1 //===-- NVPTXAsmPrinter.cpp - NVPTX LLVM assembly writer ------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains a printer that converts from our internal representation
11 // of machine-dependent LLVM code to NVPTX assembly language.
13 //===----------------------------------------------------------------------===//
15 #include "NVPTXAsmPrinter.h"
16 #include "InstPrinter/NVPTXInstPrinter.h"
17 #include "MCTargetDesc/NVPTXMCAsmInfo.h"
19 #include "NVPTXInstrInfo.h"
20 #include "NVPTXMCExpr.h"
21 #include "NVPTXMachineFunctionInfo.h"
22 #include "NVPTXRegisterInfo.h"
23 #include "NVPTXTargetMachine.h"
24 #include "NVPTXUtilities.h"
25 #include "cl_common_defines.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/CodeGen/Analysis.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineLoopInfo.h"
31 #include "llvm/CodeGen/MachineModuleInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/IR/DebugInfo.h"
34 #include "llvm/IR/DerivedTypes.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/GlobalVariable.h"
37 #include "llvm/IR/Mangler.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/MC/MCStreamer.h"
41 #include "llvm/MC/MCSymbol.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/FormattedStream.h"
45 #include "llvm/Support/Path.h"
46 #include "llvm/Support/TargetRegistry.h"
47 #include "llvm/Support/TimeValue.h"
48 #include "llvm/Target/TargetLoweringObjectFile.h"
49 #include "llvm/Transforms/Utils/UnrollLoop.h"
53 #define DEPOTNAME "__local_depot"
56 EmitLineNumbers("nvptx-emit-line-numbers", cl::Hidden,
57 cl::desc("NVPTX Specific: Emit Line numbers even without -G"),
61 InterleaveSrc("nvptx-emit-src", cl::ZeroOrMore, cl::Hidden,
62 cl::desc("NVPTX Specific: Emit source line in ptx file"),
66 /// DiscoverDependentGlobals - Return a set of GlobalVariables on which \p V
68 void DiscoverDependentGlobals(const Value *V,
69 DenseSet<const GlobalVariable *> &Globals) {
70 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
73 if (const User *U = dyn_cast<User>(V)) {
74 for (unsigned i = 0, e = U->getNumOperands(); i != e; ++i) {
75 DiscoverDependentGlobals(U->getOperand(i), Globals);
81 /// VisitGlobalVariableForEmission - Add \p GV to the list of GlobalVariable
82 /// instances to be emitted, but only after any dependents have been added
84 void VisitGlobalVariableForEmission(
85 const GlobalVariable *GV, SmallVectorImpl<const GlobalVariable *> &Order,
86 DenseSet<const GlobalVariable *> &Visited,
87 DenseSet<const GlobalVariable *> &Visiting) {
88 // Have we already visited this one?
89 if (Visited.count(GV))
92 // Do we have a circular dependency?
93 if (!Visiting.insert(GV).second)
94 report_fatal_error("Circular dependency found in global variable set");
96 // Make sure we visit all dependents first
97 DenseSet<const GlobalVariable *> Others;
98 for (unsigned i = 0, e = GV->getNumOperands(); i != e; ++i)
99 DiscoverDependentGlobals(GV->getOperand(i), Others);
101 for (DenseSet<const GlobalVariable *>::iterator I = Others.begin(),
104 VisitGlobalVariableForEmission(*I, Order, Visited, Visiting);
106 // Now we can visit ourself
113 void NVPTXAsmPrinter::emitLineNumberAsDotLoc(const MachineInstr &MI) {
114 if (!EmitLineNumbers)
119 DebugLoc curLoc = MI.getDebugLoc();
121 if (!prevDebugLoc && !curLoc)
124 if (prevDebugLoc == curLoc)
127 prevDebugLoc = curLoc;
132 DIScope Scope = cast_or_null<MDScope>(curLoc.getScope());
136 StringRef fileName(Scope.getFilename());
137 StringRef dirName(Scope.getDirectory());
138 SmallString<128> FullPathName = dirName;
139 if (!dirName.empty() && !sys::path::is_absolute(fileName)) {
140 sys::path::append(FullPathName, fileName);
141 fileName = FullPathName;
144 if (filenameMap.find(fileName) == filenameMap.end())
147 // Emit the line from the source file.
149 this->emitSrcInText(fileName, curLoc.getLine());
151 std::stringstream temp;
152 temp << "\t.loc " << filenameMap[fileName] << " " << curLoc.getLine()
153 << " " << curLoc.getCol();
154 OutStreamer.EmitRawText(temp.str());
157 void NVPTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
158 SmallString<128> Str;
159 raw_svector_ostream OS(Str);
160 if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == NVPTX::CUDA)
161 emitLineNumberAsDotLoc(*MI);
164 lowerToMCInst(MI, Inst);
165 EmitToStreamer(OutStreamer, Inst);
168 // Handle symbol backtracking for targets that do not support image handles
169 bool NVPTXAsmPrinter::lowerImageHandleOperand(const MachineInstr *MI,
170 unsigned OpNo, MCOperand &MCOp) {
171 const MachineOperand &MO = MI->getOperand(OpNo);
172 const MCInstrDesc &MCID = MI->getDesc();
174 if (MCID.TSFlags & NVPTXII::IsTexFlag) {
175 // This is a texture fetch, so operand 4 is a texref and operand 5 is
177 if (OpNo == 4 && MO.isImm()) {
178 lowerImageHandleSymbol(MO.getImm(), MCOp);
181 if (OpNo == 5 && MO.isImm() && !(MCID.TSFlags & NVPTXII::IsTexModeUnifiedFlag)) {
182 lowerImageHandleSymbol(MO.getImm(), MCOp);
187 } else if (MCID.TSFlags & NVPTXII::IsSuldMask) {
189 1 << (((MCID.TSFlags & NVPTXII::IsSuldMask) >> NVPTXII::IsSuldShift) - 1);
191 // For a surface load of vector size N, the Nth operand will be the surfref
192 if (OpNo == VecSize && MO.isImm()) {
193 lowerImageHandleSymbol(MO.getImm(), MCOp);
198 } else if (MCID.TSFlags & NVPTXII::IsSustFlag) {
199 // This is a surface store, so operand 0 is a surfref
200 if (OpNo == 0 && MO.isImm()) {
201 lowerImageHandleSymbol(MO.getImm(), MCOp);
206 } else if (MCID.TSFlags & NVPTXII::IsSurfTexQueryFlag) {
207 // This is a query, so operand 1 is a surfref/texref
208 if (OpNo == 1 && MO.isImm()) {
209 lowerImageHandleSymbol(MO.getImm(), MCOp);
219 void NVPTXAsmPrinter::lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp) {
221 TargetMachine &TM = const_cast<TargetMachine&>(MF->getTarget());
222 NVPTXTargetMachine &nvTM = static_cast<NVPTXTargetMachine&>(TM);
223 const NVPTXMachineFunctionInfo *MFI = MF->getInfo<NVPTXMachineFunctionInfo>();
224 const char *Sym = MFI->getImageHandleSymbol(Index);
225 std::string *SymNamePtr =
226 nvTM.getManagedStrPool()->getManagedString(Sym);
227 MCOp = GetSymbolRef(OutContext.GetOrCreateSymbol(
228 StringRef(SymNamePtr->c_str())));
231 void NVPTXAsmPrinter::lowerToMCInst(const MachineInstr *MI, MCInst &OutMI) {
232 OutMI.setOpcode(MI->getOpcode());
233 // Special: Do not mangle symbol operand of CALL_PROTOTYPE
234 if (MI->getOpcode() == NVPTX::CALL_PROTOTYPE) {
235 const MachineOperand &MO = MI->getOperand(0);
236 OutMI.addOperand(GetSymbolRef(
237 OutContext.GetOrCreateSymbol(Twine(MO.getSymbolName()))));
241 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
242 const MachineOperand &MO = MI->getOperand(i);
245 if (!nvptxSubtarget->hasImageHandles()) {
246 if (lowerImageHandleOperand(MI, i, MCOp)) {
247 OutMI.addOperand(MCOp);
252 if (lowerOperand(MO, MCOp))
253 OutMI.addOperand(MCOp);
257 bool NVPTXAsmPrinter::lowerOperand(const MachineOperand &MO,
259 switch (MO.getType()) {
260 default: llvm_unreachable("unknown operand type");
261 case MachineOperand::MO_Register:
262 MCOp = MCOperand::CreateReg(encodeVirtualRegister(MO.getReg()));
264 case MachineOperand::MO_Immediate:
265 MCOp = MCOperand::CreateImm(MO.getImm());
267 case MachineOperand::MO_MachineBasicBlock:
268 MCOp = MCOperand::CreateExpr(MCSymbolRefExpr::Create(
269 MO.getMBB()->getSymbol(), OutContext));
271 case MachineOperand::MO_ExternalSymbol:
272 MCOp = GetSymbolRef(GetExternalSymbolSymbol(MO.getSymbolName()));
274 case MachineOperand::MO_GlobalAddress:
275 MCOp = GetSymbolRef(getSymbol(MO.getGlobal()));
277 case MachineOperand::MO_FPImmediate: {
278 const ConstantFP *Cnt = MO.getFPImm();
279 APFloat Val = Cnt->getValueAPF();
281 switch (Cnt->getType()->getTypeID()) {
282 default: report_fatal_error("Unsupported FP type"); break;
283 case Type::FloatTyID:
284 MCOp = MCOperand::CreateExpr(
285 NVPTXFloatMCExpr::CreateConstantFPSingle(Val, OutContext));
287 case Type::DoubleTyID:
288 MCOp = MCOperand::CreateExpr(
289 NVPTXFloatMCExpr::CreateConstantFPDouble(Val, OutContext));
298 unsigned NVPTXAsmPrinter::encodeVirtualRegister(unsigned Reg) {
299 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
300 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
302 DenseMap<unsigned, unsigned> &RegMap = VRegMapping[RC];
303 unsigned RegNum = RegMap[Reg];
305 // Encode the register class in the upper 4 bits
306 // Must be kept in sync with NVPTXInstPrinter::printRegName
308 if (RC == &NVPTX::Int1RegsRegClass) {
310 } else if (RC == &NVPTX::Int16RegsRegClass) {
312 } else if (RC == &NVPTX::Int32RegsRegClass) {
314 } else if (RC == &NVPTX::Int64RegsRegClass) {
316 } else if (RC == &NVPTX::Float32RegsRegClass) {
318 } else if (RC == &NVPTX::Float64RegsRegClass) {
321 report_fatal_error("Bad register class");
324 // Insert the vreg number
325 Ret |= (RegNum & 0x0FFFFFFF);
328 // Some special-use registers are actually physical registers.
329 // Encode this as the register class ID of 0 and the real register ID.
330 return Reg & 0x0FFFFFFF;
334 MCOperand NVPTXAsmPrinter::GetSymbolRef(const MCSymbol *Symbol) {
336 Expr = MCSymbolRefExpr::Create(Symbol, MCSymbolRefExpr::VK_None,
338 return MCOperand::CreateExpr(Expr);
341 void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
342 const DataLayout *TD = TM.getDataLayout();
343 const TargetLowering *TLI = nvptxSubtarget->getTargetLowering();
345 Type *Ty = F->getReturnType();
347 bool isABI = (nvptxSubtarget->getSmVersion() >= 20);
349 if (Ty->getTypeID() == Type::VoidTyID)
355 if (Ty->isFloatingPointTy() || Ty->isIntegerTy()) {
357 if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty)) {
358 size = ITy->getBitWidth();
362 assert(Ty->isFloatingPointTy() && "Floating point type expected here");
363 size = Ty->getPrimitiveSizeInBits();
366 O << ".param .b" << size << " func_retval0";
367 } else if (isa<PointerType>(Ty)) {
368 O << ".param .b" << TLI->getPointerTy().getSizeInBits()
370 } else if ((Ty->getTypeID() == Type::StructTyID) || isa<VectorType>(Ty)) {
371 unsigned totalsz = TD->getTypeAllocSize(Ty);
372 unsigned retAlignment = 0;
373 if (!llvm::getAlign(*F, 0, retAlignment))
374 retAlignment = TD->getABITypeAlignment(Ty);
375 O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
378 llvm_unreachable("Unknown return type");
380 SmallVector<EVT, 16> vtparts;
381 ComputeValueVTs(*TLI, Ty, vtparts);
383 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
385 EVT elemtype = vtparts[i];
386 if (vtparts[i].isVector()) {
387 elems = vtparts[i].getVectorNumElements();
388 elemtype = vtparts[i].getVectorElementType();
391 for (unsigned j = 0, je = elems; j != je; ++j) {
392 unsigned sz = elemtype.getSizeInBits();
393 if (elemtype.isInteger() && (sz < 32))
395 O << ".reg .b" << sz << " func_retval" << idx;
408 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
410 const Function *F = MF.getFunction();
411 printReturnValStr(F, O);
414 // Return true if MBB is the header of a loop marked with
415 // llvm.loop.unroll.disable.
416 // TODO: consider "#pragma unroll 1" which is equivalent to "#pragma nounroll".
417 bool NVPTXAsmPrinter::isLoopHeaderOfNoUnroll(
418 const MachineBasicBlock &MBB) const {
419 MachineLoopInfo &LI = getAnalysis<MachineLoopInfo>();
420 // TODO: isLoopHeader() should take "const MachineBasicBlock *".
421 // We insert .pragma "nounroll" only to the loop header.
422 if (!LI.isLoopHeader(const_cast<MachineBasicBlock *>(&MBB)))
425 // llvm.loop.unroll.disable is marked on the back edges of a loop. Therefore,
426 // we iterate through each back edge of the loop with header MBB, and check
427 // whether its metadata contains llvm.loop.unroll.disable.
428 for (auto I = MBB.pred_begin(); I != MBB.pred_end(); ++I) {
429 const MachineBasicBlock *PMBB = *I;
430 if (LI.getLoopFor(PMBB) != LI.getLoopFor(&MBB)) {
431 // Edges from other loops to MBB are not back edges.
434 if (const BasicBlock *PBB = PMBB->getBasicBlock()) {
435 if (MDNode *LoopID = PBB->getTerminator()->getMetadata("llvm.loop")) {
436 if (GetUnrollMetadata(LoopID, "llvm.loop.unroll.disable"))
444 void NVPTXAsmPrinter::EmitBasicBlockStart(const MachineBasicBlock &MBB) const {
445 AsmPrinter::EmitBasicBlockStart(MBB);
446 if (isLoopHeaderOfNoUnroll(MBB))
447 OutStreamer.EmitRawText(StringRef("\t.pragma \"nounroll\";\n"));
450 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
451 SmallString<128> Str;
452 raw_svector_ostream O(Str);
454 if (!GlobalsEmitted) {
455 emitGlobals(*MF->getFunction()->getParent());
456 GlobalsEmitted = true;
460 MRI = &MF->getRegInfo();
461 F = MF->getFunction();
462 emitLinkageDirective(F, O);
463 if (llvm::isKernelFunction(*F))
467 printReturnValStr(*MF, O);
472 emitFunctionParamList(*MF, O);
474 if (llvm::isKernelFunction(*F))
475 emitKernelFunctionDirectives(*F, O);
477 OutStreamer.EmitRawText(O.str());
479 prevDebugLoc = DebugLoc();
482 void NVPTXAsmPrinter::EmitFunctionBodyStart() {
484 OutStreamer.EmitRawText(StringRef("{\n"));
485 setAndEmitFunctionVirtualRegisters(*MF);
487 SmallString<128> Str;
488 raw_svector_ostream O(Str);
489 emitDemotedVars(MF->getFunction(), O);
490 OutStreamer.EmitRawText(O.str());
493 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
494 OutStreamer.EmitRawText(StringRef("}\n"));
498 void NVPTXAsmPrinter::emitImplicitDef(const MachineInstr *MI) const {
499 unsigned RegNo = MI->getOperand(0).getReg();
500 if (TargetRegisterInfo::isVirtualRegister(RegNo)) {
501 OutStreamer.AddComment(Twine("implicit-def: ") +
502 getVirtualRegisterName(RegNo));
504 OutStreamer.AddComment(Twine("implicit-def: ") +
505 nvptxSubtarget->getRegisterInfo()->getName(RegNo));
507 OutStreamer.AddBlankLine();
510 void NVPTXAsmPrinter::emitKernelFunctionDirectives(const Function &F,
511 raw_ostream &O) const {
512 // If the NVVM IR has some of reqntid* specified, then output
513 // the reqntid directive, and set the unspecified ones to 1.
514 // If none of reqntid* is specified, don't output reqntid directive.
515 unsigned reqntidx, reqntidy, reqntidz;
516 bool specified = false;
517 if (!llvm::getReqNTIDx(F, reqntidx))
521 if (!llvm::getReqNTIDy(F, reqntidy))
525 if (!llvm::getReqNTIDz(F, reqntidz))
531 O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
534 // If the NVVM IR has some of maxntid* specified, then output
535 // the maxntid directive, and set the unspecified ones to 1.
536 // If none of maxntid* is specified, don't output maxntid directive.
537 unsigned maxntidx, maxntidy, maxntidz;
539 if (!llvm::getMaxNTIDx(F, maxntidx))
543 if (!llvm::getMaxNTIDy(F, maxntidy))
547 if (!llvm::getMaxNTIDz(F, maxntidz))
553 O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
557 if (llvm::getMinCTASm(F, mincta))
558 O << ".minnctapersm " << mincta << "\n";
562 NVPTXAsmPrinter::getVirtualRegisterName(unsigned Reg) const {
563 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
566 raw_string_ostream NameStr(Name);
568 VRegRCMap::const_iterator I = VRegMapping.find(RC);
569 assert(I != VRegMapping.end() && "Bad register class");
570 const DenseMap<unsigned, unsigned> &RegMap = I->second;
572 VRegMap::const_iterator VI = RegMap.find(Reg);
573 assert(VI != RegMap.end() && "Bad virtual register");
574 unsigned MappedVR = VI->second;
576 NameStr << getNVPTXRegClassStr(RC) << MappedVR;
582 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr,
584 O << getVirtualRegisterName(vr);
587 void NVPTXAsmPrinter::printVecModifiedImmediate(
588 const MachineOperand &MO, const char *Modifier, raw_ostream &O) {
589 static const char vecelem[] = { '0', '1', '2', '3', '0', '1', '2', '3' };
590 int Imm = (int) MO.getImm();
591 if (0 == strcmp(Modifier, "vecelem"))
592 O << "_" << vecelem[Imm];
593 else if (0 == strcmp(Modifier, "vecv4comm1")) {
594 if ((Imm < 0) || (Imm > 3))
596 } else if (0 == strcmp(Modifier, "vecv4comm2")) {
597 if ((Imm < 4) || (Imm > 7))
599 } else if (0 == strcmp(Modifier, "vecv4pos")) {
602 O << "_" << vecelem[Imm % 4];
603 } else if (0 == strcmp(Modifier, "vecv2comm1")) {
604 if ((Imm < 0) || (Imm > 1))
606 } else if (0 == strcmp(Modifier, "vecv2comm2")) {
607 if ((Imm < 2) || (Imm > 3))
609 } else if (0 == strcmp(Modifier, "vecv2pos")) {
612 O << "_" << vecelem[Imm % 2];
614 llvm_unreachable("Unknown Modifier on immediate operand");
619 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
621 emitLinkageDirective(F, O);
622 if (llvm::isKernelFunction(*F))
626 printReturnValStr(F, O);
627 O << *getSymbol(F) << "\n";
628 emitFunctionParamList(F, O);
632 static bool usedInGlobalVarDef(const Constant *C) {
636 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
637 if (GV->getName() == "llvm.used")
642 for (const User *U : C->users())
643 if (const Constant *C = dyn_cast<Constant>(U))
644 if (usedInGlobalVarDef(C))
650 static bool usedInOneFunc(const User *U, Function const *&oneFunc) {
651 if (const GlobalVariable *othergv = dyn_cast<GlobalVariable>(U)) {
652 if (othergv->getName() == "llvm.used")
656 if (const Instruction *instr = dyn_cast<Instruction>(U)) {
657 if (instr->getParent() && instr->getParent()->getParent()) {
658 const Function *curFunc = instr->getParent()->getParent();
659 if (oneFunc && (curFunc != oneFunc))
667 for (const User *UU : U->users())
668 if (!usedInOneFunc(UU, oneFunc))
674 /* Find out if a global variable can be demoted to local scope.
675 * Currently, this is valid for CUDA shared variables, which have local
676 * scope and global lifetime. So the conditions to check are :
677 * 1. Is the global variable in shared address space?
678 * 2. Does it have internal linkage?
679 * 3. Is the global variable referenced only in one function?
681 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
682 if (!gv->hasInternalLinkage())
684 const PointerType *Pty = gv->getType();
685 if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
688 const Function *oneFunc = nullptr;
690 bool flag = usedInOneFunc(gv, oneFunc);
699 static bool useFuncSeen(const Constant *C,
700 llvm::DenseMap<const Function *, bool> &seenMap) {
701 for (const User *U : C->users()) {
702 if (const Constant *cu = dyn_cast<Constant>(U)) {
703 if (useFuncSeen(cu, seenMap))
705 } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
706 const BasicBlock *bb = I->getParent();
709 const Function *caller = bb->getParent();
712 if (seenMap.find(caller) != seenMap.end())
719 void NVPTXAsmPrinter::emitDeclarations(const Module &M, raw_ostream &O) {
720 llvm::DenseMap<const Function *, bool> seenMap;
721 for (Module::const_iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
722 const Function *F = FI;
724 if (F->isDeclaration()) {
727 if (F->getIntrinsicID())
729 emitDeclaration(F, O);
732 for (const User *U : F->users()) {
733 if (const Constant *C = dyn_cast<Constant>(U)) {
734 if (usedInGlobalVarDef(C)) {
735 // The use is in the initialization of a global variable
736 // that is a function pointer, so print a declaration
737 // for the original function
738 emitDeclaration(F, O);
741 // Emit a declaration of this function if the function that
742 // uses this constant expr has already been seen.
743 if (useFuncSeen(C, seenMap)) {
744 emitDeclaration(F, O);
749 if (!isa<Instruction>(U))
751 const Instruction *instr = cast<Instruction>(U);
752 const BasicBlock *bb = instr->getParent();
755 const Function *caller = bb->getParent();
759 // If a caller has already been seen, then the caller is
760 // appearing in the module before the callee. so print out
761 // a declaration for the callee.
762 if (seenMap.find(caller) != seenMap.end()) {
763 emitDeclaration(F, O);
771 void NVPTXAsmPrinter::recordAndEmitFilenames(Module &M) {
772 DebugInfoFinder DbgFinder;
773 DbgFinder.processModule(M);
776 for (const MDCompileUnit *DIUnit : DbgFinder.compile_units()) {
777 StringRef Filename = DIUnit->getFilename();
778 StringRef Dirname = DIUnit->getDirectory();
779 SmallString<128> FullPathName = Dirname;
780 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
781 sys::path::append(FullPathName, Filename);
782 Filename = FullPathName;
784 if (filenameMap.find(Filename) != filenameMap.end())
786 filenameMap[Filename] = i;
787 OutStreamer.EmitDwarfFileDirective(i, "", Filename);
791 for (MDSubprogram *SP : DbgFinder.subprograms()) {
792 StringRef Filename = SP->getFilename();
793 StringRef Dirname = SP->getDirectory();
794 SmallString<128> FullPathName = Dirname;
795 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
796 sys::path::append(FullPathName, Filename);
797 Filename = FullPathName;
799 if (filenameMap.find(Filename) != filenameMap.end())
801 filenameMap[Filename] = i;
806 bool NVPTXAsmPrinter::doInitialization(Module &M) {
807 // Construct a default subtarget off of the TargetMachine defaults. The
808 // rest of NVPTX isn't friendly to change subtargets per function and
809 // so the default TargetMachine will have all of the options.
810 StringRef TT = TM.getTargetTriple();
811 StringRef CPU = TM.getTargetCPU();
812 StringRef FS = TM.getTargetFeatureString();
813 const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
814 const NVPTXSubtarget STI(TT, CPU, FS, NTM);
816 SmallString<128> Str1;
817 raw_svector_ostream OS1(Str1);
819 MMI = getAnalysisIfAvailable<MachineModuleInfo>();
820 MMI->AnalyzeModule(M);
822 // We need to call the parent's one explicitly.
823 //bool Result = AsmPrinter::doInitialization(M);
825 // Initialize TargetLoweringObjectFile.
826 const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
827 .Initialize(OutContext, TM);
829 Mang = new Mangler(TM.getDataLayout());
831 // Emit header before any dwarf directives are emitted below.
832 emitHeader(M, OS1, STI);
833 OutStreamer.EmitRawText(OS1.str());
835 // Already commented out
836 //bool Result = AsmPrinter::doInitialization(M);
838 // Emit module-level inline asm if it exists.
839 if (!M.getModuleInlineAsm().empty()) {
840 OutStreamer.AddComment("Start of file scope inline assembly");
841 OutStreamer.AddBlankLine();
842 OutStreamer.EmitRawText(StringRef(M.getModuleInlineAsm()));
843 OutStreamer.AddBlankLine();
844 OutStreamer.AddComment("End of file scope inline assembly");
845 OutStreamer.AddBlankLine();
848 // If we're not NVCL we're CUDA, go ahead and emit filenames.
849 if (Triple(TM.getTargetTriple()).getOS() != Triple::NVCL)
850 recordAndEmitFilenames(M);
852 GlobalsEmitted = false;
854 return false; // success
857 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
858 SmallString<128> Str2;
859 raw_svector_ostream OS2(Str2);
861 emitDeclarations(M, OS2);
863 // As ptxas does not support forward references of globals, we need to first
864 // sort the list of module-level globals in def-use order. We visit each
865 // global variable in order, and ensure that we emit it *after* its dependent
866 // globals. We use a little extra memory maintaining both a set and a list to
867 // have fast searches while maintaining a strict ordering.
868 SmallVector<const GlobalVariable *, 8> Globals;
869 DenseSet<const GlobalVariable *> GVVisited;
870 DenseSet<const GlobalVariable *> GVVisiting;
872 // Visit each global variable, in order
873 for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
875 VisitGlobalVariableForEmission(I, Globals, GVVisited, GVVisiting);
877 assert(GVVisited.size() == M.getGlobalList().size() &&
878 "Missed a global variable");
879 assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
881 // Print out module-level global variables in proper order
882 for (unsigned i = 0, e = Globals.size(); i != e; ++i)
883 printModuleLevelGV(Globals[i], OS2);
887 OutStreamer.EmitRawText(OS2.str());
890 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O,
891 const NVPTXSubtarget &STI) {
893 O << "// Generated by LLVM NVPTX Back-End\n";
897 unsigned PTXVersion = STI.getPTXVersion();
898 O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
901 O << STI.getTargetName();
903 const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
904 if (NTM.getDrvInterface() == NVPTX::NVCL)
905 O << ", texmode_independent";
907 if (!STI.hasDouble())
908 O << ", map_f64_to_f32";
911 if (MAI->doesSupportDebugInformation())
916 O << ".address_size ";
926 bool NVPTXAsmPrinter::doFinalization(Module &M) {
927 // If we did not emit any functions, then the global declarations have not
929 if (!GlobalsEmitted) {
931 GlobalsEmitted = true;
934 // XXX Temproarily remove global variables so that doFinalization() will not
935 // emit them again (global variables are emitted at beginning).
937 Module::GlobalListType &global_list = M.getGlobalList();
938 int i, n = global_list.size();
939 GlobalVariable **gv_array = new GlobalVariable *[n];
941 // first, back-up GlobalVariable in gv_array
943 for (Module::global_iterator I = global_list.begin(), E = global_list.end();
947 // second, empty global_list
948 while (!global_list.empty())
949 global_list.remove(global_list.begin());
951 // call doFinalization
952 bool ret = AsmPrinter::doFinalization(M);
954 // now we restore global variables
955 for (i = 0; i < n; i++)
956 global_list.insert(global_list.end(), gv_array[i]);
958 clearAnnotationCache(&M);
963 //bool Result = AsmPrinter::doFinalization(M);
964 // Instead of calling the parents doFinalization, we may
965 // clone parents doFinalization and customize here.
966 // Currently, we if NVISA out the EmitGlobals() in
967 // parent's doFinalization, which is too intrusive.
969 // Same for the doInitialization.
973 // This function emits appropriate linkage directives for
974 // functions and global variables.
976 // extern function declaration -> .extern
977 // extern function definition -> .visible
978 // external global variable with init -> .visible
979 // external without init -> .extern
980 // appending -> not allowed, assert.
981 // for any linkage other than
982 // internal, private, linker_private,
983 // linker_private_weak, linker_private_weak_def_auto,
986 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
988 if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == NVPTX::CUDA) {
989 if (V->hasExternalLinkage()) {
990 if (isa<GlobalVariable>(V)) {
991 const GlobalVariable *GVar = cast<GlobalVariable>(V);
993 if (GVar->hasInitializer())
998 } else if (V->isDeclaration())
1002 } else if (V->hasAppendingLinkage()) {
1004 msg.append("Error: ");
1005 msg.append("Symbol ");
1007 msg.append(V->getName());
1008 msg.append("has unsupported appending linkage type");
1009 llvm_unreachable(msg.c_str());
1010 } else if (!V->hasInternalLinkage() &&
1011 !V->hasPrivateLinkage()) {
1017 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
1019 bool processDemoted) {
1022 if (GVar->hasSection()) {
1023 if (GVar->getSection() == StringRef("llvm.metadata"))
1027 // Skip LLVM intrinsic global variables
1028 if (GVar->getName().startswith("llvm.") ||
1029 GVar->getName().startswith("nvvm."))
1032 const DataLayout *TD = TM.getDataLayout();
1034 // GlobalVariables are always constant pointers themselves.
1035 const PointerType *PTy = GVar->getType();
1036 Type *ETy = PTy->getElementType();
1038 if (GVar->hasExternalLinkage()) {
1039 if (GVar->hasInitializer())
1043 } else if (GVar->hasLinkOnceLinkage() || GVar->hasWeakLinkage() ||
1044 GVar->hasAvailableExternallyLinkage() ||
1045 GVar->hasCommonLinkage()) {
1049 if (llvm::isTexture(*GVar)) {
1050 O << ".global .texref " << llvm::getTextureName(*GVar) << ";\n";
1054 if (llvm::isSurface(*GVar)) {
1055 O << ".global .surfref " << llvm::getSurfaceName(*GVar) << ";\n";
1059 if (GVar->isDeclaration()) {
1060 // (extern) declarations, no definition or initializer
1061 // Currently the only known declaration is for an automatic __local
1062 // (.shared) promoted to global.
1063 emitPTXGlobalVariable(GVar, O);
1068 if (llvm::isSampler(*GVar)) {
1069 O << ".global .samplerref " << llvm::getSamplerName(*GVar);
1071 const Constant *Initializer = nullptr;
1072 if (GVar->hasInitializer())
1073 Initializer = GVar->getInitializer();
1074 const ConstantInt *CI = nullptr;
1076 CI = dyn_cast<ConstantInt>(Initializer);
1078 unsigned sample = CI->getZExtValue();
1083 addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1085 O << "addr_mode_" << i << " = ";
1091 O << "clamp_to_border";
1094 O << "clamp_to_edge";
1105 O << "filter_mode = ";
1106 switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1114 llvm_unreachable("Anisotropic filtering is not supported");
1119 if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1120 O << ", force_unnormalized_coords = 1";
1129 if (GVar->hasPrivateLinkage()) {
1131 if (!strncmp(GVar->getName().data(), "unrollpragma", 12))
1134 // FIXME - need better way (e.g. Metadata) to avoid generating this global
1135 if (!strncmp(GVar->getName().data(), "filename", 8))
1137 if (GVar->use_empty())
1141 const Function *demotedFunc = nullptr;
1142 if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
1143 O << "// " << GVar->getName() << " has been demoted\n";
1144 if (localDecls.find(demotedFunc) != localDecls.end())
1145 localDecls[demotedFunc].push_back(GVar);
1147 std::vector<const GlobalVariable *> temp;
1148 temp.push_back(GVar);
1149 localDecls[demotedFunc] = temp;
1155 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1157 if (isManaged(*GVar)) {
1158 O << " .attribute(.managed)";
1161 if (GVar->getAlignment() == 0)
1162 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1164 O << " .align " << GVar->getAlignment();
1166 if (ETy->isFloatingPointTy() || ETy->isIntegerTy() || ETy->isPointerTy()) {
1168 // Special case: ABI requires that we use .u8 for predicates
1169 if (ETy->isIntegerTy(1))
1172 O << getPTXFundamentalTypeStr(ETy, false);
1174 O << *getSymbol(GVar);
1176 // Ptx allows variable initilization only for constant and global state
1178 if (GVar->hasInitializer()) {
1179 if ((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1180 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) {
1181 const Constant *Initializer = GVar->getInitializer();
1182 // 'undef' is treated as there is no value spefied.
1183 if (!Initializer->isNullValue() && !isa<UndefValue>(Initializer)) {
1185 printScalarConstant(Initializer, O);
1188 // The frontend adds zero-initializer to variables that don't have an
1189 // initial value, so skip warning for this case.
1190 if (!GVar->getInitializer()->isNullValue()) {
1191 std::string warnMsg =
1192 ("initial value of '" + GVar->getName() +
1193 "' is not allowed in addrspace(" +
1194 Twine(llvm::utostr_32(PTy->getAddressSpace())) + ")").str();
1195 report_fatal_error(warnMsg.c_str());
1200 unsigned int ElementSize = 0;
1202 // Although PTX has direct support for struct type and array type and
1203 // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1204 // targets that support these high level field accesses. Structs, arrays
1205 // and vectors are lowered into arrays of bytes.
1206 switch (ETy->getTypeID()) {
1207 case Type::StructTyID:
1208 case Type::ArrayTyID:
1209 case Type::VectorTyID:
1210 ElementSize = TD->getTypeStoreSize(ETy);
1211 // Ptx allows variable initilization only for constant and
1212 // global state spaces.
1213 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1214 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1215 GVar->hasInitializer()) {
1216 const Constant *Initializer = GVar->getInitializer();
1217 if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1218 AggBuffer aggBuffer(ElementSize, O, *this);
1219 bufferAggregateConstant(Initializer, &aggBuffer);
1220 if (aggBuffer.numSymbols) {
1221 if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit()) {
1222 O << " .u64 " << *getSymbol(GVar) << "[";
1223 O << ElementSize / 8;
1225 O << " .u32 " << *getSymbol(GVar) << "[";
1226 O << ElementSize / 4;
1230 O << " .b8 " << *getSymbol(GVar) << "[";
1238 O << " .b8 " << *getSymbol(GVar);
1246 O << " .b8 " << *getSymbol(GVar);
1255 llvm_unreachable("type not supported yet");
1262 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1263 if (localDecls.find(f) == localDecls.end())
1266 std::vector<const GlobalVariable *> &gvars = localDecls[f];
1268 for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1269 O << "\t// demoted variable\n\t";
1270 printModuleLevelGV(gvars[i], O, true);
1274 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1275 raw_ostream &O) const {
1276 switch (AddressSpace) {
1277 case llvm::ADDRESS_SPACE_LOCAL:
1280 case llvm::ADDRESS_SPACE_GLOBAL:
1283 case llvm::ADDRESS_SPACE_CONST:
1286 case llvm::ADDRESS_SPACE_SHARED:
1290 report_fatal_error("Bad address space found while emitting PTX");
1296 NVPTXAsmPrinter::getPTXFundamentalTypeStr(const Type *Ty, bool useB4PTR) const {
1297 switch (Ty->getTypeID()) {
1299 llvm_unreachable("unexpected type");
1301 case Type::IntegerTyID: {
1302 unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1305 else if (NumBits <= 64) {
1306 std::string name = "u";
1307 return name + utostr(NumBits);
1309 llvm_unreachable("Integer too large");
1314 case Type::FloatTyID:
1316 case Type::DoubleTyID:
1318 case Type::PointerTyID:
1319 if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit())
1329 llvm_unreachable("unexpected type");
1333 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1336 const DataLayout *TD = TM.getDataLayout();
1338 // GlobalVariables are always constant pointers themselves.
1339 const PointerType *PTy = GVar->getType();
1340 Type *ETy = PTy->getElementType();
1343 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1344 if (GVar->getAlignment() == 0)
1345 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1347 O << " .align " << GVar->getAlignment();
1349 if (ETy->isFloatingPointTy() || ETy->isIntegerTy() || ETy->isPointerTy()) {
1351 O << getPTXFundamentalTypeStr(ETy);
1353 O << *getSymbol(GVar);
1357 int64_t ElementSize = 0;
1359 // Although PTX has direct support for struct type and array type and LLVM IR
1360 // is very similar to PTX, the LLVM CodeGen does not support for targets that
1361 // support these high level field accesses. Structs and arrays are lowered
1362 // into arrays of bytes.
1363 switch (ETy->getTypeID()) {
1364 case Type::StructTyID:
1365 case Type::ArrayTyID:
1366 case Type::VectorTyID:
1367 ElementSize = TD->getTypeStoreSize(ETy);
1368 O << " .b8 " << *getSymbol(GVar) << "[";
1370 O << itostr(ElementSize);
1375 llvm_unreachable("type not supported yet");
1380 static unsigned int getOpenCLAlignment(const DataLayout *TD, Type *Ty) {
1381 if (Ty->isSingleValueType())
1382 return TD->getPrefTypeAlignment(Ty);
1384 const ArrayType *ATy = dyn_cast<ArrayType>(Ty);
1386 return getOpenCLAlignment(TD, ATy->getElementType());
1388 const StructType *STy = dyn_cast<StructType>(Ty);
1390 unsigned int alignStruct = 1;
1391 // Go through each element of the struct and find the
1392 // largest alignment.
1393 for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1394 Type *ETy = STy->getElementType(i);
1395 unsigned int align = getOpenCLAlignment(TD, ETy);
1396 if (align > alignStruct)
1397 alignStruct = align;
1402 const FunctionType *FTy = dyn_cast<FunctionType>(Ty);
1404 return TD->getPointerPrefAlignment();
1405 return TD->getPrefTypeAlignment(Ty);
1408 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1409 int paramIndex, raw_ostream &O) {
1410 O << *getSymbol(I->getParent()) << "_param_" << paramIndex;
1413 void NVPTXAsmPrinter::printParamName(int paramIndex, raw_ostream &O) {
1414 O << *CurrentFnSym << "_param_" << paramIndex;
1417 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1418 const DataLayout *TD = TM.getDataLayout();
1419 const AttributeSet &PAL = F->getAttributes();
1420 const TargetLowering *TLI = nvptxSubtarget->getTargetLowering();
1421 Function::const_arg_iterator I, E;
1422 unsigned paramIndex = 0;
1424 bool isKernelFunc = llvm::isKernelFunction(*F);
1425 bool isABI = (nvptxSubtarget->getSmVersion() >= 20);
1426 MVT thePointerTy = TLI->getPointerTy();
1430 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1431 Type *Ty = I->getType();
1438 // Handle image/sampler parameters
1439 if (isKernelFunction(*F)) {
1440 if (isSampler(*I) || isImage(*I)) {
1442 std::string sname = I->getName();
1443 if (isImageWriteOnly(*I) || isImageReadWrite(*I)) {
1444 if (nvptxSubtarget->hasImageHandles())
1445 O << "\t.param .u64 .ptr .surfref ";
1447 O << "\t.param .surfref ";
1448 O << *CurrentFnSym << "_param_" << paramIndex;
1450 else { // Default image is read_only
1451 if (nvptxSubtarget->hasImageHandles())
1452 O << "\t.param .u64 .ptr .texref ";
1454 O << "\t.param .texref ";
1455 O << *CurrentFnSym << "_param_" << paramIndex;
1458 if (nvptxSubtarget->hasImageHandles())
1459 O << "\t.param .u64 .ptr .samplerref ";
1461 O << "\t.param .samplerref ";
1462 O << *CurrentFnSym << "_param_" << paramIndex;
1468 if (!PAL.hasAttribute(paramIndex + 1, Attribute::ByVal)) {
1469 if (Ty->isAggregateType() || Ty->isVectorTy()) {
1470 // Just print .param .align <a> .b8 .param[size];
1471 // <a> = PAL.getparamalignment
1472 // size = typeallocsize of element type
1473 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1475 align = TD->getABITypeAlignment(Ty);
1477 unsigned sz = TD->getTypeAllocSize(Ty);
1478 O << "\t.param .align " << align << " .b8 ";
1479 printParamName(I, paramIndex, O);
1480 O << "[" << sz << "]";
1485 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1488 // Special handling for pointer arguments to kernel
1489 O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1491 if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() !=
1493 Type *ETy = PTy->getElementType();
1494 int addrSpace = PTy->getAddressSpace();
1495 switch (addrSpace) {
1499 case llvm::ADDRESS_SPACE_CONST:
1500 O << ".ptr .const ";
1502 case llvm::ADDRESS_SPACE_SHARED:
1503 O << ".ptr .shared ";
1505 case llvm::ADDRESS_SPACE_GLOBAL:
1506 O << ".ptr .global ";
1509 O << ".align " << (int) getOpenCLAlignment(TD, ETy) << " ";
1511 printParamName(I, paramIndex, O);
1515 // non-pointer scalar to kernel func
1517 // Special case: predicate operands become .u8 types
1518 if (Ty->isIntegerTy(1))
1521 O << getPTXFundamentalTypeStr(Ty);
1523 printParamName(I, paramIndex, O);
1526 // Non-kernel function, just print .param .b<size> for ABI
1527 // and .reg .b<size> for non-ABI
1529 if (isa<IntegerType>(Ty)) {
1530 sz = cast<IntegerType>(Ty)->getBitWidth();
1533 } else if (isa<PointerType>(Ty))
1534 sz = thePointerTy.getSizeInBits();
1536 sz = Ty->getPrimitiveSizeInBits();
1538 O << "\t.param .b" << sz << " ";
1540 O << "\t.reg .b" << sz << " ";
1541 printParamName(I, paramIndex, O);
1545 // param has byVal attribute. So should be a pointer
1546 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1547 assert(PTy && "Param with byval attribute should be a pointer type");
1548 Type *ETy = PTy->getElementType();
1550 if (isABI || isKernelFunc) {
1551 // Just print .param .align <a> .b8 .param[size];
1552 // <a> = PAL.getparamalignment
1553 // size = typeallocsize of element type
1554 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1556 align = TD->getABITypeAlignment(ETy);
1558 unsigned sz = TD->getTypeAllocSize(ETy);
1559 O << "\t.param .align " << align << " .b8 ";
1560 printParamName(I, paramIndex, O);
1561 O << "[" << sz << "]";
1564 // Split the ETy into constituent parts and
1565 // print .param .b<size> <name> for each part.
1566 // Further, if a part is vector, print the above for
1567 // each vector element.
1568 SmallVector<EVT, 16> vtparts;
1569 ComputeValueVTs(*TLI, ETy, vtparts);
1570 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1572 EVT elemtype = vtparts[i];
1573 if (vtparts[i].isVector()) {
1574 elems = vtparts[i].getVectorNumElements();
1575 elemtype = vtparts[i].getVectorElementType();
1578 for (unsigned j = 0, je = elems; j != je; ++j) {
1579 unsigned sz = elemtype.getSizeInBits();
1580 if (elemtype.isInteger() && (sz < 32))
1582 O << "\t.reg .b" << sz << " ";
1583 printParamName(I, paramIndex, O);
1599 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1601 const Function *F = MF.getFunction();
1602 emitFunctionParamList(F, O);
1605 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1606 const MachineFunction &MF) {
1607 SmallString<128> Str;
1608 raw_svector_ostream O(Str);
1610 // Map the global virtual register number to a register class specific
1611 // virtual register number starting from 1 with that class.
1612 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1613 //unsigned numRegClasses = TRI->getNumRegClasses();
1615 // Emit the Fake Stack Object
1616 const MachineFrameInfo *MFI = MF.getFrameInfo();
1617 int NumBytes = (int) MFI->getStackSize();
1619 O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
1620 << getFunctionNumber() << "[" << NumBytes << "];\n";
1621 if (static_cast<const NVPTXTargetMachine &>(MF.getTarget()).is64Bit()) {
1622 O << "\t.reg .b64 \t%SP;\n";
1623 O << "\t.reg .b64 \t%SPL;\n";
1625 O << "\t.reg .b32 \t%SP;\n";
1626 O << "\t.reg .b32 \t%SPL;\n";
1630 // Go through all virtual registers to establish the mapping between the
1632 // register number and the per class virtual register number.
1633 // We use the per class virtual register number in the ptx output.
1634 unsigned int numVRs = MRI->getNumVirtRegs();
1635 for (unsigned i = 0; i < numVRs; i++) {
1636 unsigned int vr = TRI->index2VirtReg(i);
1637 const TargetRegisterClass *RC = MRI->getRegClass(vr);
1638 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1639 int n = regmap.size();
1640 regmap.insert(std::make_pair(vr, n + 1));
1643 // Emit register declarations
1644 // @TODO: Extract out the real register usage
1645 // O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
1646 // O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
1647 // O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
1648 // O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
1649 // O << "\t.reg .s64 %rd<" << NVPTXNumRegisters << ">;\n";
1650 // O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
1651 // O << "\t.reg .f64 %fd<" << NVPTXNumRegisters << ">;\n";
1653 // Emit declaration of the virtual registers or 'physical' registers for
1654 // each register class
1655 for (unsigned i=0; i< TRI->getNumRegClasses(); i++) {
1656 const TargetRegisterClass *RC = TRI->getRegClass(i);
1657 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1658 std::string rcname = getNVPTXRegClassName(RC);
1659 std::string rcStr = getNVPTXRegClassStr(RC);
1660 int n = regmap.size();
1662 // Only declare those registers that may be used.
1664 O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
1669 OutStreamer.EmitRawText(O.str());
1672 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
1673 APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
1675 unsigned int numHex;
1678 if (Fp->getType()->getTypeID() == Type::FloatTyID) {
1681 APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &ignored);
1682 } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
1685 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &ignored);
1687 llvm_unreachable("unsupported fp type");
1689 APInt API = APF.bitcastToAPInt();
1690 std::string hexstr(utohexstr(API.getZExtValue()));
1692 if (hexstr.length() < numHex)
1693 O << std::string(numHex - hexstr.length(), '0');
1694 O << utohexstr(API.getZExtValue());
1697 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
1698 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
1699 O << CI->getValue();
1702 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
1703 printFPConstant(CFP, O);
1706 if (isa<ConstantPointerNull>(CPV)) {
1710 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1711 PointerType *PTy = dyn_cast<PointerType>(GVar->getType());
1712 bool IsNonGenericPointer = false;
1713 if (PTy && PTy->getAddressSpace() != 0) {
1714 IsNonGenericPointer = true;
1716 if (EmitGeneric && !isa<Function>(CPV) && !IsNonGenericPointer) {
1718 O << *getSymbol(GVar);
1721 O << *getSymbol(GVar);
1725 if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1726 const Value *v = Cexpr->stripPointerCasts();
1727 PointerType *PTy = dyn_cast<PointerType>(Cexpr->getType());
1728 bool IsNonGenericPointer = false;
1729 if (PTy && PTy->getAddressSpace() != 0) {
1730 IsNonGenericPointer = true;
1732 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
1733 if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
1735 O << *getSymbol(GVar);
1738 O << *getSymbol(GVar);
1742 O << *lowerConstant(CPV);
1746 llvm_unreachable("Not scalar type found in printScalarConstant()");
1749 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
1750 AggBuffer *aggBuffer) {
1752 const DataLayout *TD = TM.getDataLayout();
1754 if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
1755 int s = TD->getTypeAllocSize(CPV->getType());
1758 aggBuffer->addZeros(s);
1763 switch (CPV->getType()->getTypeID()) {
1765 case Type::IntegerTyID: {
1766 const Type *ETy = CPV->getType();
1767 if (ETy == Type::getInt8Ty(CPV->getContext())) {
1768 unsigned char c = (unsigned char)cast<ConstantInt>(CPV)->getZExtValue();
1770 aggBuffer->addBytes(ptr, 1, Bytes);
1771 } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
1772 short int16 = (short)cast<ConstantInt>(CPV)->getZExtValue();
1773 ptr = (unsigned char *)&int16;
1774 aggBuffer->addBytes(ptr, 2, Bytes);
1775 } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
1776 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1777 int int32 = (int)(constInt->getZExtValue());
1778 ptr = (unsigned char *)&int32;
1779 aggBuffer->addBytes(ptr, 4, Bytes);
1781 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1782 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1783 ConstantFoldConstantExpression(Cexpr, *TD))) {
1784 int int32 = (int)(constInt->getZExtValue());
1785 ptr = (unsigned char *)&int32;
1786 aggBuffer->addBytes(ptr, 4, Bytes);
1789 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1790 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1791 aggBuffer->addSymbol(v);
1792 aggBuffer->addZeros(4);
1796 llvm_unreachable("unsupported integer const type");
1797 } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
1798 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1799 long long int64 = (long long)(constInt->getZExtValue());
1800 ptr = (unsigned char *)&int64;
1801 aggBuffer->addBytes(ptr, 8, Bytes);
1803 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1804 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1805 ConstantFoldConstantExpression(Cexpr, *TD))) {
1806 long long int64 = (long long)(constInt->getZExtValue());
1807 ptr = (unsigned char *)&int64;
1808 aggBuffer->addBytes(ptr, 8, Bytes);
1811 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1812 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1813 aggBuffer->addSymbol(v);
1814 aggBuffer->addZeros(8);
1818 llvm_unreachable("unsupported integer const type");
1820 llvm_unreachable("unsupported integer const type");
1823 case Type::FloatTyID:
1824 case Type::DoubleTyID: {
1825 const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
1826 const Type *Ty = CFP->getType();
1827 if (Ty == Type::getFloatTy(CPV->getContext())) {
1828 float float32 = (float) CFP->getValueAPF().convertToFloat();
1829 ptr = (unsigned char *)&float32;
1830 aggBuffer->addBytes(ptr, 4, Bytes);
1831 } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
1832 double float64 = CFP->getValueAPF().convertToDouble();
1833 ptr = (unsigned char *)&float64;
1834 aggBuffer->addBytes(ptr, 8, Bytes);
1836 llvm_unreachable("unsupported fp const type");
1840 case Type::PointerTyID: {
1841 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1842 aggBuffer->addSymbol(GVar);
1843 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1844 const Value *v = Cexpr->stripPointerCasts();
1845 aggBuffer->addSymbol(v);
1847 unsigned int s = TD->getTypeAllocSize(CPV->getType());
1848 aggBuffer->addZeros(s);
1852 case Type::ArrayTyID:
1853 case Type::VectorTyID:
1854 case Type::StructTyID: {
1855 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV) ||
1856 isa<ConstantStruct>(CPV) || isa<ConstantDataSequential>(CPV)) {
1857 int ElementSize = TD->getTypeAllocSize(CPV->getType());
1858 bufferAggregateConstant(CPV, aggBuffer);
1859 if (Bytes > ElementSize)
1860 aggBuffer->addZeros(Bytes - ElementSize);
1861 } else if (isa<ConstantAggregateZero>(CPV))
1862 aggBuffer->addZeros(Bytes);
1864 llvm_unreachable("Unexpected Constant type");
1869 llvm_unreachable("unsupported type");
1873 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
1874 AggBuffer *aggBuffer) {
1875 const DataLayout *TD = TM.getDataLayout();
1879 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
1880 if (CPV->getNumOperands())
1881 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
1882 bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
1886 if (const ConstantDataSequential *CDS =
1887 dyn_cast<ConstantDataSequential>(CPV)) {
1888 if (CDS->getNumElements())
1889 for (unsigned i = 0; i < CDS->getNumElements(); ++i)
1890 bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
1895 if (isa<ConstantStruct>(CPV)) {
1896 if (CPV->getNumOperands()) {
1897 StructType *ST = cast<StructType>(CPV->getType());
1898 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
1900 Bytes = TD->getStructLayout(ST)->getElementOffset(0) +
1901 TD->getTypeAllocSize(ST) -
1902 TD->getStructLayout(ST)->getElementOffset(i);
1904 Bytes = TD->getStructLayout(ST)->getElementOffset(i + 1) -
1905 TD->getStructLayout(ST)->getElementOffset(i);
1906 bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
1911 llvm_unreachable("unsupported constant type in printAggregateConstant()");
1914 // buildTypeNameMap - Run through symbol table looking for type names.
1917 bool NVPTXAsmPrinter::isImageType(const Type *Ty) {
1919 std::map<const Type *, std::string>::iterator PI = TypeNameMap.find(Ty);
1921 if (PI != TypeNameMap.end() && (!PI->second.compare("struct._image1d_t") ||
1922 !PI->second.compare("struct._image2d_t") ||
1923 !PI->second.compare("struct._image3d_t")))
1930 bool NVPTXAsmPrinter::ignoreLoc(const MachineInstr &MI) {
1931 switch (MI.getOpcode()) {
1934 case NVPTX::CallArgBeginInst:
1935 case NVPTX::CallArgEndInst0:
1936 case NVPTX::CallArgEndInst1:
1937 case NVPTX::CallArgF32:
1938 case NVPTX::CallArgF64:
1939 case NVPTX::CallArgI16:
1940 case NVPTX::CallArgI32:
1941 case NVPTX::CallArgI32imm:
1942 case NVPTX::CallArgI64:
1943 case NVPTX::CallArgParam:
1944 case NVPTX::CallVoidInst:
1945 case NVPTX::CallVoidInstReg:
1946 case NVPTX::Callseq_End:
1947 case NVPTX::CallVoidInstReg64:
1948 case NVPTX::DeclareParamInst:
1949 case NVPTX::DeclareRetMemInst:
1950 case NVPTX::DeclareRetRegInst:
1951 case NVPTX::DeclareRetScalarInst:
1952 case NVPTX::DeclareScalarParamInst:
1953 case NVPTX::DeclareScalarRegInst:
1954 case NVPTX::StoreParamF32:
1955 case NVPTX::StoreParamF64:
1956 case NVPTX::StoreParamI16:
1957 case NVPTX::StoreParamI32:
1958 case NVPTX::StoreParamI64:
1959 case NVPTX::StoreParamI8:
1960 case NVPTX::StoreRetvalF32:
1961 case NVPTX::StoreRetvalF64:
1962 case NVPTX::StoreRetvalI16:
1963 case NVPTX::StoreRetvalI32:
1964 case NVPTX::StoreRetvalI64:
1965 case NVPTX::StoreRetvalI8:
1966 case NVPTX::LastCallArgF32:
1967 case NVPTX::LastCallArgF64:
1968 case NVPTX::LastCallArgI16:
1969 case NVPTX::LastCallArgI32:
1970 case NVPTX::LastCallArgI32imm:
1971 case NVPTX::LastCallArgI64:
1972 case NVPTX::LastCallArgParam:
1973 case NVPTX::LoadParamMemF32:
1974 case NVPTX::LoadParamMemF64:
1975 case NVPTX::LoadParamMemI16:
1976 case NVPTX::LoadParamMemI32:
1977 case NVPTX::LoadParamMemI64:
1978 case NVPTX::LoadParamMemI8:
1979 case NVPTX::PrototypeInst:
1980 case NVPTX::DBG_VALUE:
1986 /// PrintAsmOperand - Print out an operand for an inline asm expression.
1988 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
1989 unsigned AsmVariant,
1990 const char *ExtraCode, raw_ostream &O) {
1991 if (ExtraCode && ExtraCode[0]) {
1992 if (ExtraCode[1] != 0)
1993 return true; // Unknown modifier.
1995 switch (ExtraCode[0]) {
1997 // See if this is a generic print operand
1998 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
2004 printOperand(MI, OpNo, O);
2009 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(
2010 const MachineInstr *MI, unsigned OpNo, unsigned AsmVariant,
2011 const char *ExtraCode, raw_ostream &O) {
2012 if (ExtraCode && ExtraCode[0])
2013 return true; // Unknown modifier
2016 printMemOperand(MI, OpNo, O);
2022 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
2023 raw_ostream &O, const char *Modifier) {
2024 const MachineOperand &MO = MI->getOperand(opNum);
2025 switch (MO.getType()) {
2026 case MachineOperand::MO_Register:
2027 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
2028 if (MO.getReg() == NVPTX::VRDepot)
2029 O << DEPOTNAME << getFunctionNumber();
2031 O << NVPTXInstPrinter::getRegisterName(MO.getReg());
2033 emitVirtualRegister(MO.getReg(), O);
2037 case MachineOperand::MO_Immediate:
2040 else if (strstr(Modifier, "vec") == Modifier)
2041 printVecModifiedImmediate(MO, Modifier, O);
2044 "Don't know how to handle modifier on immediate operand");
2047 case MachineOperand::MO_FPImmediate:
2048 printFPConstant(MO.getFPImm(), O);
2051 case MachineOperand::MO_GlobalAddress:
2052 O << *getSymbol(MO.getGlobal());
2055 case MachineOperand::MO_MachineBasicBlock:
2056 O << *MO.getMBB()->getSymbol();
2060 llvm_unreachable("Operand type not supported.");
2064 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
2065 raw_ostream &O, const char *Modifier) {
2066 printOperand(MI, opNum, O);
2068 if (Modifier && !strcmp(Modifier, "add")) {
2070 printOperand(MI, opNum + 1, O);
2072 if (MI->getOperand(opNum + 1).isImm() &&
2073 MI->getOperand(opNum + 1).getImm() == 0)
2074 return; // don't print ',0' or '+0'
2076 printOperand(MI, opNum + 1, O);
2080 void NVPTXAsmPrinter::emitSrcInText(StringRef filename, unsigned line) {
2081 std::stringstream temp;
2082 LineReader *reader = this->getReader(filename);
2084 temp << filename.str();
2088 temp << reader->readLine(line);
2090 this->OutStreamer.EmitRawText(temp.str());
2093 LineReader *NVPTXAsmPrinter::getReader(std::string filename) {
2095 reader = new LineReader(filename);
2098 if (reader->fileName() != filename) {
2100 reader = new LineReader(filename);
2106 std::string LineReader::readLine(unsigned lineNum) {
2107 if (lineNum < theCurLine) {
2109 fstr.seekg(0, std::ios::beg);
2111 while (theCurLine < lineNum) {
2112 fstr.getline(buff, 500);
2118 // Force static initialization.
2119 extern "C" void LLVMInitializeNVPTXAsmPrinter() {
2120 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2121 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);