[Perf, Test] (MG_Util): libfork execution engine for the shader compile pool, runtime-selectable

Adds libfork v3.8.0 (3rdparty submodule, header-only, wired like the asio
precedent) as a second execution engine behind ShaderCompilePool, selected
per process by MOBILEGL_ASYNC_POOL=asio|libfork (default asio; unknown
values warn and fall back). The engine boundary is deliberately tiny: the
queue, the concurrency budget and its clamping, the suspension latch,
cancel request-vs-outcome, the stopped-is-synchronous fallback and the
drain all stay in the shared Impl - an engine only answers how a
budget-cleared job reaches a worker.

The libfork engine runs detached root tasks as CHAINS: a finished body
takes the next queued job in the same coroutine on the same worker, so
the refills a worker posts are absorbed without a scheduler round trip
(the naive dispatch-thread shape measured 4x worse than asio on short
jobs). Absorption is bounded at one job per live chain - unbounded
absorption serialized bursts posted from inside the pool, which is the
shipped shape (one compile settling fans out link jobs via SubmitAfter
and the adoption map), caught by the review and pinned by a permanent
peak-concurrency regression test (pre-fix: libfork peak 1 vs asio peak 4
on a 16-job worker-posted burst). External submissions go through a
round-robin adaptor instead of lf::lazy_pool::schedule, which both
avoids a data race on lazy_pool's unsynchronized xoshiro under
concurrent submits and beats birthday-collision placement by ~1.3x at
budget == thread count.

The measured answer to "does asio scale poorly": no - the executor was
never the bottleneck. On real pack corpora extracted from the trace
fixtures (BSL 61 shaders, Complementary 277), interleaved best-of-5 per
cell, the engines are within noise of each other at every thread count
(complementary: 4965/2506/1376/824 ms at 1/2/4/8 threads for asio;
libfork within 1%), both ~6x at 8 threads. perf counters show the
flattening past 4 threads is machine-level (instructions flat at 22.1e9
from 1 to 16 threads - no added work, no lock spinning - while cycles
and LLC misses double: memory-stall bound), and the separating control
- N fully independent single-threaded processes with no shared
scheduler at all - scales WORSE than the pool (5.27x vs 5.94x at 8).
The pool microbenchmark does favor libfork on pure dispatch (518 vs
530 ns/job at 1 worker, growing with worker count), but a real compile
body is 1-100 ms, so dispatch is under 0.1% either way. asio therefore
stays the default; this branch exists to make the comparison
reproducible (MG_Test/Util/AsyncPoolBench drives either engine over a
corpus directory) and to keep the alternative viable.

613/613 unit tests in all four combos ({asio, libfork} x {async default
on, kill switch}), integration scenarios byte-identical between engines
on both backends.

.gitmodules

@@ -34,3 +34,6 @@
 [submodule "3rdparty/asio"]
 	path = 3rdparty/asio
 	url = https://github.com/chriskohlhoff/asio.git
+[submodule "3rdparty/libfork"]
+	path = 3rdparty/libfork
+	url = https://github.com/ConorWilliams/libfork.git

CMakeLists.txt

@@ -373,6 +373,13 @@ set(MOBILEGL_INCLUDE_DIR
         # MG_Util/Async/ShaderCompilePool.cpp includes it, and it stays behind that file's
         # pimpl so no consumer target needs this path.
         ${CMAKE_SOURCE_DIR}/3rdparty/asio/asio/include
+        # The second shader-compile execution engine (MOBILEGL_ASYNC_POOL=libfork), on the
+        # same terms as Asio above: header-only, no add_subdirectory (its CMakeLists only
+        # declares an INTERFACE target plus install/test scaffolding we do not want), no link
+        # target, and reachable from exactly one translation unit. libfork's own
+        # target_compile_features asks for cxx_std_23, which this project already sets
+        # globally, so its C++20 coroutines need no per-source standard override.
+        ${CMAKE_SOURCE_DIR}/3rdparty/libfork/include
 )
 
 add_library(${CMAKE_PROJECT_NAME} SHARED

MobileGL/Config.h

@@ -66,6 +66,12 @@ namespace MobileGL::MG_Config {
     //   - DISPLAY: X11 session variable, not MobileGL configuration.
     //   - MOBILEGL_LOG_FILE_PATH: log-file init runs before MG_ConfigLoader::Init
     //     (see MG_Util/Debug/Log.cpp).
+    //   - MOBILEGL_ASYNC_POOL: a ShaderCompilePool is constructed by binaries that never call
+    //     MobileGL::Initialize() and so never run MG_ConfigLoader::Init - MG_Test's
+    //     JobNodeTest builds pools directly, and it is the suite that runs the whole async
+    //     matrix against both execution engines. Mirroring it here would resolve to the
+    //     default in exactly the tests that exist to tell the engines apart (see
+    //     MG_Util/Async/ShaderCompilePool.cpp, DetectAsyncPoolEngine).
     struct FeaturesTable {
         // MOBILEGL_DISABLE_TIMERQUERY: do not advertise or use GPU timer queries.
         Bool DisableTimerQuery = false;

MobileGL/MG_Test/Util/AsyncPoolBench.cpp

@@ -0,0 +1,368 @@
+// MobileGL - MobileGL/MG_Test/Util/AsyncPoolBench.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+// A head-to-head harness for the two ShaderCompilePool execution engines
+// (MOBILEGL_ASYNC_POOL=asio|libfork). Not a gtest: it measures one wall-clock interval per
+// process, because most of what it drives is memoized per process (the shader preprocess
+// cache and the compile-adoption map both live for the life of the GL context), so a second
+// timed repetition inside one process would measure the cache, not the compiler. The driver
+// script re-executes the binary for every repetition instead.
+//
+// Two modes:
+//
+//   corpus - the REAL frontend path. glCreateShader/glShaderSource are done untimed, then
+//            the clock starts and glCompileShader/glLinkProgram submit every job, and stops
+//            once glGetProgramiv(GL_LINK_STATUS) has joined all of them. That is exactly the
+//            first-submit-to-all-joined interval a shaderpack load pays.
+//
+//   micro  - N trivial JobNodes straight through ShaderCompilePool::Post, isolating the
+//            executor's own dispatch overhead from any workload contention.
+
+#include <algorithm>
+#include <atomic>
+#include <chrono>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <filesystem>
+#include <fstream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "Includes.h"
+#include "Init.h"
+#include <Config.h>
+
+#include <MG_Impl/GLImpl/Program/GL_Program.h>
+#include <MG_Util/Async/JobNode.h>
+#include <MG_Util/Async/ShaderCompilePool.h>
+
+using namespace MobileGL;
+using namespace MobileGL::MG_Util::Async;
+namespace GLImpl = MobileGL::MG_Impl::GLImpl;
+namespace fs = std::filesystem;
+
+namespace {
+    using Clock = std::chrono::steady_clock;
+
+    double MillisSince(const Clock::time_point start) {
+        return std::chrono::duration<double, std::milli>(Clock::now() - start).count();
+    }
+
+    GLenum StageFromExtension(const std::string& ext) {
+        if (ext == ".vert") return GL_VERTEX_SHADER;
+        if (ext == ".frag") return GL_FRAGMENT_SHADER;
+        if (ext == ".geom") return GL_GEOMETRY_SHADER;
+        if (ext == ".comp") return GL_COMPUTE_SHADER;
+        if (ext == ".tesc") return GL_TESS_CONTROL_SHADER;
+        if (ext == ".tese") return GL_TESS_EVALUATION_SHADER;
+        return 0;
+    }
+
+    std::string ReadFile(const fs::path& path) {
+        std::ifstream in(path, std::ios::binary);
+        std::ostringstream buf;
+        buf << in.rdbuf();
+        return buf.str();
+    }
+
+    struct CorpusShader {
+        std::string name;
+        std::string source;
+        GLenum stage = 0;
+    };
+
+    // One program's worth of the corpus: the trace's link group. Shaders are indices into
+    // the flat shader list, because a source shared by several programs must stay ONE entry
+    // - that sharing is what the compile-adoption map sees in the real path too.
+    struct CorpusProgram {
+        std::vector<SizeT> shaders;
+    };
+
+    struct Corpus {
+        std::vector<CorpusShader> shaders;
+        std::vector<CorpusProgram> programs;
+        SizeT totalBytes = 0;
+    };
+
+    // Reads a corpus directory written by extract_corpus.py: one file per compiled shader,
+    // stage in the extension, plus manifest.txt naming the trace's link groups.
+    Corpus LoadCorpus(const fs::path& dir) {
+        Corpus corpus;
+        std::unordered_map<std::string, SizeT> byName;
+
+        const auto intern = [&](const std::string& name) -> SizeT {
+            if (const auto it = byName.find(name); it != byName.end()) return it->second;
+            const fs::path path = dir / name;
+            if (!fs::exists(path)) return static_cast<SizeT>(-1);
+            CorpusShader shader;
+            shader.name = name;
+            shader.source = ReadFile(path);
+            shader.stage = StageFromExtension(path.extension().string());
+            if (shader.stage == 0) return static_cast<SizeT>(-1);
+            corpus.totalBytes += shader.source.size();
+            corpus.shaders.push_back(Move(shader));
+            const SizeT index = corpus.shaders.size() - 1;
+            byName.emplace(name, index);
+            return index;
+        };
+
+        const fs::path manifest = dir / "manifest.txt";
+        if (fs::exists(manifest)) {
+            std::ifstream in(manifest);
+            std::string line;
+            while (std::getline(in, line)) {
+                if (line.empty() || line[0] == '#') continue;
+                CorpusProgram program;
+                std::istringstream fields(line);
+                std::string name;
+                while (fields >> name) {
+                    const SizeT index = intern(name);
+                    if (index != static_cast<SizeT>(-1)) program.shaders.push_back(index);
+                }
+                if (!program.shaders.empty()) corpus.programs.push_back(Move(program));
+            }
+        }
+
+        // Anything in the directory the manifest never linked still gets compiled, as a
+        // program-less group, so the corpus on disk and the corpus measured are the same set.
+        std::vector<fs::path> leftovers;
+        for (const auto& entry : fs::directory_iterator(dir)) {
+            if (!entry.is_regular_file()) continue;
+            const std::string name = entry.path().filename().string();
+            if (name == "manifest.txt") continue;
+            if (StageFromExtension(entry.path().extension().string()) == 0) continue;
+            if (byName.count(name) != 0) continue;
+            leftovers.push_back(entry.path());
+        }
+        std::sort(leftovers.begin(), leftovers.end());
+        for (const auto& path : leftovers) intern(path.filename().string());
+
+        return corpus;
+    }
+
+    struct CorpusResult {
+        double submitMs = 0;  // first glCompileShader -> last glLinkProgram returned
+        double joinMs = 0;    // last submit -> every program joined
+        double totalMs = 0;   // the number that matters: first submit -> all joined
+        SizeT linkFailures = 0;
+        SizeT compileFailures = 0;
+    };
+
+    CorpusResult RunCorpus(const Corpus& corpus) {
+        // ---- Untimed: create every GL object and stage every source ----------------------
+        // glShaderSource is a memcpy into the shader object and glAttachShader is a pointer
+        // append; neither touches the pool. Keeping them outside the clock makes the measured
+        // interval exactly the compile+link critical path, which is what an application's
+        // loading screen waits on.
+        std::vector<GLuint> shaderNames(corpus.shaders.size(), 0);
+        for (SizeT i = 0; i < corpus.shaders.size(); ++i) {
+            const CorpusShader& shader = corpus.shaders[i];
+            const GLuint name = GLImpl::CreateShader(shader.stage);
+            const GLchar* text = shader.source.c_str();
+            const GLint length = static_cast<GLint>(shader.source.size());
+            GLImpl::ShaderSource(name, 1, &text, &length);
+            shaderNames[i] = name;
+        }
+
+        std::vector<GLuint> programNames(corpus.programs.size(), 0);
+        for (SizeT p = 0; p < corpus.programs.size(); ++p) {
+            const GLuint program = GLImpl::CreateProgram();
+            for (const SizeT shaderIndex : corpus.programs[p].shaders) {
+                GLImpl::AttachShader(program, shaderNames[shaderIndex]);
+            }
+            programNames[p] = program;
+        }
+
+        // ---- Timed ------------------------------------------------------------------------
+        const Clock::time_point start = Clock::now();
+
+        // Submission order follows the trace: a program's shaders, then its link. That order
+        // is what exercises ProgramLinkTask::SubmitAfter's dependency chaining rather than a
+        // flat burst of independent compiles.
+        std::vector<Bool> submitted(corpus.shaders.size(), false);
+        for (SizeT p = 0; p < corpus.programs.size(); ++p) {
+            for (const SizeT shaderIndex : corpus.programs[p].shaders) {
+                if (submitted[shaderIndex]) continue;
+                submitted[shaderIndex] = true;
+                GLImpl::CompileShader(shaderNames[shaderIndex]);
+            }
+            GLImpl::LinkProgram(programNames[p]);
+        }
+        for (SizeT i = 0; i < corpus.shaders.size(); ++i) {
+            if (submitted[i]) continue;
+            submitted[i] = true;
+            GLImpl::CompileShader(shaderNames[i]);
+        }
+
+        const Clock::time_point submitted_at = Clock::now();
+
+        CorpusResult result;
+        // GL_LINK_STATUS is a joining query (GL_COMPLETION_STATUS_KHR is the one that must
+        // not join), so this loop is the all-joined barrier.
+        for (const GLuint program : programNames) {
+            GLint status = 0;
+            GLImpl::GetProgramiv(program, GL_LINK_STATUS, &status);
+            if (status == GL_FALSE) ++result.linkFailures;
+        }
+        for (const GLuint shader : shaderNames) {
+            GLint status = 0;
+            GLImpl::GetShaderiv(shader, GL_COMPILE_STATUS, &status);
+            if (status == GL_FALSE) ++result.compileFailures;
+        }
+
+        result.totalMs = MillisSince(start);
+        result.submitMs = std::chrono::duration<double, std::milli>(submitted_at - start).count();
+        result.joinMs = result.totalMs - result.submitMs;
+
+        for (const GLuint program : programNames) GLImpl::DeleteProgram(program);
+        for (const GLuint shader : shaderNames) GLImpl::DeleteShader(shader);
+        return result;
+    }
+
+    // ---- Executor microbenchmark ----------------------------------------------------------
+    // The body is deliberately near-empty: what is being measured is Post -> engine ->
+    // RunOnWorker -> next dispatch, i.e. the executor's own cost per job, with no compiler
+    // work to hide it.
+    //
+    // The barrier is an all-jobs-ran latch, and it has to be. This bench used to stop the
+    // clock at StopAndDrain(), which is not a "wait for everything" - it is the teardown path,
+    // and its contract is to ABANDON whatever the budget has not dispatched yet (see
+    // ShaderCompilePool::StopAndDrain, and the JobNodeTest case that pins exactly that). With
+    // 100k jobs behind a budget of N, most of them were therefore cancelled rather than run,
+    // and the fraction that survived was decided by how fast the engine drained the queue
+    // relative to the posting loop - i.e. by the very quantity under test. Measured on this
+    // machine at 8 workers: Asio ran 75,906 of 100,000 and libfork 99,998, and both were
+    // scored as if they had run 100,000. The reported "libfork is 1.36x faster" was libfork
+    // being charged for 32% more work than Asio.
+    class TrivialJob final : public JobNode {
+    public:
+        TrivialJob(std::atomic<Uint64>* sink, const Uint64 total, std::mutex* mutex,
+                   std::condition_variable* cv)
+            : m_sink(sink), m_total(total), m_mutex(mutex), m_cv(cv) {}
+
+    private:
+        void RunBody() override {
+            if (m_sink->fetch_add(1, std::memory_order_acq_rel) + 1 == m_total) {
+                // The last job wakes the timer. Under the lock, so the waiter cannot miss it
+                // between its predicate check and its wait.
+                const std::lock_guard<std::mutex> lock(*m_mutex);
+                m_cv->notify_all();
+            }
+        }
+
+        std::atomic<Uint64>* m_sink;
+        Uint64 m_total;
+        std::mutex* m_mutex;
+        std::condition_variable* m_cv;
+    };
+
+    struct MicroResult {
+        double ms = 0;
+        Uint64 ran = 0;
+    };
+
+    MicroResult RunMicrobench(const Uint threads, const SizeT jobs) {
+        ShaderCompilePool pool(threads);
+        std::atomic<Uint64> counter{0};
+        std::mutex mutex;
+        std::condition_variable cv;
+        const auto total = static_cast<Uint64>(jobs);
+
+        // Nodes are allocated up front: MakeShared is not what is under test, and leaving it
+        // inside the loop would put an allocator on the critical path in front of the
+        // dispatch path this is meant to isolate.
+        std::vector<SharedPtr<JobNode>> nodes;
+        nodes.reserve(jobs);
+        for (SizeT i = 0; i < jobs; ++i) {
+            nodes.push_back(MakeShared<TrivialJob>(&counter, total, &mutex, &cv));
+        }
+
+        const Clock::time_point start = Clock::now();
+        for (auto& node : nodes) pool.Post(Move(node));
+        {
+            std::unique_lock<std::mutex> lock(mutex);
+            cv.wait(lock, [&] { return counter.load(std::memory_order_acquire) >= total; });
+        }
+        const double ms = MillisSince(start);
+
+        MicroResult result;
+        result.ms = ms;
+        result.ran = counter.load(std::memory_order_acquire);
+        return result;
+    }
+
+    [[noreturn]] void Usage() {
+        std::fprintf(stderr,
+                     "usage: AsyncPoolBench --corpus DIR\n"
+                     "       AsyncPoolBench --micro JOBS --threads N\n"
+                     "env: MOBILEGL_ASYNC_POOL=asio|libfork, "
+                     "MOBILEGL_ASYNC_SHADER_COMPILE_THREADS=N\n");
+        std::exit(2);
+    }
+} // namespace
+
+int main(int argc, char** argv) {
+    std::string corpusDir;
+    SizeT microJobs = 0;
+    Uint microThreads = 0;
+
+    for (int i = 1; i < argc; ++i) {
+        const std::string arg = argv[i];
+        const auto next = [&]() -> std::string {
+            if (i + 1 >= argc) Usage();
+            return argv[++i];
+        };
+        if (arg == "--corpus") corpusDir = next();
+        else if (arg == "--micro") microJobs = static_cast<SizeT>(std::stoull(next()));
+        else if (arg == "--threads") microThreads = static_cast<Uint>(std::stoul(next()));
+        else Usage();
+    }
+    if (corpusDir.empty() && microJobs == 0) Usage();
+
+    Initialize();
+
+    const AsyncPoolEngine engine = DetectAsyncPoolEngine();
+    const char* engineName = AsyncPoolEngineName(engine);
+
+    if (microJobs != 0) {
+        const Uint threads = microThreads != 0 ? microThreads : DetectShaderCompileThreadCount();
+        const MicroResult result = RunMicrobench(threads, microJobs);
+        // `ran` is printed, not just checked, so that a run in which the arms did different
+        // amounts of work is visible in the results file rather than on a stderr the driver
+        // script redirects to /dev/null. ns_per_job divides by what actually ran.
+        std::printf("RESULT mode=micro engine=%s threads=%u jobs=%zu ran=%llu total_ms=%.3f "
+                    "ns_per_job=%.1f\n",
+                    engineName, threads, microJobs, static_cast<unsigned long long>(result.ran),
+                    result.ms, result.ms * 1e6 / static_cast<double>(result.ran));
+        return result.ran == microJobs ? 0 : 1;
+    }
+
+    const Corpus corpus = LoadCorpus(corpusDir);
+    if (corpus.shaders.empty()) {
+        std::fprintf(stderr, "AsyncPoolBench: no shaders found in %s\n", corpusDir.c_str());
+        return 1;
+    }
+
+    if (!AsyncShaderCompileActive()) {
+        std::fprintf(stderr, "AsyncPoolBench: asynchronous compilation is OFF; measuring the "
+                             "inline path\n");
+    }
+
+    const CorpusResult result = RunCorpus(corpus);
+    const Uint threads = ShaderCompilePool::Get().GetThreadCount();
+
+    std::printf("RESULT mode=corpus engine=%s threads=%u corpus=%s shaders=%zu programs=%zu "
+                "bytes=%zu total_ms=%.3f submit_ms=%.3f join_ms=%.3f link_fail=%zu "
+                "compile_fail=%zu\n",
+                engineName, threads, corpusDir.c_str(), corpus.shaders.size(),
+                corpus.programs.size(), corpus.totalBytes, result.totalMs, result.submitMs,
+                result.joinMs, result.linkFailures, result.compileFailures);
+    return 0;
+}

MobileGL/MG_Test/Util/CMakeLists.txt

@@ -10,11 +10,35 @@ target_include_directories(JobNodeTest PRIVATE
         ${MGL_ROOT}/MobileGL
 )
 
+# GTest::gtest, not GTest::gtest_main: JobNodeTest supplies its own main so that
+# MOBILEGL_LOG_FILE_PATH is set before the first log write in the process. The engine
+# -selection cases read the log back to assert that an unrecognized MOBILEGL_ASYNC_POOL value
+# warns, and the desktop log sink is the file (MOBILEGL_LOG_ENABLE_CONSOLE is 0).
 target_link_libraries(
         JobNodeTest PRIVATE
-        GTest::gtest_main
+        GTest::gtest
         ${LINK_LIBRARIES}
 )
 
 include(GoogleTest)
 gtest_discover_tests(JobNodeTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)
+
+# The engine comparison harness. Deliberately NOT registered with add_test: it measures wall
+# time, so it has no pass/fail verdict to give CI, and it is driven by a script that varies
+# MOBILEGL_ASYNC_POOL and MOBILEGL_ASYNC_SHADER_COMPILE_THREADS across a matrix. It lives
+# beside JobNodeTest because it drives the same pool through the same two engines; it links
+# MobileGL_s for the real glCompileShader/glLinkProgram frontend path.
+add_executable(
+        AsyncPoolBench
+        AsyncPoolBench.cpp
+)
+
+target_include_directories(AsyncPoolBench PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+)
+
+target_link_libraries(
+        AsyncPoolBench PRIVATE
+        ${LINK_LIBRARIES}
+)

MobileGL/MG_Test/Util/JobNodeTest.cpp

@@ -9,8 +9,19 @@
 #include <gtest/gtest.h>
 
 #include <chrono>
+#include <cstdlib>
+#include <filesystem>
+#include <fstream>
 #include <stdexcept>
 
+#ifdef _WIN32
+#include <process.h>
+#define MGL_TEST_GETPID _getpid
+#else
+#include <unistd.h>
+#define MGL_TEST_GETPID getpid
+#endif
+
 #include "Includes.h"
 #include <Config.h>
 
@@ -21,6 +32,29 @@ using namespace MobileGL;
 using namespace MobileGL::MG_Util::Async;
 
 namespace {
+    // Where this binary's MobileGL log lands, set by main() below. The engine-selection cases
+    // read it back: MobileGL's desktop log sink is the FILE, not the console
+    // (MOBILEGL_LOG_ENABLE_CONSOLE is 0 in Defines.h), so gtest's stdout capture would see
+    // nothing, and "unrecognized value warns" is a contract worth pinning rather than
+    // assuming - a silent fallback makes a misspelt engine name look exactly like an unset
+    // variable.
+    String g_logFilePath;
+
+    // Log.cpp flushes the file after every line, so everything written before this call is
+    // already visible.
+    String ReadLogFrom(const std::streamoff offset) {
+        std::ifstream file(g_logFilePath, std::ios::binary);
+        if (!file) return {};
+        file.seekg(offset);
+        return String((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+    }
+
+    std::streamoff LogSize() {
+        std::error_code error;
+        const auto size = std::filesystem::file_size(g_logFilePath, error);
+        return error ? 0 : static_cast<std::streamoff>(size);
+    }
+
     // Every test drives its own pool instance rather than ShaderCompilePool::Get(): the
     // process-wide pool is stopped permanently by StopAndDrain (that is the teardown
     // contract), so a test that drained the singleton would poison every test after it.
@@ -73,6 +107,21 @@ namespace {
         Bool m_open = false;
     };
 
+    // Live thread count of this process. Linux only - /proc/self/task has one entry per
+    // thread - and 0 where that is not available, which is how the one case that uses it
+    // decides to skip rather than to assert something it cannot see.
+    SizeT LiveThreadCount() {
+#ifdef __linux__
+        std::error_code error;
+        const auto count = static_cast<SizeT>(
+            std::distance(std::filesystem::directory_iterator("/proc/self/task", error),
+                          std::filesystem::directory_iterator()));
+        return error ? 0 : count;
+#else
+        return 0;
+#endif
+    }
+
     Bool WaitUntil(const std::function<Bool()>& predicate,
                    const std::chrono::milliseconds timeout = std::chrono::seconds(10)) {
         const auto deadline = std::chrono::steady_clock::now() + timeout;
@@ -89,11 +138,33 @@ namespace {
 // ---------------------------------------------------------------------------------------
 
 TEST(ShaderCompilePoolLifecycle, ConstructingAPoolStartsNoThreadUntilSomethingIsPosted) {
+    const SizeT before = LiveThreadCount();
+
     ShaderCompilePool pool(kTestThreads);
     EXPECT_EQ(pool.GetThreadCount(), kTestThreads);
     EXPECT_EQ(pool.GetMaxConcurrency(), kTestThreads);
-    // Nothing observable to assert about thread creation from here; what this pins is that
+
-    // construction is side-effect free and the pool destructs cleanly without ever running.
+    if (before == 0) {
+        // No thread census on this platform. The rest still holds: construction is
+        // side-effect free and the pool destructs cleanly without ever having run.
+        SUCCEED();
+        return;
+    }
+
+    // "A build that never posts pays nothing" is a real requirement, not a stylistic one -
+    // asynchronous compilation can be switched off entirely, and a switched-off pool that
+    // still spawned its workers would cost every such process its threads and their stacks.
+    // Worth asserting rather than asserting-by-comment now that an engine's thread shape is
+    // selectable: the libfork engine starts its workers AND a dispatch thread of its own, so
+    // a regression here would cost more than it used to.
+    EXPECT_EQ(LiveThreadCount(), before) << "constructing a pool started " << (LiveThreadCount() - before)
+                                         << " thread(s) before anything was posted";
+
+    auto job = MakeShared<TestJob>();
+    pool.Post(job);
+    job->Wait();
+    EXPECT_GT(LiveThreadCount(), before) << "the first Post started no thread at all, so the engine did not "
+                                            "really run the job off the calling thread";
 }
 
 TEST(ShaderCompilePoolLifecycle, StopAndDrainIsIdempotentAndSafeOnAnUnusedPool) {
@@ -217,6 +288,94 @@ TEST(JobNodeSubmit, ManyJobsAllComplete) {
     }
 }
 
+TEST(JobNodeSubmit, AJobBodyMayPostAnotherJobToTheSamePool) {
+    // The ProgramLinkTask::SubmitAfter shape, reduced to its scheduling core: the dependent is
+    // posted by whichever thread drove the dependency terminal, which for a job that finished
+    // on a worker is that WORKER. Every engine therefore has to accept a submission from
+    // inside its own pool.
+    //
+    // Not a hypothetical: libfork refuses this outright at its normal entry point
+    // (lf::schedule throws lf::schedule_in_worker, because a libfork worker may never block),
+    // which is why the libfork engine owns a dispatch thread of its own. Without this case a
+    // naive port passes every other test in the file and turns every dependency-released link
+    // job into a cancelled one on the real GL path.
+    ShaderCompilePool pool(kTestThreads);
+
+    std::atomic<Bool> innerSawPoolThread{false};
+    auto inner = MakeShared<TestJob>(
+        [&](TestJob&) { innerSawPoolThread.store(ShaderCompilePool::IsPoolThread(), std::memory_order_release); });
+
+    std::atomic<Bool> postedFromPoolThread{false};
+    auto outer = MakeShared<TestJob>([&](TestJob&) {
+        postedFromPoolThread.store(ShaderCompilePool::IsPoolThread(), std::memory_order_release);
+        pool.Post(inner);
+    });
+
+    pool.Post(outer);
+    outer->Wait();
+    inner->Wait();
+
+    EXPECT_TRUE(postedFromPoolThread.load()) << "the outer body did not run on a pool thread, so this case "
+                                                "did not exercise posting from inside the pool";
+    EXPECT_TRUE(outer->IsComplete());
+    // The load-bearing one: the inner job RAN. A dispatch the engine refused would have
+    // settled it Cancelled instead, and its body would never have executed.
+    EXPECT_TRUE(inner->IsComplete()) << "a job posted from a pool thread was not dispatched";
+    EXPECT_FALSE(inner->IsCancelled());
+    EXPECT_EQ(inner->ran.load(), 1u);
+    EXPECT_TRUE(innerSawPoolThread.load());
+}
+
+TEST(JobNodeSubmit, ABurstPostedFromInsideThePoolStillRunsInParallel) {
+    // The tail of a pack load: one compile job goes terminal and its continuations release
+    // several programs at once (ShaderCompileAdoptionMap lets one compile settle many), so a
+    // WORKER posts a burst into a pool that is otherwise idle. Every one of those posts clears
+    // the budget immediately, so the engine is handed `kBurst` runnable jobs from inside
+    // itself - and it has to spread them, not run them one behind another on the thread that
+    // submitted them.
+    //
+    // Asserting on peak concurrency rather than on wall time: the budget is the contract, and
+    // an engine that dispatches within the budget but executes serially has silently turned
+    // the budget into an upper bound nothing reaches.
+    constexpr Uint kBurst = 4; // == kTestThreads, so the budget can hold all of them at once
+    ShaderCompilePool pool(kTestThreads);
+
+    std::atomic<Uint> live{0};
+    std::atomic<Uint> peak{0};
+    std::atomic<Uint> finished{0};
+
+    Vector<SharedPtr<TestJob>> burst;
+    burst.reserve(kBurst);
+    for (Uint i = 0; i < kBurst; ++i) {
+        burst.push_back(MakeShared<TestJob>([&](TestJob&) {
+            const Uint now = live.fetch_add(1, std::memory_order_acq_rel) + 1;
+            Uint seen = peak.load(std::memory_order_acquire);
+            while (now > seen && !peak.compare_exchange_weak(seen, now, std::memory_order_acq_rel)) {
+            }
+            // Long enough that a serial engine cannot fake overlap, short enough to keep the
+            // case cheap: with any real spread every body is inside this window together.
+            std::this_thread::sleep_for(std::chrono::milliseconds(120));
+            live.fetch_sub(1, std::memory_order_acq_rel);
+            finished.fetch_add(1, std::memory_order_acq_rel);
+        }));
+    }
+
+    std::atomic<Bool> postedFromPoolThread{false};
+    auto seeder = MakeShared<TestJob>([&](TestJob&) {
+        postedFromPoolThread.store(ShaderCompilePool::IsPoolThread(), std::memory_order_release);
+        for (const auto& job : burst) pool.Post(job);
+    });
+
+    pool.Post(seeder);
+    seeder->Wait();
+    for (const auto& job : burst) job->Wait();
+
+    ASSERT_TRUE(postedFromPoolThread.load()) << "the burst was not posted from a pool thread";
+    EXPECT_EQ(finished.load(), kBurst);
+    EXPECT_GT(peak.load(), 1u) << "a burst posted from inside the pool ran strictly one at a time; the "
+                                  "engine serialized work the budget had already cleared";
+}
+
 TEST(JobNodeSubmit, ConcurrencyBudgetIsNeverExceeded) {
     constexpr Uint kBudget = 2;
     constexpr Uint kJobs = 64;
@@ -472,30 +631,58 @@ TEST(JobNodeException, AThrowingJobDoesNotPoisonTheWorkerForLaterJobs) {
 // ---------------------------------------------------------------------------------------
 
 TEST(ShaderCompilePoolDrain, StopAndDrainWithAThousandQueuedJobsLeavesNoneRunningOrPending) {
-    constexpr Uint kJobs = 1000;
+    constexpr Uint kQueued = 1000;
     ShaderCompilePool pool(kTestThreads);
-    pool.SetMaxConcurrency(1); // keep the vast majority queued behind the budget
+    pool.SetMaxConcurrency(1); // one slot, so everything behind the first job stays queued
 
-    Vector<SharedPtr<TestJob>> jobs;
+    // Pin that slot with a job that will not return until this test says so. Everything
-    jobs.reserve(kJobs);
+    // posted behind it is then PROVABLY still in the queue, which is what makes the counts
-    for (Uint i = 0; i < kJobs; ++i) {
+    // below exact.
-        jobs.push_back(MakeShared<TestJob>());
+    //
-        pool.Post(jobs.back());
+    // This case used to post a thousand trivial jobs and drain immediately, hoping the drain
+    // would beat the workers to some of them - and then assert only that "some" were
+    // cancelled. That hope does not survive an engine whose workers take their next job
+    // without a scheduler round trip: the libfork engine drained all thousand before the
+    // posting loop had finished, so the assertion failed about one run in fifty. The property
+    // being tested (a drain ABANDONS queued work rather than running it) is real and
+    // engine-independent; only the way it was provoked was a race.
+    Gate gate;
+    std::atomic<Bool> entered{false};
+    auto blocker = MakeShared<TestJob>([&](TestJob&) {
+        entered.store(true, std::memory_order_release);
+        gate.Wait();
+    });
+    pool.Post(blocker);
+    ASSERT_TRUE(WaitUntil([&] { return entered.load(); }));
+
+    Vector<SharedPtr<TestJob>> queued;
+    queued.reserve(kQueued);
+    for (Uint i = 0; i < kQueued; ++i) {
+        queued.push_back(MakeShared<TestJob>());
+        pool.Post(queued.back());
     }
+    for (const auto& job : queued) ASSERT_FALSE(job->IsTerminal());
 
-    pool.StopAndDrain();
+    std::thread drain([&] { pool.StopAndDrain(); });
+    // StopAndDrain settles the entire queue before it waits for the running body, so the
+    // first cancelled node proves it is past that point - and the gate can then be released
+    // without racing it.
+    ASSERT_TRUE(WaitUntil([&] { return queued.front()->IsTerminal(); }));
+    gate.Open();
+    drain.join();
 
-    // Every node is terminal, so nothing can be waiting on a worker that will never come.
+    // The job that was already running still finished: an in-flight body is waited for, not
-    Uint complete = 0;
+    // interrupted.
-    Uint cancelled = 0;
+    EXPECT_TRUE(blocker->IsComplete());
-    for (const auto& job : jobs) {
+    EXPECT_EQ(blocker->ran.load(), 1u);
+
+    // And every queued node is terminal, so nothing is left waiting on a worker that will
+    // never come - settled as cancelled, with its body never entered.
+    for (const auto& job : queued) {
         ASSERT_TRUE(job->IsTerminal());
-        if (job->IsComplete()) ++complete;
+        EXPECT_TRUE(job->IsCancelled());
-        if (job->IsCancelled()) ++cancelled;
+        EXPECT_EQ(job->ran.load(), 0u);
-        EXPECT_LE(job->ran.load(), 1u);
     }
-    EXPECT_EQ(complete + cancelled, kJobs);
-    EXPECT_GT(cancelled, 0u); // the drain really did abandon queued work rather than run it
 }
 
 TEST(ShaderCompilePoolDrain, StopAndDrainWaitsForARunningBodyToReturn) {
@@ -535,3 +722,141 @@ TEST(ShaderCompilePoolDrain, JobsPostedAfterADrainStillRun) {
     EXPECT_TRUE(job->IsComplete());
     EXPECT_EQ(job->ran.load(), 1u);
 }
+
+// ---------------------------------------------------------------------------------------
+// Execution engine selection (MOBILEGL_ASYNC_POOL)
+// ---------------------------------------------------------------------------------------
+//
+// The engine decides only HOW a job that the concurrency budget has already cleared reaches a
+// worker thread. Everything else in this file - the budget, cancel request-vs-outcome, the
+// continuation machinery, the inline fallback after a stop, the drain - is engine-independent
+// by construction, which is why the whole suite is expected to pass unchanged with
+// MOBILEGL_ASYNC_POOL unset and with it set to libfork. These cases pin the selection itself,
+// so that a run of the matrix cannot silently test asio twice.
+
+TEST(AsyncPoolEngineSelection, EveryAcceptedSpellingParsesToItsEngine) {
+    EXPECT_EQ(ParseAsyncPoolEngine("asio"), AsyncPoolEngine::Asio);
+    EXPECT_EQ(ParseAsyncPoolEngine("libfork"), AsyncPoolEngine::Libfork);
+    // Case-insensitive, like the other named-value variables (MOBILEGL_*_MULTIDRAW_MODE).
+    EXPECT_EQ(ParseAsyncPoolEngine("Libfork"), AsyncPoolEngine::Libfork);
+    EXPECT_EQ(ParseAsyncPoolEngine("LIBFORK"), AsyncPoolEngine::Libfork);
+    EXPECT_EQ(ParseAsyncPoolEngine("ASIO"), AsyncPoolEngine::Asio);
+
+    EXPECT_STREQ(AsyncPoolEngineName(AsyncPoolEngine::Asio), "asio");
+    EXPECT_STREQ(AsyncPoolEngineName(AsyncPoolEngine::Libfork), "libfork");
+    // Round trip: whatever the name prints is a spelling the variable accepts back.
+    EXPECT_EQ(ParseAsyncPoolEngine(AsyncPoolEngineName(AsyncPoolEngine::Asio)), AsyncPoolEngine::Asio);
+    EXPECT_EQ(ParseAsyncPoolEngine(AsyncPoolEngineName(AsyncPoolEngine::Libfork)), AsyncPoolEngine::Libfork);
+}
+
+TEST(AsyncPoolEngineSelection, EmptyAndAutoAreTheDefaultEngineAndSaySoSilently) {
+    // Unset resolves through the empty string, and "auto" is the spelling the other named
+    // -value variables accept for "no preference". Neither is a mistake, so neither warns.
+    const std::streamoff before = LogSize();
+    EXPECT_EQ(ParseAsyncPoolEngine(""), AsyncPoolEngine::Asio);
+    EXPECT_EQ(ParseAsyncPoolEngine("auto"), AsyncPoolEngine::Asio);
+    EXPECT_EQ(ReadLogFrom(before).find("MOBILEGL_ASYNC_POOL"), String::npos)
+        << "a legitimate value warned; only an unrecognized one may";
+}
+
+TEST(AsyncPoolEngineSelection, AnUnrecognizedEngineNameFallsBackToAsioAndWarns) {
+    const std::streamoff before = LogSize();
+    EXPECT_EQ(ParseAsyncPoolEngine("libfrok"), AsyncPoolEngine::Asio);
+
+// The warning is the other half of the contract: a misspelt engine name that fell back
+// silently would be indistinguishable from an unset variable, and a scaling measurement
+// taken against the wrong engine is worse than no measurement.
+//
+// Guarded because MGLOG_W is a compile-time no-op unless the build's log level admits it -
+// and the shipped level does not (Log.h orders the levels DEBUG=0, WARN=1, ERROR=2, INFO=3,
+// FATAL=4 and gates on `ACTIVE <= LEVEL`, so the default INFO build enables only INFO and
+// FATAL). Nothing is skipped: the fallback above is pinned in every build, and this half is
+// checked by a build configured with
+// -DMOBILEGL_LOG_ACTIVE_LEVEL=MOBILEGL_LOG_LEVEL_WARN. The same guard is what makes the
+// preceding "says so silently" case honest rather than vacuously true.
+#if MOBILEGL_LOG_ACTIVE_LEVEL <= MOBILEGL_LOG_LEVEL_WARN
+    const String logged = ReadLogFrom(before);
+    EXPECT_NE(logged.find("MOBILEGL_ASYNC_POOL"), String::npos) << "no warning names the variable; log tail: " << logged;
+    EXPECT_NE(logged.find("libfrok"), String::npos)
+        << "the warning does not quote the rejected value; log tail: " << logged;
+    EXPECT_NE(logged.find("asio"), String::npos)
+        << "the warning does not say what it fell back to; log tail: " << logged;
+#else
+    (void)before;
+#endif
+}
+
+TEST(AsyncPoolEngineSelection, TheDetectedEngineIsTheOneTheEnvironmentAskedFor) {
+    // Read the variable directly rather than through the pool, so this really compares the
+    // process's answer against the environment the runner exported. This is the case that
+    // makes "the suite passed with MOBILEGL_ASYNC_POOL=libfork" mean something.
+    const char* const raw = std::getenv("MOBILEGL_ASYNC_POOL");
+    const AsyncPoolEngine expected = ParseAsyncPoolEngine(raw != nullptr ? String(raw) : String());
+    EXPECT_EQ(DetectAsyncPoolEngine(), expected);
+
+    // Stable: resolved once per process, so it cannot drift between calls.
+    EXPECT_EQ(DetectAsyncPoolEngine(), DetectAsyncPoolEngine());
+
+    if (DetectAsyncPoolEngine() != AsyncPoolEngine::Asio) {
+        // Selecting a non-default engine announces itself at INFO, which the shipped log level
+        // does admit - so on the libfork half of the matrix this doubles as the positive
+        // control for the log plumbing the preceding two cases read: it proves
+        // MOBILEGL_LOG_FILE_PATH took effect and that ReadLogFrom really sees MobileGL's
+        // output, rather than passing because the file is always empty.
+        const String logged = ReadLogFrom(0);
+        EXPECT_NE(logged.find("MOBILEGL_ASYNC_POOL"), String::npos)
+            << "the selected engine was never announced, so this binary's log capture proves nothing";
+        EXPECT_NE(logged.find(AsyncPoolEngineName(DetectAsyncPoolEngine())), String::npos);
+    }
+}
+
+TEST(AsyncPoolEngineSelection, EveryPoolReportsTheProcessEngineAndRunsWorkOnIt) {
+    ShaderCompilePool first(kTestThreads);
+    ShaderCompilePool second(kTestThreads);
+    EXPECT_EQ(first.GetEngine(), DetectAsyncPoolEngine());
+    EXPECT_EQ(second.GetEngine(), first.GetEngine())
+        << "two pools in one process disagree about the engine; a process must never run both";
+
+    // And the engine it reports is the one that actually executed the work: the body ran off
+    // the calling thread, on a thread the pool owns.
+    const auto callingThread = std::this_thread::get_id();
+    std::atomic<Bool> sawPoolThread{false};
+    std::thread::id bodyThread{};
+    auto job = MakeShared<TestJob>([&](TestJob&) {
+        sawPoolThread.store(ShaderCompilePool::IsPoolThread(), std::memory_order_release);
+        bodyThread = std::this_thread::get_id();
+    });
+    first.Post(job);
+    job->Wait();
+
+    ASSERT_TRUE(job->IsComplete());
+    EXPECT_TRUE(sawPoolThread.load());
+    EXPECT_NE(bodyThread, callingThread);
+}
+
+// gtest_main is replaced here for one reason: the engine-selection cases above assert that an
+// unrecognized MOBILEGL_ASYNC_POOL value WARNS, and MobileGL's desktop log sink is the log
+// file - MOBILEGL_LOG_ENABLE_CONSOLE is 0 in Defines.h, so there is nothing on stdout to
+// capture. MOBILEGL_LOG_FILE_PATH is read by Log.cpp's InitFile() at the first log write in
+// the process, so it has to be set before any test body runs.
+int main(int argc, char** argv) {
+    const std::filesystem::path logPath =
+        std::filesystem::temp_directory_path() /
+        ("mobilegl-jobnodetest-" + std::to_string(static_cast<long long>(MGL_TEST_GETPID())) + ".log");
+    g_logFilePath = logPath.string();
+    std::filesystem::remove(logPath);
+#ifdef _WIN32
+    ::_putenv_s("MOBILEGL_LOG_FILE_PATH", g_logFilePath.c_str());
+#else
+    ::setenv("MOBILEGL_LOG_FILE_PATH", g_logFilePath.c_str(), 1);
+#endif
+
+    ::testing::InitGoogleTest(&argc, argv);
+    const int result = RUN_ALL_TESTS();
+
+    // Best-effort: leaving a log file per test process in the temp directory would be litter,
+    // and a failed run has already printed the tail it needed into the gtest output.
+    std::error_code ignored;
+    std::filesystem::remove(logPath, ignored);
+    return result;
+}

MobileGL/MG_Util/Async/ShaderCompilePool.cpp

@@ -12,8 +12,14 @@
 #include <asio/post.hpp>
 #include <asio/thread_pool.hpp>
 
+#include <libfork/core.hpp>
+#include <libfork/schedule/lazy_pool.hpp>
+
 #include <cstdio>
+#include <cstdlib>
 #include <deque>
+#include <functional>
+#include <span>
 
 namespace MobileGL::MG_Util::Async {
     namespace {
@@ -135,46 +141,506 @@ namespace MobileGL::MG_Util::Async {
         return std::clamp(DetectBigCoreCount(), 1u, kMaxAutoShaderCompileThreads);
     }
 
+    // ---- Engine selection -----------------------------------------------------------------
+
+    const char* AsyncPoolEngineName(const AsyncPoolEngine engine) {
+        switch (engine) {
+        case AsyncPoolEngine::Libfork: return "libfork";
+        case AsyncPoolEngine::Asio: break;
+        }
+        return "asio";
+    }
+
+    AsyncPoolEngine ParseAsyncPoolEngine(const String& value) {
+        String lowered = value;
+        std::transform(lowered.begin(), lowered.end(), lowered.begin(),
+                       [](const unsigned char c) { return static_cast<char>(std::tolower(c)); });
+        if (lowered == "libfork") return AsyncPoolEngine::Libfork;
+        if (lowered == "asio" || lowered == "auto" || lowered.empty()) return AsyncPoolEngine::Asio;
+        // Not silent: a misspelt engine name resolving to the default would be
+        // indistinguishable from not having set the variable at all, and the only reason to
+        // set it is to know which engine ran.
+        MGLOG_W("Config: Ignoring invalid env variable MOBILEGL_ASYNC_POOL='%s'; expected asio|libfork, "
+                "using asio",
+                value.c_str());
+        return AsyncPoolEngine::Asio;
+    }
+
+    AsyncPoolEngine DetectAsyncPoolEngine() {
+        // A live std::getenv rather than an MG_Config::Features mirror, and deliberately so:
+        // a ShaderCompilePool is constructed by binaries that never call MobileGL::Initialize()
+        // and therefore never run MG_ConfigLoader::Init() - MG_Test/Util/JobNodeTest builds
+        // pools directly, and it is the suite that exercises the engines against each other.
+        // Reading Features there would silently resolve to the default and the libfork half of
+        // the test matrix would prove nothing. See the exemption list in Config.h.
+        //
+        // Resolved once per process (a function-local static): every pool in a process gets
+        // the same engine, so a process can never end up running two.
+        static const AsyncPoolEngine engine = [] {
+            const char* value = std::getenv("MOBILEGL_ASYNC_POOL");
+            const AsyncPoolEngine resolved = ParseAsyncPoolEngine(value != nullptr ? String(value) : String());
+            if (resolved != AsyncPoolEngine::Asio) {
+                MGLOG_I("ShaderCompilePool: MOBILEGL_ASYNC_POOL selected the %s execution engine",
+                        AsyncPoolEngineName(resolved));
+            }
+            return resolved;
+        }();
+        return engine;
+    }
+
+    namespace {
+        // ---- The engine boundary ----------------------------------------------------------
+        // Submit() has exactly asio::post's contract, and ShaderCompilePool::Impl leans on all
+        // four halves of it:
+        //   * it NEVER runs `fn` on the calling thread. DispatchLocked calls it while holding
+        //     the pool's plain, non-recursive mutex, and a job body (or a terminal
+        //     continuation it releases) is free to call Post() again - an inline run would
+        //     deadlock on the lock this frame already owns.
+        //   * it is callable from ANY thread, a worker of this very pool included:
+        //     ProgramLinkTask::OnDepSettled posts the link job from whichever thread drove the
+        //     last compile terminal, which is a worker.
+        //   * it may throw, and when it does it must not have consumed the caller's job node,
+        //     so Post/DispatchLocked can settle the node instead of stranding it Pending with
+        //     a joiner blocked forever.
+        //   * once it has accepted `fn`, `fn` WILL run. A dropped callable is a node nothing
+        //     ever settles, so the engines run it themselves rather than discard it.
+        class JobExecutor {
+        public:
+            virtual ~JobExecutor() = default;
+            JobExecutor() = default;
+            JobExecutor(const JobExecutor&) = delete;
+            JobExecutor& operator=(const JobExecutor&) = delete;
+
+            virtual void Submit(std::function<void()> fn) = 0;
+
+            // Returns once every callable ever handed to Submit has finished running. The
+            // guarantee StopAndDrain sells to library teardown: after it returns, no worker is
+            // still inside a job body that could touch glslang's process globals.
+            virtual void JoinAll() = 0;
+        };
+
+        // ---- Engine 1: Asio (the shipped default) -----------------------------------------
+        class AsioJobExecutor final : public JobExecutor {
+        public:
+            explicit AsioJobExecutor(const Uint threads) : m_pool(threads) {}
+
+            // asio::post only enqueues; it never runs the handler on the calling thread, which
+            // is what makes calling it under the pool mutex safe.
+            void Submit(std::function<void()> fn) override { asio::post(m_pool, Move(fn)); }
+
+            void JoinAll() override { m_pool.join(); }
+
+        private:
+            asio::thread_pool m_pool;
+        };
+
+        // ---- Engine 2: libfork ------------------------------------------------------------
+        //
+        // libfork is a continuation-stealing fork-join runtime, and the shape that fits here is
+        // NOT fork-join: a job body is one coarse, blocking, non-forking unit (a glslang
+        // compile), and the concurrency budget that bounds peak RSS is Impl's, not the
+        // scheduler's. So libfork is used as a job executor - each dispatched job is a detached
+        // root task - and what it is being asked to beat is Asio's single scheduler queue with
+        // its per-worker work-stealing deques and sleeping workers.
+        //
+        // The one thing libfork forbids is the thing this pool does constantly: lf::schedule
+        // (which lf::detach is built on) THROWS lf::schedule_in_worker when the calling thread
+        // is a libfork worker, because workers may never block. Yet a worker submits on every
+        // job completion - RunOnWorker's tail refills the budget - and again whenever a
+        // terminal continuation posts (ProgramLinkTask::OnDepSettled). Routing those through a
+        // separate dispatch thread works but costs two thread wakeups per job, which measured
+        // 4x worse than Asio on short jobs. So instead a dispatched root is a CHAIN: when its
+        // body returns it takes the next queued job itself and runs it in the same coroutine
+        // on the same worker. The refill a worker submits is therefore absorbed by the very
+        // chain that submitted it - no scheduler round trip, no wakeup - and libfork is only
+        // entered for work that arrives from outside the pool.
+        //
+        // Absorption is bounded at one job per running chain, though, because a chain is one
+        // worker: past that bound the queue would be jobs the budget has already cleared,
+        // waiting behind each other on a single thread. See Submit.
+        //
+        // Why none of this can strand a job: the queue below is only ever added to from inside
+        // a running chain (tl_chainOwner == this), and a chain exits only when it finds the
+        // queue empty - unconditionally, whatever the bound says. Every other submitter goes
+        // to the dispatch thread or straight to lf::detach.
+        class LibforkJobExecutor;
+
+        // Which executor's chain, if any, is running on this thread. Deliberately narrower
+        // than ShaderCompilePool::IsPoolThread(): that flag is process-wide and latched
+        // forever, so a worker of a DIFFERENT pool would read as "mine" and queue a job into a
+        // chain that will never drain it. This says exactly "a chain of *this* executor is
+        // executing on this thread, and it will look at the queue again before it exits".
+        thread_local LibforkJobExecutor* tl_chainOwner = nullptr;
+
+        // One dispatched job, heap-owned. It reaches its coroutine as a POINTER passed BY
+        // VALUE: libfork forwards a root task's arguments into the coroutine frame, so a
+        // by-value pointer is copied into the frame, whereas anything passed by reference
+        // would dangle the moment lf::detach returns - and detach, unlike sync_wait, does not
+        // outlive the task.
+        struct LibforkJob {
+            std::function<void()> body;
+            LibforkJobExecutor* owner;
+        };
+
+        // A scheduler adaptor for lf::detach: it places external submissions round-robin over
+        // lf::lazy_pool's worker contexts instead of letting the pool pick one at random.
+        // Both reasons are load-bearing, and the second was worth 1.3x at a budget equal to
+        // the worker count - the configuration MobileGL actually ships, since maxConcurrency
+        // is clamped to the thread count:
+        //   * lf::lazy_pool::schedule chooses its victim with a
+        //     std::uniform_int_distribution over a lazy_pool-member xoshiro generator -
+        //     unsynchronized mutable state, so two concurrent submissions are a data race
+        //     inside libfork itself. An atomic cursor is not.
+        //   * A worker's SUBMISSION list is drained only by that worker
+        //     (worker_context::try_pop_all is documented "for use only by the owning worker
+        //     thread"); a thief takes from the task deque, which is a different queue. So a
+        //     job placed on a worker that is inside a long blocking body waits for that body
+        //     rather than being stolen - and random placement of `budget` submissions over
+        //     `budget` workers collides by the birthday rule. Round-robin lands the GL
+        //     thread's burst one per worker, which is exactly the intended shape.
+        struct RoundRobinSubmitter {
+            std::span<lf::worker_context*> contexts;
+            std::atomic<Uint64>* cursor;
+
+            void schedule(const lf::submit_handle job) const {
+                const Uint64 index = cursor->fetch_add(1, std::memory_order_relaxed);
+                contexts[static_cast<SizeT>(index % contexts.size())]->schedule(job);
+            }
+        };
+
+        void RunLibforkChain(LibforkJob* raw) noexcept;
+
+        // The root task every dispatched chain runs as. libfork async function objects are
+        // copyable, captureless callables returning lf::task<>, whose first parameter is the
+        // combinator's synthesized first argument (unused here: this task neither forks nor
+        // joins). The coroutine exists purely as libfork's entry protocol; the loop is in
+        // RunLibforkChain.
+        inline constexpr auto kLibforkChainTask = [](auto /*self*/, LibforkJob* job) -> lf::task<void> {
+            RunLibforkChain(job);
+            co_return;
+        };
+
+        class LibforkJobExecutor final : public JobExecutor {
+        public:
+            explicit LibforkJobExecutor(const Uint threads)
+                : m_pool(static_cast<std::size_t>(std::max(1u, threads))), m_contexts(m_pool.contexts()),
+                  m_fallback([this] { FallbackLoop(); }) {}
+
+            ~LibforkJobExecutor() override {
+                JoinAll();
+                {
+                    const std::lock_guard<std::mutex> lock(m_mutex);
+                    m_fallbackStop = true;
+                }
+                m_fallbackCv.notify_all();
+                if (m_fallback.joinable()) m_fallback.join();
+                // m_pool is destroyed last, and only here: lf::lazy_pool may not be destructed
+                // while any submitted task can still run or submit more. JoinAll() has
+                // established the first and the joined fallback thread the second. Its
+                // destructor then joins the worker threads, so a worker still unwinding a
+                // finished coroutine frame is waited for rather than pulled out from under.
+            }
+
+            void Submit(std::function<void()> fn) override {
+                if (tl_chainOwner == this) {
+                    const std::lock_guard<std::mutex> lock(m_mutex);
+                    // The hot path: ONE job per running chain. A chain picks up exactly one
+                    // queued job each time its body returns, so a queue no longer than the
+                    // number of live chains is a queue every entry of which has a distinct
+                    // worker waiting to take it - which is precisely the steady state this
+                    // absorption exists for (every worker finishes a job and refills its own
+                    // slot, all at once, with no scheduler round trip between them).
+                    //
+                    // Past that it is oversubscription, and absorbing it would be a
+                    // correctness-preserving way to destroy the pool's parallelism: the
+                    // budget would still say `maxConcurrency` jobs are in flight while one
+                    // worker ran them one behind another. That is not hypothetical - it is
+                    // the tail of a pack load, where one compile going terminal releases
+                    // several programs at once (ShaderCompileAdoptionMap lets a single
+                    // compile settle many) and the worker that drove it posts the whole
+                    // burst into an otherwise idle pool. Measured before this branch existed:
+                    // four such jobs took 4x one job's wall time on libfork and 1x on Asio.
+                    //
+                    // The overflow cannot go to lf::detach from here - a libfork worker may
+                    // not schedule - so it goes to the dispatch thread, which detaches it to
+                    // a worker of its own. That costs one thread wakeup; serializing costs a
+                    // whole compile.
+                    //
+                    // The count is taken AFTER the push, not before: deque::push_back is
+                    // strongly exception-safe, so an allocation failure here leaves `fn`
+                    // intact for DispatchLocked to settle - but a count incremented in front
+                    // of it would be a count nothing ever gives back, and JoinAll would wait
+                    // on it forever.
+                    const Bool takeable = m_chainQueue.size() < m_liveChains;
+                    if (takeable) {
+                        m_chainQueue.push_back(Move(fn));
+                        ++m_outstanding;
+                    } else {
+                        m_fallbackQueue.push_back(Move(fn));
+                        ++m_outstanding;
+                        m_fallbackCv.notify_one();
+                    }
+                    return;
+                }
+
+                {
+                    // Counted before anything can run it, so JoinAll cannot observe a zero
+                    // that this job would have broken.
+                    const std::lock_guard<std::mutex> lock(m_mutex);
+                    ++m_outstanding;
+                }
+                try {
+                    DetachChain(Move(fn));
+                } catch (const lf::schedule_in_worker&) {
+                    // Submitted from a libfork worker that is not running one of my chains -
+                    // a worker of another ShaderCompilePool. libfork will not take a
+                    // submission from there at all, and the queue above is not safe for it
+                    // (no chain of mine is running on that thread to drain it), so it goes to
+                    // the fallback thread, which is neither. DetachChain restored `fn` before
+                    // it threw.
+                    const std::lock_guard<std::mutex> lock(m_mutex);
+                    m_fallbackQueue.push_back(Move(fn));
+                    m_fallbackCv.notify_one();
+                } catch (...) {
+                    // Out of memory. Give the count back and let the caller settle its node:
+                    // that is Submit's contract and what DispatchLocked is written against.
+                    Retire();
+                    throw;
+                }
+            }
+
+            void JoinAll() override {
+                std::unique_lock<std::mutex> lock(m_mutex);
+                m_idleCv.wait(lock, [this] { return m_outstanding == 0; });
+            }
+
+            // A chain announces itself before it runs its first body, so that Submit's
+            // absorption rule can count the workers that are going to come back and ask for
+            // more. Under-counting is the only direction this can be wrong in (a detached
+            // chain is not counted until it starts), and under-counting only sends work to
+            // the dispatch thread that a chain could have taken - never the reverse.
+            void EnterChain() noexcept {
+                const std::lock_guard<std::mutex> lock(m_mutex);
+                ++m_liveChains;
+            }
+
+            // The end of one job in a chain. Returns true having loaded `body` with the next
+            // job to run on this same worker, false when there is nothing left - after which
+            // the caller must touch neither `this` nor anything owned by it, because the
+            // count this drops to zero may be the one JoinAll is waiting for.
+            //
+            // `body` must arrive empty: the finished job's captures (a strong reference to its
+            // JobNode) are released by the chain, outside this lock, so that no JobNode
+            // destructor ever runs inside the executor's critical section.
+            Bool RetireAndTakeNext(std::function<void()>& body) noexcept {
+                const std::lock_guard<std::mutex> lock(m_mutex);
+                --m_outstanding;
+                if (!m_chainQueue.empty()) {
+                    // Unconditional, and it has to stay that way: a chain that exited while
+                    // the queue was non-empty could be the last one, and the entry would then
+                    // be waiting on a worker that never comes. That is what makes the
+                    // absorption bound in Submit a scheduling policy rather than a liveness
+                    // requirement.
+                    //
+                    // swap, not move-assign: std::function's move assignment is not noexcept,
+                    // and this function is.
+                    body.swap(m_chainQueue.front());
+                    m_chainQueue.pop_front();
+                    return true; // the taken job's own count stays held
+                }
+                --m_liveChains;
+                // Notified while STILL HOLDING the lock, which is the whole reason this is not
+                // the usual notify-after-unlock. The wakeup this sends can be the one that
+                // lets JoinAll return and ~LibforkJobExecutor destroy m_idleCv - and a
+                // std::condition_variable may not be destroyed while another thread is inside
+                // notify_all() on it. Holding the lock across the notify means the waiter
+                // cannot re-acquire the mutex, and therefore cannot leave wait(), until this
+                // thread is out of both the notify and the unlock. ThreadSanitizer catches the
+                // other order immediately (pthread_cond_destroy vs pthread_cond_broadcast).
+                if (m_outstanding == 0) m_idleCv.notify_all();
+                return false;
+            }
+
+        private:
+            // Builds the root task and hands it to libfork. On any failure `fn` is restored,
+            // so the caller can still decide what to do with the job.
+            void DetachChain(std::function<void()>&& fn) {
+                // `new T{...}` allocates before it constructs, so a throwing operator new
+                // leaves `fn` untouched; the member move is std::function's noexcept one.
+                LibforkJob* job = new LibforkJob{Move(fn), this};
+                try {
+                    lf::detach(RoundRobinSubmitter{m_contexts, &m_cursor}, kLibforkChainTask, job);
+                } catch (...) {
+                    // lf::schedule upholds the strong exception guarantee, so nothing was
+                    // scheduled and the payload is still ours.
+                    const UniquePtr<LibforkJob> owned(job);
+                    fn = Move(owned->body);
+                    throw;
+                }
+            }
+
+            void Retire() noexcept {
+                // Under the lock, for the reason RetireAndTakeNext spells out.
+                const std::lock_guard<std::mutex> lock(m_mutex);
+                if (--m_outstanding == 0) m_idleCv.notify_all();
+            }
+
+            // The dispatch thread. It exists because lf::detach is illegal on a libfork worker
+            // and legal here, and it serves the two cases Submit cannot take itself: a
+            // submission from another pool's worker, and a chain's overflow past the
+            // one-job-per-chain bound. It sleeps otherwise, and it dispatches rather than
+            // executes - a body only ever runs here if libfork refuses the job outright.
+            void FallbackLoop() {
+                for (;;) {
+                    std::function<void()> fn;
+                    {
+                        std::unique_lock<std::mutex> lock(m_mutex);
+                        m_fallbackCv.wait(lock, [this] { return !m_fallbackQueue.empty() || m_fallbackStop; });
+                        // Emptiness is checked before the stop flag so that a stop can never
+                        // strand accepted work: an accepted job always runs, because the node
+                        // behind it has a joiner that would otherwise block forever.
+                        if (m_fallbackQueue.empty()) return;
+                        fn.swap(m_fallbackQueue.front());
+                        m_fallbackQueue.pop_front();
+                    }
+                    try {
+                        DetachChain(Move(fn));
+                    } catch (...) {
+                        MGLOG_E("ShaderCompilePool: libfork refused a fallback dispatch; running the job on "
+                                "the dispatch thread instead of dropping it");
+                        RunHere(Move(fn));
+                    }
+                }
+            }
+
+            // Last resort. Running the body here costs this engine its parallelism for one
+            // job; dropping it would cost a joiner its wakeup forever.
+            void RunHere(std::function<void()>&& fn) noexcept {
+                try {
+                    if (fn) fn();
+                } catch (...) {
+                    MGLOG_E("ShaderCompilePool: a job body escaped its own containment on the dispatch "
+                            "thread; it has been swallowed to keep the thread alive");
+                }
+                fn = nullptr;
+                Retire();
+            }
+
+            lf::lazy_pool m_pool;
+            // Fixed for the pool's lifetime, so it is read once rather than per submission.
+            std::span<lf::worker_context*> m_contexts;
+            std::atomic<Uint64> m_cursor{0};
+
+            std::mutex m_mutex;
+            std::condition_variable m_fallbackCv;
+            std::condition_variable m_idleCv;
+            // Refills and continuations submitted from inside a chain: drained by the chains.
+            std::deque<std::function<void()>> m_chainQueue;
+            // Chains currently executing, i.e. workers that will look at m_chainQueue again
+            // before they exit. The bound on how much Submit may absorb into a chain.
+            Uint m_liveChains = 0;
+            // Submissions from another pool's libfork worker, and the overflow of the rule
+            // above: drained by m_fallback, which detaches each one to a worker.
+            std::deque<std::function<void()>> m_fallbackQueue;
+            // Everything submitted and not yet finished, whichever queue it is in and whether
+            // or not it has reached a worker, so JoinAll needs a single predicate.
+            Uint m_outstanding = 0;
+            Bool m_fallbackStop = false;
+            std::thread m_fallback;
+        };
+
+        void RunLibforkChain(LibforkJob* const raw) noexcept {
+            UniquePtr<LibforkJob> job(raw);
+            LibforkJobExecutor* const owner = job->owner;
+            std::function<void()> body;
+            body.swap(job->body);
+            job.reset();
+
+            LibforkJobExecutor* const savedOwner = tl_chainOwner;
+            tl_chainOwner = owner;
+            owner->EnterChain();
+
+            for (;;) {
+                try {
+                    if (body) body();
+                } catch (...) {
+                    // JobNode::Run contains every body exception already; this is the backstop
+                    // for the wrapper itself. An exception escaping here would be stashed in
+                    // the root task's shared state, which lf::detach discards - i.e. silently
+                    // lost - and would abandon the rest of the chain.
+                    MGLOG_E("ShaderCompilePool: a job body escaped its own containment on a libfork worker; "
+                            "it has been swallowed to keep the chain alive");
+                }
+                // Release the finished job's captures (its strong JobNode reference) HERE,
+                // outside the executor's lock: a JobNode destructor is arbitrary code.
+                body = nullptr;
+                if (!owner->RetireAndTakeNext(body)) break;
+            }
+
+            // `owner` may already be destroyed - RetireAndTakeNext returning false can be the
+            // call that releases a JoinAll. Nothing below touches it.
+            tl_chainOwner = savedOwner;
+        }
+
+        UniquePtr<JobExecutor> MakeJobExecutor(const AsyncPoolEngine engine, const Uint threads) {
+            switch (engine) {
+            case AsyncPoolEngine::Libfork: return MakeUnique<LibforkJobExecutor>(threads);
+            case AsyncPoolEngine::Asio: break;
+            }
+            return MakeUnique<AsioJobExecutor>(threads);
+        }
+    } // namespace
+
     struct ShaderCompilePool::Impl {
-        explicit Impl(const Uint threads) : threadCount(std::max(1u, threads)), maxConcurrency(threadCount) {}
+        explicit Impl(const Uint threads)
+            : threadCount(std::max(1u, threads)), engine(DetectAsyncPoolEngine()), maxConcurrency(threadCount) {}
 
         const Uint threadCount;
+        // Latched at construction, not re-read: a pool may not change engines under its own
+        // workers, and GetEngine() is what the tests compare against the environment.
+        const AsyncPoolEngine engine;
 
         std::mutex mutex;
-        // Created on the first dispatched Post, never in the constructor: asio::thread_pool
+        // Created on the first dispatched Post, never in the constructor: both engines spawn
-        // spawns its threads eagerly, and a build with async off must not pay for threads it
+        // their threads eagerly (asio::thread_pool its workers, lf::lazy_pool its workers plus
+        // this file's dispatch thread), and a build with async off must not pay for threads it
         // will never use.
-        UniquePtr<asio::thread_pool> pool;
+        UniquePtr<JobExecutor> executor;
         std::deque<SharedPtr<JobNode>> queue;
         Uint inFlight = 0;
         Uint maxConcurrency;
         std::atomic<Bool> stopped{false};
 
-        // Callers hold `mutex`. Hands as many queued nodes to Asio as the concurrency budget
+        // Callers hold `mutex`. Hands as many queued nodes to the engine as the concurrency
-        // allows. Posting under the lock is safe and is what keeps `pool` from being moved
+        // budget allows. Submitting under the lock is safe and is what keeps `executor` from
-        // out by a concurrent StopAndDrain between the decision and the dispatch: asio::post
+        // being moved out by a concurrent StopAndDrain between the decision and the dispatch:
-        // only enqueues, it never runs the handler on the calling thread, so it cannot
+        // Submit only enqueues, it never runs the callable on the calling thread, so it cannot
         // re-enter this mutex.
         //
-        // A node asio::post fails to hand off is appended to `toCancel` instead of being
+        // A node the engine fails to accept is appended to `toCancel` instead of being
         // Cancel()'d here: Cancel() runs the node's OnTerminal continuations inline (stage 4
         // added ProgramLinkTask::OnDepSettled as a real one), and a continuation is free to
         // call ShaderCompilePool::Post() again. Every caller of DispatchLocked holds `mutex`
         // (a plain, non-recursive std::mutex) - Cancel()'ing in here would let that
         // re-entrant Post() deadlock on the very lock this frame already owns. The caller
         // drains `toCancel` after releasing the lock.
+        //
+        // The `stopped` check is also what keeps this loop from dereferencing a null
+        // `executor`: StopAndDrain sets the flag and moves the executor out in the same
+        // critical section, so a stopped pool never reaches the Submit below.
         void DispatchLocked(Vector<SharedPtr<JobNode>>& toCancel) {
             while (!queue.empty() && inFlight < maxConcurrency && !stopped.load(std::memory_order_acquire)) {
-                // Copy rather than move into the handler: if asio::post throws (it allocates)
+                // Copy rather than move into the callable: if Submit throws (both engines
-                // the local SharedPtr is still valid, so the node can be settled instead of
+                // allocate) the local SharedPtr is still valid, so the node can be settled
-                // being stranded Pending in a queue nothing will dispatch from again - a
+                // instead of being stranded Pending in a queue nothing will dispatch from
-                // joiner would block on it forever. Reclaiming the slot matters just as much:
+                // again - a joiner would block on it forever. Reclaiming the slot matters just
-                // a leaked `inFlight` shrinks the pool's concurrency budget permanently.
+                // as much: a leaked `inFlight` shrinks the pool's concurrency budget
+                // permanently.
                 SharedPtr<JobNode> node = queue.front();
                 queue.pop_front();
                 ++inFlight;
                 try {
-                    asio::post(*pool, [this, node]() mutable { RunOnWorker(Move(node)); });
+                    executor->Submit([this, node]() mutable { RunOnWorker(Move(node)); });
                 } catch (...) {
                     --inFlight;
                     toCancel.push_back(Move(node));
@@ -184,7 +650,7 @@ namespace MobileGL::MG_Util::Async {
 
         void RunOnWorker(SharedPtr<JobNode> node) {
             tl_isPoolThread = true;
-            // A node that was already handed to Asio when StopAndDrain ran still arrives
+            // A node that was already handed to the engine when StopAndDrain ran still arrives
             // here; cancelling it first turns the dispatch into a state transition instead of
             // a full compile, so the drain's join() returns promptly. This Cancel() runs
             // before `mutex` is ever taken in this frame, so it is not subject to the
@@ -234,13 +700,15 @@ namespace MobileGL::MG_Util::Async {
         return m_impl->maxConcurrency;
     }
 
+    AsyncPoolEngine ShaderCompilePool::GetEngine() const { return m_impl->engine; }
+
     void ShaderCompilePool::SetMaxConcurrency(const Uint n) {
         Vector<SharedPtr<JobNode>> toCancel;
         {
             const std::lock_guard<std::mutex> lock(m_impl->mutex);
             m_impl->maxConcurrency = std::clamp(n, 1u, m_impl->threadCount);
             // Raising the budget releases whatever the old one was holding back.
-            if (m_impl->pool) m_impl->DispatchLocked(toCancel);
+            if (m_impl->executor) m_impl->DispatchLocked(toCancel);
         }
         // Outside the lock: see DispatchLocked's comment.
         for (const auto& n2 : toCancel) {
@@ -252,23 +720,23 @@ namespace MobileGL::MG_Util::Async {
         if (!node) return;
         EnsureProcessTeardownSentinel();
 
-        // Enqueueing can throw: the thread_pool construction and asio::post both allocate,
+        // Enqueueing can throw: building the engine and submitting to it both allocate (and
-        // and under memory pressure a throw here would escape glCompileShader leaving the
+        // both spawn threads), and under memory pressure a throw here would escape
-        // node Pending with nothing left to dispatch it - the first observable read would
+        // glCompileShader leaving the node Pending with nothing left to dispatch it - the
-        // then block the GL thread forever. Settle the node instead: a cancelled node is a
+        // first observable read would then block the GL thread forever. Settle the node
-        // state every joiner already handles.
+        // instead: a cancelled node is a state every joiner already handles.
         //
-        // `node` is still valid in the catch for every throw this try can produce. The
+        // `node` is still valid in the catch for every throw this try can produce. The engine
-        // thread_pool construction runs before the move; deque::push_back is strongly
+        // construction runs before the move; deque::push_back is strongly exception-safe and
-        // exception-safe and SharedPtr's move constructor is noexcept, so a throwing
+        // SharedPtr's move constructor is noexcept, so a throwing push_back never consumed it;
-        // push_back never consumed it; and DispatchLocked contains its own asio::post
+        // and DispatchLocked contains its own Submit failures rather than propagating them
-        // failures rather than propagating them (see above). Keep it that way.
+        // (see above). Keep it that way.
         Bool enqueued = false;
         Vector<SharedPtr<JobNode>> toCancel;
         try {
             const std::lock_guard<std::mutex> lock(m_impl->mutex);
             if (!m_impl->stopped.load(std::memory_order_acquire) && !InProcessTeardown()) {
-                if (!m_impl->pool) m_impl->pool = MakeUnique<asio::thread_pool>(m_impl->threadCount);
+                if (!m_impl->executor) m_impl->executor = MakeJobExecutor(m_impl->engine, m_impl->threadCount);
                 m_impl->queue.push_back(Move(node));
                 m_impl->DispatchLocked(toCancel);
                 enqueued = true;
@@ -304,17 +772,17 @@ namespace MobileGL::MG_Util::Async {
     }
 
     void ShaderCompilePool::StopAndDrain() {
-        // asio::thread_pool::join() from a pool thread would deadlock on itself, and the
+        // Waiting for the workers from a worker would deadlock on itself (asio's join() says
-        // whole point of this call is that the GL thread waits for the workers.
+        // so outright), and the whole point of this call is that the GL thread waits.
         MOBILEGL_ASSERT(!IsPoolThread(), "ShaderCompilePool::StopAndDrain() called from a pool thread");
 
         std::deque<SharedPtr<JobNode>> abandoned;
-        UniquePtr<asio::thread_pool> pool;
+        UniquePtr<JobExecutor> executor;
         {
             const std::lock_guard<std::mutex> lock(m_impl->mutex);
             m_impl->stopped.store(true, std::memory_order_release);
             abandoned.swap(m_impl->queue);
-            pool = Move(m_impl->pool);
+            executor = Move(m_impl->executor);
         }
 
         // Queued but never dispatched: settle them so anything chained behind them is
@@ -323,9 +791,9 @@ namespace MobileGL::MG_Util::Async {
             if (node) node->Cancel();
         }
 
-        if (pool) {
+        if (executor) {
-            pool->join(); // returns once every handler already handed to Asio has finished
+            executor->JoinAll(); // returns once every job already handed to the engine is done
-            pool.reset();
+            executor.reset();    // and this stops the engine's threads
         }
 
         const std::lock_guard<std::mutex> lock(m_impl->mutex);

MobileGL/MG_Util/Async/ShaderCompilePool.h

@@ -11,11 +11,12 @@
 #include <MG_Util/Types.h>
 #include <MG_Util/Async/JobNode.h>
 
-// This header deliberately includes NO Asio header: asio::thread_pool lives behind the pimpl
+// This header deliberately includes NO Asio and NO libfork header: both execution engines
-// in ShaderCompilePool.cpp. Asio stays a private implementation detail of one translation
+// live behind the pimpl in ShaderCompilePool.cpp. They stay private implementation details of
-// unit, so no consumer target (MG_Test, MG_IntegrationTest, MG_Benchmark - each with its own
+// one translation unit, so no consumer target (MG_Test, MG_IntegrationTest, MG_Benchmark -
-// target_include_directories) needs the Asio include path, and no consumer pays its compile
+// each with its own target_include_directories) needs either include path, and no consumer
-// time. Do not add one here.
+// pays their compile time. libfork in particular is a C++20-coroutine header set whose
+// instantiation cost nothing outside the pool has any reason to carry. Do not add one here.
 
 namespace MobileGL::MG_Util::Async {
     // Stage 7: on by default. The gate behind the flip (2026-08-09, headless Mesa, both
@@ -66,6 +67,31 @@ namespace MobileGL::MG_Util::Async {
     // MOBILEGL_ASYNC_SHADER_COMPILE_THREADS overrides it outright.
     Uint DetectShaderCompileThreadCount();
 
+    // ---- MOBILEGL_ASYNC_POOL: which engine drives the worker threads ----------------------
+    // The engine is ONLY the execution engine. The job queue, the concurrency budget and its
+    // clamping, the suspension latch, cancel request-vs-outcome, the stopped-is-synchronous
+    // fallback and the drain are all engine-independent - they live in ShaderCompilePool::Impl
+    // and are shared verbatim by both engines, which is what lets the whole async suite run
+    // unchanged against either one. An engine answers exactly one question: how does a job
+    // that the budget has already cleared reach a worker thread?
+    enum class AsyncPoolEngine : Uint8 {
+        Asio,    // asio::thread_pool: one shared queue behind Asio's scheduler lock
+        Libfork, // lf::lazy_pool: per-worker work-stealing deques, workers sleep when idle
+    };
+
+    // "asio" / "libfork" - the spelling the environment variable accepts and the log prints.
+    const char* AsyncPoolEngineName(AsyncPoolEngine engine);
+
+    // Parses one MOBILEGL_ASYNC_POOL value. Case-insensitive; empty, "auto" and anything
+    // unrecognized resolve to Asio, and an unrecognized value warns (a misspelt engine name
+    // would otherwise be indistinguishable from the default, and the whole point of the
+    // variable is to know which engine ran).
+    AsyncPoolEngine ParseAsyncPoolEngine(const String& value);
+
+    // The process's engine, resolved from MOBILEGL_ASYNC_POOL on first call and cached. Every
+    // pool constructed afterwards reports the same answer, so a process never mixes engines.
+    AsyncPoolEngine DetectAsyncPoolEngine();
+
     class ShaderCompilePool {
     public:
         explicit ShaderCompilePool(Uint threadCount);
@@ -97,6 +123,11 @@ namespace MobileGL::MG_Util::Async {
         Uint GetThreadCount() const;
         Uint GetMaxConcurrency() const;
 
+        // The engine this pool was built with, latched at construction from
+        // DetectAsyncPoolEngine(). Reported rather than re-resolved so that a pool cannot
+        // change engines under its own workers.
+        AsyncPoolEngine GetEngine() const;
+
         // Bounded concurrency doubles as the memory bound, and is how
         // glMaxShaderCompilerThreadsKHR(n) is honoured: a 300-program pack load cannot put
         // 300 glslang arenas in flight at once. Clamped to [1, thread count].

MobileGL/MG_Util/SelfTest/DriverPost.cpp

@@ -151,12 +151,20 @@ namespace MobileGL::MG_Util::SelfTest {
                 return;
             }
             const Uint threads = MG_Util::Async::DetectShaderCompileThreadCount();
+            // The execution engine is named here too. It changes no observable GL behaviour -
+            // both engines run the same job queue under the same budget - but when a scaling
+            // or stall report comes back from a device, "which engine was this?" is the first
+            // question, and a POST page is the one artefact that always accompanies it.
+            const char* const engineName =
+                MG_Util::Async::AsyncPoolEngineName(MG_Util::Async::DetectAsyncPoolEngine());
             builder.Pass(rowName,
-                         format("on with {} compiler thread{}; GL_KHR_parallel_shader_compile is advertised "
+                         format("on with {} compiler thread{} on the {} execution engine; "
+                                "GL_KHR_parallel_shader_compile is advertised "
                                 "and GL_MAX_SHADER_COMPILER_THREADS_KHR = {} (set environment variable "
-                                "MOBILEGL_ASYNC_SHADER_COMPILE=0 to disable it, or "
+                                "MOBILEGL_ASYNC_SHADER_COMPILE=0 to disable it, "
-                                "MOBILEGL_ASYNC_SHADER_COMPILE_THREADS=n to change the count)",
+                                "MOBILEGL_ASYNC_SHADER_COMPILE_THREADS=n to change the count, or "
-                                threads, threads == 1 ? "" : "s", threads));
+                                "MOBILEGL_ASYNC_POOL=asio|libfork to change the engine)",
+                                threads, threads == 1 ? "" : "s", engineName, threads));
         }
 
         // Appends the four "MobileGL reported ..." rows for one backend section.