otomatalabs/furnace
7
0
Fork 0
Code Issues Pull requests Actions Packages Projects Releases Wiki Activity

fix build failure

Browse source
This commit is contained in:
tildearrow 2025-04-07 20:44:25 -05:00
parent 914ba11e3b
commit 584305f1fa
1 changed files with 116 additions and 126 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

242
src/engine/cmdStreamOps.cpp
View file

@ -678,8 +678,7 @@ struct BlockMatch {
#define MIN_MATCH_SIZE 16 #define MIN_MATCH_SIZE 16
// TODO: // TODO:
// - check whether a block consists only of calls // - see if we can optimize even more
// - see if we can optimize better
SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlocks, unsigned char* speedDial) { SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlocks, unsigned char* speedDial) {
unsigned char* buf=stream->getFinalBuf(); unsigned char* buf=stream->getFinalBuf();
size_t matchSize=MIN_MATCH_SIZE; size_t matchSize=MIN_MATCH_SIZE;
@ -709,7 +708,6 @@ SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlock
// search for bigger matches // search for bigger matches
bool wantMore=true; bool wantMore=true;
do { do {
size_t matchCount=0;
wantMore=false; wantMore=false;
matchSize+=8; matchSize+=8;
for (size_t i=0; i<matches.size(); i++) { for (size_t i=0; i<matches.size(); i++) {
@ -729,13 +727,11 @@ SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlock
// this match may be bigger // this match may be bigger
b.len=matchSize; b.len=matchSize;
wantMore=true; wantMore=true;
matchCount++;
} else { } else {
// this is the max size // this is the max size
b.done=true; b.done=true;
} }
} }
//logV("size %d: %d matches",(int)matchSize,(int)matchCount);
} while (wantMore); } while (wantMore);
// first stage done // first stage done
@ -776,144 +772,138 @@ SafeWriter* findSubBlocks(SafeWriter* stream, std::vector<SafeWriter*>& subBlock
std::vector<BlockMatch> workMatches; std::vector<BlockMatch> workMatches;
bool newBlocks=false; bool newBlocks=false;
// try with a smaller size // try with a smaller size
matchSize=0; matchSize=0;
for (BlockMatch& i: matches) { for (BlockMatch& i: matches) {
if (i.done) continue; if (i.done) continue;
if (i.len>matchSize) matchSize=i.len; if (i.len>matchSize) matchSize=i.len;
}
workMatches.clear();
// find matches with matching size
for (BlockMatch& i: matches) {
if (i.len==matchSize) {
// mark it as done and push it
workMatches.push_back(i);
i.done=true;
} }
}
//while (matchSize>=MIN_MATCH_SIZE) { // check which sub-blocks are viable to make
workMatches.clear(); size_t lastOrig=SIZE_MAX;
// find matches with matching size size_t lastOrigOff=0;
for (BlockMatch& i: matches) { int gains=0;
if (i.len==matchSize) { int blockSize=0;
// mark it as done and push it for (size_t i=0; i<=workMatches.size(); i++) {
workMatches.push_back(i); BlockMatch b(SIZE_MAX,SIZE_MAX,0);
i.done=true; if (i<workMatches.size()) b=workMatches[i];
} // unlikely
} if (b.done) continue;
// check which sub-blocks are viable to make if (b.orig!=lastOrig) {
size_t lastOrig=SIZE_MAX; if (lastOrig!=SIZE_MAX) {
size_t lastOrigOff=0; // commit previous block and start new one
int gains=0; logV("%x gains: %d",(int)lastOrig,gains);
int blockSize=0; if (gains<=0) {
for (size_t i=0; i<=workMatches.size(); i++) { // don't make a sub-block for these matches since we only have loss
BlockMatch b(SIZE_MAX,SIZE_MAX,0); logV("(LOSSES!)");
if (i<workMatches.size()) b=workMatches[i]; for (size_t j=lastOrigOff; j<i; j++) {
// unlikely workMatches[j].done=true;
if (b.done) continue;
if (b.orig!=lastOrig) {
if (lastOrig!=SIZE_MAX) {
// commit previous block and start new one
logV("%x gains: %d",(int)lastOrig,gains);
if (gains<=0) {
// don't make a sub-block for these matches since we only have loss
logV("(LOSSES!)");
for (size_t j=lastOrigOff; j<i; j++) {
workMatches[j].done=true;
}
} }
} }
lastOrig=b.orig;
lastOrigOff=i;
if (lastOrig!=SIZE_MAX) {
blockSize=estimateBlockSize(&buf[b.orig],b.len,speedDial);
} else {
blockSize=0;
}
gains=-4;
} }
gains+=(blockSize-3); lastOrig=b.orig;
lastOrigOff=i;
if (lastOrig!=SIZE_MAX) {
blockSize=estimateBlockSize(&buf[b.orig],b.len,speedDial);
} else {
blockSize=0;
}
gains=-4;
} }
gains+=(blockSize-3);
}
// make sub-blocks // make sub-blocks
lastOrig=SIZE_MAX; lastOrig=SIZE_MAX;
size_t subBlockID=subBlocks.size(); size_t subBlockID=subBlocks.size();
for (BlockMatch& i: workMatches) { for (BlockMatch& i: workMatches) {
// skip invalid matches (yes, this can happen) // skip invalid matches (yes, this can happen)
if (i.done) continue; if (i.done) continue;
// create new sub-block if necessary // create new sub-block if necessary
if (i.orig!=lastOrig) { if (i.orig!=lastOrig) {
subBlockID=subBlocks.size(); subBlockID=subBlocks.size();
newBlocks=true; newBlocks=true;
logV("new sub-block %d",(int)subBlockID); logV("new sub-block %d",(int)subBlockID);
// isolate this sub-block // isolate this sub-block
SafeWriter* newBlock=new SafeWriter; SafeWriter* newBlock=new SafeWriter;
newBlock->init(); newBlock->init();
newBlock->write(&buf[i.orig],i.len); newBlock->write(&buf[i.orig],i.len);
newBlock->writeC(0xf9); // ret newBlock->writeC(0xf9); // ret
// padding // padding
newBlock->writeC(0); newBlock->writeC(0);
newBlock->writeC(0); newBlock->writeC(0);
newBlock->writeC(0); newBlock->writeC(0);
newBlock->writeC(0); newBlock->writeC(0);
newBlock->writeC(0); newBlock->writeC(0);
newBlock->writeC(0); newBlock->writeC(0);
newBlock->writeC(0); newBlock->writeC(0);
subBlocks.push_back(newBlock); subBlocks.push_back(newBlock);
lastOrig=i.orig; lastOrig=i.orig;
// insert call on the original block // insert call on the original block
buf[i.orig]=0xf4; buf[i.orig]=0xf4;
buf[i.orig+1]=subBlockID&0xff; buf[i.orig+1]=subBlockID&0xff;
buf[i.orig+2]=(subBlockID>>8)&0xff; buf[i.orig+2]=(subBlockID>>8)&0xff;
buf[i.orig+3]=(subBlockID>>16)&0xff; buf[i.orig+3]=(subBlockID>>16)&0xff;
buf[i.orig+4]=(subBlockID>>24)&0xff; buf[i.orig+4]=(subBlockID>>24)&0xff;
buf[i.orig+5]=0; buf[i.orig+5]=0;
buf[i.orig+6]=0; buf[i.orig+6]=0;
buf[i.orig+7]=0; buf[i.orig+7]=0;
// replace the rest with nop
for (size_t j=i.orig+8; j<i.orig+i.len; j++) {
buf[j]=0xf1;
}
}
// set match to the last sub-block
buf[i.block]=0xf4;
buf[i.block+1]=subBlockID&0xff;
buf[i.block+2]=(subBlockID>>8)&0xff;
buf[i.block+3]=(subBlockID>>16)&0xff;
buf[i.block+4]=(subBlockID>>24)&0xff;
buf[i.block+5]=0;
buf[i.block+6]=0;
buf[i.block+7]=0;
// replace the rest with nop // replace the rest with nop
for (size_t j=i.block+8; j<i.block+i.len; j++) { for (size_t j=i.orig+8; j<i.orig+i.len; j++) {
buf[j]=0xf1; buf[j]=0xf1;
} }
// invalidate overlapping work matches
for (BlockMatch& j: workMatches) {
if (j.orig!=i.orig) {
j.done=true;
}
if (OVERLAPS(i.block,i.block+i.len,j.block,j.block+j.len)) {
j.done=true;
}
}
// invalidate overlapping matches
for (BlockMatch& j: matches) {
if (OVERLAPS(i.orig,i.orig+i.len,j.orig,j.orig+j.len) ||
OVERLAPS(i.orig,i.orig+i.len,j.block,j.block+j.len) ||
OVERLAPS(i.block,i.block+i.len,j.orig,j.orig+j.len) ||
OVERLAPS(i.block,i.block+i.len,j.block,j.block+j.len)) {
j.done=true;
}
}
} }
//if (matchSize>=MIN_MATCH_SIZE) { // set match to the last sub-block
//logV("trying next size %d",matchSize); buf[i.block]=0xf4;
//} buf[i.block+1]=subBlockID&0xff;
//} buf[i.block+2]=(subBlockID>>8)&0xff;
buf[i.block+3]=(subBlockID>>16)&0xff;
buf[i.block+4]=(subBlockID>>24)&0xff;
buf[i.block+5]=0;
buf[i.block+6]=0;
buf[i.block+7]=0;
// replace the rest with nop
for (size_t j=i.block+8; j<i.block+i.len; j++) {
buf[j]=0xf1;
}
// invalidate overlapping work matches
for (BlockMatch& j: workMatches) {
if (j.orig!=i.orig) {
j.done=true;
}
if (OVERLAPS(i.block,i.block+i.len,j.block,j.block+j.len)) {
j.done=true;
}
}
// invalidate overlapping matches
for (BlockMatch& j: matches) {
if (OVERLAPS(i.orig,i.orig+i.len,j.orig,j.orig+j.len) ||
OVERLAPS(i.orig,i.orig+i.len,j.block,j.block+j.len) ||
OVERLAPS(i.block,i.block+i.len,j.orig,j.orig+j.len) ||
OVERLAPS(i.block,i.block+i.len,j.block,j.block+j.len)) {
j.done=true;
}
}
}
logV("done!"); logV("done!");