dynamorio: adopt better patch for fp saving issue

patches/dynamorio-10.0.patch

@@ -1,100 +1,3 @@
-diff --git a/api/samples/memtrace_x86.c b/api/samples/memtrace_x86.c
-index 4e726c5b2..b229086bf 100644
---- a/api/samples/memtrace_x86.c
-+++ b/api/samples/memtrace_x86.c
-@@ -260,6 +260,20 @@ event_thread_exit(void *drcontext)
-
-     memtrace(drcontext);
-     data = drmgr_get_tls_field(drcontext, tls_index);
-+
-+#ifdef OUTPUT_TEXT
-+    const char *dump_modules[0x4] = { "libc.so.6", "daydream", "ld-linux-x86-64.so.2", NULL };
-+    for (int i = 0; i < 0x4; i++) {
-+        if (dump_modules[i] == NULL) break;
-+        module_data_t *mod = dr_lookup_module_by_name(dump_modules[i]);
-+        if (mod) {
-+            fprintf(data->logf, "BASE %s %p %p\n", dump_modules[i], mod->start, mod->end);
-+        } else {
-+            fprintf(data->logf, "Could not get base for %s\n", dump_modules[i]);
-+        }
-+    }
-+#endif
-+
-     dr_mutex_lock(mutex);
-     global_num_refs += data->num_refs;
-     dr_mutex_unlock(mutex);
-diff --git a/core/arch/interp.c b/core/arch/interp.c
-index dc786a904..80e350399 100644
---- a/core/arch/interp.c
-+++ b/core/arch/interp.c
-@@ -2700,6 +2700,8 @@ client_check_syscall(instrlist_t *ilist, instr_t *inst, bool *found_syscall,
-     return true;
- }
-
-+void    print_xmm0(int);
-+
- /* Pass bb to client, and afterward check for criteria we require and rescan for
-  * eflags and other flags that might have changed.
-  * Returns true normally; returns false to indicate "go native".
-@@ -2707,6 +2709,7 @@ client_check_syscall(instrlist_t *ilist, instr_t *inst, bool *found_syscall,
- static bool
- client_process_bb(dcontext_t *dcontext, build_bb_t *bb)
- {
-+
-     dr_emit_flags_t emitflags = DR_EMIT_DEFAULT;
-     instr_t *inst;
-     bool found_exit_cti = false;
-@@ -5110,6 +5113,7 @@ build_basic_block_fragment(dcontext_t *dcontext, app_pc start, uint initial_flag
-     KSTART(bb_building);
-     dcontext->whereami = DR_WHERE_INTERP;
-
-+
-     /* Neither thin_client nor hotp_only should be building any bbs. */
-     ASSERT(!RUNNING_WITHOUT_CODE_CACHE());
-
-@@ -5118,7 +5122,10 @@ build_basic_block_fragment(dcontext_t *dcontext, app_pc start, uint initial_flag
-      */
-     image_entry = check_for_image_entry(start);
-
-+
-     init_interp_build_bb(dcontext, &bb, start, initial_flags, for_trace, unmangled_ilist);
-+
-+
-     if (at_native_exec_gateway(dcontext, start,
-                                &bb.native_call _IF_DEBUG(false /*not xfer tgt*/))) {
-         DODEBUG({ report_native_module(dcontext, bb.start_pc); });
-@@ -5161,11 +5168,13 @@ build_basic_block_fragment(dcontext_t *dcontext, app_pc start, uint initial_flag
-         jitopt_add_dgc_bb(bb.start_pc, bb.end_pc, TEST(FRAG_IS_TRACE_HEAD, bb.flags));
-     }
-
-+
-     /* emit fragment into fcache */
-     KSTART(bb_emit);
-     f = emit_fragment_ex(dcontext, start, bb.ilist, bb.flags, bb.vmlist, link, visible);
-     KSTOP(bb_emit);
-
-+
- #ifdef CUSTOM_TRACES_RET_REMOVAL
-     f->num_calls = dcontext->num_calls;
-     f->num_rets = dcontext->num_rets;
-diff --git a/core/dispatch.c b/core/dispatch.c
-index 37e6bb531..c91394ea9 100644
---- a/core/dispatch.c
-+++ b/core/dispatch.c
-@@ -121,6 +121,13 @@ exited_due_to_ni_syscall(dcontext_t *dcontext)
-     return false;
- }
-
-+void print_xmm0(int id) {
-+    unsigned long xmm0, xmm1;
-+    asm volatile("movq %%xmm0, %0" : "=r" (xmm0));
-+    asm volatile("movq %%xmm1, %0" : "=r" (xmm1));
-+    dr_fprintf(STDERR, "xmm0 @ %d: %lx, xmm1: %lx\n", id, xmm0, xmm1);
-+    dr_flush_file(STDERR);
-+}
- /* This is the central hub of control management in DynamoRIO.
-  * It is entered with a clean dstack at startup and after every cache
-  * exit, whether normal or kernel-mediated via a trampoline context switch.
 diff --git a/core/heap.c b/core/heap.c
 index 8a0c440cf..f596108da 100644
 --- a/core/heap.c
@@ -108,42 +11,6 @@ index 8a0c440cf..f596108da 100644
          });
  #endif
 
-diff --git a/core/ir/disassemble_shared.c b/core/ir/disassemble_shared.c
-index 19241476c..778b0532f 100644
---- a/core/ir/disassemble_shared.c
-+++ b/core/ir/disassemble_shared.c
-@@ -611,6 +611,13 @@ print_known_pc_target(char *buf, size_t bufsz, size_t *sofar INOUT, dcontext_t *
-     return printed;
- }
-
-+static void double_print2(opnd_t opnd, uint *top, uint *bottom, const char **sign, bool double_type) {
-+    if (double_type)
-+        double_print(opnd_get_immed_double(opnd), 6, top, bottom, sign);
-+    else
-+        double_print(opnd_get_immed_float(opnd), 6, top, bottom, sign);
-+}
-+
- void
- internal_opnd_disassemble(char *buf, size_t bufsz, size_t *sofar INOUT,
-                           dcontext_t *dcontext, opnd_t opnd, bool use_size_sfx)
-@@ -667,7 +674,7 @@ internal_opnd_disassemble(char *buf, size_t bufsz, size_t *sofar INOUT,
-             uint top;
-             uint bottom;
-             const char *sign;
--            double_print(opnd_get_immed_float(opnd), 6, &top, &bottom, &sign);
-+            double_print2(opnd, &top, &bottom, &sign, false);
-             print_to_buffer(buf, bufsz, sofar, "%s%s%u.%.6u", immed_prefix(), sign, top,
-                             bottom);
-         });
-@@ -683,7 +690,7 @@ internal_opnd_disassemble(char *buf, size_t bufsz, size_t *sofar INOUT,
-             uint top;
-             uint bottom;
-             const char *sign;
--            double_print(opnd_get_immed_double(opnd), 6, &top, &bottom, &sign);
-+            double_print2(opnd, &top, &bottom, &sign, true);
-             print_to_buffer(buf, bufsz, sofar, "%s%s%u.%.6u", immed_prefix(), sign, top,
-                             bottom);
-         });
 diff --git a/core/lib/dr_tools.h b/core/lib/dr_tools.h
 index efbb1c636..b7ab5d060 100644
 --- a/core/lib/dr_tools.h
@@ -378,7 +245,7 @@ index f5eef1e5c..91f5a16bd 100644
      /* grab all_threads_synch_lock */
      /* since all_threads synch doesn't give any permissions this is necessary
 diff --git a/core/unix/loader.c b/core/unix/loader.c
-index 10c4518b0..37aa2aa0c 100644
+index 10c4518b0..b27284a8e 100644
 --- a/core/unix/loader.c
 +++ b/core/unix/loader.c
 @@ -158,7 +158,7 @@ privload_locate_and_load(const char *impname, privmod_t *dependent, bool reachab
@@ -489,13 +356,10 @@ index 10c4518b0..37aa2aa0c 100644
      { "dlsym", (app_pc)redirect_dlsym },
      /* We need these for clients that don't use libc (i#1747) */
      { "strlen", (app_pc)strlen },
-@@ -1569,7 +1577,15 @@ static const redirect_import_t redirect_imports[] = {
+@@ -1569,7 +1577,12 @@ static const redirect_import_t redirect_imports[] = {
      { "memset_chk", (app_pc)memset },
      { "memmove_chk", (app_pc)memmove },
      { "strncpy_chk", (app_pc)strncpy },
-+    { "__memcpy_chk", (app_pc)memcpy },
-+    { "__memset_chk", (app_pc)memset },
-+    { "__memmove_chk", (app_pc)memmove },
 +    /* { "__errno_location", (app_pc)__errno_location } */
  };
 +
@@ -505,7 +369,7 @@ index 10c4518b0..37aa2aa0c 100644
  #define REDIRECT_IMPORTS_NUM (sizeof(redirect_imports) / sizeof(redirect_imports[0]))
 
  #ifdef DEBUG
-@@ -1599,6 +1615,15 @@ privload_redirect_sym(os_privmod_data_t *opd, ptr_uint_t *r_addr, const char *na
+@@ -1599,6 +1612,15 @@ privload_redirect_sym(os_privmod_data_t *opd, ptr_uint_t *r_addr, const char *na
          }
      }
  #endif
@@ -817,14 +681,14 @@ index d98df70db..f106ce611 100644
  bool
  privload_redirect_sym(os_privmod_data_t *opd, ptr_uint_t *r_addr, const char *name);
 diff --git a/core/unix/os.c b/core/unix/os.c
-index d5133bf16..0efaa2140 100644
+index d5133bf16..106d0672a 100644
 --- a/core/unix/os.c
 +++ b/core/unix/os.c
 @@ -3761,8 +3761,13 @@ os_thread_sleep(uint64 milliseconds)
          /* not unusual for client threads to use itimers and have their run
           * routine sleep forever
           */
-+        
++
          if (count++ > 3 && !IS_CLIENT_THREAD(get_thread_private_dcontext())) {
 -            ASSERT_NOT_REACHED();
 +            // ATH: We are hitting this case in release builds, and it seems to be breaking
@@ -869,43 +733,38 @@ index 4d3b9e60f..4ce713450 100644
 
  int
 diff --git a/core/unix/signal.c b/core/unix/signal.c
-index 0cbc94337..a73aa85bf 100644
+index 0cbc94337..51ef52f80 100644
 --- a/core/unix/signal.c
 +++ b/core/unix/signal.c
-@@ -3197,12 +3197,12 @@ thread_set_self_context(void *cxt)
-     dcontext_t *dcontext = get_thread_private_dcontext();
+@@ -3193,20 +3193,10 @@ thread_set_self_context(void *cxt)
+     ASSERT_NOT_IMPLEMENTED(false); /* PR 405694: can't use regular sigreturn! */
  #endif
+     memset(&frame, 0, sizeof(frame));
+-#if defined(X86)
+-    dcontext_t *dcontext = get_thread_private_dcontext();
+-#endif
  #ifdef LINUX
-+    frame.uc.uc_mcontext = *sc;
- #    ifdef X86
+-#    ifdef X86
 -    byte *xstate = get_and_initialize_xstate_buffer(dcontext);
 -    frame.uc.uc_mcontext.fpstate = &((kernel_xstate_t *)xstate)->fpstate;
-+    frame.uc.uc_mcontext.fpstate = (void*)get_and_initialize_xstate_buffer(dcontext);
- #    endif /* X86 */
--    frame.uc.uc_mcontext = *sc;
--#endif
-+#endif /* inlux*/
-+
+-#    endif /* X86 */
+     frame.uc.uc_mcontext = *sc;
+ #endif
      IF_ARM(ASSERT_NOT_TESTED());
- #if defined(X86)
-     save_fpstate(dcontext, &frame);
-@@ -8552,15 +8552,19 @@ handle_suspend_signal(dcontext_t *dcontext, kernel_siginfo_t *siginfo,
-     dcontext->whereami = prior_whereami;
-
-     if (ostd->retakeover) {
-+        LOG(THREAD, LOG_ASYNCH, 2, "->retakeover\n");
-         ostd->retakeover = false;
-         sig_take_over(ucxt); /* shouldn't return for this case */
-         ASSERT_NOT_REACHED();
-     } else if (ostd->do_detach) {
-+        LOG(THREAD, LOG_ASYNCH, 2, "->do_detach\n");
-         ostd->do_detach = false;
-         sig_detach(dcontext, frame, &ostd->detached); /* no return */
-         ASSERT_NOT_REACHED();
-     }
-
-+    LOG(THREAD, LOG_ASYNCH, 2, "->return\n");
-+
-     return false; /* do not pass to app */
- }
-
+-#if defined(X86)
+-    save_fpstate(dcontext, &frame);
+-#endif
+     /* The kernel calls do_sigaltstack on sys_rt_sigreturn primarily to ensure
+      * the frame is ok, but the side effect is we can mess up our own altstack
+      * settings if we're not careful.  Having invalid ss_size looks good for
+@@ -3295,7 +3285,9 @@ thread_set_self_mcontext(priv_mcontext_t *mc)
+     sig_full_cxt_t sc_full;
+     sig_full_initialize(&sc_full, &ucxt);
+ #if defined(LINUX) && defined(X86)
+-    sc_full.sc->fpstate = NULL; /* for mcontext_to_sigcontext */
++    /* for mcontext_to_sigcontext to fill in with saved fp state */
++     sc_full.sc->fpstate = (kernel_fpstate_t *)get_and_initialize_xstate_buffer(
++         get_thread_private_dcontext());
+ #endif
+     mcontext_to_sigcontext(&sc_full, mc, DR_MC_ALL);
+     thread_set_segment_registers(sc_full.sc);