Viper_VRTE/mlirlite/__function.vp

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,
    MlirTypeText, MlirTypeTextIsPointer, MlirFail, MlirSetContextLine,
)


# dump sentinel: IR list traversal step cap (cycle detection, fail-fast, aligned with __module.vp).
# A broken Next cycle would hang traversal; over the cap it prints the list name and exit(1).
define IR_DUMP_CHAIN_LIMIT = 100000

# Peephole master switch (T2 forwarding/ret zeroing, T3 no-op bitcast removal).
# --no-llvmpeephole zeroes it so FunctionPrint skips local peepholes for IR diffing (RenumberSSA still runs).
G_PeepEnable: bool = True

# Function-text truncation flag: _append_cstr sets it to 1 when output hits the buffer cap.
# Checked before FunctionPrint returns and MlirFail'd: truncated MLIR must fail loudly, never silently.
_MlirOutTrunc: int = 0


# ============================================================
# Line: instruction text line
#
# One LLVM IR instruction serializes to one text line, held in a linked list.
# Extends GSListNode[Line] for typed Next: Line|CPtr
#
# NOTE: must carry @t.NoVTable, else TransPyC adds a vtable ptr (T*) but drops
# the parent GSListNode[T] Next field, misaligning the struct so field reads return garbage.
# ============================================================
novirtual class Line(GSListNode[Line]):
    Buf: str       # line text (NUL-terminated)
    Len: sizet               # text length (excluding NUL)


# ============================================================
# BasicBlock: basic block
#
# Holds name + instruction-line list + terminator flag.
# Extends GSListNode[BasicBlock] for typed Next: BasicBlock|CPtr
# Lines is a GSList[Line] pointer providing O(1) append.
#
# NOTE: must carry @t.NoVTable (same reason as Line).
# ============================================================
novirtual class BasicBlock(GSListNode[BasicBlock]):
    Name: str          # block name (e.g. "entry"/"if.then")
    Lines: rawptr[GSList[Line]]    # instruction-line list container (Head/Tail/Count)
    IsTerminated: int            # 1=terminator present (br/ret)


# ============================================================
# Param: function parameter
#
# Extends GSListNode[Param] for typed Next: Param|CPtr
# Attrs holds parameter attribute text (e.g. "noundef"/"nocapture"/"readonly"), None=none.
#
# NOTE: must carry @t.NoVTable (same reason as Line).
# ============================================================
novirtual class Param(GSListNode[Param]):
    Ty: rawptr[LLVMType]           # parameter type
    Name: str          # parameter name (e.g. "%0"/"%a")
    Attrs: str         # parameter attribute text (None=none)


# ============================================================
# Function: function
#
# Holds name + return type + parameter list + basic-block list.
# Extends GSListNode[Function] for typed Next: Function|CPtr
# Params/Blocks are GSList pointers providing O(1) append.
# Attrs holds function-level attribute text (e.g. "nounwind"/"noinline"), None=none.
#
# NOTE: must carry @t.NoVTable (same reason as Line).
# ============================================================
novirtual class Function(GSListNode[Function]):
    Name: str              # function name
    RetTy: rawptr[LLVMType]            # return type
    Params: rawptr[GSList[Param]]      # parameter list container
    Blocks: rawptr[GSList[BasicBlock]] # basic-block list container
    IsDeclared: int                  # 1=declaration only (declare), 0=definition (define)
    IsVarArg: int                    # 1=variadic function (...)
    Attrs: str             # function attribute text (None=none)
    AttrGroup: str         # string attribute group content (e.g. '"section"=".text.startup"', None=none)
    AttrGroupId: int                 # attribute group id (>0 emits #N on the function line and attributes #N at module end)
    NameLen: sizet                # function name length (computed once at creation, always == strlen(Name); appended last so existing field offsets stay)
    DbgRef: str            # DISubprogram ref (kept as-is; ignored by MLIR printing, e.g. "!3")


# ============================================================
# Factory functions
# ============================================================
# Line hardcoded size: Next(8) + Buf(8) + Len(8) = 24 bytes
define LINE_SIZE = 24

def new_line(pool: memhub.MemBuddy*, text: str) -> rawptr[Line]:
    """Create an instruction line, copying text into a newly allocated buffer"""
    ptr: rawptr[Line] = pool.alloc(LINE_SIZE)
    if ptr is None: return None
    string.memset(ptr, 0, LINE_SIZE)
    if text is not None:
        slen: sizet = 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: memhub.MemBuddy*, list_size: sizet) -> t.CPtr:
    """Allocate and zero-init a GSList container from pool (Head/Tail/Count all zero)"""
    gsl: t.CPtr = pool.alloc(list_size)
    if gsl is None: return None
    string.memset(gsl, 0, list_size)
    return gsl


# BasicBlock hardcoded size: Next(8) + Name(8) + Lines(8) + IsTerminated(4) = 32 bytes
define BLOCK_SIZE = 32

def new_basic_block(pool: memhub.MemBuddy*, name: str) -> rawptr[BasicBlock]:
    """Create a basic block"""
    ptr: rawptr[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 hardcoded size: Next(8) + Ty(8) + Name(8) + Attrs(8) = 32 bytes
define PARAM_SIZE = 32

def new_param(pool: memhub.MemBuddy*, ty: rawptr[LLVMType],
              name: str) -> rawptr[Param]:
    """Create a function parameter"""
    ptr: rawptr[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 hardcoded size: Next(8) + Name(8) + RetTy(8) + Params(8) + Blocks(8) + IsDeclared(4) + IsVarArg(4) + Attrs(8) + AttrGroup(8) + AttrGroupId(8) + NameLen(8) + DbgRef(8) = 88 bytes (same layout as llvmlite)
define FUNC_SIZE = 88

def new_function(pool: memhub.MemBuddy*, name: str,
                 ret_ty: rawptr[LLVMType]) -> rawptr[Function]:
    """Create a function"""
    ptr: rawptr[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
    # NameLen computed once: Name never changes, so NameLen == strlen(Name) (avoids repeated strlen).
    # name=None sets 0, matching new_line's null convention (explicit for fixed semantics, not allocator behavior).
    if name is not None:
        ptr.NameLen = string.strlen(name)
    else:
        ptr.NameLen = 0
    return ptr


# ============================================================
# List operations (delegating to GSList.append)
# ============================================================
def function_add_param(func: rawptr[Function], param: rawptr[Param]):
    """Append a parameter to the function's parameter list"""
    if func is None or param is None: return
    # NOTE: use a local variable + type parameter (see function_move_block_to_end): chaining
    # func.Params.append() mis-resolves to LinkedNode.append and corrupts memory 48 bytes before the container.
    params: rawptr[GSList[Param]] = func.Params
    if params is None: return
    params.append(param)


def function_add_block(func: rawptr[Function], block: rawptr[BasicBlock]):
    """Append a basic block to the function's block list"""
    if func is None or block is None: return
    blocks: rawptr[GSList[BasicBlock]] = func.Blocks
    if blocks is None: return
    blocks.append(block)


def function_move_block_to_end(func: rawptr[Function], block: rawptr[BasicBlock]):
    """Move block to the end of the function's block list (BB text order == control-flow order).

    Fixes non-monotonic SSA numbering (BBs are created in order but instructions are emitted in
    control-flow order) that makes llc error; called at position_at_end so BB order matches position order.
    NOTE: must use the typed GSList[BasicBlock] parameter, else TransPyC builds an incomplete struct
    type and blocks.Head/Tail/Count field reads return garbage.
    """
    if func is None or block is None: return
    blocks: rawptr[GSList[BasicBlock]] = func.Blocks
    if blocks is None: return

    # Already Tail; nothing to move
    if blocks.Tail is block:
        return

    # Remove block from the list
    if blocks.Head is block:
        # block is Head; advance Head directly
        blocks.Head = block.Next
    else:
        # Walk to find the predecessor
        prev: rawptr[BasicBlock] = blocks.Head
        while prev is not None and prev.Next is not block:
            prev = prev.Next
        if prev is None:
            # block not in list; append directly
            blocks.append(block)
            return
        prev.Next = block.Next

    # block.Next handled by the removal logic; reset to None
    block.Next = None

    # Append at the end (no blocks.append call to avoid double-incrementing Count)
    if blocks.Tail is not None:
        blocks.Tail.Next = block
    blocks.Tail = block


def block_append_line(block: rawptr[BasicBlock], line: rawptr[Line]):
    """Append an instruction line to the basic block"""
    if block is None or line is None: return
    lines: rawptr[GSList[Line]] = block.Lines
    if lines is None: return
    lines.append(line)


def block_set_terminated(block: rawptr[BasicBlock], val: int):
    """Set the block terminator flag (cross-module call to avoid opaque chained field access)"""
    if block is None: return
    block.IsTerminated = val


def block_get_terminated(block: rawptr[BasicBlock]) -> int:
    """Get the block terminator flag (cross-module call to avoid opaque chained field access)"""
    if block is None:
        return 1
    return block.IsTerminated


def block_append_text(pool: memhub.MemBuddy*, block: rawptr[BasicBlock],
                      text: str):
    """Create an instruction line and append it to the basic block"""
    if block is None or text is None: return
    line: rawptr[Line] = new_line(pool, text)
    if line is not None:
        block_append_line(block, line)


def block_insert_text_before_terminator(pool: memhub.MemBuddy*,
                                        block: rawptr[BasicBlock],
                                        text: str):
    """Create an instruction line and insert it before the block terminator.

    If the block is not terminated, append at the end; if terminated (br/ret), insert before
    the terminator so alloca and similar instructions do not land after it.
    """
    if block is None or text is None: return
    if block.Lines is None: return
    line: rawptr[Line] = new_line(pool, text)
    if line is None: return
    # Block not terminated: append directly
    if block.IsTerminated == 0:
        block_append_line(block, line)
        return
    # Block terminated: insert before Tail (the terminator)
    lines: rawptr[GSList[Line]] = block.Lines
    if lines.Head is None:
        block_append_line(block, line)
        return
    # Single node: insert before Head
    if lines.Head is lines.Tail:
        line.Next = lines.Head
        lines.Head = line
        lines.Count += 1
        return
    # Find the predecessor of Tail
    prev: rawptr[Line] = lines.Head
    while prev is not None and prev.Next is not lines.Tail:
        prev = prev.Next
    if prev is None:
        # No predecessor found; fall back to append
        block_append_line(block, line)
        return
    # Insert line between prev and Tail
    line.Next = lines.Tail
    prev.Next = line
    lines.Count += 1


# ============================================================
# Attribute setters
# ============================================================
def function_set_attrs(func: rawptr[Function], attrs: str):
    """Set function-level attribute text (e.g. "nounwind"/"noinline")"""
    if func is None: return
    func.Attrs = attrs


_g_attr_group_counter: int = 0

def function_set_attr_group(func: rawptr[Function], group: str):
    """Set function string attribute group (e.g. '"section"=".text.startup"') and assign #N.

    LLVM string attributes (with a value, e.g. section) cannot be inlined on the signature
    line; they must use a group: the function line emits #N and the group is emitted at the end.
    """
    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: rawptr[Param], attrs: str):
    """Set parameter attribute text (e.g. "noundef"/"nocapture"/"readonly")"""
    if param is None: return
    param.Attrs = attrs


def function_set_dbg(func: rawptr[Function], ref: str):
    """Set the function-level debug ref (DISubprogram node ref, e.g. "!3").

    Kept as-is: MLIR printing ignores it (FunctionPrint emits no !dbg), so this only
    records the model without changing output. ref=None clears it.
    """
    if func is None: return
    func.DbgRef = ref


# ============================================================
# _ll_name_needs_quote - whether a symbol name needs @"..." quoting in MLIR
#
# WARNING: the rule differs from LLVM (measured on mlir-opt 22.1.8):
#   - A bare identifier's first char must be a letter or '_': leading digit, '.', '$' must be quoted.
#     => TPV's {sha1}.{name} all start with a hex digit, so all need quotes;
#     => the string-constant global @.str starts with '.', so it needs quotes too.
#   - Later chars allow only [A-Za-z0-9_$] and '.'.
#   - '-' is illegal in a bare MLIR name (LLVM allows it); measured error "expected '('".
#   - '['/']' (generic specialization names) are likewise illegal.
# This predicate takes the conservative direction: when unsure, quote (@"..." is legal for any name).
# ============================================================
def _ll_name_needs_quote(name: str) -> int:
    """MLIR bare-name predicate: returns 1 (needs quotes) unless it matches [A-Za-z_][A-Za-z0-9_$.]*"""
    if name is None:
        return 0
    name_len: sizet = string.strlen(name)
    if name_len == 0:
        return 1
    c0: char = name[0]
    if not ((c0 >= 'a' and c0 <= 'z') or (c0 >= 'A' and c0 <= 'Z') or c0 == '_'):
        return 1
    i: sizet = 1
    while i < name_len:
        ch: char = name[i]
        if not ((ch >= 'a' and ch <= 'z') or (ch >= 'A' and ch <= 'Z') or
                (ch >= '0' and ch <= '9') or ch == '_' or ch == '$' or ch == '.'):
            return 1
        i += 1
    return 0


# ============================================================
# _MlirSymRefAppend - emit a symbol reference (unified quoting rule)
#
# name does NOT include a leading '@'; an already-quoted form (first char '"') is emitted as-is.
# Used where a symbol appears as an operand (e.g. llvm.mlir.addressof), sharing the same
# predicate with _MlirFuncSig / call target / global name.
# ============================================================
def _MlirSymRefAppend(out: str, outsz: sizet, name: str):
    if name is None or name[0] == '\0':
        MlirFail("MLIR 渲染: 空符号名")
    if name[0] == '"':
        _append_cstr(out, outsz, "@")
        _append_cstr(out, outsz, name)
        return
    if _ll_name_needs_quote(name) != 0:
        _append_cstr(out, outsz, "@\"")
        _append_cstr(out, outsz, name)
        _append_cstr(out, outsz, "\"")
        return
    _append_cstr(out, outsz, "@")
    _append_cstr(out, outsz, name)


# ============================================================
# RenumberSSA: renumber numeric SSA names within a function
#
# Fixes the non-monotonic SSA names in text output caused by build_alloca_at_entry
# inserting allocas before the entry block terminator (llc error:
# "instruction expected to be numbered '%N' or greater").
#
# Example: entry originally %0-%6 + br; alloca_at_entry inserts %28/%41 before br,
#          so text order is entry(%0..%6, %28, %41) -> if.then(%7..%27),
#          placing %28 before %7, which llc rejects.
#
# Algorithm:
#   1. Scan all BB instruction lines, collect numeric SSA names %N in text-output order
#   2. Build an old->new map by first-occurrence order (incrementing from 0)
#   3. Replace %N in every instruction line text with the new number
#
# NOTE:
#   - Only pure numeric SSA names (%0, %1, ...) are renumbered
#   - Named SSA names (%entry, %.tmp, %self), block labels and global symbols @N are not
#   - Only called for define (with a body); declare has no body to process
#   - TransPyV params are mostly named (%self/%env/%n), so no numeric param renumbering needed
# ============================================================

# Whether ch is a decimal digit
def _is_digit_ch(ch: char) -> int:
    """Return 1 if ch is '0'-'9'.

    ch is a value type (char = i8); TPC has fixed is-None handling for value types
    (always false, generating an i1 0), so no null deref occurs.
    """
    if ch is None:
        return 0
    if ch >= '0' and ch <= '9':
        return 1
    return 0


# ============================================================
# Safe i32 array access (explicit byte offset)
#
# TransPyC's ptr[i] pointer indexing may be treated as byte indexing (i8*) instead of i32*
# indexing, corrupting memory. Must compute the byte offset explicitly: addr = base + idx * 4,
# then cast to i32*. See HandlesTranslator.py:358's t.CUInt64T(ptr) + offset * sizeof usage.
# ============================================================
def _map_get_i32(map_arr: str, idx: int) -> int:
    """Read i32 array map_arr[idx] (explicit byte offset idx*4).

    Casts the byte offset to i32* via (rawptr[int])(uint64), then dereferences with c.Deref.
    TPC has fixed type-union casts (_HandleTypeUnionCastLlvm fallback), so this form works.
    """
    if map_arr is None:
        return -1
    byte_off: uint64 = uint64(map_arr) + idx * 4
    elem_ptr: rawptr[int] = (rawptr[int])(byte_off)
    return *elem_ptr


def _map_set_i32(map_arr: str, idx: int, val: int):
    """Write map_arr[idx] = val (i32 array, explicit byte offset idx*4).

    Casts via (rawptr[int])(uint64) and assigns through c.DerefAs.
    TPC has fixed type-union casts, so this form works.
    """
    if map_arr is None:
        return
    byte_off: uint64 = uint64(map_arr) + idx * 4
    elem_ptr: rawptr[int] = (rawptr[int])(byte_off)
    c.DerefAs(elem_ptr, val)


def _scan_max_ssa_in_text(text: str, cur_max: int) -> int:
    """Scan text for the largest numeric %N and return max(cur_max, found).

    Only matches '%' immediately followed by a digit; skips named SSA names (%entry etc.).
    """
    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: rawptr[Function]) -> int:
    """Walk all BB instruction lines for the largest %N; returns -1 if no numeric SSA names"""
    max_n: int = -1
    if func is None or func.Blocks is None:
        return max_n
    blocks: rawptr[GSList[BasicBlock]] = func.Blocks
    blk: rawptr[BasicBlock] = blocks.Head
    blk_steps: sizet = 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: rawptr[GSList[Line]] = blk.Lines
            line: rawptr[Line] = lines.Head
            line_steps: sizet = 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:
    """Scan text; for each numeric %N, assign next_new if map_arr[N]==-1.

    Returns the updated next_new. Mapping is built in first-occurrence order.
    """
    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: memhub.MemBuddy*,
                         map_arr: str,
                         max_n: int) -> str:
    """Build new text with each %N replaced by its mapped new number.

    Char-by-char scan: a numeric %N is looked up and substituted, others copied as-is.
    Returns a newly allocated buffer (caller need not free the old Buf; pool manages it).
    """
    src_len: sizet = string.strlen(text)
    # New buffer: worst case each number grows in digits; allocate src_len*2 + 64
    new_size: sizet = src_len * 2 + 64
    new_buf: str = pool.alloc(new_size)
    if new_buf is None:
        return None
    # Allocate the digit buffer once (reused in the loop)
    num_buf: str = pool.alloc(16)
    if num_buf is None:
        new_buf[0] = '\0'
        return new_buf

    dst_idx: sizet = 0
    src_p: str = text
    while src_p is not None and src_p[0] != '\0':
        if dst_idx + 16 >= new_size:
            # Buffer too small; truncate
            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)
                # Range-check before reading map_arr: when num > max_n the array holds only
                # (max_n+1)*4 bytes, so a read-before-check would read out of bounds.
                if num >= 0 and num <= max_n:
                    new_num: int = _map_get_i32(map_arr, num)
                    if new_num != -1:
                        # Write '%'
                        new_buf[dst_idx] = '%'
                        dst_idx += 1
                        # Write the new number
                        num_buf[0] = '\0'
                        viperlib.snprintf(num_buf, 16, "%d", new_num)
                        k: sizet = 0
                        while num_buf[k] != '\0' and dst_idx + 1 < new_size:
                            new_buf[dst_idx] = num_buf[k]
                            dst_idx += 1
                            k += 1
                        # Skip the old digit chars
                        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
        # Ordinary char: copy directly
        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: rawptr[Function], pool: memhub.MemBuddy*):
    """Renumber all numeric SSA names in the function so text-output order is monotonic.

    Called before FunctionPrint serializes the body; only applies to define (IsDeclared==0)
    functions with non-empty Blocks.
    """
    if func is None or pool is None:
        return
    if func.IsDeclared == 1:
        return  # declare has no body
    if func.Blocks is None:
        return

    # 1. First pass: find the largest SSA number
    max_n: int = _find_max_ssa_number(func)
    if max_n < 0:
        return  # no numeric SSA names

    # 2. Allocate and init map_arr[0..max_n] = -1; stored as i8* (pool.alloc), accessed via
    #    _map_get_i32/_map_set_i32; memset to 0xFF (-1 per 4-byte i32) to avoid ptr[i] indexing issues.
    map_bytes: sizet = (max_n + 1) * 4  # int = 4 bytes
    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. Second pass: assign new numbers in text order
    next_new: int = 0
    blocks: rawptr[GSList[BasicBlock]] = func.Blocks
    blk: rawptr[BasicBlock] = blocks.Head
    blk_steps: sizet = 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: rawptr[GSList[Line]] = blk.Lines
            line: rawptr[Line] = lines.Head
            line_steps: sizet = 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. Check whether any renumbering is needed (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  # already monotonic; no replacement needed

    # 5. Third pass: replace every instruction line text
    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: rawptr[GSList[Line]] = blk.Lines
            line: rawptr[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


# ============================================================
# Local peephole optimization (text layer, before output, semantics-preserving)
#   T1: ret <iN> zeroinitializer -> ret <iN> 0; ret <ptr*> zeroinitializer -> null
#   T2: when an entry alloca slot has exactly one store with a side-effect-free value
#       (param/%SSA/int/float literal), forward store->load to that value and drop
#       alloca/store/load ("virtual register optimization").
# Safety: only when the alloca and its single store are both in entry, loads anywhere,
# and no other address-taking/aliasing use.
# ============================================================
def _peep_is_startwith(s: str, pfx: str) -> int:
    if s is None or pfx is None:
        return 0
    pl: sizet = 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: memhub.MemBuddy*,
                        line_buf: str, tok: str, repl: str,
                        tlen: sizet, rlen: sizet) -> str:
    """Replace the standalone token %<tok> in line_buf with repl; returns a new buffer (original if unchanged).

    tlen/rlen = strlen(tok)/strlen(repl): computed by the caller outside the line loop
    (the same ltok/safe_val applies to all lines here) to avoid recomputing lengths per line.
    """
    if line_buf is None or tok is None or repl is None:
        return line_buf
    if tlen == 0:
        return line_buf
    # Pre-scan for a hit: no strlen, single NUL-bounded pass (most calls return with no hit,
    # avoiding per-line strlen). Per-input equivalent: tok has no NUL, so strncmp equality
    # implies the token fits; dropping the old i+tlen<=ln bound does not change hit detection.
    hit: int = 0
    i: sizet = 0
    while line_buf[i] != '\0':
        if line_buf[i] == '%' and line_buf[i + 1] == tok[0] \
                and string.strncmp(line_buf + i + 1, tok, tlen) == 0:
            nxt: int = 0
            if line_buf[i + 1 + tlen] != '\0':
                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
    ln: sizet = string.strlen(line_buf)
    # Buffer size bound: at most floor(ln/2) replacements per line, each removing >=2 bytes
    # (%+name) and inserting rlen; worst output <= ln + (ln/2)*rlen + NUL. ln + 1 + rlen*(ln+1)
    # always covers it, preventing overflow (random memory corruption) when rlen is larger.
    out: str = pool.alloc(ln + 1 + rlen * (ln + 1))
    if out is None:
        return line_buf
    op: sizet = 0
    oq: sizet = 0
    i = 0
    while i < ln:
        if line_buf[i] == '%' and i + 1 + tlen <= ln \
                and line_buf[i + 1] == tok[0] \
                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:
                for j in range(rlen):
                    out[oq] = repl[j]
                    oq += 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, nl: sizet) -> int:
    """Whether the line's last '%name' token equals a_name (store/load pointer operand is at line end).

    nl = strlen(a_name): computed by the caller outside the alloca-slot scan (a_name is
    constant within one slot scan) to avoid repeated strlen per line.
    """
    if sb is None or a_name is None:
        return 0
    # Single pass to locate the last '%' (avoids repeated strchr calls, one per '%').
    r: str = None
    probe: str = sb
    while probe[0] != '\0':
        if probe[0] == '%':
            r = probe
        probe += 1
    if r is None:
        return 0
    # Compare the first char first (cheap reject); only strncmp when it matches
    if r[1] == a_name[0] and 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: memhub.MemBuddy*) -> str:
    """Extract name (without %) from '%name = alloca ...'; returns None on failure"""
    if eb is None or eb[0] != '%':
        return None
    eqpos: str = string.strstr(eb, " = alloca ")
    if eqpos is None:
        return None
    nlen: sizet = sizet(uint64(eqpos) - uint64(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 _peep_is_alloc_of(eb: str, a_name: str, nl: sizet) -> int:
    """Whether eb is a_name's alloca definition line; equivalent to strcmp(_peep_alloc_name(eb), a_name)==0.

    Compares the name first (strncmp, reject on mismatch), then locates " = alloca " only when
    bytes match, reducing strstr count from O(slots*lines) to only hit lines per slot.
    nl = strlen(a_name); nlen==0 (empty name) returned None before, so it is rejected here too.
    Equivalence: _peep_alloc_name takes the bytes before the first " = alloca " and after '%',
    so equal-length equal name <=> strncmp match and the substring starts at eb+1+nl.
    """
    if eb is None or a_name is None or nl == 0 or eb[0] != '%':
        return 0
    # First-char prefilter: nl>=1, so strncmp equality implies equal first chars; a mismatch
    # cheaply rejects, keeping most lines out of strncmp. Per-input equivalent: this branch
    # rejects iff the old strncmp would too. eb[0]=='%' is non-NUL, so eb[1] is in-bounds.
    if eb[1] != a_name[0]:
        return 0
    if string.strncmp(eb + 1, a_name, nl) != 0:
        return 0
    eqpos: str = string.strstr(eb, " = alloca ")
    if eqpos is not None and uint64(eqpos) == uint64(eb) + 1 + nl:
        return 1
    return 0


def _peephole_opt(func: rawptr[Function], pool: memhub.MemBuddy*):
    """Text-layer local peephole: T1 ret zeroinitializer; T2 entry-slot store->load forwarding"""
    if func is None or func.Blocks is None or pool is None:
        return
    # NOTE: TPC silently drops assignments for a two-level "var = ptr.field.member" auto-deref
    # chain (the IR reads an uninitialized stack var -> random stack errors). Split into one level:
    # fetch the field into a local rawptr, then null-check before .Head (same pattern as FunctionPrint).
    func_blocks: rawptr[GSList[BasicBlock]] = func.Blocks
    blk: rawptr[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: rawptr[Line] = None
        if blk_t1.Lines is not None:
            t1_lines: rawptr[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: sizet = sizet(uint64(zi) - uint64(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 slot forwarding ----------
    entry_blk: rawptr[BasicBlock] = blk
    if entry_blk is None:
        return
    entry_lines: rawptr[GSList[Line]] = entry_blk.Lines
    if entry_lines is None:
        return
    e_line: rawptr[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:
                # a_name is constant across the slot's full line scan; strlen computed once (not per line)
                a_name_len: sizet = string.strlen(a_name)
                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: int64 = 0
                alloca_order: int = -1
                order: int = 0
                unsafe: int = 0
                # First pass: count this alloca's store/load/other refs and line order
                blk_s = blk
                while blk_s is not None:
                    l_s: rawptr[Line] = None
                    if blk_s.Lines is not None:
                        scan_lines: rawptr[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] == '%':
                                if _peep_is_alloc_of(sb, a_name, a_name_len) != 0:
                                    is_alloc_def = 1
                            if is_alloc_def == 0 and _peep_last_ref_eq(sb, a_name, a_name_len) != 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: sizet = sizet(uint64(vcomma) - uint64(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: sizet = sizet(
                                            uint64(string.strstr(sb, " = load ")) - uint64(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* = (str*)(uint64(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 alloc failed: this load won't enter load_ltoks, but the
                                            # delete phase still removes its def line by slot name, leaving a
                                            # dangling use. Conservatively abandon T2 forwarding for this slot.
                                            unsafe = 1
                                    else:
                                        # Loads exceed record capacity (load_ltoks=8*64, cap 64) or the line has
                                        # no '%' to parse a def name: unrecorded loads would still have their
                                        # " = load " def lines deleted by slot name, leaving dangling uses
                                        # (llc: "use of undefined value", e.g. strcmp short-circuit >64 branch).
                                        # Only a fully recorded load set allows deleting all defs, so
                                        # conservatively abandon the whole T2 forwarding for this slot (IR stays valid).
                                        unsafe = 1
                                else:
                                    unsafe = 1
                        if sb is not None:
                            order += 1
                        l_s = l_s.Next
                    blk_s = blk_s.Next
                # Criteria
                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:
                    # Second pass: replace load-name references
                    # safe_val is constant during this slot's replacement; strlen computed once per slot
                    safe_len: sizet = string.strlen(safe_val)
                    for li2 in range(load_cnt):
                        slotsp: str* = (str*)(uint64(load_ltoks) + li2 * 8)
                        ltok: str = slotsp[0]
                        if ltok is not None:
                            # ltok is constant across the load name's function-wide scan; strlen computed once
                            ltok_len: sizet = string.strlen(ltok)
                            blk_r = blk
                            while blk_r is not None:
                                lr: rawptr[Line] = None
                                if blk_r.Lines is not None:
                                    repl_lines: rawptr[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,
                                                                       ltok_len, safe_len)
                                        if nb2 is not None and nb2 is not lr.Buf:
                                            lr.Buf = nb2
                                    lr = lr.Next
                                blk_r = blk_r.Next
                    # Delete: alloca def / the single store / all load lines
                    blk_d = blk
                    while blk_d is not None:
                        ld: rawptr[Line] = None
                        if blk_d.Lines is not None:
                            kill_lines: rawptr[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
                                if db[0] == '%' and db[1] == a_name[0]:
                                    if _peep_is_alloc_of(db, a_name, a_name_len) != 0:
                                        kill = 1
                                if kill == 0 and _peep_last_ref_eq(db, a_name, a_name_len) != 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 removal ----------
    _peep_noop_bitcast(func, pool)


def _peep_noop_bitcast(func: rawptr[Function], pool: memhub.MemBuddy*):
    """T3: remove no-op bitcast (bitcast <T> %x to <T>, same type on both sides).

    The emitter (virtual calls / union width casts) produces same-type identity bitcasts
    (e.g. `%6 = bitcast i8* %ok_msg to i8*`), purely redundant. Replace all uses of %x with
    the source operand (%SSA/@global/constant token), then delete the line.
    Conservative limit: both type and operand must be simple whitespace-free tokens.
    """
    if func is None or pool is None or func.Blocks is None:
        return
    # Single forward pass: on a valid candidate, replace refs function-wide + delete the line,
    # resuming from the next line; stop at the first non-matching bitcast. Equivalent deletion
    # result to the old rescan-from-block-head approach (replacement only touches operand tokens),
    # but removes the O(K*C) quadratic rescan.
    blk: rawptr[BasicBlock] = func.Blocks.Head
    while blk is not None:
        ln: rawptr[Line] = None
        if blk.Lines is not None:
            ln = blk.Lines.Head
        while ln is not None:
            nxt: rawptr[Line] = ln.Next
            if ln.Buf is not None:
                eqpos: str = string.strstr(ln.Buf, " = bitcast ")
                if eqpos is not None:
                    nlen: sizet = sizet(uint64(eqpos) - uint64(ln.Buf)) - 1
                    nm_buf: str = None
                    src_buf: str = None
                    if nlen > 0:
                        after: str = eqpos + 11  # skip " = bitcast "
                        sp1: str = string.strchr(after, ' ')
                        tpos: str = string.strstr(after, " to ")
                        if sp1 is not None and tpos is not None \
                                and uint64(sp1) < uint64(tpos):
                            tyA_len: sizet = sizet(uint64(sp1) - uint64(after))
                            src_st: str = sp1 + 1
                            src_len: sizet = sizet(uint64(tpos) - uint64(src_st))
                            tyB: str = tpos + 4
                            tyB_len: sizet = string.strlen(tyB)
                            if src_len > 0 and string.strchr(tyB, ' ') is None \
                                    and uint64(string.strchr(src_st, ' ')) == uint64(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'
                    if nm_buf is None or src_buf is None:
                        # First non-matching bitcast: stop (same semantics as before)
                        return
                    # Replace %nm references with the source operand function-wide
                    # nm_buf/src_buf are constant across the scan; strlen computed once (not per line)
                    nm_buf_len: sizet = string.strlen(nm_buf)
                    src_buf_len: sizet = string.strlen(src_buf)
                    blk_r: rawptr[BasicBlock] = func.Blocks.Head
                    while blk_r is not None:
                        if blk_r.Lines is not None:
                            ln2: rawptr[Line] = blk_r.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,
                                                                  nm_buf_len, src_buf_len)
                                    if rb is not None and rb is not ln2.Buf:
                                        ln2.Buf = rb
                                        ln2.Len = string.strlen(rb)
                                ln2 = ln2.Next
                        blk_r = blk_r.Next
                    ln.Buf = None
            ln = nxt
        blk = blk.Next


# ============================================================
# MLIR (llvm dialect) render stage
#
# Located inside FunctionPrint. Builders stay unchanged and keep writing LLVM-syntax
# instruction text into Line; this stage re-parses that text and handles the three
# MLIR-specific differences:
#   1) constant materialization -- MLIR operands cannot be literals: non-SSA operands
#      (int/float/null/zeroinitializer/@symbol) become llvm.mlir.constant / llvm.mlir.zero /
#      llvm.mlir.addressof ops at the entry block head, assigned %cN (once per (type,text) per function)
#   2) phi -> block parameters -- leading phis become block params ^blk(%0: i32); phi lines are
#      not emitted; br/cond_br into that block get arguments matched by source block
#   3) opaque pointers -- all pointers unify to !llvm.ptr (handled by TypePrint/MlirTypeText)
#
# Two passes: pass1 only registers (const table / bitcast no-op alias table) + shape checks;
#             pass2 writes text (const table is complete, so entry-head const ops can be emitted first).
# ============================================================

# Table capacity
define MLIR_CONST_CAP = 512
define MLIR_ALIAS_CAP = 512

# Render pass: 1=scan(register) 2=emit(write text)
_MLIR_Pass: int = 0
_MLIR_Pool: memhub.MemBuddy* = None
_MLIR_Func: rawptr[Function] = None
_MLIR_SrcBlk: rawptr[BasicBlock] = None

# Constant table: (MLIR type text, constant text) -> materialized SSA name
_MLIR_ConstTy: t.CPtr = None
_MLIR_ConstVal: t.CPtr = None
_MLIR_ConstName: t.CPtr = None
_MLIR_ConstCount: int = 0

# bitcast no-op alias table + cmpxchg result binding table: result name -> source operand / extractvalue result name
# (cmpxchg's extractvalue result name is deterministically derived from the result name, not a counter; see _MlirCxResName)
_MLIR_AliasFrom: t.CPtr = None
_MLIR_AliasTo: t.CPtr = None
_MLIR_AliasCount: int = 0

# The intermediate instruction text line currently being rendered (set at _MlirRenderLine entry).
# Supplies line context to type-text conversion failures: MlirFail terminates immediately with no
# stack, so the line text is stashed in __types.vp (via MlirSetContextLine) and printed with the error.
_MLIR_CurLine: str = None

# Parse-time scratch buffers (allocated once per FunctionPrint, reused throughout;
# text that must outlive a line is dup'd on table/alias insertion, see _mlir_dup)
_SC_Op: str = None
_SC_Tb: str = None
_SC_Ty: str = None
_SC_Ty2: str = None
_SC_V1: str = None
_SC_V2: str = None
_SC_V3: str = None
_SC_Callee: str = None
_SC_Args: str = None
_SC_ArgTs: str = None
_SC_IdxV: str = None
_SC_IdxT: str = None
_SC_Mt: str = None
_SC_Tmp: str = None
_SC_Tmp2: str = None
_SC_P1: str = None
_SC_P2: str = None
_SC_B1: str = None
_SC_B2: str = None
_SC_Buf: str = None


def _mlir_slotp(buf: t.CPtr, idx: int) -> str*:
    """Table slot address (buf + idx*8)"""
    return (str*)(uint64(buf) + idx * 8)


def _mlir_dup(s: str) -> str:
    """Copy text into the pool (for table entries that must outlive a line)"""
    if s is None:
        return None
    n: sizet = string.strlen(s)
    b: str = _MLIR_Pool.alloc(n + 1)
    if b is None:
        MlirFail("MLIR 渲染: 文本 dup 分配失败")
    string.strncpy(b, s, n)
    b[n] = '\0'
    return b


def _MlirTablesInit(pool: memhub.MemBuddy*):
    """Initialize the two tables and the parse scratch buffers"""
    global _MLIR_Pool, _MLIR_Pass, _MLIR_ConstTy, _MLIR_ConstVal, _MLIR_ConstName, _MLIR_ConstCount, _MLIR_AliasFrom, _MLIR_AliasTo, _MLIR_AliasCount, _SC_Op, _SC_Tb, _SC_Ty, _SC_Ty2, _SC_V1, _SC_V2, _SC_V3, _SC_Callee, _SC_Args, _SC_ArgTs, _SC_IdxV, _SC_IdxT, _SC_Mt, _SC_Tmp, _SC_Tmp2, _SC_P1, _SC_P2, _SC_B1, _SC_B2, _SC_Buf
    if pool is None:
        MlirFail("MLIR 渲染: 内存池为空")
    _MLIR_Pool = pool
    _MLIR_Pass = 0
    _MLIR_ConstCount = 0
    _MLIR_AliasCount = 0
    _MLIR_ConstTy = pool.alloc(MLIR_CONST_CAP * 8)
    _MLIR_ConstVal = pool.alloc(MLIR_CONST_CAP * 8)
    _MLIR_ConstName = pool.alloc(MLIR_CONST_CAP * 8)
    _MLIR_AliasFrom = pool.alloc(MLIR_ALIAS_CAP * 8)
    _MLIR_AliasTo = pool.alloc(MLIR_ALIAS_CAP * 8)
    if _MLIR_ConstTy is None or _MLIR_ConstVal is None or _MLIR_ConstName is None \
            or _MLIR_AliasFrom is None or _MLIR_AliasTo is None:
        MlirFail("MLIR 渲染: 常量/别名表分配失败")
    string.memset(_MLIR_ConstTy, 0, MLIR_CONST_CAP * 8)
    string.memset(_MLIR_ConstVal, 0, MLIR_CONST_CAP * 8)
    string.memset(_MLIR_ConstName, 0, MLIR_CONST_CAP * 8)
    string.memset(_MLIR_AliasFrom, 0, MLIR_ALIAS_CAP * 8)
    string.memset(_MLIR_AliasTo, 0, MLIR_ALIAS_CAP * 8)
    _SC_Op = pool.alloc(64)
    _SC_Tb = pool.alloc(512)
    _SC_Ty = pool.alloc(4096)
    _SC_Ty2 = pool.alloc(4096)
    _SC_V1 = pool.alloc(512)
    _SC_V2 = pool.alloc(512)
    _SC_V3 = pool.alloc(512)
    _SC_Callee = pool.alloc(512)
    _SC_Args = pool.alloc(2048)
    _SC_ArgTs = pool.alloc(8192)
    _SC_IdxV = pool.alloc(1024)
    _SC_IdxT = pool.alloc(4096)
    _SC_Mt = pool.alloc(4096)
    _SC_Tmp = pool.alloc(8192)
    _SC_Tmp2 = pool.alloc(4096)
    _SC_P1 = pool.alloc(512)
    _SC_P2 = pool.alloc(512)
    _SC_B1 = pool.alloc(4096)
    _SC_B2 = pool.alloc(4096)
    _SC_Buf = pool.alloc(4096)
    if _SC_Op is None or _SC_Ty is None or _SC_Mt is None or _SC_Tmp is None:
        MlirFail("MLIR 渲染: 解析 scratch 分配失败")
    _SC_Op[0] = '\0'
    _SC_Ty[0] = '\0'
    _SC_Mt[0] = '\0'
    _SC_Tmp[0] = '\0'


# ============================================================
# Constant table / alias table
# ============================================================
def _MlirAliasLookup(name: str) -> str:
    """Look up the alias table; returns None on miss"""
    if name is None:
        return None
    for i in range(_MLIR_AliasCount):
        sp: str* = _mlir_slotp(_MLIR_AliasFrom, i)
        if sp[0] is not None and string.strcmp(sp[0], name) == 0:
            st: str* = _mlir_slotp(_MLIR_AliasTo, i)
            return st[0]
    return None


def _MlirAliasAdd(from_name: str, to_name: str):
    """Register a no-op bitcast alias (from is replaced by to)"""
    global _MLIR_AliasCount
    if from_name is None or to_name is None:
        return
    if _MLIR_AliasCount >= MLIR_ALIAS_CAP:
        MlirFail("MLIR 渲染: bitcast 别名表溢出")
    sp: str* = _mlir_slotp(_MLIR_AliasFrom, _MLIR_AliasCount)
    st: str* = _mlir_slotp(_MLIR_AliasTo, _MLIR_AliasCount)
    sp[0] = _mlir_dup(from_name)
    st[0] = _mlir_dup(to_name)
    _MLIR_AliasCount += 1


def _MlirResolve(name: str) -> str:
    """Resolve the alias chain (bitcast no-op result name -> real operand)"""
    if name is None:
        return None
    cur: str = name
    for g in range(64):
        nx: str = _MlirAliasLookup(cur)
        if nx is None:
            return cur
        cur = nx
    return cur


def _MlirConstGet(tytext: str, valtext: str) -> str:
    """Get (or create) the constant's materialized name %cN"""
    global _MLIR_ConstCount
    for i in range(_MLIR_ConstCount):
        sp: str* = _mlir_slotp(_MLIR_ConstTy, i)
        if sp[0] is not None and string.strcmp(sp[0], tytext) == 0:
            sv: str* = _mlir_slotp(_MLIR_ConstVal, i)
            if sv[0] is not None and string.strcmp(sv[0], valtext) == 0:
                sn: str* = _mlir_slotp(_MLIR_ConstName, i)
                return sn[0]
    if _MLIR_Pass == 2:
        # pass1 already scanned all instruction lines; no new constant should appear here
        MlirFail("MLIR 渲染: 发射阶段出现未物化常量（扫描遗漏）")
    if _MLIR_ConstCount >= MLIR_CONST_CAP:
        MlirFail("MLIR 渲染: 常量表溢出")
    nm: str = _MLIR_Pool.alloc(32)
    if nm is None:
        MlirFail("MLIR 渲染: 常量名分配失败")
    viperlib.snprintf(nm, 32, "%%c%d", _MLIR_ConstCount)
    sp2: str* = _mlir_slotp(_MLIR_ConstTy, _MLIR_ConstCount)
    sv2: str* = _mlir_slotp(_MLIR_ConstVal, _MLIR_ConstCount)
    sn2: str* = _mlir_slotp(_MLIR_ConstName, _MLIR_ConstCount)
    sp2[0] = _mlir_dup(tytext)
    sv2[0] = _mlir_dup(valtext)
    sn2[0] = nm
    _MLIR_ConstCount += 1
    return nm


def _MlirCxResName(resname: str, suffix: str) -> str:
    """cmpxchg's extractvalue result name: deterministically derived from the struct result name.

    A bare MLIR SSA name's first char must be a letter or `_` (mlir-opt 22.1.8 measured: `%0_v`
    errors "expected '=' after SSA name"), so a `_` is prefixed to the result name body:
      `%0` -> `%_0_v` (old value [0]) / `%_0_ok` (success bit [1])
      `%a` -> `%_a_v` / `%_a_ok`
    Names map one-to-one to result names with no counter, so pass1 (alias registration) and
    pass2 (emission) produce identical names and references never dangle.
    """
    nm: str = _MLIR_Pool.alloc(64)
    if nm is None:
        MlirFail("MLIR 渲染: cmpxchg 结果名分配失败")
    nm[0] = '\0'
    _append_cstr(nm, 4096, "%_")
    _append_cstr(nm, 4096, resname + 1)
    _append_cstr(nm, 4096, suffix)
    return nm


def _MlirFloatLiteral(val: str, mt: str) -> str:
    """Float constant text -> MLIR float literal text.

    MLIR float literals have two forms (Builtin Dialect FloatAttr syntax:
    `float-literal (':' float-type)?` | `hexadecimal-literal ':' float-type`):
      - decimal float literal (e.g. `3.14 : f64`);
      - hexadecimal bit pattern (e.g. `0x40091EB851EB851F : f64`, interpreted bitwise).

    So the `0x...` bit pattern from ConstFloat passes through as-is (no precision loss);
    converting it to `%e` (7 significant digits) would truncate and defeat the bit-pattern intent.
    The pattern width must match the type (f16/f32/f64/f128 <-> 4/8/16/32 hex digits), else a loud
    error, since MLIR interprets the hex value by the type width.
    """
    if val[0] == '0' and (val[1] == 'x' or val[1] == 'X'):
        want: int = 0
        if string.strcmp(mt, "f16") == 0:
            want = 4
        elif string.strcmp(mt, "f32") == 0:
            want = 8
        elif string.strcmp(mt, "f64") == 0:
            want = 16
        elif string.strcmp(mt, "f128") == 0:
            want = 32
        if want == 0:
            MlirFail("MLIR 渲染: 十六进制浮点字面量的类型既非 f16/f32/f64/f128")
        nd: int = 0
        p: str = val + 2
        while (p[0] >= '0' and p[0] <= '9') \
                or (p[0] >= 'a' and p[0] <= 'f') \
                or (p[0] >= 'A' and p[0] <= 'F'):
            nd += 1
            p = p + 1
        if nd != want:
            MlirFail("MLIR 渲染: 浮点位模式位数与类型不符（f16/f32/f64/f128 依次需 4/8/16/32 位十六进制）")
        return val
    # Already decimal text: use as-is
    return val


def _MlirEmitOneConst(out: str, outsz: sizet, mt: str, val: str, nm: str):
    """Emit one constant-materialization op"""
    _append_cstr(out, outsz, "  ")
    _append_cstr(out, outsz, nm)
    if val[0] == '@':
        # Symbol reference (global/function address): use the unified quoting rule
        # (LLVM and MLIR bare-name rules differ: MLIR disallows leading digit/'.' and '-')
        _append_cstr(out, outsz, " = llvm.mlir.addressof ")
        _MlirSymRefAppend(out, outsz, val + 1)
        _append_cstr(out, outsz, " : ")
        _append_cstr(out, outsz, mt)
    elif string.strcmp(val, "null") == 0 or string.strcmp(val, "zeroinitializer") == 0:
        _append_cstr(out, outsz, " = llvm.mlir.zero : ")
        _append_cstr(out, outsz, mt)
    else:
        lit: str = val
        if mt[0] == 'f':
            lit = _MlirFloatLiteral(val, mt)
        elif mt[0] != 'i':
            MlirFail("MLIR 渲染: 不支持的常量操作数形态")
        _append_cstr(out, outsz, " = llvm.mlir.constant(")
        _append_cstr(out, outsz, lit)
        # i1's `true`/`false` are MLIR bare boolean keywords, no type suffix allowed:
        # `llvm.mlir.constant(false : i1)` makes mlir-opt error "expected ')'".
        # Other scalars must be written `value : type` (otherwise treated as an untyped attr, mismatching the result type).
        if string.strcmp(lit, "true") != 0 and string.strcmp(lit, "false") != 0:
            _append_cstr(out, outsz, " : ")
            _append_cstr(out, outsz, mt)
        _append_cstr(out, outsz, ") : ")
        _append_cstr(out, outsz, mt)
    _append_cstr(out, outsz, "\n")


def _MlirEmitConsts(out: str, outsz: sizet):
    """Emit all constant-materialization ops at the entry block head in registration order"""
    for i in range(_MLIR_ConstCount):
        st: str* = _mlir_slotp(_MLIR_ConstTy, i)
        sv: str* = _mlir_slotp(_MLIR_ConstVal, i)
        sn: str* = _mlir_slotp(_MLIR_ConstName, i)
        _MlirEmitOneConst(out, outsz, st[0], sv[0], sn[0])


# ============================================================
# Parsing primitives for intermediate instruction text
# ============================================================
def _mlir_ws(p: str) -> str:
    """Skip whitespace (space/tab/newline)"""
    if p is None:
        return None
    while p[0] == ' ' or p[0] == '\t' or p[0] == '\n' or p[0] == '\r':
        p = p + 1
    return p


def _mlir_delim(ch: int) -> int:
    """Whether ch ends a token"""
    if ch == 0 or ch == 32 or ch == 9 or ch == 10 or ch == 13:
        return 1
    if ch == ',' or ch == '(' or ch == ')' or ch == '[' or ch == ']' \
            or ch == ':' or ch == '=':
        return 1
    return 0


def _mlir_read_value(p: str, out: str, outsz: sizet) -> str:
    """Read a value token (%name / @"name" / %"name" / literal); returns the new cursor"""
    p = _mlir_ws(p)
    n: sizet = 0
    if p[0] == '\0':
        out[0] = '\0'
        return p
    if p[0] == '%' or p[0] == '@':
        if n + 1 < outsz:
            out[n] = p[0]
            n += 1
        p = p + 1
    if p[0] == '"':
        if n + 1 < outsz:
            out[n] = '"'
            n += 1
        p = p + 1
        while p[0] != '\0' and p[0] != '"':
            if n + 1 < outsz:
                out[n] = p[0]
                n += 1
            p = p + 1
        if p[0] == '"':
            if n + 1 < outsz:
                out[n] = '"'
                n += 1
            p = p + 1
        out[n] = '\0'
        return p
    while p[0] != '\0' and _mlir_delim(p[0]) == 0:
        if n + 1 < outsz:
            out[n] = p[0]
            n += 1
        p = p + 1
    out[n] = '\0'
    return p


def _mlir_read_type(p: str, out: str, outsz: sizet, nofunc: int = 0) -> str:
    """Read a type token (MLIR form !llvm... or scalar iN/fN, also LLVM aggregate forms).

    Bracket balancing covers `[] {} () <>`: MLIR's !llvm.array<...>/!llvm.struct<(...)> contain
    spaces and commas and must be read whole by depth counting; bracketless types (!llvm.ptr)
    end at a top-level delimiter.

    With nofunc=1, does NOT absorb the following ` (...)` function-type suffix: in a call's return
    type position, `(...)` (LLVM syntax) is the callee's parameter type list, not part of the return
    type, so it must be left to _MlirCall's variadic branch (see the comment there).
    """
    p = _mlir_ws(p)
    n: sizet = 0
    if p[0] == '\0':
        out[0] = '\0'
        return p
    c0: int = p[0]
    if c0 == '!' or c0 == '[' or c0 == '{' or c0 == '(':
        d: int = 0
        while p[0] != '\0':
            ch: int = p[0]
            if ch == '[' or ch == '{' or ch == '(' or ch == '<':
                d += 1
            elif ch == ']' or ch == '}' or ch == ')' or ch == '>':
                if n + 1 < outsz:
                    out[n] = ch
                    n += 1
                p = p + 1
                d -= 1
                if d <= 0:
                    break
                continue
            elif d == 0 and (ch == ' ' or ch == ',' or ch == '\t'
                             or ch == '\n' or ch == '\r'):
                # Top-level delimiter: bracketless types (e.g. !llvm.ptr) end here
                break
            if n + 1 < outsz:
                out[n] = ch
                n += 1
            p = p + 1
        while p[0] == '*':
            if n + 1 < outsz:
                out[n] = '*'
                n += 1
            p = p + 1
        out[n] = '\0'
        return p
    if c0 == '%':
        p = _mlir_read_value(p, out, outsz)
        n = string.strlen(out)
        while p[0] == '*':
            if n + 1 < outsz:
                out[n] = '*'
                n += 1
            p = p + 1
        out[n] = '\0'
        return p
    # Ordinary word (i32 / i8* / void / double ...)
    while p[0] != '\0' and _mlir_delim(p[0]) == 0:
        if n + 1 < outsz:
            out[n] = p[0]
            n += 1
        p = p + 1
    # Function-type suffix " ( ... )" (not absorbed when nofunc=1, see the docstring)
    q: str = _mlir_ws(p)
    if nofunc == 0 and q[0] == '(':
        if n + 1 < outsz:
            out[n] = ' '
            n += 1
        p = q
        d2: int = 0
        while p[0] != '\0':
            ch2: int = p[0]
            if ch2 == '(':
                d2 += 1
            elif ch2 == ')':
                if n + 1 < outsz:
                    out[n] = ch2
                    n += 1
                p = p + 1
                d2 -= 1
                if d2 <= 0:
                    break
                continue
            if n + 1 < outsz:
                out[n] = ch2
                n += 1
            p = p + 1
    while p[0] == '*':
        if n + 1 < outsz:
            out[n] = '*'
            n += 1
        p = p + 1
    out[n] = '\0'
    return p


def _mlir_in_set(ob: str, setstr: str) -> int:
    """Whether opcode belongs to a "|a|b|c|"-form set"""
    tb: str = _SC_Tb
    tb[0] = '\0'
    string.strcat(tb, "|")
    string.strcat(tb, ob)
    string.strcat(tb, "|")
    if string.strstr(setstr, tb) is not None:
        return 1
    return 0


def _MlirBlkRef(t: str) -> str:
    """Block reference text: strip the leading '%' (MLIR blocks use ^name)"""
    if t is None:
        return ""
    if t[0] == '%':
        return t + 1
    return t


def _MlirFindBlock(func: rawptr[Function], name: str) -> rawptr[BasicBlock]:
    """Find a basic block by name (without %) within the function"""
    if func is None or func.Blocks is None or name is None:
        return None
    blocks: rawptr[GSList[BasicBlock]] = func.Blocks
    blk: rawptr[BasicBlock] = blocks.Head
    steps: int = 0
    while blk is not None:
        steps += 1
        if steps > IR_DUMP_CHAIN_LIMIT:
            MlirFail("MLIR 渲染: 块链疑似成环")
        if blk.Name is not None and string.strcmp(blk.Name, name) == 0:
            return blk
        blk = blk.Next
    return None


def _MlirIsPhiLine(text: str) -> int:
    """Whether the line is a phi line"""
    if text is None:
        return 0
    if string.strstr(text, " = phi ") is not None:
        return 1
    return 0


# ============================================================
# Operand rendering (with constant materialization)
# ============================================================
def _MlirOperand(val: str, tyllvm: str, out: str, outsz: sizet) -> str:
    """Render one operand and return its final text.

    SSA value -> %name after alias resolution;
    non-SSA (literal/symbol ref) -> materialized %cN (registered during the scan pass).
    """
    if val is None or val[0] == '\0':
        MlirFail("MLIR 渲染: 操作数为空")
    if val[0] == '%':
        r: str = _MlirResolve(val)
        _append_cstr(out, outsz, r)
        return r
    mt: str = _SC_Mt
    mt[0] = '\0'
    MlirTypeText(mt, 4096, tyllvm, _MLIR_Pool)
    nm: str = _MlirConstGet(mt, val)
    _append_cstr(out, outsz, nm)
    return nm


# ============================================================
# phi parsing
# ============================================================
def _MlirPhiScan(text: str):
    """Scan a phi line: register all incoming constants (scan pass only)"""
    p: str = _mlir_ws(text)
    tb: str = _SC_Buf
    p = _mlir_read_value(p, tb, 512)
    p = _mlir_ws(p)
    if p[0] == '=':
        p = _mlir_ws(p + 1)
    ob: str = _SC_Op
    p = _mlir_read_value(p, ob, 64)
    if string.strcmp(ob, "phi") != 0:
        return
    tyb: str = _SC_Ty
    p = _mlir_read_type(p, tyb, 512)
    g: int = 0
    while p[0] != '\0':
        g += 1
        if g > 512:
            MlirFail("MLIR 渲染: phi 入边过多")
        p = _mlir_ws(p)
        if p[0] != '[':
            break
        p = _mlir_ws(p + 1)
        vb: str = _SC_P1
        p = _mlir_read_value(p, vb, 512)
        p = _mlir_ws(p)
        if p[0] == ',':
            p = _mlir_ws(p + 1)
        bb: str = _SC_P2
        p = _mlir_read_value(p, bb, 512)
        p = _mlir_ws(p)
        if p[0] == ']':
            p = p + 1
        p = _mlir_ws(p)
        if p[0] == ',':
            p = p + 1
        # Register this incoming value (constants/symbol refs need materialization)
        _MlirOperand(vb, tyb, None, 0)


def _MlirPhiParam(text: str) -> int:
    """Take a phi's result name and type as block params; on success writes _SC_B1(name)/_SC_B2(type) and returns 1"""
    p: str = _mlir_ws(text)
    nm: str = _SC_B1
    p = _mlir_read_value(p, nm, 512)
    p = _mlir_ws(p)
    if p[0] != '=':
        return 0
    p = _mlir_ws(p + 1)
    ob: str = _SC_Op
    p = _mlir_read_value(p, ob, 64)
    if string.strcmp(ob, "phi") != 0:
        return 0
    tyb: str = _SC_B2
    _mlir_read_type(p, tyb, 512)
    return 1


def _MlirPhiFindIncoming(text: str, src: rawptr[BasicBlock]) -> int:
    """Find the phi incoming edge matching the source block; on success writes _SC_B1(value)/_SC_B2(type) and returns 1"""
    p: str = _mlir_ws(text)
    tb: str = _SC_Buf
    p = _mlir_read_value(p, tb, 512)
    p = _mlir_ws(p)
    if p[0] == '=':
        p = _mlir_ws(p + 1)
    ob: str = _SC_Op
    p = _mlir_read_value(p, ob, 64)
    if string.strcmp(ob, "phi") != 0:
        return 0
    tyb: str = _SC_Ty
    p = _mlir_read_type(p, tyb, 512)
    srcn: str = None
    if src is not None:
        srcn = src.Name
    g: int = 0
    while p[0] != '\0':
        g += 1
        if g > 512:
            MlirFail("MLIR 渲染: phi 入边过多")
        p = _mlir_ws(p)
        if p[0] != '[':
            break
        p = _mlir_ws(p + 1)
        vb: str = _SC_P1
        p = _mlir_read_value(p, vb, 512)
        p = _mlir_ws(p)
        if p[0] == ',':
            p = _mlir_ws(p + 1)
        bb: str = _SC_P2
        p = _mlir_read_value(p, bb, 512)
        p = _mlir_ws(p)
        if p[0] == ']':
            p = p + 1
        p = _mlir_ws(p)
        if p[0] == ',':
            p = p + 1
        if srcn is not None:
            bn: str = bb
            if bn[0] == '%':
                bn = bn + 1
            if string.strcmp(bn, srcn) == 0:
                string.strcpy(_SC_B1, vb)
                string.strcpy(_SC_B2, tyb)
                return 1
    return 0


# ============================================================
# Block parameters / block arguments
# ============================================================
def _MlirBlockHasPhi(blk: rawptr[BasicBlock]) -> int:
    """Whether the block head has a phi (i.e. the block carries block params)"""
    if blk is None or blk.Lines is None:
        return 0
    lines: rawptr[GSList[Line]] = blk.Lines
    ln: rawptr[Line] = lines.Head
    while ln is not None:
        if ln.Buf is not None:
            if _MlirIsPhiLine(ln.Buf) != 0:
                return 1
            return 0
        ln = ln.Next
    return 0


def _MlirBlockOperands(out: str, outsz: sizet,
                       target: rawptr[BasicBlock], src: rawptr[BasicBlock]):
    """Emit target's block arguments when branching to it: (%v : ty, ...).

    Incoming edges are matched to phis by source block; no match -> loud error (never silently drop args).
    """
    if target is None or src is None:
        return
    if _MlirBlockHasPhi(target) == 0:
        return
    tmp: str = _SC_Tmp
    tmp[0] = '\0'
    lines: rawptr[GSList[Line]] = target.Lines
    ln: rawptr[Line] = lines.Head
    cnt: int = 0
    while ln is not None:
        if ln.Buf is None:
            ln = ln.Next
            continue
        if _MlirIsPhiLine(ln.Buf) == 0:
            break
        if _MlirPhiFindIncoming(ln.Buf, src) == 0:
            MlirFail("MLIR 渲染: 跳转目标块带 phi，但找不到与源块匹配的入边")
        if cnt > 0:
            _append_cstr(tmp, 8192, ", ")
        _MlirOperand(_SC_B1, _SC_B2, tmp, 4096)
        _append_cstr(tmp, 8192, " : ")
        mt: str = _SC_Mt
        mt[0] = '\0'
        MlirTypeText(mt, 4096, _SC_B2, _MLIR_Pool)
        _append_cstr(tmp, 8192, mt)
        cnt += 1
        ln = ln.Next
    if cnt == 0:
        MlirFail("MLIR 渲染: 目标块带 phi 但未生成块实参")
    _append_cstr(out, outsz, "(")
    _append_cstr(out, outsz, tmp)
    _append_cstr(out, outsz, ")")


def _MlirEmitBlockHeader(out: str, outsz: sizet, blk: rawptr[BasicBlock]):
    """Emit a non-entry block header: ^name(%p: ty, ...):"""
    _append_cstr(out, outsz, "^")
    if blk.Name is not None:
        _append_cstr(out, outsz, blk.Name)
    if _MlirBlockHasPhi(blk) != 0:
        tmp: str = _SC_Tmp
        tmp[0] = '\0'
        lines: rawptr[GSList[Line]] = blk.Lines
        ln: rawptr[Line] = lines.Head
        cnt: int = 0
        while ln is not None:
            if ln.Buf is None:
                ln = ln.Next
                continue
            if _MlirIsPhiLine(ln.Buf) == 0:
                break
            if _MlirPhiParam(ln.Buf) != 0:
                if cnt > 0:
                    _append_cstr(tmp, 8192, ", ")
                _append_cstr(tmp, 8192, _SC_B1)
                _append_cstr(tmp, 8192, ": ")
                mt: str = _SC_Mt
                mt[0] = '\0'
                MlirTypeText(mt, 4096, _SC_B2, _MLIR_Pool)
                _append_cstr(tmp, 8192, mt)
                cnt += 1
            ln = ln.Next
        if cnt == 0:
            MlirFail("MLIR 渲染: 块首 phi 解析失败")
        _append_cstr(out, outsz, "(")
        _append_cstr(out, outsz, tmp)
        _append_cstr(out, outsz, ")")
    _append_cstr(out, outsz, ":\n")


def _MlirEmitBlockLines(out: str, outsz: sizet, blk: rawptr[BasicBlock]):
    """Emit in-block instructions (skip leading phis, which became block params)"""
    if blk is None or blk.Lines is None:
        return
    lines: rawptr[GSList[Line]] = blk.Lines
    ln: rawptr[Line] = lines.Head
    skipping: int = 1
    steps: int = 0
    while ln is not None:
        steps += 1
        if steps > IR_DUMP_CHAIN_LIMIT:
            MlirFail("MLIR 渲染: 块指令链疑似成环")
        if ln.Buf is not None:
            if skipping != 0 and _MlirIsPhiLine(ln.Buf) != 0:
                ln = ln.Next
                continue
            skipping = 0
            _MlirRenderLine(ln.Buf, out, outsz)
        ln = ln.Next


def _MlirCheckEntryNoPhi(blk: rawptr[BasicBlock]):
    """entry block must not have phis (no predecessor, cannot become block params)"""
    if _MlirBlockHasPhi(blk) != 0:
        MlirFail("MLIR 渲染: entry 块不允许 phi（MLIR entry 块不接受块参数）")


# ============================================================
# Per-instruction MLIR rendering
# ============================================================
def _MlirEmitRes(out: str, outsz: sizet, resname: str):
    """Emit the "  [%N = ]" prefix"""
    _append_cstr(out, outsz, "  ")
    if resname is not None:
        _append_cstr(out, outsz, resname)
        _append_cstr(out, outsz, " = ")


def _MlirBinop(p: str, resname: str, ob: str, out: str, outsz: sizet):
    """Integer/float binary op: llvm.<op> %a, %b : ty"""
    tyb: str = _SC_Ty
    p = _mlir_read_type(p, tyb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    lb: str = _SC_V1
    p = _mlir_read_value(p, lb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    rb: str = _SC_V2
    _mlir_read_value(p, rb, 512)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.")
    _append_cstr(out, outsz, ob)
    _append_cstr(out, outsz, " ")
    _MlirOperand(lb, tyb, out, outsz)
    _append_cstr(out, outsz, ", ")
    _MlirOperand(rb, tyb, out, outsz)
    mt: str = _SC_Mt
    mt[0] = '\0'
    MlirTypeText(mt, 4096, tyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, mt)
    _append_cstr(out, outsz, "\n")


def _MlirFneg(p: str, resname: str, out: str, outsz: sizet):
    """fneg: llvm.fneg %a : ty"""
    tyb: str = _SC_Ty
    p = _mlir_read_type(p, tyb, 512)
    vb: str = _SC_V1
    _mlir_read_value(p, vb, 512)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.fneg ")
    _MlirOperand(vb, tyb, out, outsz)
    mt: str = _SC_Mt
    mt[0] = '\0'
    MlirTypeText(mt, 4096, tyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, mt)
    _append_cstr(out, outsz, "\n")


def _MlirCmp(p: str, resname: str, ob: str, out: str, outsz: sizet):
    """icmp/fcmp: llvm.icmp "sgt" %a, %b : ty

    The intermediate text is LLVM syntax `icmp <pred> <type> <lhs>, <rhs>` -- predicate before type,
    so the read order must be predicate -> type -> operands.
    Do not read the type first: the predicate (e.g. `sgt`) would be passed as type text, triggering
    "cannot convert to MLIR type text" and aborting.
    """
    pb: str = _SC_P1
    p = _mlir_read_value(p, pb, 512)
    p = _mlir_ws(p)
    tyb: str = _SC_Ty
    p = _mlir_read_type(p, tyb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    lb: str = _SC_V1
    p = _mlir_read_value(p, lb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    rb: str = _SC_V2
    _mlir_read_value(p, rb, 512)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.")
    _append_cstr(out, outsz, ob)
    _append_cstr(out, outsz, " \"")
    _append_cstr(out, outsz, pb)
    _append_cstr(out, outsz, "\" ")
    _MlirOperand(lb, tyb, out, outsz)
    _append_cstr(out, outsz, ", ")
    _MlirOperand(rb, tyb, out, outsz)
    mt: str = _SC_Mt
    mt[0] = '\0'
    MlirTypeText(mt, 4096, tyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, mt)
    _append_cstr(out, outsz, "\n")


def _MlirAlloca(p: str, resname: str, out: str, outsz: sizet):
    """alloca: llvm.alloca %cN x ty {alignment = N : i64} : (i32) -> !llvm.ptr

    The MLIR form of the alignment attribute was determined by measuring mlir-opt 22.1.8
    (do NOT rely on memory): `{alignment = 16 : i64}` -> exit 0 (adopted);
    `align(16)` / `, align 16` (LLVM syntax) -> exit 1 (expected ':'); `{alignment = 16 : i32}`
    -> exit 1 (requires i64). So the intermediate `, align N` must be rewritten to `{alignment = N : i64}`.

    Alignment accepts only a decimal constant that is a power of two: mlir-opt does not check
    power-of-two (`{alignment = 3 : i64}` exits 0), but non-powers are illegal in LLVM semantics, so this
    layer gates it: non-constant/non-power/out-of-range (> 2^30) always errors loudly.
    """
    tyb: str = _SC_Ty
    p = _mlir_read_type(p, tyb, 512)
    p = _mlir_ws(p)
    aligntxt: str = _SC_P2
    aligntxt[0] = '\0'
    if p[0] == ',':
        # ", align N" (the intermediate-text form from build_alloca_aligned)
        p = _mlir_ws(p + 1)
        if string.strncmp(p, "align", 5) != 0:
            MlirFail("MLIR 渲染: alloca 只支持 align 属性（未知属性）")
        p = _mlir_ws(p + 5)
        nd: int = 0
        while p[0] >= '0' and p[0] <= '9':
            if nd + 1 >= 512:
                MlirFail("MLIR 渲染: alloca 对齐值过长（超范围）")
            aligntxt[nd] = p[0]
            nd += 1
            p = p + 1
        aligntxt[nd] = '\0'
        p = _mlir_ws(p)
        if p[0] != '\0':
            MlirFail("MLIR 渲染: alloca 的 align 之后出现多余记号")
        if nd == 0:
            MlirFail("MLIR 渲染: alloca 对齐值必须是常量（常量十进制，非 SSA）")
        if nd > 10:
            MlirFail("MLIR 渲染: alloca 对齐值超范围（须为 2 的幂且 <= 2^30）")
        av: int = string.atoi(aligntxt)
        if av <= 0 or av > 1073741824:
            MlirFail("MLIR 渲染: alloca 对齐值超范围（须为 2 的幂且 <= 2^30）")
        if (av & (av - 1)) != 0:
            MlirFail("MLIR 渲染: alloca 对齐值必须是 2 的幂")
    mt: str = _SC_Mt
    mt[0] = '\0'
    MlirTypeText(mt, 4096, tyb, _MLIR_Pool)
    cnt: str = _MlirConstGet("i32", "1")
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.alloca ")
    _append_cstr(out, outsz, cnt)
    _append_cstr(out, outsz, " x ")
    _append_cstr(out, outsz, mt)
    if aligntxt[0] != '\0':
        _append_cstr(out, outsz, " {alignment = ")
        _append_cstr(out, outsz, aligntxt)
        _append_cstr(out, outsz, " : i64}")
    _append_cstr(out, outsz, " : (i32) -> !llvm.ptr\n")


def _MlirLoad(p: str, resname: str, out: str, outsz: sizet):
    """load: llvm.load %p : !llvm.ptr -> ty"""
    p = _mlir_ws(p)
    vol: int = 0
    if string.strncmp(p, "volatile ", 9) == 0:
        vol = 1
        p = _mlir_ws(p + 9)
    tyb: str = _SC_Ty
    p = _mlir_read_type(p, tyb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    ptyb: str = _SC_Ty2
    p = _mlir_read_type(p, ptyb, 512)
    p = _mlir_ws(p)
    pb: str = _SC_V1
    _mlir_read_value(p, pb, 512)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.load ")
    if vol != 0:
        _append_cstr(out, outsz, "volatile ")
    _MlirOperand(pb, ptyb, out, outsz)
    tmt: str = _SC_Mt
    tmt[0] = '\0'
    MlirTypeText(tmt, 4096, tyb, _MLIR_Pool)
    pmt: str = _SC_Tmp2
    pmt[0] = '\0'
    MlirTypeText(pmt, 4096, ptyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, pmt)
    _append_cstr(out, outsz, " -> ")
    _append_cstr(out, outsz, tmt)
    _append_cstr(out, outsz, "\n")


def _MlirStore(p: str, out: str, outsz: sizet):
    """store: llvm.store %v, %p : ty, !llvm.ptr"""
    p = _mlir_ws(p)
    vol: int = 0
    if string.strncmp(p, "volatile ", 9) == 0:
        vol = 1
        p = _mlir_ws(p + 9)
    vtyb: str = _SC_Ty
    p = _mlir_read_type(p, vtyb, 512)
    p = _mlir_ws(p)
    vb: str = _SC_V1
    p = _mlir_read_value(p, vb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    ptyb: str = _SC_Ty2
    p = _mlir_read_type(p, ptyb, 512)
    p = _mlir_ws(p)
    pb: str = _SC_V2
    _mlir_read_value(p, pb, 512)
    _MlirEmitRes(out, outsz, None)
    _append_cstr(out, outsz, "llvm.store ")
    if vol != 0:
        _append_cstr(out, outsz, "volatile ")
    _MlirOperand(vb, vtyb, out, outsz)
    _append_cstr(out, outsz, ", ")
    _MlirOperand(pb, ptyb, out, outsz)
    vmt: str = _SC_Mt
    vmt[0] = '\0'
    MlirTypeText(vmt, 1024, vtyb, _MLIR_Pool)
    pmt: str = _SC_Tmp2
    pmt[0] = '\0'
    MlirTypeText(pmt, 4096, ptyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, vmt)
    _append_cstr(out, outsz, ", ")
    _append_cstr(out, outsz, pmt)
    _append_cstr(out, outsz, "\n")


def _MlirBr(p: str, out: str, outsz: sizet):
    """br: in LLVM syntax the conditional branch opcode is also `br`, dispatched by whether the first token is `label`.

        br label %x                    -> llvm.br ^x
        br i1 %c, label %a, label %b   -> llvm.cond_br %c, ^a, ^b

    Note: build_cond_br emits intermediate text `br i1 ...` (NOT `cond_br`), so the conditional
    form must be forwarded here to _MlirCondBr; otherwise it is taken as unconditional and the
    type `i1` is mistaken for a block name (producing `llvm.br ^i1` and dropping the condition).
    """
    p = _mlir_ws(p)
    if string.strncmp(p, "label ", 6) != 0:
        # Conditional branch: `<type> <cond>, label %a, label %b`
        _MlirCondBr(p, out, outsz)
        return
    p = _mlir_ws(p + 6)
    tb: str = _SC_V1
    _mlir_read_value(p, tb, 512)
    br: str = _MlirBlkRef(tb)
    _MlirEmitRes(out, outsz, None)
    _append_cstr(out, outsz, "llvm.br ^")
    _append_cstr(out, outsz, br)
    _MlirBlockOperands(out, outsz, _MlirFindBlock(_MLIR_Func, br), _MLIR_SrcBlk)
    _append_cstr(out, outsz, "\n")


def _MlirCondBr(p: str, out: str, outsz: sizet):
    """cond_br: llvm.cond_br %c, ^then(%v : ty), ^else(%v : ty)"""
    ctyb: str = _SC_Ty
    p = _mlir_read_type(p, ctyb, 512)
    p = _mlir_ws(p)
    cb: str = _SC_V1
    p = _mlir_read_value(p, cb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    if string.strncmp(p, "label ", 6) == 0:
        p = _mlir_ws(p + 6)
    tb: str = _SC_V2
    p = _mlir_read_value(p, tb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    if string.strncmp(p, "label ", 6) == 0:
        p = _mlir_ws(p + 6)
    eb: str = _SC_V3
    _mlir_read_value(p, eb, 512)
    _MlirEmitRes(out, outsz, None)
    _append_cstr(out, outsz, "llvm.cond_br ")
    _MlirOperand(cb, ctyb, out, outsz)
    tr: str = _MlirBlkRef(tb)
    _append_cstr(out, outsz, ", ^")
    _append_cstr(out, outsz, tr)
    _MlirBlockOperands(out, outsz, _MlirFindBlock(_MLIR_Func, tr), _MLIR_SrcBlk)
    er: str = _MlirBlkRef(eb)
    _append_cstr(out, outsz, ", ^")
    _append_cstr(out, outsz, er)
    _MlirBlockOperands(out, outsz, _MlirFindBlock(_MLIR_Func, er), _MLIR_SrcBlk)
    _append_cstr(out, outsz, "\n")


def _MlirRet(p: str, out: str, outsz: sizet):
    """ret: llvm.return %v : ty (void omits operand and type)"""
    p = _mlir_ws(p)
    if string.strncmp(p, "void", 4) == 0 and (p[4] == '\0' or p[4] == ' '):
        _MlirEmitRes(out, outsz, None)
        _append_cstr(out, outsz, "llvm.return\n")
        return
    tyb: str = _SC_Ty
    p = _mlir_read_type(p, tyb, 512)
    vb: str = _SC_V1
    _mlir_read_value(p, vb, 512)
    _MlirEmitRes(out, outsz, None)
    _append_cstr(out, outsz, "llvm.return ")
    _MlirOperand(vb, tyb, out, outsz)
    mt: str = _SC_Mt
    mt[0] = '\0'
    MlirTypeText(mt, 4096, tyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, mt)
    _append_cstr(out, outsz, "\n")


def _MlirUnreachable(out: str, outsz: sizet):
    """unreachable: llvm.unreachable"""
    _MlirEmitRes(out, outsz, None)
    _append_cstr(out, outsz, "llvm.unreachable\n")


def _MlirCall(p: str, resname: str, out: str, outsz: sizet):
    """call: llvm.call @f(%a) : (i32) -> i32 / indirect llvm.call %fp(%a) : !llvm.ptr, (i32) -> i32

    Variadic calls (intermediate text `call <ret> (...) @f(...)`) append `vararg(!llvm.func<...>)`.
    """
    p = _mlir_ws(p)
    variadic: int = 0
    vftys: str = _SC_Tmp
    vftys[0] = '\0'
    rtyb: str = _SC_Ty
    # The return type must be read with nofunc=1: in variadic-call text, the `(...)` after the
    # return type is the callee's parameter type list (LLVM syntax), not part of the return type.
    # Reading it as a function-type suffix would leave p at `@`/`%`, never entering the variadic
    # branch, turning rtyb into `i32 (!llvm.ptr, i32, ...)`, which MlirTypeText splits on commas
    # and misreports on the trailing `...`.
    p = _mlir_read_type(p, rtyb, 512, 1)
    p = _mlir_ws(p)
    if string.strncmp(p, "asm", 3) == 0:
        # Inline asm: intermediate text is `[%N = ]call <ret_ty> asm [sideeffect]
        # [inteldialect] "<asm>", "<constraints>"(<ty> %v, ...)` (LLVM syntax);
        # the MLIR form is rewritten by _MlirAsm.
        _MlirAsm(p, resname, rtyb, out, outsz)
        return
    if p[0] == '(':
        # Variadic call: intermediate text is `call <ret> (<arg types...>, ...) @f(<args>)`
        # (builder's build_call is_variadic=1 form).
        # MLIR form (mlir-opt 22.1.8 measured exit 0):
        #   %0 = llvm.call @printf(%p, %n) vararg(!llvm.func<i32 (!llvm.ptr, i32, ...)>)
        #          : (!llvm.ptr, i32) -> i32
        # `vararg(...)` holds the callee's full MLIR function type (with `...`);
        # for void returns that return type is `void` (writing `none` errors
        # "expected var_callee_type to return void").
        # Note: buffers like vftys/calb avoid _SC_Ty (rtyb holds it) and _SC_Tmp2 (used by the arg loop).
        variadic = 1
        p = p + 1
        g0: int = 0
        while p[0] != '\0':
            g0 += 1
            if g0 > 256:
                MlirFail("MLIR 渲染: 变参 call 的函数类型参数过多")
            p = _mlir_ws(p)
            if p[0] == ')':
                p = p + 1
                break
            if vftys[0] != '\0':
                _append_cstr(vftys, 8192, ", ")
            if p[0] == '.':
                # `...` variadic token passes through (not a type; MlirTypeText rejects it)
                _append_cstr(vftys, 8192, "...")
                while p[0] == '.':
                    p = p + 1
            else:
                ftb: str = _SC_V2
                ftb[0] = '\0'
                p = _mlir_read_type(p, ftb, 512)
                fmt: str = _SC_IdxT
                fmt[0] = '\0'
                MlirTypeText(fmt, 4096, ftb, _MLIR_Pool)
                _append_cstr(vftys, 8192, fmt)
            p = _mlir_ws(p)
            if p[0] == ',':
                p = p + 1
    calb: str = _SC_Callee
    p = _mlir_read_value(p, calb, 512)
    p = _mlir_ws(p)
    if p[0] != '(':
        MlirFail("MLIR 渲染: call 缺少实参列表")
    p = p + 1
    args: str = _SC_Args
    argts: str = _SC_ArgTs
    args[0] = '\0'
    argts[0] = '\0'
    nargs: int = 0
    g: int = 0
    while p[0] != '\0':
        g += 1
        if g > 256:
            MlirFail("MLIR 渲染: call 实参过多")
        p = _mlir_ws(p)
        if p[0] == ')':
            break
        atyb: str = _SC_Ty2
        p = _mlir_read_type(p, atyb, 512)
        p = _mlir_ws(p)
        avb: str = _SC_V1
        p = _mlir_read_value(p, avb, 512)
        if nargs > 0:
            _append_cstr(args, 2048, ", ")
            _append_cstr(argts, 8192, ", ")
        _MlirOperand(avb, atyb, args, 2048)
        amt: str = _SC_Tmp2
        amt[0] = '\0'
        MlirTypeText(amt, 4096, atyb, _MLIR_Pool)
        _append_cstr(argts, 8192, amt)
        nargs += 1
        p = _mlir_ws(p)
        if p[0] == ',':
            p = p + 1
    rmt: str = _SC_Mt
    rmt[0] = '\0'
    MlirTypeText(rmt, 1024, rtyb, _MLIR_Pool)
    # LLVM intrinsics (bare symbol @llvm.xxx, not llvm.mlir.*) must use the generic op
    # `llvm.call_intrinsic "llvm.xxx"(...) : (...) -> ...`: mlir-translate mangles the intrinsic
    # name by operand types (e.g. llvm.va_start -> llvm.va_start.p0). A plain llvm.call would emit
    # an unmangled @llvm.va_start and LLVM verification errors
    # "Intrinsic name not mangled correctly for type arguments!".
    if calb[0] == '@' and string.strncmp(calb + 1, "llvm.", 5) == 0 \
            and string.strncmp(calb + 1, "llvm.mlir.", 10) != 0:
        if string.strcmp(rmt, "none") == 0:
            _MlirEmitRes(out, outsz, None)
        else:
            _MlirEmitRes(out, outsz, resname)
        _append_cstr(out, outsz, "llvm.call_intrinsic \"")
        _append_cstr(out, outsz, calb + 1)
        _append_cstr(out, outsz, "\"(")
        _append_cstr(out, outsz, args)
        _append_cstr(out, outsz, ") : (")
        _append_cstr(out, outsz, argts)
        _append_cstr(out, outsz, ") -> ")
        if string.strcmp(rmt, "none") == 0:
            _append_cstr(out, outsz, "()")
        else:
            _append_cstr(out, outsz, rmt)
        _append_cstr(out, outsz, "\n")
        return
    if string.strcmp(rmt, "none") == 0:
        _MlirEmitRes(out, outsz, None)
    else:
        _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.call ")
    _append_cstr(out, outsz, calb)
    _append_cstr(out, outsz, "(")
    _append_cstr(out, outsz, args)
    _append_cstr(out, outsz, ")")
    if variadic != 0:
        # Variadic call: vararg(the callee's full MLIR function type) (includes `...`)
        _append_cstr(out, outsz, " vararg(!llvm.func<")
        if string.strcmp(rmt, "none") == 0:
            # Void-returning variadic function type: write `void`, not `none`
            _append_cstr(out, outsz, "void")
        else:
            _append_cstr(out, outsz, rmt)
        _append_cstr(out, outsz, " (")
        _append_cstr(out, outsz, vftys)
        # Close three levels in order: the param list `)`, `!llvm.func<`'s `>`, `vararg(`'s `)`
        # (omitting vararg's `)` made mlir-opt 22.1.8 report "expected ')'" verbatim)
        _append_cstr(out, outsz, ")>)")
    _append_cstr(out, outsz, " : ")
    if calb[0] != '@':
        # Indirect call: the callee is a !llvm.ptr operand
        _append_cstr(out, outsz, "!llvm.ptr, ")
    _append_cstr(out, outsz, "(")
    _append_cstr(out, outsz, argts)
    _append_cstr(out, outsz, ") -> ")
    if string.strcmp(rmt, "none") == 0:
        _append_cstr(out, outsz, "()")
    else:
        _append_cstr(out, outsz, rmt)
    _append_cstr(out, outsz, "\n")


def _MlirAsm(p: str, resname: str, rtyb: str, out: str, outsz: sizet):
    """Inline asm: llvm.inline_asm [has_side_effects] [asm_dialect = intel] "...", "..." %a, %b : (i32, i32) -> i32

    Intermediate text (LLVM syntax) has two sources:
      builders' build_inline_asm -> `call <ret> asm sideeffect "...", "..."()`
      TPV CallGenCBuiltin injected -> `call <ret> asm sideeffect inteldialect "...", "..."(<ty> %v, ...)`
    The MLIR (llvm dialect) form was determined by measuring mlir-opt 22.1.8 (do NOT rely on memory):
      * The dialect keyword is `asm_dialect = intel`: LLVM's bare `inteldialect` errors
        "expected attribute value"; `asm_dialect = inteldialect` errors with the enum list [att, intel].
      * Input operands are comma-separated bare operands without parens: `"...", "=r,r"(%x) : (i32) -> i32`
        errors "expected ':'" (that is LLVM syntax and must be rewritten).
      * No return => `: (<operand types...>) -> ()`; with return => `-> ty`.
      * Multi-result inline asm is a single struct-typed result in MLIR: TPV's
        `%N = call {i32, i32} asm sideeffect "...", "=r,r"()` renders as
        `%N = llvm.inline_asm has_side_effects "...", "=r,r" : () -> !llvm.struct<(i32, i32)>`
        (mlir-opt / mlir-translate 22.1.8 measured exit 0), then fields are extracted with
        `llvm.extractvalue` (see _MlirExtractValue). No special branch is needed here: the return type
        text goes through MlirTypeText (`{...}` -> `!llvm.struct<(...)>`), nested structs included.
      * Measured passing combinations: has_side_effects x asm_dialect = intel x 0/1/2 operands x return/no-return.
    Illegal forms always error loudly: unknown tokens besides `sideeffect`/`inteldialect`, missing
    constraint string, non-string literal, unpaired args, or a return value without a result name.
    """
    p = _mlir_ws(p + 3)
    side: int = 0
    intel: int = 0
    g: int = 0
    while g < 8:
        g += 1
        if string.strncmp(p, "sideeffect", 10) == 0 and (p[10] == ' ' or p[10] == '\t'):
            side = 1
            p = _mlir_ws(p + 10)
            continue
        if string.strncmp(p, "inteldialect", 12) == 0 and (p[12] == ' ' or p[12] == '\t'):
            intel = 1
            p = _mlir_ws(p + 12)
            continue
        break
    # Asm text can reach several KB (TPV's ASM_LINE_BUF_SIZE=4096); use a 4096 buffer, not a
    # small one, to avoid truncation; escape pairs (\22/\5C/\0A) pass through as MLIR accepts them.
    asmb: str = _SC_Tmp
    p = _mlir_read_value(p, asmb, 4096)
    if asmb[0] != '"':
        MlirFail("MLIR 渲染: inline_asm 缺少汇编文本（或含未知关键字）")
    p = _mlir_ws(p)
    if p[0] != ',':
        MlirFail("MLIR 渲染: inline_asm 缺少约束串（形态须为 \"asm\", \"constraints\"）")
    p = _mlir_ws(p + 1)
    conb: str = _SC_Args
    p = _mlir_read_value(p, conb, 2048)
    if conb[0] != '"':
        MlirFail("MLIR 渲染: inline_asm 约束串必须是字符串字面量")
    p = _mlir_ws(p)
    if p[0] != '(':
        MlirFail("MLIR 渲染: inline_asm 缺少实参列表 '(...)'")
    p = p + 1
    ops: str = _SC_ArgTs
    ops[0] = '\0'
    tys: str = _SC_IdxV
    tys[0] = '\0'
    nops: int = 0
    g = 0
    while p[0] != '\0':
        g += 1
        if g > 256:
            MlirFail("MLIR 渲染: inline_asm 实参过多")
        p = _mlir_ws(p)
        if p[0] == ')':
            break
        atyb: str = _SC_Ty2
        p = _mlir_read_type(p, atyb, 512)
        p = _mlir_ws(p)
        avb: str = _SC_V1
        p = _mlir_read_value(p, avb, 512)
        if nops > 0:
            _append_cstr(ops, 8192, ", ")
            _append_cstr(tys, 1024, ", ")
        _MlirOperand(avb, atyb, ops, 2048)
        amt: str = _SC_Mt
        amt[0] = '\0'
        MlirTypeText(amt, 4096, atyb, _MLIR_Pool)
        _append_cstr(tys, 1024, amt)
        nops += 1
        p = _mlir_ws(p)
        if p[0] == ',':
            p = p + 1
    if p[0] == ')':
        p = p + 1
    p = _mlir_ws(p)
    if p[0] != '\0':
        MlirFail("MLIR 渲染: inline_asm 实参列表之后出现多余记号")
    rmt: str = _SC_Buf
    rmt[0] = '\0'
    MlirTypeText(rmt, 1024, rtyb, _MLIR_Pool)
    if string.strcmp(rmt, "none") == 0:
        _MlirEmitRes(out, outsz, None)
    else:
        if resname is None:
            MlirFail("MLIR 渲染: inline_asm 有返回值但缺少结果名（%N = call …）")
        _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.inline_asm")
    if side != 0:
        _append_cstr(out, outsz, " has_side_effects")
    if intel != 0:
        _append_cstr(out, outsz, " asm_dialect = intel")
    _append_cstr(out, outsz, " ")
    _append_cstr(out, outsz, asmb)
    _append_cstr(out, outsz, ", ")
    _append_cstr(out, outsz, conb)
    if nops > 0:
        _append_cstr(out, outsz, " ")
        _append_cstr(out, outsz, ops)
    _append_cstr(out, outsz, " : (")
    _append_cstr(out, outsz, tys)
    _append_cstr(out, outsz, ") -> ")
    if string.strcmp(rmt, "none") == 0:
        _append_cstr(out, outsz, "()")
    else:
        _append_cstr(out, outsz, rmt)
    _append_cstr(out, outsz, "\n")


def _MlirCast(p: str, resname: str, ob: str, out: str, outsz: sizet):
    """Cast instruction: llvm.zext %v : i32 to i64"""
    ftyb: str = _SC_Ty
    p = _mlir_read_type(p, ftyb, 512)
    p = _mlir_ws(p)
    vb: str = _SC_V1
    p = _mlir_read_value(p, vb, 512)
    p = _mlir_ws(p)
    if string.strncmp(p, "to ", 3) == 0:
        p = _mlir_ws(p + 3)
    ttyb: str = _SC_Ty2
    _mlir_read_type(p, ttyb, 512)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.")
    _append_cstr(out, outsz, ob)
    _append_cstr(out, outsz, " ")
    _MlirOperand(vb, ftyb, out, outsz)
    fmt: str = _SC_Mt
    fmt[0] = '\0'
    MlirTypeText(fmt, 4096, ftyb, _MLIR_Pool)
    tmt: str = _SC_Tmp2
    tmt[0] = '\0'
    MlirTypeText(tmt, 4096, ttyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, fmt)
    _append_cstr(out, outsz, " to ")
    _append_cstr(out, outsz, tmt)
    _append_cstr(out, outsz, "\n")


def _MlirBitcast(p: str, resname: str, out: str, outsz: sizet):
    """bitcast: ptr->ptr is treated as a no-op (records an alias); otherwise llvm.bitcast %v : T to U"""
    ftyb: str = _SC_Ty
    p = _mlir_read_type(p, ftyb, 512)
    p = _mlir_ws(p)
    vb: str = _SC_V1
    p = _mlir_read_value(p, vb, 512)
    p = _mlir_ws(p)
    if string.strncmp(p, "to ", 3) == 0:
        p = _mlir_ws(p + 3)
    ttyb: str = _SC_Ty2
    _mlir_read_type(p, ttyb, 512)
    if MlirTypeTextIsPointer(ftyb) != 0 and MlirTypeTextIsPointer(ttyb) != 0:
        # With opaque pointers, a ptr->ptr bitcast is meaningless -> no-op:
        # register the result name as an alias of the source operand (later refs auto-rewritten), emit nothing
        if resname is None:
            MlirFail("MLIR 渲染: no-op bitcast 缺少结果名")
        opnd: str = _MlirOperand(vb, ftyb, None, 0)
        _MlirAliasAdd(resname, opnd)
        return
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.bitcast ")
    _MlirOperand(vb, ftyb, out, outsz)
    fmt: str = _SC_Mt
    fmt[0] = '\0'
    MlirTypeText(fmt, 4096, ftyb, _MLIR_Pool)
    tmt: str = _SC_Tmp2
    tmt[0] = '\0'
    MlirTypeText(tmt, 4096, ttyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, fmt)
    _append_cstr(out, outsz, " to ")
    _append_cstr(out, outsz, tmt)
    _append_cstr(out, outsz, "\n")


def _MlirSelect(p: str, resname: str, out: str, outsz: sizet):
    """select: llvm.select %c, %a, %b : i1, i32"""
    ctyb: str = _SC_Ty
    p = _mlir_read_type(p, ctyb, 512)
    p = _mlir_ws(p)
    cb: str = _SC_V1
    p = _mlir_read_value(p, cb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    ttyb: str = _SC_Ty2
    p = _mlir_read_type(p, ttyb, 512)
    p = _mlir_ws(p)
    tb: str = _SC_V2
    p = _mlir_read_value(p, tb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    ftyb: str = _SC_Buf
    p = _mlir_read_type(p, ftyb, 512)
    p = _mlir_ws(p)
    fb: str = _SC_V3
    _mlir_read_value(p, fb, 512)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.select ")
    _MlirOperand(cb, ctyb, out, outsz)
    _append_cstr(out, outsz, ", ")
    _MlirOperand(tb, ttyb, out, outsz)
    _append_cstr(out, outsz, ", ")
    _MlirOperand(fb, ttyb, out, outsz)
    cmt: str = _SC_Mt
    cmt[0] = '\0'
    MlirTypeText(cmt, 4096, ctyb, _MLIR_Pool)
    vmt: str = _SC_Tmp2
    vmt[0] = '\0'
    MlirTypeText(vmt, 1024, ttyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, cmt)
    _append_cstr(out, outsz, ", ")
    _append_cstr(out, outsz, vmt)
    _append_cstr(out, outsz, "\n")


def _MlirGep(p: str, resname: str, out: str, outsz: sizet):
    """getelementptr: llvm.getelementptr %p[i0, i1] : (!llvm.ptr, t0, t1) -> !llvm.ptr, base"""
    baseb: str = _SC_Ty2
    p = _mlir_read_type(p, baseb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    ptyb: str = _SC_Buf
    p = _mlir_read_type(p, ptyb, 512)
    p = _mlir_ws(p)
    pb: str = _SC_V3
    p = _mlir_read_value(p, pb, 512)
    idxv: str = _SC_IdxV
    idxt: str = _SC_IdxT
    idxv[0] = '\0'
    idxt[0] = '\0'
    p = _mlir_ws(p)
    nidx: int = 0
    tmpv: str = _SC_Tmp
    tmpv[0] = '\0'
    g: int = 0
    while p[0] == ',':
        g += 1
        if g > 32:
            MlirFail("MLIR 渲染: GEP 索引过多")
        p = _mlir_ws(p + 1)
        ityb: str = _SC_Ty
        p = _mlir_read_type(p, ityb, 512)
        p = _mlir_ws(p)
        ivb: str = _SC_V1
        p = _mlir_read_value(p, ivb, 512)
        # Constant indices must be written as integer literals (e.g. `[0, 1]`), not materialized
        # as SSA operands: MLIR's llvm.getelementptr requires struct-field indices to be constant,
        # else the verifier errors 'expected index N indexing a struct to be constant'
        # (operand form only populates rawConstantIndices' dynamic slot). Only non-constant (SSA,
        # starting with '%') indices are materialized and take a type slot; constants do not.
        if nidx > 0:
            _append_cstr(idxv, 1024, ", ")
        if ivb[0] == '%':
            _MlirOperand(ivb, ityb, tmpv, 4096)
            _append_cstr(idxv, 1024, tmpv)
            tmpv[0] = '\0'
            imt: str = _SC_Mt
            imt[0] = '\0'
            MlirTypeText(imt, 4096, ityb, _MLIR_Pool)
            if idxt[0] != '\0':
                _append_cstr(idxt, 4096, ", ")
            _append_cstr(idxt, 4096, imt)
        else:
            _append_cstr(idxv, 1024, ivb)
        nidx += 1
        p = _mlir_ws(p)
    if nidx == 0:
        MlirFail("MLIR 渲染: GEP 缺少索引")
    bmt: str = _SC_Tmp2
    bmt[0] = '\0'
    MlirTypeText(bmt, 4096, baseb, _MLIR_Pool)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.getelementptr ")
    _MlirOperand(pb, ptyb, out, outsz)
    _append_cstr(out, outsz, "[")
    _append_cstr(out, outsz, idxv)
    _append_cstr(out, outsz, "] : (!llvm.ptr")
    # One type per index
    _MlirAppendCommaTypes(out, outsz, idxt)
    _append_cstr(out, outsz, ") -> !llvm.ptr, ")
    _append_cstr(out, outsz, bmt)
    _append_cstr(out, outsz, "\n")


def _MlirAppendCommaTypes(out: str, outsz: sizet, tylist: str):
    """Append ", t0, t1" to out (tylist is a comma-separated type string)"""
    if tylist is None or tylist[0] == '\0':
        return
    _append_cstr(out, outsz, ", ")
    _append_cstr(out, outsz, tylist)


def _MlirSwitch(p: str, out: str, outsz: sizet):
    """switch: llvm.switch %v : i32, ^default [0: ^case0, 1: ^case1]

    The form was determined by measuring mlir-opt 22.1.8 (the design-doc form
    `llvm.switch %v, ^default [0: ^case0] : i32` is REJECTED: "expected ':'"):
      `llvm.switch %v : i32, ^default [0: ^c0, 1: ^c1]` -> exit 0 (adopted)
      `llvm.switch %v : i32, ^default(%r : i32) [ ... ]` -> exit 0 (default block args)
      `llvm.switch %v : i32, ^default [0: ^c0(%r : i32)]` -> exit 0 (case block args)
      `llvm.switch %v : i64, ^default [4294967296: ^c0]` -> exit 0 (case value follows value type)
      `llvm.switch %v : i32, ^default [-1: ^c0]` -> exit 0
      `llvm.switch %v : i32, ^default []` -> exit 0 (empty case table)
    i.e. the value type follows the value (`%v : i32,`); block args are written exactly like
    br/cond_br, so _MlirBlockOperands' phi->block-arg facility is reused as-is => target blocks may
    carry phis (a missing incoming edge errors loudly, never silently dropping args).
    Case values must still be constants: in MLIR they are attributes (case_values), not SSA values.
    """
    tyb: str = _SC_Ty
    p = _mlir_read_type(p, tyb, 512)
    p = _mlir_ws(p)
    vb: str = _SC_V1
    p = _mlir_read_value(p, vb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    if string.strncmp(p, "label ", 6) == 0:
        p = _mlir_ws(p + 6)
    db: str = _SC_V2
    p = _mlir_read_value(p, db, 512)
    p = _mlir_ws(p)
    if p[0] == '[':
        p = p + 1
    # Materialize the value operand and its MLIR type text first: _MlirBlockOperands inside the
    # case loop reuses _SC_Ty/_SC_Buf/_SC_P1/_SC_P2 scratch, so capture into a separate buffer first.
    vop: str = _SC_IdxT
    vop[0] = '\0'
    _MlirOperand(vb, tyb, vop, 1024)
    vmt: str = _SC_IdxV
    vmt[0] = '\0'
    MlirTypeText(vmt, 1024, tyb, _MLIR_Pool)
    cases: str = _SC_Tmp2
    cases[0] = '\0'
    ncase: int = 0
    g: int = 0
    while p[0] != '\0':
        g += 1
        if g > 1024:
            MlirFail("MLIR 渲染: switch 分支过多")
        p = _mlir_ws(p)
        if p[0] == ']' or p[0] == '\0':
            break
        ctb: str = _SC_P1
        p = _mlir_read_type(p, ctb, 512)
        p = _mlir_ws(p)
        cvb: str = _SC_P2
        p = _mlir_read_value(p, cvb, 512)
        p = _mlir_ws(p)
        if p[0] == ',':
            p = _mlir_ws(p + 1)
        if string.strncmp(p, "label ", 6) == 0:
            p = _mlir_ws(p + 6)
        cb: str = _SC_Ty2
        p = _mlir_read_value(p, cb, 512)
        if cvb[0] == '%':
            MlirFail("MLIR 渲染: switch case 值必须是常量（SSA 值不支持）")
        cbr: str = _MlirBlkRef(cb)
        if ncase > 0:
            _append_cstr(cases, 4096, ", ")
        _append_cstr(cases, 4096, cvb)
        _append_cstr(cases, 4096, ": ^")
        _append_cstr(cases, 4096, cbr)
        # Target block has phis => append block arguments (same facility as br/cond_br)
        _MlirBlockOperands(cases, 1024, _MlirFindBlock(_MLIR_Func, cbr), _MLIR_SrcBlk)
        ncase += 1
        p = _mlir_ws(p)
        if p[0] == ',':
            p = p + 1
    dr: str = _MlirBlkRef(db)
    _MlirEmitRes(out, outsz, None)
    _append_cstr(out, outsz, "llvm.switch ")
    _append_cstr(out, outsz, vop)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, vmt)
    _append_cstr(out, outsz, ", ^")
    _append_cstr(out, outsz, dr)
    _MlirBlockOperands(out, outsz, _MlirFindBlock(_MLIR_Func, dr), _MLIR_SrcBlk)
    _append_cstr(out, outsz, " [")
    _append_cstr(out, outsz, cases)
    _append_cstr(out, outsz, "]\n")


def _MlirAtomicRmw(p: str, resname: str, out: str, outsz: sizet):
    """atomicrmw: llvm.atomicrmw add %p, %v monotonic : !llvm.ptr, i32"""
    aop: str = _SC_P1
    p = _mlir_read_value(p, aop, 512)
    p = _mlir_ws(p)
    ptyb: str = _SC_Ty
    p = _mlir_read_type(p, ptyb, 512)
    p = _mlir_ws(p)
    pb: str = _SC_V1
    p = _mlir_read_value(p, pb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    vtyb: str = _SC_Ty2
    p = _mlir_read_type(p, vtyb, 512)
    p = _mlir_ws(p)
    vb: str = _SC_V2
    p = _mlir_read_value(p, vb, 512)
    p = _mlir_ws(p)
    ordb: str = _SC_P2
    _mlir_read_value(p, ordb, 512)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.atomicrmw ")
    _append_cstr(out, outsz, aop)
    _append_cstr(out, outsz, " ")
    _MlirOperand(pb, ptyb, out, outsz)
    _append_cstr(out, outsz, ", ")
    _MlirOperand(vb, vtyb, out, outsz)
    _append_cstr(out, outsz, " ")
    _append_cstr(out, outsz, ordb)
    pmt: str = _SC_Mt
    pmt[0] = '\0'
    MlirTypeText(pmt, 4096, ptyb, _MLIR_Pool)
    vmt: str = _SC_Tmp2
    vmt[0] = '\0'
    MlirTypeText(vmt, 1024, vtyb, _MLIR_Pool)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, pmt)
    _append_cstr(out, outsz, ", ")
    _append_cstr(out, outsz, vmt)
    _append_cstr(out, outsz, "\n")


def _MlirCmpXchg(p: str, resname: str, resname2: str,
                 out: str, outsz: sizet):
    """cmpxchg: MLIR single result struct{T, i1} + two extractvalues for old value / success bit.

    Intermediate text (LLVM syntax, two forms: builder single result / two-result prefix)
      %N = cmpxchg <ptr_ty> <ptr>, <cmp_ty> <cmp>, <new_ty> <new> <succ> <fail>
      %a, %b = cmpxchg ...
    MLIR form (mlir-opt 22.1.8 measured exit 0; cmpxchg is a single result in MLIR, yielding
    struct{T, i1}, not two SSA values):
      %N = llvm.cmpxchg %p, %c, %n acq_rel monotonic : !llvm.ptr, i32
      %_N_v = llvm.extractvalue %N[0] : !llvm.struct<(i32, i1)>   ; old value (compared value)
      %_N_ok = llvm.extractvalue %N[1] : !llvm.struct<(i32, i1)>  ; i1 success bit
    The two extractvalue result names are deterministically derived from the struct result name
    (_MlirCxResName, e.g. `%0` -> `%_0_v`/`%_0_ok`).
    Bindings: `%N` (single result) / `%a` (two-result first name) -> old value `[0]`;
    two-result second name `%b` -> success bit `[1]`. Bindings use the same alias facility as
    bitcast no-op (_MlirAliasAdd); later references are auto-rewritten by _MlirResolve to the
    extractvalue result names. This preserves builder-side build_cmpxchg's "T-typed Value" (old value) semantics.
    """
    ptyb: str = _SC_Ty
    p = _mlir_read_type(p, ptyb, 512)
    p = _mlir_ws(p)
    pb: str = _SC_V1
    p = _mlir_read_value(p, pb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    ctyb: str = _SC_Ty2
    p = _mlir_read_type(p, ctyb, 512)
    p = _mlir_ws(p)
    cb: str = _SC_V2
    p = _mlir_read_value(p, cb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    ntyb: str = _SC_B1
    p = _mlir_read_type(p, ntyb, 512)
    p = _mlir_ws(p)
    nb: str = _SC_B2
    p = _mlir_read_value(p, nb, 512)
    p = _mlir_ws(p)
    succ: str = _SC_P1
    p = _mlir_read_value(p, succ, 512)
    p = _mlir_ws(p)
    fail: str = _SC_P2
    _mlir_read_value(p, fail, 512)
    if resname is None:
        MlirFail("MLIR 渲染: cmpxchg 缺少结果名")
    # The two extractvalue result names derive deterministically from the struct name (identical across both passes)
    onm: str = _MlirCxResName(resname, "_v")
    snm: str = _MlirCxResName(resname, "_ok")
    _MlirAliasAdd(resname, onm)
    if resname2 is not None:
        _MlirAliasAdd(resname2, snm)
    # MLIR type text (use a non-_SC_Mt buffer: _MlirOperand occupies _SC_Mt for literals)
    pmt: str = _SC_Tmp2
    pmt[0] = '\0'
    MlirTypeText(pmt, 4096, ptyb, _MLIR_Pool)
    vmt: str = _SC_IdxV
    vmt[0] = '\0'
    MlirTypeText(vmt, 1024, ctyb, _MLIR_Pool)
    stmt: str = _SC_IdxT
    stmt[0] = '\0'
    _append_cstr(stmt, 4096, "!llvm.struct<(")
    _append_cstr(stmt, 4096, vmt)
    _append_cstr(stmt, 4096, ", i1)>")
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.cmpxchg ")
    _MlirOperand(pb, ptyb, out, outsz)
    _append_cstr(out, outsz, ", ")
    _MlirOperand(cb, ctyb, out, outsz)
    _append_cstr(out, outsz, ", ")
    _MlirOperand(nb, ntyb, out, outsz)
    _append_cstr(out, outsz, " ")
    _append_cstr(out, outsz, succ)
    _append_cstr(out, outsz, " ")
    _append_cstr(out, outsz, fail)
    _append_cstr(out, outsz, " : ")
    _append_cstr(out, outsz, pmt)
    _append_cstr(out, outsz, ", ")
    _append_cstr(out, outsz, vmt)
    _append_cstr(out, outsz, "\n")
    _MlirEmitRes(out, outsz, onm)
    _append_cstr(out, outsz, "llvm.extractvalue ")
    _append_cstr(out, outsz, resname)
    _append_cstr(out, outsz, "[0] : ")
    _append_cstr(out, outsz, stmt)
    _append_cstr(out, outsz, "\n")
    _MlirEmitRes(out, outsz, snm)
    _append_cstr(out, outsz, "llvm.extractvalue ")
    _append_cstr(out, outsz, resname)
    _append_cstr(out, outsz, "[1] : ")
    _append_cstr(out, outsz, stmt)
    _append_cstr(out, outsz, "\n")


def _MlirVaArg(p: str, resname: str, out: str, outsz: sizet):
    """va_arg: use the generic op syntax (deliberate choice, see below).

    Intermediate text (LLVM syntax): %N = va_arg <va_list_ty> <va_list_ptr>, <arg_ty>
    MLIR text:                       %N = "llvm.va_arg"(%ap) : (!llvm.ptr) -> i32

    Why generic form is required (mlir-opt 22.1.8 measured): the custom assembly syntax
    `llvm.va_arg %ap, i32` requires an explicit va_list type (in LLVM it is `ptr`), which cannot
    be expressed with opaque pointers -- mlir-opt errors "expected builtin.function". So the
    generic op form is used, described only by operand/result types. Down-level verification
    (mlir-translate --mlir-to-llvmir) yields `%2 = va_arg ptr %0, i32`, isomorphic to LLVM's va_arg.
    """
    vtyb: str = _SC_Ty
    p = _mlir_read_type(p, vtyb, 512)
    p = _mlir_ws(p)
    vb: str = _SC_V1
    p = _mlir_read_value(p, vb, 512)
    p = _mlir_ws(p)
    if p[0] == ',':
        p = _mlir_ws(p + 1)
    atyb: str = _SC_Ty2
    _mlir_read_type(p, atyb, 512)
    if resname is None:
        MlirFail("MLIR 渲染: va_arg 缺少结果名")
    # Operand/result types (_MlirOperand only uses _SC_Mt, so take separate buffers here)
    omt: str = _SC_Tmp2
    omt[0] = '\0'
    MlirTypeText(omt, 4096, vtyb, _MLIR_Pool)
    rmt: str = _SC_IdxV
    rmt[0] = '\0'
    MlirTypeText(rmt, 1024, atyb, _MLIR_Pool)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "\"llvm.va_arg\"(")
    _MlirOperand(vb, vtyb, out, outsz)
    _append_cstr(out, outsz, ") : (")
    _append_cstr(out, outsz, omt)
    _append_cstr(out, outsz, ") -> ")
    _append_cstr(out, outsz, rmt)
    _append_cstr(out, outsz, "\n")


def _MlirExtractValue(p: str, resname: str, out: str, outsz: sizet):
    """extractvalue: llvm.extractvalue %v[i0, i1] : <aggregate type>

    Intermediate text is LLVM syntax (aggregate type BEFORE the value, indices after commas):
      %r = extractvalue {i32, i32} %v, 0
      %r = extractvalue {i32, {i32, i32}} %v, 1, 0        ; multi-level index
    MLIR form (mlir-opt 22.1.8 measured exit 0):
      %r = llvm.extractvalue %v[0] : !llvm.struct<(i32, i32)>
      %r = llvm.extractvalue %v[1, 0] : !llvm.struct<(i32, !llvm.struct<(i32, i32)>)>
    The index list is written inside `[]` separated by ", " in MLIR (different from LLVM's
    `<aggregate>, i0, i1` syntax, so it must be rewritten); the aggregate type text goes through
    MlirTypeText (anonymous `{...}` -> !llvm.struct<(...)>, named `%"name"` -> inlined named body,
    array `[N x T]` -> !llvm.array<N x T>, all measured exit 0).
    Real source: TPV's multi-result inline asm (`%asm_result = call {i32, i32} asm ...`) followed by
    per-field extraction (`%asm_out_0 = extractvalue {i32, i32} %asm_result, 0`); multi-level indices
    are also supported (measured legal). Indices must be constant literals (LLVM syntax) -- SSA index => loud error.
    """
    atyb: str = _SC_Ty
    p = _mlir_read_type(p, atyb, 512)
    p = _mlir_ws(p)
    vb: str = _SC_V1
    p = _mlir_read_value(p, vb, 512)
    idxs: str = _SC_IdxV
    idxs[0] = '\0'
    nidx: int = 0
    g: int = 0
    # Index list: comma-separated in LLVM syntax (`<aggregate>, i0, i1`); read each and rewrite.
    # Loop bound matches other parse loops in this file (bounded; exceeding errors loudly).
    while g < 64:
        p = _mlir_ws(p)
        if p[0] != ',':
            break
        g += 1
        if g > 32:
            MlirFail("MLIR 渲染: extractvalue 索引过多")
        p = _mlir_ws(p + 1)
        ib: str = _SC_IdxT
        p = _mlir_read_value(p, ib, 512)
        if ib[0] == '\0':
            MlirFail("MLIR 渲染: extractvalue 缺少索引")
        if ib[0] == '%':
            MlirFail("MLIR 渲染: extractvalue 索引必须是常量（SSA 索引不支持）")
        if nidx > 0:
            _append_cstr(idxs, 1024, ", ")
        _append_cstr(idxs, 1024, ib)
        nidx += 1
    if nidx == 0:
        MlirFail("MLIR 渲染: extractvalue 缺少索引")
    if p[0] != '\0':
        MlirFail("MLIR 渲染: extractvalue 索引之后出现多余记号")
    # Aggregate type text goes into a separate buffer first (_MlirOperand only uses _SC_Mt)
    amt: str = _SC_Tmp2
    amt[0] = '\0'
    MlirTypeText(amt, 4096, atyb, _MLIR_Pool)
    _MlirEmitRes(out, outsz, resname)
    _append_cstr(out, outsz, "llvm.extractvalue ")
    _MlirOperand(vb, atyb, out, outsz)
    _append_cstr(out, outsz, "[")
    _append_cstr(out, outsz, idxs)
    _append_cstr(out, outsz, "] : ")
    _append_cstr(out, outsz, amt)
    _append_cstr(out, outsz, "\n")


# ============================================================
# Instruction line -> MLIR dispatch
# ============================================================
def _MlirRenderLine(text: str, out: str, outsz: sizet):
    """Parse one intermediate instruction text and render it to MLIR (out=None: register/validate only)"""
    global _MLIR_Pass, _MLIR_CurLine
    if text is None:
        return
    # Stash the "current line" at entry: printed on type-text conversion failure (both passes
    # use the same line, so pass1 failures are also locatable). text points to Line.Buf, live for the whole render.
    _MLIR_CurLine = text
    MlirSetContextLine(_MLIR_CurLine)
    p: str = _mlir_ws(text)
    if p[0] == '\0':
        return
    resname: str = None
    resname2: str = None
    if p[0] == '%':
        tb: str = _SC_Tb
        q: str = _mlir_read_value(p, tb, 512)
        q2: str = _mlir_ws(q)
        if q2[0] == '=':
            resname = _mlir_dup(tb)
            p = _mlir_ws(q2 + 1)
        elif q2[0] == ',':
            # Two-result prefix `%a, %b = <op>` (cmpxchg only): MLIR's cmpxchg is a single
            # struct{T, i1} result; the two names each bind to one extractvalue (see _MlirCmpXchg).
            resname = _mlir_dup(tb)
            q3: str = _mlir_ws(q2 + 1)
            tb2: str = _SC_V3
            q3 = _mlir_read_value(q3, tb2, 512)
            q4: str = _mlir_ws(q3)
            if q4[0] != '=':
                MlirFail("MLIR 渲染: 双结果前缀（%a, %b =）缺少 '='")
            resname2 = _mlir_dup(tb2)
            p = _mlir_ws(q4 + 1)
    ob: str = _SC_Op
    p = _mlir_read_value(p, ob, 64)
    if ob[0] == '\0':
        MlirFail("MLIR 渲染: 空指令")
    if resname2 is not None and string.strcmp(ob, "cmpxchg") != 0:
        MlirFail("MLIR 渲染: 双结果前缀（%a, %b =）只支持 cmpxchg")
    if _mlir_in_set(ob, "|add|sub|mul|sdiv|udiv|srem|urem|and|or|xor|shl|lshr|ashr|") != 0:
        _MlirBinop(p, resname, ob, out, outsz)
        return
    if _mlir_in_set(ob, "|fadd|fsub|fmul|fdiv|frem|") != 0:
        _MlirBinop(p, resname, ob, out, outsz)
        return
    if string.strcmp(ob, "fneg") == 0:
        _MlirFneg(p, resname, out, outsz)
        return
    if string.strcmp(ob, "icmp") == 0:
        _MlirCmp(p, resname, ob, out, outsz)
        return
    if string.strcmp(ob, "fcmp") == 0:
        _MlirCmp(p, resname, ob, out, outsz)
        return
    if string.strcmp(ob, "alloca") == 0:
        _MlirAlloca(p, resname, out, outsz)
        return
    if string.strcmp(ob, "load") == 0:
        _MlirLoad(p, resname, out, outsz)
        return
    if string.strcmp(ob, "store") == 0:
        _MlirStore(p, out, outsz)
        return
    if string.strcmp(ob, "br") == 0:
        _MlirBr(p, out, outsz)
        return
    if string.strcmp(ob, "ret") == 0:
        _MlirRet(p, out, outsz)
        return
    if string.strcmp(ob, "unreachable") == 0:
        _MlirUnreachable(out, outsz)
        return
    if string.strcmp(ob, "call") == 0:
        _MlirCall(p, resname, out, outsz)
        return
    if string.strcmp(ob, "select") == 0:
        _MlirSelect(p, resname, out, outsz)
        return
    if string.strcmp(ob, "phi") == 0:
        if out is not None:
            # Leading phis are skipped in _MlirEmitBlockLines; reaching here means phi is not at block head
            MlirFail("MLIR 渲染: phi 不在块首（v1 只支持块首连续 phi → 块参数）")
        _MlirPhiScan(text)
        return
    if string.strcmp(ob, "getelementptr") == 0:
        _MlirGep(p, resname, out, outsz)
        return
    if string.strcmp(ob, "bitcast") == 0:
        _MlirBitcast(p, resname, out, outsz)
        return
    if _mlir_in_set(ob, "|sext|trunc|zext|ptrtoint|inttoptr|fpext|fptrunc|fptosi|sitofp|fptoui|uitofp|") != 0:
        _MlirCast(p, resname, ob, out, outsz)
        return
    if string.strcmp(ob, "switch") == 0:
        _MlirSwitch(p, out, outsz)
        return
    if string.strcmp(ob, "atomicrmw") == 0:
        _MlirAtomicRmw(p, resname, out, outsz)
        return
    if string.strcmp(ob, "cmpxchg") == 0:
        _MlirCmpXchg(p, resname, resname2, out, outsz)
        return
    if string.strcmp(ob, "va_arg") == 0:
        _MlirVaArg(p, resname, out, outsz)
        return
    if string.strcmp(ob, "extractvalue") == 0:
        # Real source: TPV's multi-result inline-asm result extraction (see _MlirExtractValue).
        # Note: other unsupported opcodes like `insertvalue` never occur in the corpus and fall
        # through to the loud fallback error ("unknown instruction opcode").
        _MlirExtractValue(p, resname, out, outsz)
        return
    # Note: inline asm's opcode in intermediate text is "call" (call <ret> asm ...); _MlirCall
    # recognizes it and forwards to _MlirAsm (llvm.inline_asm).
    MlirFail("MLIR 渲染: 未知指令 opcode")


# ============================================================
# Function signature and body
# ============================================================
def _MlirFuncSig(buf: str, size: sizet,
                 func: rawptr[Function], pool: memhub.MemBuddy*):
    """llvm.func @name(%a: i32, %b: i32) -> i32 (void return omits -> ty)"""
    _append_cstr(buf, size, "llvm.func @")
    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, "(")
    params: rawptr[GSList[Param]] = func.Params
    cur: rawptr[Param] = None
    if params is not None:
        cur = params.Head
    first: int = 1
    cur_steps: int = 0
    while cur is not None:
        cur_steps += 1
        if cur_steps > IR_DUMP_CHAIN_LIMIT:
            MlirFail("MLIR 渲染: 参数链疑似成环")
        if first == 0:
            _append_cstr(buf, size, ", ")
        if cur.Name is not None:
            if cur.Name[0] != '%':
                _append_cstr(buf, size, "%")
            _append_cstr(buf, size, cur.Name)
            _append_cstr(buf, size, ": ")
        ty_buf: str = _SC_Ty
        ty_buf[0] = '\0'
        TypePrint(ty_buf, 4096, cur.Ty, pool)
        _append_cstr(buf, size, ty_buf)
        cur = cur.Next
        first = 0
    # Variadic function: append `...` at the end of the parameter list (mlir-opt 22.1.8 measured exit 0:
    #   `llvm.func @f(%p: !llvm.ptr, ...) -> i32` / with no fixed params `llvm.func @f(...) -> i32`)
    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, ")")
    # Return type (Void omitted)
    is_void: int = 0
    match func.RetTy:
        case LLVMType.Void():
            is_void = 1
        case _:
            is_void = 0
    if is_void == 0:
        ret_buf: str = _SC_Ty2
        ret_buf[0] = '\0'
        TypePrint(ret_buf, 4096, func.RetTy, pool)
        _append_cstr(buf, size, " -> ")
        _append_cstr(buf, size, ret_buf)


def FunctionPrint(buf: str, size: sizet,
                  func: rawptr[Function], pool: memhub.MemBuddy*):
    """Serialize function func as MLIR (llvm dialect) text APPENDED to buf.

    NOTE: like the llvmlite version, this appends (so the module printer can concatenate
    functions); callers wanting single-function text must set buf[0]='\\0' themselves.
    """
    global _MLIR_Pass, _MLIR_Func, _MLIR_SrcBlk, _MLIR_CurLine, _MlirOutTrunc
    if func is None or buf is None or size == 0:
        return
    _MlirOutTrunc = 0

    # Local peephole first (ret zeroinitializer->0/null, entry-slot store->load forwarding,
    # no-op bitcast removal): peepholes match by name and delete lines by setting Buf=None.
    # They act on the builders' intermediate instruction text and are orthogonal to MLIR rendering;
    # RenumberSSA internally skips declare functions (IsDeclared==1).
    if func.IsDeclared == 0 and G_PeepEnable != 0:
        _peephole_opt(func, pool)
    RenumberSSA(func, pool)

    _MlirTablesInit(pool)
    _MLIR_Func = func
    # Clear the previous function's line context: signature/declare and other non-line-driven
    # positions must not carry an unrelated stale line, avoiding misleading diagnostics.
    _MLIR_CurLine = None
    MlirSetContextLine(None)

    if func.IsDeclared == 1:
        # LLVM intrinsic declarations are NOT emitted: on the MLIR side intrinsics are expressed
        # via llvm.call_intrinsic (see _MlirCall). Emitting `llvm.func @llvm.va_start(!llvm.ptr)`
        # would be down-leveled to an unmangled `declare void @llvm.va_start(ptr)`, and LLVM
        # verification errors "Intrinsic name not mangled correctly for type arguments!".
        # Rule: name starts with "llvm." and not "llvm.mlir.*" (the latter are MLIR-internal op names).
        if func.Name is not None and string.strncmp(func.Name, "llvm.", 5) == 0 \
                and string.strncmp(func.Name, "llvm.mlir.", 10) != 0:
            return
        # declare (no body)
        _MlirFuncSig(buf, size, func, pool)
        _append_cstr(buf, size, "\n")
        _MlirOutTruncFail()
        return

    blocks: rawptr[GSList[BasicBlock]] = func.Blocks
    if blocks is None:
        _MlirFuncSig(buf, size, func, pool)
        _append_cstr(buf, size, "\n")
        _MlirOutTruncFail()
        return

    # ---- pass1: scan (register constants/aliases + shape checks) ----
    _MLIR_Pass = 1
    blk: rawptr[BasicBlock] = blocks.Head
    blk_steps: int = 0
    while blk is not None:
        blk_steps += 1
        if blk_steps > IR_DUMP_CHAIN_LIMIT:
            MlirFail("MLIR 渲染: 函数块链疑似成环")
        _MLIR_SrcBlk = blk
        if blk.Lines is not None:
            lines: rawptr[GSList[Line]] = blk.Lines
            ln: rawptr[Line] = lines.Head
            ln_steps: int = 0
            while ln is not None:
                ln_steps += 1
                if ln_steps > IR_DUMP_CHAIN_LIMIT:
                    MlirFail("MLIR 渲染: 块指令链疑似成环")
                if ln.Buf is not None:
                    _MlirRenderLine(ln.Buf, None, 0)
                ln = ln.Next
        blk = blk.Next

    # ---- pass2: emit ----
    _MLIR_Pass = 2
    _MlirFuncSig(buf, size, func, pool)
    _append_cstr(buf, size, " {\n")
    blk = blocks.Head
    first: int = 1
    blk_steps = 0
    while blk is not None:
        blk_steps += 1
        if blk_steps > IR_DUMP_CHAIN_LIMIT:
            MlirFail("MLIR 渲染: 函数块链疑似成环")
        _MLIR_SrcBlk = blk
        if first != 0:
            # entry block has no label; constant-materialization ops go at the entry block head
            # (entry dominates the function, so constants are visible everywhere without breaking the phi-at-head rule)
            _MlirCheckEntryNoPhi(blk)
            _MlirEmitConsts(buf, size)
        else:
            _MlirEmitBlockHeader(buf, size, blk)
        _MlirEmitBlockLines(buf, size, blk)
        blk = blk.Next
        first = 0
    _append_cstr(buf, size, "}\n")
    _MlirOutTruncFail()


# ============================================================
# Internal helpers
# ============================================================
# _append_cstr write-position cache (perf: O(1) when appending into the same buffer,
# avoiding O(n^2) strlen(dst) rescans -- an IR-serialization hot spot).
# Falls back to strlen on buffer switch (different dst) or buffer reset (dst[0]=0 and pos>0).
_g_last_dst: uint64 = 0
_g_last_pos: sizet = 0

def _append_cstr(dst: str, dst_size: sizet, src: str):
    """Append C string src to the end of dst"""
    global _g_last_dst, _g_last_pos, _MlirOutTrunc
    if dst is None or src is None: return
    dlen: sizet = 0
    if _g_last_dst == uint64(dst):
        dlen = _g_last_pos
        # Buffer reset (caller reuses the same address after buf[0]=0) -> restart from head
        if dlen > 0 and dst[0] == '\0':
            dlen = 0
        # Cache-consistency check: the cached end must still be the current text end (NUL at the
        # end, non-NUL before it). If several modules write the same buf or pool reuse leaves old
        # content, the cached pos goes stale and appends past the first NUL become invisible
        # (LLvmLiteTest Test5/9 root cause); on failure fall back to strlen.
        if dlen == 0:
            if dst[0] != '\0':
                dlen = string.strlen(dst)
        elif dst[dlen] != '\0' or dst[dlen - 1] == '\0':
            dlen = string.strlen(dst)
    else:
        dlen = string.strlen(dst)
        _g_last_dst = uint64(dst)
    # Unsigned-underflow guard: if dlen exceeds the buffer size (corrupted/dangling buffer),
    # return; otherwise remain = dst_size - dlen underflows to a huge value and writes out of bounds.
    if dlen >= dst_size:
        _MlirOutTrunc = 1
        return
    slen: sizet = string.strlen(src)
    remain: sizet = dst_size - dlen
    if remain <= 0:
        _MlirOutTrunc = 1
        return
    i: sizet = 0
    while i < slen and i + 1 < remain:
        dst[dlen + i] = src[i]
        i += 1
    if i < slen:
        _MlirOutTrunc = 1
    dst[dlen + i] = '\0'
    _g_last_pos = dlen + i


def _MlirOutTruncFail():
    """Output-truncation check before printing finishes (truncation errors loudly, never emits partial MLIR silently)"""
    global _MlirOutTrunc
    if _MlirOutTrunc != 0:
        _MlirOutTrunc = 0
        MlirFail("MLIR 渲染: 输出文本超出缓冲被截断（需增大 dump_ir/FunctionPrint 的缓冲）")


# ============================================================
# Debug call-site recording (--traceback only; zero overhead when off)
#
# The MLIR backend emits no DWARF (see README 3.3): the final exe has no .debug_line records
# for our .vp functions, so runtime stackmaps cannot map a caller frame's PC (= return address)
# back to a call-site line. Here, at BUILD time, each define (F record) and each DIRECT call
# (C record) is recorded into a text buffer; after each module is translated TransPyV writes it
# to a <sha1>.vpcall side file, and BuildPipeline.WriteStackMap combines it with
# `llvm-objdump -d` for "return address -> (file, line, function)" address normalization.
#
# Record format (fields separated by \t to avoid ambiguity from spaces/backslashes in Windows paths):
#   F\t<function symbol>\t<definition line>\t<absolute file path>\n
#   C\t<caller symbol>\t<callee symbol>\t<call-site line>\t<absolute file path>\n
#
# Function symbol = Function.Name (no @ prefix), matching objdump labels; the callee symbol comes
# from build_call's callee (also no @); LLVM intrinsics (llvm.*) never become named callq targets and are skipped.
# ============================================================

# Master switch: mirrors TransPyV's --traceback (relayed via DebugInfoGen.SetEnabled)
_g_dbg_on: int = 0

# Text buffer (stdlib-allocated, survives pool reset; cleared after DebugTakeText hands off ownership)
_g_dbg_buf: bytes = None
_g_dbg_len: sizet = 0
_g_dbg_cap: sizet = 0

# Current function context (C records pull caller symbol/file/line from here); nested defines
# (lambdas etc.) use a fixed-size Push/Pop stack.
define DBGC_STACK_MAX = 256
_g_dbg_func: str = None
_g_dbg_file: str = None
_g_dbg_defline: int = 0
_g_dbg_line: int = 0
_g_dbg_depth: int = 0
_g_dbg_stk_func: bytes = None   # 8 bytes per slot (char*)
_g_dbg_stk_file: bytes = None   # 8 bytes per slot (char*)
_g_dbg_stk_def: bytes = None    # 8 bytes per slot (sizet)
_g_dbg_stk_ln: bytes = None     # 8 bytes per slot (sizet)


def DebugSetEnabled(flag: bool) -> int:
    """Enable/disable debug call-site recording (relayed from DebugInfoGen.SetEnabled --traceback)."""
    global _g_dbg_on
    if flag:
        _g_dbg_on = 1
    else:
        _g_dbg_on = 0
    return 0


def DebugIsEnabled() -> int:
    return _g_dbg_on


def _DbgReserve(extra: sizet) -> int:
    """Ensure the text buffer can hold at least extra more bytes (excluding the trailing NUL)."""
    global _g_dbg_buf, _g_dbg_cap
    if _g_dbg_buf is None:
        cap0: sizet = 262144
        b: bytes = stdlib.malloc(cap0)
        if b is None:
            return 1
        b[0] = '\0'
        _g_dbg_buf = b
        _g_dbg_cap = cap0
    need: sizet = _g_dbg_len + extra + 1
    if need <= _g_dbg_cap:
        return 0
    ncap: sizet = _g_dbg_cap
    while ncap < need:
        ncap = ncap * 2
    nb: bytes = stdlib.realloc(_g_dbg_buf, ncap)
    if nb is None:
        return 1
    _g_dbg_buf = nb
    _g_dbg_cap = ncap
    return 0


def _DbgEnsureStack() -> int:
    """Lazily allocate the Push/Pop context stack (only when enabled and actual nesting occurs)."""
    global _g_dbg_stk_func, _g_dbg_stk_file, _g_dbg_stk_def, _g_dbg_stk_ln
    if _g_dbg_stk_func is not None:
        return 0
    b1: bytes = stdlib.malloc(sizet(DBGC_STACK_MAX) * 8)
    b2: bytes = stdlib.malloc(sizet(DBGC_STACK_MAX) * 8)
    b3: bytes = stdlib.malloc(sizet(DBGC_STACK_MAX) * 8)
    b4: bytes = stdlib.malloc(sizet(DBGC_STACK_MAX) * 8)
    if b1 is None or b2 is None or b3 is None or b4 is None:
        return 1
    _g_dbg_stk_func = b1
    _g_dbg_stk_file = b2
    _g_dbg_stk_def = b3
    _g_dbg_stk_ln = b4
    return 0


def DebugNoteFunc(sym: str, file: str, line: int) -> int:
    """Enter a define: push the outer context, set as current function, and write an F record."""
    global _g_dbg_func, _g_dbg_file, _g_dbg_defline, _g_dbg_line, _g_dbg_depth, _g_dbg_len
    if _g_dbg_on == 0:
        return 0
    if _g_dbg_depth >= DBGC_STACK_MAX:
        MlirFail("MLIR 调试记录: 函数嵌套深度超上限（DBGC_STACK_MAX）")
    if _DbgEnsureStack() != 0:
        MlirFail("MLIR 调试记录: 上下文栈分配失败")
    fs: str* = (str*)(uint64(_g_dbg_stk_func) + sizet(_g_dbg_depth) * 8)
    fs[0] = _g_dbg_func
    ff: str* = (str*)(uint64(_g_dbg_stk_file) + sizet(_g_dbg_depth) * 8)
    ff[0] = _g_dbg_file
    ds: sizet* = (sizet*)(uint64(_g_dbg_stk_def) + sizet(_g_dbg_depth) * 8)
    ds[0] = sizet(_g_dbg_defline)
    ls: sizet* = (sizet*)(uint64(_g_dbg_stk_ln) + sizet(_g_dbg_depth) * 8)
    ls[0] = sizet(_g_dbg_line)
    _g_dbg_depth += 1
    _g_dbg_func = sym
    _g_dbg_file = file
    _g_dbg_defline = line
    _g_dbg_line = line
    if sym is None or file is None:
        return 0
    need: sizet = string.strlen(sym) + string.strlen(file) + 64
    if _DbgReserve(need) != 0:
        MlirFail("MLIR 调试记录: F 记录缓冲分配失败")
    viperlib.snprintf(_g_dbg_buf + _g_dbg_len, _g_dbg_cap - _g_dbg_len,
                      "F\t%s\t%d\t%s\n", sym, line, file)
    _g_dbg_len = _g_dbg_len + string.strlen(_g_dbg_buf + _g_dbg_len)
    return 0


def DebugClearFunc() -> int:
    """Leave a define: restore the outer context (clear if none)."""
    global _g_dbg_func, _g_dbg_file, _g_dbg_defline, _g_dbg_line, _g_dbg_depth
    if _g_dbg_on == 0:
        return 0
    if _g_dbg_depth <= 0:
        _g_dbg_func = None
        _g_dbg_file = None
        _g_dbg_defline = 0
        _g_dbg_line = 0
        return 0
    _g_dbg_depth -= 1
    fs: str* = (str*)(uint64(_g_dbg_stk_func) + sizet(_g_dbg_depth) * 8)
    _g_dbg_func = fs[0]
    ff: str* = (str*)(uint64(_g_dbg_stk_file) + sizet(_g_dbg_depth) * 8)
    _g_dbg_file = ff[0]
    ds: sizet* = (sizet*)(uint64(_g_dbg_stk_def) + sizet(_g_dbg_depth) * 8)
    _g_dbg_defline = int(ds[0])
    ls: sizet* = (sizet*)(uint64(_g_dbg_stk_ln) + sizet(_g_dbg_depth) * 8)
    _g_dbg_line = int(ls[0])
    return 0


def DebugNoteLine(line: int) -> int:
    """Record the current statement line (used by the following build_call to write a C record)."""
    global _g_dbg_line
    if _g_dbg_on == 0:
        return 0
    if line > 0:
        _g_dbg_line = line
    return 0


def DebugResetFunc() -> int:
    """Clear the current function context (module-start fallback: a mid-translation failure may leave an un-Popped context)."""
    global _g_dbg_func, _g_dbg_file, _g_dbg_defline, _g_dbg_line, _g_dbg_depth
    if _g_dbg_on == 0:
        return 0
    _g_dbg_func = None
    _g_dbg_file = None
    _g_dbg_defline = 0
    _g_dbg_line = 0
    _g_dbg_depth = 0
    return 0


def _DebugNoteCall(callee: str) -> int:
    """Write a C record (caller = current function, line = current statement line)."""
    global _g_dbg_len
    if _g_dbg_on == 0:
        return 0
    if _g_dbg_func is None or _g_dbg_file is None or _g_dbg_line <= 0:
        return 0
    if callee is None or callee[0] == '\0':
        return 0
    if string.strncmp(callee, "llvm.", 5) == 0:
        return 0
    need: sizet = string.strlen(_g_dbg_func) + string.strlen(callee) \
        + string.strlen(_g_dbg_file) + 64
    if _DbgReserve(need) != 0:
        MlirFail("MLIR 调试记录: C 记录缓冲分配失败")
    viperlib.snprintf(_g_dbg_buf + _g_dbg_len, _g_dbg_cap - _g_dbg_len,
                      "C\t%s\t%s\t%d\t%s\n",
                      _g_dbg_func, callee, _g_dbg_line, _g_dbg_file)
    _g_dbg_len = _g_dbg_len + string.strlen(_g_dbg_buf + _g_dbg_len)
    return 0


def DebugTakeText() -> str:
    """Take and clear the accumulated text (once per module); buffer ownership passes to the caller."""
    global _g_dbg_buf, _g_dbg_len, _g_dbg_cap
    if _g_dbg_buf is None or _g_dbg_len == 0:
        if _g_dbg_buf is not None:
            _g_dbg_len = 0
            _g_dbg_buf[0] = '\0'
        return None
    out: str = _g_dbg_buf
    _g_dbg_buf = None
    _g_dbg_len = 0
    _g_dbg_cap = 0
    return out


# ============================================================
# Field accessors (let other modules bypass stub type limits)
#
# Cross-module, the user module only has stub types for Function/Param (missing fields, possibly
# wrong types), so TransPyC silently skips field access (no GEP, no error) and fields stay null.
# These accessors run inside this module where the full type definitions exist, so they can access
# all fields correctly; other modules get field values via function calls, bypassing the stub limits.
# ============================================================
def function_get_name(func: rawptr[Function]) -> str:
    """Get the function name"""
    if func is None:
        return None
    return func.Name


def function_get_name_len(func: rawptr[Function]) -> sizet:
    """Get the function name length (computed once at creation, always == strlen(Name)).

    Mirrors function_get_name's cross-module stub-bypass pattern: the user module only has a stub
    Function type (missing fields, possibly wrong types), so reading Name/NameLen directly is
    silently skipped. This accessor reads fields inside the full-type module and returns the length
    via a call, avoiding repeated strlen on the same name (a program-wide strlen hot spot).
    """
    if func is None:
        return 0
    return func.NameLen


def function_get_ret_ty(func: rawptr[Function]) -> rawptr[LLVMType]:
    """Get the function return type"""
    if func is None:
        return None
    return func.RetTy


def function_get_params(func: rawptr[Function]) -> rawptr[GSList[Param]]:
    """Get the function parameter list container"""
    if func is None:
        return None
    return func.Params


def function_get_param_head(func: rawptr[Function]) -> rawptr[Param]:
    """Get the function parameter list head (one step, avoiding GSList.Head access by callers)"""
    if func is None:
        return None
    params: rawptr[GSList[Param]] = func.Params
    if params is None:
        return None
    return params.Head


def function_get_next(func: rawptr[Function]) -> rawptr[Function]:
    """Get the next function in the list"""
    if func is None:
        return None
    return func.Next


def function_is_declared(func: rawptr[Function]) -> int:
    """Get whether the function is a declaration (1=declare, 0=define)"""
    if func is None:
        return 0
    return func.IsDeclared


def param_get_name(param: rawptr[Param]) -> str:
    """Get the parameter name"""
    if param is None:
        return None
    return param.Name


def param_get_ty(param: rawptr[Param]) -> rawptr[LLVMType]:
    """Get the parameter type"""
    if param is None:
        return None
    return param.Ty


def param_get_next(param: rawptr[Param]) -> rawptr[Param]:
    """Get the next parameter in the list"""
    if param is None:
        return None
    return param.Next