Move heap/stack tracking to before init phase

Start tracking during the setup of the algo, so we can capture the memory usage of the algo init functions.
2025-10-17 19:27:59 +01:00
parent 036c66c777
commit 65bb68b6c0
1 changed files with 258 additions and 8 deletions
--- a/wolfcrypt/benchmark/benchmark.c
+++ b/wolfcrypt/benchmark/benchmark.c
@@ -342,6 +342,38 @@ static WC_INLINE long bench_stack_checkpoint_measure(void)
 }
 #endif

+#if defined(WC_BENCH_HEAP_TRACKING) || defined(WC_BENCH_STACK_TRACKING)
+static int bench_stats_resources_prepared;
+#ifdef WC_BENCH_HEAP_TRACKING
+static long bench_stats_heap_setup_allocs;
+static long bench_stats_heap_setup_bytes;
+#endif
+#ifdef WC_BENCH_STACK_TRACKING
+static long bench_stats_stack_setup_bytes;
+#endif
+
+static WC_INLINE void bench_stats_reset_setup_offsets(void)
+{
+#ifdef WC_BENCH_HEAP_TRACKING
+    bench_stats_heap_setup_allocs = 0;
+    bench_stats_heap_setup_bytes  = 0;
+#endif
+#ifdef WC_BENCH_STACK_TRACKING
+    bench_stats_stack_setup_bytes = 0;
+#endif
+}
+#endif
+
+static WC_INLINE void bench_stats_prepare(void)
+{
+    bench_heap_checkpoint_prepare();
+    bench_stack_checkpoint_prepare();
+#if defined(WC_BENCH_HEAP_TRACKING) || defined(WC_BENCH_STACK_TRACKING)
+    bench_stats_reset_setup_offsets();
+    bench_stats_resources_prepared = 1;
+#endif
+}
+
 #if defined(WC_BENCH_HEAP_TRACKING) || defined(WC_BENCH_STACK_TRACKING)
 static WC_INLINE void bench_append_memory_info(char* buffer, size_t size,
    int asCsv, long heapAllocs, long heapBytes, long stackBytes)
@@ -2526,6 +2558,10 @@ static WC_INLINE void bench_stats_init(void)
 #ifdef WC_BENCH_TRACK_STATS
    bench_stats_head = NULL;
    bench_stats_tail = NULL;
+#endif
+#if defined(WC_BENCH_HEAP_TRACKING) || defined(WC_BENCH_STACK_TRACKING)
+    bench_stats_reset_setup_offsets();
+    bench_stats_resources_prepared = 0;
 #endif
    INIT_CYCLE_COUNTER
 }
@@ -2535,8 +2571,35 @@ static WC_INLINE void bench_stats_start(int* count, double* start)
    *count = 0;
    *start = current_time(1);

+#if defined(WC_BENCH_HEAP_TRACKING) || defined(WC_BENCH_STACK_TRACKING)
+    if (bench_stats_resources_prepared) {
+#ifdef WC_BENCH_HEAP_TRACKING
+        long prepAllocs = 0;
+        long prepBytes  = 0;
+        bench_heap_checkpoint_measure(&prepAllocs, &prepBytes);
+        bench_stats_heap_setup_allocs += prepAllocs;
+        bench_stats_heap_setup_bytes  += prepBytes;
+#else
+        (void)bench_heap_checkpoint_measure(NULL, NULL);
+#endif
+#ifdef WC_BENCH_STACK_TRACKING
+        bench_stats_stack_setup_bytes += bench_stack_checkpoint_measure();
+#else
+        (void)bench_stack_checkpoint_measure();
+#endif
+        bench_heap_checkpoint_prepare();
+        bench_stack_checkpoint_prepare();
+    }
+    else {
+        bench_heap_checkpoint_prepare();
+        bench_stack_checkpoint_prepare();
+        bench_stats_reset_setup_offsets();
+    }
+    bench_stats_resources_prepared = 0;
+#else
    bench_heap_checkpoint_prepare();
    bench_stack_checkpoint_prepare();
+#endif

 #ifdef WOLFSSL_ESPIDF
    #ifdef DEBUG_WOLFSSL_BENCHMARK_TIMING
@@ -2869,8 +2932,19 @@ static void bench_stats_sym_finish(const char* desc, int useDeviceID,
        NULL
 #endif
    );
+#ifdef WC_BENCH_HEAP_TRACKING
+    heapAllocs += bench_stats_heap_setup_allocs;
+    heapBytes  += bench_stats_heap_setup_bytes;
+    bench_stats_heap_setup_allocs = 0;
+    bench_stats_heap_setup_bytes  = 0;
+#endif
 #ifdef WC_BENCH_STACK_TRACKING
-    bench_stack_checkpoint_measure();
+    {
+        long stackUsed = bench_stack_checkpoint_measure();
+        stackUsed += bench_stats_stack_setup_bytes;
+        bench_last_stack_bytes = stackUsed;
+    }
+    bench_stats_stack_setup_bytes = 0;
 #else
    (void)bench_stack_checkpoint_measure();
 #endif
@@ -3023,6 +3097,7 @@ static void bench_stats_sym_finish(const char* desc, int useDeviceID,
    kernel_fpu_end();
 #endif

+    bench_stats_prepare();
    TEST_SLEEP();
 } /* bench_stats_sym_finish */

@@ -3099,7 +3174,22 @@ static void bench_stats_asym_finish_ex(const char* algo, int strength,
        NULL
 #endif
    );
-    bench_stack_checkpoint_measure();
+#ifdef WC_BENCH_HEAP_TRACKING
+    heapAllocs += bench_stats_heap_setup_allocs;
+    heapBytes  += bench_stats_heap_setup_bytes;
+    bench_stats_heap_setup_allocs = 0;
+    bench_stats_heap_setup_bytes  = 0;
+#endif
+#ifdef WC_BENCH_STACK_TRACKING
+    {
+        long stackUsed = bench_stack_checkpoint_measure();
+        stackUsed += bench_stats_stack_setup_bytes;
+        bench_last_stack_bytes = stackUsed;
+    }
+    bench_stats_stack_setup_bytes = 0;
+#else
+    (void)bench_stack_checkpoint_measure();
+#endif

    SLEEP_ON_ERROR(ret);

@@ -3361,6 +3451,7 @@ static void bench_stats_asym_finish_ex(const char* algo, int strength,
    kernel_fpu_end();
 #endif

+    bench_stats_prepare();
    TEST_SLEEP();
 } /* bench_stats_asym_finish_ex */

@@ -4637,6 +4728,8 @@ void bench_rng(void)
    WC_RNG myrng;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
 #ifndef HAVE_FIPS
    ret = wc_InitRng_ex(&myrng, HEAP_HINT, devId);
 #else
@@ -4700,6 +4793,8 @@ static void bench_aescbc_internal(int useDeviceID,
    double start;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(enc, Aes, BENCH_MAX_PENDING,
                     sizeof(Aes), HEAP_HINT);

@@ -4863,6 +4958,8 @@ static void bench_aesgcm_internal(int useDeviceID,
    WC_DECLARE_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_DECLARE_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_ALLOC_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);
    WC_CALLOC_ARRAY(enc, Aes, BENCH_MAX_PENDING,
@@ -5045,6 +5142,8 @@ static void bench_aesgcm_stream_internal(int useDeviceID,
    WC_DECLARE_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_DECLARE_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_ALLOC_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);

@@ -5274,6 +5373,8 @@ void bench_gmac(int useDeviceID)
    const char* gmacStr = "GMAC Default";
 #endif

+    bench_stats_prepare();
+
 /* Implementations of /Dev/Crypto will error out if the size of Auth in is */
 /* greater than the system's page size */
 #if defined(WOLFSSL_DEVCRYPTO) && defined(WOLFSSL_AUTHSZ_BENCH)
@@ -5340,6 +5441,8 @@ static void bench_aesecb_internal(int useDeviceID,
    const word32 benchSz = bench_size;
 #endif

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(enc, Aes, BENCH_MAX_PENDING,
                     sizeof(Aes), HEAP_HINT);

@@ -5503,6 +5606,8 @@ static void bench_aescfb_internal(const byte* key,
    int    i, ret, count;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    ret = wc_AesInit(&enc, HEAP_HINT, INVALID_DEVID);
    if (ret != 0) {
        printf("AesInit failed at L%d, ret = %d\n", __LINE__, ret);
@@ -5568,6 +5673,8 @@ static void bench_aesofb_internal(const byte* key,
    int    i, ret, count;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    ret = wc_AesInit(&enc, NULL, INVALID_DEVID);
    if (ret != 0) {
        printf("AesInit failed at L%d, ret = %d\n", __LINE__, ret);
@@ -5628,6 +5735,8 @@ void bench_aesxts(void)
    int    i, count, ret;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    static const unsigned char k1[] = {
        0xa1, 0xb9, 0x0c, 0xba, 0x3f, 0x06, 0xac, 0x35,
        0x3b, 0x2c, 0x34, 0x38, 0x76, 0x08, 0x17, 0x62,
@@ -5724,6 +5833,8 @@ static void bench_aesctr_internal(const byte* key, word32 keySz,
    int    i, count, ret = 0;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    if ((ret = wc_AesInit(&enc, HEAP_HINT,
        useDeviceID ? devId : INVALID_DEVID)) != 0) {
        printf("wc_AesInit failed, ret = %d\n", ret);
@@ -5786,6 +5897,8 @@ void bench_aesccm(int useDeviceID)
    WC_DECLARE_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_DECLARE_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_ALLOC_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);

@@ -5883,6 +5996,8 @@ static void bench_aessiv_internal(const byte* key, word32 keySz, const char*
    double start;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    bench_stats_start(&count, &start);
    do {
        for (i = 0; i < numBlocks; i++) {
@@ -5953,6 +6068,8 @@ void bench_poly1305(void)
    int      ret = 0, i, count;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    if (digest_stream) {
        ret = wc_Poly1305SetKey(&enc, bench_key, 32);
        if (ret != 0) {
@@ -6018,6 +6135,8 @@ void bench_camellia(void)
    int      ret, i, count;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    ret = wc_CamelliaSetKey(&cam, bench_key, 16, bench_iv);
    if (ret != 0) {
        printf("CamelliaSetKey failed, ret = %d\n", ret);
@@ -6059,6 +6178,8 @@ void bench_sm4_cbc(void)
    int    count;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    ret = wc_Sm4SetKey(&sm4, bench_key, SM4_KEY_SIZE);
    if (ret != 0) {
        printf("Sm4SetKey failed, ret = %d\n", ret);
@@ -6131,6 +6252,8 @@ void bench_sm4_gcm(void)
    WC_DECLARE_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_DECLARE_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_ALLOC_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);

@@ -6208,6 +6331,8 @@ void bench_sm4_ccm(void)
    WC_DECLARE_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_DECLARE_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_ALLOC_VAR(bench_tag, byte, AES_AUTH_TAG_SZ, HEAP_HINT);

@@ -6284,6 +6409,8 @@ void bench_des(int useDeviceID)
    double start;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(enc, Des3, BENCH_MAX_PENDING,
                     sizeof(Des3), HEAP_HINT);

@@ -6356,6 +6483,8 @@ void bench_arc4(int useDeviceID)
    double start;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(enc, Arc4, BENCH_MAX_PENDING,
                     sizeof(Arc4), HEAP_HINT);

@@ -6426,6 +6555,8 @@ void bench_chacha(void)
    int    ret, i, count;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(enc, ChaCha, 1, HEAP_HINT);

    XMEMSET(enc, 0, sizeof(ChaCha));
@@ -6499,6 +6630,9 @@ void bench_chacha20_poly1305_aead(void)

    WC_DECLARE_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_DECLARE_VAR(authTag, byte, CHACHA20_POLY1305_AEAD_AUTHTAG_SIZE, HEAP_HINT);
+
+    bench_stats_prepare();
+
    WC_ALLOC_VAR(bench_additional, byte, AES_AUTH_ADD_SZ, HEAP_HINT);
    WC_ALLOC_VAR(authTag, byte, CHACHA20_POLY1305_AEAD_AUTHTAG_SIZE, HEAP_HINT);
    XMEMSET(bench_additional, 0, AES_AUTH_ADD_SZ);
@@ -6547,6 +6681,8 @@ void bench_ascon_aead(void)
    WC_DECLARE_VAR(enc, wc_AsconAEAD128, 1, HEAP_HINT);
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(authTag, byte, ASCON_AEAD128_TAG_SZ, HEAP_HINT);
    XMEMSET(authTag, 0, ASCON_AEAD128_TAG_SZ);

@@ -6608,6 +6744,8 @@ void bench_md5(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_MD5_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Md5, BENCH_MAX_PENDING,
                     sizeof(wc_Md5), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -6724,6 +6862,8 @@ void bench_sha(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha, BENCH_MAX_PENDING,
                     sizeof(wc_Sha), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -6838,6 +6978,8 @@ void bench_sha224(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA224_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha224, BENCH_MAX_PENDING,
                     sizeof(wc_Sha224), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -6948,6 +7090,8 @@ void bench_sha256(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA256_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha256, BENCH_MAX_PENDING,
                     sizeof(wc_Sha256), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -7059,6 +7203,8 @@ void bench_sha384(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA384_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha384, BENCH_MAX_PENDING,
                     sizeof(wc_Sha384), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -7169,6 +7315,8 @@ void bench_sha512(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA512_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha512, BENCH_MAX_PENDING,
                     sizeof(wc_Sha512), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -7279,6 +7427,8 @@ void bench_sha512_224(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA512_224_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha512_224, BENCH_MAX_PENDING,
                     sizeof(wc_Sha512_224), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -7390,6 +7540,8 @@ void bench_sha512_256(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA512_256_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha512_256, BENCH_MAX_PENDING,
                     sizeof(wc_Sha512_256), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -7504,6 +7656,8 @@ void bench_sha3_224(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA3_224_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha3, BENCH_MAX_PENDING,
                     sizeof(wc_Sha3), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -7614,6 +7768,8 @@ void bench_sha3_256(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA3_256_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha3, BENCH_MAX_PENDING,
                     sizeof(wc_Sha3), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -7724,6 +7880,8 @@ void bench_sha3_384(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA3_384_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha3, BENCH_MAX_PENDING,
                     sizeof(wc_Sha3), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -7834,6 +7992,8 @@ void bench_sha3_512(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA3_512_DIGEST_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sha3, BENCH_MAX_PENDING,
                     sizeof(wc_Sha3), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -7944,6 +8104,8 @@ void bench_shake128(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA3_128_BLOCK_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Shake, BENCH_MAX_PENDING,
                     sizeof(wc_Shake), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -8056,6 +8218,8 @@ void bench_shake256(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING,
                     WC_SHA3_256_BLOCK_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Shake, BENCH_MAX_PENDING,
                     sizeof(wc_Shake), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING,
@@ -8169,6 +8333,8 @@ void bench_sm3(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING, WC_SM3_DIGEST_SIZE,
        HEAP_HINT);

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hash, wc_Sm3, BENCH_MAX_PENDING,
                     sizeof(wc_Sm3), HEAP_HINT);
    WC_ALLOC_ARRAY(digest, byte, BENCH_MAX_PENDING, WC_SM3_DIGEST_SIZE,
@@ -8275,6 +8441,8 @@ void bench_ripemd(void)
    int    i, count, ret = 0;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    if (digest_stream) {
        ret = wc_InitRipeMd(&hash);
        if (ret != 0) {
@@ -8352,6 +8520,8 @@ void bench_blake2b(void)
    int     ret = 0, i, count;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    if (digest_stream) {
        ret = wc_InitBlake2b(&b2b, 64);
        if (ret != 0) {
@@ -8425,6 +8595,8 @@ void bench_blake2s(void)
    int     ret = 0, i, count;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    if (digest_stream) {
        ret = wc_InitBlake2s(&b2s, 32);
        if (ret != 0) {
@@ -8497,6 +8669,8 @@ void bench_ascon_hash(void)
    double  start;
    int     ret = 0, i, count;

+    bench_stats_prepare();
+
    if (digest_stream) {
        ret = wc_AsconHash256_Init(&ascon);
        if (ret != 0) {
@@ -8579,6 +8753,8 @@ static void bench_cmac_helper(word32 keySz, const char* outMsg, int useDeviceID)
 #endif
    (void)useDeviceID;

+    bench_stats_prepare();
+
    bench_stats_start(&count, &start);
    do {
    #ifdef HAVE_FIPS
@@ -8646,6 +8822,8 @@ void bench_scrypt(void)
    int    ret, i, count;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    bench_stats_start(&count, &start);
    do {
        for (i = 0; i < scryptCnt; i++) {
@@ -8695,6 +8873,8 @@ static void bench_hmac(int useDeviceID, int type, int digestSz,

    (void)digestSz;

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(hmac, Hmac, BENCH_MAX_PENDING,
                     sizeof(Hmac), HEAP_HINT);

@@ -8930,6 +9110,8 @@ void bench_pbkdf2(void)
    byte derived[32];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    bench_stats_start(&count, &start);
    PRIVATE_KEY_UNLOCK();
    do {
@@ -8963,6 +9145,8 @@ void bench_siphash(void)
    int    i;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    bench_stats_start(&count, &start);
    do {
        for (i = 0; i < numBlocks; i++) {
@@ -9022,6 +9206,8 @@ void bench_srtpkdf(void)
    int i;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    bench_stats_start(&count, &start);
    PRIVATE_KEY_UNLOCK();
    do {
@@ -9128,6 +9314,8 @@ static void bench_rsaKeyGen_helper(int useDeviceID, word32 keySz)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(genKey, RsaKey, BENCH_MAX_PENDING,
                     sizeof(RsaKey), HEAP_HINT);

@@ -9363,6 +9551,8 @@ static void bench_rsa_helper(int useDeviceID,
    byte* out[BENCH_MAX_PENDING];
 #endif

+    bench_stats_prepare();
+
 #if !defined(WOLFSSL_RSA_VERIFY_ONLY) || !defined(WOLFSSL_RSA_PUBLIC_ONLY)
    WC_CALLOC_HEAP_ARRAY(out, byte, BENCH_MAX_PENDING,
                                    rsaKeySz/8, HEAP_HINT);
@@ -9819,6 +10009,8 @@ void bench_dh(int useDeviceID)
    WC_DECLARE_VAR(priv2, byte,
                     BENCH_DH_PRIV_SIZE, HEAP_HINT);

+    bench_stats_prepare();
+
    /* old scan-build misfires -Wmaybe-uninitialized on these. */
    XMEMSET(pub, 0, sizeof(pub));
    XMEMSET(agree, 0, sizeof(agree));
@@ -10048,6 +10240,8 @@ static void bench_mlkem_keygen(int type, const char* name, int keySize,
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    /* MLKEM Make Key */
    bench_stats_start(&count, &start);
    do {
@@ -10105,6 +10299,8 @@ static void bench_mlkem_encap(int type, const char* name, int keySize,
    word32 ctSz;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    ret = wc_KyberKey_PublicKeySize(key1, &pubLen);
    if (ret != 0) {
        return;
@@ -10467,6 +10663,8 @@ static void bench_lms_keygen(enum wc_LmsParm parm, byte* pub)
    const char* str = wc_LmsKey_ParmToStr(parm);
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
 #ifndef HAVE_FIPS
    ret = wc_InitRng_ex(&rng, HEAP_HINT, INVALID_DEVID);
 #else
@@ -10580,6 +10778,8 @@ static void bench_lms_sign_verify(enum wc_LmsParm parm, byte* pub)
    const char * str = wc_LmsKey_ParmToStr(parm);
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    ret = wc_LmsKey_Init(&key, NULL, INVALID_DEVID);
    if (ret) {
        printf("wc_LmsKey_Init failed: %d\n", ret);
@@ -10997,6 +11197,8 @@ static void bench_xmss_sign_verify(const char * params)
    int             count = 0;
    double          start = 0.0F;

+    bench_stats_prepare();
+
 #ifndef HAVE_FIPS
    ret = wc_InitRng_ex(&rng, HEAP_HINT, INVALID_DEVID);
 #else
@@ -11508,6 +11710,8 @@ void bench_eccMakeKey(int useDeviceID, int curveId)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_CALLOC_ARRAY(genKey, ecc_key, BENCH_MAX_PENDING,
                     sizeof(ecc_key), HEAP_HINT);

@@ -11605,6 +11809,8 @@ void bench_ecc(int useDeviceID, int curveId)
                     BENCH_MAX_PENDING, MAX_ECC_BYTES, HEAP_HINT);
 #endif

+    bench_stats_prepare();
+
 #if !defined(NO_ASN) && defined(HAVE_ECC_SIGN)
    /* old scan-build misfires -Wmaybe-uninitialized on these. */
    XMEMSET(sig, 0, sizeof(sig));
@@ -11883,6 +12089,8 @@ void bench_eccEncrypt(int curveId)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
 #ifdef WOLFSSL_SMALL_STACK
    userA = (ecc_key *)XMALLOC(sizeof(*userA),
                               HEAP_HINT, DYNAMIC_TYPE_TMP_BUFFER);
@@ -12032,6 +12240,8 @@ static void bench_sm2_MakeKey(int useDeviceID)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    deviceID = useDeviceID ? devId : INVALID_DEVID;
    keySize = wc_ecc_get_curve_size_from_id(ECC_SM2P256V1);

@@ -12122,6 +12332,8 @@ void bench_sm2(int useDeviceID)
    WC_DECLARE_ARRAY(digest, byte, BENCH_MAX_PENDING, MAX_ECC_BYTES, HEAP_HINT);
 #endif

+    bench_stats_prepare();
+
 #ifdef HAVE_ECC_DHE
    WC_ALLOC_ARRAY(shared, byte, BENCH_MAX_PENDING, MAX_ECC_BYTES, HEAP_HINT);
 #endif
@@ -12368,6 +12580,8 @@ void bench_curve25519KeyGen(int useDeviceID)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    /* Key Gen */
    bench_stats_start(&count, &start);
    do {
@@ -12412,6 +12626,8 @@ void bench_curve25519KeyAgree(int useDeviceID)
    word32 x = 0;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    wc_curve25519_init_ex(&genKey,  HEAP_HINT,
        useDeviceID ? devId : INVALID_DEVID);
    wc_curve25519_init_ex(&genKey2, HEAP_HINT,
@@ -12430,18 +12646,18 @@ void bench_curve25519KeyAgree(int useDeviceID)
    }

 #ifdef WOLFSSL_CURVE25519_BLINDING
-   ret = wc_curve25519_set_rng(&genKey, &gRng);
-   if (ret != 0) {
+    ret = wc_curve25519_set_rng(&genKey, &gRng);
+    if (ret != 0) {
        wc_curve25519_free(&genKey);
        wc_curve25519_free(&genKey2);
        return;
-   }
-   ret = wc_curve25519_set_rng(&genKey2, &gRng);
-   if (ret != 0) {
+    }
+    ret = wc_curve25519_set_rng(&genKey2, &gRng);
+    if (ret != 0) {
        wc_curve25519_free(&genKey);
        wc_curve25519_free(&genKey2);
        return;
-   }
+    }
 #endif

    /* Shared secret */
@@ -12486,6 +12702,8 @@ void bench_ed25519KeyGen(void)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    /* Key Gen */
    bench_stats_start(&count, &start);
    do {
@@ -12526,6 +12744,8 @@ void bench_ed25519KeySign(void)
    DECLARE_MULTI_VALUE_STATS_VARS()
 #endif

+    bench_stats_prepare();
+
    wc_ed25519_init(&genKey);

 #ifdef HAVE_ED25519_MAKE_KEY
@@ -12608,6 +12828,8 @@ void bench_curve448KeyGen(void)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    /* Key Gen */
    bench_stats_start(&count, &start);
    do {
@@ -12644,6 +12866,8 @@ void bench_curve448KeyAgree(void)
    word32 x = 0;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    wc_curve448_init(&genKey);
    wc_curve448_init(&genKey2);

@@ -12699,6 +12923,8 @@ void bench_ed448KeyGen(void)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    /* Key Gen */
    bench_stats_start(&count, &start);
    do {
@@ -12735,6 +12961,8 @@ void bench_ed448KeySign(void)
    DECLARE_MULTI_VALUE_STATS_VARS()
 #endif

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(genKey, ed448_key, 1, HEAP_HINT);

    wc_ed448_init(genKey);
@@ -12821,6 +13049,8 @@ void bench_eccsiKeyGen(void)
    int    ret;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(genKey, EccsiKey, 1, HEAP_HINT);

    /* Key Gen */
@@ -12865,6 +13095,8 @@ void bench_eccsiPairGen(void)
    int ret;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(genKey, EccsiKey, 1, HEAP_HINT);
    WC_ALLOC_VAR(ssk, mp_int, 1, HEAP_HINT);

@@ -12922,6 +13154,8 @@ void bench_eccsiValidate(void)
    int ret;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(genKey, EccsiKey, 1, HEAP_HINT);
    WC_ALLOC_VAR(ssk, mp_int, 1, HEAP_HINT);

@@ -12985,6 +13219,8 @@ void bench_eccsi(void)
    int verified;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(genKey, EccsiKey, 1, HEAP_HINT);
    WC_ALLOC_VAR(ssk, mp_int, 1, HEAP_HINT);

@@ -13073,6 +13309,8 @@ void bench_sakkeKeyGen(void)
    int    ret;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(genKey, SakkeKey, 1, HEAP_HINT);

    /* Key Gen */
@@ -13116,6 +13354,8 @@ void bench_sakkeRskGen(void)
    int ret;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(genKey, SakkeKey, 1, HEAP_HINT);

    rsk = wc_ecc_new_point();
@@ -13167,6 +13407,8 @@ void bench_sakkeValidate(void)
    int ret;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(genKey, SakkeKey, 1, HEAP_HINT);

    rsk = wc_ecc_new_point();
@@ -13227,6 +13469,8 @@ void bench_sakke(void)
    word32 iTableLen = 0;
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    WC_ALLOC_VAR(genKey, SakkeKey, 1, HEAP_HINT);

    XMEMCPY(ssv, ssv_init, sizeof ssv);
@@ -13457,6 +13701,8 @@ void bench_falconKeySign(byte level)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    ret = wc_falcon_init(&key);
    if (ret != 0) {
        printf("wc_falcon_init failed %d\n", ret);
@@ -14653,6 +14899,8 @@ void bench_dilithiumKeySign(byte level)
    DECLARE_MULTI_VALUE_STATS_VARS()
    byte params = 0;

+    bench_stats_prepare();
+
 #ifdef WOLFSSL_SMALL_STACK
    key = (dilithium_key *)XMALLOC(sizeof(*key), HEAP_HINT,
        DYNAMIC_TYPE_TMP_BUFFER);
@@ -14918,6 +15166,8 @@ void bench_sphincsKeySign(byte level, byte optim)
    const char**desc = bench_desc_words[lng_index];
    DECLARE_MULTI_VALUE_STATS_VARS()

+    bench_stats_prepare();
+
    ret = wc_sphincs_init(&key);
    if (ret != 0) {
        printf("wc_sphincs_init failed %d\n", ret);