diff --git a/backend.h b/backend.h
index 6e050f1..71e4b7c 100644
--- a/backend.h
+++ b/backend.h
@@ -75,15 +75,18 @@ arch_addr_t get_instruction_pointer(struct task *task);
 void set_instruction_pointer(struct task *task, arch_addr_t addr);
 
 /* do a single step */
-int do_singlestep(struct task *task);
+int do_singlestep(struct task *task, struct  breakpoint *bp);
 
 /* handle a single step event */
-int handle_singlestep(struct task *task, int (*singlestep)(struct task *task));
+int handle_singlestep(struct task *task, int (*singlestep)(struct task *task), struct breakpoint *bp);
 
 /* Find and return caller address, i.e. the address where the current
  * function returns.  */
 arch_addr_t get_return_addr(struct task *task);
 
+/* get address of IP register */
+unsigned int ip_reg_addr(void);
+
 #if HW_BREAKPOINTS > 0
 /* returns true if the hw breakpoint is pendig */
 int get_hw_bp_state(struct task *task, unsigned int n);
diff --git a/breakpoint.c b/breakpoint.c
index aae569d..43588e9 100644
--- a/breakpoint.c
+++ b/breakpoint.c
@@ -291,6 +291,9 @@ void breakpoint_hw_destroy(struct task *task)
 {
 	unsigned int i;
 
+	if (options.nohwbp)
+		return;
+
 	for(i = 0; i < HW_BREAKPOINTS; ++i)
 		task->hw_bp[i] = NULL;
 
diff --git a/client/process.c b/client/process.c
index 334e4e6..25dcb8d 100644
--- a/client/process.c
+++ b/client/process.c
@@ -680,7 +680,7 @@ static int process_rb_insert_block(struct process *process, unsigned long addr,
 		if ((addr <= this->addr) && (addr + n > this->addr)) {
 			process_dump_collision(process, this, addr, size, operation);
 
-			if (options.kill)
+			if (unlikely(options.kill))
 				abort();
 		}
 
@@ -1262,7 +1262,7 @@ void process_munmap(struct process *process, struct mt_msg *mt_msg, void *payloa
 		if (!is_mmap(block->stack_node->stack->operation)) {
 			fprintf(stderr, ">>> block missmatch pid:%d MAP<>MALLOC %#lx\n", process->pid, ptr);
 
-			if (options.kill)
+			if (unlikely(options.kill))
 				abort();
 
 			break;
@@ -1387,17 +1387,10 @@ void process_free(struct process *process, struct mt_msg *mt_msg, void *payload)
 
 	block = process_rb_search(&process->block_table, ptr);
 	if (block) {
-		if (block->addr != ptr) {
-			fprintf(stderr, ">>> block invalid free %#lx pid:%d\n", ptr, process->pid);
-
-			if (options.kill)
-				abort();
-		}
-
 		if (is_mmap(block->stack_node->stack->operation)) {
 			fprintf(stderr, ">>> block missmatch pid:%d MAP<>MALLOC %#lx\n", process->pid, ptr);
 
-			if (options.kill)
+			if (unlikely(options.kill))
 				abort();
 		}
 
@@ -1427,10 +1420,10 @@ void process_free(struct process *process, struct mt_msg *mt_msg, void *payload)
 	}
 	else {
 		if (!process->attached) {
-			if (unlikely(options.kill)) {
-				fprintf(stderr, ">>> block %#lx not found pid:%d\n", ptr, process->pid);
+			fprintf(stderr, ">>> block %#lx not found pid:%d\n", ptr, process->pid);
+
+			if (unlikely(options.kill))
 				abort();
-			}
 		}
 	}
 }
@@ -1438,7 +1431,7 @@ void process_free(struct process *process, struct mt_msg *mt_msg, void *payload)
 void process_realloc_done(struct process *process, struct mt_msg *mt_msg, void *payload)
 {
 	unsigned long ptr;
-	unsigned long pid;
+	unsigned int pid;
 	struct list_head *it;
 
 	if (!process->tracing)
@@ -1467,11 +1460,11 @@ void process_realloc_done(struct process *process, struct mt_msg *mt_msg, void *
 				process_rb_insert_block(process, re->addr, re->size, re->stack, re->flags, re->operation);
 
 			realloc_del(re);
-
 			return;
 		}
 	}
 
+//	fprintf(stderr, ">>> unexpected realloc done pid: %u\n", pid);
 	return;
 }
 
@@ -1517,7 +1510,7 @@ void process_alloc(struct process *process, struct mt_msg *mt_msg, void *payload
 
 		process_dump_collision(process, block, ptr, size, mt_msg->operation);
 
-		if (options.kill)
+		if (unlikely(options.kill))
 			abort();
 
 		process_rb_delete_block(process, block);
diff --git a/event.c b/event.c
index 8546d66..28b74ea 100644
--- a/event.c
+++ b/event.c
@@ -312,7 +312,7 @@ static void handle_breakpoint(struct task *task)
 
 		save_param_context(task);
 
-		if (libsym->func->report_out || options.kill) {
+		if (libsym->func->report_out || !options.nocpp) {
 			task->breakpoint = breakpoint_insert(task, get_return_addr(task), NULL, BP_HW_SCRATCH);
 			if (likely(task->breakpoint)) {
 				task->libsym = libsym;
diff --git a/report.c b/report.c
index b8b5c25..1101f8b 100644
--- a/report.c
+++ b/report.c
@@ -113,24 +113,19 @@ static void report_alloc32(struct task *task, enum mt_operation op, unsigned lon
 	server_send_msg(op, task->leader->pid, alloc, sizeof(*alloc) + i * sizeof(uint32_t));
 }
 
-static void _report_alloc(struct task *task, enum mt_operation op, unsigned long ptr, unsigned long size, int depth, struct library_symbol *libsym)
+static void report_alloc(struct task *task, enum mt_operation op, unsigned long ptr, unsigned long size, int depth, struct library_symbol *libsym)
 {
 	debug(DEBUG_FUNCTION, "%d [%d]: %#lx %lu", op, task->pid, ptr, size);
 
+	if (!ptr)
+		return;
+
 	if (task_is_64bit(task))
 		report_alloc64(task, op, ptr, size, depth, libsym);
 	else
 		report_alloc32(task, op, ptr, size, depth, libsym);
 }
 
-static void report_alloc(struct task *task, enum mt_operation op, unsigned long ptr, unsigned long size, int depth, struct library_symbol *libsym)
-{
-	if (!ptr)
-		return;
-
-	_report_alloc(task, op, ptr, size, depth, libsym);
-}
-
 static void _report_alloc_op(struct task *task, struct library_symbol *libsym, enum mt_operation op)
 {
 	unsigned long size = fetch_param(task, 0);
@@ -186,21 +181,23 @@ static void _report_realloc(struct task *task, struct library_symbol *libsym)
 		report_alloc(task, MT_REALLOC, ret, size, options.bt_depth, libsym);
 	}
 
-	if (task_is_64bit(task)) {
-		struct mt_alloc_payload_64 *alloc = alloca(sizeof(*alloc));
+	if (fetch_param(task, 0)) {
+		if (task_is_64bit(task)) {
+			struct mt_alloc_payload_64 *alloc = alloca(sizeof(*alloc));
 
-		alloc->ptr = (uint64_t)ret;
-		alloc->size = (uint64_t)task->pid;
+			alloc->ptr = (uint64_t)ret;
+			alloc->size = (uint64_t)task->pid;
 
-		server_send_msg(MT_REALLOC_DONE, task->leader->pid, alloc, sizeof(*alloc));
-	}
-	else {
-		struct mt_alloc_payload_32 *alloc = alloca(sizeof(*alloc));
+			server_send_msg(MT_REALLOC_DONE, task->leader->pid, alloc, sizeof(*alloc));
+		}
+		else {
+			struct mt_alloc_payload_32 *alloc = alloca(sizeof(*alloc));
 
-		alloc->ptr = (uint32_t)ret;
-		alloc->size = (uint32_t)task->pid;
+			alloc->ptr = (uint32_t)ret;
+			alloc->size = (uint32_t)task->pid;
 
-		server_send_msg(MT_REALLOC_DONE, task->leader->pid, alloc, sizeof(*alloc));
+			server_send_msg(MT_REALLOC_DONE, task->leader->pid, alloc, sizeof(*alloc));
+		}
 	}
 }
 
@@ -208,7 +205,7 @@ static void report_realloc(struct task *task, struct library_symbol *libsym)
 {
 	unsigned long addr = fetch_param(task, 0);
 
-	_report_alloc(task, MT_REALLOC_ENTER, addr, task->pid, options.sanity ? options.bt_depth : 0, libsym);
+	report_alloc(task, MT_REALLOC_ENTER, addr, task->pid, options.sanity ? options.bt_depth : 0, libsym);
 }
 
 static void _report_calloc(struct task *task, struct library_symbol *libsym)
@@ -368,6 +365,10 @@ static const struct function flist[] = {
 	{ "delete[](void*)",					"_ZdaPv",		1,	report_delete_array,	NULL },
 	{ "delete(void*, std::nothrow_t const&)",		"_ZdlPvRKSt9nothrow_t",	1,	report_delete,		NULL },
 	{ "delete[](void*, std::nothrow_t const&)",		"_ZdaPvRKSt9nothrow_t",	1,	report_delete_array,	NULL },
+	{ "delete(void*, unsigned long)",			"_ZdlPvm",		1,	report_delete,		NULL },
+	{ "delete[](void*, unsigned long)",			"_ZdaPvj",		1,	report_delete_array,	NULL },
+	{ "delete(void*, unsigned long)",			"_ZdlPvj",		1,	report_delete,		NULL },
+	{ "delete[](void*, unsigned long)",			"_ZdaPvm",		1,	report_delete_array,	NULL },
 };
 
 const struct function *flist_matches_symbol(const char *sym_name)
diff --git a/sysdeps/linux-gnu/arm/arch.c b/sysdeps/linux-gnu/arm/arch.c
index 9979fc3..bf18192 100644
--- a/sysdeps/linux-gnu/arm/arch.c
+++ b/sysdeps/linux-gnu/arm/arch.c
@@ -669,7 +669,7 @@ static int ptrace_cont(struct task *task)
 	return 0;
 }
 
-int do_singlestep(struct task *task)
+int do_singlestep(struct task *task, struct breakpoint *bp)
 {
 	const uint32_t pc = get_instruction_pointer(task);
 	uint32_t cpsr;
@@ -714,16 +714,13 @@ int do_singlestep(struct task *task)
 	else
 		bp2 = NULL;
 
-	ret = handle_singlestep(task, ptrace_cont);
+	ret = handle_singlestep(task, ptrace_cont, bp);
 
 	if (bp1)
 		breakpoint_disable(task, bp1);
 	if (bp2)
 		breakpoint_disable(task, bp2);
 
-	if (ret)
-		return ret;
-
-	return 0;
+	return ret;
 }
 
diff --git a/sysdeps/linux-gnu/arm/regs.c b/sysdeps/linux-gnu/arm/regs.c
index c33bb4a..47ad44f 100644
--- a/sysdeps/linux-gnu/arm/regs.c
+++ b/sysdeps/linux-gnu/arm/regs.c
@@ -59,6 +59,11 @@ void set_instruction_pointer(struct task *task, arch_addr_t addr)
 	}
 }
 
+unsigned int ip_reg_addr(void)
+{
+	return offsetof(struct pt_regs, ARM_pc);
+}
+
 arch_addr_t get_return_addr(struct task *task)
 {
 	return ARCH_ADDR_T(task->context.regs.ARM_lr);
diff --git a/sysdeps/linux-gnu/ppc/regs.c b/sysdeps/linux-gnu/ppc/regs.c
index b40de17..adecbf0 100644
--- a/sysdeps/linux-gnu/ppc/regs.c
+++ b/sysdeps/linux-gnu/ppc/regs.c
@@ -74,6 +74,11 @@ void set_instruction_pointer(struct task *task, arch_addr_t addr)
 	}
 }
 
+unsigned int ip_reg_addr(void)
+{
+	return sizeof(unsigned long) * PT_NIP;
+}
+
 arch_addr_t get_return_addr(struct task *task)
 {
 #ifdef __powerpc64__
diff --git a/sysdeps/linux-gnu/trace.c b/sysdeps/linux-gnu/trace.c
index 92f14a9..97eb2e2 100644
--- a/sysdeps/linux-gnu/trace.c
+++ b/sysdeps/linux-gnu/trace.c
@@ -233,6 +233,7 @@ static void process_event(struct task *task, int status)
 	stop_signal = _process_event(task, status);
 
 	if (stop_signal == -1) {
+fprintf(stderr, "%s:%d\n", __FUNCTION__, __LINE__);
 		task->event.type = EVENT_NONE;
 		continue_task(task, 0);
 		return;
@@ -245,6 +246,7 @@ static void process_event(struct task *task, int status)
 		return;
 
 	if (unlikely(fetch_context(task) == -1)) {
+fprintf(stderr, "%s:%d\n", __FUNCTION__, __LINE__);
 		task->event.type = EVENT_NONE;
 		continue_task(task, 0);
 		return;
@@ -259,7 +261,6 @@ static void process_event(struct task *task, int status)
 		if (task->hw_bp[i] && task->hw_bp[i]->addr == ip) {
 			if (likely(get_hw_bp_state(task, i)))
 				bp = task->hw_bp[i];
-
 			break;
 		}
 	}
@@ -273,8 +274,9 @@ static void process_event(struct task *task, int status)
 	{
 		bp = breakpoint_find(leader, ip - DECR_PC_AFTER_BREAK);
 		if (unlikely(!bp)) {
-			task->event.type = EVENT_NONE;
-			continue_task(task, 0);
+fprintf(stderr, "%s:%d\n", __FUNCTION__, __LINE__);
+//			task->event.type = EVENT_NONE;
+//			continue_task(task, 0);
 			return;
 		}
 #if HW_BREAKPOINTS > 0
@@ -459,10 +461,12 @@ void stop_threads(struct task *task)
 	}
 }
 
-int handle_singlestep(struct task *task, int (*singlestep)(struct task *task))
+int handle_singlestep(struct task *task, int (*singlestep)(struct task *task), struct breakpoint *bp)
 {
 	int status;
 	int stop_signal;
+	unsigned long ip;
+	int ret;
 
 	for(;;) {
 		if (unlikely(singlestep(task) == -1))
@@ -470,26 +474,41 @@ int handle_singlestep(struct task *task, int (*singlestep)(struct task *task))
 
 		if (unlikely(TEMP_FAILURE_RETRY(waitpid(task->pid, &status, __WALL)) != task->pid)) {
 			fprintf(stderr, "%s waitpid pid=%d %s\n", __FUNCTION__,  task->pid, strerror(errno));
-			return -1;
+			return 0;
 		}
 
 		stop_signal = _process_event(task, status);
 
 		if (stop_signal == -1)
-			return -1;
+			return 0;
 
-		if (likely(stop_signal == SIGTRAP))
-			return 0; /* check if was a real breakpoint code there */
+		ip = ptrace(PTRACE_PEEKUSER, task->pid, ip_reg_addr(), 0);
+		if (ip == (unsigned long)-1) {
+			fprintf(stderr, "%s ptrace get IP pid=%d %s\n", __FUNCTION__,  task->pid, strerror(errno));
+			return 0;
+		}
+
+		if (ip != bp->addr) {
+			ret = 0;
+
+			if (likely(stop_signal == SIGTRAP))
+				return ret;
+		}
+		else
+			ret = 1;
 
 		if (likely(!stop_signal)) {
 			queue_event(task);
-			return 0;
+			return ret;
 		}
 
 		if (fix_signal(task, stop_signal) > 0) {
 			queue_event(task);
-			return 1;
+			return ret;
 		}
+
+		if (!ret)
+			return ret;
 	}
 }
 
@@ -504,9 +523,9 @@ static int ptrace_singlestep(struct task *task)
 	return 0;
 }
 
-int do_singlestep(struct task *task)
+int do_singlestep(struct task *task, struct breakpoint *bp)
 {
-	return handle_singlestep(task, ptrace_singlestep);
+	return handle_singlestep(task, ptrace_singlestep, bp);
 }
 #endif
 
diff --git a/sysdeps/linux-gnu/x86/regs.c b/sysdeps/linux-gnu/x86/regs.c
index 1547be7..f718d27 100644
--- a/sysdeps/linux-gnu/x86/regs.c
+++ b/sysdeps/linux-gnu/x86/regs.c
@@ -91,6 +91,15 @@ void set_instruction_pointer(struct task *task, arch_addr_t addr)
 		fprintf(stderr, "pid=%d Couldn't set instruction pointer: %s\n", task->pid, strerror(errno));
 }
 
+unsigned int ip_reg_addr(void)
+{
+#ifdef __x86_64__
+	return sizeof(unsigned long) * RIP;
+#else
+	return sizeof(unsigned long) * EIP;
+#endif
+}
+
 arch_addr_t get_return_addr(struct task *task)
 {
 	long a;
diff --git a/trace.c b/trace.c
index 4f0da0e..e244cae 100644
--- a/trace.c
+++ b/trace.c
@@ -53,19 +53,21 @@ int skip_breakpoint(struct task *task, struct breakpoint *bp)
 		int ret = 0;
 		struct timespec start;
 
+		if (task->skip_bp)
+			return 1;
+
 		if (unlikely(options.verbose > 1))
 			start_time(&start);
 
 		breakpoint_disable(task, bp);
-		ret = do_singlestep(task);
+		ret = do_singlestep(task, bp);
 		breakpoint_enable(task, bp);
 
 		if (unlikely(options.verbose > 1))
 			set_timer(&start, &skip_bp_time);
 
 		if (unlikely(ret)) {
-			if (unlikely(ret == 1))
-				task->skip_bp = breakpoint_get(bp);
+			task->skip_bp = breakpoint_get(bp);
 			return ret;
 		}
 	}