vendor : update subprocess.h (#27409)

* cuda : add POOL_1D support

* fix: add missing trailing newline for editorconfig compliance

common/json.cpp

@@ -78,19 +78,21 @@ common_json_value::common_json_value(const common_json & val) :
 common_json_value::common_json_value(common_json && val) :
     type(VAL_JSON), val_json(std::make_shared<common_json>(std::move(val))) {}
 
+// the ctors and get<T>() below are explicit specializations, giving strong symbols
+// an explicit instantiation is a weak symbol, dropped by some LTO builds (clang-cl)
 template <typename T>
-common_json_value::common_json_value(const std::set<T> & vals) : type(VAL_JSON) {
+static std::shared_ptr<common_json> set_json(const std::set<T> & vals) {
     common_json out = common_json::array();
 
     for (const auto & val : vals) {
         out.push_back(val);
     }
 
-    val_json = std::make_shared<common_json>(std::move(out));
+    return std::make_shared<common_json>(std::move(out));
 }
 
 // a set value is usable only for the types below
-#define COMMON_JSON_SET(...) template common_json_value::common_json_value(const std::set<__VA_ARGS__> &);
+#define COMMON_JSON_SET(...) template <> common_json_value::common_json_value(const std::set<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(set_json(vals)) {}
 
 COMMON_JSON_SET(int)
 COMMON_JSON_SET(std::string)
@@ -98,56 +100,45 @@ COMMON_JSON_SET(std::string)
 #undef COMMON_JSON_SET
 
 template <typename T>
-common_json_value::common_json_value(const std::map<std::string, T> & vals) : type(VAL_JSON) {
+static std::shared_ptr<common_json> map_json(const T & vals) {
     common_json out = common_json::object();
 
     for (const auto & val : vals) {
         out.set({ val.first, val.second });
     }
 
-    val_json = std::make_shared<common_json>(std::move(out));
+    return std::make_shared<common_json>(std::move(out));
 }
 
 // a map value is usable only for the types below
-#define COMMON_JSON_MAP(...) template common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> &);
+#define COMMON_JSON_MAP(...) template <> common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
 
 COMMON_JSON_MAP(bool)
 COMMON_JSON_MAP(std::string)
 
 #undef COMMON_JSON_MAP
 
-template <typename T>
-common_json_value::common_json_value(const std::unordered_map<std::string, T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::object();
-
-    for (const auto & val : vals) {
-        out.set({ val.first, val.second });
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
 // an unordered map value is usable only for the types below
-#define COMMON_JSON_UMAP(...) template common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> &);
+#define COMMON_JSON_UMAP(...) template <> common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
 
 COMMON_JSON_UMAP(size_t)
 
 #undef COMMON_JSON_UMAP
 
 template <typename T>
-common_json_value::common_json_value(const std::vector<T> & vals) : type(VAL_JSON) {
+static std::shared_ptr<common_json> vec_json(const std::vector<T> & vals) {
     common_json out = common_json::array();
 
     for (const auto & val : vals) {
         out.push_back(val);
     }
 
-    val_json = std::make_shared<common_json>(std::move(out));
+    return std::make_shared<common_json>(std::move(out));
 }
 
 // a vector value is usable only for the types below
 // note: std::vector<bool> is not here, its proxy reference does not convert
-#define COMMON_JSON_VEC(...) template common_json_value::common_json_value(const std::vector<__VA_ARGS__> &);
+#define COMMON_JSON_VEC(...) template <> common_json_value::common_json_value(const std::vector<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(vec_json(vals)) {}
 
 COMMON_JSON_VEC(int)
 COMMON_JSON_VEC(unsigned char)
@@ -404,10 +395,6 @@ common_json::items_view common_json::items() const {
     return items_view(const_cast<common_json *>(this), size());
 }
 
-template <typename T> T common_json::get() const {
-    return guard([&] { return as_json(this).get<T>(); });
-}
-
 // the backing library cannot build a common_json, so this one is just a copy
 template <> common_json common_json::get<common_json>() const {
     return *this;
@@ -415,7 +402,7 @@ template <> common_json common_json::get<common_json>() const {
 
 // get<T>() is usable only for the types below
 
-#define COMMON_JSON_GET(...) template __VA_ARGS__ common_json::get<__VA_ARGS__>() const;
+#define COMMON_JSON_GET(...) template <> __VA_ARGS__ common_json::get<__VA_ARGS__>() const { return guard([&] { return as_json(this).get<__VA_ARGS__>(); }); }
 
 COMMON_JSON_GET(bool)
 COMMON_JSON_GET(int)
@@ -435,3 +422,12 @@ COMMON_JSON_GET(std::vector<size_t>)
 COMMON_JSON_GET(std::unordered_map<std::string, size_t>)
 
 #undef COMMON_JSON_GET
+
+// must stay below the get<std::string> specialization
+common_json::operator std::string() const {
+    return get<std::string>();
+}
+
+std::string common_json::value(const std::string & key, const char * def) const {
+    return contains(key) ? at(key).get<std::string>() : std::string(def);
+}

common/json.h

@@ -221,16 +221,14 @@ class common_json {
     // implicit get<T>() for plain values, so they can be assigned to their C++ type directly
     // note: kept to this short list on purpose, a wider one makes j["key"] ambiguous
     // note: a numeric one would make "str = json;" ambiguous, a number converts to char too
-    operator std::string() const { return get<std::string>(); }
+    operator std::string() const;
 
     template <typename T>
     T value(const std::string & key, T def) const {
         return contains(key) ? at(key).get<T>() : def;
     }
 
-    std::string value(const std::string & key, const char * def) const {
-        return contains(key) ? at(key).get<std::string>() : std::string(def);
-    }
+    std::string value(const std::string & key, const char * def) const;
 
     // a JSON default needs no get<T>(), it is already the right type
     common_json value(const std::string & key, const common_json & def) const {

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -38,6 +38,7 @@
 #include "ggml-cuda/out-prod.cuh"
 #include "ggml-cuda/pad.cuh"
 #include "ggml-cuda/pool2d.cuh"
+#include "ggml-cuda/pool1d.cuh"
 #include "ggml-cuda/quantize.cuh"
 #include "ggml-cuda/rope.cuh"
 #include "ggml-cuda/roll.cuh"
@@ -2326,6 +2327,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_POOL_2D:
             ggml_cuda_op_pool2d(ctx, dst);
             break;
+        case GGML_OP_POOL_1D:
+            ggml_cuda_op_pool1d(ctx, dst);
+            break;
         case GGML_OP_SUM:
             ggml_cuda_op_sum(ctx, dst);
             break;
@@ -5245,6 +5249,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_2D_DW:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_TRANSPOSE_2D:
+        case GGML_OP_POOL_1D:
         case GGML_OP_POOL_2D:
             return true;
         case GGML_OP_ACC:

ggml/src/ggml-cuda/pool1d.cu

@@ -0,0 +1,85 @@
+#include "pool1d.cuh"
+
+static __global__ void pool1d_nchw_kernel(
+        const int iw, const int ow,
+        const int kw, const int sw, const int pw,
+        const int parallel_elements,
+        const float * src, float * dst, const enum ggml_op_pool op) {
+    const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (idx >= parallel_elements) {
+        return;
+    }
+
+    const int nc = idx / ow;
+    const int cur_ow = idx % ow;
+
+    const float * i_ptr = src + nc * iw;
+    float * o_ptr = dst + nc * ow;
+
+    const int start = cur_ow * sw - pw;
+    const int b = max(0, start);
+    const int e = min(iw, start + kw);
+
+    float res;
+    switch (op) {
+        case GGML_OP_POOL_AVG: res = 0.0f;     break;
+        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+        default: return;
+    }
+
+    int count = 0;
+    for (int i = b; i < e; i++) {
+#if __CUDA_ARCH__ >= 350
+        float cur = __ldg(i_ptr + i);
+#else
+        float cur = i_ptr[i];
+#endif
+        switch (op) {
+            case GGML_OP_POOL_AVG: res += cur;                break;
+            case GGML_OP_POOL_MAX: res = max(res, cur);       break;
+            default: break;
+        }
+        count++;
+    }
+
+    if (op == GGML_OP_POOL_AVG) {
+        res = (count > 0) ? (res / count) : 0.0f;
+    }
+
+    o_ptr[cur_ow] = res;
+}
+
+static void pool1d_nchw_kernel_f32_f32_cuda(
+        const int iw, const int ow,
+        const int kw, const int sw, const int pw,
+        const int parallel_elements,
+        const float * src, float * dst, const enum ggml_op_pool op,
+        cudaStream_t stream) {
+    const int num_blocks = (parallel_elements + CUDA_POOL1D_BLOCK_SIZE - 1) / CUDA_POOL1D_BLOCK_SIZE;
+    dim3 block_nums(num_blocks);
+    pool1d_nchw_kernel<<<block_nums, CUDA_POOL1D_BLOCK_SIZE, 0, stream>>>(iw, ow, kw, sw, pw, parallel_elements, src, dst, op);
+}
+
+void ggml_cuda_op_pool1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    const int k0 = opts[1];
+    const int s0 = opts[2];
+    const int p0 = opts[3];
+
+    const int64_t IW = src0->ne[0];
+    const int64_t OW = dst->ne[0];
+    const int64_t nr = ggml_nrows(src0);
+
+    const int parallel_elements = (int)(nr * OW);
+
+    pool1d_nchw_kernel_f32_f32_cuda(IW, OW, k0, s0, p0, parallel_elements, src0_d, dst_d, op, stream);
+}

ggml/src/ggml-cuda/pool1d.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_POOL1D_BLOCK_SIZE 256
+
+void ggml_cuda_op_pool1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -955,6 +955,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_diag[2];
     vk_pipeline pipeline_clamp[2];
     vk_pipeline pipeline_pad_f32;
+    vk_pipeline pipeline_pad_reflect_1d_f32;
     vk_pipeline pipeline_roll_f32;
     vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32;
     vk_pipeline pipeline_repeat_i16;
@@ -5630,6 +5631,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_pad_reflect_1d_f32, "pad_reflect_1d_f32", pad_reflect_1d_f32_len, pad_reflect_1d_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -11336,6 +11338,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_pad_f32;
         }
         return nullptr;
+    case GGML_OP_PAD_REFLECT_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pad_reflect_1d_f32;
+        }
+        return nullptr;
     case GGML_OP_ROLL:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_roll_f32;
@@ -12239,6 +12246,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CLAMP:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_PAD:
+    case GGML_OP_PAD_REFLECT_1D:
     case GGML_OP_ROLL:
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
@@ -13111,6 +13119,17 @@ static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p));
 }
 
+static void ggml_vk_pad_reflect_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t p0 = (uint32_t)dst->op_params[0];
+    const uint32_t p1 = (uint32_t)dst->op_params[1];
+
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
+    memcpy(&p.param1, &p0, sizeof(float));
+    memcpy(&p.param2, &p1, sizeof(float));
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD_REFLECT_1D, std::move(p));
+}
+
 static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const int32_t s0 = ggml_get_op_params_i32(dst, 0);
     const int32_t s1 = ggml_get_op_params_i32(dst, 1);
@@ -15520,6 +15539,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_PAD:
         ggml_vk_pad(ctx, compute_ctx, src0, node);
 
+        break;
+    case GGML_OP_PAD_REFLECT_1D:
+        ggml_vk_pad_reflect_1d(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_ROLL:
         ggml_vk_roll(ctx, compute_ctx, src0, node);
@@ -18446,6 +18469,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SCALE:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ROLL:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_DIAG_MASK_INF:
@@ -19228,6 +19252,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else if (tensor->op == GGML_OP_PAD) {
             tensor_clone = ggml_pad_ext(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3],
                                                                 tensor->op_params[4], tensor->op_params[5], tensor->op_params[6], tensor->op_params[7]);
+        } else if (tensor->op == GGML_OP_PAD_REFLECT_1D) {
+            tensor_clone = ggml_pad_reflect_1d(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1]);
         } else if (tensor->op == GGML_OP_REPEAT) {
             tensor_clone = ggml_repeat(ggml_ctx, src_clone[0], tensor);
         } else if (tensor->op == GGML_OP_REPEAT_BACK) {

ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp

@@ -0,0 +1,43 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_unary_head.glsl" // included to use functions like fastdiv etc.
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint p0 = floatBitsToUint(p.param1);
+    const uint p1 = floatBitsToUint(p.param2);
+
+    const uint i3 = fastdiv(idx, p.ne1_012mp, fastdiv_L(p.ne1_Ls, 0));
+    const uint i3_offset = i3 * p.ne12 * p.ne11 * p.ne10;
+
+    const uint i2 = fastdiv(idx - i3_offset, p.ne1_01mp, fastdiv_L(p.ne1_Ls, 1));
+    const uint i2_offset = i2 * p.ne11 * p.ne10;
+
+    const uint i1 = fastdiv(idx - i3_offset - i2_offset, p.ne1_0mp, fastdiv_L(p.ne1_Ls, 2));
+    const uint i0 = idx - i3_offset - i2_offset - i1 * p.ne10;
+
+    uint src_col;
+
+    if (i0 < p0) {
+        src_col = p0 - i0;    // left pad area
+    } else if (i0 < p0 + p.ne00) {
+        src_col = i0 - p0;    // center area
+    } else {
+        src_col = 2u * p.ne00 - 2u - (i0 - p0);  // right pad area
+    }
+
+    const uint src_idx = i3 * p.nb03 + i2 * p.nb02 + i1 * p.nb01 + src_col * p.nb00;
+    const uint d_idx = i3 * p.nb13 + i2 * p.nb12 + i1 * p.nb11 + i0 * p.nb10;
+
+    // copy the computed value to the destination tensor
+    data_d[get_doffset() + d_idx] = D_TYPE(data_a[get_aoffset() + src_idx]);
+}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -896,6 +896,7 @@ void process_shaders() {
     string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
     string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    string_to_spv("pad_reflect_1d_f32", "pad_reflect_1d.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
 
     string_to_spv("concat_i8", "concat.comp", {{"A_TYPE", "uint8_t"}, {"B_TYPE", "uint8_t"}, {"D_TYPE", "uint8_t"}});
     string_to_spv("concat_i16", "concat.comp", {{"A_TYPE", "uint16_t"}, {"B_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});

scripts/sync_vendor.py

@@ -27,7 +27,7 @@ vendor = {
     f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/split.py":  "split.py",
     f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/LICENSE":   "vendor/cpp-httplib/LICENSE",
 
-    "https://raw.githubusercontent.com/sheredom/subprocess.h/9ce0d701b6fb10f8f8c4445edd31e7c60a1237e3/subprocess.h": "vendor/sheredom/subprocess.h",
+    "https://raw.githubusercontent.com/sheredom/subprocess.h/0dccaa9aa176dd6d7ef8afeca3c18d6e80a32795/subprocess.h": "vendor/sheredom/subprocess.h",
 
     f"https://raw.githubusercontent.com/Cyan4973/xxHash/{XXHASH_COMMIT}/xxhash.c":      "vendor/hash/xxhash/xxhash.c",
     f"https://raw.githubusercontent.com/Cyan4973/xxHash/{XXHASH_COMMIT}/xxhash.h":      "vendor/hash/xxhash/xxhash.h",

vendor/sheredom/subprocess.h

@@ -275,13 +275,46 @@ subprocess_weak int subprocess_alive(struct subprocess_s *const process);
 #include <unistd.h>
 #endif
 
+#if defined(__NetBSD__)
+#include <sys/param.h>
+#endif
+
+/* Which spelling of the chdir file action the platform provides, if any.
+   POSIX 2024 standardised posix_spawn_file_actions_addchdir; implementations
+   that shipped it earlier called it ..._np. macOS 26 and NetBSD 10 use the
+   standard name, glibc 2.29+, macOS 10.15+ and FreeBSD 13.1+ use the _np name,
+   and AIX, NetBSD 9 and older, and OpenBSD provide neither. */
+#if !defined(SUBPROCESS_ADDCHDIR_IS_POSIX)
+#if (defined(__APPLE__) && MAC_OS_X_VERSION_MIN_REQUIRED >= 260000) ||        \
+    (defined(__NetBSD__) && __NetBSD_Version__ >= 1000000000)
+#define SUBPROCESS_ADDCHDIR_IS_POSIX 1
+#else
+#define SUBPROCESS_ADDCHDIR_IS_POSIX 0
+#endif
+#endif
+
+/* Whether to launch the child with fork()+exec() instead of posix_spawn(),
+   for platforms with no posix_spawn_file_actions_addchdir under either
+   spelling: the child chdir()s before exec, and a close-on-exec pipe carries
+   exec's errno back. Define this yourself to force either implementation. */
+#if !defined(SUBPROCESS_SPAWN_VIA_FORK)
+#if defined(_AIX) || defined(__OpenBSD__) ||                                  \
+    (defined(__NetBSD__) && (__NetBSD_Version__ < 1000000000))
+#define SUBPROCESS_SPAWN_VIA_FORK 1
+#else
+#define SUBPROCESS_SPAWN_VIA_FORK 0
+#endif
+#endif
+
 /* Whether subprocess_create_ex can honour process_cwd. glibc only gained
    posix_spawn_file_actions_addchdir_np in 2.29, and macOS in 10.15; the SDKs
    mark it unavailable on iOS, tvOS and watchOS, where the undefined version
    macro folds to 0 and so answers correctly. Define this yourself to override
    the detection, for instance on musl older than 1.1.24. */
 #if !defined(SUBPROCESS_HAVE_CWD)
-#if defined(__GLIBC__)
+#if SUBPROCESS_SPAWN_VIA_FORK
+#define SUBPROCESS_HAVE_CWD 1
+#elif defined(__GLIBC__)
 #if __GLIBC_PREREQ(2, 29)
 #define SUBPROCESS_HAVE_CWD 1
 #else
@@ -294,10 +327,13 @@ subprocess_weak int subprocess_alive(struct subprocess_s *const process);
 #endif
 #endif
 
-/* Whether posix_spawn reports a failed exec back to the caller. glibc only
-   started doing so in 2.24; before that the child silently exits with 127. */
+/* Whether a failed exec is reported back to the caller. The fork() path always
+   reports it through its error pipe. glibc's posix_spawn only started doing so
+   in 2.24; before that the child silently exits with 127. */
 #if !defined(SUBPROCESS_SPAWN_REPORTS_EXEC_ERRORS)
-#if defined(__GLIBC__)
+#if SUBPROCESS_SPAWN_VIA_FORK
+#define SUBPROCESS_SPAWN_REPORTS_EXEC_ERRORS 1
+#elif defined(__GLIBC__)
 #if __GLIBC_PREREQ(2, 24)
 #define SUBPROCESS_SPAWN_REPORTS_EXEC_ERRORS 1
 #else
@@ -342,6 +378,14 @@ typedef intptr_t subprocess_intptr_t;
 typedef size_t subprocess_size_t;
 #endif
 
+/* SIZE_T is ULONG_PTR, which is not size_t: on Win32 both are 32 bits wide but
+   unsigned long and unsigned int are still distinct types. */
+#ifdef _WIN64
+typedef subprocess_size_t subprocess_ulongptr_t;
+#else
+typedef unsigned long subprocess_ulongptr_t;
+#endif
+
 #ifdef __clang__
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wreserved-identifier"
@@ -351,6 +395,7 @@ typedef struct _PROCESS_INFORMATION *LPPROCESS_INFORMATION;
 typedef struct _SECURITY_ATTRIBUTES *LPSECURITY_ATTRIBUTES;
 typedef struct _STARTUPINFOW *LPSTARTUPINFOW;
 typedef struct _OVERLAPPED *LPOVERLAPPED;
+typedef struct _PROC_THREAD_ATTRIBUTE_LIST *LPPROC_THREAD_ATTRIBUTE_LIST;
 
 #ifdef __clang__
 #pragma clang diagnostic pop
@@ -402,6 +447,11 @@ struct subprocess_startup_info_s {
   void *hStdError;
 };
 
+struct subprocess_startup_info_ex_s {
+  struct subprocess_startup_info_s startupInfo;
+  void *attributeList;
+};
+
 struct subprocess_overlapped_s {
   uintptr_t Internal;
   uintptr_t InternalHigh;
@@ -451,6 +501,14 @@ __declspec(dllimport) int __stdcall CreateProcessW(
     const subprocess_wchar_t *, subprocess_wchar_t *, LPSECURITY_ATTRIBUTES,
     LPSECURITY_ATTRIBUTES, int, unsigned long, void *,
     const subprocess_wchar_t *, LPSTARTUPINFOW, LPPROCESS_INFORMATION);
+__declspec(dllimport) int __stdcall
+InitializeProcThreadAttributeList(LPPROC_THREAD_ATTRIBUTE_LIST, unsigned long,
+                                  unsigned long, subprocess_ulongptr_t *);
+__declspec(dllimport) int __stdcall UpdateProcThreadAttribute(
+    LPPROC_THREAD_ATTRIBUTE_LIST, unsigned long, subprocess_ulongptr_t, void *,
+    subprocess_ulongptr_t, void *, subprocess_ulongptr_t *);
+__declspec(dllimport) void __stdcall
+DeleteProcThreadAttributeList(LPPROC_THREAD_ATTRIBUTE_LIST);
 __declspec(dllimport) int __stdcall MultiByteToWideChar(
     unsigned int, unsigned long, const char *, int, subprocess_wchar_t *, int);
 __declspec(dllimport) int __stdcall CloseHandle(void *);
@@ -667,12 +725,104 @@ int subprocess_create_named_pipe_helper(void **rd, void **wr) {
 }
 #endif
 
+#if !defined(_WIN32)
+/* Move a pipe end off 0, 1 or 2. Duplicating a descriptor onto itself is a
+   no-op, so a pipe end already sitting on a standard descriptor would keep its
+   FD_CLOEXEC and be closed by exec, leaving the child without that stream. */
+static int subprocess_fds_above_std(int fds[2]) {
+  int fd_flags;
+  int index;
+  int moved;
+  int saved_errno;
+
+  for (index = 0; index < 2; index++) {
+    if (fds[index] > STDERR_FILENO) {
+      continue;
+    }
+
+    moved = fcntl(fds[index], F_DUPFD, STDERR_FILENO + 1);
+    if (-1 != moved) {
+      fd_flags = fcntl(moved, F_GETFD, 0);
+      if ((-1 == fd_flags) ||
+          (-1 == fcntl(moved, F_SETFD, fd_flags | FD_CLOEXEC))) {
+        saved_errno = errno;
+        close(moved);
+        errno = saved_errno;
+        moved = -1;
+      }
+    }
+
+    if (-1 == moved) {
+      saved_errno = errno;
+      close(fds[0]);
+      close(fds[1]);
+      fds[0] = -1;
+      fds[1] = -1;
+      errno = saved_errno;
+      return -1;
+    }
+
+    close(fds[index]);
+    fds[index] = moved;
+  }
+
+  return 0;
+}
+
+/* Create pipes with close-on-exec set so later subprocesses do not inherit
+   descriptors belonging to subprocesses which are already running. */
+static int subprocess_pipe_cloexec(int fds[2]) {
+  int fd_flags;
+  int index;
+  int saved_errno;
+
+#if defined(__linux__) || defined(__FreeBSD__) || defined(__NetBSD__) ||       \
+    defined(__OpenBSD__) || defined(__DragonFly__) ||                         \
+    (defined(__sun) && defined(__SVR4))
+  if (0 == pipe2(fds, O_CLOEXEC)) {
+    return subprocess_fds_above_std(fds);
+  }
+
+  /* Older kernels can lack pipe2 even when the C library declares it. */
+  if (ENOSYS != errno) {
+    return -1;
+  }
+#endif
+
+  if (0 != pipe(fds)) {
+    return -1;
+  }
+
+  for (index = 0; index < 2; index++) {
+    fd_flags = fcntl(fds[index], F_GETFD, 0);
+    if ((-1 == fd_flags) ||
+        (-1 == fcntl(fds[index], F_SETFD, fd_flags | FD_CLOEXEC))) {
+      saved_errno = errno;
+      close(fds[0]);
+      close(fds[1]);
+      fds[0] = -1;
+      fds[1] = -1;
+      errno = saved_errno;
+      return -1;
+    }
+  }
+
+  return subprocess_fds_above_std(fds);
+}
+#endif
+
 int subprocess_create(const char *const commandLine[], int options,
                       struct subprocess_s *const out_process) {
   return subprocess_create_ex(commandLine, options, SUBPROCESS_NULL,
                               SUBPROCESS_NULL, out_process);
 }
 
+#if SUBPROCESS_SPAWN_VIA_FORK
+/* Not every platform declares execvpe: AIX exports it from libc without ever
+   naming it in a header, and glibc hides it behind _GNU_SOURCE. */
+extern int execvpe(const char *, char *const *, char *const *);
+#endif
+
 int subprocess_create_ex(const char *const commandLine[], int options,
                          const char *const environment[],
                          const char *const process_cwd,
@@ -692,6 +842,7 @@ int subprocess_create_ex(const char *const commandLine[], int options,
   subprocess_size_t bs_run;
   unsigned long flags = 0;
   unsigned long last_error = 0;
+  int attribute_list_initialized = 0;
   int result = subprocess_error_unknown;
   const unsigned int codePageUtf8 = 65001;
   const unsigned long mbErrInvalidChars = 0x00000008;
@@ -699,6 +850,8 @@ int subprocess_create_ex(const char *const commandLine[], int options,
   const unsigned long handleFlagInherit = 0x00000001;
   const unsigned long createNoWindow = 0x08000000;
   const unsigned long createUnicodeEnvironment = 0x00000400;
+  const unsigned long extendedStartupInfoPresent = 0x00080000;
+  const subprocess_size_t procThreadAttributeHandleList = 0x00020002;
   struct subprocess_subprocess_information_s processInfo = {SUBPROCESS_NULL,
                                                             SUBPROCESS_NULL, 0,
                                                             0};
@@ -706,6 +859,11 @@ int subprocess_create_ex(const char *const commandLine[], int options,
                                                     SUBPROCESS_NULL, 1};
   subprocess_wchar_t empty_environment[2] = {0, 0};
   subprocess_wchar_t *used_environment = SUBPROCESS_NULL;
+  subprocess_ulongptr_t attribute_list_size = 0;
+  subprocess_size_t inherited_handle_count = 0;
+  LPPROC_THREAD_ATTRIBUTE_LIST attribute_list = SUBPROCESS_NULL;
+  void *inherited_handles[3];
+  struct subprocess_startup_info_ex_s startInfoEx;
   struct subprocess_startup_info_s startInfo = {0,
                                                 SUBPROCESS_NULL,
                                                 SUBPROCESS_NULL,
@@ -1080,6 +1238,44 @@ int subprocess_create_ex(const char *const commandLine[], int options,
     }
   }
 
+  /* Restrict inheritance to this subprocess's standard streams. Without a
+     handle list, concurrent subprocess_create calls can inherit each other's
+     temporarily-inheritable child pipe handles. */
+  inherited_handles[inherited_handle_count++] = startInfo.hStdInput;
+  inherited_handles[inherited_handle_count++] = startInfo.hStdOutput;
+  if (startInfo.hStdError != startInfo.hStdOutput) {
+    inherited_handles[inherited_handle_count++] = startInfo.hStdError;
+  }
+
+  InitializeProcThreadAttributeList(SUBPROCESS_NULL, 1, 0,
+                                    &attribute_list_size);
+  if (0 == attribute_list_size) {
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+
+  attribute_list = SUBPROCESS_PTR_CAST(LPPROC_THREAD_ATTRIBUTE_LIST,
+                                       _alloca(attribute_list_size));
+  if (!attribute_list || !InitializeProcThreadAttributeList(
+                             attribute_list, 1, 0, &attribute_list_size)) {
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+  attribute_list_initialized = 1;
+
+  if (!UpdateProcThreadAttribute(
+          attribute_list, 0, procThreadAttributeHandleList, inherited_handles,
+          inherited_handle_count * sizeof(inherited_handles[0]),
+          SUBPROCESS_NULL, SUBPROCESS_NULL)) {
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+
+  startInfoEx.startupInfo = startInfo;
+  startInfoEx.startupInfo.cb = sizeof(startInfoEx);
+  startInfoEx.attributeList = attribute_list;
+  flags |= extendedStartupInfoPresent;
+
   if (!CreateProcessW(
           SUBPROCESS_NULL,
           commandLineCombinedWide, // command line
@@ -1090,7 +1286,7 @@ int subprocess_create_ex(const char *const commandLine[], int options,
           used_environment,        // used environment
           process_cwd_wide,        // use specified current directory
           SUBPROCESS_PTR_CAST(LPSTARTUPINFOW,
-                              &startInfo), // STARTUPINFO pointer
+                              &startInfoEx), // STARTUPINFOEX pointer
           SUBPROCESS_PTR_CAST(LPPROCESS_INFORMATION, &processInfo))) {
     result = subprocess_error_from_windows_error(GetLastError());
     if (subprocess_error_unknown == result) {
@@ -1099,6 +1295,9 @@ int subprocess_create_ex(const char *const commandLine[], int options,
     goto cleanup;
   }
 
+  DeleteProcThreadAttributeList(attribute_list);
+  attribute_list_initialized = 0;
+
   out_process->hProcess = processInfo.hProcess;
   processInfo.hProcess = SUBPROCESS_NULL;
 
@@ -1128,6 +1327,10 @@ int subprocess_create_ex(const char *const commandLine[], int options,
 cleanup:
   last_error = GetLastError();
 
+  if (attribute_list_initialized) {
+    DeleteProcThreadAttributeList(attribute_list);
+  }
+
   if (subprocess_error_unknown == result) {
     result = subprocess_error_from_windows_error(last_error);
   }
@@ -1173,15 +1376,20 @@ cleanup:
   int stderrfd[2] = {-1, -1};
   int fd, fd_flags;
   int async_no_wait;
-  int actions_created = 0;
   int result = subprocess_error_unknown;
   int saved_errno = 0;
-  int posix_error;
   pid_t child = 0;
   extern char **environ;
   char *const empty_environment[1] = {SUBPROCESS_NULL};
-  posix_spawn_file_actions_t actions;
   char *const *used_environment;
+#if SUBPROCESS_SPAWN_VIA_FORK
+  /* Pipe used to relay the child's exec() errno back to the parent. */
+  int exec_errfd[2] = {-1, -1};
+#else
+  int actions_created = 0;
+  int posix_error;
+  posix_spawn_file_actions_t actions;
+#endif
 
   async_no_wait = subprocess_option_enable_async_no_wait ==
                   (options & subprocess_option_enable_async_no_wait);
@@ -1202,13 +1410,13 @@ cleanup:
 
   memset(out_process, 0, sizeof(*out_process));
 
-  if (0 != pipe(stdinfd)) {
+  if (0 != subprocess_pipe_cloexec(stdinfd)) {
     saved_errno = errno;
     result = subprocess_error_pipe;
     goto cleanup;
   }
 
-  if (0 != pipe(stdoutfd)) {
+  if (0 != subprocess_pipe_cloexec(stdoutfd)) {
     saved_errno = errno;
     result = subprocess_error_pipe;
     goto cleanup;
@@ -1216,7 +1424,7 @@ cleanup:
 
   if (subprocess_option_combined_stdout_stderr !=
       (options & subprocess_option_combined_stdout_stderr)) {
-    if (0 != pipe(stderrfd)) {
+    if (0 != subprocess_pipe_cloexec(stderrfd)) {
       saved_errno = errno;
       result = subprocess_error_pipe;
       goto cleanup;
@@ -1240,6 +1448,136 @@ cleanup:
     used_environment = empty_environment;
   }
 
+#if SUBPROCESS_SPAWN_VIA_FORK
+  /* fork()+exec() instead of posix_spawn, so the child can chdir() first.
+     exec_errfd[1] is close-on-exec: a successful exec closes it and the parent
+     reads EOF; a failed exec writes errno through it before _exit. */
+  if (0 != pipe(exec_errfd)) {
+    saved_errno = errno;
+    result = subprocess_error_pipe;
+    goto cleanup;
+  }
+
+  if (-1 == fcntl(exec_errfd[1], F_SETFD, FD_CLOEXEC)) {
+    saved_errno = errno;
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+
+  child = fork();
+
+  if (child < 0) {
+    saved_errno = errno;
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+
+  if (0 == child) {
+    /* Child. Everything below must stay async-signal-safe: after fork() in a
+       threaded process only such functions may be called before exec. */
+    int child_errno;
+
+    close(exec_errfd[0]);
+
+    if ((-1 == dup2(stdinfd[0], STDIN_FILENO)) ||
+        (-1 == dup2(stdoutfd[1], STDOUT_FILENO))) {
+      goto child_failed;
+    }
+
+    if (subprocess_option_combined_stdout_stderr ==
+        (options & subprocess_option_combined_stdout_stderr)) {
+      if (-1 == dup2(STDOUT_FILENO, STDERR_FILENO)) {
+        goto child_failed;
+      }
+    } else {
+      if (-1 == dup2(stderrfd[1], STDERR_FILENO)) {
+        goto child_failed;
+      }
+    }
+
+    /* The originals are only closed once they have been duplicated, so that a
+       pipe end that already sits on 0, 1 or 2 is not closed out from under us. */
+    if (stdinfd[0] > STDERR_FILENO) {
+      close(stdinfd[0]);
+    }
+    if (stdinfd[1] > STDERR_FILENO) {
+      close(stdinfd[1]);
+    }
+    if (stdoutfd[0] > STDERR_FILENO) {
+      close(stdoutfd[0]);
+    }
+    if (stdoutfd[1] > STDERR_FILENO) {
+      close(stdoutfd[1]);
+    }
+    if (stderrfd[0] > STDERR_FILENO) {
+      close(stderrfd[0]);
+    }
+    if (stderrfd[1] > STDERR_FILENO) {
+      close(stderrfd[1]);
+    }
+
+    if (process_cwd && (0 != chdir(process_cwd))) {
+      goto child_failed;
+    }
+
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wcast-qual"
+#pragma clang diagnostic ignored "-Wold-style-cast"
+#endif
+    if (subprocess_option_search_user_path ==
+        (options & subprocess_option_search_user_path)) {
+      execvpe(commandLine[0],
+              SUBPROCESS_CONST_CAST(char *const *, commandLine),
+              SUBPROCESS_CONST_CAST(char *const *, used_environment));
+    } else {
+      execve(commandLine[0],
+             SUBPROCESS_CONST_CAST(char *const *, commandLine),
+             SUBPROCESS_CONST_CAST(char *const *, used_environment));
+    }
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif
+
+  child_failed:
+    child_errno = errno;
+    /* Nothing useful can be done if this write fails; the parent then sees EOF
+       and reports success, exactly as posix_spawn would without exec reporting. */
+    (void)!write(exec_errfd[1], &child_errno, sizeof(child_errno));
+    /* 127 is what POSIX requires posix_spawn's child to exit with when exec
+       fails, so both implementations look the same to a caller. */
+    _exit(127);
+  }
+
+  /* Parent. */
+  close(exec_errfd[1]);
+  exec_errfd[1] = -1;
+
+  {
+    int child_errno = 0;
+    ssize_t bytes_read;
+
+    do {
+      bytes_read = read(exec_errfd[0], &child_errno, sizeof(child_errno));
+    } while ((-1 == bytes_read) && (EINTR == errno));
+
+    close(exec_errfd[0]);
+    exec_errfd[0] = -1;
+
+    if (bytes_read == (ssize_t)sizeof(child_errno)) {
+      /* exec failed in the child. Reap it and surface the reason. */
+      while ((-1 == waitpid(child, SUBPROCESS_NULL, 0)) && (EINTR == errno)) {
+      }
+      child = 0;
+      saved_errno = child_errno;
+      result = subprocess_error_from_errno(child_errno);
+      if (subprocess_error_unknown == result) {
+        result = subprocess_error_spawn;
+      }
+      goto cleanup;
+    }
+  }
+#else
   posix_error = posix_spawn_file_actions_init(&actions);
   if (0 != posix_error) {
     saved_errno = posix_error;
@@ -1253,7 +1591,7 @@ cleanup:
 
   // Set working directory
   if (process_cwd) {
-#if defined(__NetBSD__) || (defined(__APPLE__) && MAC_OS_X_VERSION_MIN_REQUIRED >= 260000)
+#if SUBPROCESS_ADDCHDIR_IS_POSIX
     posix_error = posix_spawn_file_actions_addchdir(&actions, process_cwd);
 #elif !SUBPROCESS_HAVE_CWD
     posix_error = ENOSYS;
@@ -1406,6 +1744,7 @@ cleanup:
 #ifdef __clang__
 #pragma clang diagnostic pop
 #endif
+#endif /* SUBPROCESS_SPAWN_VIA_FORK */
 
   // Close the stdin read end
   close(stdinfd[0]);
@@ -1480,9 +1819,21 @@ cleanup:
     result = subprocess_error_from_errno(saved_errno);
   }
 
+#if SUBPROCESS_SPAWN_VIA_FORK
+  if (-1 != exec_errfd[0]) {
+    close(exec_errfd[0]);
+    exec_errfd[0] = -1;
+  }
+
+  if (-1 != exec_errfd[1]) {
+    close(exec_errfd[1]);
+    exec_errfd[1] = -1;
+  }
+#else
   if (actions_created) {
     posix_spawn_file_actions_destroy(&actions);
   }
+#endif
 
   if (0 != result) {
     if (child) {