From 3504ce942735ae4283cd47ca6defaae87b550c57 Mon Sep 17 00:00:00 2001
From: notaz <notasas@gmail.com>
Date: Mon, 9 Sep 2013 02:55:17 +0300
Subject: [PATCH] avoid using mrs

it's even worse on newer ARMs
only partially converted..
---
 Ea.cpp      | 45 +++++++++++++++++++++++++++++++--------------
 OpAny.cpp   |  6 ++++++
 OpArith.cpp | 51 ++++++++++++++++++++++++---------------------------
 OpLogic.cpp | 33 ++++++++++++++-------------------
 OpMove.cpp  |  7 +++----
 app.h       |  5 +++--
 6 files changed, 81 insertions(+), 66 deletions(-)

diff --git a/Ea.cpp b/Ea.cpp
index 234d469..a67f675 100644
--- a/Ea.cpp
+++ b/Ea.cpp
@@ -305,11 +305,15 @@ int EaCalc(int a,int mask,int ea,int size,int top,int sign_extend)
 // If top is 0 and sign_extend is not, then ARM register v is sign extended,
 // e.g. 0xc000 -> 0xffffc000 (else it may or may not be sign extended)
 
-int EaRead(int a,int v,int ea,int size,int mask,int top,int sign_extend)
+int EaRead(int a,int v,int ea,int size,int mask,int top,int sign_extend,int set_nz)
 {
   char text[32]="";
+  const char *s="";
+  int flags_set=0;
   int shift=0;
 
+  if (set_nz) s="s";
+
   shift=32-(8<<size);
 
   DisaPc=2; DisaGetEa(text,ea,size); // Get text version of the effective address
@@ -331,7 +335,8 @@ int EaRead(int a,int v,int ea,int size,int mask,int top,int sign_extend)
     else if (lsl<0) ot("  ldr%s r%d,[r7,r%d,lsr #%i]\n",Narm[nsarm],v,a,-lsl);
     else            ot("  ldr%s r%d,[r7,r%d]\n",Sarm[nsarm],v,a);
 
-    if (top&&shift) ot("  mov r%d,r%d,asl #%d\n",v,v,shift);
+    if (top&&shift) ot("  mov%s r%d,r%d,asl #%d\n",s,v,v,shift);
+    else if(set_nz) ot("  tst r%d,r%d\n",v,v);
 
     ot("\n"); return 0;
   }
@@ -344,8 +349,9 @@ int EaRead(int a,int v,int ea,int size,int mask,int top,int sign_extend)
 
     if (top) asl=shift;
 
-    if (asl) ot("  mov r%d,r%d,asl #%d\n",v,a,asl);
-    else if (v!=a) ot("  mov r%d,r%d\n",v,a);
+    if (asl)         ot("  mov%s r%d,r%d,asl #%d\n",s,v,a,asl);
+    else if (v!=a)   ot("  mov%s r%d,r%d\n",s,v,a);
+    else if (set_nz) ot("  tst r%d,r%d\n",v,v);
     ot("\n"); return 0;
   }
 
@@ -359,24 +365,35 @@ int EaRead(int a,int v,int ea,int size,int mask,int top,int sign_extend)
   {
     int d_reg=0;
     if (shift) {
-      ot("  mov r%d,r%d,asl #%d\n",v,d_reg,shift);
+      ot("  mov%s r%d,r%d,asl #%d\n",s,v,d_reg,shift);
       d_reg=v;
+      flags_set=1;
     }
     if (!top && shift) {
-      ot("  mov r%d,r%d,asr #%d\n",v,d_reg,shift);
+      ot("  mov%s r%d,r%d,asr #%d\n",s,v,d_reg,shift);
       d_reg=v;
+      flags_set=1;
+    }
+    if (d_reg != v) {
+      ot("  mov%s r%d,r%d\n",s,v,d_reg);
+      flags_set=1;
     }
-    if (d_reg != v)
-      ot("  mov r%d,r%d\n",v,d_reg);
   }
   else
   {
-    if (top && shift)
-      ot("  mov r%d,r0,asl #%d\n",v,shift);
-    else if (v!=0)
-      ot("  mov r%d,r0\n",v);
+    if (top && shift) {
+      ot("  mov%s r%d,r0,asl #%d\n",s,v,shift);
+      flags_set=1;
+    }
+    else if (v!=0) {
+      ot("  mov%s r%d,r0\n",s,v);
+      flags_set=1;
+    }
   }
 
+  if (set_nz&&!flags_set)
+    ot("  tst r%d,r%d\n",v,v);
+
   ot("\n"); return 0;
 }
 
@@ -387,7 +404,7 @@ int EaRead(int a,int v,int ea,int size,int mask,int top,int sign_extend)
 // size values 0, 1, 2 ~ byte, word, long
 // r_ea is reg to store ea in (-1 means ea is not needed), r is dst reg
 // if sign_extend is 0, non-32bit values will have MS bits undefined
-int EaCalcRead(int r_ea,int r,int ea,int size,int mask,int sign_extend)
+int EaCalcRead(int r_ea,int r,int ea,int size,int mask,int sign_extend,int set_nz)
 {
   if (ea<0x10)
   {
@@ -407,7 +424,7 @@ int EaCalcRead(int r_ea,int r,int ea,int size,int mask,int sign_extend)
   }
 
   EaCalc (r_ea,mask,ea,size,0,sign_extend);
-  EaRead (r_ea,   r,ea,size,mask,0,sign_extend);
+  EaRead (r_ea,   r,ea,size,mask,0,sign_extend,set_nz);
 
   return 0;
 }
diff --git a/OpAny.cpp b/OpAny.cpp
index 63a7e1f..8cb8c09 100644
--- a/OpAny.cpp
+++ b/OpAny.cpp
@@ -159,6 +159,12 @@ int OpGetFlags(int subtract,int xbit,int specialz)
   return 0;
 }
 
+void OpGetFlagsNZ(int rd)
+{
+  ot("  and r10,r%d,#0x80000000 ;@ r10=N_flag\n",rd);
+  ot("  orreq r10,r10,#0x40000000 ;@ get NZ, clear CV\n");
+}
+
 // -----------------------------------------------------------------
 
 int g_op;
diff --git a/OpArith.cpp b/OpArith.cpp
index ca279da..e956831 100644
--- a/OpArith.cpp
+++ b/OpArith.cpp
@@ -45,19 +45,17 @@ int OpArith(int op)
 
   ot(";@ Do arithmetic:\n");
 
-  if (type==0) ot("  orr r1,r10,r0%s\n",shiftstr);
-  if (type==1) ot("  and r1,r10,r0%s\n",shiftstr);
+  if (type==0) ot("  orrs r1,r10,r0%s\n",shiftstr);
+  if (type==1) ot("  ands r1,r10,r0%s\n",shiftstr);
   if (type==2||type==6)
                ot("  rsbs r1,r10,r0%s ;@ Defines NZCV\n",shiftstr);
   if (type==3) ot("  adds r1,r10,r0%s ;@ Defines NZCV\n",shiftstr);
-  if (type==5) ot("  eor r1,r10,r0%s\n",shiftstr);
+  if (type==5) ot("  eors r1,r10,r0%s\n",shiftstr);
 
-  if (type<2 || type==5) ot("  adds r1,r1,#0 ;@ Defines NZ, clears CV\n"); // 0,1,5
-
-  if (type< 2) OpGetFlags(0,0); // Ori/And
+  if (type< 2) OpGetFlagsNZ(1); // Ori/And
   if (type==2) OpGetFlags(1,1); // Sub: Subtract/X-bit
   if (type==3) OpGetFlags(0,1); // Add: X-bit
-  if (type==5) OpGetFlags(0,0); // Eor
+  if (type==5) OpGetFlagsNZ(1); // Eor
   if (type==6) OpGetFlags(1,0); // Cmp: Subtract
   ot("\n");
 
@@ -182,9 +180,9 @@ int OpArithReg(int op)
   EaCalcReadNoSE(dir?-1:11,1,rea,size,0x0e00);
 
   ot(";@ Do arithmetic:\n");
-  if (type==0) strop = "orr";
+  if (type==0) strop = "orrs";
   if (type==1) strop = (char *) (dir ? "subs" : "rsbs");
-  if (type==4) strop = "and";
+  if (type==4) strop = "ands";
   if (type==5) strop = "adds";
 
   if (size==0) asl=",asl #24";
@@ -193,9 +191,8 @@ int OpArithReg(int op)
   if (size<2) ot("  mov r0,r0%s\n",asl);
   ot("  %s r1,r0,r1%s\n",strop,asl);
 
-  if ((type&1)==0) ot("  adds r1,r1,#0 ;@ Defines NZ, clears CV\n");
-
-  OpGetFlags(type==1,type&1); // 1==subtract
+  if (type&1) OpGetFlags(type==1,type&1); // add/subtract
+  else        OpGetFlagsNZ(1);
   ot("\n");
 
   ot(";@ Save result:\n");
@@ -327,9 +324,8 @@ int OpMul(int op)
       ot("\n");
     }
 
-    ot("  mov r1,r3,lsl #16 ;@ Clip to 16-bits\n");
-    ot("  adds r1,r1,#0 ;@ Defines NZ, clears CV\n");
-    OpGetFlags(0,0);
+    ot("  movs r1,r3,lsl #16 ;@ Clip to 16-bits\n");
+    OpGetFlagsNZ(1);
 
     ot("  mov r1,r1,lsr #16\n");
     ot("  orr r1,r1,r2,lsl #16 ;@ Insert remainder\n");
@@ -343,9 +339,8 @@ int OpMul(int op)
     ot("  mov r2,r2,%s #16\n",sign?"asr":"lsr");
     ot("\n");
 
-    ot("  mul r1,r2,r0\n");
-    ot("  adds r1,r1,#0 ;@ Defines NZ, clears CV\n");
-    OpGetFlags(0,0);
+    ot("  muls r1,r2,r0\n");
+    OpGetFlagsNZ(1);
   }
   ot("\n");
 
@@ -716,14 +711,16 @@ int OpCmpEor(int op)
   if (size<2) asl=(char *)(size?",asl #16":",asl #24");
 
   ot(";@ Do arithmetic:\n");
-  if (eor==0) ot("  rsbs r1,r0,r1%s\n",asl);
   if (eor)
   {
-    ot("  eor r1,r0,r1%s\n",asl);
-    ot("  adds r1,r1,#0 ;@ Defines NZ, clears CV\n");
+    ot("  eors r1,r0,r1%s\n",asl);
+    OpGetFlagsNZ(1);
+  }
+  else
+  {
+    ot("  rsbs r1,r0,r1%s\n",asl);
+    OpGetFlags(1,0); // Cmp like subtract
   }
-
-  OpGetFlags(eor==0,0); // Cmp like subtract
   ot("\n");
 
   if (eor) EaWrite(11, 1,ea,size,0x003f,1);
@@ -800,18 +797,18 @@ int OpChk(int op)
   EaCalcReadNoSE(-1,1,rea,size,0x0e00);
 
   if (size<2) ot("  mov r0,r0,asl #%d\n",size?16:24);
-  if (size<2) ot("  mov r1,r1,asl #%d\n\n",size?16:24);
+
+  if (size<2) ot("  movs r1,r1,asl #%d\n\n",size?16:24);
+  else        ot("  adds r1,r1,#0 ;@ Define flags\n");
 
   ot(";@ get flags, including undocumented ones\n");
   ot("  and r3,r10,#0x80000000\n");
-  ot("  adds r1,r1,#0 ;@ Defines NZ, clears CV\n");
-  OpGetFlags(0,0);
+  OpGetFlagsNZ(1);
 
   ot(";@ is reg negative?\n");
   ot("  bmi chktrap%.4x\n",op);
 
   ot(";@ Do arithmetic:\n");
-  ot("  bic r10,r10,#0x80000000 ;@ N\n");
   ot("  cmp r1,r0\n");
   ot("  bgt chktrap%.4x\n",op);
 
diff --git a/OpLogic.cpp b/OpLogic.cpp
index 19ec569..4d4312c 100644
--- a/OpLogic.cpp
+++ b/OpLogic.cpp
@@ -285,12 +285,11 @@ int OpNeg(int op)
     ot(";@ Not:\n");
     if(size!=2) {
       ot("  mov r0,r0,asl #%i\n",size?16:24);
-      ot("  mvn r1,r0,asr #%i\n",size?16:24);
+      ot("  mvns r1,r0,asr #%i\n",size?16:24);
     }
     else
-      ot("  mvn r1,r0\n");
-    ot("  adds r1,r1,#0 ;@ Defines NZ, clears CV\n");
-    OpGetFlags(0,0);
+      ot("  mvns r1,r0\n");
+    OpGetFlagsNZ(1);
     ot("\n");
   }
 
@@ -318,9 +317,8 @@ int OpSwap(int op)
   EaCalc (11,0x0007,ea,2,1);
   EaRead (11,     0,ea,2,0x0007,1);
 
-  ot("  mov r1,r0,ror #16\n");
-  ot("  adds r1,r1,#0 ;@ Defines NZ, clears CV\n");
-  OpGetFlags(0,0);
+  ot("  movs r1,r0,ror #16\n");
+  OpGetFlagsNZ(1);
 
   EaWrite(11,     1,8,2,0x0007,1);
 
@@ -347,11 +345,10 @@ int OpTst(int op)
 
   OpStart(op,sea); Cycles=4;
 
-  EaCalc ( 0,0x003f,sea,size,1);
-  EaRead ( 0,     0,sea,size,0x003f,1);
+  EaCalc (0,0x003f,sea,size,1);
+  EaRead (0,     0,sea,size,0x003f,1,0,1);
 
-  ot("  adds r0,r0,#0 ;@ Defines NZ, clears CV\n");
-  ot("  mrs r10,cpsr ;@ r10=flags\n");
+  OpGetFlagsNZ(0);
   ot("\n");
 
   OpEnd(sea);
@@ -378,9 +375,8 @@ int OpExt(int op)
   EaCalc (11,0x0007,ea,size+1,0,0);
   EaRead (11,     0,ea,size+1,0x0007,0,0);
 
-  ot("  mov r0,r0,asl #%d\n",shift);
-  ot("  adds r0,r0,#0 ;@ Defines NZ, clears CV\n");
-  ot("  mrs r10,cpsr ;@ r10=flags\n");
+  ot("  movs r0,r0,asl #%d\n",shift);
+  OpGetFlagsNZ(0);
   ot("  mov r1,r0,asr #%d\n",shift);
   ot("\n");
 
@@ -596,8 +592,8 @@ static int EmitAsr(int op,int type,int dir,int count,int size,int usereg)
       ot("  b nozerox%.4x\n",op);
       ot("norotx_%.4x%s\n",op,ms?"":":");
       ot("  ldr r2,[r7,#0x4c]\n");
-      ot("  adds r0,r0,#0 ;@ Defines NZ, clears CV\n");
-      OpGetFlags(0,0);
+      ot("  adds r0,r0,#0 ;@ Define flags\n");
+      OpGetFlagsNZ(0);
       ot("  and r2,r2,#0x20000000\n");
       ot("  orr r10,r10,r2 ;@ C = old_X\n");
       ot("nozerox%.4x%s\n",op,ms?"":":");
@@ -746,10 +742,9 @@ int OpTas(int op, int gen_special)
   if(ea>=8) Cycles+=10;
 
   EaCalc (11,0x003f,ea,0,1);
-  EaRead (11,     1,ea,0,0x003f,1);
+  EaRead (11,     1,ea,0,0x003f,1,0,1);
 
-  ot("  adds r1,r1,#0 ;@ Defines NZ, clears CV\n");
-  OpGetFlags(0,0);
+  OpGetFlagsNZ(1);
   ot("\n");
 
 #if CYCLONE_FOR_GENESIS
diff --git a/OpMove.cpp b/OpMove.cpp
index 1e24566..82c589c 100644
--- a/OpMove.cpp
+++ b/OpMove.cpp
@@ -133,9 +133,8 @@ int OpMove(int op)
 
   if (movea==0)
   {
-    EaCalcRead(-1,0,sea,size,0x003f);
-    ot("  adds r1,r0,#0 ;@ Defines NZ, clears CV\n");
-    ot("  mrs r10,cpsr ;@ r10=NZCV flags\n");
+    EaCalcRead(-1,1,sea,size,0x003f,1,1);
+    OpGetFlagsNZ(1);
     ot("\n");
   }
   else
@@ -530,7 +529,7 @@ int OpMoveq(int op)
   ot("  movs r0,r8,asl #24\n");
   ot("  and r1,r8,#0x0e00\n");
   ot("  mov r0,r0,asr #24 ;@ Sign extended Quick value\n");
-  ot("  mrs r10,cpsr ;@ r10=NZ flags\n");
+  OpGetFlagsNZ(0);
   ot("  str r0,[r7,r1,lsr #7] ;@ Store into Dn\n");
   ot("\n");
 
diff --git a/app.h b/app.h
index 4a23630..f041dff 100644
--- a/app.h
+++ b/app.h
@@ -33,8 +33,8 @@ extern int g_pea_cycle_table[];
 extern int g_movem_cycle_table[];
 int Ea_add_ns(int *tab, int ea); // add nonstandard EA cycles
 int EaCalc(int a,int mask,int ea,int size,int top=0,int sign_extend=1); // 6
-int EaRead(int a,int v,int ea,int size,int mask,int top=0,int sign_extend=1); // 7
-int EaCalcRead(int r_ea,int r,int ea,int size,int mask,int sign_extend=1); // 6
+int EaRead(int a,int v,int ea,int size,int mask,int top=0,int sign_extend=1,int set_nz=0); // 8
+int EaCalcRead(int r_ea,int r,int ea,int size,int mask,int sign_extend=1,int set_nz=0); // 7
 int EaCalcReadNoSE(int r_ea,int r,int ea,int size,int mask);
 int EaCanRead(int ea,int size);
 int EaWrite(int a,int v,int ea,int size,int mask,int top=0,int sign_extend_ea=1);
@@ -58,6 +58,7 @@ void FlushPC(void);
 extern int g_op;
 extern int opend_op_changes_cycles, opend_check_interrupt, opend_check_trace;
 int OpGetFlags(int subtract,int xbit,int sprecialz=0);
+void OpGetFlagsNZ(int rd);
 void OpUse(int op,int use);
 void OpStart(int op,int sea=0,int tea=0,int op_changes_cycles=0,int supervisor_check=0);
 void OpEnd(int sea=0,int tea=0);
-- 
2.39.5