Merge pull request #701 from reicast/feat/rec-cpp

Initial implementation of a "dynarec-structured" cached shop interpreter

core/core.mk

@@ -37,6 +37,10 @@ ifdef X64_REC
     RZDCY_MODULES += rec-x64/
 endif
 
+ifdef CPP_REC
+    RZDCY_MODULES += rec-cpp/
+endif
+
 ifndef NO_REND
     RZDCY_MODULES += rend/gles/
 else

core/hw/pvr/spg.cpp

@@ -86,6 +86,8 @@ extern double speed_load_mspdf;
 double speed_load_mspdf;
 #endif
 
+int mips_counter;
+
 double full_rps;
 
 u32 fskip=0;
@@ -184,12 +186,13 @@ int spg_line_sched(int tag, int cycl, int jit)
 					spd_cpu*100/200,spd_vbs,
 					mode,res,fullvbs,
 					spd_fps,fskip/ts);
-				#else					
+				#else
-				sprintf(fpsStr,"%s/%c - %4.2f (%4.2f) - %4.2f - V: %4.2f (%.2f, %s%s%4.2f) R: %4.2f+%4.2f VTX: %4.2f%c", 
+				sprintf(fpsStr,"%s/%c - %4.2f (%4.2f) - %4.2f - V: %4.2f (%.2f, %s%s%4.2f) R: %4.2f+%4.2f VTX: %4.2f%c, MIPS: %.2f", 
 					VER_SHORTNAME,'n',mspdf,speed_load_mspdf,spd_cpu*100/200,spd_vbs,
 					spd_vbs/full_rps,mode,res,fullvbs,
 					spd_fps,fskip/ts
-					,mv,mv_c);
+					, mv, mv_c, mips_counter/ 1024.0 / 1024.0);
+					mips_counter = 0;
 				#endif
 				
 				fskip=0;

core/hw/sh4/dyna/decoder.cpp

@@ -15,6 +15,9 @@
 #include "hw/sh4/sh4_mem.h"
 #include "decoder_opcodes.h"
 
+#define BLOCK_MAX_SH_OPS_SOFT 500
+#define BLOCK_MAX_SH_OPS_HARD 511
+
 RuntimeBlockInfo* blk;
 
 
@@ -1054,7 +1057,10 @@ void dec_DecodeBlock(RuntimeBlockInfo* rbi,u32 max_cycles)
 			//there is no break here by design
 		case NDO_NextOp:
 			{
-				if (blk->guest_cycles>=max_cycles && !state.cpu.is_delayslot)
+				if ( 
+					( (blk->oplist.size() >= BLOCK_MAX_SH_OPS_SOFT) || (blk->guest_cycles >= max_cycles) )
+					&& !state.cpu.is_delayslot
+					)
 				{
 					dec_End(state.cpu.rpc,BET_StaticJump,false);
 				}
@@ -1142,6 +1148,8 @@ _end:
 	blk->NextBlock=state.NextAddr;
 	blk->BranchBlock=state.JumpAddr;
 	blk->BlockType=state.BlockType;
+
+	verify(blk->oplist.size() <= BLOCK_MAX_SH_OPS_HARD);
 	
 #if HOST_OS == OS_WINDOWS
 	switch(rbi->addr)

core/hw/sh4/dyna/driver.cpp

@@ -28,12 +28,13 @@
 #if FEAT_SHREC != DYNAREC_NONE
 //uh uh
 
-#if HOST_OS == OS_WINDOWS
+u8 SH4_TCB[2*CODE_SIZE+4096]
-	u8 SH4_TCB[2*CODE_SIZE+4096];
+#if HOST_OS == OS_WINDOWS || FEAT_SHREC != DYNAREC_JIT
+	;
 #elif HOST_OS == OS_LINUX
-	u8 SH4_TCB[2*CODE_SIZE+4096] __attribute__((section(".text")));
+	__attribute__((section(".text")));
 #elif HOST_OS==OS_DARWIN
-	u8 SH4_TCB[2*CODE_SIZE+4096] __attribute__((section("__TEXT,.text")));
+	__attribute__((section("__TEXT,.text")));
 #else
 	#error SH4_TCB ALLOC
 #endif

core/hw/sh4/dyna/shil_canonical.h

@@ -42,7 +42,7 @@ extern "C" f32 fipr_asm(float* fn, float* fm);
 	#define shil_opc(name) struct shil_opcl_##name { 
 	#define shil_opc_end() };
 
-	#define shil_canonical(rv,name,args,code) static rv name args { code }
+	#define shil_canonical(rv,name,args,code) struct name { static rv impl args { code } };
 	
 	#define shil_cf_arg_u32(x) ngen_CC_Param(op,&op->x,CPT_u32);
 	#define shil_cf_arg_f32(x) ngen_CC_Param(op,&op->x,CPT_f32);
@@ -50,7 +50,8 @@ extern "C" f32 fipr_asm(float* fn, float* fm);
 	#define shil_cf_rv_u32(x) ngen_CC_Param(op,&op->x,CPT_u32rv);
 	#define shil_cf_rv_f32(x) ngen_CC_Param(op,&op->x,CPT_f32rv);
 	#define shil_cf_rv_u64(x) ngen_CC_Param(op,&op->rd,CPT_u64rvL); ngen_CC_Param(op,&op->rd2,CPT_u64rvH);
-	#define shil_cf(x) ngen_CC_Call(op,(void*)x);
+	#define shil_cf_ext(x) ngen_CC_Call(op,(void*)&x);
+	#define shil_cf(x) shil_cf_ext(x::impl)
 
 	#define shil_compile(code) static void compile(shil_opcode* op) { ngen_CC_Start(op); code ngen_CC_Finish(op); }
 #elif  SHIL_MODE==2
@@ -61,7 +62,7 @@ extern "C" f32 fipr_asm(float* fn, float* fm);
 	#define shil_opc(name) struct shil_opcl_##name { 
 	#define shil_opc_end() };
 
-	#define shil_canonical(rv,name,args,code) static rv cimpl_##name args;
+	#define shil_canonical(rv,name,args,code) struct name { static rv impl args; };
 	#define shil_compile(code) static void compile(shil_opcode* op);
 #elif  SHIL_MODE==3
 	//generate struct list ...
@@ -93,7 +94,7 @@ extern "C" f32 fipr_asm(float* fn, float* fm);
 
 
 
-#if SHIL_MODE==1
+#if SHIL_MODE==1 || SHIL_MODE==2
 //only in structs we use the code :)
 #include <math.h>
 #include "types.h"
@@ -206,18 +207,26 @@ shil_opc_end()
 
 //Canonical impl. opcodes !
 shil_opc(sync_sr)
+shil_canonical
+(
+void, f1, (),
+	UpdateSR();
+)
 shil_compile
 (
-	shil_cf(UpdateSR);
+	shil_cf(f1);
-	//die();
 )
 shil_opc_end()
 
 shil_opc(sync_fpscr)
+shil_canonical
+(
+void, f1, (),
+	UpdateFPSCR();
+)
 shil_compile
 (
-	shil_cf(UpdateFPSCR);
+	shil_cf(f1);
-	//die();
 )
 shil_opc_end()
 
@@ -981,7 +990,7 @@ shil_opc_end()
 
 
 
-//shop_ftrv
+//shop_frswap
 shil_opc(frswap)
 shil_canonical
 (

core/hw/sh4/sh4_core_regs.cpp

@@ -13,23 +13,23 @@ u8* sh4_dyna_rcb;
 
 INLINE void ChangeGPR()
 {
-	u32 temp[8];
+	u32 temp;
 	for (int i=0;i<8;i++)
 	{
-		temp[i]=r[i];
+		temp=r[i];
 		r[i]=r_bank[i];
-		r_bank[i]=temp[i];
+		r_bank[i]=temp;
 	}
 }
 
 INLINE void ChangeFP()
 {
-	u32 temp[16];
+	u32 temp;
 	for (int i=0;i<16;i++)
 	{
-		temp[i]=fr_hex[i];
+		temp=fr_hex[i];
 		fr_hex[i]=xf_hex[i];
-		xf_hex[i]=temp[i];
+		xf_hex[i]=temp;
 	}
 }
 

core/rec-x86/rec_x86_driver.cpp

@@ -265,6 +265,7 @@ u32* GetRegPtr(u32 reg)
 u32 cvld;
 u32 rdmt[6];
 extern u32 memops_t,memops_l;
+extern int mips_counter;
 
 void CheckBlock(RuntimeBlockInfo* block,x86_ptr_imm place)
 {
@@ -286,6 +287,8 @@ void CheckBlock(RuntimeBlockInfo* block,x86_ptr_imm place)
 	}
 	
 }
+
+
 void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool staging,bool optimise)
 {
 	//initialise stuff
@@ -305,7 +308,11 @@ void ngen_Compile(RuntimeBlockInfo* block,bool force_checks, bool reset, bool st
 
 	x86e->Emit(op_add32,&memops_t,block->memops);
 	x86e->Emit(op_add32,&memops_l,block->linkedmemops);
-	
+
+#ifdef MIPS_COUNTER
+	x86e->Emit(op_add32, &mips_counter, block->oplist.size());
+#endif
+
 	//run register allocator
 	reg.DoAlloc(block,alloc_regs,xmm_alloc_regs);
 	

core/rend/gles/gles.cpp

@@ -1034,7 +1034,7 @@ void tryfit(float* x,float* y)
 	for (int i=0;i<128;i++)
 	{
 		float diff=min(max(b*logf(x[i])/logf(2.0)+a,(double)0),(double)1)-y[i];
-		maxdev=max((float)abs(diff),(float)maxdev);
+		maxdev=max((float)fabs((float)diff),(float)maxdev);
 	}
 	printf("FOG TABLE Curve match: maxdev: %.02f cents\n",maxdev*100);
 	fog_coefs[0]=a;

shell/linCPP/Makefile

@@ -0,0 +1,122 @@
+
+LOCAL_PATH := $(call my-dir)
+FOR_LINUX :=1
+NOT_ARM := 1
+CPP_REC := 1
+#NO_REC := 1
+#NO_REND := 1
+WEBUI :=1 
+USE_ALSA := 1
+USE_OSS := 1
+#USE_PULSEAUDIO := 1
+
+RZDCY_SRC_DIR = ../../core
+
+include $(RZDCY_SRC_DIR)/core.mk
+
+
+CXX=${CC_PREFIX}g++
+CC=${CC_PREFIX}gcc
+AS=${CC_PREFIX}as
+STRIP=${CC_PREFIX}strip
+
+LD=${CC}
+
+MFLAGS := #-m32
+#-marm -march=armv7-a -mtune=cortex-a9 -mfpu=neon -mfloat-abi=softfp -funroll-loops
+ASFLAGS := 
+#-march=armv7-a -mfpu=neon -mfloat-abi=softfp
+
+LDFLAGS	:= 	-g -Wl,-Map,$(notdir $@).map,--gc-sections -Wl,-O3 -Wl,--sort-common 
+
+CFLAGS :=  -g -O3  -D RELEASE -c -D TARGET_LINUX_x64 -D USES_HOMEDIR -D TARGET_NO_JIT
+CFLAGS += -D SUPPORT_X11
+CFLAGS += -frename-registers -fno-strict-aliasing #-fsingle-precision-constant 
+CFLAGS += -ffast-math -ftree-vectorize 
+
+
+#-fprefetch-loop-arrays 
+#-std=c++0x
+CXXFLAGS += $(CFLAGS) $(MFLAGS) -fexceptions -fno-rtti -fpermissive -std=gnu++11
+CXXFLAGS += -D SUPPORT_X11 
+CXXFLAGS += -fno-operator-names
+
+ifdef PGO_MAKE
+	CFLAGS += -fprofile-generate -pg
+	LDFLAGS	+= -fprofile-generate
+else
+	CFLAGS += -fomit-frame-pointer
+endif
+
+ifdef PGO_USE
+	CFLAGS += -fprofile-use
+endif
+
+
+ifdef LTO_TEST
+	CFLAGS += -flto -fwhole-program 
+	LDFLAGS +=-flto -fwhole-program 
+endif
+
+INCS	:= -I$(RZDCY_SRC_DIR) -I$(RZDCY_SRC_DIR)/deps -I$(RZDCY_SRC_DIR)/khronos
+
+LIBS    := # use system libs
+LIBS    += -lm -lrt -ldl
+LIBS    += -lpthread -lX11
+
+ifdef USE_ALSA
+	CXXFLAGS += -D USE_ALSA
+	LIBS += -lasound
+endif
+
+ifdef USE_OSS
+	CXXFLAGS += -D USE_OSS
+endif
+
+ifdef USE_PULSEAUDIO
+	CXXFLAGS += -D USE_PULSEAUDIO
+	LIBS += -lpulse-simple
+endif
+
+ifdef USE_GLES
+	CXXFLAGS += -DGLES
+	LIBS	+= -lEGL -lGLESv2
+else
+	LIBS	+= -ldl -lGL #for desktop gl
+endif
+
+
+OBJECTS=$(RZDCY_FILES:.cpp=.build_obj)
+OBJECTS:=$(OBJECTS:.c=.build_obj)
+OBJECTS:=$(OBJECTS:.S=.build_obj)
+OBJECTS:=$(patsubst $(RZDCY_SRC_DIR)/%,obj/%,$(OBJECTS))
+
+
+EXECUTABLE_STRIPPED=nosym-reicast.elf
+EXECUTABLE=reicast.elf
+
+PACKAGE_FILES=$(EXECUTABLE_STRIPPED) default.gcw0.desktop icon-32.png
+
+all: $(CPPFILES) $(EXECUTABLE) $(EXECUTABLE_STRIPPED)
+	
+$(EXECUTABLE): $(OBJECTS)
+	$(CXX) $(MFLAGS) $(EXTRAFLAGS) $(LDFLAGS)  $(OBJECTS) $(LIBS) -o $@
+
+$(EXECUTABLE_STRIPPED): $(EXECUTABLE)
+	cp $< $@ && $(STRIP) $@
+
+obj/%.build_obj : $(RZDCY_SRC_DIR)/%.cpp
+	mkdir -p $(dir $@)
+	$(CXX) $(EXTRAFLAGS) $(INCS) $(CFLAGS) $(CXXFLAGS) $< -o $@
+	
+obj/%.build_obj : $(RZDCY_SRC_DIR)/%.c
+	mkdir -p $(dir $@)
+	$(CC) $(EXTRAFLAGS) $(INCS) $(CFLAGS) $< -o $@
+
+obj/%.build_obj : $(RZDCY_SRC_DIR)/%.S
+	mkdir -p $(dir $@)	
+	$(AS) $(ASFLAGS) $(INCS) $< -o $@
+
+
+clean:
+	rm $(OBJECTS) $(EXECUTABLE) -f

shell/reicast.sln

@@ -1,8 +1,15 @@
 
-Microsoft Visual Studio Solution File, Format Version 11.00
+Microsoft Visual Studio Solution File, Format Version 12.00
-# Visual Studio 2010
+# Visual Studio 2013
+VisualStudioVersion = 12.0.31101.0
+MinimumVisualStudioVersion = 10.0.40219.1
 Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "reicast", "reicast.vcxproj", "{58B14048-EACB-4780-8B1E-9C84C2C30A8E}"
 EndProject
+Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Solution Items", "Solution Items", "{B03DF793-41BA-4F47-A4F2-C06E52FAFB13}"
+	ProjectSection(SolutionItems) = preProject
+		Performance2.psess = Performance2.psess
+	EndProjectSection
+EndProject
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
 		Debug|Win32 = Debug|Win32
@@ -23,4 +30,7 @@ Global
 	GlobalSection(SolutionProperties) = preSolution
 		HideSolutionNode = FALSE
 	EndGlobalSection
+	GlobalSection(Performance) = preSolution
+		HasPerformanceSessions = true
+	EndGlobalSection
 EndGlobal

shell/reicast.vcxproj

@@ -162,6 +162,10 @@
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Fast|Win32'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Fast|x64'">true</ExcludedFromBuild>
     </ClCompile>
+    <ClCompile Include="..\core\rec-cpp\rec_cpp.cpp">
+      <AdditionalOptions Condition="'$(Configuration)|$(Platform)'=='Fast|x64'">/bigobj %(AdditionalOptions)</AdditionalOptions>
+      <AdditionalOptions Condition="'$(Configuration)|$(Platform)'=='Slow|x64'">/bigobj %(AdditionalOptions)</AdditionalOptions>
+    </ClCompile>
     <ClCompile Include="..\core\rec-x64\rec_x64.cpp" />
     <ClCompile Include="..\core\rec-x86\rec_x86_driver.cpp">
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Fast|x64'">true</ExcludedFromBuild>

shell/reicast.vcxproj.filters

@@ -414,6 +414,9 @@
     <ClCompile Include="..\core\rec-x64\rec_x64.cpp">
       <Filter>rec-x64</Filter>
     </ClCompile>
+    <ClCompile Include="..\core\rec-cpp\rec_cpp.cpp">
+      <Filter>rec-cpp</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="hw">
@@ -542,6 +545,9 @@
     <Filter Include="rec-x64">
       <UniqueIdentifier>{f73263e9-dbe8-4a6f-8b73-335af8307551}</UniqueIdentifier>
     </Filter>
+    <Filter Include="rec-cpp">
+      <UniqueIdentifier>{63d1fcf2-64b4-4973-995f-cd471f51117c}</UniqueIdentifier>
+    </Filter>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\core\hw\aica\aica.h">