X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Fassem_arm64.c;h=bad2854cfd39b1f6f0039849ffead9218b2b79a3;hb=0c2126b9446fea1eb2a4e4c84fcb5ac1f364e81c;hp=670f3799e1425c1bd6dababed928b0132c5361de;hpb=277718fa66c96f64360b2c97a5dfa3ef3e6f1711;p=pcsx_rearmed.git

diff --git a/libpcsxcore/new_dynarec/assem_arm64.c b/libpcsxcore/new_dynarec/assem_arm64.c
index 670f3799..bad2854c 100644
--- a/libpcsxcore/new_dynarec/assem_arm64.c
+++ b/libpcsxcore/new_dynarec/assem_arm64.c
@@ -101,9 +101,17 @@ static void alloc_arm_reg(struct regstat *cur,int i,signed char reg,int hr)
 }
 
 // Alloc cycle count into dedicated register
-static void alloc_cc(struct regstat *cur,int i)
+static void alloc_cc(struct regstat *cur, int i)
 {
-  alloc_arm_reg(cur,i,CCREG,HOST_CCREG);
+  alloc_arm_reg(cur, i, CCREG, HOST_CCREG);
+}
+
+static void alloc_cc_optional(struct regstat *cur, int i)
+{
+  if (cur->regmap[HOST_CCREG] < 0) {
+    alloc_arm_reg(cur, i, CCREG, HOST_CCREG);
+    cur->noevict &= ~(1u << HOST_CCREG);
+  }
 }
 
 /* Special alloc */
@@ -484,7 +492,6 @@ static void emit_loadreg(u_int r, u_int hr)
     //case HIREG: addr = &hi; break;
     //case LOREG: addr = &lo; break;
     case CCREG: addr = &cycle_count; break;
-    case CSREG: addr = &psxRegs.CP0.n.SR; break;
     case INVCP: addr = &invc_ptr; is64 = 1; break;
     case ROREG: addr = &ram_offset; is64 = 1; break;
     default:
@@ -1458,6 +1465,7 @@ static void do_readstub(int n)
   int i = stubs[n].a;
   int rs = stubs[n].b;
   const struct regstat *i_regs = (void *)stubs[n].c;
+  int adj = (int)stubs[n].d;
   u_int reglist = stubs[n].e;
   const signed char *i_regmap = i_regs->regmap;
   int rt;
@@ -1520,12 +1528,22 @@ static void do_readstub(int n)
     handler=jump_handler_read32;
   assert(handler);
   pass_args64(rs,temp2);
-  int cc=get_reg(i_regmap,CCREG);
-  if(cc<0)
-    emit_loadreg(CCREG,2);
-  emit_addimm(cc<0?2:cc,(int)stubs[n].d,2);
+  int cc, cc_use;
+  cc = cc_use = get_reg(i_regmap, CCREG);
+  if (cc < 0)
+    emit_loadreg(CCREG, (cc_use = 2));
+  emit_addimm(cc_use, adj, 2);
+
   emit_far_call(handler);
-  // (no cycle reload after read)
+
+#if 0
+  // cycle reload for read32 only (value in w2 both in and out)
+  if (type == LOADW_STUB) {
+    emit_addimm(2, -adj, cc_use);
+    if (cc < 0)
+      emit_storereg(CCREG, cc_use);
+  }
+#endif
   if(dops[i].itype==C2LS||(rt>=0&&dops[i].rt1!=0)) {
     loadstore_extend(type,0,rt);
   }
@@ -1544,7 +1562,8 @@ static void inline_readstub(enum stub_type type, int i, u_int addr,
   u_int is_dynamic=0;
   uintptr_t host_addr = 0;
   void *handler;
-  int cc=get_reg(regmap,CCREG);
+  int cc, cc_use;
+  cc = cc_use = get_reg(regmap, CCREG);
   //if(pcsx_direct_read(type,addr,adj,cc,target?ra:-1,rt))
   //  return;
   handler = get_direct_memhandler(mem_rtab, addr, type, &host_addr);
@@ -1581,9 +1600,9 @@ static void inline_readstub(enum stub_type type, int i, u_int addr,
     emit_movimm(addr,0);
   else if(ra!=0)
     emit_mov(ra,0);
-  if(cc<0)
-    emit_loadreg(CCREG,2);
-  emit_addimm(cc<0?2:cc,adj,2);
+  if (cc < 0)
+    emit_loadreg(CCREG, (cc_use = 2));
+  emit_addimm(cc_use, adj, 2);
   if(is_dynamic) {
     uintptr_t l1 = ((uintptr_t *)mem_rtab)[addr>>12] << 1;
     intptr_t offset = (l1 & ~0xfffl) - ((intptr_t)out & ~0xfffl);
@@ -1599,7 +1618,16 @@ static void inline_readstub(enum stub_type type, int i, u_int addr,
 
   emit_far_call(handler);
 
-  // (no cycle reload after read)
+#if 0
+  // cycle reload for read32 only (value in w2 both in and out)
+  if (type == LOADW_STUB) {
+    if (!is_dynamic)
+      emit_far_call(do_memhandler_post);
+    emit_addimm(2, -adj, cc_use);
+    if (cc < 0)
+      emit_storereg(CCREG, cc_use);
+  }
+#endif
   if(rt>=0&&dops[i].rt1!=0)
     loadstore_extend(type, 0, rt);
   restore_regs(reglist);
@@ -1613,6 +1641,7 @@ static void do_writestub(int n)
   int i=stubs[n].a;
   int rs=stubs[n].b;
   struct regstat *i_regs=(struct regstat *)stubs[n].c;
+  int adj = (int)stubs[n].d;
   u_int reglist=stubs[n].e;
   signed char *i_regmap=i_regs->regmap;
   int rt,r;
@@ -1680,16 +1709,19 @@ static void do_writestub(int n)
     emit_mov64(temp2,3);
     host_tempreg_release();
   }
-  int cc=get_reg(i_regmap,CCREG);
-  if(cc<0)
-    emit_loadreg(CCREG,2);
-  emit_addimm(cc<0?2:cc,(int)stubs[n].d,2);
-  // returns new cycle_count
+  int cc, cc_use;
+  cc = cc_use = get_reg(i_regmap, CCREG);
+  if (cc < 0)
+    emit_loadreg(CCREG, (cc_use = 2));
+  emit_addimm(cc_use, adj, 2);
+
   emit_far_call(handler);
-  emit_addimm(0,-(int)stubs[n].d,cc<0?2:cc);
-  if(cc<0)
-    emit_storereg(CCREG,2);
-  if(restore_jump)
+
+  // new cycle_count returned in x2
+  emit_addimm(2, -adj, cc_use);
+  if (cc < 0)
+    emit_storereg(CCREG, cc_use);
+  if (restore_jump)
     set_jump_target(restore_jump, out);
   restore_regs(reglist);
   emit_jmp(stubs[n].retaddr);
@@ -1729,7 +1761,7 @@ static void inline_writestub(enum stub_type type, int i, u_int addr,
   emit_far_call(do_memhandler_pre);
   emit_far_call(handler);
   emit_far_call(do_memhandler_post);
-  emit_addimm(0, -adj, cc_use);
+  emit_addimm(2, -adj, cc_use);
   if (cc < 0)
     emit_storereg(CCREG, cc_use);
   restore_regs(reglist);
@@ -1943,6 +1975,65 @@ static void multdiv_assemble_arm64(int i, const struct regstat *i_regs)
 }
 #define multdiv_assemble multdiv_assemble_arm64
 
+// wb_dirtys making use of stp when possible
+static void wb_dirtys(const signed char i_regmap[], u_int i_dirty)
+{
+  signed char mregs[34+1];
+  int r, hr;
+  memset(mregs, -1, sizeof(mregs));
+  for (hr = 0; hr < HOST_REGS; hr++) {
+    r = i_regmap[hr];
+    if (hr == EXCLUDE_REG || r <= 0 || r == CCREG)
+      continue;
+    if (!((i_dirty >> hr) & 1))
+      continue;
+    assert(r < 34u);
+    mregs[r] = hr;
+  }
+  for (r = 1; r < 34; r++) {
+    if (mregs[r] < 0)
+      continue;
+    if (mregs[r+1] >= 0) {
+      uintptr_t offset = (u_char *)&psxRegs.GPR.r[r] - (u_char *)&dynarec_local;
+      emit_ldstp(1, 0, mregs[r], mregs[r+1], FP, offset);
+      r++;
+    }
+    else
+      emit_storereg(r, mregs[r]);
+  }
+}
+#define wb_dirtys wb_dirtys
+
+static void load_all_regs(const signed char i_regmap[])
+{
+  signed char mregs[34+1];
+  int r, hr;
+  memset(mregs, -1, sizeof(mregs));
+  for (hr = 0; hr < HOST_REGS; hr++) {
+    r = i_regmap[hr];
+    if (hr == EXCLUDE_REG || r < 0 || r == CCREG)
+      continue;
+    if ((u_int)r < 34u)
+      mregs[r] = hr;
+    else if (r < TEMPREG)
+      emit_loadreg(r, hr);
+  }
+  if (mregs[0] >= 0)
+    emit_zeroreg(mregs[0]); // we could use arm64's ZR instead of reg alloc
+  for (r = 1; r < 34; r++) {
+    if (mregs[r] < 0)
+      continue;
+    if (mregs[r+1] >= 0) {
+      uintptr_t offset = (u_char *)&psxRegs.GPR.r[r] - (u_char *)&dynarec_local;
+      emit_ldstp(0, 0, mregs[r], mregs[r+1], FP, offset);
+      r++;
+    }
+    else
+      emit_loadreg(r, mregs[r]);
+  }
+}
+#define load_all_regs load_all_regs
+
 static void do_jump_vaddr(u_int rs)
 {
   if (rs != 0)