From 6b9680711235cee63f40f2c39977d8cc0fa54c43 Mon Sep 17 00:00:00 2001
From: Nekotekina <nekotekina@gmail.com>
Date: Tue, 29 Dec 2020 15:28:02 +0300
Subject: [PATCH] Reimplement utils::popcnt64

Implement utils::popcnt128
---
 rpcs3/Emu/Cell/Modules/cellSpurs.cpp | 11 +++-------
 rpcs3/util/asm.hpp                   | 33 ++++++++++++++++++++++++++++
 rpcs3/util/atomic.cpp                | 10 +--------
 3 files changed, 37 insertions(+), 17 deletions(-)

diff --git a/rpcs3/Emu/Cell/Modules/cellSpurs.cpp b/rpcs3/Emu/Cell/Modules/cellSpurs.cpp
index e8f43583fb..39f70f4efa 100644
--- a/rpcs3/Emu/Cell/Modules/cellSpurs.cpp
+++ b/rpcs3/Emu/Cell/Modules/cellSpurs.cpp
@@ -15,6 +15,7 @@
 #include "sysPrxForUser.h"
 #include "cellSpurs.h"
 
+#include "util/asm.hpp"
 #include "util/v128.hpp"
 #include "util/v128sse.hpp"
 
@@ -3930,14 +3931,8 @@ s32 _spurs::create_task(vm::ptr<CellSpursTaskset> taskset, vm::ptr<u32> task_id,
 		if (ls_pattern)
 		{
 			v128 ls_pattern_128 = v128::from64r(ls_pattern->_u64[0], ls_pattern->_u64[1]);
-			u32 ls_blocks       = 0;
-			for (auto i = 0; i < 128; i++)
-			{
-				if (ls_pattern_128._bit[i])
-				{
-					ls_blocks++;
-				}
-			}
+
+			const u32 ls_blocks = utils::popcnt128(ls_pattern_128._u);
 
 			if (ls_blocks > alloc_ls_blocks)
 			{
diff --git a/rpcs3/util/asm.hpp b/rpcs3/util/asm.hpp
index 4f08439afa..c4b8f1f4b9 100644
--- a/rpcs3/util/asm.hpp
+++ b/rpcs3/util/asm.hpp
@@ -19,6 +19,7 @@ extern "C"
 	ushort _rotl16(ushort, uchar);
 	uint _rotl(uint, int);
 	u64 _rotl64(u64, int);
+	u64 __popcnt64(u64);
 
 	s64 __mulh(s64, s64);
 	u64 __umulh(u64, u64);
@@ -211,6 +212,38 @@ namespace utils
 #endif
 	}
 
+	constexpr u32 popcnt64(u64 v)
+	{
+#if !defined(_MSC_VER) || defined(__SSE4_2__)
+		if (std::is_constant_evaluated())
+#endif
+		{
+			v = (v & 0xaaaaaaaaaaaaaaaa) / 2 + (v & 0x5555555555555555);
+			v = (v & 0xcccccccccccccccc) / 4 + (v & 0x3333333333333333);
+			v = (v & 0xf0f0f0f0f0f0f0f0) / 16 + (v & 0x0f0f0f0f0f0f0f0f);
+			v = (v & 0xff00ff00ff00ff00) / 256 + (v & 0x00ff00ff00ff00ff);
+			v = ((v & 0xffff0000ffff0000) >> 16) + (v & 0x0000ffff0000ffff);
+			return static_cast<u32>((v >> 32) + v);
+		}
+
+#if !defined(_MSC_VER) || defined(__SSE4_2__)
+#ifdef _MSC_VER
+		return static_cast<u32>(__popcnt64(v));
+#else
+		return __builtin_popcountll(v);
+#endif
+#endif
+	}
+
+	constexpr u32 popcnt128(const u128& v)
+	{
+#ifdef _MSC_VER
+		return popcnt64(v.lo) + popcnt64(v.hi);
+#else
+		return popcnt64(v) + popcnt64(v >> 64);
+#endif
+	}
+
 	constexpr u64 umulh64(u64 x, u64 y)
 	{
 #ifdef _MSC_VER
diff --git a/rpcs3/util/atomic.cpp b/rpcs3/util/atomic.cpp
index bff204a8c2..32107e4d2f 100644
--- a/rpcs3/util/atomic.cpp
+++ b/rpcs3/util/atomic.cpp
@@ -162,15 +162,7 @@ static NEVER_INLINE bool ptr_cmp(const void* data, u32 _size, u128 old128, u128
 			// Count is taken from least significant byte and ignores some flags
 			const u64 count = static_cast<u64>(old128) & 0xff;
 
-			u64 bitc = new_value;
-			bitc = (bitc & 0xaaaaaaaaaaaaaaaa) / 2 + (bitc & 0x5555555555555555);
-			bitc = (bitc & 0xcccccccccccccccc) / 4 + (bitc & 0x3333333333333333);
-			bitc = (bitc & 0xf0f0f0f0f0f0f0f0) / 16 + (bitc & 0x0f0f0f0f0f0f0f0f);
-			bitc = (bitc & 0xff00ff00ff00ff00) / 256 + (bitc & 0x00ff00ff00ff00ff);
-			bitc = ((bitc & 0xffff0000ffff0000) >> 16) + (bitc & 0x0000ffff0000ffff);
-			bitc = (bitc >> 32) + bitc;
-
-			result = count < bitc;
+			result = count < utils::popcnt64(new_value);
 			break;
 		}
 		default: