Merge pull request #334 from snowstormxd/codex/upstream-mkw-test

MKW performance improvements

2 files changed, 16 insertions, 8 deletions

diff --git a/src/core/hle/kernel/k_priority_queue.h b/src/core/hle/kernel/k_priority_queue.h
index 89a0c5bbea..5d93455d33 100644
--- a/src/core/hle/kernel/k_priority_queue.h
+++ b/src/core/hle/kernel/k_priority_queue.h
@@ -11,6 +11,7 @@
 #include <bitset>
 #include <concepts>
 #include <cstddef>
+#include <type_traits>
 
 #include "common/alignment.h"
 #include "common/assert.h"
@@ -196,12 +197,7 @@ public:
         constexpr Member* GetFront(s32 core) const {
             ASSERT(IsValidCore(core));
 
-            const s32 priority = s32([](auto const& e) {
-                for (size_t i = 0; i < e.size(); ++i)
-                    if (e[i])
-                        return i;
-                return e.size();
-            }(m_available_priorities[core]));
+            const s32 priority = s32(GetNextSet(m_available_priorities[core], size_t(-1)));
             if (priority <= LowestPriority) {
                 return m_queues[priority].GetFront(core);
             } else {
@@ -222,7 +218,17 @@ public:
 
         template<size_t N>
         constexpr size_t GetNextSet(std::bitset<N> const& bit, size_t n) const {
-            for (size_t i = n + 1; i < bit.size(); i++)
+            const size_t first = n + 1;
+            if (first >= bit.size()) {
+                return bit.size();
+            }
+            if constexpr (N <= 64) {
+                if (!std::is_constant_evaluated()) {
+                    const u64 remaining = bit.to_ullong() >> first;
+                    return remaining != 0 ? first + size_t(std::countr_zero(remaining)) : bit.size();
+                }
+            }
+            for (size_t i = first; i < bit.size(); i++)
                 if (bit[i])
                     return i;
             return bit.size();
diff --git a/src/core/hle/kernel/kernel.cpp b/src/core/hle/kernel/kernel.cpp
index f006d129a7..9e6ccc2d89 100644
--- a/src/core/hle/kernel/kernel.cpp
+++ b/src/core/hle/kernel/kernel.cpp
@@ -79,7 +79,7 @@ struct KernelCore::Impl {
     // And we have the guarantee that the data won't move out of the way so we can safely
     // take a reference to it. This isn't always universally true but this is "global" data
     // so it will be statically given a TLS slot anyways.
-    static inline thread_local ThreadLocalData tls_data = {};
+    static thread_local ThreadLocalData tls_data;
 
     explicit Impl(Core::System& system_, KernelCore& kernel_) : system{system_} {
         tls_data.lock = true;
@@ -861,6 +861,8 @@ struct KernelCore::Impl {
     Core::System& system;
 };
 
+thread_local ThreadLocalData KernelCore::Impl::tls_data{};
+
 KernelCore::KernelCore(Core::System& system) : impl{std::make_unique<Impl>(system, *this)} {}
 KernelCore::~KernelCore() = default;