Merge pull request #13353 from unknownbrackets/texcache

TexCache: Improve QuickTexHash a little, cleanup ARM64 defines

CMakeLists.txt

@@ -45,7 +45,6 @@ if(CMAKE_SYSTEM_PROCESSOR)
 		endif()
 	elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "^aarch64")
 		set(ARM64 ON)
-		add_definitions(-DARM64)
 	elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "^amd64" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "^x86_64" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "^AMD64")
 		set(X86_DEVICE ON)
 		set(X86_64_DEVICE ON)

Core/Config.cpp

@@ -22,6 +22,7 @@
 #include <set>
 #include <sstream>
 
+#include "ppsspp_config.h"
 #include "base/display.h"
 #include "base/NativeApp.h"
 #include "file/ini_file.h"
@@ -393,15 +394,15 @@ static int DefaultNumWorkers() {
 }
 
 static int DefaultCpuCore() {
-#if defined(ARM) || defined(ARM64) || defined(_M_IX86) || defined(_M_X64)
+#if PPSSPP_ARCH(ARM) || PPSSPP_ARCH(ARM64) || PPSSPP_ARCH(X86) || PPSSPP_ARCH(AMD64)
 	return (int)CPUCore::JIT;
 #else
 	return (int)CPUCore::INTERPRETER;
 #endif
 }
 
 static bool DefaultCodeGen() {
-#if defined(ARM) || defined(ARM64) || defined(_M_IX86) || defined(_M_X64)
+#if PPSSPP_ARCH(ARM) || PPSSPP_ARCH(ARM64) || PPSSPP_ARCH(X86) || PPSSPP_ARCH(AMD64)
 	return true;
 #else
 	return false;

Core/MIPS/ARM64/Arm64Jit.cpp

@@ -472,7 +472,7 @@ void Arm64Jit::UnlinkBlock(u8 *checkedEntry, u32 originalAddress) {
 }
 
 bool Arm64Jit::ReplaceJalTo(u32 dest) {
-#ifdef ARM64
+#if PPSSPP_ARCH(ARM64)
 	const ReplacementTableEntry *entry = nullptr;
 	u32 funcSize = 0;
 	if (!CanReplaceJalTo(dest, &entry, &funcSize)) {

Core/MIPS/JitCommon/JitBlockCache.cpp

@@ -601,12 +601,12 @@ void JitBlockCache::InvalidateChangedBlocks() {
 }
 
 int JitBlockCache::GetBlockExitSize() {
-#if defined(ARM)
+#if PPSSPP_ARCH(ARM)
 	// Will depend on the sequence found to encode the destination address.
 	return 0;
-#elif defined(_M_IX86) || defined(_M_X64)
+#elif PPSSPP_ARCH(X86) || PPSSPP_ARCH(AMD64)
 	return 15;
-#elif defined(ARM64)
+#elif PPSSPP_ARCH(ARM64)
 	// Will depend on the sequence found to encode the destination address.
 	return 0;
 #else
@@ -654,11 +654,11 @@ JitBlockDebugInfo JitBlockCache::GetBlockDebugInfo(int blockNum) const {
 		debugInfo.origDisasm.push_back(mipsDis);
 	}
 
-#if defined(ARM)
+#if PPSSPP_ARCH(ARM)
 	debugInfo.targetDisasm = DisassembleArm2(block->normalEntry, block->codeSize);
-#elif defined(ARM64)
+#elif PPSSPP_ARCH(ARM64)
 	debugInfo.targetDisasm = DisassembleArm64(block->normalEntry, block->codeSize);
-#elif defined(_M_IX86) || defined(_M_X64)
+#elif PPSSPP_ARCH(X86) || PPSSPP_ARCH(AMD64)
 	debugInfo.targetDisasm = DisassembleX86(block->normalEntry, block->codeSize);
 #endif
 

Core/MIPS/JitCommon/JitBlockCache.h

@@ -22,12 +22,13 @@
 #include <vector>
 #include <string>
 
+#include "ppsspp_config.h"
 #include "Common/CommonTypes.h"
 #include "Common/CodeBlock.h"
 #include "Core/MIPS/MIPSAnalyst.h"
 #include "Core/MIPS/MIPS.h"
 
-#if defined(ARM) || defined(ARM64)
+#if PPSSPP_ARCH(ARM) || PPSSPP_ARCH(ARM64)
 const int MAX_JIT_BLOCK_EXITS = 2;
 #else
 const int MAX_JIT_BLOCK_EXITS = 8;

GPU/Common/TextureDecoder.h

@@ -23,6 +23,7 @@ enum CheckAlphaResult {
 	CHECKALPHA_ANY = 4,
 };
 
+#include "ppsspp_config.h"
 #include "Common/Common.h"
 #include "Common/Swap.h"
 #include "Core/MemMap.h"
@@ -46,7 +47,7 @@ void DoUnswizzleTex16Basic(const u8 *texptr, u32 *ydestp, int bxc, int byc, u32
 #define DoUnswizzleTex16 DoUnswizzleTex16Basic
 
 // For ARM64, NEON is mandatory, so we also statically link.
-#elif PPSSPP_ARCH(ARM64) || defined(ARM64)
+#elif PPSSPP_ARCH(ARM64)
 #define DoQuickTexHash QuickTexHashNEON
 #define StableQuickTexHash QuickTexHashNEON
 #define DoUnswizzleTex16 DoUnswizzleTex16NEON

GPU/Common/TextureDecoderNEON.cpp

@@ -41,13 +41,14 @@ u32 QuickTexHashNEON(const void *checkp, u32 size) {
 	__builtin_prefetch(checkp, 0, 0);
 
 	if (((intptr_t)checkp & 0xf) == 0 && (size & 0x3f) == 0) {
-#if defined(IOS) || PPSSPP_ARCH(ARM64) || defined(_MSC_VER)
+#if defined(IOS) || PPSSPP_ARCH(ARM64) || defined(_MSC_VER) || !PPSSPP_ARCH(ARMV7)
 		uint32x4_t cursor = vdupq_n_u32(0);
 		uint16x8_t cursor2 = vld1q_u16(QuickTexHashInitial);
 		uint16x8_t update = vdupq_n_u16(0x2455U);
 
 		const u32 *p = (const u32 *)checkp;
-		for (u32 i = 0; i < size / 16; i += 4) {
+		const u32 *pend = p + size / 4;
+		while (p < pend) {
 			cursor = vreinterpretq_u32_u16(vmlaq_u16(vreinterpretq_u16_u32(cursor), vreinterpretq_u16_u32(vld1q_u32(&p[4 * 0])), cursor2));
 			cursor = veorq_u32(cursor, vld1q_u32(&p[4 * 1]));
 			cursor = vaddq_u32(cursor, vld1q_u32(&p[4 * 2]));
@@ -58,10 +59,12 @@ u32 QuickTexHashNEON(const void *checkp, u32 size) {
 		}
 
 		cursor = vaddq_u32(cursor, vreinterpretq_u32_u16(cursor2));
-		check = vgetq_lane_u32(cursor, 0) + vgetq_lane_u32(cursor, 1) + vgetq_lane_u32(cursor, 2) + vgetq_lane_u32(cursor, 3);
+		uint32x2_t mixed = vadd_u32(vget_high_u32(cursor), vget_low_u32(cursor));
+		check = vget_lane_u32(mixed, 0) + vget_lane_u32(mixed, 1);
 #else
 		// TODO: Why does this crash on iOS, but only certain devices?
 		// It's faster than the above, but I guess it sucks to be using an iPhone.
+		// As of 2020 clang, it's still faster by ~1.4%.
 
 		// d0/d1 (q0) - cursor
 		// d2/d3 (q1) - cursor2

GPU/Common/VertexDecoderArm64.cpp

@@ -18,11 +18,6 @@
 #include "ppsspp_config.h"
 #if PPSSPP_ARCH(ARM64)
 
-// This allows highlighting to work.  Yay.
-#ifdef __INTELLISENSE__
-#define ARM64
-#endif
-
 #include "Common/CPUDetect.h"
 #include "Common/Log.h"
 #include "Core/Config.h"

GPU/Common/VertexDecoderCommon.cpp

@@ -18,6 +18,7 @@
 #include <algorithm>
 #include <cstdio>
 
+#include "ppsspp_config.h"
 #include "base/basictypes.h"
 
 #include "Common/Log.h"
@@ -1345,11 +1346,11 @@ std::string VertexDecoder::GetString(DebugShaderStringType stringType) {
 			if (!jitted_)
 				return "Not compiled";
 			std::vector<std::string> lines;
-#if defined(ARM64)
+#if PPSSPP_ARCH(ARM64)
 			lines = DisassembleArm64((const u8 *)jitted_, jittedSize_);
-#elif defined(ARM)
+#elif PPSSPP_ARCH(ARM)
 			lines = DisassembleArm2((const u8 *)jitted_, jittedSize_);
-#elif defined(MIPS)
+#elif PPSSPP_ARCH(MIPS) || PPSSPP_ARCH(MIPS64)
 			// No MIPS disassembler defined
 #else
 			lines = DisassembleX86((const u8 *)jitted_, jittedSize_);

UI/DevScreens.cpp

@@ -17,6 +17,7 @@
 
 #include <algorithm>
 
+#include "ppsspp_config.h"
 #include "base/display.h"
 #include "base/stringutil.h"
 #include "gfx_es2/gpu_features.h"
@@ -458,7 +459,7 @@ void SystemInfoScreen::CreateViews() {
 
 	deviceSpecs->Add(new ItemHeader(si->T("CPU Information")));
 	deviceSpecs->Add(new InfoItem(si->T("CPU Name", "Name"), cpu_info.brand_string));
-#if defined(ARM) || defined(ARM64) || defined(MIPS)
+#if PPSSPP_ARCH(ARM) || PPSSPP_ARCH(ARM64) || PPSSPP_ARCH(MIPS) || PPSSPP_ARCH(MIPS64)
 	deviceSpecs->Add(new InfoItem(si->T("Cores"), StringFromInt(cpu_info.num_cores)));
 #else
 	int totalThreads = cpu_info.num_cores * cpu_info.logical_cpu_count;

UI/NativeApp.cpp

@@ -125,9 +125,9 @@ static UI::Theme ui_theme;
 
 Atlas g_ui_atlas;
 
-#if defined(ARM) && defined(__ANDROID__)
+#if PPSSPP_ARCH(ARM) && defined(__ANDROID__)
 #include "../../android/jni/ArmEmitterTest.h"
-#elif defined(ARM64) && defined(__ANDROID__)
+#elif PPSSPP_ARCH(ARM64) && defined(__ANDROID__)
 #include "../../android/jni/Arm64EmitterTest.h"
 #endif
 
@@ -291,9 +291,9 @@ void NativeGetAppInfo(std::string *app_dir_name, std::string *app_nice_name, boo
 	*landscape = true;
 	*version = PPSSPP_GIT_VERSION;
 
-#if defined(ARM) && defined(__ANDROID__)
+#if PPSSPP_ARCH(ARM) && defined(__ANDROID__)
 	ArmEmitterTest();
-#elif defined(ARM64) && defined(__ANDROID__)
+#elif PPSSPP_ARCH(ARM64) && defined(__ANDROID__)
 	Arm64EmitterTest();
 #endif
 }

Windows/GPU/D3D11Context.cpp

@@ -54,7 +54,7 @@ HRESULT D3D11Context::CreateTheDevice(IDXGIAdapter *adapter) {
 	// D3D11 has no need for display rotation.
 	g_display_rotation = DisplayRotation::ROTATE_0;
 	g_display_rot_matrix.setIdentity();
-#if defined(_DEBUG) && !defined(_M_ARM) && !defined(_M_ARM64)
+#if defined(_DEBUG) && !PPSSPP_ARCH(ARM) && !PPSSPP_ARCH(ARM64)
 	UINT createDeviceFlags = D3D11_CREATE_DEVICE_DEBUG;
 #else
 	UINT createDeviceFlags = 0;

android/jni/Locals.mk

@@ -70,7 +70,7 @@ ifeq ($(TARGET_ARCH_ABI),x86_64)
   LOCAL_LDLIBS += $(LOCAL_PATH)/../../ffmpeg/android/x86_64/lib/libavutil.a
   LOCAL_C_INCLUDES += $(LOCAL_PATH)/../../ffmpeg/android/x86_64/include
 
-  LOCAL_CFLAGS := $(LOCAL_CFLAGS) -D_ARCH_64 -D_M_X64 -fomit-frame-pointer -mtune=atom -mfpmath=sse -mssse3 -mstackrealign
+  LOCAL_CFLAGS := $(LOCAL_CFLAGS) -D_M_X64 -fomit-frame-pointer -mtune=atom -mfpmath=sse -mssse3 -mstackrealign
 endif
 
 ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
@@ -80,8 +80,6 @@ ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
   LOCAL_LDLIBS += $(LOCAL_PATH)/../../ffmpeg/android/arm64/lib/libswscale.a
   LOCAL_LDLIBS += $(LOCAL_PATH)/../../ffmpeg/android/arm64/lib/libavutil.a
   LOCAL_C_INCLUDES += $(LOCAL_PATH)/../../ffmpeg/android/arm64/include
-
-  LOCAL_CFLAGS := $(LOCAL_CFLAGS) -D_ARCH_64 -DARM64
 endif
 
 # Compile with profiling.

cmake/Toolchains/beagleboard.cmake

@@ -1,4 +1,3 @@
-set(ARMV7 ON)
 set(USING_FBDEV ON)
 
 set(ARCH_FLAGS "-march=armv7-a -mfpu=neon -mcpu=cortex-a8")

cmake/Toolchains/generic.armv7.cmake

@@ -1,5 +1,3 @@
-set(ARMV7 ON)
-
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=armv7-a -mfpu=neon -mcpu=cortex-a9")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv7-a -mfpu=neon -mcpu=cortex-a9")
 set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} -march=armv7-a -mfpu=neon -mcpu=cortex-a9")

cmake/Toolchains/raspberry.armv7.cmake

@@ -38,5 +38,4 @@ if(NOT PPSSPP_PI_MODEL4)
   set(USING_FBDEV ON)
 endif()
 set(USING_GLES2 ON)
-set(ARMV7 ON)
 set(USE_WAYLAND_WSI OFF)

cmake/Toolchains/raspberry.armv8.cmake

@@ -38,5 +38,4 @@ if(NOT PPSSPP_PI_MODEL4)
   set(USING_FBDEV ON)
 endif()
 set(USING_GLES2 ON)
-set(ARMV7 ON)
 set(USE_WAYLAND_WSI OFF)

cmake/Toolchains/vero4k.armv8.cmake

@@ -17,6 +17,5 @@ set(EGL_LIBRARIES /opt/vero3/lib/libEGL.so)
 set(USING_GLES2 ON)
 set(USING_EGL ON)
 set(USING_FBDEV ON)
-set(ARMV7 ON)
 set(FORCED_CPU armv7)
 set(USING_X11_VULKAN OFF CACHE BOOL "" FORCE)

ext/glslang-build/Android.mk

@@ -64,8 +64,6 @@ ifeq ($(findstring armeabi-v7a,$(TARGET_ARCH_ABI)),armeabi-v7a)
 LOCAL_CFLAGS := $(LOCAL_CFLAGS) -DARM -DARMEABI_V7A
 else ifeq ($(TARGET_ARCH_ABI),armeabi)
 LOCAL_CFLAGS := $(LOCAL_CFLAGS) -DARM -DARMEABI -march=armv6
-else ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
-LOCAL_CFLAGS := $(LOCAL_CFLAGS) -D_ARCH_64 -DARM64
 else ifeq ($(TARGET_ARCH_ABI),x86)
 LOCAL_CFLAGS := $(LOCAL_CFLAGS) -D_M_IX86
 else ifeq ($(TARGET_ARCH_ABI),x86_64)

ext/miniupnp-build/Android.mk

@@ -31,8 +31,6 @@ ifeq ($(findstring armeabi-v7a,$(TARGET_ARCH_ABI)),armeabi-v7a)
 LOCAL_CFLAGS := $(LOCAL_CFLAGS) -DARM -DARMEABI_V7A
 else ifeq ($(TARGET_ARCH_ABI),armeabi)
 LOCAL_CFLAGS := $(LOCAL_CFLAGS) -DARM -DARMEABI -march=armv6
-else ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
-LOCAL_CFLAGS := $(LOCAL_CFLAGS) -D_ARCH_64 -DARM64
 else ifeq ($(TARGET_ARCH_ABI),x86)
 LOCAL_CFLAGS := $(LOCAL_CFLAGS) -D_M_IX86
 else ifeq ($(TARGET_ARCH_ABI),x86_64)

ext/native/Android.mk

@@ -118,8 +118,6 @@ LOCAL_SRC_FILES := $(LOCAL_SRC_FILES) \
 
 else ifeq ($(TARGET_ARCH_ABI),armeabi)
 LOCAL_CFLAGS := $(LOCAL_CFLAGS) -DARM -DARMEABI -march=armv6
-else ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
-LOCAL_CFLAGS := $(LOCAL_CFLAGS) -D_ARCH_64 -DARM64
 else ifeq ($(TARGET_ARCH_ABI),x86)
 LOCAL_CFLAGS := $(LOCAL_CFLAGS) -D_M_IX86
 LOCAL_SRC_FILES := $(LOCAL_SRC_FILES) \

ext/native/gfx/gl_common.h

@@ -1,5 +1,7 @@
 #pragma once
 
+#include "ppsspp_config.h"
+
 #ifdef IOS
 #include <OpenGLES/ES3/gl.h>
 #include <OpenGLES/ES3/glext.h>
@@ -51,7 +53,7 @@ extern PFNGLMAPBUFFERPROC glMapBuffer;
 
 typedef void (EGLAPIENTRYP PFNGLDRAWTEXTURENVPROC) (GLuint texture, GLuint sampler, GLfloat x0, GLfloat y0, GLfloat x1, GLfloat y1, GLfloat z, GLfloat s0, GLfloat t0, GLfloat s1, GLfloat t1);
 extern PFNGLDRAWTEXTURENVPROC glDrawTextureNV;
-#ifndef ARM64
+#if !PPSSPP_ARCH(ARM64)
 typedef void (EGLAPIENTRYP PFNGLBLITFRAMEBUFFERNVPROC) (
 	GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1,
 	GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1,

ppsspp_config.h

@@ -58,10 +58,6 @@
     #define PPSSPP_ARCH_ARM64 1
     #define PPSSPP_ARCH_64BIT 1
     #define PPSSPP_ARCH_ARM_NEON 1
-    //TODO: Remove this compat define
-    #ifndef ARM64
-        #define ARM64 1
-    #endif
 #endif
 
 #if defined(__mips64__)

unittest/JitHarness.cpp

@@ -17,6 +17,7 @@
 
 #include <algorithm>
 
+#include "ppsspp_config.h"
 #include "base/NativeApp.h"
 #include "Common/TimeUtil.h"
 #include "Core/ConfigValues.h"
@@ -172,9 +173,9 @@ bool TestJit() {
 		// Disassemble
 		JitBlockCache *cache = MIPSComp::jit->GetBlockCache();
 		JitBlock *block = cache->GetBlock(0);  // Should only be one block.
-#if defined(ARM)
+#if PPSSPP_ARCH(ARM)
 		std::vector<std::string> lines = DisassembleArm2(block->normalEntry, block->codeSize);
-#elif defined(ARM64)
+#elif PPSSPP_ARCH(ARM64)
 		std::vector<std::string> lines = DisassembleArm64(block->normalEntry, block->codeSize);
 #else
 		std::vector<std::string> lines = DisassembleX86(block->normalEntry, block->codeSize);

unittest/UnitTest.cpp

@@ -33,6 +33,7 @@
 #include <jni.h>
 #endif
 
+#include "ppsspp_config.h"
 #include "base/NativeApp.h"
 #include "input/input_state.h"
 #include "ext/disarm.h"
@@ -559,13 +560,13 @@ bool TestArm64Emitter();
 bool TestX64Emitter();
 
 TestItem availableTests[] = {
-#if defined(ARM64) || defined(_M_X64) || defined(_M_IX86)
+#if PPSSPP_ARCH(ARM64) || PPSSPP_ARCH(AMD64) || PPSSPP_ARCH(X86)
 	TEST_ITEM(Arm64Emitter),
 #endif
-#if defined(ARM) || defined(_M_X64) || defined(_M_IX86)
+#if PPSSPP_ARCH(ARM) || PPSSPP_ARCH(AMD64) || PPSSPP_ARCH(X86)
 	TEST_ITEM(ArmEmitter),
 #endif
-#if defined(_M_X64) || defined(_M_IX86)
+#if PPSSPP_ARCH(AMD64) || PPSSPP_ARCH(X86)
 	TEST_ITEM(X64Emitter),
 #endif
 	TEST_ITEM(VertexJit),