Reserve and allocate maximum RAM/VRAM/ARAM in all cases

Reserve enough virtual memory space for DC and Naomi
Allocate dynarec entry point tables for max possible ram
Free mem and release vmem on exit

core/build.h

@@ -286,6 +286,10 @@
 	#define FEAT_HAS_SOFTREND BUILD_COMPILER == COMPILER_VC	//GCC wants us to enable sse4 globaly to enable intrins
 #endif
 
+#define RAM_SIZE_MAX (32*1024*1024)
+#define VRAM_SIZE_MAX (16*1024*1024)
+#define ARAM_SIZE_MAX (8*1024*1024)
+
 //Depricated build configs
 #ifdef HOST_NO_REC
 #error Dont use HOST_NO_REC

core/hw/arm7/arm7.cpp

@@ -686,7 +686,7 @@ u8 ARM7_TCB[ICacheSize+4096] __attribute__((section("__TEXT, .text")));
 using namespace ARM;
 
 
-void* EntryPoints[ARAM_SIZE/4];
+void* EntryPoints[ARAM_SIZE_MAX/4];
 
 enum OpType
 {
@@ -1509,7 +1509,7 @@ naked void arm_dispatch()
 #if HOST_OS == OS_LINUX
 	__asm ( "arm_dispatch:				\n\t"
 				"mov %0, %%eax			\n\t"
-			 	"and $0x1FFFFC, %%eax	\n\t"
+			 	"and $0x7FFFFC, %%eax	\n\t"
 				"cmp $0, %1				\n\t"
 			 	"jne arm_dofiq			\n\t"
 			 	"jmp *%2(%%eax)			\n"
@@ -1529,7 +1529,7 @@ naked void arm_dispatch()
 	{
 arm_disp:
 		mov eax,reg[R15_ARM_NEXT*4].I
-		and eax,0x1FFFFC
+		and eax,0x7FFFFC
 		cmp reg[INTR_PEND*4].I,0
 		jne arm_dofiq
 		jmp [EntryPoints+eax]
@@ -2146,8 +2146,8 @@ extern "C" void CompileCode()
 void FlushCache()
 {
 	icPtr=ICache;
-	for (u32 i=0;i<ARAM_SIZE/4;i++)
+	for (u32 i = 0; i < ARRAY_SIZE(EntryPoints); i++)
-		EntryPoints[i]=(void*)&arm_compilecode;
+		EntryPoints[i] = &arm_compilecode;
 }
 
 

core/hw/mem/_vmem.cpp

@@ -396,10 +396,15 @@ void _vmem_term()
 u8* virt_ram_base;
 
 void* malloc_pages(size_t size) {
-
+#ifdef _ISOC11_SOURCE
-	u8* rv = (u8*)malloc(size + PAGE_SIZE);
+	return aligned_alloc(PAGE_SIZE, size);
-
+#else
-	return rv + PAGE_SIZE - ((unat)rv % PAGE_SIZE);
+	void *data;
+	if (posix_memalign(&data, PAGE_SIZE, size) != 0)
+		return NULL;
+	else
+		return data;
+#endif
 }
 
 bool _vmem_reserve_nonvmem()
@@ -437,6 +442,14 @@ void _vmem_bm_reset() {
 	}
 }
 
+static void _vmem_release_nonvmem()
+{
+	free(p_sh4rcb);
+	free(vram.data);
+	free(aica_ram.data);
+	free(mem_b.data);
+}
+
 #if !defined(TARGET_NO_NVMEM)
 
 #define MAP_RAM_START_OFFSET  0
@@ -483,9 +496,9 @@ void* _nvmem_unused_buffer(u32 start,u32 end)
 
 void* _nvmem_alloc_mem()
 {
-	mem_handle=CreateFileMapping(INVALID_HANDLE_VALUE,0,PAGE_READWRITE ,0,RAM_SIZE + VRAM_SIZE +ARAM_SIZE,0);
+	mem_handle = CreateFileMapping(INVALID_HANDLE_VALUE, 0, PAGE_READWRITE, 0, RAM_SIZE_MAX + VRAM_SIZE_MAX + ARAM_SIZE_MAX, 0);
 
-	void* rv=(u8*)VirtualAlloc(0,512*1024*1024 + sizeof(Sh4RCB) + ARAM_SIZE,MEM_RESERVE,PAGE_NOACCESS);
+	void* rv= (u8*)VirtualAlloc(0, 512*1024*1024 + sizeof(Sh4RCB) + ARAM_SIZE_MAX, MEM_RESERVE, PAGE_NOACCESS);
 	if (rv) VirtualFree(rv,0,MEM_RELEASE);
 	return rv;
 }
@@ -585,7 +598,7 @@ error:
 		string path = get_writable_data_path("/dcnzorz_mem");
         fd = open(path.c_str(),O_CREAT|O_RDWR|O_TRUNC,S_IRWXU|S_IRWXG|S_IRWXO);
         unlink(path.c_str());
-        verify(ftruncate(fd,RAM_SIZE + VRAM_SIZE +ARAM_SIZE)==0);
+        verify(ftruncate(fd, RAM_SIZE_MAX + VRAM_SIZE_MAX + ARAM_SIZE_MAX) == 0);
 #elif !defined(_ANDROID)
 		fd = shm_open("/dcnzorz_mem", O_CREAT | O_EXCL | O_RDWR,S_IREAD | S_IWRITE);
 		shm_unlink("/dcnzorz_mem");
@@ -595,10 +608,10 @@ error:
 			unlink("dcnzorz_mem");
 		}
 
-		verify(ftruncate(fd,RAM_SIZE + VRAM_SIZE +ARAM_SIZE)==0);
+		verify(ftruncate(fd, RAM_SIZE_MAX + VRAM_SIZE_MAX + ARAM_SIZE_MAX) == 0);
 #else
 
-		fd = ashmem_create_region(0,RAM_SIZE + VRAM_SIZE +ARAM_SIZE);
+		fd = ashmem_create_region(0, RAM_SIZE_MAX + VRAM_SIZE_MAX + ARAM_SIZE_MAX);
 		if (false)//this causes writebacks to flash -> slow and stuttery 
 		{
 		fd = open("/data/data/com.reicast.emulator/files/dcnzorz_mem",O_CREAT|O_RDWR|O_TRUNC,S_IRWXU|S_IRWXG|S_IRWXO);
@@ -608,7 +621,7 @@ error:
 
 		
 
-		u32 sz= 512*1024*1024 + sizeof(Sh4RCB) + ARAM_SIZE + 0x10000;
+		u32 sz = 512*1024*1024 + sizeof(Sh4RCB) + ARAM_SIZE_MAX + 0x10000;
 		void* rv=mmap(0, sz, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0);
 		verify(rv != NULL);
 		munmap(rv,sz);
@@ -695,6 +708,7 @@ bool _vmem_reserve()
 	
 	p_sh4rcb=(Sh4RCB*)virt_ram_base;
 
+	// Map the sh4 context but protect access to Sh4RCB.fpcb[]
 #if HOST_OS==OS_WINDOWS
 	//verify(p_sh4rcb==VirtualAlloc(p_sh4rcb,sizeof(Sh4RCB),MEM_RESERVE|MEM_COMMIT,PAGE_READWRITE));
 	verify(p_sh4rcb==VirtualAlloc(p_sh4rcb,sizeof(Sh4RCB),MEM_RESERVE,PAGE_NOACCESS));
@@ -771,15 +785,30 @@ bool _vmem_reserve()
 
 	return virt_ram_base!=0;
 }
+
+void _vmem_release()
+{
+	if (!_nvmem_enabled())
+		_vmem_release_nonvmem();
+	else
+	{
+		if (virt_ram_base != NULL)
+		{
+			munmap(virt_ram_base, 0x20000000);
+			virt_ram_base = NULL;
+		}
+		close(fd);
+	}
+}
+
 #else
 
 bool _vmem_reserve()
 {
 	return _vmem_reserve_nonvmem();
 }
-#endif
-
 void _vmem_release()
 {
-	//TODO
+	_vmem_release_nonvmem();
 }
+#endif

core/hw/sh4/sh4_if.h

@@ -299,7 +299,7 @@ void DYNACALL do_sqw_nommu_full(u32 dst, u8* sqb);
 typedef void DYNACALL sqw_fp(u32 dst,u8* sqb);
 typedef void DYNACALL TaListVoidFP(void* data);
 
-#define FPCB_SIZE (RAM_SIZE/2)
+#define FPCB_SIZE (RAM_SIZE_MAX/2)
 #define FPCB_MASK (FPCB_SIZE -1)
 //#defeine FPCB_PAD 0x40000
 #define FPCB_PAD 0x100000

core/rec-ARM/ngen_arm.S

@@ -177,14 +177,14 @@ CSYM(no_update):              @ next_pc _MUST_ be on r4 *R4 NOT R0 anymore*
 	cmp r0,#0
 	beq CSYM(cleanup)
 
-#if DC_PLATFORM == DC_PLATFORM_NAOMI
+#if RAM_SIZE_MAX == 33554432
 	sub r2,r8,#0x4100000
-	ubfx r1,r4,#1,#24
+	ubfx r1,r4,#1,#24			@ 24+1 bits: 32 MB
-#elif DC_PLATFORM == DC_PLATFORM_DREAMCAST
+#elif RAM_SIZE_MAX == 16777216
 	sub r2,r8,#0x2100000
-	ubfx r1,r4,#1,#23
+	ubfx r1,r4,#1,#23			@ 23+1 bits: 16 MB
 #else
-#error "Define DC_PLATFORM"
+#error "Define RAM_SIZE_MAX"
 #endif
 	ldr pc,[r2,r1,lsl #2]
 
@@ -241,7 +241,7 @@ HIDDEN(arm_dispatch)
 CSYM(arm_dispatch):
 	ldrd r0,r1,[r8,#184]		@load: Next PC, interrupt
 
-	ubfx r2,r0,#2,#21
+	ubfx r2,r0,#2,#21		@ assuming 8 MB address space max (23 bits)
 	cmp r1,#0
 	bne arm_dofiq
 

core/rec-ARM/rec_arm.cpp

@@ -402,7 +402,7 @@ u32 DynaRBI::Relink()
 #ifdef CALLSTACK
 #error offset broken
 		SUB(r2, r8, -FPCB_OFFSET);
-#if RAM_SIZE == 33554432
+#if RAM_SIZE_MAX == 33554432
 		UBFX(r1, r4, 1, 24);
 #else
 		UBFX(r1, r4, 1, 23);
@@ -429,7 +429,7 @@ u32 DynaRBI::Relink()
 			//this is faster
 			//why ? (Icache ?)
 			SUB(r2, r8, -FPCB_OFFSET);
-#if RAM_SIZE == 33554432
+#if RAM_SIZE_MAX == 33554432
 			UBFX(r1, r4, 1, 24);
 #else
 			UBFX(r1, r4, 1, 23);
@@ -448,7 +448,7 @@ u32 DynaRBI::Relink()
 			{
 				SUB(r2, r8, -FPCB_OFFSET);
 
-#if RAM_SIZE == 33554432
+#if RAM_SIZE_MAX == 33554432
 				UBFX(r1, r4, 1, 24);
 #else
 				UBFX(r1, r4, 1, 23);
@@ -463,7 +463,7 @@ u32 DynaRBI::Relink()
 			verify(pBranchBlock==0);
 			SUB(r2, r8, -FPCB_OFFSET);
 
-#if RAM_SIZE == 33554432
+#if RAM_SIZE_MAX == 33554432
 			UBFX(r1, r4, 1, 24);
 #else
 			UBFX(r1, r4, 1, 23);

core/rec-ARM64/rec_arm64.cpp

@@ -212,12 +212,12 @@ void ngen_mainloop(void* v_cntx)
 		"movz x2, %[RCB_SIZE], lsl #16	\n\t"
 		"sub x2, x28, x2			\n\t"
 		"add x2, x2, %[SH4CTX_SIZE]	\n\t"
-#if RAM_SIZE == 33554432
+#if RAM_SIZE_MAX == 33554432
-		"ubfx w1, w29, #1, #24		\n\t"
+		"ubfx w1, w29, #1, #24		\n\t"	// 24+1 bits: 32 MB
-#elif RAM_SIZE == 16777216
+#elif RAM_SIZE_MAX == 16777216
-		"ubfx w1, w29, #1, #23		\n\t"
+		"ubfx w1, w29, #1, #23		\n\t"	// 23+1 bits: 16 MB
 #else
-#error "Define RAM_SIZE"
+#error "Define RAM_SIZE_MAX"
 #endif
 		"ldr x0, [x2, x1, lsl #3]	\n\t"
 		"br x0						\n"
@@ -995,7 +995,7 @@ public:
 			Mov(x2, sizeof(Sh4RCB));
 			Sub(x2, x28, x2);
 			Add(x2, x2, sizeof(Sh4Context));		// x2 now points to FPCB
-#if RAM_SIZE == 33554432
+#if RAM_SIZE_MAX == 33554432
 			Ubfx(w1, w29, 1, 24);
 #else
 			Ubfx(w1, w29, 1, 23);

core/rec-x64/rec_x64.cpp

@@ -66,14 +66,14 @@ static __attribute((used)) void end_slice()
 
 #define STRINGIFY(x) #x
 #define _S(x) STRINGIFY(x)
-#if RAM_SIZE == 16*1024*1024
+#if RAM_SIZE_MAX == 16*1024*1024
 #define CPU_RUNNING 68157284
 #define PC 68157256
-#elif RAM_SIZE == 32*1024*1024
+#elif RAM_SIZE_MAX == 32*1024*1024
 #define CPU_RUNNING 135266148
 #define PC 135266120
 #else
-#error RAM_SIZE unknown
+#error RAM_SIZE_MAX unknown
 #endif
 
 #ifdef _WIN32

core/types.h

@@ -1004,3 +1004,8 @@ struct OnLoad
 	typedef void OnLoadFP();
 	OnLoad(OnLoadFP* fp) { fp(); }
 };
+
+#ifndef ARRAY_SIZE
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
+#endif
+