TO DO:

- FIX CIELOS
- FIX HUGE STACKS

BYE

src/engine/cmdStream.cpp

@@ -24,7 +24,7 @@
 #include "../ta-log.h"
 
 bool DivCSChannelState::doCall(unsigned int addr) {
-  if (callStackPos>=16) {
+  if (callStackPos>=DIV_MAX_CSSTACK) {
     readPos=0;
     return false;
   }

src/engine/cmdStream.h

@@ -24,6 +24,7 @@
 #include "safeReader.h"
 
 #define DIV_MAX_CSTRACE 64
+#define DIV_MAX_CSSTACK 128
 
 class DivEngine;
 
@@ -39,7 +40,7 @@ struct DivCSChannelState {
   int portaTarget, portaSpeed;
   unsigned char arp, arpStage, arpTicks, loopCount;
 
-  unsigned int callStack[16];
+  unsigned int callStack[DIV_MAX_CSSTACK];
   unsigned char callStackPos;
 
   unsigned int trace[DIV_MAX_CSTRACE];

src/engine/cmdStreamOps.cpp

@@ -662,6 +662,7 @@ SafeWriter* stripNops(SafeWriter* s) {
       case 0xf5: // calli
       case 0xfa: { // jmp
         unsigned int addr=buf[i+1]|(buf[i+2]<<8)|(buf[i+3]<<16)|(buf[i+4]<<24);
+        assert(!(addr&7));
         try {
           addr=addrTable[addr];
           buf[i+1]=addr&0xff;
@@ -780,14 +781,11 @@ struct BlockMatch {
 
 // TODO:
 // - see if we can optimize even more
-SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlocks, unsigned char* speedDial) {
+SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlocks, unsigned char* speedDial, DivCSProgress* progress) {
   unsigned char* buf=stream->getFinalBuf();
   size_t matchSize=MIN_MATCH_SIZE;
   std::vector<BlockMatch> matches;
 
-  // repeat until we run out of matches
-  while (true) {
-    matchSize=MIN_MATCH_SIZE;
   matches.clear();
 
   // fast match algorithm
@@ -842,6 +840,7 @@ SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlock
   // set done to false unless:
   // - this match overlaps with itself
   // - this block only consists of calls
+  // - this block contains a ret or jmp
   size_t nonOverlapCount=0;
   for (BlockMatch& i: matches) {
     i.done=false;
@@ -851,11 +850,20 @@ SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlock
     } else {
       bool onlyCalls=true;
       for (size_t j=i.orig; j<i.orig+i.len; j+=8) {
-          if (buf[j]!=0xf4) {
+        if (buf[j]!=0xf4 && buf[j]!=0xf5) {
           onlyCalls=false;
           break;
         }
       }
+      if (!onlyCalls) {
+        // check whether there's call or ret
+        for (size_t j=i.orig; j<i.orig+i.len; j+=8) {
+          if (buf[j]==0xf9 || buf[j]==0xfa) {
+            onlyCalls=true;
+            break;
+          }
+        }
+      }
       if (onlyCalls) {
         i.done=true;
       } else {
@@ -866,10 +874,14 @@ SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlock
 
   logD("%d good candidates",(int)nonOverlapCount);
 
+  if (progress!=NULL) {
+    progress->count=nonOverlapCount;
+  }
+
   // NEW STUFF
   // find and sort matches by benefit
   size_t bestBenefitIndex=0;
-    int bestBenefit=-1000000;
+  int bestBenefit=-1;
   size_t lastOrig=SIZE_MAX;
   size_t lastLen=SIZE_MAX;
   size_t lastOrigOff=0;
@@ -885,9 +897,11 @@ SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlock
       if (lastOrig!=SIZE_MAX) {
         // commit previous block and start new one
         //logV("%x gains: %d",(int)lastOrig,gains);
-          if (gains>bestBenefit) {
+        int finalBenefit=gains*2+lastLen*3;
+        if (gains<1) finalBenefit=-1;
+        if (finalBenefit>bestBenefit) {
           bestBenefitIndex=lastOrigOff;
-            bestBenefit=gains;
+          bestBenefit=finalBenefit;
         }
         if (gains<=0) {
           // don't make a sub-block for these matches since we only have loss
@@ -1020,15 +1034,12 @@ SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlock
   logV("done!");
 
   // get out if we haven't made any blocks
-   if (!newBlocks) break;
+  if (!newBlocks) return stream;
 
   // remove nop's
   stream=stripNops(stream);
   buf=stream->getFinalBuf();
 
-    logV("doing it again...");
-  }
-
   return stream;
 }
 
@@ -1059,19 +1070,20 @@ SafeWriter* packStream(SafeWriter* s, unsigned char* speedDial) {
         try {
           addr=addrTable[addr];
           // check whether we have sufficient room to turn this into a 16-bit call
+          /*
           if (addr<0xff00) {
             buf[i]=0xf8;
             buf[i+1]=addr&0xff;
             buf[i+2]=(addr>>8)&0xff;
             buf[i+3]=0xf1;
             buf[i+4]=0xf1;
-          } else {
+          } else {*/
             buf[i]=0xf5;
             buf[i+1]=addr&0xff;
             buf[i+2]=(addr>>8)&0xff;
             buf[i+3]=(addr>>16)&0xff;
             buf[i+4]=(addr>>24)&0xff;
-          }
+          //}
         } catch (std::out_of_range& e) {
           logW("address %x is not mappable!",addr);
         }
@@ -1458,21 +1470,15 @@ SafeWriter* DivEngine::saveCommand(DivCSProgress* progress, unsigned int disable
     
     // 6 is the minimum size that can be reliably optimized
     logI("finding sub-blocks");
-    globalStream=findSubBlocks(globalStream,subBlocks,sortedCmd);
-    // find sub-blocks within sub-blocks
-    size_t subBlocksLast=0;
-    size_t subBlocksLen=subBlocks.size();
-    logI("finding sub-blocks within sub-blocks");
-    while (subBlocksLast!=subBlocksLen) {
-      logI("got %d blocks... starting from %d",(int)subBlocksLen,(int)subBlocksLast);
-      for (size_t i=subBlocksLast; i<subBlocksLen; i++) {
-        SafeWriter* newBlock=findSubBlocks(subBlocks[i],subBlocks,sortedCmd);
-        subBlocks[i]=newBlock;
-      }
-      subBlocksLast=subBlocksLen;
-      subBlocksLen=subBlocks.size();
-    }
 
+    bool haveBlocks=false;
+    subBlocks.clear();
+    // repeat until no more sub-blocks are produced
+    do {
+      logD("iteration...");
+      globalStream=findSubBlocks(globalStream,subBlocks,sortedCmd,progress);
+
+      haveBlocks=!subBlocks.empty();
       // insert sub-blocks and resolve symbols
       logI("%d sub-blocks total",(int)subBlocks.size());
       std::vector<size_t> blockOff;
@@ -1528,6 +1534,7 @@ SafeWriter* DivEngine::saveCommand(DivCSProgress* progress, unsigned int disable
           }
         }
       }
+    } while (haveBlocks);
 
     size_t afterSize=globalStream->size();
     logI("(before: %d - after: %d)",(int)beforeSize,(int)afterSize);

src/gui/gui.cpp

@@ -6064,7 +6064,7 @@ bool FurnaceGUI::loop() {
         }
       } else {
         WAKE_UP;
-        ImGui::Text("Exporting...");
+        ImGui::Text("Exporting... %d",csProgress.count);
 
         // check whether we're done
         if (csExportDone) {