From: kub <derkub@gmail.com>
Date: Mon, 13 Apr 2020 20:20:13 +0000 (+0200)
Subject: sh2 timer optimization
X-Git-Tag: v2.00~752
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=74cc7aebf6a63af0506d311353585329c00f616f;p=picodrive.git

sh2 timer optimization
---

diff --git a/cpu/sh2/sh2.h b/cpu/sh2/sh2.h
index 2f2dfd92..aabe45be 100644
--- a/cpu/sh2/sh2.h
+++ b/cpu/sh2/sh2.h
@@ -48,6 +48,7 @@ typedef struct SH2_
 #define SH2_STATE_CPOLL (1 << 2)	// polling comm regs
 #define SH2_STATE_VPOLL (1 << 3)	// polling VDP
 #define SH2_STATE_RPOLL (1 << 4)	// polling address in SDRAM
+#define SH2_TIMER_RUN   (1 << 8)	// SOC WDT timer is running
 	unsigned int	state;
 	uint32_t	poll_addr;
 	int		poll_cycles;
diff --git a/pico/32x/32x.c b/pico/32x/32x.c
index 0f0cc4f5..ddd03fa8 100644
--- a/pico/32x/32x.c
+++ b/pico/32x/32x.c
@@ -508,12 +508,18 @@ void sync_sh2s_normal(unsigned int m68k_target)
           now = ssh2.m68krcycles_done;
       }
       if (CYCLES_GT(now, timer_cycles+STEP_N)) {
-        p32x_timers_do(now - timer_cycles);
+        if  (msh2.state & SH2_TIMER_RUN)
+          p32x_timer_do(&msh2, now - timer_cycles);
+        if  (ssh2.state & SH2_TIMER_RUN)
+          p32x_timer_do(&ssh2, now - timer_cycles);
         timer_cycles = now;
       }
     }
 
-    p32x_timers_do(now - timer_cycles);
+    if  (msh2.state & SH2_TIMER_RUN)
+      p32x_timer_do(&msh2, now - timer_cycles);
+    if  (ssh2.state & SH2_TIMER_RUN)
+      p32x_timer_do(&ssh2, now - timer_cycles);
     timer_cycles = now;
   }
   pprof_end_sub(m68k);
diff --git a/pico/32x/memory.c b/pico/32x/memory.c
index f4f0a18b..3f597288 100644
--- a/pico/32x/memory.c
+++ b/pico/32x/memory.c
@@ -111,7 +111,7 @@ void p32x_m68k_poll_event(u32 flags)
   m68k_poll.addr1 = m68k_poll.addr2 = m68k_poll.cnt = 0;
 }
 
-static void NOINLINE sh2_poll_detect(u32 a, SH2 *sh2, u32 flags, int maxcnt)
+void NOINLINE p32x_sh2_poll_detect(u32 a, SH2 *sh2, u32 flags, int maxcnt)
 {
   u32 cycles_done = sh2_cycles_done_t(sh2);
 
@@ -275,7 +275,7 @@ u32 REGPARM(3) p32x_sh2_poll_memory16(u32 a, u32 d, SH2 *sh2)
     d = (s16)sh2_poll_read(a, d, cycles, sh2);
   }
 
-  sh2_poll_detect(a, sh2, SH2_STATE_RPOLL, 5);
+  p32x_sh2_poll_detect(a, sh2, SH2_STATE_RPOLL, 5);
 
   DRC_RESTORE_SR(sh2);
   return d;
@@ -296,7 +296,7 @@ u32 REGPARM(3) p32x_sh2_poll_memory32(u32 a, u32 d, SH2 *sh2)
         ((u16)sh2_poll_read(a+2, d, cycles, sh2));
   }
 
-  sh2_poll_detect(a, sh2, SH2_STATE_RPOLL, 5);
+  p32x_sh2_poll_detect(a, sh2, SH2_STATE_RPOLL, 5);
 
   DRC_RESTORE_SR(sh2);
   return d;
@@ -735,7 +735,7 @@ static u32 p32x_sh2reg_read16(u32 a, SH2 *sh2)
       return (r[0] & P32XS_FM) | Pico32x.sh2_regs[0]
         | Pico32x.sh2irq_mask[sh2->is_slave];
     case 0x04/2: // H count (often as comm too)
-      sh2_poll_detect(a, sh2, SH2_STATE_CPOLL, 9);
+      p32x_sh2_poll_detect(a, sh2, SH2_STATE_CPOLL, 9);
       cycles = sh2_cycles_done_m68k(sh2);
       sh2s_sync_on_read(sh2, cycles);
       return sh2_poll_read(a, Pico32x.sh2_regs[4 / 2], cycles, sh2);
@@ -769,7 +769,7 @@ static u32 p32x_sh2reg_read16(u32 a, SH2 *sh2)
     case 0x2a/2:
     case 0x2c/2:
     case 0x2e/2:
-      sh2_poll_detect(a, sh2, SH2_STATE_CPOLL, 9);
+      p32x_sh2_poll_detect(a, sh2, SH2_STATE_CPOLL, 9);
       cycles = sh2_cycles_done_m68k(sh2);
       sh2s_sync_on_read(sh2, cycles);
       return sh2_poll_read(a, r[a / 2], cycles, sh2);
@@ -1456,7 +1456,7 @@ static u32 REGPARM(2) sh2_read8_cs0(u32 a, SH2 *sh2)
 
   if ((a & 0x3fff0) == 0x4100) {
     d = p32x_vdp_read16(a);
-    sh2_poll_detect(a, sh2, SH2_STATE_VPOLL, 9);
+    p32x_sh2_poll_detect(a, sh2, SH2_STATE_VPOLL, 9);
     goto out_16to8;
   }
 
@@ -1519,7 +1519,7 @@ static u32 REGPARM(2) sh2_read16_cs0(u32 a, SH2 *sh2)
 
   if ((a & 0x3fff0) == 0x4100) {
     d = p32x_vdp_read16(a);
-    sh2_poll_detect(a, sh2, SH2_STATE_VPOLL, 9);
+    p32x_sh2_poll_detect(a, sh2, SH2_STATE_VPOLL, 9);
     goto out;
   }
 
diff --git a/pico/32x/sh2soc.c b/pico/32x/sh2soc.c
index cf11666d..8895d49b 100644
--- a/pico/32x/sh2soc.c
+++ b/pico/32x/sh2soc.c
@@ -209,6 +209,9 @@ void p32x_timers_recalc(void)
 
   // SH2 timer step
   for (i = 0; i < 2; i++) {
+    sh2s[i].state &= ~SH2_TIMER_RUN;
+    if (PREG8(sh2s[i].peri_regs, 0x80) & 0x20) // TME
+      sh2s[i].state |= SH2_TIMER_RUN;
     tmp = PREG8(sh2s[i].peri_regs, 0x80) & 7;
     // Sclk cycles per timer tick
     if (tmp)
@@ -222,32 +225,29 @@ void p32x_timers_recalc(void)
   }
 }
 
-void p32x_timers_do(unsigned int m68k_slice)
+NOINLINE void p32x_timer_do(SH2 *sh2, unsigned int m68k_slice)
 {
   unsigned int cycles = m68k_slice * 3;
-  int cnt, i;
-
-  // WDT timers
-  for (i = 0; i < 2; i++) {
-    void *pregs = sh2s[i].peri_regs;
-    if (PREG8(pregs, 0x80) & 0x20) { // TME
-      timer_cycles[i] += cycles;
-      // cnt = timer_cycles[i] / timer_tick_cycles[i];
-      cnt = (1ULL * timer_cycles[i] * timer_tick_factor[i]) >> 32;
-      timer_cycles[i] -= timer_tick_cycles[i] * cnt;
-      if (timer_cycles[i] > timer_tick_cycles[i])
-        timer_cycles[i] -= timer_tick_cycles[i], cnt++;
-      cnt += PREG8(pregs, 0x81);
-      if (cnt >= 0x100) {
-        int level = PREG8(pregs, 0xe3) >> 4;
-        int vector = PREG8(pregs, 0xe4) & 0x7f;
-        elprintf(EL_32XP, "%csh2 WDT irq (%d, %d)",
-          i ? 's' : 'm', level, vector);
-        sh2_internal_irq(&sh2s[i], level, vector);
-        cnt &= 0xff;
-      }
-      PREG8(pregs, 0x81) = cnt;
+  void *pregs = sh2->peri_regs;
+  int cnt; int i = sh2->is_slave;
+
+  // WDT timer
+  timer_cycles[i] += cycles;
+  if (timer_cycles[i] > timer_tick_cycles[i]) {
+    // cnt = timer_cycles[i] / timer_tick_cycles[i];
+    cnt = (1ULL * timer_cycles[i] * timer_tick_factor[i]) >> 32;
+    timer_cycles[i] -= timer_tick_cycles[i] * cnt;
+
+    cnt += PREG8(pregs, 0x81);
+    if (cnt >= 0x100) {
+      int level = PREG8(pregs, 0xe3) >> 4;
+      int vector = PREG8(pregs, 0xe4) & 0x7f;
+      elprintf(EL_32XP, "%csh2 WDT irq (%d, %d)",
+        i ? 's' : 'm', level, vector);
+      sh2_internal_irq(sh2, level, vector);
+      cnt &= 0xff;
     }
+    PREG8(pregs, 0x81) = cnt;
   }
 }
 
diff --git a/pico/pico_int.h b/pico/pico_int.h
index 5fed483d..e4bd4c1e 100644
--- a/pico/pico_int.h
+++ b/pico/pico_int.h
@@ -977,6 +977,7 @@ unsigned int REGPARM(3) p32x_sh2_poll_memory8(unsigned int a, unsigned int d, SH
 unsigned int REGPARM(3) p32x_sh2_poll_memory16(unsigned int a, unsigned int d, SH2 *sh2);
 unsigned int REGPARM(3) p32x_sh2_poll_memory32(unsigned int a, unsigned int d, SH2 *sh2);
 void *p32x_sh2_get_mem_ptr(unsigned int a, unsigned int *mask, SH2 *sh2);
+void p32x_sh2_poll_detect(unsigned int a, SH2 *sh2, unsigned int flags, int maxcnt);
 void p32x_sh2_poll_event(SH2 *sh2, unsigned int flags, unsigned int m68k_cycles);
 int p32x_sh2_memcpy(unsigned int dst, unsigned int src, int count, int size, SH2 *sh2);
 
@@ -1012,7 +1013,7 @@ void p32x_pwm_state_loaded(void);
 void p32x_dreq0_trigger(void);
 void p32x_dreq1_trigger(void);
 void p32x_timers_recalc(void);
-void p32x_timers_do(unsigned int m68k_slice);
+void p32x_timer_do(SH2 *sh2, unsigned int m68k_slice);
 void sh2_peripheral_reset(SH2 *sh2);
 unsigned int REGPARM(2) sh2_peripheral_read8(unsigned int a, SH2 *sh2);
 unsigned int REGPARM(2) sh2_peripheral_read16(unsigned int a, SH2 *sh2);