Fix(sim): pass context helpers to AICore SO instead of dlsym(RTLD_DEFAULT)

src/a2a3/platform/sim/aicore/inner_kernel.h

@@ -186,31 +186,34 @@ inline uint32_t get_physical_core_id() { return g_sim_physical_core_id; }
 // =============================================================================
 // CPU Simulation Context Hooks
 // =============================================================================
+//
+// These functions bridge the AICore kernel SO → host runtime SO gap.
+// The host runtime's DeviceRunner calls set_sim_context_helpers() after
+// dlopen to pass function pointers, avoiding dlsym(RTLD_DEFAULT) which
+// fails when the host SO is loaded with RTLD_LOCAL.
+
+// Function pointer storage — set by set_sim_context_helpers(), used by macros below.
+using SimSetExecCtxFn = void (*)(uint32_t, uint32_t, uint32_t);
+using SimSetTaskCookieFn = void (*)(uint64_t);
+using SimGetTaskCookieFn = uint64_t (*)(uint32_t, uint32_t);
+
+extern SimSetExecCtxFn g_sim_set_exec_ctx_fn;
+extern SimSetTaskCookieFn g_sim_set_task_cookie_fn;
+extern SimGetTaskCookieFn g_sim_get_task_cookie_fn;
 
-// CPU_SIM_SET_EXECUTION_CONTEXT — set block/subblock context for CANN intrinsic emulation.
-// Resolves pto_cpu_sim_set_execution_context (defined in cpu_sim_context.cpp) via dlsym.
 inline void cpu_sim_set_execution_context(uint32_t block_idx, uint32_t subblock_id, uint32_t subblock_dim) {
-    using Fn = void (*)(uint32_t, uint32_t, uint32_t);
-    static auto fn = reinterpret_cast<Fn>(dlsym(RTLD_DEFAULT, "pto_cpu_sim_set_execution_context"));
-    if (fn != nullptr) fn(block_idx, subblock_id, subblock_dim);
+    if (g_sim_set_exec_ctx_fn != nullptr) g_sim_set_exec_ctx_fn(block_idx, subblock_id, subblock_dim);
 }
 #define CPU_SIM_SET_EXECUTION_CONTEXT(block_idx, subblock_id, subblock_dim) \
     cpu_sim_set_execution_context(block_idx, subblock_id, subblock_dim)
 
-// CPU_SIM_SET_TASK_COOKIE — set task cookie for simulation tracing.
-// Resolves pto_cpu_sim_set_task_cookie (defined in cpu_sim_context.cpp) via dlsym.
 inline void cpu_sim_set_task_cookie(uint64_t cookie) {
-    using Fn = void (*)(uint64_t);
-    static auto fn = reinterpret_cast<Fn>(dlsym(RTLD_DEFAULT, "pto_cpu_sim_set_task_cookie"));
-    if (fn != nullptr) fn(cookie);
+    if (g_sim_set_task_cookie_fn != nullptr) g_sim_set_task_cookie_fn(cookie);
 }
 #define CPU_SIM_SET_TASK_COOKIE(cookie) cpu_sim_set_task_cookie(cookie)
 
-// platform_get_cpu_sim_task_cookie — resolve per-dispatch logical task identity.
 inline uint64_t platform_get_cpu_sim_task_cookie(uint32_t core_id, uint32_t reg_task_id) {
-    using Fn = uint64_t (*)(uint32_t, uint32_t);
-    static auto fn = reinterpret_cast<Fn>(dlsym(RTLD_DEFAULT, "platform_get_cpu_sim_task_cookie"));
-    return (fn != nullptr) ? fn(core_id, reg_task_id) : 0;
+    return (g_sim_get_task_cookie_fn != nullptr) ? g_sim_get_task_cookie_fn(core_id, reg_task_id) : 0;
 }
 
 #endif  // PLATFORM_A2A3SIM_AICORE_INNER_KERNEL_H_

src/a2a3/platform/sim/aicore/kernel.cpp

@@ -16,6 +16,7 @@
  */
 
 #include <cstdint>
+#include "inner_kernel.h"  // NOLINT(build/include_subdir)
 #include "aicore/aicore.h"
 #include "common/core_type.h"
 #include "common/platform_config.h"
@@ -27,6 +28,17 @@ thread_local volatile uint8_t *g_sim_reg_base = nullptr;
 // Thread-local simulated physical core ID (declared in inner_kernel.h)
 thread_local uint32_t g_sim_physical_core_id = 0;
 
+// Sim context function pointers — set by DeviceRunner after dlopen.
+SimSetExecCtxFn g_sim_set_exec_ctx_fn = nullptr;
+SimSetTaskCookieFn g_sim_set_task_cookie_fn = nullptr;
+SimGetTaskCookieFn g_sim_get_task_cookie_fn = nullptr;
+
+extern "C" void set_sim_context_helpers(void *set_exec_ctx, void *set_task_cookie, void *get_task_cookie) {
+    g_sim_set_exec_ctx_fn = reinterpret_cast<SimSetExecCtxFn>(set_exec_ctx);
+    g_sim_set_task_cookie_fn = reinterpret_cast<SimSetTaskCookieFn>(set_task_cookie);
+    g_sim_get_task_cookie_fn = reinterpret_cast<SimGetTaskCookieFn>(get_task_cookie);
+}
+
 // Declare the original function (defined in aicore_executor.cpp with weak linkage)
 void aicore_execute(__gm__ Runtime *runtime, int block_idx, CoreType core_type);
 

src/a2a3/platform/sim/host/cpu_sim_context.h

@@ -19,9 +19,23 @@
 
 #pragma once
 
+#include <cstdint>
+
 /**
  * Free all entries in the CPU simulation shared storage map and reset the
  * pthread-backed per-thread execution context store.
  * Called by DeviceRunner::run() at start and DeviceRunner::finalize() at end.
  */
 void clear_cpu_sim_shared_storage();
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void pto_cpu_sim_set_execution_context(uint32_t block_idx, uint32_t subblock_id, uint32_t subblock_dim);
+void pto_cpu_sim_set_task_cookie(uint64_t task_cookie);
+uint64_t platform_get_cpu_sim_task_cookie(uint32_t core_id, uint32_t reg_task_id);
+
+#ifdef __cplusplus
+}
+#endif

src/a2a3/platform/sim/host/device_runner.cpp

@@ -167,6 +167,19 @@ int DeviceRunner::ensure_binaries_loaded(
             return -1;
         }
         LOG_INFO("DeviceRunner(sim): Loaded aicore_execute_wrapper from %s", aicore_so_path_.c_str());
+
+        // Pass sim context function pointers to the AICore SO so it doesn't
+        // need dlsym(RTLD_DEFAULT) — which fails when the host runtime SO
+        // is loaded with RTLD_LOCAL.
+        auto set_helpers =
+            reinterpret_cast<void (*)(void *, void *, void *)>(dlsym(aicore_so_handle_, "set_sim_context_helpers"));
+        if (set_helpers != nullptr) {
+            set_helpers(
+                reinterpret_cast<void *>(pto_cpu_sim_set_execution_context),
+                reinterpret_cast<void *>(pto_cpu_sim_set_task_cookie),
+                reinterpret_cast<void *>(platform_get_cpu_sim_task_cookie)
+            );
+        }
     }
 
     return 0;

src/a5/platform/sim/aicore/inner_kernel.h

@@ -195,30 +195,28 @@ inline uint32_t get_physical_core_id() { return g_sim_physical_core_id; }
 // CPU Simulation Context Hooks
 // =============================================================================
 
-// CPU_SIM_SET_EXECUTION_CONTEXT — set block/subblock context for CANN intrinsic emulation.
-// Resolves pto_cpu_sim_set_execution_context (defined in cpu_sim_context.cpp) via dlsym.
+// Function pointer storage — set by set_sim_context_helpers(), used by macros below.
+using SimSetExecCtxFn = void (*)(uint32_t, uint32_t, uint32_t);
+using SimSetTaskCookieFn = void (*)(uint64_t);
+using SimGetTaskCookieFn = uint64_t (*)(uint32_t, uint32_t);
+
+extern SimSetExecCtxFn g_sim_set_exec_ctx_fn;
+extern SimSetTaskCookieFn g_sim_set_task_cookie_fn;
+extern SimGetTaskCookieFn g_sim_get_task_cookie_fn;
+
 inline void cpu_sim_set_execution_context(uint32_t block_idx, uint32_t subblock_id, uint32_t subblock_dim) {
-    using Fn = void (*)(uint32_t, uint32_t, uint32_t);
-    static auto fn = reinterpret_cast<Fn>(dlsym(RTLD_DEFAULT, "pto_cpu_sim_set_execution_context"));
-    if (fn != nullptr) fn(block_idx, subblock_id, subblock_dim);
+    if (g_sim_set_exec_ctx_fn != nullptr) g_sim_set_exec_ctx_fn(block_idx, subblock_id, subblock_dim);
 }
 #define CPU_SIM_SET_EXECUTION_CONTEXT(block_idx, subblock_id, subblock_dim) \
     cpu_sim_set_execution_context(block_idx, subblock_id, subblock_dim)
 
-// CPU_SIM_SET_TASK_COOKIE — set task cookie for simulation tracing.
-// Resolves pto_cpu_sim_set_task_cookie (defined in cpu_sim_context.cpp) via dlsym.
 inline void cpu_sim_set_task_cookie(uint64_t cookie) {
-    using Fn = void (*)(uint64_t);
-    static auto fn = reinterpret_cast<Fn>(dlsym(RTLD_DEFAULT, "pto_cpu_sim_set_task_cookie"));
-    if (fn != nullptr) fn(cookie);
+    if (g_sim_set_task_cookie_fn != nullptr) g_sim_set_task_cookie_fn(cookie);
 }
 #define CPU_SIM_SET_TASK_COOKIE(cookie) cpu_sim_set_task_cookie(cookie)
 
-// platform_get_cpu_sim_task_cookie — resolve per-dispatch logical task identity.
 inline uint64_t platform_get_cpu_sim_task_cookie(uint32_t core_id, uint32_t reg_task_id) {
-    using Fn = uint64_t (*)(uint32_t, uint32_t);
-    static auto fn = reinterpret_cast<Fn>(dlsym(RTLD_DEFAULT, "platform_get_cpu_sim_task_cookie"));
-    return (fn != nullptr) ? fn(core_id, reg_task_id) : 0;
+    return (g_sim_get_task_cookie_fn != nullptr) ? g_sim_get_task_cookie_fn(core_id, reg_task_id) : 0;
 }
 
 #endif  // PLATFORM_A5SIM_AICORE_INNER_KERNEL_H_

src/a5/platform/sim/aicore/kernel.cpp

@@ -16,6 +16,7 @@
  */
 
 #include <cstdint>
+#include "inner_kernel.h"  // NOLINT(build/include_subdir)
 #include "aicore/aicore.h"
 #include "common/core_type.h"
 #include "common/platform_config.h"
@@ -27,6 +28,17 @@ thread_local volatile uint8_t *g_sim_reg_base = nullptr;
 // Thread-local simulated physical core ID (declared in inner_kernel.h)
 thread_local uint32_t g_sim_physical_core_id = 0;
 
+// Sim context function pointers — set by DeviceRunner after dlopen.
+SimSetExecCtxFn g_sim_set_exec_ctx_fn = nullptr;
+SimSetTaskCookieFn g_sim_set_task_cookie_fn = nullptr;
+SimGetTaskCookieFn g_sim_get_task_cookie_fn = nullptr;
+
+extern "C" void set_sim_context_helpers(void *set_exec_ctx, void *set_task_cookie, void *get_task_cookie) {
+    g_sim_set_exec_ctx_fn = reinterpret_cast<SimSetExecCtxFn>(set_exec_ctx);
+    g_sim_set_task_cookie_fn = reinterpret_cast<SimSetTaskCookieFn>(set_task_cookie);
+    g_sim_get_task_cookie_fn = reinterpret_cast<SimGetTaskCookieFn>(get_task_cookie);
+}
+
 // Declare the original function (defined in aicore_executor.cpp with weak linkage)
 void aicore_execute(__gm__ Runtime *runtime, int block_idx, CoreType core_type);
 

src/a5/platform/sim/host/cpu_sim_context.h

@@ -19,9 +19,23 @@
 
 #pragma once
 
+#include <cstdint>
+
 /**
  * Free all entries in the CPU simulation shared storage map and reset the
  * pthread-backed per-thread execution context store.
  * Called by DeviceRunner::run() at start and DeviceRunner::finalize() at end.
  */
 void clear_cpu_sim_shared_storage();
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void pto_cpu_sim_set_execution_context(uint32_t block_idx, uint32_t subblock_id, uint32_t subblock_dim);
+void pto_cpu_sim_set_task_cookie(uint64_t task_cookie);
+uint64_t platform_get_cpu_sim_task_cookie(uint32_t core_id, uint32_t reg_task_id);
+
+#ifdef __cplusplus
+}
+#endif

src/a5/platform/sim/host/device_runner.cpp

@@ -167,6 +167,19 @@ int DeviceRunner::ensure_binaries_loaded(
             return -1;
         }
         LOG_INFO("DeviceRunner(sim): Loaded aicore_execute_wrapper from %s", aicore_so_path_.c_str());
+
+        // Pass sim context function pointers to the AICore SO so it doesn't
+        // need dlsym(RTLD_DEFAULT) — which fails when the host runtime SO
+        // is loaded with RTLD_LOCAL.
+        auto set_helpers =
+            reinterpret_cast<void (*)(void *, void *, void *)>(dlsym(aicore_so_handle_, "set_sim_context_helpers"));
+        if (set_helpers != nullptr) {
+            set_helpers(
+                reinterpret_cast<void *>(pto_cpu_sim_set_execution_context),
+                reinterpret_cast<void *>(pto_cpu_sim_set_task_cookie),
+                reinterpret_cast<void *>(platform_get_cpu_sim_task_cookie)
+            );
+        }
     }
 
     return 0;