Arm64 Support

AI/CMakeLists.txt

@@ -31,7 +31,8 @@ set(CMAKE_C_FLAGS_RELEASE        "${CMAKE_CXX_FLAGS_RELEASE}")
 set(CMAKE_C_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO}")
 set(CMAKE_C_FLAGS_PROFILE        "${CMAKE_CXX_FLAGS_PROFILE}")
 
-remove_definitions(-DSTREFLOP_SSE)
+# remove_definitions(-DSTREFLOP_SSE)
+remove_definitions(-DSTREFLOP_NEON)
 add_definitions(${PIC_FLAG} -D_REENTRANT -D_GNU_SOURCE=1)
 
 if    (MINGW)

AI/Interfaces/C/CMakeLists.txt

@@ -43,7 +43,7 @@ macro    (configure_native_skirmish_ai mySourceDirRel_var additionalSources_var
 
 		# Create a list of all the AIs own source files
 		get_native_sources_recursive(mySources "${mySourceDir}" "${myDir}")
-
+		enable_language(ASM)
 		# Compile the library
 		add_library(${myTarget} MODULE ${mySources} ${additionalSources} ${myVersionDepFile})
 		set_target_properties(${myTarget} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/data)

CMakeLists.txt

@@ -418,7 +418,8 @@ option(ENABLE_STREFLOP "build Spring with streflop support (REQUIRED FOR MULTIPL
 
 if (ENABLE_STREFLOP)
 ## note: is removed in ./AI/CMakeLists.txt since it would break AI compilation
-add_definitions(-DSTREFLOP_SSE)
+# add_definitions(-DSTREFLOP_SSE)
+add_definitions(-DSTREFLOP_NEON)
 else ()
 add_definitions(-DNOT_USING_STREFLOP)
 endif ()
@@ -462,10 +463,11 @@ if (CMAKE_COMPILER_IS_GNUCXX)
 			set(FALLBACK_MARCH "i686")
 		endif (MARCH_BITS EQUAL 64)
 
-		set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=${FALLBACK_MARCH} -mtune=generic ${FALLBACK_SSE_FLAGS}")
+		# set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=${FALLBACK_MARCH} -mtune=generic ${FALLBACK_SSE_FLAGS}")
+		set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv8-a+simd -ffp-contract=off")
 	endif (NOT MARCH_FLAG STREQUAL "generic")
 
-	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SSE_FLAGS} -mfpmath=sse") #SSE_FLAGS are now user input only
+	# set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SSE_FLAGS} -mfpmath=sse") #SSE_FLAGS are now user input only
 elseif (MSVC)
 	set(MSVC_CXX_FLAGS "/arch:SSE2") #default, but still set explicitly
 	# silence warnings

pr.patch

@@ -0,0 +1,36 @@
+diff --git a/src/Version.cpp b/src/Version.cpp
+index 300fa83..2ba270d 100644
+--- a/src/Version.cpp
++++ b/src/Version.cpp
+@@ -32,6 +32,8 @@ const char* platformToString(Platform platform)
+ 			return "linux64";
+ 		case Platform::Windows_x64:
+ 			return "windows64";
++		case Platform::Linux_arm64:
++			return "linuxArm64";
+ 		default:
+ 			// unreachable
+ 			std::abort();
+diff --git a/src/Version.h b/src/Version.h
+index 4c114df..c13d844 100644
+--- a/src/Version.h
++++ b/src/Version.h
+@@ -8,12 +8,17 @@ const char* getAgent();
+ enum class Platform {
+ 	Windows_x64,
+ 	Linux_x64,
++	Linux_arm64
+ };
+
+ #if defined(_WIN64)
+ constexpr Platform PRD_CURRENT_PLATFORM = Platform::Windows_x64;
+-#elif defined(__linux__) && defined(__x86_64__)
++#elif defined(__linux__)
++#if defined(__x86_64__)
+ constexpr Platform PRD_CURRENT_PLATFORM = Platform::Linux_x64;
++#else
++constexpr Platform PRD_CURRENT_PLATFORM = Platform::Linux_arm64;
++#endif
+ #endif
+
+ const char* platformToString(Platform Platform);

rts/Sim/Path/QTPFS/NodeLayer.cpp

@@ -4,20 +4,21 @@
 
 #include <limits>
 
-#if defined(_MSC_VER)
-#include <intrin.h>
-// visual c
-inline int __bsfd (int mask)
-{
-	unsigned long index;
-	_BitScanForward(&index, mask);
-	return index;
-}
-#elif defined(__GNUC__)
-#include <x86intrin.h>
-#else
-#error no bsfd intrinsic currently set
-#endif
+// #if defined(_MSC_VER)
+// #include <intrin.h>
+// // visual c
+// inline int __bsfd (int mask)
+// {
+// 	unsigned long index;
+// 	_BitScanForward(&index, mask);
+// 	return index;
+// }
+// #elif defined(__GNUC__)
+// #include <x86intrin.h>
+// #else
+// #error no bsfd intrinsic currently set
+// #endif
+#include "System/sse2neon.h"
 
 #include "NodeLayer.h"
 #include "PathManager.h"

rts/System/FastMath.h

@@ -4,7 +4,8 @@
 #define FASTMATH_H
 
 #ifndef DEDICATED_NOSSE
-#include <xmmintrin.h>
+// #include <xmmintrin.h>
+#include "System/sse2neon.h"
 #endif
 #include <cinttypes>
 

rts/System/Matrix44f.cpp

@@ -10,8 +10,9 @@
 #include <algorithm>
 #include <cstring>
 
-#include <xmmintrin.h>
-#include <emmintrin.h>
+// #include <xmmintrin.h>
+// #include <emmintrin.h>
+#include "System/sse2neon.h"
 
 CR_BIND(CMatrix44f, )
 

rts/System/Platform/Linux/CrashHandler.cpp

@@ -709,20 +709,20 @@ namespace CrashHandler
 
 		unw_cursor_t cursor;
 
-#if (defined(__arm__) || defined(__APPLE__))
+// #if (defined(__arm__) || defined(__APPLE__))
 		// ucontext_t and unw_context_t are not aliases here
 		unw_context_t thisctx;
 		unw_getcontext(&thisctx);
-#else
-		// Effective ucontext_t. If uc not supplied, use unw_getcontext
-		// locally. This is appropriate inside signal handlers.
-		ucontext_t thisctx;
-
-		if (uc == nullptr) {
-			unw_getcontext(&thisctx);
-			uc = &thisctx;
-		}
-#endif
+// #else
+// 		// Effective ucontext_t. If uc not supplied, use unw_getcontext
+// 		// locally. This is appropriate inside signal handlers.
+// 		ucontext_t thisctx;
+//
+// 		if (uc == nullptr) {
+// 			unw_getcontext(&thisctx);
+// 			uc = &thisctx;
+// 		}
+// #endif
 
 
 		char procbuffer[1024];
@@ -742,11 +742,11 @@ namespace CrashHandler
 		}
 		*/
 
-#if (defined(__arm__) || defined(__APPLE__))
+// #if (defined(__arm__) || defined(__APPLE__))
 		const int err = unw_init_local(&cursor, &thisctx);
-#else
-		const int err = unw_init_local(&cursor, uc);
-#endif
+// #else
+// 		const int err = unw_init_local(&cursor, uc);
+// #endif
 
 		if (err != 0) {
 			LOG_L(L_ERROR, "unw_init_local returned %d", err);

rts/System/ScopedFPUSettings.h

@@ -18,7 +18,7 @@ class ScopedDisableFpuExceptions {
 		streflop::fesetenv(&fenv);
 	}
 private:
-	streflop::fpenv_t fenv;
+	streflop::fenv_t fenv;
 };
 
 #else

rts/System/Sync/FPUCheck.cpp

@@ -19,6 +19,7 @@ void good_fpu_init() { LOG_L(L_WARNING, "[%s] streflop is disabled", __func__);
 
 #ifdef STREFLOP_SSE
 #elif STREFLOP_X87
+#elif STREFLOP_NEON
 #else
 	#error "streflop FP-math mode must be either SSE or X87"
 #endif
@@ -92,7 +93,7 @@ void good_fpu_control_registers(const char* text)
 	constexpr int x87_c = 0x0032; // signan
 
 #ifdef STREFLOP_H
-	streflop::fpenv_t fenv;
+	streflop::fenv_t fenv;
 	streflop::fegetenv(&fenv);
 
 	#if defined(STREFLOP_SSE)
@@ -139,6 +140,8 @@ void good_fpu_init()
 	LOG("[%s][STREFLOP_SSE]", __func__);
 	#elif (defined(STREFLOP_X87))
 	LOG("[%s][STREFLOP_X87]", __func__);
+	#elif (defined(STREFLOP_NEON))
+	LOG("[%s][STREFLOP_NEON]", __func__);
 	#else
 	#error
 	#endif
@@ -157,6 +160,8 @@ void good_fpu_init()
 	LOG_L(L_WARNING, "\tStreflop floating-point math is set to X87 mode");
 	LOG_L(L_WARNING, "\tThis may cause desyncs during multi-player games");
 	LOG_L(L_WARNING, "\tYour CPU is %s SSE-capable; consider %s", (sseFlag == 0)? "not": "", (sseFlag == 1)? "recompiling": "upgrading");
+	#elif (defined(STREFLOP_NEON))
+	LOG_L(L_WARNING, "\tStreflop NEON mode, glhf");
 	#else
 	#error
 	#endif

rts/System/Threading/SpringThreading.h

@@ -9,7 +9,8 @@
 #include <atomic>
 #include <thread>
 #include <condition_variable>
-#include <immintrin.h>
+// #include <immintrin.h>
+#include "System/sse2neon.h"
 
 
 #if   defined(_WIN32)