From 49b28f17578d47534dd31762d79751d430323b9c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Micha=C5=82=20Janiszewski?=
 <janisozaur@users.noreply.github.com>
Date: Thu, 25 Apr 2024 16:14:17 +0200
Subject: [PATCH] Prefetch paint struct bounds (#21894)

With restored `benchspritesort` this yields around 5% better results.
---
 src/openrct2/libopenrct2.vcxproj |  1 +
 src/openrct2/paint/Paint.cpp     |  2 ++
 src/openrct2/util/Prefetch.h     | 34 ++++++++++++++++++++++++++++++++
 3 files changed, 37 insertions(+)
 create mode 100644 src/openrct2/util/Prefetch.h

diff --git a/src/openrct2/libopenrct2.vcxproj b/src/openrct2/libopenrct2.vcxproj
index 3dda36af25..be99f4ae3e 100644
--- a/src/openrct2/libopenrct2.vcxproj
+++ b/src/openrct2/libopenrct2.vcxproj
@@ -567,6 +567,7 @@
     <ClInclude Include="ui\UiContext.h" />
     <ClInclude Include="ui\WindowManager.h" />
     <ClInclude Include="util\Math.hpp" />
+    <ClInclude Include="util/Prefetch.h" />
     <ClInclude Include="util\SawyerCoding.h" />
     <ClInclude Include="util\Util.h" />
     <ClInclude Include="Version.h" />
diff --git a/src/openrct2/paint/Paint.cpp b/src/openrct2/paint/Paint.cpp
index 53f82b608f..76ddd1c0ee 100644
--- a/src/openrct2/paint/Paint.cpp
+++ b/src/openrct2/paint/Paint.cpp
@@ -20,6 +20,7 @@
 #include "../paint/Painter.h"
 #include "../profiling/Profiling.h"
 #include "../util/Math.hpp"
+#include "../util/Prefetch.h"
 #include "Boundbox.h"
 #include "Paint.Entity.h"
 #include "tile_element/Paint.TileElement.h"
@@ -391,6 +392,7 @@ template<uint8_t TRotation> static void PaintStructsSortQuadrant(PaintStruct* pa
         auto* ps = child;
         child = child->NextQuadrantEntry;
 
+        PREFETCH(&child->Bounds);
         if (child == nullptr || child->SortFlags & PaintSortFlags::OutsideQuadrant)
         {
             break;
diff --git a/src/openrct2/util/Prefetch.h b/src/openrct2/util/Prefetch.h
new file mode 100644
index 0000000000..8fbeb40556
--- /dev/null
+++ b/src/openrct2/util/Prefetch.h
@@ -0,0 +1,34 @@
+/*****************************************************************************
+ * Copyright (c) 2014-2024 OpenRCT2 developers
+ *
+ * For a complete list of all authors, please refer to contributors.md
+ * Interested in contributing? Visit https://github.com/OpenRCT2/OpenRCT2
+ *
+ * OpenRCT2 is licensed under the GNU General Public License version 3.
+ *****************************************************************************/
+
+#if defined(__amd64__) || defined(_M_AMD64) || defined(__i386__) || defined(_M_IX86)
+// Don't bother checking for CPUID, prefetch is available since Pentium 4
+#    include <xmmintrin.h>
+// This cannot be expressed as `constexpr` function, exclude it from clang-tidy check
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#    define PREFETCH(x) _mm_prefetch(reinterpret_cast<const char*>(x), _MM_HINT_T0)
+
+#elif defined(_MSC_VER) && defined(_M_ARM64)
+// ARM64 prefetch is available since ARMv8.
+// MSVC's help (https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-170)
+// is incorrect and points to a different header (arm64_neon.h), which doesn't contain prefetch.
+// Correct header info taken from
+// https://github.com/microsoft/wdkmetadata/blob/1ac0dd0719f19196334de12cf2a6dec20316d440/generation/WDK/IdlHeaders/km/crt/intrin.h#L2
+// and
+// https://github.com/microsoft/wdkmetadata/blob/1ac0dd0719f19196334de12cf2a6dec20316d440/generation/WDK/IdlHeaders/km/crt/intrin.h#L411
+#    include <intrin.h>
+#    define PREFETCH(x) __prefetch(x)
+
+#elif defined(__GNUC__)
+// Let the compiler handle prefetch instruction
+#    define PREFETCH(x) __builtin_prefetch(x)
+
+#else
+#    define PREFETCH(x)
+#endif