i#5383 mac a64, part 5: Get hello,world to run (#6184)

Fixes global references in aarch64.asm which were loading the value
instead of the address (the Linux a64 assembler auto-magically gives the
address whether using ADD or LDR!).

Fixes and clarifies the Mac64 x86 and arm TLS slots vs offsets.

Uses gettimeofday() for query_time_seconds() instead of
SYS_gettimeofday.

Removes an invalid hardcoded 0 sysnum for global_do_syscall.

Issue: #5383

core/arch/aarch64/aarch64.asm

@@ -1,5 +1,5 @@
 /* **********************************************************
- * Copyright (c) 2019-2022 Google, Inc. All rights reserved.
+ * Copyright (c) 2019-2023 Google, Inc. All rights reserved.
  * Copyright (c) 2016 ARM Limited. All rights reserved.
  * **********************************************************/
 
@@ -105,6 +105,12 @@ DECL_EXTERN(dr_setjmp_sigmask)
 
 DECL_EXTERN(d_r_internal_error)
 
+DECL_EXTERN(exiting_thread_count)
+DECL_EXTERN(d_r_initstack)
+DECL_EXTERN(initstack_mutex)
+DECL_EXTERN(icache_op_struct)
+DECL_EXTERN(linkstub_selfmod)
+
 /* For debugging: report an error if the function called by call_switch_stack()
  * unexpectedly returns.  Also used elsewhere.
  */
@@ -325,7 +331,7 @@ GLOBAL_LABEL(cleanup_and_terminate:)
 #endif
 
         /* inc exiting_thread_count to avoid being killed once off all_threads list */
-        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(exiting_thread_count), x0)
+        AARCH64_ADRP_GOT(GLOBAL_REF(exiting_thread_count), x0)
         CALLC2(GLOBAL_REF(atomic_add), x0, #1)
 
         /* save dcontext->dstack for freeing later and set dcontext->is_exiting */
@@ -349,7 +355,7 @@ cat_thread_only:
         CALLC0(GLOBAL_REF(dynamo_thread_exit))
 cat_no_thread:
         /* switch to d_r_initstack for cleanup of dstack */
-        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(initstack_mutex), x26)
+        AARCH64_ADRP_GOT(GLOBAL_REF(initstack_mutex), x26)
 cat_spin:
         CALLC2(GLOBAL_REF(atomic_swap), x26, #1)
         cbz      w0, cat_have_lock
@@ -358,20 +364,20 @@ cat_spin:
 
 cat_have_lock:
         /* switch stack */
-        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(d_r_initstack), x0)
+        AARCH64_ADRP_GOT(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 */
-        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(initstack_mutex), x0)
+        AARCH64_ADRP_GOT(GLOBAL_REF(initstack_mutex), x0)
         mov      x1, #0
         str      x1, [x0]
 
         /* dec exiting_thread_count (allows another thread to kill us) */
-        AARCH64_ADRP_GOT_LDR(GLOBAL_REF(exiting_thread_count), x0)
+        AARCH64_ADRP_GOT(GLOBAL_REF(exiting_thread_count), x0)
         CALLC2(GLOBAL_REF(atomic_add), x0, #-1)
 
         /* put system call number in x8 */
@@ -401,7 +407,6 @@ 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 */

core/arch/aarchxx/emit_utils.c

@@ -1,5 +1,5 @@
 /* **********************************************************
- * Copyright (c) 2014-2022 Google, Inc.  All rights reserved.
+ * Copyright (c) 2014-2023 Google, Inc.  All rights reserved.
  * Copyright (c) 2016 ARM Limited. All rights reserved.
  * **********************************************************/
 
@@ -66,9 +66,6 @@ insert_load_dr_tls_base(dcontext_t *dcontext, instrlist_t *ilist, instr_t *where
     /* ldr dr_reg_stolen, [reg_base, DR_TLS_BASE_OFFSET] */
     PRE(ilist, where,
         XINST_CREATE_load(dcontext, opnd_create_reg(dr_reg_stolen),
-                          /* TODO i#5383: For MACOS && AARCH64 we need to verify the
-                           * offset: PR #5497 had an 8x multiplier here.
-                           */
                           OPND_CREATE_MEMPTR(reg_base, DR_TLS_BASE_OFFSET)));
 }
 

core/ir/aarch64/codec.c

@@ -9168,7 +9168,7 @@ decode_common(dcontext_t *dcontext, byte *pc, byte *orig_pc, instr_t *instr)
          */
         if ((enc & 0xfffff01f) == 0xd503201f) {
             SYSLOG_INTERNAL_WARNING("Undefined HINT instruction found: "
-                                    "encoding 0x%x (CRm:op2 0x%x)\n",
+                                    "encoding 0x%x (CRm:op2 0x%x)",
                                     enc, (enc & 0xfe0) >> 5);
             instr_set_opcode(instr, OP_nop);
             instr_set_num_opnds(dcontext, instr, 0, 0);

core/unix/os.c

@@ -1329,14 +1329,22 @@ uint
 query_time_seconds(void)
 {
     struct timeval current_time;
+#if defined(MACOS) && defined(AARCH64)
+    /* TODO i#5383: Replace with a system call (unless we're sure this library call
+     * is re-entrant, just a simple load from commpage).
+     */
+#    undef gettimeofday /* Remove "gettimeofday_forbidden_function". */
+    uint64 val = gettimeofday(&current_time, NULL);
+#else
     uint64 val = dynamorio_syscall(SYS_gettimeofday, 2, &current_time, NULL);
-#ifdef MACOS
+#    ifdef MACOS
     /* MacOS before Sierra returns usecs:secs and does not set the timeval struct. */
     if (macos_version < MACOS_VERSION_SIERRA) {
         if ((int)val < 0)
             return 0;
         return (uint)val + UTC_TO_EPOCH_SECONDS;
     }
+#    endif
 #endif
     if ((int)val >= 0) {
         return current_time.tv_sec + UTC_TO_EPOCH_SECONDS;
@@ -1358,7 +1366,9 @@ query_time_millis()
 #if !(defined(MACOS) && defined(AARCH64))
     uint64 val = dynamorio_syscall(SYS_gettimeofday, 2, &current_time, NULL);
 #else
-    /* TODO i#5383: Replace with a system call. */
+    /* TODO i#5383: Replace with a system call (unless we're sure this library call
+     * is re-entrant, just a simple load from commpage).
+     */
 #    undef gettimeofday /* Remove "gettimeofday_forbidden_function". */
     uint64 val = gettimeofday(&current_time, NULL);
 #endif
@@ -2103,7 +2113,7 @@ get_segment_base(uint seg)
 {
 #if defined(MACOS64) && defined(X86)
     ptr_uint_t *pthread_self = (ptr_uint_t *)read_thread_register(seg);
-    return (byte *)&pthread_self[SEG_TLS_BASE_OFFSET];
+    return (byte *)&pthread_self[SEG_TLS_BASE_SLOT];
 #elif defined(X86)
     if (seg == SEG_CS || seg == SEG_SS || seg == SEG_DS || seg == SEG_ES)
         return NULL;

core/unix/os_exports.h

@@ -152,10 +152,11 @@
  * limited interoperability w/ code targeting the Windows x64 ABI. We steal slot 6
  * for our own use.
  */
-#    define SEG_TLS_BASE_OFFSET 28 /* offset from pthread_t struct to segment base */
-#    define DR_TLS_BASE_SLOT 6     /* the TLS slot for DR's TLS base */
+/* XXX i#5383: This is used as *8 so it's really a slot not a byte offset. */
+#    define SEG_TLS_BASE_SLOT 28 /* offset from pthread_t struct to segment base */
+#    define DR_TLS_BASE_SLOT 6   /* the TLS slot for DR's TLS base */
 /* offset from pthread_t struct to slot 6 */
-#    define DR_TLS_BASE_OFFSET (SEG_TLS_BASE_OFFSET + DR_TLS_BASE_SLOT)
+#    define DR_TLS_BASE_OFFSET (sizeof(void *) * (SEG_TLS_BASE_SLOT + DR_TLS_BASE_SLOT))
 #endif
 
 #if defined(AARCHXX) && !defined(MACOS64)

core/unix/tls_macos.c

@@ -1,5 +1,5 @@
 /* *******************************************************************************
- * Copyright (c) 2013-2022 Google, Inc.  All rights reserved.
+ * Copyright (c) 2013-2023 Google, Inc.  All rights reserved.
  * *******************************************************************************/
 
 /*
@@ -185,7 +185,7 @@ tls_get_dr_addr(void)
 byte **
 get_app_tls_swap_slot_addr(void)
 {
-    byte **app_tls_base = (byte **)read_thread_register(TLS_REG_LIB);
+    byte *app_tls_base = (byte *)read_thread_register(TLS_REG_LIB);
     if (app_tls_base == NULL) {
         ASSERT_NOT_IMPLEMENTED(false);
     }
@@ -199,7 +199,12 @@ tls_thread_init(os_local_state_t *os_tls, byte *segment)
 #ifdef X64
     /* For now we have both a directly-addressable os_local_state_t and a pointer to
      * it in slot 6.  If we settle on always doing the full os_local_state_t in slots,
-     * we would probably get rid of the use of slot 6.
+     * we would probably get rid of the use of slot 6 on x86 (on aarch64 the
+     * os_local_state_t slots are not directly addressible; we rely on the stolen
+     * register, whose value is populated from the pointer in slot 6 -- which could
+     * be moved to a slot right before os_local_state_t or something I suppose, or
+     * we could move the whole os_local_state_t to our own mmap since we access
+     * through a pointer anyway).
      */
     byte **tls_swap_slot;
     ASSERT((byte *)(os_tls->self) == segment);