i#5383: macos M1 support

This patch adds enough support to run simple hello world on M1 macs.

Issue: DynamoRIO#5383

CMakeLists.txt

@@ -216,14 +216,14 @@ endif ()
 # For cross-compilation this should still work as you're supposed to set this var.
 # X64 mean 64-bit generically, whether AMD64 or AARCH64.
 set(TARGET_ARCH "${CMAKE_SYSTEM_PROCESSOR}" CACHE STRING "Target architecture")
-if (TARGET_ARCH MATCHES "^arm")
-  set(ARM 1) # This means AArch32.
-  set(X64 OFF)
-  message(STATUS "Building for ARM")
-elseif (TARGET_ARCH MATCHES "^aarch64")
+if (TARGET_ARCH MATCHES "^arm64" OR TARGET_ARCH MATCHES "^aarch64")
   set(AARCH64 1)
   set(X64 1)
   message(STATUS "Building for AArch64")
+elseif (TARGET_ARCH MATCHES "^arm")
+  set(ARM 1) # This means AArch32.
+  set(X64 OFF)
+  message(STATUS "Building for ARM")
 else ()
   set(X86 1) # This means IA-32 or AMD64
   message(STATUS "Building for x86")
@@ -263,17 +263,17 @@ endif ()
 
 # The execution architecture, which might differ from the target for building
 # an AArch64 decoder to execute on x86 machines (i#1684).
-if (CMAKE_SYSTEM_PROCESSOR MATCHES "^arm")
-  set(DR_HOST_ARM 1)
-  set(DR_HOST_ARCH_NAME "arm")
-  set(DR_HOST_AARCHXX 1)
-  set(DR_HOST_ARCH_NAME_SHARED aarchxx)
-elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "^aarch64")
+if (CMAKE_SYSTEM_PROCESSOR MATCHES "^aarch64" OR CMAKE_SYSTEM_PROCESSOR MATCHES "^arm64")
   set(DR_HOST_AARCH64 1)
   set(DR_HOST_ARCH_NAME "aarch64")
   set(DR_HOST_AARCHXX 1)
   set(DR_HOST_ARCH_NAME_SHARED aarchxx)
   set(DR_HOST_X64 1)
+elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "^arm")
+  set(DR_HOST_ARM 1)
+  set(DR_HOST_ARCH_NAME "arm")
+  set(DR_HOST_AARCHXX 1)
+  set(DR_HOST_ARCH_NAME_SHARED aarchxx)
 elseif (CMAKE_C_SIZEOF_DATA_PTR EQUAL 8)
   set(DR_HOST_X86 1)
   set(DR_HOST_ARCH_NAME "x86")

core/arch/aarch64/aarch64.asm

@@ -37,7 +37,11 @@
 
 #include "../asm_defines.asm"
 START_FILE
+
+
+#if !(defined(MACOS) && defined(AARCH64))
 #include "include/syscall.h"
+#endif
 
 #ifndef UNIX
 # error Non-Unix is not supported
@@ -308,17 +312,22 @@ GLOBAL_LABEL(dynamorio_app_take_over:)
 GLOBAL_LABEL(cleanup_and_terminate:)
         /* move argument registers to callee saved registers */
         mov      x19, x0  /* dcontext ptr size */
+#ifdef MACOS
+        mov      x20, x1  /* sysnr */
+        mov      x21, x2  /* arg1 */
+        mov      x22, x3  /* arg2 */
+#else
         mov      w20, w1  /* sysnum 32-bit int */
         mov      w21, w2  /* sys_arg1 32-bit int */
         mov      w22, w3  /* sys_arg2 32-bit int */
+#endif
         mov      w23, w4  /* exitproc 32-bit int */
                           /* x24 reserved for dstack ptr */
                           /* x25 reserved for syscall ptr */
 
         /* inc exiting_thread_count to avoid being killed once off all_threads list */
-        adrp     x0, :got:exiting_thread_count
-        ldr      x0, [x0, #:got_lo12:exiting_thread_count]
-        CALLC2(atomic_add, x0, #1)
+        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(exiting_thread_count), x0)
+        CALLC2(GLOBAL_REF(atomic_add), x0, #1)
 
         /* save dcontext->dstack for freeing later and set dcontext->is_exiting */
         mov      w1, #1
@@ -341,39 +350,41 @@ cat_thread_only:
         CALLC0(GLOBAL_REF(dynamo_thread_exit))
 cat_no_thread:
         /* switch to d_r_initstack for cleanup of dstack */
-        adrp     x26, :got:initstack_mutex
-        ldr      x26, [x26, #:got_lo12:initstack_mutex]
+        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(initstack_mutex), x26)
 cat_spin:
-        CALLC2(atomic_swap, x26, #1)
+        CALLC2(GLOBAL_REF(atomic_swap), x26, #1)
         cbz      w0, cat_have_lock
         yield
         b        cat_spin
 
 cat_have_lock:
         /* switch stack */
-        adrp     x0, :got:d_r_initstack
-        ldr      x0, [x0, #:got_lo12:d_r_initstack]
+        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(d_r_initstack), x0)
         ldr      x0, [x0]
         mov      sp, x0
 
         /* free dstack and call the EXIT_DR_HOOK */
         CALLC1(GLOBAL_REF(dynamo_thread_stack_free_and_exit), x24) /* pass dstack */
 
         /* give up initstack_mutex */
-        adrp     x0, :got:initstack_mutex
-        ldr      x0, [x0, #:got_lo12:initstack_mutex]
+        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(initstack_mutex), x0)
         mov      x1, #0
         str      x1, [x0]
 
         /* dec exiting_thread_count (allows another thread to kill us) */
-        adrp     x0, :got:exiting_thread_count
-        ldr      x0, [x0, #:got_lo12:exiting_thread_count]
-        CALLC2(atomic_add, x0, #-1)
+        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(exiting_thread_count), x0)
+        CALLC2(GLOBAL_REF(atomic_add), x0, #-1)
 
         /* put system call number in x8 */
+#ifdef MACOS
+        mov      x0, x20  /* sysnr */
+        mov      x1, x21  /* arg1 */
+        mov      x2, x22  /* arg2 */
+#else
         mov      w0, w21 /* sys_arg1 32-bit int */
         mov      w1, w22 /* sys_arg2 32-bit int */
         mov      w8, w20 /* int sys_call */
+#endif
 
         br       x25  /* go do the syscall! */
         bl       GLOBAL_REF(unexpected_return) /* FIXME i#1569: NYI */
@@ -392,8 +403,13 @@ GLOBAL_LABEL(atomic_add:)
 
         DECLARE_FUNC(global_do_syscall_int)
 GLOBAL_LABEL(global_do_syscall_int:)
+#ifdef MACOS
+        mov      x16, #0
+        svc      #0x80
+#else
         /* FIXME i#1569: NYI on AArch64 */
         svc      #0
+#endif
         bl       GLOBAL_REF(unexpected_return)
         END_FUNC(global_do_syscall_int)
 
@@ -414,7 +430,7 @@ DECLARE_GLOBAL(safe_read_asm_recover)
         DECLARE_FUNC(safe_read_asm)
 GLOBAL_LABEL(safe_read_asm:)
         cmp      ARG3, #0
-1:      b.eq     safe_read_asm_recover
+1:      b.eq     safe_read_asm_recover_local
 ADDRTAKEN_LABEL(safe_read_asm_pre:)
         ldrb     w3, [ARG2]
 ADDRTAKEN_LABEL(safe_read_asm_mid:)
@@ -425,6 +441,7 @@ ADDRTAKEN_LABEL(safe_read_asm_post:)
         add      ARG1, ARG1, #1
         b        1b
 ADDRTAKEN_LABEL(safe_read_asm_recover:)
+safe_read_asm_recover_local:
         mov      x0, ARG2
         ret
         END_FUNC(safe_read_asm)
@@ -564,6 +581,15 @@ GLOBAL_LABEL(main_signal_handler:)
 
 #endif /* LINUX */
 
+#ifdef MACOS
+        DECLARE_FUNC(main_signal_handler)
+GLOBAL_LABEL(main_signal_handler:)
+        /* see sendsig_set_thread_state64 in unix_signal.c */
+        mov      ARG6, sp
+        b        GLOBAL_REF(main_signal_handler_C) /* chain call */
+        END_FUNC(main_signal_handler)
+#endif
+
         DECLARE_FUNC(hashlookup_null_handler)
 GLOBAL_LABEL(hashlookup_null_handler:)
         bl       GLOBAL_REF(unexpected_return) /* FIXME i#1569: NYI */
@@ -621,8 +647,7 @@ GLOBAL_LABEL(icache_op_ic_ivau_asm:)
         /* Spill X1 and X2 to TLS_REG4_SLOT and TLS_REG5_SLOT. */
         stp      x1, x2, [x0, #spill_state_r4_OFFSET]
         /* Point X1 at icache_op_struct.lock. */
-        adrp     x1, (icache_op_struct + icache_op_struct_lock_OFFSET)
-        add      x1, x1, #:lo12:(icache_op_struct + icache_op_struct_lock_OFFSET)
+        AARCH64_ADRP_GOT((GLOBAL_REF(icache_op_struct) + icache_op_struct_lock_OFFSET), x1)
         /* Acquire lock. */
         prfm     pstl1keep, [x1]
 1:
@@ -720,8 +745,7 @@ ic_ivau_return:
         /* Load fcache_return into X1. */
         ldr      x1, [x0, #spill_state_fcache_return_OFFSET]
         /* Point X0 at fake linkstub. */
-        adrp     x0, linkstub_selfmod
-        add      x0, x0, #:lo12:linkstub_selfmod
+        AARCH64_ADRP_GOT(GLOBAL_REF(linkstub_selfmod), x0)
         /* Branch to fcache_return. */
         br       x1
 
@@ -753,8 +777,7 @@ GLOBAL_LABEL(icache_op_isb_asm:)
         ldr      x2, [x0, #spill_state_r2_OFFSET]
         stp      x1, x2, [x0, #spill_state_r4_OFFSET]
         /* Point X1 at icache_op_struct.lock. */
-        adrp     x1, (icache_op_struct + icache_op_struct_lock_OFFSET)
-        add      x1, x1, #:lo12:(icache_op_struct + icache_op_struct_lock_OFFSET)
+        AARCH64_ADRP_GOT((GLOBAL_REF(icache_op_struct) + icache_op_struct_lock_OFFSET), x1)
         /* Acquire lock. */
         prfm     pstl1keep, [x1]
 1:
@@ -781,10 +804,28 @@ GLOBAL_LABEL(icache_op_isb_asm:)
         /* Load fcache_return into X1. */
         ldr      x1, [x0, #spill_state_fcache_return_OFFSET]
         /* Point X0 at fake linkstub. */
-        adrp     x0, linkstub_selfmod
-        add      x0, x0, #:lo12:linkstub_selfmod
+        AARCH64_ADRP_GOT(GLOBAL_REF(linkstub_selfmod), x0)
         /* Branch to fcache_return. */
         br       x1
         END_FUNC(icache_op_isb_asm)
 
+#if defined(MACOS) && defined(AARCH64)
+        DECLARE_FUNC(dynamorio_sigreturn)
+GLOBAL_LABEL(dynamorio_sigreturn:)
+        brk 0xb001
+        END_FUNC(dynamorio_sigreturn)
+
+        DECLARE_FUNC(dynamorio_sys_exit)
+GLOBAL_LABEL(dynamorio_sys_exit:)
+        brk 0xb002
+        END_FUNC(dynamorio_sys_exit)
+#endif
+
+#ifdef MACOS
+        DECLARE_FUNC(new_bsdthread_intercept)
+GLOBAL_LABEL(new_bsdthread_intercept:)
+        brk 0xb003
+        END_FUNC(new_bsdthread_intercept)
+#endif
+
 END_FILE

core/arch/aarch64/memfuncs.asm

@@ -70,14 +70,14 @@ GLOBAL_LABEL(memset:)
 
 /* See x86.asm notes about needing these to avoid gcc invoking *_chk */
 .global __memcpy_chk
-.hidden __memcpy_chk
+HIDDEN(__memcpy_chk)
 WEAK(__memcpy_chk)
-.set __memcpy_chk,memcpy
+.set __memcpy_chk,GLOBAL_REF(memcpy)
 
 .global __memset_chk
-.hidden __memset_chk
+HIDDEN(__memset_chk)
 WEAK(__memset_chk)
-.set __memset_chk,memset
+.set __memset_chk,GLOBAL_REF(memset)
 
 #endif /* UNIX */
 

core/arch/aarch64/proc.c

@@ -104,6 +104,7 @@ proc_init_arch(void)
     }
 
 #ifndef DR_HOST_NOT_TARGET
+#    if !defined(MACOS)
     get_processor_specific_info();
 
     DOLOG(1, LOG_TOP, {
@@ -135,6 +136,7 @@ proc_init_arch(void)
         /* FIXME i#5474: Log all FEATURE_s for ID_AA64PFR0_EL1. */
         LOG_FEATURE(FEATURE_FP16);
     });
+#    endif
 #endif
 }
 

core/arch/aarchxx/aarchxx.asm

@@ -80,7 +80,7 @@ GLOBAL_LABEL(xfer_to_new_libdr:)
          * the current DR, but w/o clobbering ARG3 or ARG4.
          */
         adr      ARG1, .L_start_invoke_C
-        adr      ARG2, _start
+        adr      ARG2, GLOBAL_REF(_start)
         sub      ARG1, ARG1, ARG2
         add      REG_PRESERVED_1, REG_PRESERVED_1, ARG1
         /* _start expects these as 2nd & 3rd args */

core/arch/aarchxx/emit_utils.c

@@ -52,7 +52,7 @@ insert_load_dr_tls_base(dcontext_t *dcontext, instrlist_t *ilist, instr_t *where
      */
     PRE(ilist, where,
         INSTR_CREATE_mrs(dcontext, opnd_create_reg(reg_base),
-                         opnd_create_reg(DR_REG_TPIDR_EL0)));
+                         opnd_create_reg(LIB_SEG_TLS)));
 #else // ARM
     /* load TLS base from user-read-only-thread-ID register
      * mrc p15, 0, reg_base, c13, c0, 3
@@ -65,8 +65,9 @@ insert_load_dr_tls_base(dcontext_t *dcontext, instrlist_t *ilist, instr_t *where
 #endif
     /* ldr dr_reg_stolen, [reg_base, DR_TLS_BASE_OFFSET] */
     PRE(ilist, where,
-        XINST_CREATE_load(dcontext, opnd_create_reg(dr_reg_stolen),
-                          OPND_CREATE_MEMPTR(reg_base, DR_TLS_BASE_OFFSET)));
+        XINST_CREATE_load(
+            dcontext, opnd_create_reg(dr_reg_stolen),
+            OPND_CREATE_MEMPTR(reg_base, IF_MACOS64_ELSE(8, 1) * DR_TLS_BASE_OFFSET)));
 }
 
 /* Having only one thread register (TPIDRURO for ARM, TPIDR_EL0 for AARCH64) shared

core/arch/arch.c

@@ -497,6 +497,8 @@ static void shared_gencode_init(IF_X86_64_ELSE(gencode_mode_t gencode_mode, void
     bool x86_to_x64_mode = false;
 #endif
 
+    pthread_jit_write();
+
     gencode = heap_mmap_reserve(GENCODE_RESERVE_SIZE, GENCODE_COMMIT_SIZE,
                                 MEMPROT_EXEC | MEMPROT_READ | MEMPROT_WRITE,
                                 VMM_SPECIAL_MMAP | VMM_REACHABLE);
@@ -3426,15 +3428,15 @@ dr_mcontext_to_priv_mcontext(priv_mcontext_t *dst, dr_mcontext_t *src)
     if (src->size > sizeof(dr_mcontext_t))
         return false;
     if (TESTALL(DR_MC_ALL, src->flags) && src->size == sizeof(dr_mcontext_t)) {
-        *dst = *(priv_mcontext_t *)(&MCXT_FIRST_REG_FIELD(src));
+        *dst = *(priv_mcontext_t *)(MCXT_FIRST_REG_FIELD(src));
     } else {
         if (TEST(DR_MC_INTEGER, src->flags)) {
             /* xsp is in the middle of the mcxt, so we save dst->xsp here and
              * restore it later so we can use one memcpy for DR_MC_INTEGER.
              */
             reg_t save_xsp = dst->xsp;
             if (src->size >= offsetof(dr_mcontext_t, IF_X86_ELSE(xflags, pc))) {
-                memcpy(&MCXT_FIRST_REG_FIELD(dst), &MCXT_FIRST_REG_FIELD(src),
+                memcpy(MCXT_FIRST_REG_FIELD(dst), MCXT_FIRST_REG_FIELD(src),
                        /* end of the mcxt integer gpr */
                        offsetof(priv_mcontext_t, IF_X86_ELSE(xflags, pc)));
             } else
@@ -3521,15 +3523,15 @@ priv_mcontext_to_dr_mcontext(dr_mcontext_t *dst, priv_mcontext_t *src)
     if (dst->size > sizeof(dr_mcontext_t))
         return false;
     if (TESTALL(DR_MC_ALL, dst->flags) && dst->size == sizeof(dr_mcontext_t)) {
-        *(priv_mcontext_t *)(&MCXT_FIRST_REG_FIELD(dst)) = *src;
+        *(priv_mcontext_t *)(MCXT_FIRST_REG_FIELD(dst)) = *src;
     } else {
         if (TEST(DR_MC_INTEGER, dst->flags)) {
             /* xsp is in the middle of the mcxt, so we save dst->xsp here and
              * restore it later so we can use one memcpy for DR_MC_INTEGER.
              */
             reg_t save_xsp = dst->xsp;
             if (dst->size >= offsetof(dr_mcontext_t, IF_X86_ELSE(xflags, pc))) {
-                memcpy(&MCXT_FIRST_REG_FIELD(dst), &MCXT_FIRST_REG_FIELD(src),
+                memcpy(MCXT_FIRST_REG_FIELD(dst), MCXT_FIRST_REG_FIELD(src),
                        /* end of the mcxt integer gpr */
                        offsetof(priv_mcontext_t, IF_X86_ELSE(xflags, pc)));
             } else
@@ -3610,7 +3612,7 @@ priv_mcontext_t *
 dr_mcontext_as_priv_mcontext(dr_mcontext_t *mc)
 {
     /* It's up to the caller to ensure the proper DR_MC_ flags are set (i#1848) */
-    return (priv_mcontext_t *)(&MCXT_FIRST_REG_FIELD(mc));
+    return (priv_mcontext_t *)(MCXT_FIRST_REG_FIELD(mc));
 }
 
 priv_mcontext_t *

core/arch/arch.h

@@ -1336,9 +1336,20 @@ new_thread_setup(priv_mcontext_t *mc);
 #    ifdef MACOS
 void
 new_bsdthread_setup(priv_mcontext_t *mc);
+// Enable writing to MAP_JIT pages.
+#        define pthread_jit_write() pthread_jit_write_protect_np(false)
+// Enable executing MAP_JIT pages.
+#        define pthread_jit_read() pthread_jit_write_protect_np(true)
+void
+pthread_jit_write_protect_np(int);
 #    endif
 #endif
 
+#ifndef pthread_jit_write
+#    define pthread_jit_write()
+#    define pthread_jit_read()
+#endif
+
 void
 get_simd_vals(priv_mcontext_t *mc);