GVSDS/Viper_VRTE
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(mlir): 全局初值 region 草稿缓冲改为单次 PrintModule 复用，修子池耗尽

Browse Source 
每个全局初值 region 逐全局 alloc leaf/ins 各 32 KiB、每个初值节点 alloc
4 KiB 类型/符号草稿且从不释放：结构体/vtable 全局多的模块（MLIRLiteTest
#12）在约 2 MiB 模块子池上耗尽，报"零初值类型缓冲分配失败"。

改为按 PrintModule 惰性分配一组复用草稿（leaf/ins/类型/符号/路径栈），
跨全局、跨节点共享；PrintModule 入口 MlirGvResetScratch 使上次草稿失效，
下次按当前池重建（绝不复用已释放/被复用地址的池，避免悬空写穿 canary）。
下标路径改按递归深度分槽的复用栈，不再逐节点分配。
This commit is contained in:
TermiNexus committed 2026-10-04 21:09:36 +08:00
1 parent 68fe839da9
commit 5dad869a40
1 file changed
+111 -59
Download Patch File Download Diff File Expand all files Collapse all files
mlirlite/__module.vp
+111 -59
View File
@@ -340,6 +340,10 @@ class LLVMModule:
# 打印期间有效，清空后不再解引用（避免 pool reset 后悬空）。
_MlirPublishNamedTypes(self, pool)
# 使 region 渲染草稿失效：本次打印首次用到时按当前 pool 惰性重建（见
# _MlirGvEnsureScratch）。跨打印绝不复用（上一次的池可能已释放/被复用）。
MlirGvResetScratch()
_append_cstr(buf, size, "module {\n")
# 模式说明注释（MLIR 无 stub/text 二分）
@@ -1028,7 +1032,13 @@ def _MlirArrayDenseAt(buf: str, size: sizet, src: str,
# * 符号名一律经 _MlirSymRefAppend（先剥 LLVM 口径引号，再按 MLIR 口径重加）；
# * 不支持的叶子形态一律 MlirFail（带原始文本），绝不静默输出错误 MLIR。
# ============================================================
GV_REGION_BUF_SIZE: t.CDefine = 32768
GV_REGION_BUF_SIZE: t.CDefine = 32768
# 下标路径复用栈：每个递归深度一个槽（顶层 path="" → 深度 d 的路径写在槽 d）。
# 深度上限与 _MlirGvWalk 的 64 层守卫一致，留余量取 72；每槽 512 字节足够
# （64 层 × " , NN" ≈ 400 字节）。
GV_PATH_SLOT: t.CDefine = 512
GV_PATH_MAX: t.CDefine = 72
# region 渲染期状态（单线程递归渲染，逐全局重置）
_MlirGvLeafBuf: str = None # 叶子 op 缓冲
@@ -1040,6 +1050,72 @@ _MlirGvPool: memhub.MemBuddy* = None
_MlirGvLeafSeq: int = 0
_MlirGvInsSeq: int = 0
# ---- region 渲染的**复用**草稿缓冲（惰性分配一次，跨全局/跨节点共享）----
#
# 此前每个全局、每个初值节点都从模块子分配器 alloc 新缓冲且从不释放：
# 每全局 leaf/ins 各 32 KiB、每节点 4 KiB ⇒ 结构体/vtable 全局多的模块
# （MLIRLiteTest #12）把 ~2 MiB 子池耗尽，报 "零初值类型缓冲分配失败"。
# region 渲染是单线程顺序的，草稿内容用完即弃 ⇒ 分配一次复用即可。
_MlirGvSbRead1: str = None # 512B：类型前缀记号读取草稿
_MlirGvSbRead2: str = None # 512B：叶子类型记号读取草稿
_MlirGvSbName: str = None # 512B：符号名 / 常量表达式符号名草稿
_MlirGvSbVal: str = None # 512B：标量初值字面量草稿
_MlirGvTyBuf: str = None # 4096B：元素 MLIR 类型文本（zero/const）
_MlirGvNameBuf: str = None # 32B：SSA 名生成草稿
_MlirGvPathStack: str = None # GV_PATH_MAX × GV_PATH_SLOT：下标路径复用栈
def MlirGvResetScratch():
"""使 region 渲染的复用草稿失效（下次 EnsureScratch 按当前池重新分配）
草稿只在**单次 PrintModule** 内经惰性分配复用（跨全局/跨节点共享）；
打印结束由 PrintModule 调用本函数置空 ⇒ 绝不复用上一次打印的指针
（上一次的池可能已释放，或同一地址被新池复用 → 悬空写穿）。
"""
global _MlirGvLeafBuf
_MlirGvLeafBuf = None
def _MlirGvEnsureScratch(pool: memhub.MemBuddy*) -> int:
"""惰性分配本次 PrintModule 的复用草稿缓冲（跨全局/跨节点共享），返回 1=就绪"""
global _MlirGvLeafBuf, _MlirGvInsBuf, _MlirGvTopType, _MlirGvLast, _MlirGvPrev
global _MlirGvSbRead1, _MlirGvSbRead2, _MlirGvSbName, _MlirGvSbVal
global _MlirGvTyBuf, _MlirGvNameBuf, _MlirGvPathStack
if _MlirGvLeafBuf is not None:
return 1
lb: str = pool.alloc(GV_REGION_BUF_SIZE)
ib: str = pool.alloc(GV_REGION_BUF_SIZE)
tb: str = pool.alloc(4096)
la: str = pool.alloc(32)
pr: str = pool.alloc(32)
r1: str = pool.alloc(512)
r2: str = pool.alloc(512)
nm: str = pool.alloc(512)
vl: str = pool.alloc(512)
tyb: str = pool.alloc(4096)
nb: str = pool.alloc(32)
ps: str = pool.alloc(GV_PATH_SLOT * GV_PATH_MAX)
if lb is None or ib is None or tb is None or la is None or pr is None \
or r1 is None or r2 is None or nm is None or vl is None \
or tyb is None or nb is None or ps is None:
MlirFail("MLIR 渲染: 全局 region 缓冲分配失败")
lb[0] = '\0'; ib[0] = '\0'; tb[0] = '\0'; la[0] = '\0'; pr[0] = '\0'
r1[0] = '\0'; r2[0] = '\0'; nm[0] = '\0'; vl[0] = '\0'
tyb[0] = '\0'; nb[0] = '\0'; ps[0] = '\0'
_MlirGvLeafBuf = lb
_MlirGvInsBuf = ib
_MlirGvTopType = tb
_MlirGvLast = la
_MlirGvPrev = pr
_MlirGvSbRead1 = r1
_MlirGvSbRead2 = r2
_MlirGvSbName = nm
_MlirGvSbVal = vl
_MlirGvTyBuf = tyb
_MlirGvNameBuf = nb
_MlirGvPathStack = ps
return 1
def _MlirGlobalInitNeedsRegion(x: str) -> int:
"""判定全局初值是否改走 initializer region（1=region / 0=属性路径）
@@ -1110,19 +1186,18 @@ def _MlirGlobalInitRegion(buf: str, size: sizet,
每行都带 '\\n'；调用方在末尾补 '}'，模块层再统一缩进 2 空格。
"""
global _MlirGvLeafBuf, _MlirGvInsBuf, _MlirGvTopType, _MlirGvLast
global _MlirGvPrev, _MlirGvPool, _MlirGvLeafSeq, _MlirGvInsSeq
global _MlirGvPool, _MlirGvLeafSeq, _MlirGvInsSeq
x: str = gv.Initializer
if x is None:
return
topbuf: str = pool.alloc(4096)
leafbuf: str = pool.alloc(GV_REGION_BUF_SIZE)
insbuf: str = pool.alloc(GV_REGION_BUF_SIZE)
lastbuf: str = pool.alloc(32)
prevbuf: str = pool.alloc(32)
if topbuf is None or leafbuf is None or insbuf is None \
or lastbuf is None or prevbuf is None:
MlirFail("MLIR 渲染: 全局 region 缓冲分配失败")
# 复用缓冲（惰性分配一次，跨全局共享）——此前逐全局 alloc 且不释放，
# 结构体/vtable 全局多时耗尽模块子分配器。
_MlirGvEnsureScratch(pool)
topbuf: str = _MlirGvTopType
leafbuf: str = _MlirGvLeafBuf
insbuf: str = _MlirGvInsBuf
lastbuf: str = _MlirGvLast
prevbuf: str = _MlirGvPrev
topbuf[0] = '\0'
leafbuf[0] = '\0'
insbuf[0] = '\0'
@@ -1130,11 +1205,6 @@ def _MlirGlobalInitRegion(buf: str, size: sizet,
prevbuf[0] = '\0'
TypePrint(topbuf, 4096, gv.Ty, pool)
_MlirGvLeafBuf = leafbuf
_MlirGvInsBuf = insbuf
_MlirGvTopType = topbuf
_MlirGvLast = lastbuf
_MlirGvPrev = prevbuf
_MlirGvPool = pool
_MlirGvLeafSeq = 0
_MlirGvInsSeq = 0
@@ -1185,18 +1255,14 @@ def _MlirGvWalk(p: str, ty: rawptr[LLVMType], path: str, depth: int) -> str:
MlirFail("MLIR 渲染: 全局初值文本意外结束")
# 显式数组类型前缀（如字段写成 "[3 x i8] [i8 1, i8 2]"）先整体读掉
if p[0] == '[' and _MlirGvIsArrayType(p) != 0:
tb: str = _MlirGvPool.alloc(4096)
if tb is None:
MlirFail("MLIR 渲染: 数组类型前缀缓冲分配失败")
tb: str = _MlirGvSbRead1
p = _mlir_read_type(p, tb, 512)
p = _mlir_ws(p)
# 显式结构体类型前缀（如 llvm.global_ctors 元素的
# "{ i32, void ()*, i8* } { i32 65535, … }"）先整体读掉；`{…}` 既可能是
# 类型也可能是字面量，靠"其后是否紧跟另一个 `{`"区分（见 _MlirGvIsStructType）
if p[0] == '{' and _MlirGvIsStructType(p) != 0:
stb: str = _MlirGvPool.alloc(4096)
if stb is None:
MlirFail("MLIR 渲染: 结构体类型前缀缓冲分配失败")
stb: str = _MlirGvSbRead1
p = _mlir_read_type(p, stb, 512)
p = _mlir_ws(p)
# MLIR 形态聚合类型前缀（`!llvm.array<…>` / `!llvm.struct<…>`）后紧跟聚合
@@ -1205,9 +1271,7 @@ def _MlirGvWalk(p: str, ty: rawptr[LLVMType], path: str, depth: int) -> str:
# TranslatorGenMixin._build_array_initializer_text），上方 LLVM 形态前缀
# 判据（_MlirGvIsArrayType/_MlirGvIsStructType）认不出 '!' 开头
if p[0] == '!':
mtb: str = _MlirGvPool.alloc(4096)
if mtb is None:
MlirFail("MLIR 渲染: MLIR 类型前缀缓冲分配失败")
mtb: str = _MlirGvSbRead1
qm: str = _mlir_read_type(p, mtb, 512)
qm = _mlir_ws(qm)
if qm[0] == '[' or qm[0] == '{':
@@ -1276,15 +1340,19 @@ def _MlirGvIsStructType(p: str) -> int:
return 0
def _MlirGvPathChild(path: str, idx: int) -> str:
"""由父下标路径与子序号拼出子下标路径（顶层 path="" → "0"）"""
nb: str = _MlirGvPool.alloc(string.strlen(path) + 16)
if nb is None:
MlirFail("MLIR 渲染: 下标路径缓冲分配失败")
def _MlirGvPathChild(path: str, idx: int, depth: int) -> str:
"""由父下标路径与子序号拼出子下标路径（顶层 path="" → "0"）
写入按递归深度分槽的复用栈（槽 = depth）：祖先路径占 0..depth-1，
本层写槽 depth、子层写槽 depth+1，互不覆盖；避免逐节点 alloc 泄漏。
"""
if depth < 0 or depth >= GV_PATH_MAX:
MlirFail("MLIR 渲染: 下标路径深度越界")
nb: str = _MlirGvPathStack + depth * GV_PATH_SLOT
if path[0] == '\0':
viperlib.snprintf(nb, string.strlen(path) + 16, "%d", idx)
viperlib.snprintf(nb, GV_PATH_SLOT, "%d", idx)
else:
viperlib.snprintf(nb, string.strlen(path) + 16, "%s, %d", path, idx)
viperlib.snprintf(nb, GV_PATH_SLOT, "%s, %d", path, idx)
return nb
@@ -1317,7 +1385,7 @@ def _MlirGvWalkStruct(p: str, ty: rawptr[LLVMType], path: str, depth: int) -> st
fty: rawptr[LLVMType] = None
if cur is not None:
fty = paramnode_get_ty(cur)
p = _MlirGvWalk(p, fty, _MlirGvPathChild(path, idx), depth + 1)
p = _MlirGvWalk(p, fty, _MlirGvPathChild(path, idx, depth), depth + 1)
if cur is not None:
cur = paramnode_get_next(cur)
idx += 1
@@ -1351,7 +1419,7 @@ def _MlirGvWalkArray(p: str, ty: rawptr[LLVMType], path: str, depth: int) -> str
elif p[0] == ',':
p = p + 1
else:
p = _MlirGvWalk(p, elem_ty, _MlirGvPathChild(path, idx), depth + 1)
p = _MlirGvWalk(p, elem_ty, _MlirGvPathChild(path, idx, depth), depth + 1)
idx += 1
if idx == 0:
string.strcpy(_MlirGvLast, "%undef")
@@ -1395,26 +1463,20 @@ def _MlirGvWalkLeaf(p: str, ty: rawptr[LLVMType], path: str, depth: int) -> str:
p = _mlir_ws(p)
if _MlirGvIsBareValue(p) == 0:
# 元素自带类型记号（`i8* null` / `!llvm.ptr null` 等）：只用于推进游标
tb: str = _MlirGvPool.alloc(4096)
if tb is None:
MlirFail("MLIR 渲染: 叶子类型记号缓冲分配失败")
tb: str = _MlirGvSbRead2
p = _mlir_read_type(p, tb, 512)
p = _mlir_ws(p)
if p[0] == '\0':
MlirFail("MLIR 渲染: 叶子初值只有类型、缺少值")
if p[0] == '@':
# 裸符号引用（剥 LLVM 引号 → 经 _MlirSymRefAppend 按 MLIR 口径重加）
nb: str = _MlirGvPool.alloc(4096)
if nb is None:
MlirFail("MLIR 渲染: 符号名缓冲分配失败")
nb: str = _MlirGvSbName
p = _MlirGvReadSymName(p, nb, 512)
_MlirGvEmitAddressOf(nb)
elif string.strncmp(p, "bitcast", 7) == 0 \
or string.strncmp(p, "getelementptr", 13) == 0:
# 常量表达式整体视为符号引用（bitcast 不透明指针下恒等；GEP 取首元素地址）
nb2: str = _MlirGvPool.alloc(4096)
if nb2 is None:
MlirFail("MLIR 渲染: 常量表达式符号名缓冲分配失败")
nb2: str = _MlirGvSbName
p = _MlirGvScanConstExprSym(p, nb2, 512)
_MlirGvEmitAddressOf(nb2)
elif string.strncmp(p, "null", 4) == 0:
@@ -1426,9 +1488,7 @@ def _MlirGvWalkLeaf(p: str, ty: rawptr[LLVMType], path: str, depth: int) -> str:
else:
c0: int = p[0]
if c0 == '-' or (c0 >= '0' and c0 <= '9') or c0 == '.':
vb: str = _MlirGvPool.alloc(4096)
if vb is None:
MlirFail("MLIR 渲染: 标量初值缓冲分配失败")
vb: str = _MlirGvSbVal
p = _mlir_read_value(p, vb, 512)
_MlirGvEmitConst(vb, ty)
else:
@@ -1513,11 +1573,9 @@ def _MlirGvSymNameEnd(ch: int) -> int:
def _MlirGvNextLeafName() -> str:
"""分配下一个叶子 SSA 名（%v0, %v1, …）"""
"""分配下一个叶子 SSA 名（%v0, %v1, …；复用共享名缓冲，调用后立即取用）"""
global _MlirGvLeafSeq
ln: str = _MlirGvPool.alloc(32)
if ln is None:
MlirFail("MLIR 渲染: 叶子 SSA 名缓冲分配失败")
ln: str = _MlirGvNameBuf
viperlib.snprintf(ln, 32, "%%v%d", _MlirGvLeafSeq)
_MlirGvLeafSeq += 1
return ln
@@ -1536,9 +1594,7 @@ def _MlirGvEmitAddressOf(name: str):
def _MlirGvEmitZero(ty: rawptr[LLVMType]):
"""发射叶子 op: %vN = llvm.mlir.zero : <元素 MLIR 类型>"""
mt: str = _MlirGvPool.alloc(4096)
if mt is None:
MlirFail("MLIR 渲染: 零初值类型缓冲分配失败")
mt: str = _MlirGvTyBuf
mt[0] = '\0'
TypePrint(mt, 4096, ty, _MlirGvPool)
ln: str = _MlirGvNextLeafName()
@@ -1552,9 +1608,7 @@ def _MlirGvEmitZero(ty: rawptr[LLVMType]):
def _MlirGvEmitConst(val: str, ty: rawptr[LLVMType]):
"""发射叶子 op: %vN = llvm.mlir.constant(<lit> : <元素 MLIR 类型>) : <同类型>"""
mt: str = _MlirGvPool.alloc(4096)
if mt is None:
MlirFail("MLIR 渲染: 标量元素类型缓冲分配失败")
mt: str = _MlirGvTyBuf
mt[0] = '\0'
TypePrint(mt, 4096, ty, _MlirGvPool)
if mt[0] != 'i' and mt[0] != 'f':
@@ -1580,9 +1634,7 @@ def _MlirGvInsertLeaf(path: str):
"""发射 insertvalue（多下标路径；类型后缀始终是顶层聚合类型）"""
global _MlirGvInsSeq
_MlirGvInsSeq += 1
sn: str = _MlirGvPool.alloc(32)
if sn is None:
MlirFail("MLIR 渲染: insertvalue SSA 名缓冲分配失败")
sn: str = _MlirGvNameBuf
viperlib.snprintf(sn, 32, "%%s%d", _MlirGvInsSeq)
_append_cstr(_MlirGvInsBuf, GV_REGION_BUF_SIZE, " ")
_append_cstr(_MlirGvInsBuf, GV_REGION_BUF_SIZE, sn)