From: kub <derkub@gmail.com>
Date: Fri, 11 Oct 2019 22:26:11 +0000 (+0200)
Subject: sh2 drc bugfix for aarch64/mips
X-Git-Tag: v2.00~823
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=a6c0ab7d99b9d015b0d75a4455b1bf4acb9c9d6d;p=picodrive.git

sh2 drc bugfix for aarch64/mips
---

diff --git a/Makefile b/Makefile
index 15549dca..a79c054b 100644
--- a/Makefile
+++ b/Makefile
@@ -225,7 +225,7 @@ endif
 pprof: platform/linux/pprof.c
 	$(CC) $(CFLAGS) -O2 -ggdb -DPPROF -DPPROF_TOOL -I../../ -I. $^ -o $@ $(LDFLAGS) $(LDLIBS)
 
-pico/pico_int_offs.h:: tools/mkoffsets.sh
+pico/pico_int_offs.h: tools/mkoffsets.sh
 	make -C tools/ XCC="$(CC)" XCFLAGS="$(CFLAGS)"
 
 .s.o:
diff --git a/cpu/drc/emit_arm64.c b/cpu/drc/emit_arm64.c
index 3ef402b4..4bad6469 100644
--- a/cpu/drc/emit_arm64.c
+++ b/cpu/drc/emit_arm64.c
@@ -979,7 +979,7 @@ static void emith_ldst_offs(int sz, int rd, int rn, int o9, int ld, int mode)
 #define emith_save_caller_regs(mask) do { \
 	int _c, _r1, _r2; u32 _m = mask & 0x3ffff; \
 	if (__builtin_parity(_m) == 1) _m |= 0x40000; /* hardware align */ \
-	for (_c = HOST_REGS, _r1 = -1; _m && _c >= 0; _m &= ~(1 << _c), _c--) \
+	for (_c = HOST_REGS-1, _r1 = -1; _m && _c >= 0; _m &= ~(1 << _c), _c--)\
 		if (_m & (1 << _c)) { \
 			_r2 = _r1, _r1 = _c; \
 			if (_r2 != -1) { \
diff --git a/cpu/drc/emit_mips.c b/cpu/drc/emit_mips.c
index 4a452a68..38d68f40 100644
--- a/cpu/drc/emit_mips.c
+++ b/cpu/drc/emit_mips.c
@@ -1065,7 +1065,7 @@ static void emith_lohi_nops(void)
 	if (__builtin_parity(_m) == 1) _m |= 0x1; /* ABI align */ \
 	int _s = count_bits(_m) * 4, _o = _s; \
 	if (_s) emith_sub_r_imm(SP, _s); \
-	for (_c = HOST_REGS; _m && _c >= 0; _m &= ~(1 << _c), _c--) \
+	for (_c = HOST_REGS-1; _m && _c >= 0; _m &= ~(1 << _c), _c--) \
 		if (_m & (1 << _c)) \
 			{ _o -= 4; if (_c) emith_write_r_r_offs(_c, SP, _o); } \
 } while (0)
@@ -1279,7 +1279,7 @@ static int emith_cond_check(int cond, int *r)
 	if (__builtin_parity(_m) == 1) _m |= 0x1; /* ABI align for SP is 8 */ \
 	int _s = count_bits(_m) * 4 + 16, _o = _s; /* 16 byte arg save area */ \
 	if (_s) emith_sub_r_imm(SP, _s); \
-	for (_c = HOST_REGS; _m && _c >= 0; _m &= ~(1 << _c), _c--) \
+	for (_c = HOST_REGS-1; _m && _c >= 0; _m &= ~(1 << _c), _c--) \
 		if (_m & (1 << _c)) \
 			{ _o -= 4; if (_c) emith_write_r_r_offs(_c, SP, _o); } \
 } while (0)
diff --git a/cpu/drc/emit_x86.c b/cpu/drc/emit_x86.c
index 44e10ecf..212a12c5 100644
--- a/cpu/drc/emit_x86.c
+++ b/cpu/drc/emit_x86.c
@@ -1115,7 +1115,7 @@ enum { xAX = 0, xCX, xDX, xBX, xSP, xBP, xSI, xDI,	// x86-64,i386 common
 #define emith_save_caller_regs(mask) do { \
 	int _c; u32 _m = mask & 0xfc7; /* AX, CX, DX, SI, DI, 8, 9, 10, 11 */ \
 	if (__builtin_parity(_m) == 1) _m |= 0x8; /* BX for ABI align */ \
-	for (_c = HOST_REGS; _m && _c >= 0; _m &= ~(1 << _c), _c--) \
+	for (_c = HOST_REGS-1; _m && _c >= 0; _m &= ~(1 << _c), _c--) \
 		if (_m & (1 << _c)) emith_push(_c); \
 } while (0)
 
diff --git a/pico/32x/memory.c b/pico/32x/memory.c
index e139910a..60820e1a 100644
--- a/pico/32x/memory.c
+++ b/pico/32x/memory.c
@@ -347,7 +347,7 @@ static u32 p32x_reg_read16(u32 a)
   if ((a & 0x30) == 0x20) {
     unsigned int cycles = SekCyclesDone();
 
-    if (cycles - msh2.m68krcycles_done > 244)
+    if (CYCLES_GT(cycles - msh2.m68krcycles_done, 244))
       p32x_sync_sh2s(cycles);
 
     if (m68k_poll_detect(a, cycles, P32XF_68KCPOLL)) {
@@ -360,7 +360,7 @@ static u32 p32x_reg_read16(u32 a)
 
   if (a == 2) { // INTM, INTS
     unsigned int cycles = SekCyclesDone();
-    if (cycles - msh2.m68krcycles_done > 64)
+    if (CYCLES_GT(cycles - msh2.m68krcycles_done, 64))
       p32x_sync_sh2s(cycles);
     goto out;
   }
@@ -420,7 +420,7 @@ static void p32x_reg_write8(u32 a, u32 d)
       return;
     case 0x03: // irq ctl
       if ((d ^ r[0x02 / 2]) & 3) {
-        int cycles = SekCyclesDone();
+        unsigned int cycles = SekCyclesDone();
         p32x_sync_sh2s(cycles);
         r[0x02 / 2] = d & 3;
         p32x_update_cmd_irq(NULL, cycles);
@@ -610,9 +610,9 @@ static void p32x_reg_write16(u32 a, u32 d)
     case 0x2c/2:
     case 0x2e/2:
       if (r[a / 2] != d) {
-        int cycles = SekCyclesDone();
+        unsigned int cycles = SekCyclesDone();
 
-        if (cycles - (int)msh2.m68krcycles_done > 30)
+        if (CYCLES_GT(cycles - msh2.m68krcycles_done, 64))
           p32x_sync_sh2s(cycles);
 
         r[a / 2] = d;
@@ -712,7 +712,7 @@ static void p32x_vdp_write16(u32 a, u32 d, SH2 *sh2)
     }
     Pico32x.vdp_regs[0x06 / 2] = a;
     Pico32x.vdp_regs[0x08 / 2] = d;
-    if (sh2 != NULL && len > 4) {
+    if (sh2 != NULL && len > 8) {
       Pico32x.vdp_regs[0x0a / 2] |= P32XV_nFEN;
       // supposedly takes 3 bus/6 sh2 cycles? or 3 sh2 cycles?
       p32x_event_schedule_sh2(sh2, P32X_EVENT_FILLEND, 3 + len);
@@ -824,8 +824,8 @@ static void p32x_sh2reg_write8(u32 a, u32 d, SH2 *sh2)
       if (Pico32x.sh2_regs[4 / 2] != d) {
         unsigned int cycles = sh2_cycles_done_m68k(sh2);
         Pico32x.sh2_regs[4 / 2] = d;
-        sh2_end_run(sh2, 4);
         p32x_sh2_poll_event(sh2->other_sh2, SH2_STATE_CPOLL, cycles);
+        sh2_end_run(sh2, 4);
         sh2_poll_write(a & ~1, d, cycles, sh2);
       }
       return;
@@ -849,9 +849,9 @@ static void p32x_sh2reg_write8(u32 a, u32 d, SH2 *sh2)
         unsigned int cycles = sh2_cycles_done_m68k(sh2);
 
         REG8IN16(r, a) = d;
-        sh2_end_run(sh2, 1);
         p32x_m68k_poll_event(P32XF_68KCPOLL);
         p32x_sh2_poll_event(sh2->other_sh2, SH2_STATE_CPOLL, cycles);
+        sh2_end_run(sh2, 1);
         sh2_poll_write(a & ~1, r[a / 2], cycles, sh2);
       }
       return;
@@ -941,9 +941,9 @@ static void p32x_sh2reg_write16(u32 a, u32 d, SH2 *sh2)
         unsigned int cycles = sh2_cycles_done_m68k(sh2);
 
         Pico32x.regs[a / 2] = d;
-        sh2_end_run(sh2, 1);
         p32x_m68k_poll_event(P32XF_68KCPOLL);
         p32x_sh2_poll_event(sh2->other_sh2, SH2_STATE_CPOLL, cycles);
+        sh2_end_run(sh2, 1);
         sh2_poll_write(a, d, cycles, sh2);
       }
       return;
@@ -1574,10 +1574,10 @@ static void NOINLINE sh2_sdram_poll(u32 a, u32 d, SH2 *sh2)
   unsigned cycles;
 
   DRC_SAVE_SR(sh2);
-  sh2_end_run(sh2, 1);
   cycles = sh2_cycles_done_m68k(sh2);
   sh2_poll_write(a, d, cycles, sh2);
   p32x_sh2_poll_event(sh2->other_sh2, SH2_STATE_RPOLL, cycles);
+  sh2_end_run(sh2, 1);
   DRC_RESTORE_SR(sh2);
 }
 
diff --git a/tools/mkoffsets.sh b/tools/mkoffsets.sh
index 8f2d888c..2223b804 100755
--- a/tools/mkoffsets.sh
+++ b/tools/mkoffsets.sh
@@ -11,10 +11,10 @@ ENDIAN=
 # compile with target C compiler and extract value from .rodata section
 compile_rodata ()
 {
-	# $CC $CFLAGS -I .. -shared /tmp/getoffs.c -o /tmp/getoffs.o || exit 1
-	echo 'void dummy(void) { asm(""::"r" (&val)); }' >> /tmp/getoffs.c
-	$CC $CFLAGS -I .. -nostdlib -Wl,-edummy /tmp/getoffs.c \
-						-o /tmp/getoffs.o || exit 1
+	$CC $CFLAGS -I .. -c /tmp/getoffs.c -o /tmp/getoffs.o || exit 1
+	# echo 'void dummy(void) { asm(""::"r" (&val)); }' >> /tmp/getoffs.c
+	# $CC $CFLAGS -I .. -nostdlib -Wl,-edummy /tmp/getoffs.c \
+	#					-o /tmp/getoffs.o || exit 1
 	# find the name of the .rodata section (in case -fdata-sections is used)
 	rosect=$(readelf -S /tmp/getoffs.o | grep '\.rodata' |
 						sed 's/^[^.]*././;s/ .*//')
@@ -48,6 +48,7 @@ get_define () # prefix struct member member...
 	line=$(printf "#define %-20s 0x%04x" $prefix$name $rodata)
 }
 
+CFLAGS="$CFLAGS -fno-lto"
 # determine endianess
 echo "const int val = 1;" >/tmp/getoffs.c
 compile_rodata