diff --git a/Makefile b/Makefile
index a0ba536..7548bf0 100644
--- a/Makefile
+++ b/Makefile
@@ -63,6 +63,7 @@ else
 endif
 
 #MCFLAGS += -fsanitize=thread
+#MCFLAGS += -fsanitize=address -static-libasan
 
 CC ?= $(CROSS_COMPILE)gcc
 LD ?= $(CROSS_COMPILE)ld.bfd
diff --git a/arch/arm/enter.c b/arch/arm/enter.c
index db14bae..457be67 100644
--- a/arch/arm/enter.c
+++ b/arch/arm/enter.c
@@ -29,6 +29,7 @@ static void __attribute__((used)) container(void)
 	/* rt_sigprocmask(), expects pointer to area for masks in r8 */
 	asm volatile(
 		".global sigprocmask_blob			\n"
+		".align 3					\n"
 		"sigprocmask_blob:				\n"
 		"mov r7, #175					\n" /* __NR_rt_sigprocmask */
 		"mov r0, %0					\n" /* @how */
@@ -38,6 +39,7 @@ static void __attribute__((used)) container(void)
 		"svc 0x0					\n"
 		"udf #16					\n" /* SIGTRAP */
 		".global sigprocmask_blob_size			\n"
+		".align 3					\n"
 		"sigprocmask_blob_size:				\n"
 		".int sigprocmask_blob_size - sigprocmask_blob	\n"
 		:: "i" (SIG_SETMASK)
@@ -46,6 +48,7 @@ static void __attribute__((used)) container(void)
 	/* mmaps anon area for parasite_blob */
 	asm volatile(
 		".global mmap_blob				\n"
+		".align 3					\n"
 		"mmap_blob:					\n"
 		"mov r7, #192					\n" /* __NR_mmap2 */
 		"mov r0, #0					\n" /* @addr */
@@ -57,6 +60,7 @@ static void __attribute__((used)) container(void)
 		"svc 0x0					\n"
 		"udf #16					\n" /* SIGTRAP */
 		".global mmap_blob_size				\n"
+		".align 3					\n"
 		"mmap_blob_size:				\n"
 		".int mmap_blob_size - mmap_blob		\n"
 		:: "i" (PROT_EXEC | PROT_READ | PROT_WRITE),
@@ -66,6 +70,7 @@ static void __attribute__((used)) container(void)
 	/* clones parasite, expects parasite address in r8 */
 	asm volatile(
 		".global clone_blob				\n"
+		".align 3					\n"
 		"clone_blob:					\n"
 		"mov r7, #120					\n" /* __NR_clone */
 		"ldr r0, CLONE_FLAGS				\n"
@@ -79,6 +84,7 @@ static void __attribute__((used)) container(void)
 		"CLONE_FLAGS:					\n"
 		".word 0x80050f00				\n" /* (CLONE_FILES | CLONE_FS | CLONE_IO | CLONE_SIGHAND | CLONE_SYSVSEM | CLONE_THREAD | CLONE_VM) */
 		".global clone_blob_size			\n"
+		".align 3					\n"
 		"clone_blob_size:				\n"
 		".int clone_blob_size - clone_blob		\n"
 	);
@@ -86,6 +92,7 @@ static void __attribute__((used)) container(void)
 	/* munmap anon area for parasite_blob, expects addr in r8 and len in r9 */
 	asm volatile(
 		".global munmap_blob				\n"
+		".align 3					\n"
 		"munmap_blob:					\n"
 		"mov r7, #91					\n" /* __NR_munmap */
 		"mov r0, r8					\n" /* @addr */
@@ -93,6 +100,7 @@ static void __attribute__((used)) container(void)
 		"svc 0x0					\n"
 		"udf #16					\n" /* SIGTRAP */
 		".global munmap_blob_size			\n"
+		".align 3					\n"
 		"munmap_blob_size:				\n"
 		".int munmap_blob_size - munmap_blob		\n"
 	);
diff --git a/arch/arm64/enter.c b/arch/arm64/enter.c
index d47f172..272a767 100644
--- a/arch/arm64/enter.c
+++ b/arch/arm64/enter.c
@@ -30,6 +30,7 @@ static void __attribute__((used)) container(void)
 	/* rt_sigprocmask(), expects pointer to area for masks in x10 */
 	asm volatile(
 		".global sigprocmask_blob			\n"
+		".align 3					\n"
 		"sigprocmask_blob:				\n"
 		"mov x8, #135					\n" /* __NR_rt_sigprocmask */
 		"mov x0, %0					\n" /* @how */
@@ -39,6 +40,7 @@ static void __attribute__((used)) container(void)
 		"svc #0						\n"
 		"brk #0						\n" /* SIGTRAP */
 		".global sigprocmask_blob_size			\n"
+		".align 3					\n"
 		"sigprocmask_blob_size:				\n"
 		".int sigprocmask_blob_size - sigprocmask_blob	\n"
 		:: "i" (SIG_SETMASK)
@@ -47,6 +49,7 @@ static void __attribute__((used)) container(void)
 	/* mmaps anon area for parasite_blob */
 	asm volatile(
 		".global mmap_blob				\n"
+		".align 3					\n"
 		"mmap_blob:					\n"
 		"mov x8, #222					\n" /* __NR_mmap2 */
 		"mov x0, #0					\n" /* @addr */
@@ -58,6 +61,7 @@ static void __attribute__((used)) container(void)
 		"svc #0						\n"
 		"brk #0						\n" /* SIGTRAP */
 		".global mmap_blob_size				\n"
+		".align 3					\n"
 		"mmap_blob_size:				\n"
 		".int mmap_blob_size - mmap_blob		\n"
 		:: "i" (PROT_EXEC | PROT_READ | PROT_WRITE),
@@ -67,6 +71,7 @@ static void __attribute__((used)) container(void)
 	/* clones parasite, expects parasite address in x10 */
 	asm volatile(
 		".global clone_blob				\n"
+		".align 3					\n"
 		"clone_blob:					\n"
 		"mov x8, #220					\n" /* __NR_clone */
 		"ldr x0, CLONE_FLAGS				\n" /* @flags */
@@ -82,6 +87,7 @@ static void __attribute__((used)) container(void)
 		"CLONE_FLAGS:					\n"
 		".quad (%0 & 0xffffffff)			\n" /* zero high .word */
 		".global clone_blob_size			\n"
+		".align 3					\n"
 		"clone_blob_size:				\n"
 		".int clone_blob_size - clone_blob		\n"
 		:: "i" (CLONE_FILES | CLONE_FS | CLONE_IO | CLONE_SIGHAND | CLONE_SYSVSEM | CLONE_THREAD | CLONE_VM)
@@ -90,6 +96,7 @@ static void __attribute__((used)) container(void)
 	/* munmap anon area for parasite_blob, expects addr in x10 and len in x11 */
 	asm volatile(
 		".global munmap_blob				\n"
+		".align 3					\n"
 		"munmap_blob:					\n"
 		"mov x8, #215					\n" /* __NR_munmap */
 		"mov x0, x10					\n" /* @addr */
@@ -97,6 +104,7 @@ static void __attribute__((used)) container(void)
 		"svc #0						\n"
 		"brk #0						\n" /* SIGTRAP */
 		".global munmap_blob_size			\n"
+		".align 3					\n"
 		"munmap_blob_size:				\n"
 		".int munmap_blob_size - munmap_blob		\n"
 	);
diff --git a/arch/enter.h b/arch/enter.h
index 791e6e5..ca2c944 100644
--- a/arch/enter.h
+++ b/arch/enter.h
@@ -19,7 +19,7 @@
 #ifndef __ENTER_H__
 #define __ENTER_H__
 
-extern char test_blob[], sigprocmask_blob[], mmap_blob[], clone_blob[], munmap_blob[];
-extern int test_blob_size, sigprocmask_blob_size, mmap_blob_size, clone_blob_size, munmap_blob_size;
+extern char sigprocmask_blob[], mmap_blob[], clone_blob[], munmap_blob[];
+extern int sigprocmask_blob_size, mmap_blob_size, clone_blob_size, munmap_blob_size;
 
 #endif
\ No newline at end of file
diff --git a/arch/x86_64/enter.c b/arch/x86_64/enter.c
index cb23c99..741fbc4 100644
--- a/arch/x86_64/enter.c
+++ b/arch/x86_64/enter.c
@@ -30,6 +30,7 @@ static void __attribute__((used)) container(void)
 	/* rt_sigprocmask(), expects pointer to area for masks in %r15 */
 	asm volatile(
 		".global sigprocmask_blob			\n"
+		".align 8					\n"
 		"sigprocmask_blob:				\n"
 		"movq $14, %%rax				\n" /* __NR_rt_sigprocmask */
 		"movq %0, %%rdi					\n" /* @how */
@@ -39,6 +40,7 @@ static void __attribute__((used)) container(void)
 		"syscall					\n"
 		"int $0x03					\n"
 		".global sigprocmask_blob_size			\n"
+		".align 8					\n"
 		"sigprocmask_blob_size:				\n"
 		".int sigprocmask_blob_size - sigprocmask_blob	\n"
 		:: "i" (SIG_SETMASK));
@@ -46,6 +48,7 @@ static void __attribute__((used)) container(void)
 	/* mmaps anon area for parasite_blob */
 	asm volatile(
 		".global mmap_blob				\n"
+		".align 8					\n"
 		"mmap_blob:					\n"
 		"movq $9, %%rax					\n" /* mmap */
 		"movq $0, %%rdi					\n" /* @addr */
@@ -57,6 +60,7 @@ static void __attribute__((used)) container(void)
 		"syscall					\n"
 		"int $0x03					\n"
 		".global mmap_blob_size				\n"
+		".align 8					\n"
 		"mmap_blob_size:				\n"
 		".int mmap_blob_size - mmap_blob		\n"
 		:: "i" (PROT_EXEC | PROT_READ | PROT_WRITE),
@@ -65,6 +69,7 @@ static void __attribute__((used)) container(void)
 	/* clones parasite, expects parasite address in %r15 */
 	asm volatile(
 		".global clone_blob				\n"
+		".align 8					\n"
 		"clone_blob:					\n"
 		"movq $56, %%rax				\n" /* clone */
 		"movq %0, %%rdi					\n" /* @flags */
@@ -77,6 +82,7 @@ static void __attribute__((used)) container(void)
 		"jmp *%%r15					\n" /* jmp parasite */
 		"1: int $0x03					\n"
 		".global clone_blob_size			\n"
+		".align 8					\n"
 		"clone_blob_size:				\n"
 		".int clone_blob_size - clone_blob		\n"
 		:: "i" (CLONE_FILES | CLONE_FS | CLONE_IO | CLONE_SIGHAND | CLONE_SYSVSEM | CLONE_THREAD | CLONE_VM));
@@ -84,6 +90,7 @@ static void __attribute__((used)) container(void)
 	/* munmaps anon area for parasite_blob, expects mmap address in %r15 and len in %r14 */
 	asm volatile(
 		".global munmap_blob				\n"
+		".align 8					\n"
 		"munmap_blob:					\n"
 		"movq $11, %%rax				\n" /* munmap */
 		"movq %%r15, %%rdi				\n" /* @addr */
@@ -91,6 +98,7 @@ static void __attribute__((used)) container(void)
 		"syscall					\n"
 		"int $0x03					\n"
 		".global munmap_blob_size			\n"
+		".align 8					\n"
 		"munmap_blob_size:				\n"
 		".int munmap_blob_size - munmap_blob		\n"
 		::
diff --git a/memcr.c b/memcr.c
index 8325878..87ff122 100644
--- a/memcr.c
+++ b/memcr.c
@@ -1837,7 +1837,10 @@ static int peek(pid_t pid, unsigned long *addr, unsigned long *dst, size_t len)
 {
 	int i;
 
-	for (i = 0; i < DIV_ROUND_UP(len, sizeof(unsigned long)); i++) {
+	/* len must be a mutliple of CPU word size */
+	assert(len % sizeof(unsigned long) == 0);
+
+	for (i = 0; i < (len / sizeof(unsigned long)); i++) {
 		errno = 0;
 		dst[i] = ptrace(PTRACE_PEEKDATA, pid, addr + i, NULL);
 		if (errno) {
@@ -1854,7 +1857,10 @@ static int poke(pid_t pid, unsigned long *addr, unsigned long *src, size_t len)
 	int ret;
 	int i;
 
-	for (i = 0; i < DIV_ROUND_UP(len, sizeof(unsigned long)); i++) {
+	/* len must be a mutliple of CPU word size */
+	assert(len % sizeof(unsigned long) == 0);
+
+	for (i = 0; i < (len / sizeof(unsigned long)); i++) {
 		ret = ptrace(PTRACE_POKEDATA, pid, addr + i, *(src + i));
 		if (ret) {
 			fprintf(stderr, "[-] %s() failed addr %p, src %p, i %d: %m\n", __func__, addr, src, i);
@@ -2067,18 +2073,16 @@ static int signals_unblock(pid_t pid)
 
 static int ctx_save(pid_t pid)
 {
+	int max_blob_size;
 	struct registers regs;
 
 	ctx.pid = pid;
 
 	/* allocate space to save original code */
-	ctx.code_size = DIV_ROUND_UP(MAX(sigprocmask_blob_size,
-				 MAX(mmap_blob_size,
-				 MAX(clone_blob_size,
-				 munmap_blob_size))),
-				 sizeof(unsigned long));
+	max_blob_size = MAX(sigprocmask_blob_size, MAX(mmap_blob_size, MAX(clone_blob_size, munmap_blob_size)));
+	ctx.code_size = DIV_ROUND_UP(max_blob_size, sizeof(unsigned long)) * sizeof(unsigned long);
 
-	ctx.code = malloc(sizeof(unsigned long) * ctx.code_size);
+	ctx.code = malloc(ctx.code_size);
 	assert(ctx.code);
 
 	read_cpu_regs(ctx.pid, &regs);
diff --git a/memcr.h b/memcr.h
index d4a8fea..94d94cb 100644
--- a/memcr.h
+++ b/memcr.h
@@ -25,9 +25,10 @@
 #define PAGE_SIZE 4096
 #endif
 
+/* size is CPU word aligned for ptrace() peek/poke */
 struct parasite_args {
 	struct sockaddr_un addr;
-};
+} __attribute__((aligned(sizeof(unsigned long))));
 
 typedef enum {
 	CMD_MPROTECT = 1,