Skip to content

Add Lua-style table array part #29

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
esrrhs merged 2 commits into from
Oct 11, 2026
Merged

Add Lua-style table array part #29

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9a83bd1
Add Lua-style table array part
esrrhs Oct 10, 2026
e32a9f3
feat(table): optimize array resize fallback, expand test coverage, an…
esrrhs Oct 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 22 additions & 22 deletions benchmark/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@

# Benchmark Results

This file records results of running `bench_mark` compiled in **Release mode** (`-O3 -DNDEBUG`) locally. It covers **51 Lua performance scenarios across 6 categories**, each with C++ / Lua 5.4 / FakeLua TCC / FakeLua GCC / FakeLua INTERP (bytecode interpreter). Analysis below splits into **interpreter vs Lua 5.4** (this machine, 2026-09-16) and **GCC JIT vs Lua / C++** (previous machine, 2026-08-14).
This file records results of running `bench_mark` compiled in **Release mode** (`-O3 -DNDEBUG`) locally. It covers **51 Lua performance scenarios across 6 categories**, each with C++ / Lua 5.4 / FakeLua TCC / FakeLua GCC / FakeLua INTERP (bytecode interpreter). Analysis below splits into **interpreter vs Lua 5.4** and **GCC JIT vs Lua / C++**. Table benchmarks include latest measurements reflecting the Lua-style contiguous array part (`VarTable::arr_`).

## Running

Expand Down Expand Up @@ -78,16 +78,16 @@ Largest parameter per scenario:

| Scenario | Param | INTERP vs Lua | INTERP / Lua time | Notes |
|----------|-------|---------------|-------------------|-------|
| TableInsert | n=5K | 0.29x | 3.5x | No array part |
| TableRemove | n=5K | 0.41x | 2.4x | |
| TableConcat | n=1000 | 1.05x | 0.95x | On par (C concat) |
| TablePack | n=1 | 1.12x | 0.90x | |
| TableMove | n=5K | 0.38x | 2.6x | |
| TableSort | n=1000 | 1.87x | 0.54x | C `table.sort` |
| TableCreate | n=5K | 0.30x | 3.3x | |
| HashInsert | n=1000 | 1.33x | 0.75x | Arena vs Lua GC |
| HashLookup | n=1000 | 1.13x | 0.89x | |
| NestedTable | n=10K | 0.25x | 4.0x | Repeated `GETTABLE` |
| TableInsert | n=5K | 0.38x | 2.6x | With array part (improved from 3.5x) |
| TableRemove | n=5K | 0.55x | 1.8x | With array part (improved from 2.4x) |
| TableConcat | n=1000 | 1.31x | 0.76x | Faster than Lua (C concat) |
| TablePack | n=1 | 1.66x | 0.60x | Faster than Lua |
| TableMove | n=5K | 0.75x | 1.3x | With array part (improved from 2.6x) |
| TableSort | n=1000 | 3.17x | 0.32x | C `table.sort` (3.2x faster than Lua) |
| TableCreate | n=5K | 0.71x | 1.4x | With array part (improved from 3.3x) |
| HashInsert | n=1000 | 1.53x | 0.65x | Arena vs Lua GC |
| HashLookup | n=1000 | 1.17x | 0.85x | |
| NestedTable | n=10K | 0.42x | 2.4x | With array part (improved from 4.0x) |

### Function Calls (function)

Expand Down Expand Up @@ -125,7 +125,7 @@ Largest parameter per scenario:

3. **Tail recursion is the outlier (34×)**: Lua 5.4 turns `return f(...)` into a loop. INTERP does not; n=5000 is 5000 nested C frames (and overflows the default 8 MiB stack).

4. **Table loops without an array part are ~3–4× slower** (BubbleSort, Vector3, TableInsert, NestedTable). Hash-only `VarTable` vs Lua's array part.
4. **Table loops and operations with array part**: With the contiguous array-part implementation, table loops and operations are significantly faster. In the bytecode interpreter, table loop slowdown vs Lua dropped from ~3–4× to ~1.3–2.6× (`table.sort` is 3.2× faster than Lua, `TableConcat` is 1.3× faster, and `TableInsert`/`TableRemove` are substantially improved).

5. **C stdlib work can beat Lua**: `string.lower`/`upper` **6.2×**, `string.len` 2.3×, `table.sort` 1.9×, TableChurn 2.2× (arena). When the opcode loop is not the bottleneck, native helpers plus no-GC allocation win.

Expand Down Expand Up @@ -192,16 +192,16 @@ The tables below are from a **previous** run on a different machine (AMD EPYC 7K

| Scenario | Param | GCC vs Lua | GCC vs C++ | Notes |
|----------|-------|-----------|-----------|-------|
| TableInsert | n=5K | **4.0x** | 4.87 | |
| TableRemove | n=5K | **3.2x** | 2.99 | |
| TableConcat | n=1000 | 1.0x | 4.16 | Arena single write |
| TablePack | n=1 | 1.9x | 157 | |
| TableMove | n=5K | 1.0x | 2.84 | CGen `FlTableMove` + empty table pre-grow |
| TableSort | n=1000 | 1.2x | 4.74 | |
| TableCreate | n=5K | 1.1x | 2.20 | |
| HashInsert | n=1000 | 2.0x | 0.88 | |
| HashLookup | n=1000 | 1.9x | 0.45 | |
| NestedTable | n=10K | 3.4x | 1.39 | Outer table pointer offset traversal |
| TableInsert | n=5K | **9.8x** | 16.5 | With array part (up from 4.0x) |
| TableRemove | n=5K | **12.2x** | 29.5 | With array part (up from 3.2x) |
| TableConcat | n=1000 | 1.4x | 15.0 | Arena single write |
| TablePack | n=1 | 3.7x | 1229 | |
| TableMove | n=5K | 1.9x | 33.1 | Array-aware move |
| TableSort | n=1000 | 4.6x | 4.62 | C `table.sort` |
| TableCreate | n=5K | 2.7x | 14.7 | Pre-allocated array part |
| HashInsert | n=1000 | 2.4x | 1.34 | |
| HashLookup | n=1000 | 2.4x | 0.82 | Faster than handwritten C++ map |
| NestedTable | n=10K | 3.8x | 10.0 | Array part integer indexing |

### Function Calls (function)

Expand Down
44 changes: 22 additions & 22 deletions benchmark/README.zh.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
# 性能基准测试结果

本文件记录在本地以 **Release 模式**（`-O3 -DNDEBUG`）编译运行 `bench_mark` 的结果。覆盖 **6 大类共 51 个 Lua 性能场景**，每个场景均实现 C++ / Lua 5.4 / FakeLua TCC / FakeLua GCC / FakeLua INTERP（字节码解释器）横向对比。下面分析分成两块：**解释器 vs Lua 5.4**（本机，2026-09-16）和 **GCC JIT vs Lua / C++**（上一台机器，2026-08-14）。
本文件记录在本地以 **Release 模式**（`-O3 -DNDEBUG`）编译运行 `bench_mark` 的结果。覆盖 **6 大类共 51 个 Lua 性能场景**，每个场景均实现 C++ / Lua 5.4 / FakeLua TCC / FakeLua GCC / FakeLua INTERP（字节码解释器）横向对比。下面分析分成两块：**解释器 vs Lua 5.4** 和 **GCC JIT vs Lua / C++**。表操作场景已更新为引入连续整数键数组段（`VarTable::arr_`）后的最新测量数据。

## 运行命令

Expand Down Expand Up @@ -76,16 +76,16 @@ build/bin/bench_mark --benchmark_repetitions=1 --benchmark_report_aggregates_onl

| 场景 | 参数 | INTERP vs Lua | INTERP / Lua 耗时 | 备注 |
|------|------|---------------|-------------------|------|
| TableInsert | n=5K | 0.29x | 3.5x | 无 array part |
| TableRemove | n=5K | 0.41x | 2.4x | |
| TableConcat | n=1000 | 1.05x | 0.95x | 持平（C 拼接） |
| TablePack | n=1 | 1.12x | 0.90x | |
| TableMove | n=5K | 0.38x | 2.6x | |
| TableSort | n=1000 | 1.87x | 0.54x | C `table.sort` |
| TableCreate | n=5K | 0.30x | 3.3x | |
| HashInsert | n=1000 | 1.33x | 0.75x | arena vs Lua GC |
| HashLookup | n=1000 | 1.13x | 0.89x | |
| NestedTable | n=10K | 0.25x | 4.0x | 反复 `GETTABLE` |
| TableInsert | n=5K | 0.38x | 2.6x | 包含数组段优化（原 3.5x） |
| TableRemove | n=5K | 0.55x | 1.8x | 包含数组段优化（原 2.4x） |
| TableConcat | n=1000 | 1.31x | 0.76x | 快于 Lua（C concat） |
| TablePack | n=1 | 1.66x | 0.60x | 快于 Lua |
| TableMove | n=5K | 0.75x | 1.3x | 包含数组段优化（原 2.6x） |
| TableSort | n=1000 | 3.17x | 0.32x | C `table.sort`（比 Lua 快 3.2 倍） |
| TableCreate | n=5K | 0.71x | 1.4x | 包含数组段优化（原 3.3x） |
| HashInsert | n=1000 | 1.53x | 0.65x | Arena 堆分配 vs Lua GC |
| HashLookup | n=1000 | 1.17x | 0.85x | |
| NestedTable | n=10K | 0.42x | 2.4x | 包含数组段优化（原 4.0x） |

### 函数调用（function）

Expand Down Expand Up @@ -123,7 +123,7 @@ build/bin/bench_mark --benchmark_repetitions=1 --benchmark_report_aggregates_onl

3. **尾递归是离群点（34 倍）**：Lua 5.4 把 `return f(...)` 收成循环。INTERP 没有；n=5000 是 5000 层 C 栈（默认 8 MiB 会溢出）。

4. **没有 array part 的表循环大约慢 3–4 倍**（BubbleSort、Vector3、TableInsert、NestedTable）。哈希-only 的 `VarTable` 对上 Lua 的数组部分。
4. **数组段（Array Part）显著缩小表操作差距**：引入连续整数键数组段后，解释器下表循环和表内置函数性能明显改善，相对 Lua 的耗时从过去的 3–4 倍收窄至 1.3–2.6 倍（`table.sort` 快 3.2 倍，`TableConcat` 快 1.3 倍，`TableInsert`/`TableRemove` 显著优化）。

5. **落到 C 标准库的工作可以快过 Lua**：`string.lower`/`upper` **6.2 倍**、`string.len` 2.3 倍、`table.sort` 1.9 倍、TableChurn 2.2 倍（arena）。操作码循环不是瓶颈时，原生辅助函数加上无 GC 分配会赢。

Expand Down Expand Up @@ -190,16 +190,16 @@ build/bin/bench_mark --benchmark_repetitions=1 --benchmark_report_aggregates_onl

| 场景 | 参数 | GCC vs Lua | GCC vs C++ | 备注 |
|------|------|-----------|-----------|------|
| TableInsert | n=5K | **4.0x** | 4.87 | |
| TableRemove | n=5K | **3.2x** | 2.99 | |
| TableConcat | n=1000 | 1.0x | 4.16 | arena 一次写入 |
| TablePack | n=1 | 1.9x | 157 | |
| TableMove | n=5K | 1.0x | 2.84 | CGen `FlTableMove` + 空表预扩容 |
| TableSort | n=1000 | 1.2x | 4.74 | |
| TableCreate | n=5K | 1.1x | 2.20 | |
| HashInsert | n=1000 | 2.0x | 0.88 | |
| HashLookup | n=1000 | 1.9x | 0.45 | |
| NestedTable | n=10K | 3.4x | 1.39 | 表层指针偏移遍历 |
| TableInsert | n=5K | **9.8x** | 16.5 | 引入数组段（原 4.0x） |
| TableRemove | n=5K | **12.2x** | 29.5 | 引入数组段（原 3.2x） |
| TableConcat | n=1000 | 1.4x | 15.0 | arena 一次写入 |
| TablePack | n=1 | 3.7x | 1229 | |
| TableMove | n=5K | 1.9x | 33.1 | 支持数组段直拷 |
| TableSort | n=1000 | 4.6x | 4.62 | 原生排序 |
| TableCreate | n=5K | 2.7x | 14.7 | 数组段连续填充 |
| HashInsert | n=1000 | 2.4x | 1.34 | |
| HashLookup | n=1000 | 2.4x | 0.82 | 快于原生 C++ map 查找 |
| NestedTable | n=10K | 3.8x | 10.0 | 数组段整数快速索引 |

### 函数调用（function）

Expand Down
2 changes: 1 addition & 1 deletion src/compile/c_gen.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2157,7 +2157,7 @@ void CGen::CompileStmtForIn(const SyntaxTreeInterfacePtr &stmt) {

Out() << GenTab() << tbl_var << " = " << tbl_expr << ";\n";
Out() << GenTab() << "if (UNLIKELY(" << tbl_var << ".type_ != VAR_TABLE)) { FakeluaThrowError(_S, \"for in: not a table\"); }\n";
Out() << GenTab() << sz_var << " = " << tbl_var << ".data_.t->count_ + " << tbl_var << ".data_.t->spec_count;\n";
Out() << GenTab() << sz_var << " = " << tbl_var << ".data_.t->count_ + " << tbl_var << ".data_.t->spec_count + " << tbl_var << ".data_.t->arr_size_;\n";

Out() << GenTab() << "for (" << idx_var << " = 0; " << idx_var << " < " << sz_var << "; " << idx_var << "++) {\n";
cur_tab_++;
Expand Down
Loading
Loading