feat(ccr): v10 多路 match 与模式覆盖

build_selfhost_native.py

@@ -95,6 +95,7 @@
 # + rax 回存；内置体名索引扫描段留 elf.cr 原位）；跨轴消费 x86_64/instr.cr
 # 编码原语（e2_mov/e2_w8/e2_store_ret）。
 os_linux_files = [
+    'src/os/linux/file_mode.cr',
     'src/os/linux/entry.cr',
     'src/os/linux/callseq.cr',
     'src/os/linux/syscall.cr',

docs/superpowers/plans/2026-10-24-multibranch-v10.md

@@ -0,0 +1,220 @@
+# v10 多路分支 / match 决策 Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** 在 v10 `.ccr` 上实现多路 `MATCH` 节点和 CASE 表：corec 把 `match` 降为多路分支，corearch 映射为比较链/二分/跳转表，interp 与 ELF 语义一致。
+
+**Architecture:** 保留 `IR_BRANCH` 作为 `if`/`while` 的二元分支；新增 `IR_MATCH` 和 CASE 记录。CASE 是 v10 第 9 段，记录为 24B `{value, kind, str_index, target_nod, flags}`。先做整数/bool/enum 标签的多路分派；字符串/struct/ident 模式在后续任务补齐。
+
+**Spec:** `docs/superpowers/specs/2026-10-23-pattern-matching-ir-design.md`（v9.1 设计，本计划适配 v10）+ `docs/superpowers/specs/2026-10-23-lattice-ccr-v10-design.md` §11 CASE。
+
+## 全局约束
+
+- 不改变 `IR_BRANCH` 的二元编码；`if`/`while` 继续走现有路径。
+- `MATCH` 只表示多路分派；`CASE` 不存宽度、偏移、ABI 或机器码。
+- CASE 值只进入 v10 CASE 段；字符串模式后续归一化为 `str_eq`，不把字符串指针放进 CASE。
+- 每个任务单独提交；先 CASE/IR_MATCH 读写，再 corec，再 corearch/interp。
+- 自举要求 stage2/stage3 逐字节一致；`test_bootstrap` 和 canary 不因中间态长时间红。
+- `jj` only；子代理每个任务新开，禁止并行实现同一文件。
+- 所有长命令 `nice -n 19`。
+
+---
+
+### Task 1: IR_MATCH/CASE schema 与 v10 读写
+
+**Files:**
+- Modify: `src/compiler/ast.cr`
+- Modify: `src/compiler/globals.cr`
+- Modify: `src/compiler/ccr_schema.cr`
+- Modify: `src/compiler/ccr_io.cr`
+- Test: `tests/selfhost/test_multibranch_v10.py`
+
+**Interfaces:**
+- Consumes: v10 CASE 段 tag 9，24B 记录。
+- Produces:
+  - `IR_MATCH : int = 54`
+  - `g_ccr_case : string`, `g_ccr_case_count : int`
+  - CASE record offsets/stride:
+    - `CCR_CASE_VALUE=0` (8B signed)
+    - `CCR_CASE_KIND=8` (4B)
+    - `CCR_CASE_STR=12` (4B signed)
+    - `CCR_CASE_TARGET=16` (4B signed NOD label id)
+    - `CCR_CASE_FLAGS=20` (4B)
+    - `CCR_CASE_DISK=24`
+  - `CCR_CASE_KIND_VALUE=0`, `KIND_ENUM=1`, `KIND_STRING=2`, `KIND_DEFAULT=3`
+
+- [ ] **Step 1: failing test**
+
+Create a Python test that builds a `.ccr` with an `IR_MATCH` node and 3 CASE records, parses the CASE segment independently, and asserts count/order/value/target. Also assert a v10 file with CASE count > 0 is not rejected by the container (once implemented).
+
+- [ ] **Step 2: implement schema**
+
+Add `IR_MATCH`; `ccr_op_is_known` includes it; CASE segment writer reads `g_ccr_case`; loader rebuilds it. For now `g_ccr_case_count=0` unless corec populates. CASE segment size = `4 + count*24`. Add bounds/str/target checks; default exactly one optional.
+
+- [ ] **Step 3: verify**
+
+```bash
+python3 tests/selfhost/test_multibranch_v10.py
+python3 tests/selfhost/test_ccr_v10_format.py
+python3 tests/selfhost/test_ccr_v10_stream.py
+nice -n 19 python3 build_selfhost_native.py
+```
+
+- [ ] **Step 4: commit**
+
+`jj commit -m "ccr: add v10 match and case schema"`
+
+---
+
+### Task 2: corec `match` 降为 MATCH
+
+**Files:**
+- Modify: `src/compiler/ir_gen.cr`
+- Modify: `src/compiler/dump.cr`
+- Modify: `src/compiler/ccr_types.cr`
+- Modify: `src/compiler/cir_cache.cr`
+- Test: `tests/selfhost/test_multibranch_v10.py`
+
+**Interfaces:**
+- Consumes: Task 1 `IR_MATCH`, `g_ccr_case`.
+- Produces: one `IR_MATCH` for a match whose arms are int/bool/enum/wildcard; body labels are existing labels; default is a `KIND_DEFAULT` case.
+- Unsupported patterns keep the existing sequential lowering until Task 4; do not silently change semantics.
+
+- [ ] **Step 1: RED fixture**
+
+4-way int match and enum match; assert exactly one `MATCH` in the function and CASE count = arms + default; old code produces zero.
+
+- [ ] **Step 2: implement**
+
+Collect cases in arm order, dedupe by value keeping first, `KIND_DEFAULT` for `_`; emit `MATCH` with `src1=scrutinee`, `src2=case_start`, `src3=case_count`. Keep optional representation special path unchanged for now. If a match has string/struct/ident patterns, fall back to existing code and mark a TODO; do not emit a wrong MATCH.
+
+- [ ] **Step 3: CIR cache**
+
+Disable CIR cache for functions containing `MATCH` in this task (simplest correct); add a per-function `g_cir_has_match` reset and check. Task 5 may replace with CASE range persistence.
+
+- [ ] **Step 4: dump**
+
+`--dump-objects` prints `match` with case range and cases.
+
+- [ ] **Step 5: verify**
+
+```bash
+python3 tests/selfhost/test_multibranch_v10.py
+nice -n 19 python3 tests/selfhost/test_ccr_types.py
+nice -n 19 python3 build_selfhost_native.py
+./build/corec check src/compiler
+```
+
+- [ ] **Step 6: commit**
+
+`jj commit -m "ccr: lower integer and enum matches to v10 MATCH"`
+
+---
+
+### Task 3: corearch / interp / dataflow support
+
+**Files:**
+- Modify: `src/arch/x86_64/instr.cr`
+- Modify: `src/arch/x86_64/regalloc.cr`
+- Modify: `src/arch/x86_64/sizes.cr` if present
+- Modify: `src/compiler/dataflow.cr`
+- Modify: `src/compiler/interp.cr`
+- Modify: `src/compiler/ccr_io.cr` validation
+- Test: `tests/selfhost/test_multibranch_v10.py`
+
+**Interfaces:**
+- `MATCH` successors = all case targets (plus default); regalloc handles multiple successors.
+- corearch emits:
+  - count 0 -> default jump/trap
+  - count 1..2 -> cmp/jcc chain
+  - count >2 -> binary search by sorted value (Task 4 can replace with jump table)
+- interp reads the loaded case table and jumps to the target; default on no hit.
+- `MATCH` target labels must be valid NOD labels and in function span; loader validates.
+
+- [ ] **Step 1: RED** 4-way int match ELF and interp; both must return the selected arm; current no MATCH path fails.
+- [ ] **Step 2: implement**; keep `IR_BRANCH` code untouched.
+- [ ] **Step 3: verify**:
+```bash
+python3 tests/selfhost/test_multibranch_v10.py
+python3 tests/selfhost/test_interp_parity.py
+python3 tests/selfhost/test_native_memory.py
+python3 tests/selfhost/test_backend_bootstrap.py
+```
+- [ ] **Step 4: commit**
+`jj commit -m "backend: lower v10 MATCH to compare chains"`
+
+---
+
+### Task 4: dense jump table / sparse binary search
+
+**Files:**
+- Modify: `src/arch/x86_64/instr.cr`
+- Modify: `src/arch/x86_64/regalloc.cr`
+- Modify: `src/arch/x86_64/sizes.cr`
+- Test: `tests/selfhost/test_multibranch_v10.py`
+
+**Interfaces:**
+- Dense: `count >= 8 && max-min+1 <= 2*count` -> 32-bit relative table in rodata; range check; default out-of-range.
+- Sparse: sorted binary comparison tree.
+- `s3==1` with value 1 and default remains same machine shape as existing `IR_BRANCH` if possible.
+- Add 16/64-way benchmarks and assert no linear chain on dense case.
+
+- [ ] **Step 1: RED** 64-way dense fixture; assert ELF contains table path / bounded comparisons.
+- [ ] **Step 2: implement** with label patch mechanism; no PIE/dynamic relocations.
+- [ ] **Step 3: verify and commit**
+`ccr: add jump-table lowering for dense match`
+
+---
+
+### Task 5: pattern coverage
+
+**Files:**
+- Modify: `src/compiler/parser.cr` (tuple pattern if missing)
+- Modify: `src/compiler/checker.cr`
+- Modify: `src/compiler/ir_gen.cr`
+- Modify: `bootstrap/corec/frontend/desugar.py`
+- Test: `tests/selfhost/test_multibranch_v10.py`
+
+**Interfaces:**
+- string/char patterns normalize to `str_eq` and binary MATCH/default.
+- struct/tuple fields load then recursively dispatch.
+- ident binds; wildcard default.
+- literal exhaustiveness: `_` or irrefutable arm else `EC_TM_EXHAUST`.
+- corec source must avoid unsupported bootstrap match patterns; extend desugar if needed.
+
+- [ ] **Steps** TDD per pattern: string, char, struct, tuple, ident; behavior parity; no silent first-arm.
+- [ ] **Commit** `ccr: cover string struct and ident match patterns`
+
+---
+
+### Task 6: cleanup, cache, canary
+
+**Files:**
+- Modify: `src/compiler/cir_cache.cr` (persist CASE range or keep disable)
+- Modify: `tools/baseline/canary_values.tsv`
+- Modify: `docs/superpowers/specs/2026-10-23-pattern-matching-ir-design.md` (v10 adaptation)
+- Test: `tests/selfhost/test_multibranch_v10.py`
+
+- [ ] Replace temporary cache disable with CASE-range persistence if feasible; otherwise document.
+- [ ] Update bootstrap parser for V4 syntax tests if CI still red.
+- [ ] Run full gate and relock only affected `.ccr`/ELF with attribution.
+- [ ] Commit.
+
+## Final Gate
+
+```bash
+nice -n 19 python3 build_selfhost_native.py
+./build/corec check src/compiler
+python3 tests/selfhost/test_multibranch_v10.py
+python3 tests/selfhost/test_ccr_v10_format.py
+python3 tests/selfhost/test_ccr_v10_stream.py
+python3 tests/selfhost/test_ccr_types.py
+python3 tests/selfhost/test_backend_bootstrap.py
+python3 tests/selfhost/test_interp_parity.py
+python3 tests/selfhost/test_native_memory.py
+python3 tests/harness/test_ci_hook_coverage.py
+bash tools/baseline/canary_check.sh
+./build/corec build src/compiler/main.cr -o build/corec2 --static -O 0
+./build/corec2 build src/compiler/main.cr -o build/corec3 --static -O 0
+cmp build/corec2 build/corec3
+```

docs/superpowers/specs/2026-10-23-pattern-matching-ir-design.md

@@ -1,5 +1,13 @@
 # match 的决策树编译与统一分支节点
 
+> **v10 实现说明（2026-10-24）**：现行 `.ccr` 是 v10，不是本文写作时的 v9.1。
+> 多路分支在 v10 中使用 `IR_MATCH`（opcode 54）+ CASE 24B 记录
+> `{value, kind, str_index, target_nod, flags}`；`IR_BRANCH` 继续承载二元
+> `if`/`while`。本文的 v9.1 版本/段表/16B CASE 段落是历史设计，已被
+> `docs/superpowers/specs/2026-10-23-lattice-ccr-v10-design.md` §11 取代。
+> 已实现范围：int/bool/enum/string/char/struct/tuple/ident 模式、二分与密集
+> 跳转表、CIR CASE 区间持久化。下文保留历史决策与推导。
+
 ## 概述
 
 `match` 目前在 `src/compiler/ir_gen.cr:2683` 起按 arm 顺序展开为 `IR_CONST`、`IR_BINARY`、`IR_BRANCH` 组成的比较链，corec 的 IR 中只有二元条件分支。字符串、字符、结构体、标识符模式在 `ir_gen` 中没有测试或绑定生成，会静默执行第一个 arm。

src/arch/x86_64/backend_arrays.cr

@@ -215,6 +215,35 @@ fn grow_rodataref(needed: int):
     g_x86_rodataref_ro = n2
     g_x86_rodataref_cap = nc
 
+fn grow_match_table_ref(needed: int):
+    if needed < g_match_table_ref_cap:
+        return
+    nc : ., mut = g_match_table_ref_cap * 2
+    if nc < 32:
+        nc = 32
+    if nc < needed:
+        nc = needed + 32
+    n1 := alloc(nc * 8)
+    _dyncpy(g_match_table_ref_pos, g_match_table_ref_cap * 8, n1)
+    g_match_table_ref_pos = n1
+    n2 := alloc(nc * 8)
+    _dyncpy(g_match_table_ref_id, g_match_table_ref_cap * 8, n2)
+    g_match_table_ref_id = n2
+    g_match_table_ref_cap = nc
+
+fn grow_label_abs(needed: int):
+    if needed < g_label_abs_cap:
+        return
+    nc : ., mut = g_label_abs_cap * 2
+    if nc < 64:
+        nc = 64
+    if nc < needed:
+        nc = needed + 64
+    nb := alloc(nc * 8)
+    _dyncpy(g_label_abs, g_label_abs_cap * 8, nb)
+    g_label_abs = nb
+    g_label_abs_cap = nc
+
 fn grow_alloc_patch(needed: int):
     if needed < g_x86_alloc_patch_cap:
         return

src/arch/x86_64/backend_flags.cr

@@ -23,6 +23,7 @@ fn register_backend_flags():
     cli_flag_bool_debug("inject-read-gap", "", "Debug hook: truncate last version interval")
     cli_flag_bool_debug("inject-coexist-oob", "", "Debug hook: probe entries_coexist out of bounds")
     cli_flag_bool_debug("dump-objects", "", "Debug: dump NOD/EDG objects")
+    cli_flag_bool_debug("dump-match-plan", "", "Debug: dump MATCH dispatch strategy")
     cli_flag_bool_debug("dump-nod-types", "", "Debug: dump NOD type_id/flags")
     cli_flag_bool_debug("dump-alloc-strides", "", "Debug: dump ALLOC_ARRAY element type/stride")
     cli_flag_bool_debug("dump-types", "", "Debug: dump TYPE segment")

src/arch/x86_64/backend_globals.cr

@@ -66,6 +66,26 @@ g_x86_alloc_patch_pos : string, mut
 g_x86_alloc_patch_cap : int, mut
 g_x86_alloc_patch_count : int, mut
 g_x86_emit_rt_stubs : int, mut         // 1 = pure-static: backend emits g_set_curg/g_get_curg stubs
+// Count of IR_MATCH nodes that actually entered e2_emit_match_binary. This is a
+// verification counter, not a strategy predicate: a scratch-only emitter pass in
+// --dump-match-plan and the real Phase 3 emission both increment it.
+g_match_bin_emits : int, mut
+// Dense jump-table state. g_match_table_cursor is the table ordinal during code
+// emission; each emitted LEA is recorded with that ordinal and patched only when
+// the rodata writer has the final table address. This cannot use a
+// g2_rodata_sz()-based offset: Phase 3 can still intern BigInt decimal strings.
+g_match_table_emits : int, mut
+g_match_table_cursor : int, mut
+g_match_table_ro_cursor : int, mut
+g_match_table_ref_pos : string, mut
+g_match_table_ref_id : string, mut
+g_match_table_ref_count : int, mut
+g_match_table_ref_cap : int, mut
+// Label number -> absolute code-buffer position. g_label_poses is per-function
+// and overwritten; case targets are global label numbers, so the rodata phase
+// needs this global map to resolve table entries.
+g_label_abs : string, mut
+g_label_abs_cap : int, mut
 // --big-int: user int slots hold tagged values (immediate v*2 or BigInt pointer+1).
 // The compiler itself and the BigInt runtime blob stay raw.  See int_rep.cr.
 g_x86_big_enabled : int, mut