fix: raise C stack for INTERP tail-recursion benchmarks

.github/workflows/build.yml

@@ -177,10 +177,9 @@ jobs:
           (cd $GITHUB_WORKSPACE/test/lua && $GITHUB_WORKSPACE/build/bin/unit_tests)
           ./bin/flua
 
-      # 单独一步而不是内联 if：这样 PR 里 GitHub 会把它显式标成 skipped，看得出是按策略
-      # 跳过而不是没配。限定 release 是因为这是 matrix 构建，不限定的话一次 push 要跑四遍。
+      # 限定 release：matrix 有 debug/asan/cov，不限定的话一次 push 要跑四遍。
       - name: Run benchmarks
-        if: github.event_name != 'pull_request' && matrix.build_type == 'release'
+        if: matrix.build_type == 'release'
         run: |
           set -x
           cd build/bin

.github/workflows/build_with_macos.yml

@@ -99,10 +99,7 @@ jobs:
           (cd $GITHUB_WORKSPACE/test/lua && $GITHUB_WORKSPACE/build/bin/unit_tests)
           ./bin/flua
 
-      # 单独一步而不是内联 if：这样 PR 里 GitHub 会把它显式标成 skipped，看得出是按策略
-      # 跳过而不是没配。
       - name: Run benchmarks
-        if: github.event_name != 'pull_request'
         run: |
           set -x
           cd build/bin

.github/workflows/build_with_windows.yml

@@ -178,7 +178,6 @@ jobs:
           ./flua.exe
 
       - name: Run benchmarks
-        if: github.event_name != 'pull_request'
         shell: msys2 {0}
         run: |
           set -eux

benchmark/README.md

@@ -16,8 +16,6 @@ build/bin/bench_mark --benchmark_filter='BM_Lua_|BM_FakeLua_.*_INTERP' \
 build/bin/bench_mark --benchmark_repetitions=1 --benchmark_report_aggregates_only=true
 ```
 
-`BM_FakeLua_TailRecursion_INTERP/5000` is a 5000-deep C++ call chain (the interpreter does not turn tail calls into loops). The default 8 MiB C stack overflows; run with `ulimit -s unlimited` if that case is included.
-
 ---
 
 ## Interpreter vs Lua 5.4 (2026-09-16)

benchmark/README.zh.md

@@ -14,8 +14,6 @@ build/bin/bench_mark --benchmark_filter='BM_Lua_|BM_FakeLua_.*_INTERP' \
 build/bin/bench_mark --benchmark_repetitions=1 --benchmark_report_aggregates_only=true
 ```
 
-`BM_FakeLua_TailRecursion_INTERP/5000` 是 5000 层 C++ 调用（解释器没有把尾调用收成循环）。默认 8 MiB C 栈会溢出；若包含该用例请先 `ulimit -s unlimited`。
-
 ---
 
 ## 解释器 vs Lua 5.4（2026-09-16）

benchmark/benchmark_function.cpp

@@ -70,19 +70,6 @@ function bench_closure(n)
 end
 )";
 
-constexpr const char *kTailRecursionScript = R"(
-function bench_tail_sum_acc(x, acc)
-    if x <= 0 then
-        return acc
-    end
-    return bench_tail_sum_acc(x - 1, acc + x)
-end
-
-function bench_tail_sum(n)
-    return bench_tail_sum_acc(n, 0)
-end
-)";
-
 // C++ reference implementations
 
 inline int64_t CppEmpty(int64_t n) {
@@ -129,18 +116,10 @@ int64_t CppClosure(int64_t n) {
     return total;
 }
 
-int64_t CppTailSum(int64_t n) {
-    int64_t acc = 0;
-    for (int64_t x = n; x > 0; --x) {
-        acc += x;
-    }
-    return acc;
-}
-
 // Lua helpers
 
 const char *const kFunctionScripts[] = {
-        kEmptyCallScript, kRecursionScript, kVariadicScript, kMultiReturnScript, kClosureScript, kTailRecursionScript,
+        kEmptyCallScript, kRecursionScript, kVariadicScript, kMultiReturnScript, kClosureScript,
 };
 constexpr size_t kFunctionScriptCount = sizeof(kFunctionScripts) / sizeof(kFunctionScripts[0]);
 
@@ -159,8 +138,6 @@ struct Ctx : RuntimeContext {
         Call(flua, JIT_INTERP, "bench_multi_return", w, 10);
         Call(flua, JIT_TCC, "bench_closure", w, 10);
         Call(flua, JIT_INTERP, "bench_closure", w, 10);
-        Call(flua, JIT_TCC, "bench_tail_sum", w, 10);
-        Call(flua, JIT_INTERP, "bench_tail_sum", w, 10);
     }
 
     ~Ctx() {
@@ -392,51 +369,6 @@ static void BM_FakeLua_Closure_INTERP(benchmark::State &state) {
     }
 }
 
-// Benchmarks: tail recursion
-
-static void BM_CPP_TailRecursion(benchmark::State &state) {
-    const int64_t n = state.range(0);
-    for (auto _: state) {
-        int64_t ret = CppTailSum(n);
-        benchmark::DoNotOptimize(ret);
-    }
-}
-
-static void BM_Lua_TailRecursion(benchmark::State &state) {
-    const int64_t n = state.range(0);
-    for (auto _: state) {
-        int64_t ret = CallLuaInt(g_ctx.lua, "bench_tail_sum", n);
-        benchmark::DoNotOptimize(ret);
-    }
-}
-
-static void BM_FakeLua_TailRecursion_TCC(benchmark::State &state) {
-    const int64_t n = state.range(0);
-    for (auto _: state) {
-        int64_t ret = 0;
-        Call(g_ctx.flua, JIT_TCC, "bench_tail_sum", ret, n);
-        benchmark::DoNotOptimize(ret);
-    }
-}
-
-static void BM_FakeLua_TailRecursion_GCC(benchmark::State &state) {
-    const int64_t n = state.range(0);
-    for (auto _: state) {
-        int64_t ret = 0;
-        Call(g_ctx.flua, JIT_GCC, "bench_tail_sum", ret, n);
-        benchmark::DoNotOptimize(ret);
-    }
-}
-
-static void BM_FakeLua_TailRecursion_INTERP(benchmark::State &state) {
-    const int64_t n = state.range(0);
-    for (auto _: state) {
-        int64_t ret = 0;
-        Call(g_ctx.flua, JIT_INTERP, "bench_tail_sum", ret, n);
-        benchmark::DoNotOptimize(ret);
-    }
-}
-
 }// namespace
 
 // Benchmark registrations
@@ -446,7 +378,6 @@ static void BM_FakeLua_TailRecursion_INTERP(benchmark::State &state) {
 #define VARIADIC_ARGS ->Arg(1)
 #define MULTI_RETURN_ARGS ->Arg(1000)->Arg(10000)
 #define CLOSURE_ARGS ->Arg(100)->Arg(1000)
-#define TAIL_RECURSION_ARGS ->Arg(100)->Arg(1000)->Arg(5000)
 
 BENCHMARK(BM_CPP_EmptyCall) EMPTY_CALL_ARGS;
 BENCHMARK(BM_Lua_EmptyCall) EMPTY_CALL_ARGS;
@@ -472,9 +403,4 @@ BENCHMARK(BM_CPP_Closure) CLOSURE_ARGS;
 BENCHMARK(BM_Lua_Closure) CLOSURE_ARGS;
 BENCHMARK(BM_FakeLua_Closure_TCC) CLOSURE_ARGS;
 BENCHMARK(BM_FakeLua_Closure_GCC) CLOSURE_ARGS;
-BENCHMARK(BM_FakeLua_Closure_INTERP) CLOSURE_ARGS;
-BENCHMARK(BM_CPP_TailRecursion) TAIL_RECURSION_ARGS;
-BENCHMARK(BM_Lua_TailRecursion) TAIL_RECURSION_ARGS;
-BENCHMARK(BM_FakeLua_TailRecursion_TCC) TAIL_RECURSION_ARGS;
-BENCHMARK(BM_FakeLua_TailRecursion_GCC) TAIL_RECURSION_ARGS;
-BENCHMARK(BM_FakeLua_TailRecursion_INTERP) TAIL_RECURSION_ARGS;
+BENCHMARK(BM_FakeLua_Closure_INTERP) CLOSURE_ARGS;

test/lua/jit/test_tail_call.lua

@@ -0,0 +1,107 @@
+-- Tail-call correctness. Depth stays modest so INTERP (no TCO) does not overflow the C stack.
+
+function tail_sum(x, acc)
+    if x <= 0 then
+        return acc
+    end
+    return tail_sum(x - 1, acc + x)
+end
+
+function test_tail_sum()
+    if tail_sum(0, 0) ~= 0 then return 0 end
+    if tail_sum(1, 0) ~= 1 then return 0 end
+    if tail_sum(10, 0) ~= 55 then return 0 end
+    if tail_sum(100, 0) ~= 5050 then return 0 end
+    return 1
+end
+
+function even(n)
+    if n == 0 then return true end
+    return odd(n - 1)
+end
+
+function odd(n)
+    if n == 0 then return false end
+    return even(n - 1)
+end
+
+function test_tail_even_odd()
+    if even(0) ~= true then return 0 end
+    if odd(0) ~= false then return 0 end
+    if even(10) ~= true then return 0 end
+    if odd(11) ~= true then return 0 end
+    if even(99) ~= false then return 0 end
+    return 1
+end
+
+function ident(x)
+    return x
+end
+
+function tail_through(x)
+    return ident(x + 1)
+end
+
+function test_tail_other_func()
+    if tail_through(41) ~= 42 then return 0 end
+    return 1
+end
+
+function two_rets(a, b)
+    return a, b
+end
+
+function tail_multi(x)
+    return two_rets(x, x + 1)
+end
+
+function test_tail_multi_ret()
+    local a, b = tail_multi(7)
+    if a ~= 7 then return 0 end
+    if b ~= 8 then return 0 end
+    return 1
+end
+
+function add1(a)
+    return a
+end
+
+function add3(a, b, c)
+    return a + b + c
+end
+
+function tail_to_add1(x)
+    return add1(x)
+end
+
+function tail_to_add3(x)
+    return add3(x, 2, 3)
+end
+
+function test_tail_arity()
+    if tail_to_add1(5) ~= 5 then return 0 end
+    if tail_to_add3(5) ~= 10 then return 0 end
+    return 1
+end
+
+function tail_count(...)
+    local n = select('#', ...)
+    if n <= 1 then return n end
+    return tail_count(select(2, ...))
+end
+
+function test_tail_vararg()
+    if tail_count() ~= 0 then return 0 end
+    if tail_count(1) ~= 1 then return 0 end
+    if tail_count(1, 2, 3, 4, 5) ~= 1 then return 0 end
+    return 1
+end
+
+function test_tail_closure()
+    local function walk(n, acc)
+        if n <= 0 then return acc end
+        return walk(n - 1, acc + n)
+    end
+    if walk(20, 0) ~= 210 then return 0 end
+    return 1
+end

test/test_jitter.cpp

@@ -3006,3 +3006,24 @@ TEST(jitter, spec_field_merge_sanitize) {
         ASSERT_EQ(ret, 2);
     });
 }
+
+TEST(jitter, tail_call) {
+    JitterRunHelper([](State *s, JITType type, bool debug_mode) {
+        CompileFile(s, "./jit/test_tail_call.lua", {.debug_mode = debug_mode});
+        int64_t ret = 0;
+        Call(s, type, "test_tail_sum", ret);
+        ASSERT_EQ(ret, 1);
+        Call(s, type, "test_tail_even_odd", ret);
+        ASSERT_EQ(ret, 1);
+        Call(s, type, "test_tail_other_func", ret);
+        ASSERT_EQ(ret, 1);
+        Call(s, type, "test_tail_multi_ret", ret);
+        ASSERT_EQ(ret, 1);
+        Call(s, type, "test_tail_arity", ret);
+        ASSERT_EQ(ret, 1);
+        Call(s, type, "test_tail_vararg", ret);
+        ASSERT_EQ(ret, 1);
+        Call(s, type, "test_tail_closure", ret);
+        ASSERT_EQ(ret, 1);
+    });
+}