Skip to content

Commit 6d8f4f7

Browse files
rossbridgerrossbridger
committed
(WIP) See if microsoft/mimalloc#1196 fixes the test error.
Signed-off-by: Ross Charles C. <10781734+rossbridger@users.noreply.github.com>
1 parent 2629719 commit 6d8f4f7

2 files changed

Lines changed: 130 additions & 0 deletions

File tree

‎Code/Framework/AzCore/3rdParty/Findmimalloc.cmake‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -16,12 +16,14 @@ function(Getmimalloc)
1616
set(MIMALLOC_GIT_REPOSITORY "https://github.com/microsoft/mimalloc.git")
1717
set(MIMALLOC_GIT_TAG "a5969c54000f44cd45968e830f6482592373cb0e")
1818
set(MIMALLOC_VERSION_STRING "v3.2.7")
19+
set(MIMALLOC_GIT_PATCH "${CMAKE_CURRENT_LIST_DIR}/pr1196.patch")
1920

2021
FetchContent_Declare(
2122
mimalloc
2223
GIT_REPOSITORY ${MIMALLOC_GIT_REPOSITORY}
2324
GIT_TAG ${MIMALLOC_GIT_TAG}
2425
GIT_SHALLOW TRUE
26+
PATCH_COMMAND cmake -P "${LY_ROOT_FOLDER}/cmake/PatchIfNotAlreadyPatched.cmake" ${MIMALLOC_GIT_PATCH}
2527
EXCLUDE_FROM_ALL
2628
)
2729

‎Code/Framework/AzCore/3rdParty/pr1196.patch‎

Lines changed: 128 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,128 @@
1+
include/mimalloc/prim.h | 18 +++++++++++++++
2+
src/init.c | 61 ++++++++++++++++++++++++++++++++++++++++++++-----
3+
2 files changed, 73 insertions(+), 6 deletions(-)
4+
5+
diff --git a/include/mimalloc/prim.h b/include/mimalloc/prim.h
6+
index 1638235b..5caab9fb 100644
7+
--- a/include/mimalloc/prim.h
8+
+++ b/include/mimalloc/prim.h
9+
@@ -420,12 +420,30 @@ static inline mi_theap_t* _mi_theap_cached(void) {
10+
11+
extern mi_decl_hidden size_t _mi_theap_default_slot;
12+
extern mi_decl_hidden size_t _mi_theap_cached_slot;
13+
+extern mi_decl_hidden bool _mi_theap_use_win_tls_api;
14+
+extern mi_decl_hidden bool _mi_theap_use_compiler_tls;
15+
+extern mi_decl_hidden size_t _mi_theap_default_tls_index;
16+
+extern mi_decl_hidden size_t _mi_theap_cached_tls_index;
17+
+extern mi_decl_hidden mi_decl_thread mi_theap_t* __mi_theap_default_fallback;
18+
+extern mi_decl_hidden mi_decl_thread mi_theap_t* __mi_theap_cached_fallback;
19+
20+
static inline mi_theap_t* _mi_theap_default(void) {
21+
+ if mi_unlikely(_mi_theap_use_compiler_tls) {
22+
+ return __mi_theap_default_fallback;
23+
+ }
24+
+ if mi_unlikely(_mi_theap_use_win_tls_api && _mi_theap_default_tls_index != (size_t)TLS_OUT_OF_INDEXES) {
25+
+ return (mi_theap_t*)TlsGetValue((DWORD)_mi_theap_default_tls_index);
26+
+ }
27+
return (mi_theap_t*)mi_prim_tls_slot(_mi_theap_default_slot); // valid initial "last user slot" so it returns NULL at first leading to slot initialization
28+
}
29+
30+
static inline mi_theap_t* _mi_theap_cached(void) {
31+
+ if mi_unlikely(_mi_theap_use_compiler_tls) {
32+
+ return __mi_theap_cached_fallback;
33+
+ }
34+
+ if mi_unlikely(_mi_theap_use_win_tls_api && _mi_theap_cached_tls_index != (size_t)TLS_OUT_OF_INDEXES) {
35+
+ return (mi_theap_t*)TlsGetValue((DWORD)_mi_theap_cached_tls_index);
36+
+ }
37+
return (mi_theap_t*)mi_prim_tls_slot(_mi_theap_cached_slot);
38+
}
39+
40+
diff --git a/src/init.c b/src/init.c
41+
index c037c432..c3987b0d 100644
42+
--- a/src/init.c
43+
+++ b/src/init.c
44+
@@ -696,13 +696,46 @@ mi_decl_cold mi_decl_noinline mi_theap_t* _mi_theap_empty_get(void) {
45+
mi_decl_hidden size_t _mi_theap_default_slot = MI_TLS_USER_LAST_SLOT;
46+
mi_decl_hidden size_t _mi_theap_cached_slot = MI_TLS_USER_LAST_SLOT;
47+
48+
+// Fallback if we cannot allocate a TLS index in the first 63 TLS slots.
49+
+// In that case we use the Win32 TLS APIs (TlsGetValue/TlsSetValue) instead of directly
50+
+// accessing the TEB TLS user slots.
51+
+mi_decl_hidden bool _mi_theap_use_win_tls_api = false;
52+
+mi_decl_hidden size_t _mi_theap_default_tls_index = (size_t)TLS_OUT_OF_INDEXES;
53+
+mi_decl_hidden size_t _mi_theap_cached_tls_index = (size_t)TLS_OUT_OF_INDEXES;
54+
+
55+
+// Last-resort fallback if the process has exhausted TLS indices (TlsAlloc returns TLS_OUT_OF_INDEXES).
56+
+// We then store the theap pointers in compiler thread-local storage instead.
57+
+mi_decl_hidden bool _mi_theap_use_compiler_tls = false;
58+
+mi_decl_hidden mi_decl_thread mi_theap_t* __mi_theap_default_fallback = (mi_theap_t*)&_mi_theap_empty;
59+
+mi_decl_hidden mi_decl_thread mi_theap_t* __mi_theap_cached_fallback = (mi_theap_t*)&_mi_theap_empty;
60+
+
61+
mi_decl_cold mi_theap_t* _mi_tls_slots_init(void) {
62+
static mi_atomic_once_t tls_slots_init;
63+
if (mi_atomic_once(&tls_slots_init)) {
64+
- _mi_theap_default_slot = TlsAlloc() + MI_TLS_USER_BASE;
65+
- _mi_theap_cached_slot = TlsAlloc() + MI_TLS_USER_BASE;
66+
- if (_mi_theap_cached_slot >= MI_TLS_USER_LAST_SLOT) {
67+
- _mi_error_message(EFAULT, "unable to allocate fast TLS user slot (0x%zx)\n", _mi_theap_cached_slot);
68+
+ const DWORD default_index = TlsAlloc();
69+
+ const DWORD cached_index = TlsAlloc();
70+
+
71+
+ // If we cannot allocate any TLS index at all, we cannot function.
72+
+ if (default_index == TLS_OUT_OF_INDEXES || cached_index == TLS_OUT_OF_INDEXES) {
73+
+ // The process has exhausted all TLS indices. Fall back to compiler TLS.
74+
+ _mi_theap_use_compiler_tls = true;
75+
+ _mi_theap_use_win_tls_api = false;
76+
+ // keep fast slots at sentinel values; do not write to TEB slots.
77+
+ }
78+
+ // We can only use the fast TEB TLS user slots if the TLS index fits in the first 63 slots.
79+
+ // Slot 63 is reserved as the initial sentinel value (so a first read returns NULL).
80+
+ else if (default_index < 63 && cached_index < 63) {
81+
+ _mi_theap_default_slot = (size_t)default_index + MI_TLS_USER_BASE;
82+
+ _mi_theap_cached_slot = (size_t)cached_index + MI_TLS_USER_BASE;
83+
+ _mi_theap_use_win_tls_api = false;
84+
+ _mi_theap_use_compiler_tls = false;
85+
+ }
86+
+ else {
87+
+ // Fall back to Win32 TLS APIs for correctness in TLS-slot-heavy processes.
88+
+ _mi_theap_default_tls_index = (size_t)default_index;
89+
+ _mi_theap_cached_tls_index = (size_t)cached_index;
90+
+ _mi_theap_use_win_tls_api = true;
91+
+ _mi_theap_use_compiler_tls = false;
92+
}
93+
}
94+
return (mi_theap_t*)&_mi_theap_empty;
95+
@@ -732,7 +765,15 @@ void _mi_theap_cached_set(mi_theap_t* theap) {
96+
mi_prim_tls_slot_set(MI_TLS_MODEL_FIXED_SLOT_CACHED, theap);
97+
#elif MI_TLS_MODEL_DYNAMIC_WIN32
98+
_mi_tls_slots_init();
99+
- mi_prim_tls_slot_set(_mi_theap_cached_slot, theap);
100+
+ if (mi_unlikely(_mi_theap_use_compiler_tls)) {
101+
+ __mi_theap_cached_fallback = theap;
102+
+ }
103+
+ else if (mi_unlikely(_mi_theap_use_win_tls_api)) {
104+
+ TlsSetValue((DWORD)_mi_theap_cached_tls_index, theap);
105+
+ }
106+
+ else {
107+
+ mi_prim_tls_slot_set(_mi_theap_cached_slot, theap);
108+
+ }
109+
#elif MI_TLS_MODEL_DYNAMIC_PTHREADS
110+
_mi_tls_keys_init();
111+
if (_mi_theap_cached_key!=0) pthread_setspecific(_mi_theap_cached_key, theap);
112+
@@ -748,7 +789,15 @@ void _mi_theap_default_set(mi_theap_t* theap) {
113+
mi_prim_tls_slot_set(MI_TLS_MODEL_FIXED_SLOT_DEFAULT, theap);
114+
#elif MI_TLS_MODEL_DYNAMIC_WIN32
115+
_mi_tls_slots_init();
116+
- mi_prim_tls_slot_set(_mi_theap_default_slot, theap);
117+
+ if (mi_unlikely(_mi_theap_use_compiler_tls)) {
118+
+ __mi_theap_default_fallback = theap;
119+
+ }
120+
+ else if (mi_unlikely(_mi_theap_use_win_tls_api)) {
121+
+ TlsSetValue((DWORD)_mi_theap_default_tls_index, theap);
122+
+ }
123+
+ else {
124+
+ mi_prim_tls_slot_set(_mi_theap_default_slot, theap);
125+
+ }
126+
#elif MI_TLS_MODEL_DYNAMIC_PTHREADS
127+
_mi_tls_keys_init();
128+
if (_mi_theap_default_key!=0) pthread_setspecific(_mi_theap_default_key, theap);

0 commit comments

Comments
 (0)