From be2ce7908ee82d5c627e20c15432d1819a14a4af Mon Sep 17 00:00:00 2001
From: Rose <83477269+AtariDreams@users.noreply.github.com>
Date: Thu, 28 Sep 2023 20:06:50 -0400
Subject: [PATCH] Optimize _byteCount with compiler intrinsics

Use compiler intrinsics to make _byteCount faster
---
 .../Parsing.subproj/CFBinaryPList.c           | 29 ++++++++++---------
 1 file changed, 16 insertions(+), 13 deletions(-)

diff --git a/CoreFoundation/Parsing.subproj/CFBinaryPList.c b/CoreFoundation/Parsing.subproj/CFBinaryPList.c
index 54a4877c0bc..c0dbe8f7d11 100644
--- a/CoreFoundation/Parsing.subproj/CFBinaryPList.c
+++ b/CoreFoundation/Parsing.subproj/CFBinaryPList.c
@@ -519,22 +519,25 @@ static void _flattenPlist(CFPropertyListRef plist, CFMutableArrayRef objlist, CF
 /* Get the number of bytes required to hold the value in 'count'. Will return a power of 2 value big enough to hold 'count'.
  */
 CF_INLINE uint8_t _byteCount(uint64_t count) {
-    uint64_t mask = ~(uint64_t)0;
-    uint8_t size = 0;
 
-    // Find something big enough to hold 'count'
-    while (count & mask) {
-        size++;
-        mask = mask << 8;
-    }
+    if (count == 0)
+        return 1; // Special case 0 since it is undefined for __builtin_clzll
 
-    // Ensure that 'count' is a power of 2
-    // For sizes bigger than 8, just use the required count
-    while ((size != 1 && size != 2 && size != 4 && size != 8) && size <= 8) {
-        size++;
-    }
+    // Count the number of leading 0s and subtract from the max value, which is 64
+    int zeroCount = 64 - __builtin_clzll(count);
+
+    // Round to highest by 8 number
+    zeroCount += -zeroCount & 7;
+
+    // Divide by 8
+    zeroCount >>= 3;
+
+    // Anything 8 or above just use the zeroCount
+    if (zeroCount >= 8)
+        return zeroCount;
 
-    return size;
+    // calculate nearest power of 2;
+    return 1U << (32 - __builtin_ctz(zeroCount - 1));
 }
 
 // stream can be a CFWriteStreamRef (on supported platforms) or a CFMutableDataRef