fix(bench): make the cmake arm actually compile, and stop handing bench a shim

speak-agent · speak-agent · commit 181383596154 · 2026-08-13T18:43:15.000+08:00
**1. cmake + clang 建不了任何真实工程 —— 因为 `set()` 的顺序。**

`CMAKE_CXX_EXTENSIONS OFF` 写在 `project()` **之后**。CMake 为 std 模块合成的
那个目标是在 `project()` 里的编译器探测阶段建出来的,它捕获的是**那一刻**的
`CMAKE_CXX_EXTENSIONS` —— 默认 ON。于是 std.pcm 按 `gnu++23` 编,而所有真实目标
按 `c++23` 编,clang 拒绝加载:

    error: GNU extensions was enabled in precompiled file 'std.pcm'
           but is currently disabled

`import std;` 什么都没提供,构建在 19 处死于 `use of undeclared identifier 'std'`
—— 报错指着 mcpp 的源码,既不提 std.pcm 也不提 extensions。**同一个文件里紧挨着
的注释已经为实验性 key 写下了「必须在 project() 之前」,方言设置漏了同一条规矩。**

mcpp 与 xlings 两份描述都有,都已修。本地实测:改前 19 个错误 build 失败,
改后 `1 ok, 0 failed`,整棵 137 模块的树用 cmake+clang 编通。

**2. `$MCPP` 不是二进制,是 shim —— 上一次「按路径指定」的修复没有走出 shim。**

`$MCPP` = `&lt;xlings home&gt;/subos/default/bin/mcpp`,是一个**指向 `xlings` 的符号
链接**,按 argv[0] 分发并**依据被调用时的工作目录**重新决定 exec 哪个 mcpp。而
bench 刻意让每个引擎在**被测树里**跑,那些树各自带 `.xlings.json`。于是:

    mcpp@2026.8.11.2 | [error] unknown command: build

上一版把裸 `mcpp` 换成 `$MCPP`,只是把解析提前了一步,**没有离开 shim** ——
症状从 `version not found` 变成 `unknown command`,缺陷没动。现在从 xlings 的
`data/runtimedir/mcpp-&lt;ver&gt;-&lt;os&gt;-&lt;arch&gt;/` 里解析出真正的 ELF,并且**断言它自己
报出的版本号等于 pin**(靠目录名叫这个版本是不够的)。解析不到就只跑单臂并
`::warning::`,不再拿一个会失败的引擎充数。

本地直接复现过:同一个目录下 shim 报 2026.8.11.2,解析出的二进制报 2026.8.11.3。

**3. xlings 的 lua_stdlib 生成器静默 return。** 包不在就跳过生成,构建随后死于
`missing mcpplibs.xpkg.lua_stdlib dependency`,指着消费者。与今天早些时候
mcpplibs.cmdline 那个是同一个形状,改成 raise。
diff --git a/.github/workflows/bench.yml b/.github/workflows/bench.yml
@@ -243,6 +243,30 @@ jobs:
           xlings install "mcpp@$REFERENCE_MCPP" -y || echo "::warning::mcpp@$REFERENCE_MCPP unavailable; the reference column will be missing"
           echo "::endgroup::"
 
+          # Resolve the reference to a REAL BINARY, because the thing on PATH
+          # (and $MCPP) is a shim that re-picks a version from the working
+          # directory — see the engine-spec step for what that cost. xlings
+          # unpacks each version to data/runtimedir/mcpp-<ver>-<os>-<arch>/mcpp;
+          # $MCPP is <home>/subos/default/bin/mcpp, so the home is three levels up.
+          ref=""
+          xl_home="$(cd "$(dirname "$MCPP")/../../.." && pwd)"
+          for c in "$xl_home"/data/runtimedir/mcpp-"$REFERENCE_MCPP"-*/mcpp \
+                   "$xl_home"/data/runtimedir/mcpp-"$REFERENCE_MCPP"-*/mcpp.exe; do
+            [ -x "$c" ] || continue
+            # Asserted, not assumed: a binary found by glob under a versioned
+            # directory still has to SAY it is that version, or the reference
+            # column silently compares against something else.
+            got="$("$c" --version 2>/dev/null | grep -oE '[0-9]+(\.[0-9]+){2,3}' | head -1)"
+            if [ "$got" = "$REFERENCE_MCPP" ]; then ref="$c"; break; fi
+            echo "::warning::$c reports '$got', not '$REFERENCE_MCPP'; ignoring it"
+          done
+          if [ -n "$ref" ]; then
+            echo "reference mcpp binary: $ref ($REFERENCE_MCPP)"
+          else
+            echo "::warning::no mcpp@$REFERENCE_MCPP binary under $xl_home/data/runtimedir; the reference column will be missing"
+          fi
+          echo "REFERENCE_BIN=$ref" >> "$GITHUB_ENV"
+
       # Loud, because a version that quietly differs from the pin is the whole
       # class of bug this section exists to end.
       - name: Report the resolved tool versions
@@ -387,18 +411,28 @@ jobs:
               #
               # Putting it back is one line, and the §8 reproduction is the gate:
               # all six fixture scenarios green at --runs 2 before it returns.
-              # BOTH ARMS BY PATH, never the bare `mcpp` shim. The measured
-              # workloads carry their own `.xlings.json`, and the harness runs
-              # every engine with its cwd inside the workload — so a bare `mcpp`
-              # resolves against THAT workspace's pin and fails with
+              # BOTH ARMS BY REAL BINARY — and $MCPP is NOT one.
+              #
+              # $MCPP is `<xlings home>/subos/default/bin/mcpp`, which is a
+              # SYMLINK TO `xlings` that dispatches on argv[0] and re-resolves
+              # which mcpp to exec from the workspace it is invoked in. bench
+              # deliberately runs every engine with its cwd inside the measured
+              # tree, and those trees carry their own `.xlings.json` — so the
+              # shim executed a version nobody asked for and the cells failed as
               #
-              #     xlings: version '2026.8.11.2' not found for 'mcpp'
-              #       available: 2026.8.11.3
+              #     mcpp@2026.8.11.2 | [error] unknown command: build
               #
-              # $MCPP is the binary the bootstrap installed, which IS
-              # reference_mcpp (both come from .xlings.json — asserted by
-              # tests/e2e/233_bench_matrix.sh).
-              e="mcpp=$MCPP_UNDER_TEST,mcpp=$MCPP"
+              # Passing $MCPP here was a fix for the PREVIOUS spelling of this
+              # same bug (a bare `mcpp`, which failed as `version '2026.8.11.2'
+              # not found`). It moved the resolution one step earlier without
+              # leaving the shim, so the symptom changed and the defect did not.
+              # $REFERENCE_BIN is resolved in the step above out of xlings'
+              # runtimedir, i.e. an actual ELF.
+              if [ -n "$REFERENCE_BIN" ]; then
+                e="mcpp=$MCPP_UNDER_TEST,mcpp=$REFERENCE_BIN"
+              else
+                e="mcpp=$MCPP_UNDER_TEST"
+              fi
             fi
             engines="${engines:+$engines,}$e"
           done
diff --git a/bench/projects/mcpp/CMakeLists.txt b/bench/projects/mcpp/CMakeLists.txt
@@ -29,11 +29,29 @@ cmake_minimum_required(VERSION 3.30)
 # compiler-support probe that reads it runs during project().
 set(CMAKE_EXPERIMENTAL_CXX_IMPORT_STD "a9e1cf81-9932-4810-974b-6eccaf14e457")
 
-project(mcpp CXX)
-
+# The DIALECT settings belong before project() for the same reason as the key
+# above, and getting that wrong is not a style question — it is why this arm
+# could not build at all.
+#
+# CMake synthesises its own target for the std module during the compiler probe
+# inside project(). That target captures whatever CMAKE_CXX_EXTENSIONS says AT
+# THAT MOMENT, and the default is ON. Set OFF afterwards, std.pcm is built as
+# `gnu++23` while every real target compiles as `c++23`, and clang refuses the
+# mismatch:
+#
+#   error: GNU extensions was enabled in precompiled file 'std.pcm'
+#          but is currently disabled
+#   error: precompiled file 'std.pcm' cannot be loaded due to a configuration
+#          mismatch with the current compilation
+#
+# `import std;` then supplies nothing and the build dies in 19 places with
+# `use of undeclared identifier 'std'` — an error that points at mcpp's sources
+# and names neither std.pcm nor extensions.
 set(CMAKE_CXX_STANDARD 23)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
+
+project(mcpp CXX)
 # Every mcpp module says `import std;`. This asks CMake to build the standard
 # library module from the compiler's own libstdc++.modules.json, which the
 # hermetic gcc payload ships.
diff --git a/bench/projects/xlings/CMakeLists.txt b/bench/projects/xlings/CMakeLists.txt
@@ -43,11 +43,18 @@ cmake_minimum_required(VERSION 3.30)
 # CMake version — this is the CMake 4.0 key. Must be set BEFORE project().
 set(CMAKE_EXPERIMENTAL_CXX_IMPORT_STD "a9e1cf81-9932-4810-974b-6eccaf14e457")
 
-project(xlings CXX)
-
+# Before project() for the same reason as the key above: CMake builds the std
+# module through a target it synthesises during the compiler probe inside
+# project(), and that target captures CMAKE_CXX_EXTENSIONS as it stands right
+# then — default ON. Setting OFF afterwards yields a `gnu++23` std.pcm that no
+# `c++23` target can load, and the build fails with `use of undeclared
+# identifier 'std'` pointing at the project's own sources. See the identical
+# block in ../mcpp/CMakeLists.txt, where this was diagnosed.
 set(CMAKE_CXX_STANDARD 23)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
+
+project(xlings CXX)
 set(CMAKE_CXX_MODULE_STD 1)
 
 if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
diff --git a/bench/projects/xlings/xmake.lua b/bench/projects/xlings/xmake.lua
@@ -145,7 +145,17 @@ target("xlings")
         -- LUA_STDLIB_DIR is an UPVALUE resolved at description scope: the
         -- helper that produces it is not reachable from inside this callback.
         local stdlib = LUA_STDLIB_DIR
-        if not stdlib or not os.isdir(stdlib) then return end
+        -- FATAL, not a silent return. Returning here skips emitting the module
+        -- and the build dies far away with
+        --     missing mcpplibs.xpkg.lua_stdlib dependency for module ...
+        -- naming a consumer instead of the absent package — which is exactly
+        -- how this failed on CI while passing on a box that had it unpacked.
+        -- (Same defect, same day, as the mcpplibs.cmdline arm in ../mcpp/.)
+        if not stdlib or not os.isdir(stdlib) then
+            raise("bench: mcpplibs.xpkg's lua-stdlib is not unpacked (looked for "
+                  .. tostring(stdlib) .. ") — build the tree with mcpp once first, "
+                  .. "so this arm generates the same module mcpp does")
+        end
 
         local out = path.join(os.projectdir(), "build", "generated", "xpkg-lua-stdlib.cppm")
         local text = {
