scanf to printf on AArch64

m1.c

#include <stdio.h>
#include <stdint.h>
#include <sys/mman.h>

int main()
{
	mprotect((void *)((unsigned long)main & (~0xfff)), 4096,
		 PROT_EXEC|PROT_WRITE);
	{ // gcc and clang puts the "0:", "1:" etc labels below at different places
		char *p;
		__asm__("adr %[p], 0f\n"
			: /* ouputs */ [p] "=r" (p)
			: /* inputs */
			: /* clobbers */ "x10"
			);
		printf("0f: 0x%p\n", p);
		__asm__("adr %[p], 1f\n"
			: /* ouputs */ [p] "=r" (p)
			: /* inputs */
			: /* clobbers */ "x10"
			);
		printf("1f: 0x%p\n", p);
		__asm__("adr %[p], 2f\n"
			: /* ouputs */ [p] "=r" (p)
			: /* inputs */
			: /* clobbers */ "x10"
			);
		printf("2f: 0x%p\n", p);
	}

#if defined(__clang__)
	#define HELLO1_LABEL "0f"
	#define HELLO2_LABEL "1f"
#elif defined(__GNUC__)
	#define HELLO1_LABEL "(1f-4)"
	#define HELLO2_LABEL "(2f-4)"
#endif

	uint32_t dst;
	__asm__("ldr %w[dst], " HELLO1_LABEL "\n"
		: /* ouputs */ [dst] "=r" (dst)
		: /* inputs */
		: /* clobbers */
		);
	// bl PCREL26
	printf("dst: 0x%0x\n", dst);
	dst = ((int32_t)(dst << 6)) >> 4;
	printf("dst: %d\n", dst);
	__asm__(
	        "sub %w[dst], %w[dst], " HELLO2_LABEL "-" HELLO1_LABEL "\n"
		: /* ouputs */ [dst] "+r" (dst)
		: /* inputs */
		: /* clobbers */
		);
	printf("dst: %d\n", dst);
	dst = 0x94000000 | (((uint32_t)(dst >> 2)) & (0x03ffffff));
	printf("dst: 0x%0x\n", dst);
	__asm__ volatile (
	        "adr x10, " HELLO2_LABEL "\n"
		"str %w[dst], [x10]\n"
		: /* ouputs */
		: /* inputs */ [dst] "r" (dst)
		: /* clobbers */ "x10", "memory"
		);
	// Write back and invalidate data and instruction caches
	__builtin___clear_cache((char*)main, (char*)main+1024);

	// Jump to target after codegen.
	__asm__ volatile ("b 0f");
	__asm__("0:");

	// In practice, the code works on Apple M2 and Amazon Graviton
	// without "__builtin___clear_cache()" and "b 0f" above.

	printf("Hello 1\n");
	__asm__ volatile ("1:");
	scanf("Hello 2\n");
	__asm__ volatile ("2:");
	return 0;
}