const MachineRegisterInfo * MRI;
const R600RegisterInfo * TRI;
- bool isCube;
- bool isReduction;
- bool isVector;
+ bool IsCube;
+ bool IsReduction;
+ bool IsVector;
unsigned currentElement;
- bool isLast;
+ bool IsLast;
unsigned section_start;
public:
R600CodeEmitter(formatted_raw_ostream &OS) : MachineFunctionPass(ID),
- _OS(OS), TM(NULL), isCube(false), isReduction(false), isVector(false),
- isLast(true) { }
+ _OS(OS), TM(NULL), IsCube(false), IsReduction(false), IsVector(false),
+ IsLast(true) { }
const char *getPassName() const { return "AMDGPU Machine Code Emitter"; }
private:
- void emitALUInstr(MachineInstr &MI);
- void emitSrc(const MachineOperand & MO, int chan_override = -1);
- void emitDst(const MachineOperand & MO);
- void emitALU(MachineInstr &MI, unsigned numSrc);
- void emitTexInstr(MachineInstr &MI);
- void emitFCInstr(MachineInstr &MI);
+ void EmitALUInstr(MachineInstr &MI);
+ void EmitSrc(const MachineOperand & MO, int chan_override = -1);
+ void EmitDst(const MachineOperand & MO);
+ void EmitALU(MachineInstr &MI, unsigned numSrc);
+ void EmitTexInstr(MachineInstr &MI);
+ void EmitFCInstr(MachineInstr &MI);
- void emitNullBytes(unsigned int byteCount);
+ void EmitNullBytes(unsigned int byteCount);
- void emitByte(unsigned int byte);
+ void EmitByte(unsigned int byte);
- void emitTwoBytes(uint32_t bytes);
+ void EmitTwoBytes(uint32_t bytes);
- void emit(uint32_t value);
- void emit(uint64_t value);
+ void Emit(uint32_t value);
+ void Emit(uint64_t value);
unsigned getHWReg(unsigned regNo) const;
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
I != E; ++I) {
MachineInstr &MI = *I;
- isReduction = AMDGPU::isReductionOp(MI.getOpcode());
- isVector = TII->isVector(MI);
- isCube = AMDGPU::isCubeOp(MI.getOpcode());
+ IsReduction = AMDGPU::isReductionOp(MI.getOpcode());
+ IsVector = TII->isVector(MI);
+ IsCube = AMDGPU::isCubeOp(MI.getOpcode());
if (MI.getNumOperands() > 1 && MI.getOperand(0).isReg() && MI.getOperand(0).isDead()) {
continue;
}
if (AMDGPU::isTexOp(MI.getOpcode())) {
- emitTexInstr(MI);
+ EmitTexInstr(MI);
} else if (AMDGPU::isFCOp(MI.getOpcode())){
- emitFCInstr(MI);
- } else if (isReduction || isVector || isCube) {
- isLast = false;
+ EmitFCInstr(MI);
+ } else if (IsReduction || IsVector || IsCube) {
+ IsLast = false;
// XXX: On Cayman, some (all?) of the vector instructions only need
// to fill the first three slots.
for (currentElement = 0; currentElement < 4; currentElement++) {
- isLast = (currentElement == 3);
- emitALUInstr(MI);
+ IsLast = (currentElement == 3);
+ EmitALUInstr(MI);
}
- isReduction = false;
- isVector = false;
- isCube = false;
+ IsReduction = false;
+ IsVector = false;
+ IsCube = false;
} else if (MI.getOpcode() == AMDGPU::RETURN ||
MI.getOpcode() == AMDGPU::BUNDLE ||
MI.getOpcode() == AMDGPU::KILL) {
if (NextMI.getOpcode() == AMDGPU::RETURN) {
inst |= (((uint64_t)1) << 53);
}
- emitByte(INSTR_NATIVE);
- emit(inst);
+ EmitByte(INSTR_NATIVE);
+ Emit(inst);
break;
}
case AMDGPU::VTX_READ_PARAM_i32_eg:
uint64_t InstWord01 = getBinaryCodeForInstr(MI);
uint32_t InstWord2 = MI.getOperand(2).getImm(); // Offset
- emitByte(INSTR_VTX);
- emit(InstWord01);
- emit(InstWord2);
+ EmitByte(INSTR_VTX);
+ Emit(InstWord01);
+ Emit(InstWord2);
break;
}
default:
- emitALUInstr(MI);
+ EmitALUInstr(MI);
break;
}
}
return false;
}
-void R600CodeEmitter::emitALUInstr(MachineInstr &MI)
+void R600CodeEmitter::EmitALUInstr(MachineInstr &MI)
{
unsigned numOperands = MI.getNumExplicitOperands();
const MachineOperand dstOp = MI.getOperand(0);
// Emit instruction type
- emitByte(0);
+ EmitByte(0);
- if (isCube) {
+ if (IsCube) {
static const int cube_src_swz[] = {2, 2, 0, 1};
- emitSrc(MI.getOperand(1), cube_src_swz[currentElement]);
- emitSrc(MI.getOperand(1), cube_src_swz[3-currentElement]);
- emitNullBytes(SRC_BYTE_COUNT);
+ EmitSrc(MI.getOperand(1), cube_src_swz[currentElement]);
+ EmitSrc(MI.getOperand(1), cube_src_swz[3-currentElement]);
+ EmitNullBytes(SRC_BYTE_COUNT);
} else {
unsigned int opIndex;
for (opIndex = 1; opIndex < numOperands; opIndex++) {
if (MI.getOperand(opIndex).isImm() || MI.getOperand(opIndex).isFPImm()) {
break;
}
- emitSrc(MI.getOperand(opIndex));
+ EmitSrc(MI.getOperand(opIndex));
}
// Emit zeros for unused sources
for ( ; opIndex < 4; opIndex++) {
- emitNullBytes(SRC_BYTE_COUNT);
+ EmitNullBytes(SRC_BYTE_COUNT);
}
}
- emitDst(dstOp);
+ EmitDst(dstOp);
- emitALU(MI, numOperands - 1);
+ EmitALU(MI, numOperands - 1);
}
-void R600CodeEmitter::emitSrc(const MachineOperand & MO, int chan_override)
+void R600CodeEmitter::EmitSrc(const MachineOperand & MO, int chan_override)
{
uint32_t value = 0;
// Emit the source select (2 bytes). For GPRs, this is the register index.
// value of the source select is defined in the r600isa docs.
if (MO.isReg()) {
unsigned reg = MO.getReg();
- emitTwoBytes(getHWReg(reg));
+ EmitTwoBytes(getHWReg(reg));
if (reg == AMDGPU::ALU_LITERAL_X) {
const MachineInstr * parent = MO.getParent();
unsigned immOpIndex = parent->getNumExplicitOperands() - 1;
}
} else {
// XXX: Handle other operand types.
- emitTwoBytes(0);
+ EmitTwoBytes(0);
}
// Emit the source channel (1 byte)
if (chan_override != -1) {
- emitByte(chan_override);
- } else if (isReduction) {
- emitByte(currentElement);
+ EmitByte(chan_override);
+ } else if (IsReduction) {
+ EmitByte(currentElement);
} else if (MO.isReg()) {
- emitByte(TRI->getHWRegChan(MO.getReg()));
+ EmitByte(TRI->getHWRegChan(MO.getReg()));
} else {
- emitByte(0);
+ EmitByte(0);
}
// XXX: Emit isNegated (1 byte)
&& (MO.getTargetFlags() & MO_FLAG_NEG ||
(MO.isReg() &&
(MO.getReg() == AMDGPU::NEG_ONE || MO.getReg() == AMDGPU::NEG_HALF)))){
- emitByte(1);
+ EmitByte(1);
} else {
- emitByte(0);
+ EmitByte(0);
}
// Emit isAbsolute (1 byte)
if (MO.getTargetFlags() & MO_FLAG_ABS) {
- emitByte(1);
+ EmitByte(1);
} else {
- emitByte(0);
+ EmitByte(0);
}
// XXX: Emit relative addressing mode (1 byte)
- emitByte(0);
+ EmitByte(0);
// Emit kc_bank, This will be adjusted later by r600_asm
- emitByte(0);
+ EmitByte(0);
// Emit the literal value, if applicable (4 bytes).
- emit(value);
+ Emit(value);
}
-void R600CodeEmitter::emitDst(const MachineOperand & MO)
+void R600CodeEmitter::EmitDst(const MachineOperand & MO)
{
if (MO.isReg()) {
// Emit the destination register index (1 byte)
- emitByte(getHWReg(MO.getReg()));
+ EmitByte(getHWReg(MO.getReg()));
// Emit the element of the destination register (1 byte)
- if (isReduction || isCube || isVector) {
- emitByte(currentElement);
+ if (IsReduction || IsCube || IsVector) {
+ EmitByte(currentElement);
} else {
- emitByte(TRI->getHWRegChan(MO.getReg()));
+ EmitByte(TRI->getHWRegChan(MO.getReg()));
}
// Emit isClamped (1 byte)
if (MO.getTargetFlags() & MO_FLAG_CLAMP) {
- emitByte(1);
+ EmitByte(1);
} else {
- emitByte(0);
+ EmitByte(0);
}
// Emit writemask (1 byte).
- if (((isReduction || isVector) &&
+ if (((IsReduction || IsVector) &&
currentElement != TRI->getHWRegChan(MO.getReg()))
|| MO.getTargetFlags() & MO_FLAG_MASK) {
- emitByte(0);
+ EmitByte(0);
} else {
- emitByte(1);
+ EmitByte(1);
}
// XXX: Emit relative addressing mode
- emitByte(0);
+ EmitByte(0);
} else {
// XXX: Handle other operand types. Are there any for destination regs?
- emitNullBytes(DST_BYTE_COUNT);
+ EmitNullBytes(DST_BYTE_COUNT);
}
}
-void R600CodeEmitter::emitALU(MachineInstr &MI, unsigned numSrc)
+void R600CodeEmitter::EmitALU(MachineInstr &MI, unsigned numSrc)
{
// Emit the instruction (2 bytes)
- emitTwoBytes(getBinaryCodeForInstr(MI));
+ EmitTwoBytes(getBinaryCodeForInstr(MI));
- // Emit isLast (for this instruction group) (1 byte)
- if (isLast) {
- emitByte(1);
+ // Emit IsLast (for this instruction group) (1 byte)
+ if (IsLast) {
+ EmitByte(1);
} else {
- emitByte(0);
+ EmitByte(0);
}
// Emit isOp3 (1 byte)
if (numSrc == 3) {
- emitByte(1);
+ EmitByte(1);
} else {
- emitByte(0);
+ EmitByte(0);
}
// XXX: Emit predicate (1 byte)
- emitByte(0);
+ EmitByte(0);
// XXX: Emit bank swizzle. (1 byte) Do we need this? It looks like
// r600_asm.c sets it.
- emitByte(0);
+ EmitByte(0);
// XXX: Emit bank_swizzle_force (1 byte) Not sure what this is for.
- emitByte(0);
+ EmitByte(0);
// XXX: Emit OMOD (1 byte) Not implemented.
- emitByte(0);
+ EmitByte(0);
// XXX: Emit index_mode. I think this is for indirect addressing, so we
// don't need to worry about it.
- emitByte(0);
+ EmitByte(0);
}
-void R600CodeEmitter::emitTexInstr(MachineInstr &MI)
+void R600CodeEmitter::EmitTexInstr(MachineInstr &MI)
{
unsigned opcode = MI.getOpcode();
unsigned srcSelect[4] = {0, 1, 2, 3};
// Emit instruction type
- emitByte(1);
+ EmitByte(1);
// Emit instruction
- emitByte(getBinaryCodeForInstr(MI));
+ EmitByte(getBinaryCodeForInstr(MI));
// XXX: Emit resource id r600_shader.c uses sampler + 1. Why?
- emitByte(sampler + 1 + 1);
+ EmitByte(sampler + 1 + 1);
// Emit source register
- emitByte(getHWReg(MI.getOperand(1).getReg()));
+ EmitByte(getHWReg(MI.getOperand(1).getReg()));
// XXX: Emit src isRelativeAddress
- emitByte(0);
+ EmitByte(0);
// Emit destination register
- emitByte(getHWReg(MI.getOperand(0).getReg()));
+ EmitByte(getHWReg(MI.getOperand(0).getReg()));
// XXX: Emit dst isRealtiveAddress
- emitByte(0);
+ EmitByte(0);
// XXX: Emit dst select
- emitByte(0); // X
- emitByte(1); // Y
- emitByte(2); // Z
- emitByte(3); // W
+ EmitByte(0); // X
+ EmitByte(1); // Y
+ EmitByte(2); // Z
+ EmitByte(3); // W
// XXX: Emit lod bias
- emitByte(0);
+ EmitByte(0);
// XXX: Emit coord types
unsigned coordType[4] = {1, 1, 1, 1};
}
for (unsigned i = 0; i < 4; i++) {
- emitByte(coordType[i]);
+ EmitByte(coordType[i]);
}
// XXX: Emit offsets
if (hasOffsets)
for (unsigned i = 2; i < 5; i++)
- emitByte(MI.getOperand(i).getImm()<<1);
+ EmitByte(MI.getOperand(i).getImm()<<1);
else
- emitNullBytes(3);
+ EmitNullBytes(3);
// Emit sampler id
- emitByte(sampler);
+ EmitByte(sampler);
// XXX:Emit source select
if ((textureType == TEXTURE_SHADOW1D
}
for (unsigned i = 0; i < 4; i++) {
- emitByte(srcSelect[i]);
+ EmitByte(srcSelect[i]);
}
}
-void R600CodeEmitter::emitFCInstr(MachineInstr &MI)
+void R600CodeEmitter::EmitFCInstr(MachineInstr &MI)
{
// Emit instruction type
- emitByte(INSTR_FC);
+ EmitByte(INSTR_FC);
// Emit SRC
unsigned numOperands = MI.getNumOperands();
if (numOperands > 0) {
assert(numOperands == 1);
- emitSrc(MI.getOperand(0));
+ EmitSrc(MI.getOperand(0));
} else {
- emitNullBytes(SRC_BYTE_COUNT);
+ EmitNullBytes(SRC_BYTE_COUNT);
}
// Emit FC Instruction
abort();
break;
}
- emitByte(instr);
+ EmitByte(instr);
}
-void R600CodeEmitter::emitNullBytes(unsigned int byteCount)
+void R600CodeEmitter::EmitNullBytes(unsigned int byteCount)
{
for (unsigned int i = 0; i < byteCount; i++) {
- emitByte(0);
+ EmitByte(0);
}
}
-void R600CodeEmitter::emitByte(unsigned int byte)
+void R600CodeEmitter::EmitByte(unsigned int byte)
{
_OS.write((uint8_t) byte & 0xff);
}
-void R600CodeEmitter::emitTwoBytes(unsigned int bytes)
+void R600CodeEmitter::EmitTwoBytes(unsigned int bytes)
{
_OS.write((uint8_t) (bytes & 0xff));
_OS.write((uint8_t) ((bytes >> 8) & 0xff));
}
-void R600CodeEmitter::emit(uint32_t value)
+void R600CodeEmitter::Emit(uint32_t value)
{
for (unsigned i = 0; i < 4; i++) {
_OS.write((uint8_t) ((value >> (8 * i)) & 0xff));
}
}
-void R600CodeEmitter::emit(uint64_t value)
+void R600CodeEmitter::Emit(uint64_t value)
{
for (unsigned i = 0; i < 8; i++) {
- emitByte((value >> (8 * i)) & 0xff);
+ EmitByte((value >> (8 * i)) & 0xff);
}
}
unsigned R600CodeEmitter::getHWReg(unsigned regNo) const
{
- unsigned hwReg;
+ unsigned HWReg;
- hwReg = TRI->getHWRegIndex(regNo);
+ HWReg = TRI->getHWRegIndex(regNo);
if (AMDGPU::R600_CReg32RegClass.contains(regNo)) {
- hwReg += 512;
+ HWReg += 512;
}
- return hwReg;
+ return HWReg;
}
uint64_t R600CodeEmitter::getMachineOpValue(const MachineInstr &MI,
#define CONSTANT_CACHE_SIZE_DW 127
-class R600KernelParameters : public FunctionPass
-{
- const TargetData * TD;
+class R600KernelParameters : public FunctionPass {
+ const TargetData *TD;
LLVMContext* Context;
- Module *mod;
+ Module *Mod;
- struct param
- {
- param() : val(NULL), ptr_val(NULL), offset_in_dw(0), size_in_dw(0),
- indirect(true), specialID(0) {}
+ struct Param {
+ Param() : Val(NULL), PtrVal(NULL), OffsetInDW(0), SizeInDW(0),
+ IsIndirect(true), SpecialID(0) {}
- Value* val;
- Value* ptr_val;
- int offset_in_dw;
- int size_in_dw;
+ Value* Val;
+ Value* PtrVal;
+ int OffsetInDW;
+ int SizeInDW;
- bool indirect;
+ bool IsIndirect;
- std::string specialType;
- int specialID;
+ std::string SpecialType;
+ int SpecialID;
- int end() { return offset_in_dw + size_in_dw; }
+ int End() { return OffsetInDW + SizeInDW; }
// The first 9 dwords are reserved for the grid sizes.
- int get_rat_offset() { return 9 + offset_in_dw; }
+ int getRatOffset() { return 9 + OffsetInDW; }
};
- std::vector<param> params;
+ std::vector<Param> Params;
- bool isOpenCLKernel(const Function* fun);
+ bool IsOpenCLKernel(const Function *Fun);
int getLastSpecialID(const std::string& TypeName);
int getListSize();
- void AddParam(Argument* arg);
- int calculateArgumentSize(Argument* arg);
- void RunAna(Function* fun);
- void Replace(Function* fun);
- bool isIndirect(Value* val, std::set<Value*>& visited);
- void Propagate(Function* fun);
- void Propagate(Value* v, const Twine& name, bool indirect = true);
- Value* ConstantRead(Function* fun, param& p);
- Value* handleSpecial(Function* fun, param& p);
- bool isSpecialType(Type*);
- std::string getSpecialTypeName(Type*);
+ void AddParam(Argument *Arg);
+ int CalculateArgumentSize(Argument *Arg);
+ void RunAna(Function *Fun);
+ void Replace(Function *Fun);
+ bool IsIndirect(Value *Val, std::set<Value*> &Visited);
+ void Propagate(Function* Fun);
+ void Propagate(Value *V, const Twine &Name, bool IsIndirect = true);
+ Value* ConstantRead(Function *Fun, Param &P);
+ Value* handleSpecial(Function *Fun, Param &P);
+ bool IsSpecialType(Type *T);
+ std::string getSpecialTypeName(Type *T);
public:
static char ID;
R600KernelParameters() : FunctionPass(ID) {};
static RegisterPass<R600KernelParameters> X("kerparam",
"OpenCL Kernel Parameter conversion", false, false);
-bool R600KernelParameters::isOpenCLKernel(const Function* fun)
-{
- Module *mod = const_cast<Function*>(fun)->getParent();
- NamedMDNode * md = mod->getOrInsertNamedMetadata("opencl.kernels");
+bool R600KernelParameters::IsOpenCLKernel(const Function* Fun) {
+ Module *Mod = const_cast<Function*>(Fun)->getParent();
+ NamedMDNode * MD = Mod->getOrInsertNamedMetadata("opencl.kernels");
- if (!md or !md->getNumOperands())
- {
+ if (!MD or !MD->getNumOperands()) {
return false;
}
- for (int i = 0; i < int(md->getNumOperands()); i++)
- {
- if (!md->getOperand(i) or !md->getOperand(i)->getOperand(0))
- {
+ for (int i = 0; i < int(MD->getNumOperands()); i++) {
+ if (!MD->getOperand(i) or !MD->getOperand(i)->getOperand(0)) {
continue;
}
-
- assert(md->getOperand(i)->getNumOperands() == 1);
- if (md->getOperand(i)->getOperand(0)->getName() == fun->getName())
- {
+ assert(MD->getOperand(i)->getNumOperands() == 1);
+
+ if (MD->getOperand(i)->getOperand(0)->getName() == Fun->getName()) {
return true;
}
}
return false;
}
-int R600KernelParameters::getLastSpecialID(const std::string& TypeName)
-{
- int lastID = -1;
+int R600KernelParameters::getLastSpecialID(const std::string &TypeName) {
+ int LastID = -1;
- for (std::vector<param>::iterator i = params.begin(); i != params.end(); i++)
- {
- if (i->specialType == TypeName)
- {
- lastID = i->specialID;
+ for (std::vector<Param>::iterator i = Params.begin(); i != Params.end(); i++) {
+ if (i->SpecialType == TypeName) {
+ LastID = i->SpecialID;
}
}
- return lastID;
+ return LastID;
}
-int R600KernelParameters::getListSize()
-{
- if (params.size() == 0)
- {
+int R600KernelParameters::getListSize() {
+ if (Params.size() == 0) {
return 0;
}
- return params.back().end();
+ return Params.back().End();
}
-bool R600KernelParameters::isIndirect(Value* val, std::set<Value*>& visited)
-{
+bool R600KernelParameters::IsIndirect(Value *Val, std::set<Value*> &Visited) {
//XXX Direct parameters are not supported yet, so return true here.
return true;
#if 0
- if (isa<LoadInst>(val))
- {
+ if (isa<LoadInst>(Val)) {
return false;
}
- if (isa<IntegerType>(val->getType()))
- {
+ if (isa<IntegerType>(Val->getType())) {
assert(0 and "Internal error");
return false;
}
- if (visited.count(val))
- {
+ if (Visited.count(Val)) {
return false;
}
- visited.insert(val);
+ Visited.insert(Val);
- if (isa<GetElementPtrInst>(val))
- {
- GetElementPtrInst* GEP = dyn_cast<GetElementPtrInst>(val);
- GetElementPtrInst::op_iterator i = GEP->op_begin();
+ if (isa<getElementPtrInst>(Val)) {
+ getElementPtrInst* GEP = dyn_cast<getElementPtrInst>(Val);
+ getElementPtrInst::op_iterator I = GEP->op_begin();
- for (i++; i != GEP->op_end(); i++)
- {
- if (!isa<Constant>(*i))
- {
+ for (++I; I != GEP->op_end(); ++I) {
+ if (!isa<Constant>(*I)) {
return true;
}
}
}
- for (Value::use_iterator i = val->use_begin(); i != val->use_end(); i++)
- {
- Value* v2 = dyn_cast<Value>(*i);
+ for (Value::use_iterator I = Val->use_begin(); i != Val->use_end(); ++I) {
+ Value* V2 = dyn_cast<Value>(*I);
- if (v2)
- {
- if (isIndirect(v2, visited))
- {
+ if (V2) {
+ if (IsIndirect(V2, Visited)) {
return true;
}
}
#endif
}
-void R600KernelParameters::AddParam(Argument* arg)
-{
- param p;
+void R600KernelParameters::AddParam(Argument *Arg) {
+ Param P;
- p.val = dyn_cast<Value>(arg);
- p.offset_in_dw = getListSize();
- p.size_in_dw = calculateArgumentSize(arg);
+ P.Val = dyn_cast<Value>(Arg);
+ P.OffsetInDW = getListSize();
+ P.SizeInDW = CalculateArgumentSize(Arg);
- if (isa<PointerType>(arg->getType()) and arg->hasByValAttr())
- {
- std::set<Value*> visited;
- p.indirect = isIndirect(p.val, visited);
+ if (isa<PointerType>(Arg->getType()) and Arg->hasByValAttr()) {
+ std::set<Value*> Visited;
+ P.IsIndirect = IsIndirect(P.Val, Visited);
}
- params.push_back(p);
+ Params.push_back(P);
}
-int R600KernelParameters::calculateArgumentSize(Argument* arg)
-{
- Type* t = arg->getType();
+int R600KernelParameters::CalculateArgumentSize(Argument *Arg) {
+ Type* T = Arg->getType();
- if (arg->hasByValAttr() and dyn_cast<PointerType>(t))
- {
- t = dyn_cast<PointerType>(t)->getElementType();
+ if (Arg->hasByValAttr() and dyn_cast<PointerType>(T)) {
+ T = dyn_cast<PointerType>(T)->getElementType();
}
- int store_size_in_dw = (TD->getTypeStoreSize(t) + 3)/4;
+ int StoreSizeInDW = (TD->getTypeStoreSize(T) + 3)/4;
- assert(store_size_in_dw);
+ assert(StoreSizeInDW);
- return store_size_in_dw;
+ return StoreSizeInDW;
}
-void R600KernelParameters::RunAna(Function* fun)
-{
- assert(isOpenCLKernel(fun));
+void R600KernelParameters::RunAna(Function* Fun) {
+ assert(IsOpenCLKernel(Fun));
- for (Function::arg_iterator i = fun->arg_begin(); i != fun->arg_end(); i++)
- {
- AddParam(i);
+ for (Function::arg_iterator I = Fun->arg_begin(); I != Fun->arg_end(); ++I) {
+ AddParam(I);
}
}
-void R600KernelParameters::Replace(Function* fun)
-{
- for (std::vector<param>::iterator i = params.begin(); i != params.end(); i++)
- {
- Value *new_val;
+void R600KernelParameters::Replace(Function* Fun) {
+ for (std::vector<Param>::iterator I = Params.begin(); I != Params.end(); ++I) {
+ Value *NewVal;
- if (isSpecialType(i->val->getType()))
- {
- new_val = handleSpecial(fun, *i);
- }
- else
- {
- new_val = ConstantRead(fun, *i);
+ if (IsSpecialType(I->Val->getType())) {
+ NewVal = handleSpecial(Fun, *I);
+ } else {
+ NewVal = ConstantRead(Fun, *I);
}
- if (new_val)
- {
- i->val->replaceAllUsesWith(new_val);
+ if (NewVal) {
+ I->Val->replaceAllUsesWith(NewVal);
}
}
}
-void R600KernelParameters::Propagate(Function* fun)
-{
- for (std::vector<param>::iterator i = params.begin(); i != params.end(); i++)
- {
- if (i->ptr_val)
- {
- Propagate(i->ptr_val, i->val->getName(), i->indirect);
- }
+void R600KernelParameters::Propagate(Function* Fun) {
+ for (std::vector<Param>::iterator I = Params.begin(); I != Params.end(); ++I) {
+ if (I->PtrVal) {
+ Propagate(I->PtrVal, I->Val->getName(), I->IsIndirect);
+ }
}
}
-void R600KernelParameters::Propagate(Value* v, const Twine& name, bool indirect)
-{
- LoadInst* load = dyn_cast<LoadInst>(v);
- GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(v);
+void R600KernelParameters::Propagate(Value* V, const Twine& Name, bool IsIndirect) {
+ LoadInst* Load = dyn_cast<LoadInst>(V);
+ GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V);
- unsigned addrspace;
+ unsigned Addrspace;
- if (indirect)
- {
- addrspace = AMDILAS::PARAM_I_ADDRESS;
- }
- else
- {
- addrspace = AMDILAS::PARAM_D_ADDRESS;
+ if (IsIndirect) {
+ Addrspace = AMDILAS::PARAM_I_ADDRESS;
+ } else {
+ Addrspace = AMDILAS::PARAM_D_ADDRESS;
}
- if (GEP and GEP->getType()->getAddressSpace() != addrspace)
- {
- Value* op = GEP->getPointerOperand();
+ if (GEP and GEP->getType()->getAddressSpace() != Addrspace) {
+ Value *Op = GEP->getPointerOperand();
- if (dyn_cast<PointerType>(op->getType())->getAddressSpace() != addrspace)
- {
- op = new BitCastInst(op, PointerType::get(dyn_cast<PointerType>(
- op->getType())->getElementType(), addrspace),
- name, dyn_cast<Instruction>(v));
+ if (dyn_cast<PointerType>(Op->getType())->getAddressSpace() != Addrspace) {
+ Op = new BitCastInst(Op, PointerType::get(dyn_cast<PointerType>(
+ Op->getType())->getElementType(), Addrspace),
+ Name, dyn_cast<Instruction>(V));
}
- std::vector<Value*> params(GEP->idx_begin(), GEP->idx_end());
+ std::vector<Value*> Params(GEP->idx_begin(), GEP->idx_end());
- GetElementPtrInst* GEP2 = GetElementPtrInst::Create(op, params, name,
- dyn_cast<Instruction>(v));
+ GetElementPtrInst* GEP2 = GetElementPtrInst::Create(Op, Params, Name,
+ dyn_cast<Instruction>(V));
GEP2->setIsInBounds(GEP->isInBounds());
- v = dyn_cast<Value>(GEP2);
+ V = dyn_cast<Value>(GEP2);
GEP->replaceAllUsesWith(GEP2);
GEP->eraseFromParent();
- load = NULL;
+ Load = NULL;
}
- if (load)
- {
+ if (Load) {
///normally at this point we have the right address space
- if (load->getPointerAddressSpace() != addrspace)
- {
- Value *orig_ptr = load->getPointerOperand();
- PointerType *orig_ptr_type = dyn_cast<PointerType>(orig_ptr->getType());
+ if (Load->getPointerAddressSpace() != Addrspace) {
+ Value *OrigPtr = Load->getPointerOperand();
+ PointerType *OrigPtrType = dyn_cast<PointerType>(OrigPtr->getType());
- Type* new_ptr_type = PointerType::get(orig_ptr_type->getElementType(),
- addrspace);
+ Type* NewPtrType = PointerType::get(OrigPtrType->getElementType(),
+ Addrspace);
- Value* new_ptr = orig_ptr;
+ Value* NewPtr = OrigPtr;
- if (orig_ptr->getType() != new_ptr_type)
- {
- new_ptr = new BitCastInst(orig_ptr, new_ptr_type, "prop_cast", load);
+ if (OrigPtr->getType() != NewPtrType) {
+ NewPtr = new BitCastInst(OrigPtr, NewPtrType, "prop_cast", Load);
}
- Value* new_load = new LoadInst(new_ptr, name, load);
- load->replaceAllUsesWith(new_load);
- load->eraseFromParent();
+ Value* new_Load = new LoadInst(NewPtr, Name, Load);
+ Load->replaceAllUsesWith(new_Load);
+ Load->eraseFromParent();
}
return;
}
- std::vector<User*> users(v->use_begin(), v->use_end());
+ std::vector<User*> Users(V->use_begin(), V->use_end());
- for (int i = 0; i < int(users.size()); i++)
- {
- Value* v2 = dyn_cast<Value>(users[i]);
+ for (int i = 0; i < int(Users.size()); i++) {
+ Value* V2 = dyn_cast<Value>(Users[i]);
- if (v2)
- {
- Propagate(v2, name, indirect);
+ if (V2) {
+ Propagate(V2, Name, IsIndirect);
}
}
}
-Value* R600KernelParameters::ConstantRead(Function* fun, param& p)
-{
- assert(fun->front().begin() != fun->front().end());
+Value* R600KernelParameters::ConstantRead(Function *Fun, Param &P) {
+ assert(Fun->front().begin() != Fun->front().end());
- Instruction *first_inst = fun->front().begin();
- IRBuilder <> builder (first_inst);
+ Instruction *FirstInst = Fun->front().begin();
+ IRBuilder <> Builder (FirstInst);
/* First 3 dwords are reserved for the dimmension info */
- if (!p.val->hasNUsesOrMore(1))
- {
+ if (!P.Val->hasNUsesOrMore(1)) {
return NULL;
}
- unsigned addrspace;
+ unsigned Addrspace;
- if (p.indirect)
- {
- addrspace = AMDILAS::PARAM_I_ADDRESS;
- }
- else
- {
- addrspace = AMDILAS::PARAM_D_ADDRESS;
+ if (P.IsIndirect) {
+ Addrspace = AMDILAS::PARAM_I_ADDRESS;
+ } else {
+ Addrspace = AMDILAS::PARAM_D_ADDRESS;
}
- Argument *arg = dyn_cast<Argument>(p.val);
- Type * argType = p.val->getType();
- PointerType * argPtrType = dyn_cast<PointerType>(p.val->getType());
+ Argument *Arg = dyn_cast<Argument>(P.Val);
+ Type * ArgType = P.Val->getType();
+ PointerType * ArgPtrType = dyn_cast<PointerType>(P.Val->getType());
- if (argPtrType and arg->hasByValAttr())
- {
- Value* param_addr_space_ptr = ConstantPointerNull::get(
+ if (ArgPtrType and Arg->hasByValAttr()) {
+ Value* ParamAddrSpacePtr = ConstantPointerNull::get(
PointerType::get(Type::getInt32Ty(*Context),
- addrspace));
- Value* param_ptr = GetElementPtrInst::Create(param_addr_space_ptr,
+ Addrspace));
+ Value* ParamPtr = GetElementPtrInst::Create(ParamAddrSpacePtr,
ConstantInt::get(Type::getInt32Ty(*Context),
- p.get_rat_offset()), arg->getName(),
- first_inst);
- param_ptr = new BitCastInst(param_ptr,
- PointerType::get(argPtrType->getElementType(),
- addrspace),
- arg->getName(), first_inst);
- p.ptr_val = param_ptr;
- return param_ptr;
- }
- else
- {
- Value* param_addr_space_ptr = ConstantPointerNull::get(PointerType::get(
- argType, addrspace));
-
- Value* param_ptr = builder.CreateGEP(param_addr_space_ptr,
- ConstantInt::get(Type::getInt32Ty(*Context), p.get_rat_offset()),
- arg->getName());
-
- Value* param_value = builder.CreateLoad(param_ptr, arg->getName());
-
- return param_value;
+ P.getRatOffset()), Arg->getName(),
+ FirstInst);
+ ParamPtr = new BitCastInst(ParamPtr,
+ PointerType::get(ArgPtrType->getElementType(),
+ Addrspace),
+ Arg->getName(), FirstInst);
+ P.PtrVal = ParamPtr;
+ return ParamPtr;
+ } else {
+ Value *ParamAddrSpacePtr = ConstantPointerNull::get(PointerType::get(
+ ArgType, Addrspace));
+
+ Value *ParamPtr = Builder.CreateGEP(ParamAddrSpacePtr,
+ ConstantInt::get(Type::getInt32Ty(*Context), P.getRatOffset()),
+ Arg->getName());
+
+ Value *Param_Value = Builder.CreateLoad(ParamPtr, Arg->getName());
+
+ return Param_Value;
}
}
-Value* R600KernelParameters::handleSpecial(Function* fun, param& p)
-{
- std::string name = getSpecialTypeName(p.val->getType());
+Value* R600KernelParameters::handleSpecial(Function* Fun, Param& P) {
+ std::string Name = getSpecialTypeName(P.Val->getType());
int ID;
- assert(!name.empty());
+ assert(!Name.empty());
- if (name == "image2d_t" or name == "image3d_t")
- {
- int lastID = std::max(getLastSpecialID("image2d_t"),
+ if (Name == "image2d_t" or Name == "image3d_t") {
+ int LastID = std::max(getLastSpecialID("image2d_t"),
getLastSpecialID("image3d_t"));
- if (lastID == -1)
- {
+ if (LastID == -1) {
ID = 2; ///ID0 and ID1 are used internally by the driver
+ } else {
+ ID = LastID + 1;
}
- else
- {
- ID = lastID + 1;
- }
- }
- else if (name == "sampler_t")
- {
- int lastID = getLastSpecialID("sampler_t");
+ } else if (Name == "sampler_t") {
+ int LastID = getLastSpecialID("sampler_t");
- if (lastID == -1)
- {
+ if (LastID == -1) {
ID = 0;
+ } else {
+ ID = LastID + 1;
}
- else
- {
- ID = lastID + 1;
- }
- }
- else
- {
+ } else {
///TODO: give some error message
return NULL;
}
- p.specialType = name;
- p.specialID = ID;
+ P.SpecialType = Name;
+ P.SpecialID = ID;
- Instruction *first_inst = fun->front().begin();
+ Instruction *FirstInst = Fun->front().begin();
return new IntToPtrInst(ConstantInt::get(Type::getInt32Ty(*Context),
- p.specialID), p.val->getType(),
- "resourceID", first_inst);
+ P.SpecialID), P.Val->getType(),
+ "resourceID", FirstInst);
}
-bool R600KernelParameters::isSpecialType(Type* t)
-{
- return !getSpecialTypeName(t).empty();
+bool R600KernelParameters::IsSpecialType(Type* T) {
+ return !getSpecialTypeName(T).empty();
}
-std::string R600KernelParameters::getSpecialTypeName(Type* t)
-{
- PointerType *pt = dyn_cast<PointerType>(t);
- StructType *st = NULL;
+std::string R600KernelParameters::getSpecialTypeName(Type* T) {
+ PointerType *PT = dyn_cast<PointerType>(T);
+ StructType *ST = NULL;
- if (pt)
- {
- st = dyn_cast<StructType>(pt->getElementType());
+ if (PT) {
+ ST = dyn_cast<StructType>(PT->getElementType());
}
- if (st)
- {
- std::string prefix = "struct.opencl_builtin_type_";
+ if (ST) {
+ std::string Prefix = "struct.opencl_builtin_type_";
- std::string name = st->getName().str();
+ std::string Name = ST->getName().str();
- if (name.substr(0, prefix.length()) == prefix)
- {
- return name.substr(prefix.length(), name.length());
+ if (Name.substr(0, Prefix.length()) == Prefix) {
+ return Name.substr(Prefix.length(), Name.length());
}
}
}
-bool R600KernelParameters::runOnFunction (Function &F)
-{
- if (!isOpenCLKernel(&F))
- {
+bool R600KernelParameters::runOnFunction (Function &F) {
+ if (!IsOpenCLKernel(&F)) {
return false;
}
return false;
}
-void R600KernelParameters::getAnalysisUsage(AnalysisUsage &AU) const
-{
+void R600KernelParameters::getAnalysisUsage(AnalysisUsage &AU) const {
FunctionPass::getAnalysisUsage(AU);
AU.setPreservesAll();
}
-const char *R600KernelParameters::getPassName() const
-{
+const char *R600KernelParameters::getPassName() const {
return "OpenCL Kernel parameter conversion to memory";
}
-bool R600KernelParameters::doInitialization(Module &M)
-{
+bool R600KernelParameters::doInitialization(Module &M) {
Context = &M.getContext();
- mod = &M;
+ Mod = &M;
return false;
}
-bool R600KernelParameters::doFinalization(Module &M)
-{
+bool R600KernelParameters::doFinalization(Module &M) {
return false;
}
} // End anonymous namespace
-FunctionPass* llvm::createR600KernelParametersPass(const TargetData* TD)
-{
- FunctionPass *p = new R600KernelParameters(TD);
-
- return p;
+FunctionPass* llvm::createR600KernelParametersPass(const TargetData* TD) {
+ return new R600KernelParameters(TD);
}
-
-