From 2a2e857e08768e70ce46a57c497e77d51c13af2e Mon Sep 17 00:00:00 2001
From: jacob <jacob@cagori.com>
Date: Thu, 4 Apr 2024 18:47:19 -0500
Subject: [PATCH] only initialize tls storage when tls_get is called for first
 time

---
 src/playback_wasapi.c |  2 +-
 src/tls.c             | 56 +++++++++++++++++++++++--------------------
 src/tls.h             | 12 ++++------
 3 files changed, 36 insertions(+), 34 deletions(-)

diff --git a/src/playback_wasapi.c b/src/playback_wasapi.c
index b424b886..90efb163 100644
--- a/src/playback_wasapi.c
+++ b/src/playback_wasapi.c
@@ -12,8 +12,8 @@
 
 #define COBJMACROS
 #define WIN32_LEAN_AND_MEAN
-#include <initguid.h>
 #include <Windows.h>
+#include <initguid.h>
 #include <objbase.h>
 #include <uuids.h>
 #include <avrt.h>
diff --git a/src/tls.c b/src/tls.c
index 62bc8fc3..ea46bc02 100644
--- a/src/tls.c
+++ b/src/tls.c
@@ -4,7 +4,7 @@
 #include "scratch.h"
 #include "work.h"
 
-#define TLS_TABLE_RESERVE (GIGABYTE(1))
+#define TLS_TABLE_RESERVE (MEGABYTE(1))
 
 struct tls_info {
   u64 size;
@@ -22,39 +22,43 @@ GLOBAL READONLY struct tls_info g_tls_info_table[TLS_IDENTIFIER_COUNT] = {
 struct tls_table tls_table_alloc(void)
 {
     struct tls_table t = { 0 };
-
     t.arena = arena_alloc(TLS_TABLE_RESERVE);
-
-    /* Build lookup table */
-    for (u64 i = 0; i < TLS_IDENTIFIER_COUNT; ++i) {
-        struct tls_info *info = &g_tls_info_table[i];
-        arena_align(&t.arena, info->align);
-        void *ptr = arena_push_array_zero(&t.arena, u8, info->size);
-        t.lookup[i] = ptr;
-    }
-
-    /* Call alloc functions */
-    for (u64 i = 0; i < TLS_IDENTIFIER_COUNT; ++i) {
-        tls_alloc_func *alloc = g_tls_info_table[i].alloc;
-        if (alloc) {
-            void *ptr = t.lookup[i];
-            alloc(ptr);
-        }
-    }
-
     return t;
 }
 
 void tls_table_release(struct tls_table *t)
 {
-    /* Call release functions in reverse order */
-    for (u64 i = (TLS_IDENTIFIER_COUNT - 1); i <= 0; --i) {
-        tls_alloc_func *release = g_tls_info_table[i].release;
-        if (release) {
-            void *ptr = t->lookup[i];
-            release(ptr);
+    /* Release allocated tls data in reverse order */
+    for (u64 index_plus_one = t->allocation_order_count; index_plus_one > 0; --index_plus_one) {
+        enum tls_identifier identifier = t->allocation_order[index_plus_one - 1];
+        void *data = t->lookup[identifier];
+        struct tls_info *info = &g_tls_info_table[identifier];
+        if (info->release) {
+            info->release(data);
         }
     }
 
     arena_release(&t->arena);
 }
+
+void *tls_get(enum tls_identifier identifier)
+{
+    struct tls_table *t = sys_thread_get_tls();
+    void **data_slot = &t->lookup[identifier];
+
+    if (!*data_slot) {
+        /* Allocate */
+        struct tls_info *info = &g_tls_info_table[identifier];
+        arena_align(&t->arena, info->align);
+        *data_slot = arena_push_array(&t->arena, u8, info->size);
+        if (info->alloc) {
+            info->alloc(*data_slot);
+        } else {
+            MEMZERO(*data_slot, info->size);
+        }
+        t->allocation_order[t->allocation_order_count] = identifier;
+        ++t->allocation_order_count;
+    }
+
+    return *data_slot;
+}
diff --git a/src/tls.h b/src/tls.h
index b7bef0bc..a776d65b 100644
--- a/src/tls.h
+++ b/src/tls.h
@@ -5,8 +5,8 @@
 /* TLS table (X macro)                                                        */
 /* ========================================================================== */
 
-#define TLS_TABLE(X) \
-    X(SCRATCH_TLS,  struct scratch_tls, &scratch_tls_alloc, &scratch_tls_release) \
+#define TLS_TABLE(X)  \
+    X(SCRATCH_TLS,  struct scratch_tls, &scratch_tls_alloc, &scratch_tls_release)  \
     X(WORKER_TLS,   struct worker_tls,  NULL,               NULL                )
 
 /* ========================================================================== */
@@ -31,14 +31,12 @@ enum tls_identifier {
 struct tls_table {
     struct arena arena;
     void *lookup[TLS_IDENTIFIER_COUNT];
+    enum tls_identifier allocation_order[TLS_IDENTIFIER_COUNT];
+    u64 allocation_order_count;
 };
 
 struct tls_table tls_table_alloc(void);
 void tls_table_release(struct tls_table *t);
-
-INLINE void *tls_get(enum tls_identifier identifier)
-{
-    return sys_thread_get_tls()->lookup[identifier];
-}
+void *tls_get(enum tls_identifier identifier);
 
 #endif