSizeT = t.CSizeT(t.CUInt64T(eqpos) - t.CUInt64T(eb)) - 1
|
||
if nlen == 0:
|
||
return None
|
||
nm: str = pool.alloc(nlen + 1)
|
||
if nm is None:
|
||
return None
|
||
string.strncpy(nm, eb + 1, nlen)
|
||
nm[nlen] = '\0'
|
||
return nm
|
||
|
||
|
||
def _peephole_opt(func: rowptr[Function], pool: rowptr[memhub.MemBuddy]):
|
||
"""文本层本地窥孔：T1 ret zeroinitializer；T2 入口槽 store→load 转发"""
|
||
if func is None or func.Blocks is None or pool is None:
|
||
return
|
||
# 注：TPC 对 "变量 = 指针.字段.成员" 两级自动解引用链会静默丢弃赋值
|
||
# （编译出的 IR 直接读未初始化栈变量 → 随机栈错误）。必须拆成单级：
|
||
# 先取字段到局部 rowptr，再判空后取 .Head（与 FunctionPrint 同一模式）。
|
||
func_blocks: rowptr[GSList[BasicBlock]] = func.Blocks
|
||
blk: rowptr[BasicBlock] = None
|
||
if func_blocks is not None:
|
||
blk = func_blocks.Head
|
||
|
||
# ---------- T1: ret <ty> zeroinitializer ----------
|
||
blk_t1 = blk
|
||
while blk_t1 is not None:
|
||
line1: rowptr[Line] = None
|
||
if blk_t1.Lines is not None:
|
||
t1_lines: rowptr[GSList[Line]] = blk_t1.Lines
|
||
line1 = t1_lines.Head
|
||
while line1 is not None:
|
||
if line1.Buf is not None:
|
||
zi: str = string.strstr(line1.Buf, " zeroinitializer")
|
||
if zi is not None and _peep_is_startwith(line1.Buf, "ret ") != 0:
|
||
ty_len: t.CSizeT = t.CSizeT(t.CUInt64T(zi) - t.CUInt64T(line1.Buf)) - 4
|
||
if ty_len > 0:
|
||
ty_buf: str = pool.alloc(ty_len + 1)
|
||
if ty_buf is not None:
|
||
string.strncpy(ty_buf, line1.Buf + 4, ty_len)
|
||
ty_buf[ty_len] = '\0'
|
||
lit: str = None
|
||
if ty_buf[0] == 'i':
|
||
lit = "0"
|
||
elif string.strchr(ty_buf, 42) is not None: # '*'
|
||
lit = "null"
|
||
if lit is not None:
|
||
nb: str = pool.alloc(ty_len + 16)
|
||
if nb is not None:
|
||
nb[0] = '\0'
|
||
string.strcat(nb, "ret ")
|
||
string.strcat(nb, ty_buf)
|
||
string.strcat(nb, " ")
|
||
string.strcat(nb, lit)
|
||
line1.Buf = nb
|
||
line1 = line1.Next
|
||
blk_t1 = blk_t1.Next
|
||
|
||
# ---------- T2: entry alloca 槽转发 ----------
|
||
entry_blk: rowptr[BasicBlock] = blk
|
||
if entry_blk is None:
|
||
return
|
||
entry_lines: rowptr[GSList[Line]] = entry_blk.Lines
|
||
if entry_lines is None:
|
||
return
|
||
e_line: rowptr[Line] = entry_lines.Head
|
||
while e_line is not None:
|
||
eb: str = e_line.Buf
|
||
if eb is not None:
|
||
a_name: str = _peep_alloc_name(eb, pool)
|
||
if a_name is not None:
|
||
st_val: str = None
|
||
st_ok: int = 0
|
||
st_count: int = 0
|
||
st_line_order: int = -1
|
||
load_ltoks: str = pool.alloc(8 * 64)
|
||
load_cnt: int = 0
|
||
load_entry_order: t.CInt64T = 0
|
||
alloca_order: int = -1
|
||
order: int = 0
|
||
unsafe: int = 0
|
||
# 第一遍：统计本 alloca 的 store/load/其它引用及行序
|
||
blk_s = blk
|
||
while blk_s is not None:
|
||
l_s: rowptr[Line] = None
|
||
if blk_s.Lines is not None:
|
||
scan_lines: rowptr[GSList[Line]] = blk_s.Lines
|
||
l_s = scan_lines.Head
|
||
while l_s is not None:
|
||
sb: str = l_s.Buf
|
||
if sb is not None:
|
||
is_alloc_def: int = 0
|
||
if sb[0] == '%':
|
||
tmp_nm: str = _peep_alloc_name(sb, pool)
|
||
if tmp_nm is not None and string.strcmp(tmp_nm, a_name) == 0:
|
||
is_alloc_def = 1
|
||
if is_alloc_def == 0 and _peep_last_ref_eq(sb, a_name) != 0:
|
||
if _peep_is_startwith(sb, "store ") != 0:
|
||
st_count += 1
|
||
if st_count == 1:
|
||
vcomma: str = string.strstr(sb + 6, ", ")
|
||
if vcomma is not None:
|
||
vlen: t.CSizeT = t.CSizeT(t.CUInt64T(vcomma) - t.CUInt64T(sb + 6))
|
||
if vlen > 0:
|
||
vbuf: str = pool.alloc(vlen + 1)
|
||
if vbuf is not None:
|
||
string.strncpy(vbuf, sb + 6, vlen)
|
||
vbuf[vlen] = '\0'
|
||
st_val = vbuf
|
||
st_ok = 1
|
||
else:
|
||
unsafe = 1
|
||
if blk_s == entry_blk:
|
||
st_line_order = order
|
||
elif string.strstr(sb, " = load ") is not None:
|
||
ls: str = string.strstr(sb, "%")
|
||
if ls is not None and load_cnt < 64:
|
||
lname_len: t.CSizeT = t.CSizeT(
|
||
t.CUInt64T(string.strstr(sb, " = load ")) - t.CUInt64T(ls))
|
||
lname: str = pool.alloc(lname_len + 1)
|
||
if lname is not None:
|
||
string.strncpy(lname, ls + 1, lname_len - 1)
|
||
lname[lname_len - 1] = '\0'
|
||
slotp: str | t.CPtr = t.CPtr(t.CUInt64T(load_ltoks) + load_cnt * 8)
|
||
slotp[0] = lname
|
||
load_cnt += 1
|
||
if blk_s == entry_blk and st_line_order < 0:
|
||
load_entry_order += 1
|
||
else:
|
||
# lname 分配失败：该 load 不会进 load_ltoks，但删除阶段
|
||
# 仍会按槽名删其 def 行，use 无法改写为 store 值 → 悬空。
|
||
# 保守放弃本 alloca 槽的 T2 转发。
|
||
unsafe = 1
|
||
else:
|
||
# 本 alloca 的 load 超过记录容量（load_ltoks=8*64，上限 64）
|
||
# 或行内无 '%' 无法解析 def 名：超出容量的 load 不会被记录，
|
||
# 但删除阶段按槽名删除全部 " = load " def 行 → 未记录 load 的
|
||
# use 未改写为 store 值，将悬空引用已删 def（llc:
|
||
# "use of undefined value"，见 strcmp 短路链 >64 分支场景）。
|
||
# 只有 load 全集都被记录时才能安全删全量 def，故保守放弃
|
||
# 本 alloca 槽的整个 T2 转发（保留 alloca/store/load，IR 仍合法）。
|
||
unsafe = 1
|
||
else:
|
||
unsafe = 1
|
||
if sb is not None:
|
||
order += 1
|
||
l_s = l_s.Next
|
||
blk_s = blk_s.Next
|
||
# 判据
|
||
safe_val: str = None
|
||
if unsafe == 0 and st_count == 1 and st_ok != 0 and st_line_order >= 0 \
|
||
and load_cnt > 0 and load_entry_order == 0 and st_val is not None:
|
||
pct: str = string.strchr(st_val, '%')
|
||
if pct is None:
|
||
sp1: str = string.strchr(st_val, ' ')
|
||
if sp1 is not None and sp1[1] != 0:
|
||
safe_val = sp1 + 1
|
||
else:
|
||
safe_val = st_val
|
||
else:
|
||
if string.strchr(pct, ' ') is None and string.strchr(pct, ',') is None:
|
||
safe_val = pct
|
||
if safe_val is not None:
|
||
# 第二遍：替换 load 名引用
|
||
li2: int = 0
|
||
while li2 < load_cnt:
|
||
slotsp: str | t.CPtr = t.CPtr(t.CUInt64T(load_ltoks) + li2 * 8)
|
||
ltok: str = slotsp[0]
|
||
if ltok is not None:
|
||
blk_r = blk
|
||
while blk_r is not None:
|
||
lr: rowptr[Line] = None
|
||
if blk_r.Lines is not None:
|
||
repl_lines: rowptr[GSList[Line]] = blk_r.Lines
|
||
lr = repl_lines.Head
|
||
while lr is not None:
|
||
if lr.Buf is not None:
|
||
nb2: str = _peep_replace_token(pool, lr.Buf, ltok, safe_val)
|
||
if nb2 is not None and nb2 is not lr.Buf:
|
||
lr.Buf = nb2
|
||
lr = lr.Next
|
||
blk_r = blk_r.Next
|
||
li2 += 1
|
||
# 删除：alloca def / 唯一 store / 全部 load 行
|
||
blk_d = blk
|
||
while blk_d is not None:
|
||
ld: rowptr[Line] = None
|
||
if blk_d.Lines is not None:
|
||
kill_lines: rowptr[GSList[Line]] = blk_d.Lines
|
||
ld = kill_lines.Head
|
||
while ld is not None:
|
||
db: str = ld.Buf
|
||
if db is not None:
|
||
kill: int = 0
|
||
dnm: str = None
|
||
if db[0] == '%':
|
||
dnm = _peep_alloc_name(db, pool)
|
||
if dnm is not None and string.strcmp(dnm, a_name) == 0:
|
||
kill = 1
|
||
if kill == 0 and _peep_last_ref_eq(db, a_name) != 0:
|
||
if _peep_is_startwith(db, "store ") != 0 \
|
||
or string.strstr(db, " = load ") is not None:
|
||
kill = 1
|
||
if kill != 0:
|
||
ld.Buf = None
|
||
ld = ld.Next
|
||
blk_d = blk_d.Next
|
||
e_line = e_line.Next
|
||
|
||
# ---------- T3: no-op bitcast 消除 ----------
|
||
_peep_noop_bitcast(func, pool)
|
||
|
||
|
||
def _peep_noop_bitcast(func: rowptr[Function], pool: rowptr[memhub.MemBuddy]):
|
||
"""T3: 消除 no-op bitcast（bitcast <T> %x to <T>，两侧类型相同）。
|
||
|
||
生成端（虚调用/联合位宽转换）会产生同类型恒等 bitcast（如
|
||
`%6 = bitcast i8* %ok_msg to i8*`），纯冗余指令。把 %x 的所有引用
|
||
替换为源操作数（%SSA/@global/常量 token）后删除该行。
|
||
保守限制：类型与操作数均为不含空格的简单 token。
|
||
"""
|
||
if func is None or pool is None or func.Blocks is None:
|
||
return
|
||
again: int = 1
|
||
while again != 0:
|
||
again = 0
|
||
hit_line: rowptr[Line] = None
|
||
nm_buf: str = None
|
||
src_buf: str = None
|
||
blk: rowptr[BasicBlock] = func.Blocks.Head
|
||
while blk is not None and hit_line is None:
|
||
if blk.Lines is not None:
|
||
ln: rowptr[Line] = blk.Lines.Head
|
||
while ln is not None and hit_line is None:
|
||
if ln.Buf is not None:
|
||
eqpos: str = string.strstr(ln.Buf, " = bitcast ")
|
||
if eqpos is not None:
|
||
nlen: t.CSizeT = t.CSizeT(t.CUInt64T(eqpos) - t.CUInt64T(ln.Buf)) - 1
|
||
if nlen > 0:
|
||
after: str = eqpos + 11 # 跳过 " = bitcast "
|
||
sp1: str = string.strchr(after, ' ')
|
||
tpos: str = string.strstr(after, " to ")
|
||
if sp1 is not None and tpos is not None \
|
||
and t.CUInt64T(sp1) < t.CUInt64T(tpos):
|
||
tyA_len: t.CSizeT = t.CSizeT(t.CUInt64T(sp1) - t.CUInt64T(after))
|
||
src_st: str = sp1 + 1
|
||
src_len: t.CSizeT = t.CSizeT(t.CUInt64T(tpos) - t.CUInt64T(src_st))
|
||
tyB: str = tpos + 4
|
||
tyB_len: t.CSizeT = string.strlen(tyB)
|
||
if src_len > 0 and string.strchr(tyB, ' ') is None \
|
||
and t.CUInt64T(string.strchr(src_st, ' ')) == t.CUInt64T(tpos) \
|
||
and tyA_len == tyB_len \
|
||
and string.strncmp(after, tyB, tyA_len) == 0:
|
||
nm_buf = pool.alloc(nlen + 1)
|
||
if nm_buf is not None:
|
||
string.strncpy(nm_buf, ln.Buf + 1, nlen)
|
||
nm_buf[nlen] = '\0'
|
||
src_buf = pool.alloc(src_len + 1)
|
||
if src_buf is not None:
|
||
string.strncpy(src_buf, src_st, src_len)
|
||
src_buf[src_len] = '\0'
|
||
hit_line = ln
|
||
ln = ln.Next
|
||
blk = blk.Next
|
||
if hit_line is None or nm_buf is None or src_buf is None:
|
||
return
|
||
# 全函数把 %nm 的引用替换为源操作数
|
||
blk = func.Blocks.Head
|
||
while blk is not None:
|
||
if blk.Lines is not None:
|
||
ln2: rowptr[Line] = blk.Lines.Head
|
||
while ln2 is not None:
|
||
if ln2.Buf is not None:
|
||
rb: str = _peep_replace_token(pool, ln2.Buf, nm_buf, src_buf)
|
||
if rb is not None and rb is not ln2.Buf:
|
||
ln2.Buf = rb
|
||
ln2.Len = string.strlen(rb)
|
||
ln2 = ln2.Next
|
||
blk = blk.Next
|
||
hit_line.Buf = None
|
||
again = 1
|
||
|
||
|
||
# ============================================================
|
||
# FunctionPrint: 将函数序列化为 IR 文本
|
||
#
|
||
# 格式:
|
||
# define <ret_ty> @<name>(<params>) {
|
||
# <block_name>:
|
||
# <instruction>
|
||
# <instruction>
|
||
# <block_name>:
|
||
# <instruction>
|
||
# }
|
||
# ============================================================
|
||
def FunctionPrint(buf: str, size: t.CSizeT,
|
||
func: rowptr[Function], pool: rowptr[memhub.MemBuddy]):
|
||
"""将函数 func 序列化为 IR 文本写入 buf"""
|
||
if func is None or buf is None or size == 0: return
|
||
|
||
# 先本地窥孔（ret zeroinitializer→0/null、入口槽 store→load 转发、
|
||
# no-op bitcast 消除）：窥孔按"名称一致性"匹配、删除行置 Buf=None；
|
||
# RenumberSSA 遍历时已跳过 Buf=None 行 → 窥孔之后再重编号，输出 SSA
|
||
# 编号连续（窥孔删行后不再出现 %1→%5 空洞，文件更紧凑）。
|
||
# RenumberSSA 内部会跳过 declare 函数（IsDeclared==1），无需在此判断。
|
||
if func.IsDeclared == 0:
|
||
_peephole_opt(func, pool)
|
||
|
||
RenumberSSA(func, pool)
|
||
|
||
ty_buf: str = pool.alloc(128)
|
||
if ty_buf is None: return
|
||
ty_buf[0] = '\0'
|
||
|
||
# define <ret_ty> @<name>(<params>)
|
||
if func.IsDeclared == 1:
|
||
_append_cstr(buf, size, "declare ")
|
||
else:
|
||
# 泛型特化类方法（如 GSList[NameEntry].__before_init__）：多个模块各自
|
||
# 特化生成同名定义 → ld.lld duplicate symbol warning。特化方法代码由
|
||
# 相同模板+类型参数生成，定义相同，输出 linkonce_odr 让链接器合并。
|
||
if func.Name is not None and string.strchr(func.Name, 91) is not None:
|
||
_append_cstr(buf, size, "define linkonce_odr ")
|
||
else:
|
||
_append_cstr(buf, size, "define ")
|
||
ty_buf[0] = '\0'
|
||
TypePrint(ty_buf, 128, func.RetTy, pool)
|
||
_append_cstr(buf, size, ty_buf)
|
||
_append_cstr(buf, size, " @")
|
||
if func.Name is not None:
|
||
if _ll_name_needs_quote(func.Name) != 0:
|
||
_append_cstr(buf, size, "\"")
|
||
_append_cstr(buf, size, func.Name)
|
||
_append_cstr(buf, size, "\"")
|
||
else:
|
||
_append_cstr(buf, size, func.Name)
|
||
_append_cstr(buf, size, "(")
|
||
|
||
# 参数列表（从 GSList[Param].Head 遍历）
|
||
params: rowptr[GSList[Param]] = func.Params
|
||
cur: rowptr[Param] = None
|
||
if params is not None:
|
||
cur = params.Head
|
||
first: int = 1
|
||
cur_steps: t.CSizeT = 0
|
||
while cur is not None:
|
||
cur_steps += 1
|
||
if cur_steps > IR_DUMP_CHAIN_LIMIT:
|
||
if params is not None:
|
||
stdio.printf("[IR-DUMP] 链表异常: func params（函数名=%s，Params.Count=%zu，遍历超 %d 步，print 终止）\n",
|
||
func.Name, params.Count, IR_DUMP_CHAIN_LIMIT)
|
||
else:
|
||
stdio.printf("[IR-DUMP] 链表异常: func params（函数名=%s，Params=None，遍历超 %d 步，print 终止）\n",
|
||
func.Name, IR_DUMP_CHAIN_LIMIT)
|
||
stdlib.exit(1)
|
||
if first == 0:
|
||
_append_cstr(buf, size, ", ")
|
||
# 参数属性（如 "noundef"/"nocapture"）在类型之前
|
||
if cur.Attrs is not None:
|
||
_append_cstr(buf, size, cur.Attrs)
|
||
_append_cstr(buf, size, " ")
|
||
ty_buf[0] = '\0'
|
||
TypePrint(ty_buf, 128, cur.Ty, pool)
|
||
_append_cstr(buf, size, ty_buf)
|
||
# declare 和 define 都打印参数名（如果有）
|
||
if cur.Name is not None:
|
||
_append_cstr(buf, size, " ")
|
||
# LLVM IR 要求参数名以 % 开头
|
||
if cur.Name[0] != '%':
|
||
_append_cstr(buf, size, "%")
|
||
_append_cstr(buf, size, cur.Name)
|
||
cur = cur.Next
|
||
first = 0
|
||
# 变参函数追加 ...
|
||
if func.IsVarArg == 1:
|
||
if params is not None and params.Head is not None:
|
||
_append_cstr(buf, size, ", ...")
|
||
else:
|
||
_append_cstr(buf, size, "...")
|
||
_append_cstr(buf, size, ")")
|
||
|
||
# 函数级属性（如 "nounwind"/"noinline"）在 ) 之后
|
||
if func.Attrs is not None:
|
||
_append_cstr(buf, size, " ")
|
||
_append_cstr(buf, size, func.Attrs)
|
||
# B5：字符串属性（section 等）必须用属性组 #N 引用（LLVM 不接受内联
|
||
# section="..."），函数行输出 #N，组定义在函数尾/模块中输出。
|
||
if func.AttrGroup is not None and func.AttrGroupId > 0:
|
||
ag_ref: str = pool.alloc(32)
|
||
if ag_ref is not None:
|
||
viperlib.snprintf(ag_ref, 32, " #%d", func.AttrGroupId)
|
||
_append_cstr(buf, size, ag_ref)
|
||
|
||
if func.IsDeclared == 1:
|
||
_append_cstr(buf, size, "\n")
|
||
_append_attr_group_def(buf, size, pool, func)
|
||
return
|
||
|
||
# 函数体
|
||
_append_cstr(buf, size, " {\n")
|
||
|
||
blocks: rowptr[GSList[BasicBlock]] = func.Blocks
|
||
blk: rowptr[BasicBlock] = None
|
||
if blocks is not None:
|
||
blk = blocks.Head
|
||
blk_steps: t.CSizeT = 0
|
||
while blk is not None:
|
||
blk_steps += 1
|
||
if blk_steps > IR_DUMP_CHAIN_LIMIT:
|
||
stdio.printf("[IR-DUMP] 链表成环: func blocks（函数名=%s，print 终止）\n", func.Name)
|
||
stdlib.exit(1)
|
||
# 块标签
|
||
if blk.Name is not None:
|
||
_append_cstr(buf, size, blk.Name)
|
||
_append_cstr(buf, size, ":\n")
|
||
# 指令行（从 GSList[Line].Head 遍历）
|
||
line: rowptr[Line] = None
|
||
if blk.Lines is not None:
|
||
lines: rowptr[GSList[Line]] = blk.Lines
|
||
line = lines.Head
|
||
has_lines: int = 0
|
||
line_steps: t.CSizeT = 0
|
||
while line is not None:
|
||
line_steps += 1
|
||
if line_steps > IR_DUMP_CHAIN_LIMIT:
|
||
stdio.printf("[IR-DUMP] 链表成环: block lines（函数名=%s，块=%s，print 终止）\n",
|
||
func.Name, blk.Name)
|
||
stdlib.exit(1)
|
||
if line.Buf is not None:
|
||
_append_cstr(buf, size, " ")
|
||
_append_cstr(buf, size, line.Buf)
|
||
_append_cstr(buf, size, "\n")
|
||
has_lines = 1
|
||
line = line.Next
|
||
# 块未终止兜底：LLVM IR 要求每个基本块以终止指令（br/ret/unreachable）结尾。
|
||
# 两种情况需要补 unreachable：
|
||
# 1. 空块（has_lines==0）：标签后无任何指令
|
||
# 2. 有指令但未终止（IsTerminated==0）：如 break/continue 后的 dead.N 块
|
||
# 被后续语句填入指令但缺少终止符，导致下一个标签处报
|
||
# "expected instruction opcode"
|
||
if has_lines == 0 or blk.IsTerminated == 0:
|
||
_append_cstr(buf, size, " unreachable\n")
|
||
blk = blk.Next
|
||
|
||
_append_cstr(buf, size, "}\n")
|
||
_append_attr_group_def(buf, size, pool, func)
|
||
|
||
|
||
def _append_attr_group_def(buf: str, size: t.CSizeT,
|
||
pool: rowptr[memhub.MemBuddy],
|
||
func: rowptr[Function]):
|
||
"""输出字符串属性组定义: attributes #N = { "section"="..." }"""
|
||
if func is None or func.AttrGroup is None or func.AttrGroupId <= 0:
|
||
return
|
||
ag_buf: str = pool.alloc(192)
|
||
if ag_buf is None:
|
||
return
|
||
viperlib.snprintf(ag_buf, 192, "attributes #%d = { %s }\n",
|
||
func.AttrGroupId, func.AttrGroup)
|
||
_append_cstr(buf, size, ag_buf)
|
||
|
||
|
||
# ============================================================
|
||
# 内部辅助
|
||
# ============================================================
|
||
# _append_cstr 写入位置缓存（性能优化：同一缓冲连续追加时 O(1) 定位，
|
||
# 避免每次 strlen(dst) 从头扫描导致的 O(n²) —— IR 序列化热点）。
|
||
# 缓冲切换（dst 不同）或缓冲被重置（dst[0]=0 且 pos>0）时自动回退 strlen。
|
||
_g_last_dst: t.CUInt64T = 0
|
||
_g_last_pos: t.CSizeT = 0
|
||
|
||
def _append_cstr(dst: str, dst_size: t.CSizeT, src: str):
|
||
"""将 C 字符串 src 追加到 dst 末尾"""
|
||
global _g_last_dst, _g_last_pos
|
||
if dst is None or src is None: return
|
||
dlen: t.CSizeT = 0
|
||
if _g_last_dst == t.CUInt64T(dst):
|
||
dlen = _g_last_pos
|
||
# 缓冲被重置（调用方 buf[0]=0 后复用同一地址）→ 回到头部
|
||
if dlen > 0 and dst[0] == '\0':
|
||
dlen = 0
|
||
else:
|
||
dlen = string.strlen(dst)
|
||
_g_last_dst = t.CUInt64T(dst)
|
||
# 无符号下溢防护：dlen 超过缓冲大小时（缓冲已被破坏/悬空）直接返回，
|
||
# 否则 remain = dst_size - dlen 下溢成巨值会把 dst[dlen+i] 写穿到缓冲外
|
||
if dlen >= dst_size:
|
||
return
|
||
slen: t.CSizeT = string.strlen(src)
|
||
remain: t.CSizeT = dst_size - dlen
|
||
if remain <= 0: return
|
||
i: t.CSizeT = 0
|
||
while i < slen and i + 1 < remain:
|
||
dst[dlen + i] = src[i]
|
||
i += 1
|
||
dst[dlen + i] = '\0'
|
||
_g_last_pos = dlen + i
|
||
|
||
|
||
# ============================================================
|
||
# 字段访问器（供其他模块绕过 stub 类型限制使用）
|
||
#
|
||
# 跨模块编译时，使用方模块只有 Function/Param 的 stub 类型
|
||
# （字段不足且字段类型可能错误），导致 TransPyC 静默跳过
|
||
# 字段访问（不生成 GEP，不报错），字段值永远为 null。
|
||
#
|
||
# 这些访问器在拥有完整类型定义的本模块内执行，能正确访问
|
||
# 所有字段。其他模块通过函数调用获取字段值，绕过 stub 限制。
|
||
# ============================================================
|
||
def function_get_name(func: rowptr[Function]) -> str:
|
||
"""获取函数名"""
|
||
if func is None:
|
||
return None
|
||
return func.Name
|
||
|
||
|
||
def function_get_ret_ty(func: rowptr[Function]) -> rowptr[LLVMType]:
|
||
"""获取函数返回类型"""
|
||
if func is None:
|
||
return None
|
||
return func.RetTy
|
||
|
||
|
||
def function_get_params(func: rowptr[Function]) -> rowptr[GSList[Param]]:
|
||
"""获取函数参数链表容器"""
|
||
if func is None:
|
||
return None
|
||
return func.Params
|
||
|
||
|
||
def function_get_param_head(func: rowptr[Function]) -> rowptr[Param]:
|
||
"""获取函数参数链表头节点（一步到位，避免调用方访问 GSList.Head）"""
|
||
if func is None:
|
||
return None
|
||
params: rowptr[GSList[Param]] = func.Params
|
||
if params is None:
|
||
return None
|
||
return params.Head
|
||
|
||
|
||
def function_get_next(func: rowptr[Function]) -> rowptr[Function]:
|
||
"""获取链表下一个函数"""
|
||
if func is None:
|
||
return None
|
||
return func.Next
|
||
|
||
|
||
def function_is_declared(func: rowptr[Function]) -> int:
|
||
"""获取函数是否为声明（1=declare 仅声明, 0=define 定义）"""
|
||
if func is None:
|
||
return 0
|
||
return func.IsDeclared
|
||
|
||
|
||
def param_get_name(param: rowptr[Param]) -> str:
|
||
"""获取参数名"""
|
||
if param is None:
|
||
return None
|
||
return param.Name
|
||
|
||
|
||
def param_get_ty(param: rowptr[Param]) -> rowptr[LLVMType]:
|
||
"""获取参数类型"""
|
||
if param is None:
|
||
return None
|
||
return param.Ty
|
||
|
||
|
||
def param_get_next(param: rowptr[Param]) -> rowptr[Param]:
|
||
"""获取链表下一个参数"""
|
||
if param is None:
|
||
return None
|
||
return param.Next
|