From 86c16afd45de9e975445a46a96bdd43c80da3288 Mon Sep 17 00:00:00 2001
From: kub <derkub@gmail.com>
Date: Fri, 11 Oct 2019 00:56:26 +0200
Subject: [PATCH] 32x, speed improvement

---
 pico/32x/32x.c    | 28 ++++++++++++++++++----------
 pico/32x/sh2soc.c | 17 ++++++++++-------
 2 files changed, 28 insertions(+), 17 deletions(-)

diff --git a/pico/32x/32x.c b/pico/32x/32x.c
index f6d1a153..9993bfa8 100644
--- a/pico/32x/32x.c
+++ b/pico/32x/32x.c
@@ -426,7 +426,7 @@ void p32x_sync_other_sh2(SH2 *sh2, unsigned int m68k_target)
 }
 
 #define STEP_LS 24
-#define STEP_N 488 // one line
+#define STEP_N 528 // at least one line (488)
 
 #define sync_sh2s_normal p32x_sync_sh2s
 //#define sync_sh2s_lockstep p32x_sync_sh2s
@@ -434,7 +434,7 @@ void p32x_sync_other_sh2(SH2 *sh2, unsigned int m68k_target)
 /* most timing is in 68k clock */
 void sync_sh2s_normal(unsigned int m68k_target)
 {
-  unsigned int now, target, timer_cycles;
+  unsigned int now, target, next, timer_cycles;
   int cycles;
 
   elprintf(EL_32X, "sh2 sync to %u", m68k_target);
@@ -458,40 +458,44 @@ void sync_sh2s_normal(unsigned int m68k_target)
     target = m68k_target;
     if (event_time_next && CYCLES_GT(target, event_time_next))
       target = event_time_next;
-    if (CYCLES_GT(target, now + STEP_N))
-      target = now + STEP_N;
-
     while (CYCLES_GT(target, now))
     {
-      elprintf(EL_32X, "sh2 exec to %u %d,%d/%d, flags %x", target,
-        target - msh2.m68krcycles_done, target - ssh2.m68krcycles_done,
+      next = target;
+      if (CYCLES_GT(target, now + STEP_N))
+        next = now + STEP_N;
+      elprintf(EL_32X, "sh2 exec to %u %d,%d/%d, flags %x", next,
+        next - msh2.m68krcycles_done, next - ssh2.m68krcycles_done,
         m68k_target - now, Pico32x.emu_flags);
 
       pprof_start(ssh2);
       if (!(ssh2.state & SH2_IDLE_STATES)) {
-        cycles = target - ssh2.m68krcycles_done;
+        cycles = next - ssh2.m68krcycles_done;
         if (cycles > 0) {
           run_sh2(&ssh2, cycles > 20U ? cycles : 20U);
 
           if (event_time_next && CYCLES_GT(target, event_time_next))
             target = event_time_next;
+          if (CYCLES_GT(next, target))
+            next = target;
         }
       }
       pprof_end(ssh2);
 
       pprof_start(msh2);
       if (!(msh2.state & SH2_IDLE_STATES)) {
-        cycles = target - msh2.m68krcycles_done;
+        cycles = next - msh2.m68krcycles_done;
         if (cycles > 0) {
           run_sh2(&msh2, cycles > 20U ? cycles : 20U);
 
           if (event_time_next && CYCLES_GT(target, event_time_next))
             target = event_time_next;
+          if (CYCLES_GT(next, target))
+            next = target;
         }
       }
       pprof_end(msh2);
 
-      now = target;
+      now = next;
       if (!(msh2.state & SH2_IDLE_STATES)) {
         if (CYCLES_GT(now, msh2.m68krcycles_done))
           now = msh2.m68krcycles_done;
@@ -500,6 +504,10 @@ void sync_sh2s_normal(unsigned int m68k_target)
         if (CYCLES_GT(now, ssh2.m68krcycles_done))
           now = ssh2.m68krcycles_done;
       }
+      if (now - timer_cycles >= STEP_N) {
+        p32x_timers_do(now - timer_cycles);
+        timer_cycles = now;
+      }
     }
 
     p32x_timers_do(now - timer_cycles);
diff --git a/pico/32x/sh2soc.c b/pico/32x/sh2soc.c
index 2b5a126c..dd834bfb 100644
--- a/pico/32x/sh2soc.c
+++ b/pico/32x/sh2soc.c
@@ -193,8 +193,9 @@ static void dmac_trigger(SH2 *sh2, struct dma_chan *chan)
 }
 
 // timer state - FIXME
-static int timer_cycles[2];
-static int timer_tick_cycles[2];
+static u32 timer_cycles[2];
+static u32 timer_tick_cycles[2];
+static u32 timer_tick_factor[2];
 
 // timers
 void p32x_timers_recalc(void)
@@ -211,6 +212,7 @@ void p32x_timers_recalc(void)
     else
       cycles = 2;
     timer_tick_cycles[i] = cycles;
+    timer_tick_factor[i] = (1ULL << 32) / cycles;
     timer_cycles[i] = 0;
     elprintf(EL_32XP, "WDT cycles[%d] = %d", i, cycles);
   }
@@ -226,11 +228,12 @@ void p32x_timers_do(unsigned int m68k_slice)
     void *pregs = sh2s[i].peri_regs;
     if (PREG8(pregs, 0x80) & 0x20) { // TME
       timer_cycles[i] += cycles;
-      cnt = PREG8(pregs, 0x81);
-      while (timer_cycles[i] >= timer_tick_cycles[i]) {
-        timer_cycles[i] -= timer_tick_cycles[i];
-        cnt++;
-      }
+      // 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;
-- 
2.39.5