Introduce a tier-1 JIT compiler based on aarch64 architecture

[qwe661234]

We follow the template and API of X64 to implement A64 tier-1 JIT compiler.

• Perfromance

Metric	rv32emu-T1C	qemu
aes	0.034	0.045
puzzle	0.0115	0.0169
pi	0.035	0.032
dhrystone	1.914	2.005
Nqeueens	3.87	2.898
qsort-O2	7.819	11.614
miniz-O2	7.604	3.803
primes-O2	10.551	5.986
sha512-O2	6.497	2.853
stream	52.25	45.776

As demonstrated in the memory usage analysis below, the tier-1 JIT compiler utilizes less memory than QEMU across all benchmarks.

• Memory usage

Metric	rv32emu-T1C	qemu
aes	183,212	1,265,962
puzzle	145,239	891,357
pi	144,739	872,525
dhrystone	146,282	853,256
Nqeueens	146,696	854,174
qsort-O2	146,907	856,721
miniz-O2	157,475	999,897
primes-O2	142,356	851,661
sha512-O2	145,369	901,136
stream	157,975	955,809

Related: #238
Close: #296

[jserv]

Consider the following changes for Apple Silicon:

diff --git a/Makefile b/Makefile
index f8fe8e7..3bc5cc3 100644
--- a/Makefile
+++ b/Makefile
@@ -123,7 +123,7 @@ $(call set-feature, JIT)
 ifeq ($(call has, JIT), 1)
 OBJS_EXT += jit.o 
 ifneq ($(processor), x86_64)
-ifneq ($(processor), aarch64)
+ifneq ($(processor), arm64)
 $(error JIT mode only supports for x64 and arm64 target currently.)
 endif
 endif
diff --git a/src/jit.c b/src/jit.c
index 403f874..89d0fdb 100644
--- a/src/jit.c
+++ b/src/jit.c
@@ -36,6 +36,10 @@
 #include "state.h"
 #include "utils.h"
 
+#if defined(__APPLE__) && defined(__aarch64__)
+#include <libkern/OSCacheControl.h>
+#include <pthread.h>
+#endif
 
 #define JIT_CLS_MASK 0x07
 #define JIT_ALU_OP_MASK 0xf0
@@ -286,9 +290,10 @@ static inline void offset_map_insert(struct jit_state *state, int32_t target_pc)
     map_entry->offset = state->offset;
 }
 
-
+#if !defined(__APPLE__)
 #define sys_icache_invalidate(addr, size) \
     __builtin___clear_cache((char *) (addr), (char *) (addr) + (size));
+#endif
 
 static inline void emit_bytes(struct jit_state *state, void *data, uint32_t len)
 {
@@ -1505,13 +1510,17 @@ struct jit_state *init_state(size_t size)
     struct jit_state *state = malloc(sizeof(struct jit_state));
     state->offset = 0;
     state->size = size;
+#if defined(__APPLE__) && defined(__aarch64__)
+    assert(pthread_jit_write_protect_supported_np());
+    pthread_jit_write_protect_np(0);
+#endif
+#if defined(__APPLE__) && defined(__aarch64__)
+    state->buf = mmap(NULL, size, PROT_EXEC | PROT_WRITE | PROT_READ,
+                      MAP_PRIVATE | MAP_ANONYMOUS | MAP_JIT, -1, 0);
+#else
     state->buf = mmap(0, size, PROT_READ | PROT_WRITE | PROT_EXEC,
-                      MAP_PRIVATE | MAP_ANONYMOUS
-#if defined(__APPLE__)
-                          | MAP_JIT
+                      MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 #endif
-                      ,
-                      -1, 0);
     assert(state->buf != MAP_FAILED);
     prepare_translate(state);
     state->offset_map = calloc(MAX_INSNS, sizeof(struct offset_map));

build/hello.elf is known to work.

[qwe661234]

Consider the following changes for Apple Silicon:

It pass all tests now, we gain about 4 times performance improvement on macOS + Apple Silicon.

[jserv]

Update CI pipeline for JIT/Arm64-enabled build.