Merge branch 'master' into libretro

Makefile.libretro

@@ -355,7 +355,8 @@ else ifeq ($(platform), ctr)
 	endif
 	TARGET := $(TARGET_NAME)_libretro_ctr.a
 	CFLAGS += -DARM11 -D_3DS -D__3DS__
-	CFLAGS += -DNO_DYLIB -DGPU_UNAI_USE_FLOATMATH -DGPU_UNAI_USE_FLOAT_DIV_MULTINV
+	CFLAGS += -DGPU_UNAI_USE_FLOATMATH -DGPU_UNAI_USE_FLOAT_DIV_MULTINV
+	CFLAGS += -DGPU_UNAI_USE_FLOAT_DIV_MULTINV_FOR_ONE # needed on some compilers?
 	CFLAGS += -march=armv6k -mtune=mpcore -mfloat-abi=hard -marm -mfpu=vfp -mtp=soft
 	CFLAGS += -mword-relocations
 	CFLAGS += -fomit-frame-pointer
@@ -367,6 +368,7 @@ else ifeq ($(platform), ctr)
 	CFLAGS += -DHAVE_UNISTD_H
 	CFLAGS += -DZ7_DECL_Int32_AS_long
 	CFLAGS += -DUSE_CTRULIB_2
+	CFLAGS += -DNO_DYLIB
 
 	OBJS += deps/arm-mem/memcpymove-v6l.o
 	OBJS += deps/arm-mem/memset-v6l.o

frontend/main.c

@@ -513,6 +513,16 @@ int emu_core_init(void)
 	SysPrintf("Starting PCSX-ReARMed " REV "%s\n", get_build_info());
 	SysPrintf("build time: " __DATE__ " " __TIME__ "\n");
 
+#if defined(__arm__) && defined(__VFP_FP__)
+	// RunFast mode
+	u32 fpscr = ~0;
+	__asm__ volatile("fmrx %0, fpscr" : "=r"(fpscr));
+	SysPrintf("old fpscr = %08x\n", fpscr);
+	fpscr &= ~0x00009f9f;
+	fpscr |=  0x03000000; // DN | FZ
+	__asm__ volatile("fmxr fpscr, %0" :: "r"(fpscr));
+#endif
+
 #ifdef HAVE_RTHREADS
 	pcsxr_sthread_init();
 #endif

libpcsxcore/new_dynarec/assem_arm64.c

@@ -2038,6 +2038,7 @@ static void do_jump_vaddr(u_int rs)
 {
   if (rs != 0)
     emit_mov(rs, 0);
+  emit_readptr(&hash_table_ptr, 1);
   emit_far_call(ndrc_get_addr_ht);
   emit_jmpreg(0);
 }

libpcsxcore/new_dynarec/emu_if.c

@@ -356,11 +356,14 @@ static void ari64_execute_threaded_once(struct psxRegisters *regs,
 	enum blockExecCaller block_caller)
 {
 	void *drc_local = (char *)regs - LO_psxRegs;
+	struct ht_entry *hash_table =
+		*(void **)((char *)drc_local + LO_hash_table_ptr);
 	void *target;
 
 	if (likely(!ndrc_g.thread.busy)) {
 		ndrc_g.thread.addr = 0;
-		target = ndrc_get_addr_ht_param(regs->pc, ndrc_cm_no_compile);
+		target = ndrc_get_addr_ht_param(hash_table, regs->pc,
+				ndrc_cm_no_compile);
 		if (target) {
 			clear_local_cache();
 			new_dyna_start_at(drc_local, target);
@@ -428,6 +431,8 @@ static int ari64_thread_check_range(unsigned int start, unsigned int end)
 
 static void ari64_compile_thread(void *unused)
 {
+	struct ht_entry *hash_table =
+		*(void **)((char *)dynarec_local + LO_hash_table_ptr);
 	void *target;
 	u32 addr;
 
@@ -440,7 +445,8 @@ static void ari64_compile_thread(void *unused)
 		if (!ndrc_g.thread.busy || !addr || ndrc_g.thread.exit)
 			continue;
 
-		target = ndrc_get_addr_ht_param(addr, ndrc_cm_compile_in_thread);
+		target = ndrc_get_addr_ht_param(hash_table, addr,
+				ndrc_cm_compile_in_thread);
 		//printf("c  %08x -> %p\n", addr, target);
 		ndrc_g.thread.busy = 0;
 	}
@@ -520,7 +526,6 @@ static int ari64_thread_check_range(unsigned int start, unsigned int end) { retu
 
 static int ari64_init()
 {
-	static u32 scratch_buf[8*8*2] __attribute__((aligned(64)));
 	size_t i;
 
 	new_dynarec_init();
@@ -548,7 +553,6 @@ static int ari64_init()
 #endif
 	psxH_ptr = psxH;
 	zeromem_ptr = zero_mem;
-	scratch_buf_ptr = scratch_buf;
 
 	ari64_thread_init();
 
@@ -576,11 +580,6 @@ R3000Acpu psxRec = {
 #else // if DRC_DISABLE
 
 struct ndrc_globals ndrc_g; // dummy
-void *psxH_ptr;
-void *zeromem_ptr;
-u32 zero_mem[0x1000/4];
-void *mem_rtab;
-void *scratch_buf_ptr;
 void new_dynarec_init() {}
 void new_dyna_start(void *context) {}
 void new_dynarec_cleanup() {}

libpcsxcore/new_dynarec/linkage_arm.S

@@ -79,7 +79,7 @@ DRC_VAR(mem_wtab, 4)
 DRC_VAR(psxH_ptr, 4)
 DRC_VAR(zeromem_ptr, 4)
 DRC_VAR(invc_ptr, 4)
-DRC_VAR(scratch_buf_ptr, 4)
+DRC_VAR(hash_table_ptr, 4)
 DRC_VAR(ram_offset, 4)
 DRC_VAR(mini_ht, 256)
 
@@ -150,7 +150,9 @@ FUNCTION(dyna_linker):
 	mov	r5, r1
 	lsl	r6, r6, #8
 	/* must not compile - that might expire the caller block */
-	mov	r1, #0 /* ndrc_compile_mode */
+	ldr     r0, [fp, #LO_hash_table_ptr]
+	mov     r1, r4
+	mov	r2, #0 /* ndrc_compile_mode=ndrc_cm_no_compile */
 	bl	ndrc_get_addr_ht_param
 
 	movs	r8, r0
@@ -174,6 +176,7 @@ FUNCTION(dyna_linker):
 #else
 	/* XXX: should be able to do better than this... */
 #endif
+	ldr     r1, [fp, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
 	mov	pc, r0
 	.size	dyna_linker, .-dyna_linker
@@ -223,6 +226,7 @@ FUNCTION(jump_vaddr_r7):
 	add	r0, r7, #0
 	.size	jump_vaddr_r7, .-jump_vaddr_r7
 FUNCTION(jump_vaddr_r0):
+	ldr     r1, [fp, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
 	mov	pc, r0
 	.size	jump_vaddr_r0, .-jump_vaddr_r0
@@ -248,6 +252,7 @@ FUNCTION(cc_interrupt):
 	ldmfdne	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, pc}
 	cmp	r0, r9
 	moveq	pc, lr
+	ldr     r1, [fp, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
 	mov	pc, r0
 	.size	cc_interrupt, .-cc_interrupt
@@ -304,6 +309,7 @@ FUNCTION(jump_to_new_pc):
 	str	r1, [fp, #LO_last_count]
 	sub	r10, r10, r1
 	bne	new_dyna_leave
+	ldr     r1, [fp, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
 	mov	pc, r0
 	.size	jump_to_new_pc, .-jump_to_new_pc
@@ -408,6 +414,7 @@ FUNCTION(new_dyna_start):
 	stmfd	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, lr}
 	mov	fp, r0 /* dynarec_local */
 	ldr	r0, [fp, #LO_pcaddr]
+	ldr     r1, [fp, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
 new_dyna_start_at_e:
 	ldr	r1, [fp, #LO_next_interupt]

libpcsxcore/new_dynarec/linkage_arm64.S

@@ -81,7 +81,7 @@ DRC_VAR(mem_wtab, 8)
 DRC_VAR(psxH_ptr, 8)
 DRC_VAR(invc_ptr, 8)
 DRC_VAR(zeromem_ptr, 8)
-DRC_VAR(scratch_buf_ptr, 8)
+DRC_VAR(hash_table_ptr, 8)
 DRC_VAR(ram_offset, 8)
 DRC_VAR(mini_ht, 256)
 
@@ -92,6 +92,7 @@ DRC_VAR(mini_ht, 256)
 FUNCTION(dyna_linker):
 	/* r0 = virtual target address */
 	/* r1 = instruction to patch */
+	ldr     x1, [rFP, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
 	br	x0
 	ESIZE(dyna_linker, .-dyna_linker)
@@ -118,6 +119,7 @@ FUNCTION(cc_interrupt):
 	bne	2f
 	ret
 2:
+	ldr     x1, [rFP, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
 	br	x0
 	ESIZE(cc_interrupt, .-cc_interrupt)
@@ -173,6 +175,7 @@ FUNCTION(jump_to_new_pc):
 	sub	rCC, rCC, w1
 	str	w1, [rFP, #LO_last_count]
 	cbnz	w2, new_dyna_leave
+	ldr     x1, [rFP, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
 	br	x0
 	ESIZE(jump_to_new_pc, .-jump_to_new_pc)
@@ -188,8 +191,9 @@ FUNCTION(new_dyna_start):
 	stp	x29, x30, [sp, #-SSP_ALL]!
 	mov	rFP, x0
 	ldr	w0,  [rFP, #LO_pcaddr]
+	ldr     x1,  [rFP, #LO_hash_table_ptr]
 	bl	ndrc_get_addr_ht
-	mov	x1, x0
+	mov	x1,  x0
 new_dyna_start_at_e:
 	ldr	w3,  [rFP, #LO_next_interupt]
 	ldr	w2,  [rFP, #LO_cycle]

libpcsxcore/new_dynarec/linkage_offsets.h

@@ -38,8 +38,8 @@
 #define LO_psxH_ptr		(LO_mem_wtab + PTRSZ)
 #define LO_zeromem_ptr		(LO_psxH_ptr + PTRSZ)
 #define LO_invc_ptr		(LO_zeromem_ptr + PTRSZ)
-#define LO_scratch_buf_ptr	(LO_invc_ptr + PTRSZ)
-#define LO_saved_lr		(LO_scratch_buf_ptr + PTRSZ)
+#define LO_hash_table_ptr	(LO_invc_ptr + PTRSZ)
+#define LO_saved_lr		(LO_hash_table_ptr + PTRSZ)
 #define LO_ram_offset		(LO_saved_lr + PTRSZ)
 #define LO_mini_ht		(LO_ram_offset + PTRSZ)
 #define LO_dynarec_local_size	(LO_mini_ht + PTRSZ*32*2)

libpcsxcore/new_dynarec/new_dynarec.c

@@ -75,6 +75,7 @@ extern int last_count;  // last absolute target, often = next_interupt
 
 extern int reg_cop2d[], reg_cop2c[];
 
+extern void *hash_table_ptr;
 extern uintptr_t ram_offset;
 extern uintptr_t mini_ht[32][2];
 
@@ -400,7 +401,7 @@ void jump_to_new_pc();
 void call_gteStall();
 void new_dyna_leave();
 
-void *ndrc_get_addr_ht(u_int vaddr);
+void *ndrc_get_addr_ht(u_int vaddr, struct ht_entry *ht);
 void ndrc_add_jump_out(u_int vaddr, void *src);
 void ndrc_write_invalidate_one(u_int addr);
 static void ndrc_write_invalidate_many(u_int addr, u_int end);
@@ -642,9 +643,14 @@ static u_int get_page_prev(u_int vaddr)
   return page;
 }
 
+static struct ht_entry *hash_table_get_p(struct ht_entry *ht, u_int vaddr)
+{
+  return &ht[((vaddr >> 16) ^ vaddr) & 0xFFFF];
+}
+
 static struct ht_entry *hash_table_get(u_int vaddr)
 {
-  return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
+  return hash_table_get_p(hash_table, vaddr);
 }
 
 #define HASH_TABLE_BAD 0xbac
@@ -805,7 +811,8 @@ static noinline u_int generate_exception(u_int pc)
 
 // Get address from virtual address
 // This is called from the recompiled JR/JALR instructions
-static void noinline *get_addr(const u_int vaddr, enum ndrc_compile_mode compile_mode)
+static void noinline *get_addr(struct ht_entry *ht, const u_int vaddr,
+  enum ndrc_compile_mode compile_mode)
 {
   u_int start_page = get_page_prev(vaddr);
   u_int i, page, end_page = get_page(vaddr);
@@ -846,31 +853,32 @@ static void noinline *get_addr(const u_int vaddr, enum ndrc_compile_mode compile
 
   int r = new_recompile_block(vaddr);
   if (likely(r == 0))
-    return ndrc_get_addr_ht(vaddr);
+    return ndrc_get_addr_ht(vaddr, ht);
 
   if (compile_mode == ndrc_cm_compile_live)
-    return ndrc_get_addr_ht(generate_exception(vaddr));
+    return ndrc_get_addr_ht(generate_exception(vaddr), ht);
 
   return NULL;
 }
 
 // Look up address in hash table first
-void *ndrc_get_addr_ht_param(unsigned int vaddr, enum ndrc_compile_mode compile_mode)
+void *ndrc_get_addr_ht_param(struct ht_entry *ht, unsigned int vaddr,
+  enum ndrc_compile_mode compile_mode)
 {
   //check_for_block_changes(vaddr, vaddr + MAXBLOCK);
-  const struct ht_entry *ht_bin = hash_table_get(vaddr);
+  const struct ht_entry *ht_bin = hash_table_get_p(ht, vaddr);
   u_int vaddr_a = vaddr & ~3;
   stat_inc(stat_ht_lookups);
   if (ht_bin->vaddr[0] == vaddr_a) return ht_bin->tcaddr[0];
   if (ht_bin->vaddr[1] == vaddr_a) return ht_bin->tcaddr[1];
-  return get_addr(vaddr, compile_mode);
+  return get_addr(ht, vaddr, compile_mode);
 }
 
 // "usual" addr lookup for indirect branches, etc
 // to be used by currently running code only
-void *ndrc_get_addr_ht(u_int vaddr)
+void *ndrc_get_addr_ht(u_int vaddr, struct ht_entry *ht)
 {
-  return ndrc_get_addr_ht_param(vaddr, ndrc_cm_compile_live);
+  return ndrc_get_addr_ht_param(ht, vaddr, ndrc_cm_compile_live);
 }
 
 static void clear_all_regs(signed char regmap[])
@@ -1317,6 +1325,7 @@ static const char *fpofs_name(u_int ofs)
   ofscase(psxH_ptr);
   ofscase(invc_ptr);
   ofscase(ram_offset);
+  ofscase(hash_table_ptr);
   #undef ofscase
   }
   buf[0] = 0;
@@ -3439,9 +3448,10 @@ static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
     if(i_regs->regmap==regs[i].regmap) {
       load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
       wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
-      emit_movimm(start+i*4+4,0);
-      emit_writeword(0,&psxRegs.pc);
-      emit_addimm(HOST_CCREG,2,HOST_CCREG);
+      emit_readptr(&hash_table_ptr, 1);
+      emit_movimm(start+i*4+4, 0);
+      emit_writeword(0, &psxRegs.pc);
+      emit_addimm(HOST_CCREG, 2, HOST_CCREG);
       emit_far_call(ndrc_get_addr_ht);
       emit_jmpreg(0);
     }
@@ -3620,6 +3630,7 @@ static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
       emit_cmp(HOST_TEMPREG, 0);
       void *jaddr = out;
       emit_jeq(0);
+      emit_readptr(&hash_table_ptr, 1);
       emit_far_call(ndrc_get_addr_ht);
       emit_jmpreg(0);
       set_jump_target(jaddr, out);
@@ -6368,6 +6379,7 @@ void new_dynarec_init(void)
 #endif
   out = ndrc->translation_cache;
   new_dynarec_clear_full();
+  hash_table_ptr = hash_table;
 #ifdef HOST_IMM8
   // Copy this into local area so we don't have to put it in every literal pool
   invc_ptr=invalid_code;
@@ -6542,7 +6554,7 @@ void new_dynarec_load_blocks(const void *save, int size)
         psxRegs.GPR.r[i] = 0x1f800000;
     }
 
-    ndrc_get_addr_ht_param(sblocks[b].addr, ndrc_cm_compile_offline);
+    ndrc_get_addr_ht_param(hash_table, sblocks[b].addr, ndrc_cm_compile_offline);
 
     for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
       if (f & 1)
@@ -9038,6 +9050,7 @@ static int new_recompile_block(u_int addr)
     emit_addimm(0, 0x18, 0);
     emit_adds_ptr(1, 1, 1);
     emit_ldr_dualindexed(1, 0, 0);
+    emit_readptr(&hash_table_ptr, 1);
     emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
     emit_far_call(ndrc_get_addr_ht);
     emit_jmpreg(0); // jr k0

libpcsxcore/new_dynarec/new_dynarec.h

@@ -44,12 +44,14 @@ void new_dynarec_invalidate_all_pages(void);
 void new_dyna_start(void *context);
 void new_dyna_start_at(void *context, void *compiled_code);
 
+struct ht_entry;
 enum ndrc_compile_mode {
 	ndrc_cm_no_compile = 0,
 	ndrc_cm_compile_live,       // from executing code, vaddr is the current pc
 	ndrc_cm_compile_offline,
 	ndrc_cm_compile_in_thread,
 };
-void *ndrc_get_addr_ht_param(unsigned int vaddr, enum ndrc_compile_mode compile_mode);
+void *ndrc_get_addr_ht_param(struct ht_entry *ht, unsigned int vaddr,
+	enum ndrc_compile_mode compile_mode);
 
 extern unsigned int ndrc_smrv_regs[32];