frederik
/
oscar64
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Optimize sprite multiplexer

This commit is contained in:
drmortalwombat 2024-08-08 22:07:44 +02:00
parent 6f069946ba
commit d98c4b760f
4 changed files with 93 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
include/c64/sprites.c
View File

@ -270,7 +270,7 @@ void vspr_update(void)
{ {
char xymask = 0; char xymask = 0;
volatile char * vsprs = vspriteScreen; volatile char * vsprs = vspriteScreen;
char sypos[VSPRITES_MAX]; // char sypos[VSPRITES_MAX];
#pragma unroll(full) #pragma unroll(full)
@ -291,9 +291,9 @@ void vspr_update(void)
xymask = ((unsigned)xymask | (vspriteXHigh[ri] << 8)) >> 1; xymask = ((unsigned)xymask | (vspriteXHigh[ri] << 8)) >> 1;
#endif #endif
vic.spr_pos[uj].x = vspriteXLow[ri]; vic.spr_pos[uj].x = vspriteXLow[ri];
vic.spr_pos[uj].y = vspriteYLow[ri]; vic.spr_pos[uj].y = spriteYPos[ui + 1];
sypos[ui] = vspriteYLow[ri]; // sypos[ui] = vspriteYLow[ri];
} }
vic.spr_msbx = xymask; vic.spr_msbx = xymask;
@ -303,9 +303,9 @@ void vspr_update(void)
for(char ti=0; ti<VSPRITES_MAX - 8; ti++) for(char ti=0; ti<VSPRITES_MAX - 8; ti++)
{ {
byte ri = spriteOrder[ti + 8]; if (!done && spriteYPos[ti + 9] < 250)
if (!done && vspriteYLow[ri] < 250)
{ {
byte ri = spriteOrder[ti + 8];
#ifdef VSPRITE_REVERSE #ifdef VSPRITE_REVERSE
char m = 0x80 >> (ti & 7); char m = 0x80 >> (ti & 7);
#else #else
@ -316,14 +316,14 @@ void vspr_update(void)
if (!(vspriteXHigh[ri] & 1)) if (!(vspriteXHigh[ri] & 1))
xymask ^= m; xymask ^= m;
rirq_data(spirq + ti, 2, vspriteYLow[ri]); rirq_data(spirq + ti, 2, spriteYPos[ti + 9]);
rirq_data(spirq + ti, 0, vspriteColor[ri]); rirq_data(spirq + ti, 0, vspriteColor[ri]);
rirq_data(spirq + ti, 1, vspriteXLow[ri]); rirq_data(spirq + ti, 1, vspriteXLow[ri]);
rirq_data(spirq + ti, 3, vspriteImage[ri]); rirq_data(spirq + ti, 3, vspriteImage[ri]);
rirq_data(spirq + ti, 4, xymask); rirq_data(spirq + ti, 4, xymask);
rirq_move(ti, sypos[ti] + 23); rirq_move(ti, spriteYPos[ti + 1] + 23);
sypos[ti + 8] = vspriteYLow[ri]; // spriteYPos[ti + 9] = vspriteYLow[ri];
} }
else else
{ {

5
oscar64/Constexpr.cpp
View File

@ -518,7 +518,10 @@ Declaration* ConstexprInterpreter::Value::GetConst(int offset, Declaration* type
for (int i=0; i<type->mSize; i += type->mBase->mSize) for (int i=0; i<type->mSize; i += type->mBase->mSize)
{ {
Declaration* cdec = GetConst(offset + i, type->mBase, dataSection); Declaration* cdec = GetConst(offset + i, type->mBase, dataSection);
cdec->mOffset = i; if (type->mStride)
cdec->mOffset = i / type->mBase->mSize;
else
cdec->mOffset = i;
if (ldec) if (ldec)
ldec->mNext = cdec; ldec->mNext = cdec;

83
oscar64/NativeCodeGenerator.cpp
View File

@ -6407,6 +6407,34 @@ void NativeCodeBasicBlock::StoreByteOffsetIndexedValue(InterCodeProcedure* proc,
} }
} }
void NativeCodeBasicBlock::LoadStoreIndirectPair(InterCodeProcedure* proc, const InterInstruction* wins0, const InterInstruction* wins1)
{
int sreg = BC_REG_TMP + proc->mTempOffset[wins0->mSrc[0].mTemp];
int ireg0 = BC_REG_TMP + proc->mTempOffset[wins0->mSrc[1].mTemp];
int index0 = int(wins0->mSrc[1].mIntConst);
uint32 wflags0 = NCIF_LOWER | NCIF_UPPER;
if (wins0->mVolatile)
wflags0 |= NCIF_VOLATILE;
int ireg1 = BC_REG_TMP + proc->mTempOffset[wins1->mSrc[1].mTemp];
int index1 = int(wins1->mSrc[1].mIntConst);
uint32 wflags1 = NCIF_LOWER | NCIF_UPPER;
if (wins1->mVolatile)
wflags1 |= NCIF_VOLATILE;
CheckFrameIndex(wins0, ireg0, index0, 1, BC_REG_ADDR);
CheckFrameIndex(wins1, ireg1, index1, 1, BC_REG_ACCU);
mIns.Push(NativeCodeInstruction(wins0, ASMIT_LDA, ASMIM_ZERO_PAGE, sreg));
mIns.Push(NativeCodeInstruction(wins0, ASMIT_LDY, ASMIM_IMMEDIATE, index0));
mIns.Push(NativeCodeInstruction(wins0, ASMIT_STA, ASMIM_INDIRECT_Y, ireg0, nullptr, wflags0));
mIns.Push(NativeCodeInstruction(wins1, ASMIT_LDY, ASMIM_IMMEDIATE, index1));
mIns.Push(NativeCodeInstruction(wins1, ASMIT_STA, ASMIM_INDIRECT_Y, ireg1, nullptr, wflags1));
}
void NativeCodeBasicBlock::LoadStoreIndirectValue(InterCodeProcedure* proc, const InterInstruction* rins, const InterInstruction* wins) void NativeCodeBasicBlock::LoadStoreIndirectValue(InterCodeProcedure* proc, const InterInstruction* rins, const InterInstruction* wins)
{ {
int size = InterTypeSize[wins->mSrc[0].mType]; int size = InterTypeSize[wins->mSrc[0].mType];
@ -33560,7 +33588,7 @@ bool NativeCodeBasicBlock::MoveLDSTXOutOfRange(int at)
else if (mIns[j].ReferencesXReg()) else if (mIns[j].ReferencesXReg())
return false; return false;
if (mIns[j].mType == ASMIT_JSR) if (mIns[j].mType == ASMIT_JSR || mIns[j].mType == ASMIT_RTS)
return false; return false;
if (!(mIns[j].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_Z))) if (!(mIns[j].mLive & (LIVE_CPU_REG_A | LIVE_CPU_REG_Z)))
@ -47969,6 +47997,39 @@ bool NativeCodeBasicBlock::PeepHoleOptimizerIterateN(int i, int pass)
#endif #endif
} }
#if 1
if (i + 7 < mIns.Size() &&
mIns[i + 0].mType == ASMIT_STA && mIns[i + 0].mMode == ASMIM_ZERO_PAGE &&
mIns[i + 1].mType == ASMIT_CLC &&
mIns[i + 2].mType == ASMIT_ADC && mIns[i + 2].mMode == ASMIM_ZERO_PAGE &&
mIns[i + 3].mType == ASMIT_STA && mIns[i + 3].mMode == ASMIM_ZERO_PAGE && mIns[i + 2].mAddress == mIns[i + 3].mAddress &&
mIns[i + 4].mType == ASMIT_LDA &&
mIns[i + 5].mType == ASMIT_STA && mIns[i + 5].mMode == ASMIM_ZERO_PAGE && mIns[i + 5].mAddress != mIns[i + 0].mAddress &&
mIns[i + 6].mType == ASMIT_ADC && mIns[i + 6].mMode == ASMIM_ZERO_PAGE && mIns[i + 6].mAddress == mIns[i + 2].mAddress + 1 &&
mIns[i + 7].mType == ASMIT_STA && mIns[i + 7].mMode == ASMIM_ZERO_PAGE && mIns[i + 6].mAddress == mIns[i + 7].mAddress &&
!(mIns[i + 7].mLive & LIVE_CPU_REG_A))
{
int yval = RetrieveYValue(i);
mProc->ResetPatched();
if (CheckForwardSumYPointer(this, mIns[i + 2].mAddress, mIns[i + 2].mAddress, mIns[i + 0], i + 8, yval, 0))
{
mProc->ResetPatched();
if (PatchForwardSumYPointer(this, mIns[i + 2].mAddress, mIns[i + 2].mAddress, mIns[i + 0], i + 8, yval))
{
mIns[i + 2].mType = ASMIT_NOP; mIns[i + 2].mMode = ASMIM_IMPLIED;
mIns[i + 3].mType = ASMIT_NOP; mIns[i + 3].mMode = ASMIM_IMPLIED;
if (mTrueJump)
mTrueJump->CheckLive();
if (mFalseJump)
mFalseJump->CheckLive();
return true;
}
}
}
#endif
#if 1 #if 1
if (pass == 0 && if (pass == 0 &&
i + 7 < mIns.Size() && i + 7 < mIns.Size() &&
@ -50379,7 +50440,7 @@ void NativeCodeProcedure::Compile(InterCodeProcedure* proc)
mInterProc = proc; mInterProc = proc;
mInterProc->mLinkerObject->mNativeProc = this; mInterProc->mLinkerObject->mNativeProc = this;
CheckFunc = !strcmp(mInterProc->mIdent->mString, "screen_flip"); CheckFunc = !strcmp(mInterProc->mIdent->mString, "shot_add");
int nblocks = proc->mBlocks.Size(); int nblocks = proc->mBlocks.Size();
tblocks = new NativeCodeBasicBlock * [nblocks]; tblocks = new NativeCodeBasicBlock * [nblocks];
@ -51238,6 +51299,7 @@ void NativeCodeProcedure::Optimize(void)
} while (changed && t < 20); } while (changed && t < 20);
#endif #endif
BuildDataFlowSets(); BuildDataFlowSets();
ResetVisited(); ResetVisited();
mEntryBlock->RemoveUnusedResultInstructions(); mEntryBlock->RemoveUnusedResultInstructions();
@ -51283,6 +51345,7 @@ void NativeCodeProcedure::Optimize(void)
#endif #endif
if (step == 2) if (step == 2)
{ {
ResetVisited(); ResetVisited();
@ -52334,7 +52397,21 @@ void NativeCodeProcedure::CompileInterBlock(InterCodeProcedure* iproc, InterCode
switch (ins->mCode) switch (ins->mCode)
{ {
case IC_STORE: case IC_STORE:
block->StoreValue(iproc, ins); if (i + 1 < iblock->mInstructions.Size() &&
ins->mSrc[0].mType == IT_INT8 &&
ins->mSrc[0].mTemp >= 0 &&
ins->mSrc[1].mTemp >= 0 &&
iblock->mInstructions[i + 1]->mCode == IC_STORE &&
iblock->mInstructions[i + 1]->mSrc[0].mType == IT_INT8 &&
iblock->mInstructions[i + 1]->mSrc[0].mTemp == ins->mSrc[0].mTemp &&
iblock->mInstructions[i + 1]->mSrc[0].mFinal &&
iblock->mInstructions[i + 1]->mSrc[1].mTemp >= 0)
{
block->LoadStoreIndirectPair(iproc, ins, iblock->mInstructions[i + 1]);
i++;
}
else
block->StoreValue(iproc, ins);
break; break;
case IC_LOAD: case IC_LOAD:
if (i + 1 < iblock->mInstructions.Size() && if (i + 1 < iblock->mInstructions.Size() &&

1
oscar64/NativeCodeGenerator.h
View File

@ -358,6 +358,7 @@ public:
void StoreValue(InterCodeProcedure* proc, const InterInstruction * ins); void StoreValue(InterCodeProcedure* proc, const InterInstruction * ins);
void LoadValue(InterCodeProcedure* proc, const InterInstruction * ins); void LoadValue(InterCodeProcedure* proc, const InterInstruction * ins);
void LoadStoreValue(InterCodeProcedure* proc, const InterInstruction * rins, const InterInstruction * wins); void LoadStoreValue(InterCodeProcedure* proc, const InterInstruction * rins, const InterInstruction * wins);
void LoadStoreIndirectPair(InterCodeProcedure* proc, const InterInstruction* wins0, const InterInstruction* wins1);
bool LoadOpStoreIndirectValue(InterCodeProcedure* proc, const InterInstruction* rins, const InterInstruction* oins, int oindex, const InterInstruction* wins); bool LoadOpStoreIndirectValue(InterCodeProcedure* proc, const InterInstruction* rins, const InterInstruction* oins, int oindex, const InterInstruction* wins);
bool LoadUnopStoreIndirectValue(InterCodeProcedure* proc, const InterInstruction* rins, const InterInstruction* oins, const InterInstruction* wins); bool LoadUnopStoreIndirectValue(InterCodeProcedure* proc, const InterInstruction* rins, const InterInstruction* oins, const InterInstruction* wins);
bool LoadLoadOpStoreIndirectValue(InterCodeProcedure* proc, const InterInstruction* rins1, const InterInstruction* rins0, const InterInstruction* oins, const InterInstruction* wins); bool LoadLoadOpStoreIndirectValue(InterCodeProcedure* proc, const InterInstruction* rins1, const InterInstruction* rins0, const InterInstruction* oins, const InterInstruction* wins);