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

Reduce generated code size with -Os

This commit is contained in:
drmortalwombat 2024-09-27 20:01:22 +02:00 committed by John Schneiderman
parent 8eebad4d8a
commit 25cf2bf80b
12 changed files with 339 additions and 78 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
include/crt.c
View File

@ -2858,6 +2858,11 @@ W1:
__asm faddsub __asm faddsub
{ {
fsub:
lda tmp + 3
eor #$80
sta tmp + 3
fadd:
lda #$ff lda #$ff
cmp tmp + 4 cmp tmp + 4
beq INF beq INF
@ -3008,7 +3013,7 @@ fas_zero:
__asm inp_binop_add_f32 __asm inp_binop_add_f32
{ {
jsr freg.split_exp jsr freg.split_exp
jsr faddsub jsr faddsub.fadd
jmp startup.exec jmp startup.exec
} }
@ -3017,10 +3022,7 @@ __asm inp_binop_add_f32
__asm inp_binop_sub_f32 __asm inp_binop_sub_f32
{ {
jsr freg.split_exp jsr freg.split_exp
lda tmp + 3 jsr faddsub.fsub
eor #$80
sta tmp + 3
jsr faddsub
jmp startup.exec jmp startup.exec
} }
@ -3983,11 +3985,38 @@ fru3:
jmp freg.merge_aexp jmp freg.merge_aexp
} }
__asm store32
{
lda accu + 0
sta $00, x
lda accu + 1
sta $01, x
lda accu + 2
sta $02, x
lda accu + 3
sta $03, x
rts
}
__asm load32
{
lda $00, x
sta accu + 0
lda $01, x
sta accu + 1
lda $02, x
sta accu + 2
lda $03, x
sta accu + 3
rts
}
#pragma runtime(fsplita, freg.split_aexp) #pragma runtime(fsplita, freg.split_aexp)
#pragma runtime(fsplitt, freg.split_texp) #pragma runtime(fsplitt, freg.split_texp)
#pragma runtime(fsplitx, freg.split_xexp) #pragma runtime(fsplitx, freg.split_xexp)
#pragma runtime(fmergea, freg.merge_aexp) #pragma runtime(fmergea, freg.merge_aexp)
#pragma runtime(faddsub, faddsub) #pragma runtime(fadd, faddsub.fadd)
#pragma runtime(fsub, faddsub.fsub)
#pragma runtime(fmul, fmul) #pragma runtime(fmul, fmul)
#pragma runtime(fdiv, fdiv) #pragma runtime(fdiv, fdiv)
#pragma runtime(fcmp, fcmp) #pragma runtime(fcmp, fcmp)
@ -4001,6 +4030,8 @@ fru3:
#pragma runtime(ffromlu, uint32_to_float) #pragma runtime(ffromlu, uint32_to_float)
#pragma runtime(ftoli, f32_to_i32) #pragma runtime(ftoli, f32_to_i32)
#pragma runtime(ftolu, f32_to_u32) #pragma runtime(ftolu, f32_to_u32)
#pragma runtime(store32, store32)
#pragma runtime(load32, load32)
__asm inp_op_floor_f32 __asm inp_op_floor_f32
{ {

6
oscar64/Compiler.cpp
View File

@ -943,7 +943,8 @@ bool Compiler::GenerateCode(void)
RegisterRuntime(loc, Ident::Unique("fsplitt")); RegisterRuntime(loc, Ident::Unique("fsplitt"));
RegisterRuntime(loc, Ident::Unique("fsplitx")); RegisterRuntime(loc, Ident::Unique("fsplitx"));
RegisterRuntime(loc, Ident::Unique("fsplita")); RegisterRuntime(loc, Ident::Unique("fsplita"));
RegisterRuntime(loc, Ident::Unique("faddsub")); RegisterRuntime(loc, Ident::Unique("fadd"));
RegisterRuntime(loc, Ident::Unique("fsub"));
RegisterRuntime(loc, Ident::Unique("fmul")); RegisterRuntime(loc, Ident::Unique("fmul"));
RegisterRuntime(loc, Ident::Unique("fdiv")); RegisterRuntime(loc, Ident::Unique("fdiv"));
RegisterRuntime(loc, Ident::Unique("mul16")); RegisterRuntime(loc, Ident::Unique("mul16"));
@ -971,6 +972,9 @@ bool Compiler::GenerateCode(void)
RegisterRuntime(loc, Ident::Unique("divu32")); RegisterRuntime(loc, Ident::Unique("divu32"));
RegisterRuntime(loc, Ident::Unique("modu32")); RegisterRuntime(loc, Ident::Unique("modu32"));
RegisterRuntime(loc, Ident::Unique("store32"));
RegisterRuntime(loc, Ident::Unique("load32"));
RegisterRuntime(loc, Ident::Unique("malloc")); RegisterRuntime(loc, Ident::Unique("malloc"));
RegisterRuntime(loc, Ident::Unique("free")); RegisterRuntime(loc, Ident::Unique("free"));
RegisterRuntime(loc, Ident::Unique("breakpoint")); RegisterRuntime(loc, Ident::Unique("breakpoint"));

2
oscar64/CompilerTypes.h
View File

@ -47,7 +47,7 @@ static const uint64 COPT_DEFAULT = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE |
static const uint64 COPT_OPTIMIZE_DEFAULT = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS; static const uint64 COPT_OPTIMIZE_DEFAULT = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS;
static const uint64 COPT_OPTIMIZE_SIZE = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_CODE_SIZE | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL; static const uint64 COPT_OPTIMIZE_SIZE = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_AUTO_INLINE | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_CODE_SIZE | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL;
static const uint64 COPT_OPTIMIZE_SPEED = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_AUTO_INLINE | COPT_OPTIMIZE_AUTO_UNROLL | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_ASSEMBLER | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL; static const uint64 COPT_OPTIMIZE_SPEED = COPT_OPTIMIZE_BASIC | COPT_OPTIMIZE_INLINE | COPT_OPTIMIZE_AUTO_INLINE | COPT_OPTIMIZE_AUTO_UNROLL | COPT_OPTIMIZE_CONST_EXPRESSIONS | COPT_OPTIMIZE_ASSEMBLER | COPT_OPTIMIZE_CONST_PARAMS | COPT_OPTIMIZE_MERGE_CALLS | COPT_OPTIMIZE_GLOBAL;

1
oscar64/Errors.h
View File

@ -42,6 +42,7 @@ enum ErrorID
EWARN_UNDEFINED_POINTER_ARITHMETIC, EWARN_UNDEFINED_POINTER_ARITHMETIC,
EWARN_INVALID_VALUE_RANGE, EWARN_INVALID_VALUE_RANGE,
EWARN_DEFAULT_COPY_DEPRECATED, EWARN_DEFAULT_COPY_DEPRECATED,
EWARN_INSUFFICIENT_MEMORY,
EERR_GENERIC = 3000, EERR_GENERIC = 3000,
EERR_FILE_NOT_FOUND, EERR_FILE_NOT_FOUND,

37
oscar64/GlobalAnalyzer.cpp
View File

@ -147,6 +147,25 @@ void GlobalAnalyzer::TopoSort(Declaration* procDec)
} }
} }
int GlobalAnalyzer::CallerInvokes(Declaration* called)
{
int n = 0;
for (int i = 0; i < called->mCallers.Size(); i++)
{
Declaration* f = called->mCallers[i];
n += CallerInvokes(f, called);
}
return n;
}
int GlobalAnalyzer::CallerInvokes(Declaration* caller, Declaration* called)
{
int n = 1;
if (caller->mType == DT_CONST_FUNCTION && (caller->mFlags & (DTF_INLINE | DTF_REQUEST_INLINE)) && !(caller->mFlags & DTF_PREVENT_INLINE) && !(caller->mFlags & DTF_FUNC_RECURSIVE) && !(caller->mFlags & DTF_FUNC_VARIABLE) && !(caller->mFlags & DTF_EXPORT))
n = CallerInvokes(caller);
return n > 1 ? n : 1;
}
void GlobalAnalyzer::AutoInline(void) void GlobalAnalyzer::AutoInline(void)
{ {
for (int i = 0; i < mFunctions.Size(); i++) for (int i = 0; i < mFunctions.Size(); i++)
@ -181,33 +200,39 @@ void GlobalAnalyzer::AutoInline(void)
dec = dec->mNext; dec = dec->mNext;
} }
int invokes = CallerInvokes(f);
int cost = (f->mComplexity - 20 * nparams - 10); int cost = (f->mComplexity - 20 * nparams - 10);
// printf("CHECK INLINING %s (%d) %d * (%d - 1)\n", f->mIdent->mString, f->mComplexity, cost, f->mCallers.Size()); // printf("CHECK INLINING %s (%d) %d * (%d - 1)\n", f->mIdent->mString, f->mComplexity, cost, invokes);
bool doinline = false; bool doinline = false;
if ((f->mCompilerOptions & COPT_OPTIMIZE_INLINE) && (f->mFlags & DTF_REQUEST_INLINE)) if ((f->mCompilerOptions & COPT_OPTIMIZE_INLINE) && (f->mFlags & DTF_REQUEST_INLINE))
doinline = true; doinline = true;
if (f->mLocalSize < 100) if (f->mLocalSize < 100)
{ {
if ((f->mCompilerOptions & COPT_OPTIMIZE_AUTO_INLINE) && ((cost - 20) * (f->mCallers.Size() - 1) <= 20)) if ((f->mCompilerOptions & COPT_OPTIMIZE_AUTO_INLINE) && ((cost - 20) * (invokes - 1) <= 20))
{ {
if (f->mCallers.Size() == 1 && f->mComplexity > 100) if (f->mCompilerOptions & COPT_OPTIMIZE_CODE_SIZE)
{
if (invokes == 1 && f->mSection == f->mCallers[0]->mSection || cost < 0)
doinline = true;
}
else if (invokes == 1 && f->mComplexity > 100)
{ {
// printf("CHECK INLINING2 %s <- %s %d\n", f->mIdent->mString, f->mCallers[0]->mIdent->mString, f->mCallers[0]->mCalled.Size()); // printf("CHECK INLINING2 %s <- %s %d\n", f->mIdent->mString, f->mCallers[0]->mIdent->mString, f->mCallers[0]->mCalled.Size());
if (cost < 0 || f->mCallers[0]->mComplexity + cost < 1000 || f->mCallers[0]->mCalled.Size() == 1) if (cost < 0 || f->mCallers[0]->mComplexity + cost < 1000 || f->mCallers[0]->mCalled.Size() == 1)
doinline = true; doinline = true;
} }
else else
doinline = true; doinline = true;
} }
if ((f->mCompilerOptions & COPT_OPTIMIZE_AUTO_INLINE_ALL) && (cost * (f->mCallers.Size() - 1) <= 10000)) if ((f->mCompilerOptions & COPT_OPTIMIZE_AUTO_INLINE_ALL) && (cost * (invokes - 1) <= 10000))
doinline = true; doinline = true;
} }
if (doinline) if (doinline)
{ {
// printf("INLINING %s %d * (%d - 1)\n", f->mIdent->mString, cost, f->mCallers.Size()); // printf("INLINING %s %d * (%d - 1)\n", f->mIdent->mString, cost, invokes);
f->mFlags |= DTF_INLINE; f->mFlags |= DTF_INLINE;
for (int j = 0; j < f->mCallers.Size(); j++) for (int j = 0; j < f->mCallers.Size(); j++)

2
oscar64/GlobalAnalyzer.h
View File

@ -32,6 +32,8 @@ protected:
GrowingArray<Declaration*> mGlobalVariables; GrowingArray<Declaration*> mGlobalVariables;
void AnalyzeInit(Declaration* mdec); void AnalyzeInit(Declaration* mdec);
int CallerInvokes(Declaration* called);
int CallerInvokes(Declaration* caller, Declaration* called);
Declaration* Analyze(Expression* exp, Declaration* procDec, bool lhs, bool aliasing); Declaration* Analyze(Expression* exp, Declaration* procDec, bool lhs, bool aliasing);

83
oscar64/InterCode.cpp
View File

@ -23479,7 +23479,7 @@ void InterCodeProcedure::Close(void)
DisassembleDebug("mapped variabled"); DisassembleDebug("mapped variabled");
ReduceTemporaries(); ReduceTemporaries(true);
DisassembleDebug("Reduced Temporaries"); DisassembleDebug("Reduced Temporaries");
@ -24224,12 +24224,20 @@ void InterCodeProcedure::HoistCommonConditionalPath(void)
} }
void InterCodeProcedure::ReduceTemporaries(void) void InterCodeProcedure::ReduceTemporaries(bool final)
{ {
NumberSet* collisionSet; NumberSet* collisionSet;
int i, j, numRenamedTemps; int i, j, numRenamedTemps;
int numTemps = mTemporaries.Size(); int numTemps = mTemporaries.Size();
NumberSet callerSaved(numTemps);
if (final)
{
ResetVisited();
mEntryBlock->BuildCallerSaveTempSet(callerSaved);
}
ResetVisited(); ResetVisited();
mEntryBlock->BuildLocalTempSets(numTemps); mEntryBlock->BuildLocalTempSets(numTemps);
@ -24261,24 +24269,73 @@ void InterCodeProcedure::ReduceTemporaries(void)
NumberSet usedTemps(numTemps); NumberSet usedTemps(numTemps);
for (i = 0; i < numTemps; i++) if (final)
{ {
usedTemps.Clear(); for (int sz = 4; sz > 0; sz >>= 1)
for (j = 0; j < numTemps; j++)
{ {
if (mRenameTable[j] >= 0 && (collisionSet[i][j] || InterTypeSize[mTemporaries[j]] != InterTypeSize[mTemporaries[i]])) for (i = 0; i < numTemps; i++)
{ {
usedTemps += mRenameTable[j]; if (InterTypeSize[mTemporaries[i]] == sz && !callerSaved[i])
{
usedTemps.Clear();
for (j = 0; j < numTemps; j++)
{
if (mRenameTable[j] >= 0 && collisionSet[i][j])
usedTemps += mRenameTable[j];
}
j = 0;
while (usedTemps[j])
j++;
mRenameTable[i] = j;
if (j >= numRenamedTemps) numRenamedTemps = j + 1;
}
}
for (i = 0; i < numTemps; i++)
{
if (InterTypeSize[mTemporaries[i]] == sz && callerSaved[i])
{
usedTemps.Clear();
for (j = 0; j < numTemps; j++)
{
if (mRenameTable[j] >= 0 && collisionSet[i][j])
usedTemps += mRenameTable[j];
}
j = 0;
while (usedTemps[j])
j++;
mRenameTable[i] = j;
if (j >= numRenamedTemps) numRenamedTemps = j + 1;
}
} }
} }
}
else
{
for (i = 0; i < numTemps; i++)
{
usedTemps.Clear();
j = 0; for (j = 0; j < numTemps; j++)
while (usedTemps[j]) {
j++; if (mRenameTable[j] >= 0 && (collisionSet[i][j] || InterTypeSize[mTemporaries[j]] != InterTypeSize[mTemporaries[i]]))
{
usedTemps += mRenameTable[j];
}
}
mRenameTable[i] = j; j = 0;
if (j >= numRenamedTemps) numRenamedTemps = j + 1; while (usedTemps[j])
j++;
mRenameTable[i] = j;
if (j >= numRenamedTemps) numRenamedTemps = j + 1;
}
} }
mTemporaries.SetSize(numRenamedTemps, true); mTemporaries.SetSize(numRenamedTemps, true);

2
oscar64/InterCode.h
View File

@ -739,7 +739,7 @@ public:
bool ModifiesGlobal(int varindex); bool ModifiesGlobal(int varindex);
void MapVariables(void); void MapVariables(void);
void ReduceTemporaries(void); void ReduceTemporaries(bool final = false);
void Disassemble(FILE* file); void Disassemble(FILE* file);
void Disassemble(const char* name, bool dumpSets = false); void Disassemble(const char* name, bool dumpSets = false);
protected: protected:

8
oscar64/Linker.cpp
View File

@ -772,6 +772,8 @@ void LinkerRegion::PlaceStackSection(LinkerSection* stackSection, LinkerSection*
void Linker::CopyObjects(bool inlays) void Linker::CopyObjects(bool inlays)
{ {
bool errors = false;
for (int i = 0; i < mObjects.Size(); i++) for (int i = 0; i < mObjects.Size(); i++)
{ {
LinkerObject* obj = mObjects[i]; LinkerObject* obj = mObjects[i];
@ -809,7 +811,9 @@ void Linker::CopyObjects(bool inlays)
{ {
if (!obj->mRegion) if (!obj->mRegion)
{ {
mErrors->Error(obj->mLocation, ERRR_INSUFFICIENT_MEMORY, "Could not place object", obj->mIdent); mErrors->Error(obj->mLocation, errors ? EWARN_INSUFFICIENT_MEMORY : ERRR_INSUFFICIENT_MEMORY, "Could not place object", obj->mIdent);
if (mCompilerOptions & COPT_ERROR_FILES)
errors = true;
int avail = 0; int avail = 0;
for (int i = 0; i < mRegions.Size(); i++) for (int i = 0; i < mRegions.Size(); i++)
@ -1645,7 +1649,7 @@ bool Linker::WriteCrtFile(const char* filename, uint16 id)
bool Linker::WriteMapFile(const char* filename) bool Linker::WriteMapFile(const char* filename)
{ {
bool banked = mCartridgeBankUsed[0]; bool banked = mCartridgeBankUsed[0] || mCartridgeBankUsed[1];
FILE* file; FILE* file;
fopen_s(&file, filename, "wb"); fopen_s(&file, filename, "wb");

204
oscar64/NativeCodeGenerator.cpp
View File

@ -5186,7 +5186,7 @@ static AsmInsType TransposeBranchCondition(AsmInsType code)
} }
int NativeCodeBasicBlock::PutJump(NativeCodeProcedure* proc, NativeCodeBasicBlock* target, int offset) int NativeCodeBasicBlock::PutJump(NativeCodeProcedure* proc, NativeCodeBasicBlock* target, int offset, AsmInsType code)
{ {
if (target->mIns.Size() == 1 && target->mIns[0].mType == ASMIT_RTS) if (target->mIns.Size() == 1 && target->mIns[0].mType == ASMIT_RTS)
{ {
@ -5216,10 +5216,16 @@ int NativeCodeBasicBlock::PutJump(NativeCodeProcedure* proc, NativeCodeBasicBloc
PutWord(0); PutWord(0);
return 3; return 3;
} }
#if JUMP_TO_BRANCH
else if (offset >= -126 && offset <= 129) else if (offset >= -126 && offset <= 129)
{ {
if (mNDataSet.mRegs[CPU_REG_C].mMode == NRDM_IMMEDIATE) if (code != ASMIT_INV)
{
PutOpcode(AsmInsOpcodes[InvertBranchCondition(code)][ASMIM_RELATIVE]);
PutByte(offset - 2);
return 2;
}
#if JUMP_TO_BRANCH
else if (mNDataSet.mRegs[CPU_REG_C].mMode == NRDM_IMMEDIATE)
{ {
if (mNDataSet.mRegs[CPU_REG_C].mValue) if (mNDataSet.mRegs[CPU_REG_C].mValue)
PutOpcode(AsmInsOpcodes[ASMIT_BCS][ASMIM_RELATIVE]); PutOpcode(AsmInsOpcodes[ASMIT_BCS][ASMIM_RELATIVE]);
@ -5239,8 +5245,8 @@ int NativeCodeBasicBlock::PutJump(NativeCodeProcedure* proc, NativeCodeBasicBloc
PutByte(offset - 2); PutByte(offset - 2);
return 2; return 2;
} }
}
#endif #endif
}
PutByte(0x4c); PutByte(0x4c);
LinkerReference rl; LinkerReference rl;
@ -5267,7 +5273,7 @@ int NativeCodeBasicBlock::BranchByteSize(NativeCodeBasicBlock* target, int from,
} }
} }
int NativeCodeBasicBlock::JumpByteSize(NativeCodeBasicBlock* target, int offset) int NativeCodeBasicBlock::JumpByteSize(NativeCodeBasicBlock* target, int offset, bool second)
{ {
if (target->mIns.Size() == 1 && target->mIns[0].mType == ASMIT_RTS) if (target->mIns.Size() == 1 && target->mIns[0].mType == ASMIT_RTS)
{ {
@ -5280,17 +5286,19 @@ int NativeCodeBasicBlock::JumpByteSize(NativeCodeBasicBlock* target, int offset)
{ {
return 3; return 3;
} }
#if JUMP_TO_BRANCH
else if (offset >= -126 && offset <= 129) else if (offset >= -126 && offset <= 129)
{ {
if (mNDataSet.mRegs[CPU_REG_C].mMode == NRDM_IMMEDIATE) if (second)
return 2;
#if JUMP_TO_BRANCH
else if (mNDataSet.mRegs[CPU_REG_C].mMode == NRDM_IMMEDIATE)
return 2; return 2;
else if (mNDataSet.mRegs[CPU_REG_Z].mMode == NRDM_IMMEDIATE) else if (mNDataSet.mRegs[CPU_REG_Z].mMode == NRDM_IMMEDIATE)
return 2; return 2;
#endif
else else
return 3; return 3;
} }
#endif
else else
return 3; return 3;
} }
@ -7252,9 +7260,10 @@ void NativeCodeBasicBlock::LoadValueToReg(InterCodeProcedure* proc, const InterI
{ {
if (ins->mSrc[0].mMemory == IM_INDIRECT) if (ins->mSrc[0].mMemory == IM_INDIRECT)
{ {
int areg = BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp]; int areg = BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp];
int index = int(ins->mSrc[0].mIntConst); int index = int(ins->mSrc[0].mIntConst);
int stride = ins->mSrc[0].mStride; int stride = ins->mSrc[0].mStride;
bool accu = reg == areg;
if (stride * 4 <= 256) if (stride * 4 <= 256)
{ {
@ -7264,7 +7273,10 @@ void NativeCodeBasicBlock::LoadValueToReg(InterCodeProcedure* proc, const InterI
{ {
mIns.Push(NativeCodeInstruction(ins, ASMIT_LDY, ASMIM_IMMEDIATE, index + i * stride)); mIns.Push(NativeCodeInstruction(ins, ASMIT_LDY, ASMIM_IMMEDIATE, index + i * stride));
mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_INDIRECT_Y, areg)); mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_INDIRECT_Y, areg));
mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, reg + i)); if (accu)
mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + i));
else
mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, reg + i));
} }
} }
else else
@ -7275,11 +7287,23 @@ void NativeCodeBasicBlock::LoadValueToReg(InterCodeProcedure* proc, const InterI
mIns.Push(NativeCodeInstruction(ins, ASMIT_LDY, ASMIM_IMMEDIATE, index)); mIns.Push(NativeCodeInstruction(ins, ASMIT_LDY, ASMIM_IMMEDIATE, index));
mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_INDIRECT_Y, areg)); mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_INDIRECT_Y, areg));
mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, reg + i)); if (accu)
mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + i));
else
mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, reg + i));
index += stride; index += stride;
} }
} }
if (accu)
{
for (int i = 0; i < 4; i++)
{
mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_ACCU + i));
mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, reg + i));
}
}
} }
} }
} }
@ -9677,28 +9701,23 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
{ {
case IA_ADD: case IA_ADD:
{ {
NativeCodeGenerator::Runtime& art(nproc->mGenerator->ResolveRuntime(Ident::Unique("faddsub"))); NativeCodeGenerator::Runtime& art(nproc->mGenerator->ResolveRuntime(Ident::Unique("fadd")));
mIns.Push(NativeCodeInstruction(ins, ASMIT_JSR, ASMIM_ABSOLUTE, art.mOffset, art.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER)); mIns.Push(NativeCodeInstruction(ins, ASMIT_JSR, ASMIM_ABSOLUTE, art.mOffset, art.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER));
} break; } break;
case IA_SUB: case IA_SUB:
{ {
if (!changedSign) bool add = false;
if (changedSign)
add = true;
else if (flipop)
{ {
if (flipop) mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
{ mIns.Push(NativeCodeInstruction(ins, ASMIT_EOR, ASMIM_IMMEDIATE, 0x80));
mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3)); mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
mIns.Push(NativeCodeInstruction(ins, ASMIT_EOR, ASMIM_IMMEDIATE, 0x80)); add = true;
mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_ACCU + 3));
}
else
{
mIns.Push(NativeCodeInstruction(ins, ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_WORK + 3));
mIns.Push(NativeCodeInstruction(ins, ASMIT_EOR, ASMIM_IMMEDIATE, 0x80));
mIns.Push(NativeCodeInstruction(ins, ASMIT_STA, ASMIM_ZERO_PAGE, BC_REG_WORK + 3));
}
} }
NativeCodeGenerator::Runtime& art(nproc->mGenerator->ResolveRuntime(Ident::Unique("faddsub"))); NativeCodeGenerator::Runtime& art(nproc->mGenerator->ResolveRuntime(Ident::Unique(add ? "fadd" : "fsub")));
mIns.Push(NativeCodeInstruction(ins, ASMIT_JSR, ASMIM_ABSOLUTE, art.mOffset, art.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER)); mIns.Push(NativeCodeInstruction(ins, ASMIT_JSR, ASMIM_ABSOLUTE, art.mOffset, art.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER));
} break; } break;
case IA_MUL: case IA_MUL:
@ -9733,8 +9752,16 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
case IA_AND: case IA_AND:
case IA_XOR: case IA_XOR:
{ {
if (sins1) LoadValueToReg(proc, sins1, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp], nullptr, nullptr); if (sins1 && sins0 && sins1->mDst.mTemp == sins0->mSrc[0].mTemp)
if (sins0) LoadValueToReg(proc, sins0, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp], nullptr, nullptr); {
LoadValueToReg(proc, sins0, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp], nullptr, nullptr);
LoadValueToReg(proc, sins1, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp], nullptr, nullptr);
}
else
{
if (sins1) LoadValueToReg(proc, sins1, BC_REG_TMP + proc->mTempOffset[ins->mSrc[1].mTemp], nullptr, nullptr);
if (sins0) LoadValueToReg(proc, sins0, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp], nullptr, nullptr);
}
AsmInsType atype; AsmInsType atype;
switch (ins->mOperator) switch (ins->mOperator)
@ -16735,10 +16762,18 @@ bool NativeCodeBasicBlock::GlobalLoadStoreForwarding(bool zpage, const NativeCod
} }
else if (ins.mType == ASMIT_TAY) else if (ins.mType == ASMIT_TAY)
{ {
if (mALSIns.mMode == ASMIM_ABSOLUTE_Y)
mALSIns.mType = ASMIT_INV;
if (mXLSIns.mMode == ASMIM_ABSOLUTE_Y)
mXLSIns.mType = ASMIT_INV;
mYLSIns = mALSIns; mYLSIns = mALSIns;
} }
else if (ins.mType == ASMIT_TAX) else if (ins.mType == ASMIT_TAX)
{ {
if (mALSIns.mMode == ASMIM_ABSOLUTE_X)
mALSIns.mType = ASMIT_INV;
if (mYLSIns.mMode == ASMIM_ABSOLUTE_X)
mYLSIns.mType = ASMIT_INV;
mXLSIns = mALSIns; mXLSIns = mALSIns;
} }
else if (ins.mType == ASMIT_TYA) else if (ins.mType == ASMIT_TYA)
@ -29008,6 +29043,60 @@ bool NativeCodeBasicBlock::IsDominatedBy(const NativeCodeBasicBlock* block) cons
#endif #endif
} }
bool NativeCodeBasicBlock::JoinSameBranch(NativeCodeBasicBlock* block)
{
bool changed = false;
if (!mPatched)
{
mPatched = true;
if (this != block)
{
if (mBranch == block->mBranch && mTrueJump == block->mTrueJump && mFalseJump == block->mFalseJump)
{
mTrueJump->mNumEntries--;
mFalseJump->mNumEntries--;
block->mNumEntries++;
mBranch = ASMIT_JMP;
mTrueJump = block;
mFalseJump = nullptr;
changed = true;
}
}
if (mTrueJump && mTrueJump->JoinSameBranch(block))
changed = true;
if (mFalseJump && mFalseJump->JoinSameBranch(block))
changed = true;
}
return changed;
}
bool NativeCodeBasicBlock::MergeSameBranch(void)
{
bool changed = false;
if (!mVisited)
{
mVisited = true;
if (mIns.Size() == 0 && mFalseJump)
{
mProc->ResetPatched();
if (mProc->mEntryBlock->JoinSameBranch(this))
changed = true;
}
if (mTrueJump && mTrueJump->MergeSameBranch())
changed = true;
if (mFalseJump && mFalseJump->MergeSameBranch())
changed = true;
}
return changed;
}
bool NativeCodeBasicBlock::CheckPatchFailLoop(const NativeCodeBasicBlock* block, const NativeCodeBasicBlock* head, int reg, bool changed) bool NativeCodeBasicBlock::CheckPatchFailLoop(const NativeCodeBasicBlock* block, const NativeCodeBasicBlock* head, int reg, bool changed)
{ {
if (!mPatchLoop || (changed && !mPatchLoopChanged)) if (!mPatchLoop || (changed && !mPatchLoopChanged))
@ -40989,6 +41078,23 @@ bool NativeCodeBasicBlock::BlockSizeCopyReduction(NativeCodeProcedure* proc, int
while (CheckBlockCopySequence(mIns, si + 2 * k * i, i)) while (CheckBlockCopySequence(mIns, si + 2 * k * i, i))
k++; k++;
if (k == 1 && i == 4 && mIns[si + 0].mMode == ASMIM_ZERO_PAGE && mIns[si + 1].mMode == ASMIM_ZERO_PAGE)
{
if (mIns[si + 0].mAddress == BC_REG_ACCU)
{
NativeCodeGenerator::Runtime& frt(mProc->mGenerator->ResolveRuntime(Ident::Unique("store32")));
mIns[di++] = NativeCodeInstruction(mIns[si + 0].mIns, ASMIT_JSR, ASMIM_ABSOLUTE, frt.mOffset, frt.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER | NCIF_USE_ZP_32_X, mIns[si + 1].mAddress);
si += 8;
return true;
}
else if (mIns[si + 1].mAddress == BC_REG_ACCU)
{
NativeCodeGenerator::Runtime& frt(mProc->mGenerator->ResolveRuntime(Ident::Unique("load32")));
mIns[di++] = NativeCodeInstruction(mIns[si + 0].mIns, ASMIT_JSR, ASMIM_ABSOLUTE, frt.mOffset, frt.mLinkerObject, NCIF_RUNTIME | NCIF_LOWER | NCIF_UPPER | NCIF_USE_ZP_32_X, mIns[si + 0].mAddress);
si += 8;
return true;
}
}
int sz = 3 + 4 * k; int sz = 3 + 4 * k;
for (int j = 0; j < k; j++) for (int j = 0; j < k; j++)
@ -50631,6 +50737,22 @@ void NativeCodeBasicBlock::BuildPlacement(ExpandingArray<NativeCodeBasicBlock*>&
mFalseJump->BuildPlacement(placement); mFalseJump->BuildPlacement(placement);
mTrueJump->BuildPlacement(placement); mTrueJump->BuildPlacement(placement);
} }
else if (!mTrueJump->mFalseJump && mTrueJump->mTrueJump && mTrueJump->mTrueJump->mFalseJump &&
mFalseJump->mFalseJump && mFalseJump->mTrueJump == mTrueJump->mTrueJump &&
mTrueJump->mCode.Size() < 20 && mTrueJump->mTrueJump->mCode.Size() < 20 &&
mFalseJump->mFalseJump->LeadsInto(mTrueJump->mTrueJump->mTrueJump, 0) < 10 &&
mFalseJump->mFalseJump->LeadsInto(mTrueJump->mTrueJump->mFalseJump, 0) < 10)
{
mTrueJump->mPlaced = true;
mTrueJump->mPlace = placement.Size();
placement.Push(mTrueJump);
mTrueJump->mTrueJump->mPlaced = true;
mTrueJump->mTrueJump->mPlace = placement.Size();
placement.Push(mTrueJump->mTrueJump);
mFalseJump->BuildPlacement(placement);
}
else if (!mTrueJump->mFalseJump && mTrueJump->mTrueJump && mTrueJump->mCode.Size() < 100 && mFalseJump->LeadsInto(mTrueJump->mTrueJump, 0) < 100) else if (!mTrueJump->mFalseJump && mTrueJump->mTrueJump && mTrueJump->mCode.Size() < 100 && mFalseJump->LeadsInto(mTrueJump->mTrueJump, 0) < 100)
{ {
mTrueJump->mPlaced = true; mTrueJump->mPlaced = true;
@ -50759,19 +50881,19 @@ bool NativeCodeBasicBlock::CalculateOffset(int& total)
mFalseJump->mIns.Size() == 1 && mFalseJump->mIns[0].mType == ASMIT_RTS) mFalseJump->mIns.Size() == 1 && mFalseJump->mIns[0].mType == ASMIT_RTS)
{ {
total += BranchByteSize(mTrueJump, total, mTrueJump->mOffset); total += BranchByteSize(mTrueJump, total, mTrueJump->mOffset);
total += JumpByteSize(mFalseJump, mFalseJump->mOffset - total); total += JumpByteSize(mFalseJump, mFalseJump->mOffset - total, true);
} }
else if ( else if (
mFalseJump->mPlace > mTrueJump->mPlace && mFalseJump->mPlace < mPlace || mFalseJump->mPlace > mTrueJump->mPlace && mFalseJump->mPlace < mPlace ||
mFalseJump->mPlace < mTrueJump->mPlace && mFalseJump->mPlace > mPlace) mFalseJump->mPlace < mTrueJump->mPlace && mFalseJump->mPlace > mPlace)
{ {
total += BranchByteSize(mFalseJump, total, mFalseJump->mOffset); total += BranchByteSize(mFalseJump, total, mFalseJump->mOffset);
total += JumpByteSize(mTrueJump, mTrueJump->mOffset - total); total += JumpByteSize(mTrueJump, mTrueJump->mOffset - total, true);
} }
else else
{ {
total += BranchByteSize(mTrueJump, total, mTrueJump->mOffset); total += BranchByteSize(mTrueJump, total, mTrueJump->mOffset);
total += JumpByteSize(mFalseJump, mFalseJump->mOffset - total); total += JumpByteSize(mFalseJump, mFalseJump->mOffset - total, true);
} }
} }
else if (mTrueJump) else if (mTrueJump)
@ -50786,7 +50908,7 @@ bool NativeCodeBasicBlock::CalculateOffset(int& total)
mTrueJump = nullptr; mTrueJump = nullptr;
} }
else else
total += JumpByteSize(mTrueJump, mTrueJump->mOffset - total); total += JumpByteSize(mTrueJump, mTrueJump->mOffset - total, false);
} }
} }
@ -50897,19 +51019,19 @@ void NativeCodeBasicBlock::CopyCode(NativeCodeProcedure * proc, uint8* target)
mFalseJump->mIns.Size() == 1 && mFalseJump->mIns[0].mType == ASMIT_RTS) mFalseJump->mIns.Size() == 1 && mFalseJump->mIns[0].mType == ASMIT_RTS)
{ {
end += PutBranch(proc, mTrueJump, mBranch, mTrueJump->mOffset - end); end += PutBranch(proc, mTrueJump, mBranch, mTrueJump->mOffset - end);
end += PutJump(proc, mFalseJump, mFalseJump->mOffset - end); end += PutJump(proc, mFalseJump, mFalseJump->mOffset - end, mBranch);
} }
else if ( else if (
mFalseJump->mPlace > mTrueJump->mPlace && mFalseJump->mPlace < mPlace || mFalseJump->mPlace > mTrueJump->mPlace && mFalseJump->mPlace < mPlace ||
mFalseJump->mPlace < mTrueJump->mPlace && mFalseJump->mPlace > mPlace) mFalseJump->mPlace < mTrueJump->mPlace && mFalseJump->mPlace > mPlace)
{ {
end += PutBranch(proc, mFalseJump, InvertBranchCondition(mBranch), mFalseJump->mOffset - end); end += PutBranch(proc, mFalseJump, InvertBranchCondition(mBranch), mFalseJump->mOffset - end);
end += PutJump(proc, mTrueJump, mTrueJump->mOffset - end); end += PutJump(proc, mTrueJump, mTrueJump->mOffset - end, InvertBranchCondition(mBranch));
} }
else else
{ {
end += PutBranch(proc, mTrueJump, mBranch, mTrueJump->mOffset - end); end += PutBranch(proc, mTrueJump, mBranch, mTrueJump->mOffset - end);
end += PutJump(proc, mFalseJump, mFalseJump->mOffset - end); end += PutJump(proc, mFalseJump, mFalseJump->mOffset - end, mBranch);
} }
if (mBranchIns) if (mBranchIns)
@ -51201,7 +51323,7 @@ void NativeCodeProcedure::Compile(InterCodeProcedure* proc)
mInterProc = proc; mInterProc = proc;
mInterProc->mLinkerObject->mNativeProc = this; mInterProc->mLinkerObject->mNativeProc = this;
CheckFunc = !strcmp(mInterProc->mIdent->mString, "_showFrame1"); CheckFunc = !strcmp(mInterProc->mIdent->mString, "printf");
int nblocks = proc->mBlocks.Size(); int nblocks = proc->mBlocks.Size();
tblocks = new NativeCodeBasicBlock * [nblocks]; tblocks = new NativeCodeBasicBlock * [nblocks];
@ -52921,6 +53043,8 @@ void NativeCodeProcedure::Optimize(void)
} }
} }
#endif #endif
#if 1 #if 1
if (step == 14 && cnt == 0) if (step == 14 && cnt == 0)
{ {
@ -53015,6 +53139,7 @@ void NativeCodeProcedure::Optimize(void)
} while (changed); } while (changed);
#if 1 #if 1
ResetVisited(); ResetVisited();
mEntryBlock->CombineAlternateLoads(); mEntryBlock->CombineAlternateLoads();
@ -53121,6 +53246,9 @@ void NativeCodeProcedure::Optimize(void)
ResetVisited(); ResetVisited();
mEntryBlock->RemoveJumpToBranch(); mEntryBlock->RemoveJumpToBranch();
ResetVisited();
mEntryBlock->MergeSameBranch();
#if 1 #if 1
ResetVisited(); ResetVisited();
data.Reset(); data.Reset();

7
oscar64/NativeCodeGenerator.h
View File

@ -265,8 +265,8 @@ public:
NativeCodeInstruction DecodeNative(const InterInstruction* ins, LinkerObject * lobj, int& offset) const; NativeCodeInstruction DecodeNative(const InterInstruction* ins, LinkerObject * lobj, int& offset) const;
int PutBranch(NativeCodeProcedure* proc, NativeCodeBasicBlock* target, AsmInsType code, int offset); int PutBranch(NativeCodeProcedure* proc, NativeCodeBasicBlock* target, AsmInsType code, int offset);
int PutJump(NativeCodeProcedure* proc, NativeCodeBasicBlock* target, int offset); int PutJump(NativeCodeProcedure* proc, NativeCodeBasicBlock* target, int offset, AsmInsType code = ASMIT_INV);
int JumpByteSize(NativeCodeBasicBlock * target, int offset); int JumpByteSize(NativeCodeBasicBlock * target, int offset, bool second);
int BranchByteSize(NativeCodeBasicBlock* target, int from, int to); int BranchByteSize(NativeCodeBasicBlock* target, int from, int to);
NativeCodeBasicBlock* SplitAt(int at); NativeCodeBasicBlock* SplitAt(int at);
@ -746,6 +746,9 @@ public:
bool CheckPatchFailLoop(const NativeCodeBasicBlock* block, const NativeCodeBasicBlock* head, int reg, bool changed); bool CheckPatchFailLoop(const NativeCodeBasicBlock* block, const NativeCodeBasicBlock* head, int reg, bool changed);
bool JoinSameBranch(NativeCodeBasicBlock* block);
bool MergeSameBranch(void);
// reg : base register pair to replace // reg : base register pair to replace
// index: index register // index: index register
// at : start position in block // at : start position in block

22
oscar64/oscar64.cpp
View File

@ -225,19 +225,19 @@ int main2(int argc, const char** argv)
strcpy_s(cid, arg + 5); strcpy_s(cid, arg + 5);
compiler->mCartridgeID = atoi(cid); compiler->mCartridgeID = atoi(cid);
} }
else if (arg[1] == 'n') else if (arg[1] == 'n' && arg[2] == 0)
{ {
compiler->mCompilerOptions |= COPT_NATIVE; compiler->mCompilerOptions |= COPT_NATIVE;
} }
else if (arg[1] == 'b' && arg[2] == 'c') else if (arg[1] == 'b' && arg[2] == 'c' && arg[3] == 0)
{ {
compiler->mCompilerOptions &= ~COPT_NATIVE; compiler->mCompilerOptions &= ~COPT_NATIVE;
} }
else if (arg[1] == 'p' && arg[2] == 's' && arg[3] == 'c' && arg[4] == 'i') else if (arg[1] == 'p' && arg[2] == 's' && arg[3] == 'c' && arg[4] == 'i' && arg[5] == 0)
{ {
compiler->mCompilerOptions |= COPT_PETSCII; compiler->mCompilerOptions |= COPT_PETSCII;
} }
else if (arg[1] == 'O') else if (arg[1] == 'O' && arg[2] != 0 && arg[3] == 0)
{ {
if (arg[2] == '0') if (arg[2] == '0')
compiler->mCompilerOptions &= ~(COPT_OPTIMIZE_ALL); compiler->mCompilerOptions &= ~(COPT_OPTIMIZE_ALL);
@ -261,6 +261,8 @@ int main2(int argc, const char** argv)
compiler->mCompilerOptions |= COPT_OPTIMIZE_GLOBAL; compiler->mCompilerOptions |= COPT_OPTIMIZE_GLOBAL;
else if (arg[2] == 'm') else if (arg[2] == 'm')
compiler->mCompilerOptions |= COPT_OPTIMIZE_MERGE_CALLS; compiler->mCompilerOptions |= COPT_OPTIMIZE_MERGE_CALLS;
else
compiler->mErrors->Error(loc, EERR_COMMAND_LINE, "Invalid command line argument", arg);
} }
else if (arg[1] == 'e') else if (arg[1] == 'e')
{ {
@ -291,7 +293,7 @@ int main2(int argc, const char** argv)
else else
compiler->AddDefine(Ident::Unique(def), ""); compiler->AddDefine(Ident::Unique(def), "");
} }
else if (arg[1] == 'g') else if (arg[1] == 'g' && !arg[2])
{ {
compiler->mCompilerOptions |= COPT_DEBUGINFO; compiler->mCompilerOptions |= COPT_DEBUGINFO;
} }
@ -302,17 +304,19 @@ int main2(int argc, const char** argv)
compiler->mCompilerOptions |= COPT_VERBOSE2; compiler->mCompilerOptions |= COPT_VERBOSE2;
else if (arg[2] == '3') else if (arg[2] == '3')
compiler->mCompilerOptions |= COPT_VERBOSE2 | COPT_VERBOSE3; compiler->mCompilerOptions |= COPT_VERBOSE2 | COPT_VERBOSE3;
else if (!arg[2])
compiler->mErrors->Error(loc, EERR_COMMAND_LINE, "Invalid command line argument", arg);
} }
else if (arg[1] == 'x' && arg[2] == 'z') else if (arg[1] == 'x' && arg[2] == 'z' && !arg[3])
{ {
compiler->mCompilerOptions |= COPT_EXTENDED_ZERO_PAGE; compiler->mCompilerOptions |= COPT_EXTENDED_ZERO_PAGE;
} }
else if (arg[1] == 'p' && arg[2] == 'p') else if (arg[1] == 'p' && arg[2] == 'p' && !arg[3])
{ {
compiler->mCompilerOptions |= COPT_CPLUSPLUS; compiler->mCompilerOptions |= COPT_CPLUSPLUS;
compiler->AddDefine(Ident::Unique("__cplusplus"), "1"); compiler->AddDefine(Ident::Unique("__cplusplus"), "1");
} }
else if (arg[1] == 'r' && arg[2] == 'm' && arg[3] == 'p') else if (arg[1] == 'r' && arg[2] == 'm' && arg[3] == 'p' && !arg[4])
{ {
compiler->mCompilerOptions |= COPT_ERROR_FILES; compiler->mCompilerOptions |= COPT_ERROR_FILES;
} }
@ -563,6 +567,8 @@ int main2(int argc, const char** argv)
if (diskPath[0]) if (diskPath[0])
d64 = new DiskImage(diskPath); d64 = new DiskImage(diskPath);
compiler->RemoveErrorFile(targetPath);
compiler->WriteOutputFile(targetPath, d64); compiler->WriteOutputFile(targetPath, d64);
if (d64) if (d64)