Enable static PGO for Linux SGX (#2270)

Enable static PGO for Linux SGX and update the related benchmarks
test scripts and documents.

product-mini/platforms/linux-sgx/CMakeLists.txt

@@ -89,6 +89,11 @@ if (NOT DEFINED WAMR_BUILD_SGX_IPFS)
   set (WAMR_BUILD_SGX_IPFS 0)
 endif ()
 
+if (NOT DEFINED WAMR_BUILD_STATIC_PGO)
+  # Disable static PGO by default
+  set (WAMR_BUILD_STATIC_PGO 0)
+endif ()
+
 set (CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,--gc-sections")
 set (CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -std=gnu11 -ffunction-sections -fdata-sections \
                                      -Wall -Wno-unused-parameter -Wno-pedantic \
@@ -107,6 +112,18 @@ add_custom_command (
 
 add_custom_target (vmlib_untrusted ALL DEPENDS libvmlib_untrusted.a)
 
+if ((WAMR_BUILD_STATIC_PGO EQUAL 1) AND (WAMR_BUILD_AOT EQUAL 1))
+    execute_process(
+        COMMAND bash -c "sed -i -E 's/^WAMR_BUILD_STATIC_PGO = 0/WAMR_BUILD_STATIC_PGO = 1/g' ${CMAKE_CURRENT_SOURCE_DIR}/enclave-sample/Makefile"
+        OUTPUT_VARIABLE cmdOutput
+    )
+else()
+    execute_process(
+        COMMAND bash -c "sed -i -E 's/^WAMR_BUILD_STATIC_PGO = 1/WAMR_BUILD_STATIC_PGO = 0/g' ${CMAKE_CURRENT_SOURCE_DIR}/enclave-sample/Makefile"
+        OUTPUT_VARIABLE cmdOutput
+    )
+endif()
+
 if (DEFINED WAMR_BUILD_GLOBAL_HEAP_POOL)
   execute_process(
       COMMAND bash -c "sed -i -E 's/^WAMR_BUILD_GLOBAL_HEAP_POOL = .*/WAMR_BUILD_GLOBAL_HEAP_POOL = ${WAMR_BUILD_GLOBAL_HEAP_POOL}/g' ${CMAKE_CURRENT_SOURCE_DIR}/enclave-sample/Makefile"

product-mini/platforms/linux-sgx/enclave-sample/App/App.cpp

@@ -232,6 +232,9 @@ print_help()
     printf("                         for example:\n");
     printf("                           --addr-pool=1.2.3.4/15,2.3.4.5/16\n");
     printf("  --max-threads=n        Set maximum thread number per cluster, default is 4\n");
+#if WASM_ENABLE_STATIC_PGO != 0
+    printf("  --gen-prof-file=<path> Generate LLVM PGO (Profile-Guided Optimization) profile file\n");
+#endif
     printf("  --version              Show version information\n");
     return 1;
 }
@@ -294,6 +297,10 @@ typedef enum EcallCmd {
     CMD_SET_WASI_ARGS,        /* wasm_runtime_set_wasi_args() */
     CMD_SET_LOG_LEVEL,        /* bh_log_set_verbose_level() */
     CMD_GET_VERSION,          /* wasm_runtime_get_version() */
+#if WASM_ENABLE_STATIC_PGO != 0
+    CMD_GET_PGO_PROF_BUF_SIZE,  /* wasm_runtime_get_pro_prof_data_size() */
+    CMD_DUMP_PGO_PROF_BUF_DATA, /* wasm_runtime_dump_pgo_prof_data_to_buf() */
+#endif
 } EcallCmd;
 
 static void
@@ -598,6 +605,64 @@ get_version(uint64_t *major, uint64_t *minor, uint64_t *patch)
     *patch = ecall_args[2];
 }
 
+#if WASM_ENABLE_STATIC_PGO != 0
+static void
+dump_pgo_prof_data(void *module_inst, const char *path)
+{
+    char *buf;
+    uint32_t len;
+    FILE *file;
+
+    uint64_t ecall_args[1];
+    ecall_args[0] = (uint64_t)(uintptr_t)module_inst;
+    if (SGX_SUCCESS
+        != ecall_handle_command(g_eid, CMD_GET_PGO_PROF_BUF_SIZE,
+                                (uint8_t *)ecall_args, sizeof(ecall_args))) {
+        printf("Call ecall_handle_command() failed.\n");
+        return;
+    }
+    if (!(len = ecall_args[0])) {
+        printf("failed to get LLVM PGO profile data size\n");
+        return;
+    }
+
+    if (!(buf = (char *)malloc(len))) {
+        printf("allocate memory failed\n");
+        return;
+    }
+
+    uint64_t ecall_args_2[3];
+    ecall_args_2[0] = (uint64_t)(uintptr_t)module_inst;
+    ecall_args_2[1] = (uint64_t)(uintptr_t)buf;
+    ecall_args_2[2] = len;
+    if (SGX_SUCCESS
+        != ecall_handle_command(g_eid, CMD_DUMP_PGO_PROF_BUF_DATA,
+                                (uint8_t *)ecall_args_2,
+                                sizeof(ecall_args_2))) {
+        printf("Call ecall_handle_command() failed.\n");
+        free(buf);
+        return;
+    }
+    if (!(len = ecall_args_2[0])) {
+        printf("failed to dump LLVM PGO profile data\n");
+        free(buf);
+        return;
+    }
+
+    if (!(file = fopen(path, "wb"))) {
+        printf("failed to create file %s", path);
+        free(buf);
+        return;
+    }
+    fwrite(buf, len, 1, file);
+    fclose(file);
+
+    free(buf);
+
+    printf("LLVM raw profile file %s was generated.\n", path);
+}
+#endif
+
 int
 main(int argc, char *argv[])
 {
@@ -619,6 +684,9 @@ main(int argc, char *argv[])
     const char *addr_pool[8] = { NULL };
     uint32_t addr_pool_size = 0;
     uint32_t max_thread_num = 4;
+#if WASM_ENABLE_STATIC_PGO != 0
+    const char *gen_prof_file = NULL;
+#endif
 
     if (enclave_init(&g_eid) < 0) {
         std::cout << "Fail to initialize enclave." << std::endl;
@@ -718,6 +786,13 @@ main(int argc, char *argv[])
                 return print_help();
             max_thread_num = atoi(argv[0] + 14);
         }
+#if WASM_ENABLE_STATIC_PGO != 0
+        else if (!strncmp(argv[0], "--gen-prof-file=", 16)) {
+            if (argv[0][16] == '\0')
+                return print_help();
+            gen_prof_file = argv[0] + 16;
+        }
+#endif
         else if (!strncmp(argv[0], "--version", 9)) {
             uint64_t major = 0, minor = 0, patch = 0;
             get_version(&major, &minor, &patch);
@@ -779,6 +854,11 @@ main(int argc, char *argv[])
     else
         app_instance_main(wasm_module_inst, argc, argv);
 
+#if WASM_ENABLE_STATIC_PGO != 0
+    if (gen_prof_file)
+        dump_pgo_prof_data(wasm_module_inst, gen_prof_file);
+#endif
+
     ret = 0;
 
     /* Deinstantiate module */
@@ -836,7 +916,7 @@ wamr_pal_create_process(struct wamr_pal_create_process_args *args)
     int stdoutfd = -1;
     int stderrfd = -1;
 
-    int argc = 2;
+    const int argc = 2;
     char *argv[argc] = { (char *)"./iwasm", (char *)args->argv[0] };
 
     uint8_t *wasm_files_buf = NULL;

product-mini/platforms/linux-sgx/enclave-sample/Enclave/Enclave.cpp

@@ -49,6 +49,10 @@ typedef enum EcallCmd {
     CMD_SET_WASI_ARGS,        /* wasm_runtime_set_wasi_args() */
     CMD_SET_LOG_LEVEL,        /* bh_log_set_verbose_level() */
     CMD_GET_VERSION,          /* wasm_runtime_get_version() */
+#if WASM_ENABLE_STATIC_PGO != 0
+    CMD_GET_PGO_PROF_BUF_SIZE,  /* wasm_runtime_get_pro_prof_data_size() */
+    CMD_DUMP_PGO_PROF_BUF_DATA, /* wasm_runtime_dump_pgo_prof_data_to_buf() */
+#endif
 } EcallCmd;
 
 typedef struct EnclaveModule {
@@ -597,6 +601,36 @@ handle_cmd_get_version(uint64 *args, uint32 argc)
     args[2] = patch;
 }
 
+#if WASM_ENABLE_STATIC_PGO != 0
+static void
+handle_cmd_get_pgo_prof_buf_size(uint64 *args, int32 argc)
+{
+    wasm_module_inst_t module_inst = *(wasm_module_inst_t *)args;
+    uint32 buf_len;
+
+    bh_assert(argc == 1);
+
+    buf_len = wasm_runtime_get_pgo_prof_data_size(module_inst);
+    args[0] = buf_len;
+}
+
+static void
+handle_cmd_get_pro_prof_buf_data(uint64 *args, int32 argc)
+{
+    uint64 *args_org = args;
+    wasm_module_inst_t module_inst = *(wasm_module_inst_t *)args++;
+    char *buf = *(char **)args++;
+    uint32 len = *(uint32 *)args++;
+    uint32 bytes_dumped;
+
+    bh_assert(argc == 3);
+
+    bytes_dumped =
+        wasm_runtime_dump_pgo_prof_data_to_buf(module_inst, buf, len);
+    args_org[0] = bytes_dumped;
+}
+#endif
+
 void
 ecall_handle_command(unsigned cmd, unsigned char *cmd_buf,
                      unsigned cmd_buf_size)
@@ -647,6 +681,14 @@ ecall_handle_command(unsigned cmd, unsigned char *cmd_buf,
         case CMD_GET_VERSION:
             handle_cmd_get_version(args, argc);
             break;
+#if WASM_ENABLE_STATIC_PGO != 0
+        case CMD_GET_PGO_PROF_BUF_SIZE:
+            handle_cmd_get_pgo_prof_buf_size(args, argc);
+            break;
+        case CMD_DUMP_PGO_PROF_BUF_DATA:
+            handle_cmd_get_pro_prof_buf_data(args, argc);
+            break;
+#endif
         default:
             LOG_ERROR("Unknown command %d\n", cmd);
             break;

product-mini/platforms/linux-sgx/enclave-sample/Makefile

@@ -15,6 +15,7 @@ WAMR_BUILD_SGX_IPFS = 0
 WAMR_BUILD_LIB_RATS = 0
 WAMR_BUILD_GLOBAL_HEAP_POOL = 0
 WAMR_BUILD_GLOBAL_HEAP_SIZE = 10485760
+WAMR_BUILD_STATIC_PGO = 0
 
 VMLIB_BUILD_DIR ?= $(CURDIR)/../build
 LIB_RATS_SRC ?= $(VMLIB_BUILD_DIR)/_deps/librats-build
@@ -65,7 +66,7 @@ ifeq ($(WAMR_BUILD_LIB_RATS), 1)
 	App_Include_Paths += -I$(LIB_RATS_INCLUDE_DIR)
 endif
 
-App_C_Flags := $(SGX_COMMON_CFLAGS) -fPIC -Wno-attributes $(App_Include_Paths)
+App_C_Flags := $(SGX_COMMON_CFLAGS) -fPIC -Wno-attributes $(App_Include_Paths) -DWASM_ENABLE_STATIC_PGO=$(WAMR_BUILD_STATIC_PGO)
 
 # Three configuration modes - Debug, prerelease, release
 #   Debug - Macro DEBUG enabled.
@@ -134,7 +135,7 @@ ifeq ($(WAMR_BUILD_LIB_RATS), 1)
 	Enclave_Include_Paths += -I$(LIB_RATS_INCLUDE_DIR) -I$(SGX_SSL)/include
 endif
 
-Enclave_C_Flags := $(SGX_COMMON_CFLAGS) -nostdinc -fvisibility=hidden -fpie -fstack-protector $(Enclave_Include_Paths) -DWASM_GLOBAL_HEAP_SIZE=$(WAMR_BUILD_GLOBAL_HEAP_SIZE) -DWASM_ENABLE_GLOBAL_HEAP_POOL=$(WAMR_BUILD_GLOBAL_HEAP_POOL) -DWASM_ENABLE_LIB_RATS=$(WAMR_BUILD_LIB_RATS)
+Enclave_C_Flags := $(SGX_COMMON_CFLAGS) -nostdinc -fvisibility=hidden -fpie -fstack-protector $(Enclave_Include_Paths) -DWASM_GLOBAL_HEAP_SIZE=$(WAMR_BUILD_GLOBAL_HEAP_SIZE) -DWASM_ENABLE_GLOBAL_HEAP_POOL=$(WAMR_BUILD_GLOBAL_HEAP_POOL) -DWASM_ENABLE_LIB_RATS=$(WAMR_BUILD_LIB_RATS) -DWASM_ENABLE_STATIC_PGO=$(WAMR_BUILD_STATIC_PGO)
 ifeq ($(SPEC_TEST), 1)
 	Enclave_C_Flags += -DWASM_ENABLE_SPEC_TEST=1
 else

tests/benchmarks/coremark/README.md

@@ -19,3 +19,7 @@ And then run `./build.sh` to build the source code, file `coremark.exe`, `corema
 Run `./run.sh` to test the benchmark, the native mode, iwasm aot mode and iwasm interpreter mode will be tested respectively.
 
 Run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled, please refer [here](../README.md#install-llvm-profdata) to install tool `llvm-profdata` and build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`.
+
+- For Linux, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled.
+
+- For Linux-sgx, similarly, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then `make` in the directory `enclave-sample`. And run `./test_pgo.sh --sgx` to test the benchmark.

tests/benchmarks/coremark/test_pgo.sh

@@ -5,8 +5,13 @@
 
 PLATFORM=$(uname -s | tr A-Z a-z)
 
-IWASM="../../../product-mini/platforms/${PLATFORM}/build/iwasm"
+if [ "$1" = "--sgx" ] && [ "$PLATFORM" = "linux" ]; then
-WAMRC="../../../wamr-compiler/build/wamrc"
+    IWASM="../../../product-mini/platforms/${PLATFORM}-sgx/enclave-sample/iwasm"
+    WAMRC="../../../wamr-compiler/build/wamrc -sgx"
+else
+    IWASM="../../../product-mini/platforms/${PLATFORM}/build/iwasm"
+    WAMRC="../../../wamr-compiler/build/wamrc"
+fi
 
 if [ ! -e "coremark.wasm" ]; then
     echo "coremark.wasm doesn't exist, please run build.sh first"

tests/benchmarks/dhrystone/test_pgo.sh

@@ -5,8 +5,13 @@
 
 PLATFORM=$(uname -s | tr A-Z a-z)
 
-IWASM="../../../product-mini/platforms/${PLATFORM}/build/iwasm"
+if [ "$1" = "--sgx" ] && [ "$PLATFORM" = "linux" ]; then
-WAMRC="../../../wamr-compiler/build/wamrc"
+    IWASM="../../../product-mini/platforms/${PLATFORM}-sgx/enclave-sample/iwasm"
+    WAMRC="../../../wamr-compiler/build/wamrc -sgx"
+else
+    IWASM="../../../product-mini/platforms/${PLATFORM}/build/iwasm"
+    WAMRC="../../../wamr-compiler/build/wamrc"
+fi
 
 if [ ! -e "dhrystone.wasm" ]; then
     echo "dhrystone.wasm doesn't exist, please run build.sh first"

tests/benchmarks/jetstream/README.md

@@ -29,3 +29,7 @@ And then run `./build.sh` to build the source code, the folder `out` will be cre
 Run `./run_aot.sh` to test the benchmark, the native mode and iwasm aot mode will be tested for each workload, and the file `report.txt` will be generated.
 
 Run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled, please refer [here](../README.md#install-llvm-profdata) to install tool `llvm-profdata` and build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`.
+
+- For Linux, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled.
+
+- For Linux-sgx, similarly, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then `make` in the directory `enclave-sample`. And run `./test_pgo.sh --sgx` to test the benchmark.

tests/benchmarks/jetstream/test_pgo.sh

@@ -9,8 +9,13 @@ REPORT=$CUR_DIR/report.txt
 TIME=/usr/bin/time
 
 PLATFORM=$(uname -s | tr A-Z a-z)
-IWASM_CMD=$CUR_DIR/../../../product-mini/platforms/${PLATFORM}/build/iwasm
+if [ "$1" = "--sgx" ] && [ "$PLATFORM" = "linux" ]; then
-WAMRC_CMD=$CUR_DIR/../../../wamr-compiler/build/wamrc
+    IWASM_CMD="$CUR_DIR/../../../product-mini/platforms/${PLATFORM}-sgx/enclave-sample/iwasm"
+    WAMRC_CMD="$CUR_DIR/../../../wamr-compiler/build/wamrc -sgx"
+else
+    IWASM_CMD="$CUR_DIR/../../../product-mini/platforms/${PLATFORM}/build/iwasm"
+    WAMRC_CMD="$CUR_DIR/../../../wamr-compiler/build/wamrc"
+fi
 
 BENCH_NAME_MAX_LEN=20
 

tests/benchmarks/libsodium/README.md

@@ -18,6 +18,12 @@ And then run `./build.sh` to build the source code, the libsodium source code wi
 
 Run `./run_aot.sh` to test the benchmark, the native mode and iwasm aot mode will be tested respectively.
 
+Run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled, please refer [here](../README.md#install-llvm-profdata) to install tool `llvm-profdata` and build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`.
+
+- For Linux, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled.
+
+- For Linux-sgx, similarly, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then `make` in the directory `enclave-sample`. And run `./test_pgo.sh --sgx` to test the benchmark.
+
 # Others
 
 Refer to [Performance of WebAssembly runtimes in 2023](https://00f.net/2023/01/04/webassembly-benchmark-2023) for more about the performance comparison of wasm runtimes on running the libsodium benchmarks.

tests/benchmarks/libsodium/test_pgo.sh

@@ -19,8 +19,13 @@ PLATFORM=$(uname -s | tr A-Z a-z)
 
 readonly OUT_DIR=$PWD/libsodium/zig-out/bin
 readonly REPORT=$PWD/report.txt
-readonly IWASM_CMD=$PWD/../../../product-mini/platforms/${PLATFORM}/build/iwasm
+if [ "$1" = "--sgx" ] && [ "$PLATFORM" = "linux" ]; then
-readonly WAMRC_CMD=$PWD/../../../wamr-compiler/build/wamrc
+    readonly IWASM_CMD="$PWD/../../../product-mini/platforms/${PLATFORM}-sgx/enclave-sample/iwasm"
+    readonly WAMRC_CMD="$PWD/../../../wamr-compiler/build/wamrc -sgx"
+else
+    readonly IWASM_CMD="$PWD/../../../product-mini/platforms/${PLATFORM}/build/iwasm"
+    readonly WAMRC_CMD="$PWD/../../../wamr-compiler/build/wamrc"
+fi
 readonly TIME=/usr/bin/time
 
 BENCH_NAME_MAX_LEN=20

tests/benchmarks/polybench/README.md

@@ -19,3 +19,9 @@ And then run `./build.sh` to build the source code, the folder `out` will be cre
 Run `./run_aot.sh` to test the benchmark, the native mode and iwasm aot mode will be tested for each workload, and the file `report.txt` will be generated.
 
 Run `./run_interp.sh` to test the benchmark, the native mode and iwasm interpreter mode will be tested for each workload, and the file `report.txt` will be generated.
+
+Run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled, please refer [here](../README.md#install-llvm-profdata) to install tool `llvm-profdata` and build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`.
+
+- For Linux, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled.
+
+- For Linux-sgx, similarly, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then `make` in the directory `enclave-sample`. And run `./test_pgo.sh --sgx` to test the benchmark.

tests/benchmarks/polybench/test_pgo.sh

@@ -9,8 +9,13 @@ REPORT=$CUR_DIR/report.txt
 TIME=/usr/bin/time
 
 PLATFORM=$(uname -s | tr A-Z a-z)
-IWASM_CMD=$CUR_DIR/../../../product-mini/platforms/${PLATFORM}/build/iwasm
+if [ "$1" = "--sgx" ] && [ "$PLATFORM" = "linux" ]; then
-WAMRC_CMD=$CUR_DIR/../../../wamr-compiler/build/wamrc
+    IWASM_CMD="$CUR_DIR/../../../product-mini/platforms/${PLATFORM}-sgx/enclave-sample/iwasm"
+    WAMRC_CMD="$CUR_DIR/../../../wamr-compiler/build/wamrc -sgx"
+else
+    IWASM_CMD="$CUR_DIR/../../../product-mini/platforms/${PLATFORM}/build/iwasm"
+    WAMRC_CMD="$CUR_DIR/../../../wamr-compiler/build/wamrc"
+fi
 
 BENCH_NAME_MAX_LEN=20
 

tests/benchmarks/sightglass/README.md

@@ -21,3 +21,7 @@ Run `./run_aot.sh` to test the benchmark, the native mode and iwasm aot mode wil
 Run `./run_interp.sh` to test the benchmark, the native mode and iwasm interpreter mode will be tested for each workload, and the file `report.txt` will be generated.
 
 Run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled, please refer [here](../README.md#install-llvm-profdata) to install tool `llvm-profdata` and build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`.
+
+- For Linux, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then run `./test_pgo.sh` to test the benchmark with AOT static PGO (Profile-Guided Optimization) enabled.
+
+- For Linux-sgx, similarly, build `iwasm` with `cmake -DWAMR_BUILD_STATIC_PGO=1`, then `make` in the directory `enclave-sample`. And run `./test_pgo.sh --sgx` to test the benchmark.

tests/benchmarks/sightglass/test_pgo.sh

@@ -9,8 +9,13 @@ REPORT=$CUR_DIR/report.txt
 TIME=/usr/bin/time
 
 PLATFORM=$(uname -s | tr A-Z a-z)
-IWASM_CMD=$CUR_DIR/../../../product-mini/platforms/${PLATFORM}/build/iwasm
+if [ "$1" = "--sgx" ] && [ "$PLATFORM" = "linux" ]; then
-WAMRC_CMD=$CUR_DIR/../../../wamr-compiler/build/wamrc
+    IWASM_CMD="$CUR_DIR/../../../product-mini/platforms/${PLATFORM}-sgx/enclave-sample/iwasm"
+    WAMRC_CMD="$CUR_DIR/../../../wamr-compiler/build/wamrc -sgx"
+else
+    IWASM_CMD="$CUR_DIR/../../../product-mini/platforms/${PLATFORM}/build/iwasm"
+    WAMRC_CMD="$CUR_DIR/../../../wamr-compiler/build/wamrc"
+fi
 
 BENCH_NAME_MAX_LEN=20