From f827d9141913b95a96340bccc360d21475d4a3a2 Mon Sep 17 00:00:00 2001
From: Joel Rosdahl <joel@rosdahl.net>
Date: Thu, 28 Sep 2023 18:05:25 +0200
Subject: [PATCH] bump: Upgrade to BLAKE3 1.5.0

---
 LICENSE.adoc                             |  2 +-
 src/third_party/blake3/blake3.h          |  2 +-
 src/third_party/blake3/blake3_dispatch.c | 39 +++++++++++++++++++++---
 src/third_party/blake3/blake3_impl.h     |  6 +++-
 4 files changed, 41 insertions(+), 8 deletions(-)

diff --git a/LICENSE.adoc b/LICENSE.adoc
index 01cac3dfc..9a8b702f3 100644
--- a/LICENSE.adoc
+++ b/LICENSE.adoc
@@ -72,7 +72,7 @@ along with this program.  If not, see <http://www.gnu.org/licenses/>.
 
 === src/third_party/blake3/blake3_*
 
-This is a subset of https://github.com/BLAKE3-team/BLAKE3[BLAKE3] 1.4.1 with the
+This is a subset of https://github.com/BLAKE3-team/BLAKE3[BLAKE3] 1.5.0 with the
 following license:
 
 ----
diff --git a/src/third_party/blake3/blake3.h b/src/third_party/blake3/blake3.h
index 21e0d7b9d..f694dcf27 100644
--- a/src/third_party/blake3/blake3.h
+++ b/src/third_party/blake3/blake3.h
@@ -30,7 +30,7 @@
 extern "C" {
 #endif
 
-#define BLAKE3_VERSION_STRING "1.4.1"
+#define BLAKE3_VERSION_STRING "1.5.0"
 #define BLAKE3_KEY_LEN 32
 #define BLAKE3_OUT_LEN 32
 #define BLAKE3_BLOCK_LEN 64
diff --git a/src/third_party/blake3/blake3_dispatch.c b/src/third_party/blake3/blake3_dispatch.c
index 2ab0093ee..af6c3dadc 100644
--- a/src/third_party/blake3/blake3_dispatch.c
+++ b/src/third_party/blake3/blake3_dispatch.c
@@ -6,6 +6,7 @@
 
 #if defined(IS_X86)
 #if defined(_MSC_VER)
+#include <Windows.h>
 #include <intrin.h>
 #elif defined(__GNUC__)
 #include <immintrin.h>
@@ -14,6 +15,32 @@
 #endif
 #endif
 
+#if !defined(BLAKE3_ATOMICS)
+#if defined(__has_include)
+#if __has_include(<stdatomic.h>) && !defined(_MSC_VER)
+#define BLAKE3_ATOMICS 1
+#else
+#define BLAKE3_ATOMICS 0
+#endif /* __has_include(<stdatomic.h>) && !defined(_MSC_VER) */
+#else
+#define BLAKE3_ATOMICS 0
+#endif /* defined(__has_include) */
+#endif /* BLAKE3_ATOMICS */
+
+#if BLAKE3_ATOMICS
+#define ATOMIC_INT _Atomic int
+#define ATOMIC_LOAD(x) x
+#define ATOMIC_STORE(x, y) x = y
+#elif defined(_MSC_VER)
+#define ATOMIC_INT LONG
+#define ATOMIC_LOAD(x) InterlockedOr(&x, 0)
+#define ATOMIC_STORE(x, y) InterlockedExchange(&x, y)
+#else
+#define ATOMIC_INT int
+#define ATOMIC_LOAD(x) x
+#define ATOMIC_STORE(x, y) x = y
+#endif
+
 #define MAYBE_UNUSED(x) (void)((x))
 
 #if defined(IS_X86)
@@ -76,7 +103,7 @@ enum cpu_feature {
 #if !defined(BLAKE3_TESTING)
 static /* Allow the variable to be controlled manually for testing */
 #endif
-    enum cpu_feature g_cpu_features = UNDEFINED;
+    ATOMIC_INT g_cpu_features = UNDEFINED;
 
 #if !defined(BLAKE3_TESTING)
 static
@@ -84,14 +111,16 @@ static
     enum cpu_feature
     get_cpu_features(void) {
 
-  if (g_cpu_features != UNDEFINED) {
-    return g_cpu_features;
+  /* If TSAN detects a data race here, try compiling with -DBLAKE3_ATOMICS=1 */
+  enum cpu_feature features = ATOMIC_LOAD(g_cpu_features);
+  if (features != UNDEFINED) {
+    return features;
   } else {
 #if defined(IS_X86)
     uint32_t regs[4] = {0};
     uint32_t *eax = &regs[0], *ebx = &regs[1], *ecx = &regs[2], *edx = &regs[3];
     (void)edx;
-    enum cpu_feature features = 0;
+    features = 0;
     cpuid(regs, 0);
     const int max_id = *eax;
     cpuid(regs, 1);
@@ -124,7 +153,7 @@ static
         }
       }
     }
-    g_cpu_features = features;
+    ATOMIC_STORE(g_cpu_features, features);
     return features;
 #else
     /* How to detect NEON? */
diff --git a/src/third_party/blake3/blake3_impl.h b/src/third_party/blake3/blake3_impl.h
index 3ba9ceb04..beab5cf53 100644
--- a/src/third_party/blake3/blake3_impl.h
+++ b/src/third_party/blake3/blake3_impl.h
@@ -51,7 +51,11 @@ enum blake3_flags {
 #if !defined(BLAKE3_USE_NEON) 
   // If BLAKE3_USE_NEON not manually set, autodetect based on AArch64ness
   #if defined(IS_AARCH64)
-    #define BLAKE3_USE_NEON 1
+    #if defined(__ARM_BIG_ENDIAN)
+      #define BLAKE3_USE_NEON 0
+    #else
+      #define BLAKE3_USE_NEON 1
+    #endif
   #else
     #define BLAKE3_USE_NEON 0
   #endif
-- 
2.47.2