ShuriZma
/
pcsx2
mirror of https://github.com/PCSX2/pcsx2.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

mVU: Removed full flag optimisations 

They were unused, broken and cluttering up the code, so gotten rid.
This commit is contained in:
refractionpcsx2 2021-01-03 16:17:19 +00:00
parent c9bc6eac69
commit 0f7044a90f
6 changed files with 6 additions and 82 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
pcsx2/x86/microVU.h
View File

@ -76,8 +76,7 @@ public:
microBlock* add(microBlock* pBlock) {
microBlock* thisBlock = search(&pBlock->pState);
if (!thisBlock) {
u8 doFF = doFullFlagOpt && (pBlock->pState.flagInfo&1);
u8 fullCmp = pBlock->pState.needExactMatch || doFF;
u8 fullCmp = pBlock->pState.needExactMatch;
if (fullCmp) fListI++; else qListI++;
microBlockLink*& blockList = fullCmp ? fBlockList : qBlockList;
@ -100,8 +99,7 @@ public:
return thisBlock;
}
__ri microBlock* search(microRegInfo* pState) {
u8 doFF = doFullFlagOpt && (pState->flagInfo&1);
if (pState->needExactMatch || doFF) { // Needs Detailed Search (Exact Match of Pipeline State)
if (pState->needExactMatch) { // Needs Detailed Search (Exact Match of Pipeline State)
for(microBlockLink* linkI = fBlockList; linkI != NULL; linkI = linkI->next) {
if (mVUquickSearch((void*)pState, (void*)&linkI->block.pState, sizeof(microRegInfo)))
return &linkI->block;

2
pcsx2/x86/microVU_Analyze.inl
View File

@ -521,8 +521,6 @@ __ri int mVUbranchCheck(mV) {
mVUregs.needExactMatch |= 7; // This might not be necessary, but w/e...
mVUregs.flagInfo = 0;
mVUregs.fullFlags0 = 0;
mVUregs.fullFlags1 = 0;
DevCon.Warning("microVU%d: %s in %s delay slot! [%04x] - If game broken report to PCSX2 Team", mVU.index,
branchSTR[mVUlow.branch&0xf], branchSTR[branchType&0xf], xPC);
return 1;

2
pcsx2/x86/microVU_Compile.inl
View File

@ -444,8 +444,6 @@ __fi void mVUinitFirstPass(microVU& mVU, uptr pState, u8* thisPtr) {
mVUregs.blockType = 0;
mVUregs.viBackUp = 0;
mVUregs.flagInfo = 0;
mVUregs.fullFlags0 = 0;
mVUregs.fullFlags1 = 0;
mVUsFlagHack = CHECK_VU_FLAGHACK;
mVUinitConstValues(mVU);
}

68
pcsx2/x86/microVU_Flags.inl
View File

@ -124,53 +124,28 @@ __fi void mVUsetFlags(mV, microFlagCycles& mFC) {
// Status/Mac Flags Setup Code
int xS = 0, xM = 0, xC = 0;
u32 ff0=0, ff1=0, ffOn=0, fInfo=0;
if (doFullFlagOpt) {
ff0 = mVUpBlock->pState.fullFlags0;
ff1 = mVUpBlock->pState.fullFlags1;
ffOn = mVUpBlock->pState.flagInfo&1;
fInfo = mVUpBlock->pState.flagInfo;
}
for(int i = 0; i < 4; i++) {
mFC.xStatus[i] = i;
mFC.xMac [i] = i;
mFC.xClip [i] = i;
}
if (ffOn) { // Full Flags Enabled
xS = (fInfo >> 2) & 3;
xM = (fInfo >> 4) & 3;
xC = (fInfo >> 6) & 3;
mFC.xStatus[0] = ((ff0 >> (3*0+ 0)) & 7) - 1;
mFC.xStatus[1] = ((ff0 >> (3*1+ 0)) & 7) - 1;
mFC.xStatus[2] = ((ff0 >> (3*2+ 0)) & 7) - 1;
mFC.xStatus[3] = ((ff0 >> (3*3+ 0)) & 7) - 1;
mFC.xMac [0] = ((ff0 >> (3*0+12)) & 7) - 1;
mFC.xMac [1] = ((ff0 >> (3*1+12)) & 7) - 1;
mFC.xMac [2] = ((ff0 >> (3*2+12)) & 7) - 1;
mFC.xMac [3] = ((ff0 >> (3*3+12)) & 7) - 1;
mFC.xClip [0] = ((ff0 >> (3*0+24)) & 7) - 1;
mFC.xClip [1] = ((ff0 >> (3*1+24)) & 7) - 1;
mFC.xClip [2] = ((ff1 >> (3*0+ 0)) & 7) - 1;
mFC.xClip [3] = ((ff1 >> (3*1+ 0)) & 7) - 1;
}
if(!ffOn && !(mVUpBlock->pState.needExactMatch & 1)) {
if(!(mVUpBlock->pState.needExactMatch & 1)) {
xS = (mVUpBlock->pState.flagInfo >> 2) & 3;
mFC.xStatus[0] = -1; mFC.xStatus[1] = -1;
mFC.xStatus[2] = -1; mFC.xStatus[3] = -1;
mFC.xStatus[(xS-1)&3] = 0;
}
if(!ffOn && !(mVUpBlock->pState.needExactMatch & 2)) {
if(!(mVUpBlock->pState.needExactMatch & 2)) {
//xM = (mVUpBlock->pState.flagInfo >> 4) & 3;
mFC.xMac[0] = -1; mFC.xMac[1] = -1;
mFC.xMac[2] = -1; mFC.xMac[3] = -1;
//mFC.xMac[(xM-1)&3] = 0;
}
if(!ffOn && !(mVUpBlock->pState.needExactMatch & 4)) {
if(!(mVUpBlock->pState.needExactMatch & 4)) {
xC = (mVUpBlock->pState.flagInfo >> 6) & 3;
mFC.xClip[0] = -1; mFC.xClip[1] = -1;
mFC.xClip[2] = -1; mFC.xClip[3] = -1;
@ -235,22 +210,6 @@ __fi void mVUsetFlags(mV, microFlagCycles& mFC) {
mVUregs.flagInfo |= ((__Mac||1) ? 0 : (xM << 4));
mVUregs.flagInfo |= ((__Clip) ? 0 : (xC << 6));
iPC = endPC;
if (doFullFlagOpt && (mVUregs.flagInfo & 1)) {
//if (mVUregs.needExactMatch) DevCon.Error("mVU ERROR!!!");
int bS[4], bM[4], bC[4];
sortFullFlag(mFC.xStatus, bS);
sortFullFlag(mFC.xMac, bM);
sortFullFlag(mFC.xClip, bC);
mVUregs.flagInfo = (xC<<6) | (xM<<4) | (xS<<2) | 1;
mVUregs.fullFlags0 = ((bS[3]<<9)|(bS[2]<<6)|(bS[1]<<3)|(bS[0]<<0)) << (12*0);
mVUregs.fullFlags0 |= ((bM[3]<<9)|(bM[2]<<6)|(bM[1]<<3)|(bM[0]<<0)) << (12*1);
mVUregs.fullFlags0 |= ((bC[1]<<3)|(bC[0]<<0)) << (12*2);
mVUregs.fullFlags1 = ((bC[3]<<3)|(bC[2]<<0)) << (12*0);
mVUregs.needExactMatch = 0;
DevCon.WriteLn("MVU FULL FLAG!!!!!!!! [0x%04x][0x%08x][0x%02x]",
xPC, mVUregs.fullFlags0, (u32)mVUregs.fullFlags1);
}
}
#define getFlagReg2(x) ((bStatus[0] == x) ? getFlagReg(x) : gprT1)
@ -386,51 +345,32 @@ void mVUflagPass(mV, u32 startPC, u32 sCount = 0, u32 found = 0) {
_mVUflagPass(mVU, startPC, sCount, found, v);
}
__fi void checkFFblock(mV, u32 addr, int& ffOpt) {
if (ffOpt && doFullFlagOpt) {
blockCreate(addr/8);
ffOpt = mVUblocks[addr/8]->getFullListCount() <= doFullFlagOpt;
}
}
// Checks if the first ~4 instructions of a block will read flags
void mVUsetFlagInfo(mV) {
if (noFlagOpts) {
mVUregs.needExactMatch = 0x7;
mVUregs.fullFlags0 = 0x0;
mVUregs.fullFlags1 = 0x0;
mVUregs.flagInfo = 0x0;
return;
}
int ffOpt = doFullFlagOpt;
if (mVUbranch <= 2) { // B/BAL
incPC(-1);
mVUflagPass (mVU, branchAddr(mVU));
checkFFblock(mVU, branchAddr(mVU), ffOpt);
incPC(1);
mVUregs.needExactMatch &= 0x7;
if (mVUregs.needExactMatch && ffOpt) {
mVUregs.flagInfo |= 1;
}
}
else if (mVUbranch <= 8) { // Conditional Branch
incPC(-1); // Branch Taken
mVUflagPass (mVU, branchAddr(mVU));
checkFFblock(mVU, branchAddr(mVU), ffOpt);
int backupFlagInfo = mVUregs.needExactMatch;
mVUregs.needExactMatch = 0;
incPC(4); // Branch Not Taken
mVUflagPass (mVU, xPC);
checkFFblock(mVU, xPC, ffOpt);
incPC(-3);
mVUregs.needExactMatch |= backupFlagInfo;
mVUregs.needExactMatch &= 0x7;
if (mVUregs.needExactMatch && ffOpt) {
mVUregs.flagInfo |= 1;
}
}
else { // JR/JALR
if (!doConstProp || !mVUlow.constJump.isValid) { mVUregs.needExactMatch |= 0x7; }

2
pcsx2/x86/microVU_IR.h
View File

@ -47,8 +47,6 @@ union __aligned16 microRegInfo {
u32 quick32[2];
};
u32 fullFlags0; // clip * 6 | mac * 12 | status * 12
u8 fullFlags1; // clip * 6
u8 vi15v; // 'vi15' constant is valid
u16 vi15; // Constant Prop Info for vi15

8
pcsx2/x86/microVU_Misc.h
View File

@ -303,14 +303,6 @@ static const bool doCFlagInsts = true; // Set to true to enable multiple clip
// Setting one of these to 0 acts as if there is only 1 instance of the
// corresponding flag, which may be useful when debugging flag pipeline bugs.
static const int doFullFlagOpt = false; // Set above to false to enable full flag optimization
// This attempts to eliminate some flag shuffling at the end of blocks, but
// can end up creating more recompiled code. The max amount of times this optimization
// is performed per block can be set by changing the doFullFlagOpt value to be that limit.
// i.e. setting doFullFlagOpt to 2 will recompile the current block at-most 2 times with
// the full flag optimization.
// Note: This optimization doesn't really seem to be benefitial and is buggy...
// Branch in Branch Delay Slots
static const bool doBranchInDelaySlot = true; // Set to true to enable evil-branches
// This attempts to emulate the correct behavior for branches in branch delay