From bd07808352da7c6d0d7f67b086319c2e681cd273 Mon Sep 17 00:00:00 2001
From: kub <derkub@gmail.com>
Date: Mon, 28 Jun 2021 22:58:04 +0200
Subject: [PATCH] 32x, improve poll detection

---
 cpu/sh2/compiler.c     | 1 +
 cpu/sh2/mame/sh2.c     | 1 +
 cpu/sh2/mame/sh2pico.c | 6 ++++--
 cpu/sh2/sh2.h          | 3 +++
 pico/32x/memory.c      | 6 ++++--
 pico/pico_int.h        | 4 ++++
 6 files changed, 17 insertions(+), 4 deletions(-)

diff --git a/cpu/sh2/compiler.c b/cpu/sh2/compiler.c
index 66400c77..bb2206c1 100644
--- a/cpu/sh2/compiler.c
+++ b/cpu/sh2/compiler.c
@@ -4417,6 +4417,7 @@ static void REGPARM(2) *sh2_translate(SH2 *sh2, int tcache_id)
           EMITH_HINT_COND(DCOND_EQ);
           emith_subf_r_r_imm(tmp, tmp2, 1);
           emith_set_t_cond(sr, DCOND_EQ);
+          emith_or_r_imm(sr, SH2_NO_POLLING);
           goto end_op;
         }
         goto default_;
diff --git a/cpu/sh2/mame/sh2.c b/cpu/sh2/mame/sh2.c
index fa49153a..17b96a31 100644
--- a/cpu/sh2/mame/sh2.c
+++ b/cpu/sh2/mame/sh2.c
@@ -800,6 +800,7 @@ INLINE void DT(sh2_state *sh2, UINT32 n)
 		sh2->sr |= T;
 	else
 		sh2->sr &= ~T;
+	sh2->no_polling = SH2_NO_POLLING;
 #if BUSY_LOOP_HACKS
 	{
 		UINT32 next_opcode = (UINT32)(UINT16)RW( sh2, sh2->ppc & AM );
diff --git a/cpu/sh2/mame/sh2pico.c b/cpu/sh2/mame/sh2pico.c
index f4ae85cb..65f4757e 100644
--- a/cpu/sh2/mame/sh2pico.c
+++ b/cpu/sh2/mame/sh2pico.c
@@ -24,9 +24,10 @@ typedef u8  UINT8;
 static __inline unsigned int name(SH2 *sh2, unsigned int a) \
 { \
 	unsigned int ret; \
-	sh2->sr |= sh2->icount << 12; \
+	sh2->sr |= (sh2->icount << 12) | (sh2->no_polling); \
 	ret = cname(a, sh2); \
 	sh2->icount = (signed int)sh2->sr >> 12; \
+	sh2->no_polling = (sh2->sr & SH2_NO_POLLING); \
 	sh2->sr &= 0x3f3; \
 	return ret; \
 }
@@ -34,9 +35,10 @@ static __inline unsigned int name(SH2 *sh2, unsigned int a) \
 #define MAKE_WRITEFUNC(name, cname) \
 static __inline void name(SH2 *sh2, unsigned int a, unsigned int d) \
 { \
-	sh2->sr |= sh2->icount << 12; \
+	sh2->sr |= (sh2->icount << 12) | (sh2->no_polling); \
 	cname(a, d, sh2); \
 	sh2->icount = (signed int)sh2->sr >> 12; \
+	sh2->no_polling = (sh2->sr & SH2_NO_POLLING); \
 	sh2->sr &= 0x3f3; \
 }
 
diff --git a/cpu/sh2/sh2.h b/cpu/sh2/sh2.h
index 614e7de1..bb8debe0 100644
--- a/cpu/sh2/sh2.h
+++ b/cpu/sh2/sh2.h
@@ -56,6 +56,9 @@ typedef struct SH2_
 	uint32_t	poll_addr;
 	int		poll_cycles;
 	int		poll_cnt;
+// NB MUST be a bit unused in SH2 SR, see also cpu/sh2/compiler.c!
+#define SH2_NO_POLLING	(1 << 10)	// poll detection control
+	int		no_polling;
 
 	// DRC branch cache. size must be 2^n and <=128
 	int rts_cache_idx;
diff --git a/pico/32x/memory.c b/pico/32x/memory.c
index ac26ba4b..d49cd2a5 100644
--- a/pico/32x/memory.c
+++ b/pico/32x/memory.c
@@ -119,8 +119,9 @@ void NOINLINE p32x_sh2_poll_detect(u32 a, SH2 *sh2, u32 flags, int maxcnt)
   // reading 2 consecutive 16bit values is probably a 32bit access. detect this
   // by checking address (max 2 bytes away) and cycles (max 2 cycles later).
   // no polling if more than 20 cycles have passed since last detect call.
-  if (a - sh2->poll_addr <= 2 && CYCLES_GE(10, cycles_diff)) {
-    if (CYCLES_GT(cycles_diff, 2) && ++sh2->poll_cnt >= maxcnt) {
+  if (a - sh2->poll_addr <= 2 && CYCLES_GE(20, cycles_diff)) {
+    if (!sh2_not_polling(sh2) && CYCLES_GT(cycles_diff, 2) &&
+                ++sh2->poll_cnt >= maxcnt) {
       if (!(sh2->state & flags))
         elprintf_sh2(sh2, EL_32X, "state: %02x->%02x",
           sh2->state, sh2->state | flags);
@@ -144,6 +145,7 @@ void NOINLINE p32x_sh2_poll_detect(u32 a, SH2 *sh2, u32 flags, int maxcnt)
     sh2->poll_addr = a;
   }
   sh2->poll_cycles = cycles_done;
+  sh2_set_polling(sh2);
 }
 
 void NOINLINE p32x_sh2_poll_event(SH2 *sh2, u32 flags, u32 m68k_cycles)
diff --git a/pico/pico_int.h b/pico/pico_int.h
index 18e33418..83e81418 100644
--- a/pico/pico_int.h
+++ b/pico/pico_int.h
@@ -224,6 +224,8 @@ extern SH2 sh2s[2];
 # define sh2_cycles_left(sh2) (sh2)->icount
 # define sh2_burn_cycles(sh2, n) (sh2)->icount -= n
 # define sh2_pc(sh2) (sh2)->ppc
+# define sh2_not_polling(sh2) (sh2)->no_polling
+# define sh2_set_polling(sh2) (sh2)->no_polling = 0
 #else
 # define sh2_end_run(sh2, after_) do { \
   int left_ = ((signed int)(sh2)->sr >> 12) - (after_); \
@@ -235,6 +237,8 @@ extern SH2 sh2s[2];
 # define sh2_cycles_left(sh2) ((signed int)(sh2)->sr >> 12)
 # define sh2_burn_cycles(sh2, n) (sh2)->sr -= ((n) << 12)
 # define sh2_pc(sh2) (sh2)->pc
+# define sh2_not_polling(sh2) ((sh2)->sr & SH2_NO_POLLING)
+# define sh2_set_polling(sh2) ((sh2)->sr &= ~SH2_NO_POLLING)
 #endif
 
 #define sh2_cycles_done(sh2) (unsigned)((int)(sh2)->cycles_timeslice - sh2_cycles_left(sh2))
-- 
2.39.2