32x drc functional on ARM, random adjustments

git-svn-id: file:///home/notaz/opt/svn/PicoDrive@824 be3aeb3a-fb24-0410-a615-afba39da0efa

cpu/drc/emit_arm.c

@@ -10,6 +10,7 @@
 	do { \
 		*(u32 *)ptr = x; \
 		ptr = (void *)((u8 *)ptr + sizeof(u32)); \
+		COUNT_OP; \
 	} while (0)
 
 #define EMIT(x) EMIT_PTR(tcache_ptr, x)
@@ -170,29 +171,26 @@
 
 static void emith_op_imm(int cond, int op, int r, unsigned int imm)
 {
-	u32 v, ror2;
+	int ror2, rn = r;
+	u32 v;
 
-	if (imm == 0 && op != A_OP_MOV)
+	if (op == A_OP_MOV)
+		rn = 0;
+	else if (imm == 0)
 		return;
 
-	/* shift down to get starting rot2 */
+	for (v = imm, ror2 = 0; v != 0 || op == A_OP_MOV; v >>= 8, ror2 -= 8/2) {
-	for (v = imm, ror2 = 0; v && !(v & 3); v >>= 2)
+		/* shift down to get 'best' rot2 */
-		ror2++;
+		for (; v && !(v & 3); v >>= 2)
-	ror2 = 16 - ror2;
+			ror2--;
 
-	EOP_C_DOP_IMM(cond, op, 0, op == A_OP_MOV ? 0 : r, r, ror2 & 0x0f, v & 0xff);
+		EOP_C_DOP_IMM(cond, op, 0, rn, r, ror2 & 0x0f, v & 0xff);
-	if (op == A_OP_MOV)
-		op = A_OP_ORR;
 
-	v >>= 8;
+		if (op == A_OP_MOV) {
-	if (v & 0xff)
+			op = A_OP_ORR;
-		EOP_C_DOP_IMM(cond, op, 0, r, r, (ror2 - 8/2) & 0x0f, v & 0xff);
+			rn = r;
-	v >>= 8;
+		}
-	if (v & 0xff)
+	}
-		EOP_C_DOP_IMM(cond, op, 0, r, r, (ror2 - 16/2) & 0x0f, v & 0xff);
-	v >>= 8;
-	if (v & 0xff)
-		EOP_C_DOP_IMM(cond, op, 0, r, r, (ror2 - 24/2) & 0x0f, v & 0xff);
 }
 
 #define is_offset_24(val) \
@@ -227,14 +225,6 @@ static int emith_xbranch(int cond, void *target, int is_call)
 	return (u32 *)tcache_ptr - start_ptr;
 }
 
-static void handle_caches(void)
-{
-#ifdef ARM
-	extern void cache_flush_d_inval_i(const void *start_addr, const void *end_addr);
-	cache_flush_d_inval_i(tcache, tcache_ptr);
-#endif
-}
-
 
 #define EMITH_CONDITIONAL(code, is_nonzero) { \
 	u32 val, cond, *ptr; \

cpu/drc/emit_x86.c

@@ -1,12 +1,5 @@
 #include <stdarg.h>
 
-#if (DRC_DEBUG & 1)
-#define COUNT_OP \
-	host_insn_count++
-#else
-#define COUNT_OP
-#endif
-
 enum { xAX = 0, xCX, xDX, xBX, xSP, xBP, xSI, xDI };
 
 #define CONTEXT_REG xBP

cpu/sh2/compiler.c

@@ -14,16 +14,22 @@
 #define DRC_DEBUG 0
 #endif
 
+#if DRC_DEBUG
 #define dbg(l,...) { \
   if ((l) & DRC_DEBUG) \
     elprintf(EL_STATUS, ##__VA_ARGS__); \
 }
 
-#if DRC_DEBUG
 #include "mame/sh2dasm.h"
 #include <platform/linux/host_dasm.h>
 static int insns_compiled, hash_collisions, host_insn_count;
+#define COUNT_OP \
+	host_insn_count++
+#else // !DRC_DEBUG
+#define COUNT_OP
+#define dbg(...)
 #endif
+
 #if (DRC_DEBUG & 2)
 static u8 *tcache_dsm_ptrs[3];
 static char sh2dasm_buff[64];
@@ -111,12 +117,12 @@ extern void sh2_drc_entry(SH2 *sh2, void *block);
 extern void sh2_drc_exit(void);
 
 // tmp
-extern void __attribute__((regparm(2))) sh2_do_op(SH2 *sh2, int opcode);
+extern void REGPARM(2) sh2_do_op(SH2 *sh2, int opcode);
-static void __attribute__((regparm(1))) sh2_test_irq(SH2 *sh2);
+static void REGPARM(1) sh2_test_irq(SH2 *sh2);
 
 static void flush_tcache(int tcid)
 {
-  printf("tcache #%d flush! (%d/%d, bds %d/%d)\n", tcid,
+  dbg(1, "tcache #%d flush! (%d/%d, bds %d/%d)", tcid,
     tcache_ptrs[tcid] - tcache_bases[tcid], tcache_sizes[tcid],
     block_counts[tcid], block_max_counts[tcid]);
 
@@ -469,6 +475,10 @@ end_block:
   emith_jump(sh2_drc_exit);
   tcache_ptrs[tcache_id] = tcache_ptr;
 
+#ifdef ARM
+  cache_flush_d_inval_i(block_entry, tcache_ptr);
+#endif
+
   do_host_disasm(tcache_id);
   dbg(1, " block #%d,%d tcache %d/%d, insns %d -> %d %.3f",
     tcache_id, block_counts[tcache_id],
@@ -476,6 +486,10 @@ end_block:
     insns_compiled, host_insn_count, (double)host_insn_count / insns_compiled);
   if ((sh2->pc & 0xc6000000) == 0x02000000) // ROM
     dbg(1, "  hash collisions %d/%d", hash_collisions, block_counts[tcache_id]);
+#if (DRC_DEBUG & 2)
+  fflush(stdout);
+#endif
+
   return block_entry;
 /*
 unimplemented:
@@ -592,7 +606,7 @@ void sh2_execute(SH2 *sh2, int cycles)
   sh2->cycles_done += cycles - ((signed int)sh2->sr >> 12);
 }
 
-static void __attribute__((regparm(1))) sh2_test_irq(SH2 *sh2)
+static void REGPARM(1) sh2_test_irq(SH2 *sh2)
 {
   if (sh2->pending_level > ((sh2->sr >> 4) & 0x0f))
   {
@@ -634,9 +648,23 @@ static void block_stats(void)
       (double)maxb->refcount / total * 100.0);
     maxb->refcount = 0;
   }
+
+  for (b = 0; b < ARRAY_SIZE(block_tables); b++)
+    for (i = 0; i < block_counts[b]; i++)
+      block_tables[b][i].refcount = 0;
 }
+#else
+#define block_stats()
 #endif
 
+void sh2_drc_flush_all(void)
+{
+  block_stats();
+  flush_tcache(0);
+  flush_tcache(1);
+  flush_tcache(2);
+}
+
 int sh2_drc_init(SH2 *sh2)
 {
   if (block_tables[0] == NULL) {
@@ -657,6 +685,9 @@ int sh2_drc_init(SH2 *sh2)
       tcache_bases[i] = tcache_ptrs[i] = tcache_bases[i - 1] + tcache_sizes[i - 1];
     }
 
+    // tmp
+    PicoOpt |= POPT_DIS_VDP_FIFO;
+
 #if (DRC_DEBUG & 2)
     for (i = 0; i < ARRAY_SIZE(block_tables); i++)
       tcache_dsm_ptrs[i] = tcache_bases[i];
@@ -678,9 +709,7 @@ int sh2_drc_init(SH2 *sh2)
 void sh2_drc_finish(SH2 *sh2)
 {
   if (block_tables[0] != NULL) {
-#if (DRC_DEBUG & 1)
     block_stats();
-#endif
     free(block_tables[0]);
     memset(block_tables, 0, sizeof(block_tables));
 

cpu/sh2/mame/sh2pico.c

@@ -104,8 +104,14 @@ void sh2_execute(SH2 *sh2_, int cycles)
 
 #else // DRC_TMP
 
+#ifdef __i386__
+#define REGPARM(x) __attribute__((regparm(x)))
+#else
+#define REGPARM(x)
+#endif
+
 // tmp
-void __attribute__((regparm(2))) sh2_do_op(SH2 *sh2_, int opcode)
+void REGPARM(2) sh2_do_op(SH2 *sh2_, int opcode)
 {
 	sh2 = sh2_;
 	sh2->pc += 2;

pico/32x/memory.c

@@ -735,7 +735,7 @@ static void PicoWrite8_32x_on(u32 a, u32 d)
 static void PicoWrite16_32x_on(u32 a, u32 d)
 {
   if ((a & 0xfc00) == 0x5000)
-    elprintf(EL_UIO, "m68k 32x w16 [%06x] %04x @%06x", a, d & 0xffff, SekPc);
+    elprintf(EL_32X, "m68k 32x w16 [%06x] %04x @%06x", a, d & 0xffff, SekPc);
 
   if ((a & 0xffc0) == 0x5100) { // a15100
     p32x_reg_write16(a, d);

pico/carthw/svp/compiler.c

@@ -35,6 +35,7 @@ void ssp_drc_next_patch(void){}
 void ssp_drc_end(void){}
 #endif
 
+#define COUNT_OP
 #include "../../../cpu/drc/emit_arm.c"
 
 // -----------------------------------------------------
@@ -1787,7 +1788,9 @@ void *ssp_translate_block(int pc)
 	exit(0);
 #endif
 
-	handle_caches();
+#ifdef ARM
+	cache_flush_d_inval_i(tcache, tcache_ptr);
+#endif
 
 	return block_start;
 }

pico/memory.c

@@ -957,7 +957,7 @@ static int ym2612_write_local(u32 a, u32 d, int is_from_z80)
   if (xcycles >= timer_b_next_oflow) \
     ym2612.OPN.ST.status |= (ym2612.OPN.ST.mode >> 2) & 2
 
-static u32 MEMH_FUNC ym2612_read_local_z80(void)
+static u32 ym2612_read_local_z80(void)
 {
   int xcycles = z80_cyclesDone() << 8;
 
@@ -1055,14 +1055,14 @@ void ym2612_unpack_state(void)
 // -----------------------------------------------------------------
 //                        z80 memhandlers
 
-static unsigned char MEMH_FUNC z80_md_vdp_read(unsigned short a)
+static unsigned char z80_md_vdp_read(unsigned short a)
 {
   // TODO?
   elprintf(EL_ANOMALY, "z80 invalid r8 [%06x] %02x", a, 0xff);
   return 0xff;
 }
 
-static unsigned char MEMH_FUNC z80_md_bank_read(unsigned short a)
+static unsigned char z80_md_bank_read(unsigned short a)
 {
   unsigned int addr68k;
   unsigned char ret;
@@ -1076,13 +1076,13 @@ static unsigned char MEMH_FUNC z80_md_bank_read(unsigned short a)
   return ret;
 }
 
-static void MEMH_FUNC z80_md_ym2612_write(unsigned int a, unsigned char data)
+static void z80_md_ym2612_write(unsigned int a, unsigned char data)
 {
   if (PicoOpt & POPT_EN_FM)
     emustatus |= ym2612_write_local(a, data, 1) & 1;
 }
 
-static void MEMH_FUNC z80_md_vdp_br_write(unsigned int a, unsigned char data)
+static void z80_md_vdp_br_write(unsigned int a, unsigned char data)
 {
   // TODO: allow full VDP access
   if ((a&0xfff9) == 0x7f11) // 7f11 7f13 7f15 7f17
@@ -1103,7 +1103,7 @@ static void MEMH_FUNC z80_md_vdp_br_write(unsigned int a, unsigned char data)
   elprintf(EL_ANOMALY, "z80 invalid w8 [%06x] %02x", a, data);
 }
 
-static void MEMH_FUNC z80_md_bank_write(unsigned int a, unsigned char data)
+static void z80_md_bank_write(unsigned int a, unsigned char data)
 {
   unsigned int addr68k;
 

pico/pico_int.h

@@ -789,10 +789,10 @@ extern void lprintf(const char *fmt, ...);
 #define cdprintf(x...)
 #endif
 
-#if defined(__GNUC__) && __GNUC__ >= 4 && __GNUC_MINOR__ >= 3
+#ifdef __i386__
-#define MEMH_FUNC __attribute__((aligned(4)))
+#define REGPARM(x) __attribute__((regparm(x)))
 #else
-#define MEMH_FUNC
+#define REGPARM(x)
 #endif
 
 #ifdef __GNUC__

pico/sms.c

@@ -169,7 +169,7 @@ static void write_bank(unsigned short a, unsigned char d)
   }
 }
 
-static void MEMH_FUNC xwrite(unsigned int a, unsigned char d)
+static void xwrite(unsigned int a, unsigned char d)
 {
   elprintf(EL_IO, "z80 write [%04x] %02x", a, d);
   if (a >= 0xc000)

platform/gp2x/Makefile

@@ -15,14 +15,15 @@ amalgamate = 0
 use_sh2drc = 1
 drc_debug = 1
 
+-include Makefile.local
 
 ifeq "$(debug_cyclone)" "1"
 use_cyclone = 1
 use_musashi = 1
 endif
 ifeq "$(use_musashi)" "1"
+# due to CPU stop flag acces
 asm_cdpico = 0
-asm_memory = 0
 asm_cdmemory = 0
 else
 use_cyclone = 1
@@ -119,8 +120,8 @@ DEFINES += DRC_SH2 DRC_TMP
 OBJS += cpu/sh2/mame/sh2pico.o
 OBJS += cpu/sh2/compiler.o
 OBJS += cpu/sh2/stub_arm.o
-ifeq "$(drc_debug)" "1"
+ifdef drc_debug
-DEFINES += DRC_DEBUG=1
+DEFINES += DRC_DEBUG=$(drc_debug)
 OBJS += cpu/sh2/mame/sh2dasm.o
 OBJS += platform/linux/host_dasm.o
 LDFLAGS += -lbfd -lopcodes -liberty

platform/linux/Makefile

@@ -22,13 +22,13 @@ CFLAGS += -ggdb -Wall -falign-functions=2
 endif
 DEFINES = _UNZIP_SUPPORT IO_STATS IN_EVDEV
 CFLAGS += -I../.. -I.
-LDFLAGS += -lpthread
+LDFLAGS += -lm -lpng
 ifeq "$(ARCH)" "arm"
 CFLAGS += -mcpu=arm920t
 DEFINES += ARM
-else
-LDFLAGS += -lX11
 endif
+LDFLAGS += -lpthread
+LDFLAGS += -lX11
 
 CC = $(CROSS)gcc
 
@@ -95,8 +95,8 @@ DEFINES += DRC_SH2 DRC_TMP
 OBJS += cpu/sh2/mame/sh2pico.o
 OBJS += cpu/sh2/compiler.o
 OBJS += cpu/sh2/stub_$(ARCH).o
-ifeq "$(drc_debug)" "1"
+ifdef drc_debug
-DEFINES += DRC_DEBUG=1
+DEFINES += DRC_DEBUG=$(drc_debug)
 OBJS += cpu/sh2/mame/sh2dasm.o
 OBJS += host_dasm.o
 LDFLAGS += -lbfd -lopcodes -liberty
@@ -132,7 +132,7 @@ tidy:
 
 PicoDrive : $(OBJS)
 	@echo ">>>" $@
-	$(CC) $(CFLAGS) $^ $(LDFLAGS) -lm -lpng -Wl,-Map=PicoDrive.map -o $@
+	$(CC) $(CFLAGS) $^ $(LDFLAGS) -Wl,-Map=PicoDrive.map -o $@
 
 mkdirs:
 	@mkdir -p $(DIRS)

platform/linux/host_dasm.c

@@ -37,7 +37,7 @@ remove_useless_symbols (asymbol **symbols, long count)
     {
       asymbol *sym = *in_ptr++;
 
-      if (sym->name == NULL || sym->name[0] == '\0')
+      if (sym->name == NULL || sym->name[0] == '\0' || sym->name[0] == '$')
         continue;
       if (sym->flags & (BSF_DEBUGGING | BSF_SECTION_SYM))
         continue;
@@ -53,6 +53,7 @@ remove_useless_symbols (asymbol **symbols, long count)
 */
       *out_ptr++ = sym;
     }
+
   return out_ptr - symbols;
 }