Bochs/bochs/patches/patch.fetchdecode-cache

----------------------------------------------------------------------
Patch name: patche-fetchdecode-cache
Author: tld
Date: May, 29th 2002

Detailed description:

I somehow managed to implement the fetchdecode caching system I wrote 
about last night (and to which I didn't get any answer)
I couldn't benchmark it as I wanted (mainly, I lack the tools) so I 
don't really know what performance change it brought.

It has issues with code morphing (and some other things I can't 
recognize in the code), but if the cache is "small enough" it works.
For those interested, I left the source code on
http://tld.digitalcurse.com/bochs/cpu.cc which is meant to
replace the 
file cpu/cpu.cc in the CVS snapshot 20020527 which you can find at
http://tld.digitalcurse.com/bochs/bochs-20020527.tar.bz2 if
you don't 
have CVS access.

I'd like to hear comments on this.

PS. Of course, my code does NOT follow bochs' code standard which is not 
intended to be inserted into the release. This is just intended as a 
prototype did in 2 hours (of which, 1.5 for understanding the 
surrounding code...)

Patch was created with:
  cvs diff -u
Apply patch to what version:
  cvs checked out on May, 29th 2002
Instructions:
  To patch, go to main bochs directory.
  Type "patch -p0 < THIS_PATCH_FILE".
----------------------------------------------------------------------
diff -u -r1.28 cpu.cc
--- cpu/cpu.cc	18 Apr 2002 00:22:19 -0000	1.28
+++ cpu/cpu.cc	29 May 2002 08:56:13 -0000
@@ -37,6 +37,27 @@
 
 //unsigned counter[2] = { 0, 0 };
 
+#define TLD_DECODE_CACHE 1
+
+#if TLD_DECODE_CACHE
+  #define TLDs_array_size_1 0x00ff
+  #define TLD_debug 0
+  #define TLD_stats 0x100000
+
+  struct TLDs_ARRAY {
+    unsigned long eip;
+    BxInstruction_t i;
+  } TLDs_array[TLDs_array_size_1+1];
+
+  unsigned long TLD_cs;
+  unsigned long TLD_SEL;
+  unsigned long TLD_EIP;
+
+  #if TLD_stats
+    int TLD_CACHE_HIT = 0;
+    int TLD_CACHE_ACC = TLD_stats;
+  #endif // TLD_stats
+#endif // TLD_DECODE_CACHE
 
 
 #if BX_SIM_ID == 0   // only need to define once
@@ -106,7 +127,7 @@
 BX_CPU_C::cpu_loop(Bit32s max_instr_count)
 {
   unsigned ret;
-  BxInstruction_t i;
+  BxInstruction_t bxinstruction_dummy, *i = &bxinstruction_dummy;
   unsigned maxisize;
   Bit8u *fetch_ptr;
   Boolean is_32;
@@ -217,15 +238,65 @@
   maxisize = 16;
   if (BX_CPU_THIS_PTR bytesleft < 16)
     maxisize = BX_CPU_THIS_PTR bytesleft;
-  ret = FetchDecode(fetch_ptr, &i, maxisize, is_32);
+
+#if TLD_DECODE_CACHE
+  TLD_EIP = BX_CPU_THIS_PTR sregs[BX_SEG_REG_CS].cache.u.segment.base + EIP;
+  TLD_SEL = TLD_EIP & TLDs_array_size_1;
+
+  i = &TLDs_array[TLD_SEL].i;
+/*
+  if (TLD_cs != BX_CPU_THIS_PTR sregs[BX_SEG_REG_CS].cache.u.segment.base) {
+    for (TLD_cs = TLDs_array_size_1; (signed)TLD_cs >= 0; --TLD_cs) {
+      TLDs_array[TLD_cs].eip = 0xFFFFFFFF;
+    }
+    TLD_cs = BX_CPU_THIS_PTR sregs[BX_SEG_REG_CS].cache.u.segment.base;
+  }
+*/
+  if (TLDs_array[TLD_SEL].eip == TLD_EIP) {
+    // HIT!
+    #if TLD_stats
+      TLD_CACHE_HIT++;
+    #endif    
+
+    #if TLD_debug
+      printf("%8.8x:%8.8x !\n", BX_CPU_THIS_PTR sregs[BX_SEG_REG_CS].cache.u.segment.base, EIP);
+    #endif
+    ret = 1;
+  } else {
+    // MISS :(
+    ret = FetchDecode(fetch_ptr, i, maxisize, is_32);
+
+    #if TLD_debug
+      printf("%8.8x:%8.8x\n", BX_CPU_THIS_PTR sregs[BX_SEG_REG_CS].cache.u.segment.base, EIP);
+    #endif
+
+    if (ret) {
+      TLDs_array[TLD_SEL].eip = TLD_EIP;
+    } else {
+      TLDs_array[TLD_SEL].eip = 0xFFFFFFFF;
+    }
+  }
+
+  #if TLD_stats
+    if (!--TLD_CACHE_ACC) {
+      TLD_CACHE_ACC = TLD_stats;
+      printf("%6.6x\n", TLD_CACHE_HIT);
+      TLD_CACHE_HIT = 0;
+    }
+  #endif
+
+
+#else
+  ret = FetchDecode(fetch_ptr, i, maxisize, is_32);
+#endif // TLD_DECODE_CACHE
 
   if (ret) {
-    if (i.ResolveModrm) {
+    if (i->ResolveModrm) {
       // call method on BX_CPU_C object
-      BX_CPU_CALL_METHOD(i.ResolveModrm, (&i));
+      BX_CPU_CALL_METHOD(i->ResolveModrm, (i));
       }
-    BX_CPU_THIS_PTR fetch_ptr += i.ilen;
-    BX_CPU_THIS_PTR bytesleft -= i.ilen;
+    BX_CPU_THIS_PTR fetch_ptr += i->ilen;
+    BX_CPU_THIS_PTR bytesleft -= i->ilen;
 fetch_decode_OK:
 
 #if BX_DEBUGGER
@@ -239,34 +310,34 @@
     }
 #endif
 
-    if (i.rep_used && (i.attr & BxRepeatable)) {
+    if (i->rep_used && (i->attr & BxRepeatable)) {
 repeat_loop:
-      if (i.attr & BxRepeatableZF) {
-        if (i.as_32) {
+      if (i->attr & BxRepeatableZF) {
+        if (i->as_32) {
           if (ECX != 0) {
-            BX_CPU_CALL_METHOD(i.execute, (&i));
+            BX_CPU_CALL_METHOD(i->execute, (i));
             ECX -= 1;
             }
-          if ((i.rep_used==0xf3) && (get_ZF()==0)) goto repeat_done;
-          if ((i.rep_used==0xf2) && (get_ZF()!=0)) goto repeat_done;
+          if ((i->rep_used==0xf3) && (get_ZF()==0)) goto repeat_done;
+          if ((i->rep_used==0xf2) && (get_ZF()!=0)) goto repeat_done;
           if (ECX == 0) goto repeat_done;
           goto repeat_not_done;
           }
         else {
           if (CX != 0) {
-            BX_CPU_CALL_METHOD(i.execute, (&i));
+            BX_CPU_CALL_METHOD(i->execute, (i));
             CX -= 1;
             }
-          if ((i.rep_used==0xf3) && (get_ZF()==0)) goto repeat_done;
-          if ((i.rep_used==0xf2) && (get_ZF()!=0)) goto repeat_done;
+          if ((i->rep_used==0xf3) && (get_ZF()==0)) goto repeat_done;
+          if ((i->rep_used==0xf2) && (get_ZF()!=0)) goto repeat_done;
           if (CX == 0) goto repeat_done;
           goto repeat_not_done;
           }
         }
       else { // normal repeat, no concern for ZF
-        if (i.as_32) {
+        if (i->as_32) {
           if (ECX != 0) {
-            BX_CPU_CALL_METHOD(i.execute, (&i));
+            BX_CPU_CALL_METHOD(i->execute, (i));
             ECX -= 1;
             }
           if (ECX == 0) goto repeat_done;
@@ -274,7 +345,7 @@
           }
         else { // 16bit addrsize
           if (CX != 0) {
-            BX_CPU_CALL_METHOD(i.execute, (&i));
+            BX_CPU_CALL_METHOD(i->execute, (i));
             CX -= 1;
             }
           if (CX == 0) goto repeat_done;
@@ -302,12 +373,12 @@
 
 
 repeat_done:
-      BX_CPU_THIS_PTR eip += i.ilen;
+      BX_CPU_THIS_PTR eip += i->ilen;
       }
     else {
       // non repeating instruction
-      BX_CPU_THIS_PTR eip += i.ilen;
-      BX_CPU_CALL_METHOD(i.execute, (&i));
+      BX_CPU_THIS_PTR eip += i->ilen;
+      BX_CPU_CALL_METHOD(i->execute, (i));
       }
 
     BX_CPU_THIS_PTR prev_eip = EIP; // commit new EIP
@@ -410,13 +481,13 @@
     for (; j<16; j++) {
       FetchBuffer[j] = *temp_ptr++;
       }
-    ret = FetchDecode(FetchBuffer, &i, 16, is_32);
+    ret = FetchDecode(FetchBuffer, i, 16, is_32);
     if (ret==0)
       BX_PANIC(("fetchdecode: cross boundary: ret==0"));
-    if (i.ResolveModrm) {
-      BX_CPU_CALL_METHOD(i.ResolveModrm, (&i));
+    if (i->ResolveModrm) {
+      BX_CPU_CALL_METHOD(i->ResolveModrm, (i));
       }
-    remain = i.ilen - remain;
+    remain = i->ilen - remain;
 
     // note: eip has already been advanced to beginning of page
     BX_CPU_THIS_PTR fetch_ptr = fetch_ptr + remain;