1 //===-- NVPTXAsmPrinter.cpp - NVPTX LLVM assembly writer ------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains a printer that converts from our internal representation
11 // of machine-dependent LLVM code to NVPTX assembly language.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "NVPTXAsmPrinter.h"
16 #include "InstPrinter/NVPTXInstPrinter.h"
17 #include "MCTargetDesc/NVPTXMCAsmInfo.h"
18 #include "NVPTX.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/MCInst.h"
41 #include "llvm/MC/MCStreamer.h"
42 #include "llvm/MC/MCSymbol.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/FormattedStream.h"
46 #include "llvm/Support/Path.h"
47 #include "llvm/Support/TargetRegistry.h"
48 #include "llvm/Support/TimeValue.h"
49 #include "llvm/Target/TargetLoweringObjectFile.h"
50 #include "llvm/Transforms/Utils/UnrollLoop.h"
51 #include <sstream>
52 using namespace llvm;
53 
54 #define DEPOTNAME "__local_depot"
55 
56 static cl::opt<bool>
57 EmitLineNumbers("nvptx-emit-line-numbers", cl::Hidden,
58                 cl::desc("NVPTX Specific: Emit Line numbers even without -G"),
59                 cl::init(true));
60 
61 static cl::opt<bool>
62 InterleaveSrc("nvptx-emit-src", cl::ZeroOrMore, cl::Hidden,
63               cl::desc("NVPTX Specific: Emit source line in ptx file"),
64               cl::init(false));
65 
66 namespace {
67 /// DiscoverDependentGlobals - Return a set of GlobalVariables on which \p V
68 /// depends.
69 void DiscoverDependentGlobals(const Value *V,
70                               DenseSet<const GlobalVariable *> &Globals) {
71   if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
72     Globals.insert(GV);
73   else {
74     if (const User *U = dyn_cast<User>(V)) {
75       for (unsigned i = 0, e = U->getNumOperands(); i != e; ++i) {
76         DiscoverDependentGlobals(U->getOperand(i), Globals);
77       }
78     }
79   }
80 }
81 
82 /// VisitGlobalVariableForEmission - Add \p GV to the list of GlobalVariable
83 /// instances to be emitted, but only after any dependents have been added
84 /// first.
85 void VisitGlobalVariableForEmission(
86     const GlobalVariable *GV, SmallVectorImpl<const GlobalVariable *> &Order,
87     DenseSet<const GlobalVariable *> &Visited,
88     DenseSet<const GlobalVariable *> &Visiting) {
89   // Have we already visited this one?
90   if (Visited.count(GV))
91     return;
92 
93   // Do we have a circular dependency?
94   if (!Visiting.insert(GV).second)
95     report_fatal_error("Circular dependency found in global variable set");
96 
97   // Make sure we visit all dependents first
98   DenseSet<const GlobalVariable *> Others;
99   for (unsigned i = 0, e = GV->getNumOperands(); i != e; ++i)
100     DiscoverDependentGlobals(GV->getOperand(i), Others);
101 
102   for (DenseSet<const GlobalVariable *>::iterator I = Others.begin(),
103                                                   E = Others.end();
104        I != E; ++I)
105     VisitGlobalVariableForEmission(*I, Order, Visited, Visiting);
106 
107   // Now we can visit ourself
108   Order.push_back(GV);
109   Visited.insert(GV);
110   Visiting.erase(GV);
111 }
112 }
113 
114 void NVPTXAsmPrinter::emitLineNumberAsDotLoc(const MachineInstr &MI) {
115   if (!EmitLineNumbers)
116     return;
117   if (ignoreLoc(MI))
118     return;
119 
120   DebugLoc curLoc = MI.getDebugLoc();
121 
122   if (!prevDebugLoc && !curLoc)
123     return;
124 
125   if (prevDebugLoc == curLoc)
126     return;
127 
128   prevDebugLoc = curLoc;
129 
130   if (!curLoc)
131     return;
132 
133   auto *Scope = cast_or_null<DIScope>(curLoc.getScope());
134   if (!Scope)
135      return;
136 
137   StringRef fileName(Scope->getFilename());
138   StringRef dirName(Scope->getDirectory());
139   SmallString<128> FullPathName = dirName;
140   if (!dirName.empty() && !sys::path::is_absolute(fileName)) {
141     sys::path::append(FullPathName, fileName);
142     fileName = FullPathName;
143   }
144 
145   if (filenameMap.find(fileName) == filenameMap.end())
146     return;
147 
148   // Emit the line from the source file.
149   if (InterleaveSrc)
150     this->emitSrcInText(fileName, curLoc.getLine());
151 
152   std::stringstream temp;
153   temp << "\t.loc " << filenameMap[fileName] << " " << curLoc.getLine()
154        << " " << curLoc.getCol();
155   OutStreamer->EmitRawText(temp.str());
156 }
157 
158 void NVPTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
159   SmallString<128> Str;
160   raw_svector_ostream OS(Str);
161   if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == NVPTX::CUDA)
162     emitLineNumberAsDotLoc(*MI);
163 
164   MCInst Inst;
165   lowerToMCInst(MI, Inst);
166   EmitToStreamer(*OutStreamer, Inst);
167 }
168 
169 // Handle symbol backtracking for targets that do not support image handles
170 bool NVPTXAsmPrinter::lowerImageHandleOperand(const MachineInstr *MI,
171                                            unsigned OpNo, MCOperand &MCOp) {
172   const MachineOperand &MO = MI->getOperand(OpNo);
173   const MCInstrDesc &MCID = MI->getDesc();
174 
175   if (MCID.TSFlags & NVPTXII::IsTexFlag) {
176     // This is a texture fetch, so operand 4 is a texref and operand 5 is
177     // a samplerref
178     if (OpNo == 4 && MO.isImm()) {
179       lowerImageHandleSymbol(MO.getImm(), MCOp);
180       return true;
181     }
182     if (OpNo == 5 && MO.isImm() && !(MCID.TSFlags & NVPTXII::IsTexModeUnifiedFlag)) {
183       lowerImageHandleSymbol(MO.getImm(), MCOp);
184       return true;
185     }
186 
187     return false;
188   } else if (MCID.TSFlags & NVPTXII::IsSuldMask) {
189     unsigned VecSize =
190       1 << (((MCID.TSFlags & NVPTXII::IsSuldMask) >> NVPTXII::IsSuldShift) - 1);
191 
192     // For a surface load of vector size N, the Nth operand will be the surfref
193     if (OpNo == VecSize && MO.isImm()) {
194       lowerImageHandleSymbol(MO.getImm(), MCOp);
195       return true;
196     }
197 
198     return false;
199   } else if (MCID.TSFlags & NVPTXII::IsSustFlag) {
200     // This is a surface store, so operand 0 is a surfref
201     if (OpNo == 0 && MO.isImm()) {
202       lowerImageHandleSymbol(MO.getImm(), MCOp);
203       return true;
204     }
205 
206     return false;
207   } else if (MCID.TSFlags & NVPTXII::IsSurfTexQueryFlag) {
208     // This is a query, so operand 1 is a surfref/texref
209     if (OpNo == 1 && MO.isImm()) {
210       lowerImageHandleSymbol(MO.getImm(), MCOp);
211       return true;
212     }
213 
214     return false;
215   }
216 
217   return false;
218 }
219 
220 void NVPTXAsmPrinter::lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp) {
221   // Ewwww
222   TargetMachine &TM = const_cast<TargetMachine&>(MF->getTarget());
223   NVPTXTargetMachine &nvTM = static_cast<NVPTXTargetMachine&>(TM);
224   const NVPTXMachineFunctionInfo *MFI = MF->getInfo<NVPTXMachineFunctionInfo>();
225   const char *Sym = MFI->getImageHandleSymbol(Index);
226   std::string *SymNamePtr =
227     nvTM.getManagedStrPool()->getManagedString(Sym);
228   MCOp = GetSymbolRef(OutContext.getOrCreateSymbol(
229     StringRef(SymNamePtr->c_str())));
230 }
231 
232 void NVPTXAsmPrinter::lowerToMCInst(const MachineInstr *MI, MCInst &OutMI) {
233   OutMI.setOpcode(MI->getOpcode());
234   // Special: Do not mangle symbol operand of CALL_PROTOTYPE
235   if (MI->getOpcode() == NVPTX::CALL_PROTOTYPE) {
236     const MachineOperand &MO = MI->getOperand(0);
237     OutMI.addOperand(GetSymbolRef(
238       OutContext.getOrCreateSymbol(Twine(MO.getSymbolName()))));
239     return;
240   }
241 
242   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
243     const MachineOperand &MO = MI->getOperand(i);
244 
245     MCOperand MCOp;
246     if (!nvptxSubtarget->hasImageHandles()) {
247       if (lowerImageHandleOperand(MI, i, MCOp)) {
248         OutMI.addOperand(MCOp);
249         continue;
250       }
251     }
252 
253     if (lowerOperand(MO, MCOp))
254       OutMI.addOperand(MCOp);
255   }
256 }
257 
258 bool NVPTXAsmPrinter::lowerOperand(const MachineOperand &MO,
259                                    MCOperand &MCOp) {
260   switch (MO.getType()) {
261   default: llvm_unreachable("unknown operand type");
262   case MachineOperand::MO_Register:
263     MCOp = MCOperand::createReg(encodeVirtualRegister(MO.getReg()));
264     break;
265   case MachineOperand::MO_Immediate:
266     MCOp = MCOperand::createImm(MO.getImm());
267     break;
268   case MachineOperand::MO_MachineBasicBlock:
269     MCOp = MCOperand::createExpr(MCSymbolRefExpr::create(
270         MO.getMBB()->getSymbol(), OutContext));
271     break;
272   case MachineOperand::MO_ExternalSymbol:
273     MCOp = GetSymbolRef(GetExternalSymbolSymbol(MO.getSymbolName()));
274     break;
275   case MachineOperand::MO_GlobalAddress:
276     MCOp = GetSymbolRef(getSymbol(MO.getGlobal()));
277     break;
278   case MachineOperand::MO_FPImmediate: {
279     const ConstantFP *Cnt = MO.getFPImm();
280     APFloat Val = Cnt->getValueAPF();
281 
282     switch (Cnt->getType()->getTypeID()) {
283     default: report_fatal_error("Unsupported FP type"); break;
284     case Type::FloatTyID:
285       MCOp = MCOperand::createExpr(
286         NVPTXFloatMCExpr::createConstantFPSingle(Val, OutContext));
287       break;
288     case Type::DoubleTyID:
289       MCOp = MCOperand::createExpr(
290         NVPTXFloatMCExpr::createConstantFPDouble(Val, OutContext));
291       break;
292     }
293     break;
294   }
295   }
296   return true;
297 }
298 
299 unsigned NVPTXAsmPrinter::encodeVirtualRegister(unsigned Reg) {
300   if (TargetRegisterInfo::isVirtualRegister(Reg)) {
301     const TargetRegisterClass *RC = MRI->getRegClass(Reg);
302 
303     DenseMap<unsigned, unsigned> &RegMap = VRegMapping[RC];
304     unsigned RegNum = RegMap[Reg];
305 
306     // Encode the register class in the upper 4 bits
307     // Must be kept in sync with NVPTXInstPrinter::printRegName
308     unsigned Ret = 0;
309     if (RC == &NVPTX::Int1RegsRegClass) {
310       Ret = (1 << 28);
311     } else if (RC == &NVPTX::Int16RegsRegClass) {
312       Ret = (2 << 28);
313     } else if (RC == &NVPTX::Int32RegsRegClass) {
314       Ret = (3 << 28);
315     } else if (RC == &NVPTX::Int64RegsRegClass) {
316       Ret = (4 << 28);
317     } else if (RC == &NVPTX::Float32RegsRegClass) {
318       Ret = (5 << 28);
319     } else if (RC == &NVPTX::Float64RegsRegClass) {
320       Ret = (6 << 28);
321     } else {
322       report_fatal_error("Bad register class");
323     }
324 
325     // Insert the vreg number
326     Ret |= (RegNum & 0x0FFFFFFF);
327     return Ret;
328   } else {
329     // Some special-use registers are actually physical registers.
330     // Encode this as the register class ID of 0 and the real register ID.
331     return Reg & 0x0FFFFFFF;
332   }
333 }
334 
335 MCOperand NVPTXAsmPrinter::GetSymbolRef(const MCSymbol *Symbol) {
336   const MCExpr *Expr;
337   Expr = MCSymbolRefExpr::create(Symbol, MCSymbolRefExpr::VK_None,
338                                  OutContext);
339   return MCOperand::createExpr(Expr);
340 }
341 
342 void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
343   const DataLayout &DL = getDataLayout();
344   const TargetLowering *TLI = nvptxSubtarget->getTargetLowering();
345 
346   Type *Ty = F->getReturnType();
347 
348   bool isABI = (nvptxSubtarget->getSmVersion() >= 20);
349 
350   if (Ty->getTypeID() == Type::VoidTyID)
351     return;
352 
353   O << " (";
354 
355   if (isABI) {
356     if (Ty->isFloatingPointTy() || Ty->isIntegerTy()) {
357       unsigned size = 0;
358       if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty)) {
359         size = ITy->getBitWidth();
360         if (size < 32)
361           size = 32;
362       } else {
363         assert(Ty->isFloatingPointTy() && "Floating point type expected here");
364         size = Ty->getPrimitiveSizeInBits();
365       }
366 
367       O << ".param .b" << size << " func_retval0";
368     } else if (isa<PointerType>(Ty)) {
369       O << ".param .b" << TLI->getPointerTy(DL).getSizeInBits()
370         << " func_retval0";
371     } else if ((Ty->getTypeID() == Type::StructTyID) || isa<VectorType>(Ty)) {
372       unsigned totalsz = DL.getTypeAllocSize(Ty);
373        unsigned retAlignment = 0;
374        if (!llvm::getAlign(*F, 0, retAlignment))
375          retAlignment = DL.getABITypeAlignment(Ty);
376        O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
377          << "]";
378     } else
379       llvm_unreachable("Unknown return type");
380   } else {
381     SmallVector<EVT, 16> vtparts;
382     ComputeValueVTs(*TLI, DL, Ty, vtparts);
383     unsigned idx = 0;
384     for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
385       unsigned elems = 1;
386       EVT elemtype = vtparts[i];
387       if (vtparts[i].isVector()) {
388         elems = vtparts[i].getVectorNumElements();
389         elemtype = vtparts[i].getVectorElementType();
390       }
391 
392       for (unsigned j = 0, je = elems; j != je; ++j) {
393         unsigned sz = elemtype.getSizeInBits();
394         if (elemtype.isInteger() && (sz < 32))
395           sz = 32;
396         O << ".reg .b" << sz << " func_retval" << idx;
397         if (j < je - 1)
398           O << ", ";
399         ++idx;
400       }
401       if (i < e - 1)
402         O << ", ";
403     }
404   }
405   O << ") ";
406   return;
407 }
408 
409 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
410                                         raw_ostream &O) {
411   const Function *F = MF.getFunction();
412   printReturnValStr(F, O);
413 }
414 
415 // Return true if MBB is the header of a loop marked with
416 // llvm.loop.unroll.disable.
417 // TODO: consider "#pragma unroll 1" which is equivalent to "#pragma nounroll".
418 bool NVPTXAsmPrinter::isLoopHeaderOfNoUnroll(
419     const MachineBasicBlock &MBB) const {
420   MachineLoopInfo &LI = getAnalysis<MachineLoopInfo>();
421   // We insert .pragma "nounroll" only to the loop header.
422   if (!LI.isLoopHeader(&MBB))
423     return false;
424 
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.
432       continue;
433     }
434     if (const BasicBlock *PBB = PMBB->getBasicBlock()) {
435       if (MDNode *LoopID = PBB->getTerminator()->getMetadata("llvm.loop")) {
436         if (GetUnrollMetadata(LoopID, "llvm.loop.unroll.disable"))
437           return true;
438       }
439     }
440   }
441   return false;
442 }
443 
444 void NVPTXAsmPrinter::EmitBasicBlockStart(const MachineBasicBlock &MBB) const {
445   AsmPrinter::EmitBasicBlockStart(MBB);
446   if (isLoopHeaderOfNoUnroll(MBB))
447     OutStreamer->EmitRawText(StringRef("\t.pragma \"nounroll\";\n"));
448 }
449 
450 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
451   SmallString<128> Str;
452   raw_svector_ostream O(Str);
453 
454   if (!GlobalsEmitted) {
455     emitGlobals(*MF->getFunction()->getParent());
456     GlobalsEmitted = true;
457   }
458 
459   // Set up
460   MRI = &MF->getRegInfo();
461   F = MF->getFunction();
462   emitLinkageDirective(F, O);
463   if (llvm::isKernelFunction(*F))
464     O << ".entry ";
465   else {
466     O << ".func ";
467     printReturnValStr(*MF, O);
468   }
469 
470   CurrentFnSym->print(O, MAI);
471 
472   emitFunctionParamList(*MF, O);
473 
474   if (llvm::isKernelFunction(*F))
475     emitKernelFunctionDirectives(*F, O);
476 
477   OutStreamer->EmitRawText(O.str());
478 
479   prevDebugLoc = DebugLoc();
480 }
481 
482 void NVPTXAsmPrinter::EmitFunctionBodyStart() {
483   VRegMapping.clear();
484   OutStreamer->EmitRawText(StringRef("{\n"));
485   setAndEmitFunctionVirtualRegisters(*MF);
486 
487   SmallString<128> Str;
488   raw_svector_ostream O(Str);
489   emitDemotedVars(MF->getFunction(), O);
490   OutStreamer->EmitRawText(O.str());
491 }
492 
493 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
494   OutStreamer->EmitRawText(StringRef("}\n"));
495   VRegMapping.clear();
496 }
497 
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));
503   } else {
504     OutStreamer->AddComment(Twine("implicit-def: ") +
505                             nvptxSubtarget->getRegisterInfo()->getName(RegNo));
506   }
507   OutStreamer->AddBlankLine();
508 }
509 
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))
518     reqntidx = 1;
519   else
520     specified = true;
521   if (!llvm::getReqNTIDy(F, reqntidy))
522     reqntidy = 1;
523   else
524     specified = true;
525   if (!llvm::getReqNTIDz(F, reqntidz))
526     reqntidz = 1;
527   else
528     specified = true;
529 
530   if (specified)
531     O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
532       << "\n";
533 
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;
538   specified = false;
539   if (!llvm::getMaxNTIDx(F, maxntidx))
540     maxntidx = 1;
541   else
542     specified = true;
543   if (!llvm::getMaxNTIDy(F, maxntidy))
544     maxntidy = 1;
545   else
546     specified = true;
547   if (!llvm::getMaxNTIDz(F, maxntidz))
548     maxntidz = 1;
549   else
550     specified = true;
551 
552   if (specified)
553     O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
554       << "\n";
555 
556   unsigned mincta;
557   if (llvm::getMinCTASm(F, mincta))
558     O << ".minnctapersm " << mincta << "\n";
559 }
560 
561 std::string
562 NVPTXAsmPrinter::getVirtualRegisterName(unsigned Reg) const {
563   const TargetRegisterClass *RC = MRI->getRegClass(Reg);
564 
565   std::string Name;
566   raw_string_ostream NameStr(Name);
567 
568   VRegRCMap::const_iterator I = VRegMapping.find(RC);
569   assert(I != VRegMapping.end() && "Bad register class");
570   const DenseMap<unsigned, unsigned> &RegMap = I->second;
571 
572   VRegMap::const_iterator VI = RegMap.find(Reg);
573   assert(VI != RegMap.end() && "Bad virtual register");
574   unsigned MappedVR = VI->second;
575 
576   NameStr << getNVPTXRegClassStr(RC) << MappedVR;
577 
578   NameStr.flush();
579   return Name;
580 }
581 
582 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr,
583                                           raw_ostream &O) {
584   O << getVirtualRegisterName(vr);
585 }
586 
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))
595       O << "//";
596   } else if (0 == strcmp(Modifier, "vecv4comm2")) {
597     if ((Imm < 4) || (Imm > 7))
598       O << "//";
599   } else if (0 == strcmp(Modifier, "vecv4pos")) {
600     if (Imm < 0)
601       Imm = 0;
602     O << "_" << vecelem[Imm % 4];
603   } else if (0 == strcmp(Modifier, "vecv2comm1")) {
604     if ((Imm < 0) || (Imm > 1))
605       O << "//";
606   } else if (0 == strcmp(Modifier, "vecv2comm2")) {
607     if ((Imm < 2) || (Imm > 3))
608       O << "//";
609   } else if (0 == strcmp(Modifier, "vecv2pos")) {
610     if (Imm < 0)
611       Imm = 0;
612     O << "_" << vecelem[Imm % 2];
613   } else
614     llvm_unreachable("Unknown Modifier on immediate operand");
615 }
616 
617 
618 
619 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
620 
621   emitLinkageDirective(F, O);
622   if (llvm::isKernelFunction(*F))
623     O << ".entry ";
624   else
625     O << ".func ";
626   printReturnValStr(F, O);
627   getSymbol(F)->print(O, MAI);
628   O << "\n";
629   emitFunctionParamList(F, O);
630   O << ";\n";
631 }
632 
633 static bool usedInGlobalVarDef(const Constant *C) {
634   if (!C)
635     return false;
636 
637   if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
638     if (GV->getName() == "llvm.used")
639       return false;
640     return true;
641   }
642 
643   for (const User *U : C->users())
644     if (const Constant *C = dyn_cast<Constant>(U))
645       if (usedInGlobalVarDef(C))
646         return true;
647 
648   return false;
649 }
650 
651 static bool usedInOneFunc(const User *U, Function const *&oneFunc) {
652   if (const GlobalVariable *othergv = dyn_cast<GlobalVariable>(U)) {
653     if (othergv->getName() == "llvm.used")
654       return true;
655   }
656 
657   if (const Instruction *instr = dyn_cast<Instruction>(U)) {
658     if (instr->getParent() && instr->getParent()->getParent()) {
659       const Function *curFunc = instr->getParent()->getParent();
660       if (oneFunc && (curFunc != oneFunc))
661         return false;
662       oneFunc = curFunc;
663       return true;
664     } else
665       return false;
666   }
667 
668   for (const User *UU : U->users())
669     if (!usedInOneFunc(UU, oneFunc))
670       return false;
671 
672   return true;
673 }
674 
675 /* Find out if a global variable can be demoted to local scope.
676  * Currently, this is valid for CUDA shared variables, which have local
677  * scope and global lifetime. So the conditions to check are :
678  * 1. Is the global variable in shared address space?
679  * 2. Does it have internal linkage?
680  * 3. Is the global variable referenced only in one function?
681  */
682 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
683   if (!gv->hasInternalLinkage())
684     return false;
685   const PointerType *Pty = gv->getType();
686   if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
687     return false;
688 
689   const Function *oneFunc = nullptr;
690 
691   bool flag = usedInOneFunc(gv, oneFunc);
692   if (!flag)
693     return false;
694   if (!oneFunc)
695     return false;
696   f = oneFunc;
697   return true;
698 }
699 
700 static bool useFuncSeen(const Constant *C,
701                         llvm::DenseMap<const Function *, bool> &seenMap) {
702   for (const User *U : C->users()) {
703     if (const Constant *cu = dyn_cast<Constant>(U)) {
704       if (useFuncSeen(cu, seenMap))
705         return true;
706     } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
707       const BasicBlock *bb = I->getParent();
708       if (!bb)
709         continue;
710       const Function *caller = bb->getParent();
711       if (!caller)
712         continue;
713       if (seenMap.find(caller) != seenMap.end())
714         return true;
715     }
716   }
717   return false;
718 }
719 
720 void NVPTXAsmPrinter::emitDeclarations(const Module &M, raw_ostream &O) {
721   llvm::DenseMap<const Function *, bool> seenMap;
722   for (Module::const_iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
723     const Function *F = FI;
724 
725     if (F->isDeclaration()) {
726       if (F->use_empty())
727         continue;
728       if (F->getIntrinsicID())
729         continue;
730       emitDeclaration(F, O);
731       continue;
732     }
733     for (const User *U : F->users()) {
734       if (const Constant *C = dyn_cast<Constant>(U)) {
735         if (usedInGlobalVarDef(C)) {
736           // The use is in the initialization of a global variable
737           // that is a function pointer, so print a declaration
738           // for the original function
739           emitDeclaration(F, O);
740           break;
741         }
742         // Emit a declaration of this function if the function that
743         // uses this constant expr has already been seen.
744         if (useFuncSeen(C, seenMap)) {
745           emitDeclaration(F, O);
746           break;
747         }
748       }
749 
750       if (!isa<Instruction>(U))
751         continue;
752       const Instruction *instr = cast<Instruction>(U);
753       const BasicBlock *bb = instr->getParent();
754       if (!bb)
755         continue;
756       const Function *caller = bb->getParent();
757       if (!caller)
758         continue;
759 
760       // If a caller has already been seen, then the caller is
761       // appearing in the module before the callee. so print out
762       // a declaration for the callee.
763       if (seenMap.find(caller) != seenMap.end()) {
764         emitDeclaration(F, O);
765         break;
766       }
767     }
768     seenMap[F] = true;
769   }
770 }
771 
772 void NVPTXAsmPrinter::recordAndEmitFilenames(Module &M) {
773   DebugInfoFinder DbgFinder;
774   DbgFinder.processModule(M);
775 
776   unsigned i = 1;
777   for (const DICompileUnit *DIUnit : DbgFinder.compile_units()) {
778     StringRef Filename = DIUnit->getFilename();
779     StringRef Dirname = DIUnit->getDirectory();
780     SmallString<128> FullPathName = Dirname;
781     if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
782       sys::path::append(FullPathName, Filename);
783       Filename = FullPathName;
784     }
785     if (filenameMap.find(Filename) != filenameMap.end())
786       continue;
787     filenameMap[Filename] = i;
788     OutStreamer->EmitDwarfFileDirective(i, "", Filename);
789     ++i;
790   }
791 
792   for (DISubprogram *SP : DbgFinder.subprograms()) {
793     StringRef Filename = SP->getFilename();
794     StringRef Dirname = SP->getDirectory();
795     SmallString<128> FullPathName = Dirname;
796     if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
797       sys::path::append(FullPathName, Filename);
798       Filename = FullPathName;
799     }
800     if (filenameMap.find(Filename) != filenameMap.end())
801       continue;
802     filenameMap[Filename] = i;
803     ++i;
804   }
805 }
806 
807 bool NVPTXAsmPrinter::doInitialization(Module &M) {
808   // Construct a default subtarget off of the TargetMachine defaults. The
809   // rest of NVPTX isn't friendly to change subtargets per function and
810   // so the default TargetMachine will have all of the options.
811   const Triple &TT = TM.getTargetTriple();
812   StringRef CPU = TM.getTargetCPU();
813   StringRef FS = TM.getTargetFeatureString();
814   const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
815   const NVPTXSubtarget STI(TT, CPU, FS, NTM);
816 
817   SmallString<128> Str1;
818   raw_svector_ostream OS1(Str1);
819 
820   MMI = getAnalysisIfAvailable<MachineModuleInfo>();
821 
822   // We need to call the parent's one explicitly.
823   //bool Result = AsmPrinter::doInitialization(M);
824 
825   // Initialize TargetLoweringObjectFile.
826   const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
827       .Initialize(OutContext, TM);
828 
829   Mang = new Mangler();
830 
831   // Emit header before any dwarf directives are emitted below.
832   emitHeader(M, OS1, STI);
833   OutStreamer->EmitRawText(OS1.str());
834 
835   // Already commented out
836   //bool Result = AsmPrinter::doInitialization(M);
837 
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();
846   }
847 
848   // If we're not NVCL we're CUDA, go ahead and emit filenames.
849   if (TM.getTargetTriple().getOS() != Triple::NVCL)
850     recordAndEmitFilenames(M);
851 
852   GlobalsEmitted = false;
853 
854   return false; // success
855 }
856 
857 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
858   SmallString<128> Str2;
859   raw_svector_ostream OS2(Str2);
860 
861   emitDeclarations(M, OS2);
862 
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;
871 
872   // Visit each global variable, in order
873   for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
874        I != E; ++I)
875     VisitGlobalVariableForEmission(I, Globals, GVVisited, GVVisiting);
876 
877   assert(GVVisited.size() == M.getGlobalList().size() &&
878          "Missed a global variable");
879   assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
880 
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);
884 
885   OS2 << '\n';
886 
887   OutStreamer->EmitRawText(OS2.str());
888 }
889 
890 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O,
891                                  const NVPTXSubtarget &STI) {
892   O << "//\n";
893   O << "// Generated by LLVM NVPTX Back-End\n";
894   O << "//\n";
895   O << "\n";
896 
897   unsigned PTXVersion = STI.getPTXVersion();
898   O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
899 
900   O << ".target ";
901   O << STI.getTargetName();
902 
903   const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
904   if (NTM.getDrvInterface() == NVPTX::NVCL)
905     O << ", texmode_independent";
906   else {
907     if (!STI.hasDouble())
908       O << ", map_f64_to_f32";
909   }
910 
911   if (MAI->doesSupportDebugInformation())
912     O << ", debug";
913 
914   O << "\n";
915 
916   O << ".address_size ";
917   if (NTM.is64Bit())
918     O << "64";
919   else
920     O << "32";
921   O << "\n";
922 
923   O << "\n";
924 }
925 
926 bool NVPTXAsmPrinter::doFinalization(Module &M) {
927   // If we did not emit any functions, then the global declarations have not
928   // yet been emitted.
929   if (!GlobalsEmitted) {
930     emitGlobals(M);
931     GlobalsEmitted = true;
932   }
933 
934   // XXX Temproarily remove global variables so that doFinalization() will not
935   // emit them again (global variables are emitted at beginning).
936 
937   Module::GlobalListType &global_list = M.getGlobalList();
938   int i, n = global_list.size();
939   GlobalVariable **gv_array = new GlobalVariable *[n];
940 
941   // first, back-up GlobalVariable in gv_array
942   i = 0;
943   for (Module::global_iterator I = global_list.begin(), E = global_list.end();
944        I != E; ++I)
945     gv_array[i++] = &*I;
946 
947   // second, empty global_list
948   while (!global_list.empty())
949     global_list.remove(global_list.begin());
950 
951   // call doFinalization
952   bool ret = AsmPrinter::doFinalization(M);
953 
954   // now we restore global variables
955   for (i = 0; i < n; i++)
956     global_list.insert(global_list.end(), gv_array[i]);
957 
958   clearAnnotationCache(&M);
959 
960   delete[] gv_array;
961   return ret;
962 
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.
968   //
969   // Same for the doInitialization.
970   //return Result;
971 }
972 
973 // This function emits appropriate linkage directives for
974 // functions and global variables.
975 //
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,
984 // we emit                                -> .weak.
985 
986 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
987                                            raw_ostream &O) {
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);
992         if (GVar) {
993           if (GVar->hasInitializer())
994             O << ".visible ";
995           else
996             O << ".extern ";
997         }
998       } else if (V->isDeclaration())
999         O << ".extern ";
1000       else
1001         O << ".visible ";
1002     } else if (V->hasAppendingLinkage()) {
1003       std::string msg;
1004       msg.append("Error: ");
1005       msg.append("Symbol ");
1006       if (V->hasName())
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()) {
1012       O << ".weak ";
1013     }
1014   }
1015 }
1016 
1017 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
1018                                          raw_ostream &O,
1019                                          bool processDemoted) {
1020 
1021   // Skip meta data
1022   if (GVar->hasSection()) {
1023     if (GVar->getSection() == StringRef("llvm.metadata"))
1024       return;
1025   }
1026 
1027   // Skip LLVM intrinsic global variables
1028   if (GVar->getName().startswith("llvm.") ||
1029       GVar->getName().startswith("nvvm."))
1030     return;
1031 
1032   const DataLayout *TD = TM.getDataLayout();
1033 
1034   // GlobalVariables are always constant pointers themselves.
1035   const PointerType *PTy = GVar->getType();
1036   Type *ETy = PTy->getElementType();
1037 
1038   if (GVar->hasExternalLinkage()) {
1039     if (GVar->hasInitializer())
1040       O << ".visible ";
1041     else
1042       O << ".extern ";
1043   } else if (GVar->hasLinkOnceLinkage() || GVar->hasWeakLinkage() ||
1044              GVar->hasAvailableExternallyLinkage() ||
1045              GVar->hasCommonLinkage()) {
1046     O << ".weak ";
1047   }
1048 
1049   if (llvm::isTexture(*GVar)) {
1050     O << ".global .texref " << llvm::getTextureName(*GVar) << ";\n";
1051     return;
1052   }
1053 
1054   if (llvm::isSurface(*GVar)) {
1055     O << ".global .surfref " << llvm::getSurfaceName(*GVar) << ";\n";
1056     return;
1057   }
1058 
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);
1064     O << ";\n";
1065     return;
1066   }
1067 
1068   if (llvm::isSampler(*GVar)) {
1069     O << ".global .samplerref " << llvm::getSamplerName(*GVar);
1070 
1071     const Constant *Initializer = nullptr;
1072     if (GVar->hasInitializer())
1073       Initializer = GVar->getInitializer();
1074     const ConstantInt *CI = nullptr;
1075     if (Initializer)
1076       CI = dyn_cast<ConstantInt>(Initializer);
1077     if (CI) {
1078       unsigned sample = CI->getZExtValue();
1079 
1080       O << " = { ";
1081 
1082       for (int i = 0,
1083                addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1084            i < 3; i++) {
1085         O << "addr_mode_" << i << " = ";
1086         switch (addr) {
1087         case 0:
1088           O << "wrap";
1089           break;
1090         case 1:
1091           O << "clamp_to_border";
1092           break;
1093         case 2:
1094           O << "clamp_to_edge";
1095           break;
1096         case 3:
1097           O << "wrap";
1098           break;
1099         case 4:
1100           O << "mirror";
1101           break;
1102         }
1103         O << ", ";
1104       }
1105       O << "filter_mode = ";
1106       switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1107       case 0:
1108         O << "nearest";
1109         break;
1110       case 1:
1111         O << "linear";
1112         break;
1113       case 2:
1114         llvm_unreachable("Anisotropic filtering is not supported");
1115       default:
1116         O << "nearest";
1117         break;
1118       }
1119       if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1120         O << ", force_unnormalized_coords = 1";
1121       }
1122       O << " }";
1123     }
1124 
1125     O << ";\n";
1126     return;
1127   }
1128 
1129   if (GVar->hasPrivateLinkage()) {
1130 
1131     if (!strncmp(GVar->getName().data(), "unrollpragma", 12))
1132       return;
1133 
1134     // FIXME - need better way (e.g. Metadata) to avoid generating this global
1135     if (!strncmp(GVar->getName().data(), "filename", 8))
1136       return;
1137     if (GVar->use_empty())
1138       return;
1139   }
1140 
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);
1146     else {
1147       std::vector<const GlobalVariable *> temp;
1148       temp.push_back(GVar);
1149       localDecls[demotedFunc] = temp;
1150     }
1151     return;
1152   }
1153 
1154   O << ".";
1155   emitPTXAddressSpace(PTy->getAddressSpace(), O);
1156 
1157   if (isManaged(*GVar)) {
1158     O << " .attribute(.managed)";
1159   }
1160 
1161   if (GVar->getAlignment() == 0)
1162     O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1163   else
1164     O << " .align " << GVar->getAlignment();
1165 
1166   if (ETy->isFloatingPointTy() || ETy->isIntegerTy() || ETy->isPointerTy()) {
1167     O << " .";
1168     // Special case: ABI requires that we use .u8 for predicates
1169     if (ETy->isIntegerTy(1))
1170       O << "u8";
1171     else
1172       O << getPTXFundamentalTypeStr(ETy, false);
1173     O << " ";
1174     getSymbol(GVar)->print(O, MAI);
1175 
1176     // Ptx allows variable initilization only for constant and global state
1177     // spaces.
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 specified.
1183         if (!Initializer->isNullValue() && !isa<UndefValue>(Initializer)) {
1184           O << " = ";
1185           printScalarConstant(Initializer, O);
1186         }
1187       } else {
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           report_fatal_error("initial value of '" + GVar->getName() +
1192                              "' is not allowed in addrspace(" +
1193                              Twine(PTy->getAddressSpace()) + ")");
1194         }
1195       }
1196     }
1197   } else {
1198     unsigned int ElementSize = 0;
1199 
1200     // Although PTX has direct support for struct type and array type and
1201     // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1202     // targets that support these high level field accesses. Structs, arrays
1203     // and vectors are lowered into arrays of bytes.
1204     switch (ETy->getTypeID()) {
1205     case Type::StructTyID:
1206     case Type::ArrayTyID:
1207     case Type::VectorTyID:
1208       ElementSize = TD->getTypeStoreSize(ETy);
1209       // Ptx allows variable initilization only for constant and
1210       // global state spaces.
1211       if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1212            (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1213           GVar->hasInitializer()) {
1214         const Constant *Initializer = GVar->getInitializer();
1215         if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1216           AggBuffer aggBuffer(ElementSize, O, *this);
1217           bufferAggregateConstant(Initializer, &aggBuffer);
1218           if (aggBuffer.numSymbols) {
1219             if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit()) {
1220               O << " .u64 ";
1221               getSymbol(GVar)->print(O, MAI);
1222               O << "[";
1223               O << ElementSize / 8;
1224             } else {
1225               O << " .u32 ";
1226               getSymbol(GVar)->print(O, MAI);
1227               O << "[";
1228               O << ElementSize / 4;
1229             }
1230             O << "]";
1231           } else {
1232             O << " .b8 ";
1233             getSymbol(GVar)->print(O, MAI);
1234             O << "[";
1235             O << ElementSize;
1236             O << "]";
1237           }
1238           O << " = {";
1239           aggBuffer.print();
1240           O << "}";
1241         } else {
1242           O << " .b8 ";
1243           getSymbol(GVar)->print(O, MAI);
1244           if (ElementSize) {
1245             O << "[";
1246             O << ElementSize;
1247             O << "]";
1248           }
1249         }
1250       } else {
1251         O << " .b8 ";
1252         getSymbol(GVar)->print(O, MAI);
1253         if (ElementSize) {
1254           O << "[";
1255           O << ElementSize;
1256           O << "]";
1257         }
1258       }
1259       break;
1260     default:
1261       llvm_unreachable("type not supported yet");
1262     }
1263 
1264   }
1265   O << ";\n";
1266 }
1267 
1268 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1269   if (localDecls.find(f) == localDecls.end())
1270     return;
1271 
1272   std::vector<const GlobalVariable *> &gvars = localDecls[f];
1273 
1274   for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1275     O << "\t// demoted variable\n\t";
1276     printModuleLevelGV(gvars[i], O, true);
1277   }
1278 }
1279 
1280 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1281                                           raw_ostream &O) const {
1282   switch (AddressSpace) {
1283   case llvm::ADDRESS_SPACE_LOCAL:
1284     O << "local";
1285     break;
1286   case llvm::ADDRESS_SPACE_GLOBAL:
1287     O << "global";
1288     break;
1289   case llvm::ADDRESS_SPACE_CONST:
1290     O << "const";
1291     break;
1292   case llvm::ADDRESS_SPACE_SHARED:
1293     O << "shared";
1294     break;
1295   default:
1296     report_fatal_error("Bad address space found while emitting PTX");
1297     break;
1298   }
1299 }
1300 
1301 std::string
1302 NVPTXAsmPrinter::getPTXFundamentalTypeStr(const Type *Ty, bool useB4PTR) const {
1303   switch (Ty->getTypeID()) {
1304   default:
1305     llvm_unreachable("unexpected type");
1306     break;
1307   case Type::IntegerTyID: {
1308     unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1309     if (NumBits == 1)
1310       return "pred";
1311     else if (NumBits <= 64) {
1312       std::string name = "u";
1313       return name + utostr(NumBits);
1314     } else {
1315       llvm_unreachable("Integer too large");
1316       break;
1317     }
1318     break;
1319   }
1320   case Type::FloatTyID:
1321     return "f32";
1322   case Type::DoubleTyID:
1323     return "f64";
1324   case Type::PointerTyID:
1325     if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit())
1326       if (useB4PTR)
1327         return "b64";
1328       else
1329         return "u64";
1330     else if (useB4PTR)
1331       return "b32";
1332     else
1333       return "u32";
1334   }
1335   llvm_unreachable("unexpected type");
1336   return nullptr;
1337 }
1338 
1339 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1340                                             raw_ostream &O) {
1341 
1342   const DataLayout *TD = TM.getDataLayout();
1343 
1344   // GlobalVariables are always constant pointers themselves.
1345   const PointerType *PTy = GVar->getType();
1346   Type *ETy = PTy->getElementType();
1347 
1348   O << ".";
1349   emitPTXAddressSpace(PTy->getAddressSpace(), O);
1350   if (GVar->getAlignment() == 0)
1351     O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1352   else
1353     O << " .align " << GVar->getAlignment();
1354 
1355   if (ETy->isFloatingPointTy() || ETy->isIntegerTy() || ETy->isPointerTy()) {
1356     O << " .";
1357     O << getPTXFundamentalTypeStr(ETy);
1358     O << " ";
1359     getSymbol(GVar)->print(O, MAI);
1360     return;
1361   }
1362 
1363   int64_t ElementSize = 0;
1364 
1365   // Although PTX has direct support for struct type and array type and LLVM IR
1366   // is very similar to PTX, the LLVM CodeGen does not support for targets that
1367   // support these high level field accesses. Structs and arrays are lowered
1368   // into arrays of bytes.
1369   switch (ETy->getTypeID()) {
1370   case Type::StructTyID:
1371   case Type::ArrayTyID:
1372   case Type::VectorTyID:
1373     ElementSize = TD->getTypeStoreSize(ETy);
1374     O << " .b8 ";
1375     getSymbol(GVar)->print(O, MAI);
1376     O << "[";
1377     if (ElementSize) {
1378       O << ElementSize;
1379     }
1380     O << "]";
1381     break;
1382   default:
1383     llvm_unreachable("type not supported yet");
1384   }
1385   return;
1386 }
1387 
1388 static unsigned int getOpenCLAlignment(const DataLayout *TD, Type *Ty) {
1389   if (Ty->isSingleValueType())
1390     return TD->getPrefTypeAlignment(Ty);
1391 
1392   const ArrayType *ATy = dyn_cast<ArrayType>(Ty);
1393   if (ATy)
1394     return getOpenCLAlignment(TD, ATy->getElementType());
1395 
1396   const StructType *STy = dyn_cast<StructType>(Ty);
1397   if (STy) {
1398     unsigned int alignStruct = 1;
1399     // Go through each element of the struct and find the
1400     // largest alignment.
1401     for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1402       Type *ETy = STy->getElementType(i);
1403       unsigned int align = getOpenCLAlignment(TD, ETy);
1404       if (align > alignStruct)
1405         alignStruct = align;
1406     }
1407     return alignStruct;
1408   }
1409 
1410   const FunctionType *FTy = dyn_cast<FunctionType>(Ty);
1411   if (FTy)
1412     return TD->getPointerPrefAlignment();
1413   return TD->getPrefTypeAlignment(Ty);
1414 }
1415 
1416 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1417                                      int paramIndex, raw_ostream &O) {
1418   getSymbol(I->getParent())->print(O, MAI);
1419   O << "_param_" << paramIndex;
1420 }
1421 
1422 void NVPTXAsmPrinter::printParamName(int paramIndex, raw_ostream &O) {
1423   CurrentFnSym->print(O, MAI);
1424   O << "_param_" << paramIndex;
1425 }
1426 
1427 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1428   const DataLayout *TD = TM.getDataLayout();
1429   const AttributeSet &PAL = F->getAttributes();
1430   const TargetLowering *TLI = nvptxSubtarget->getTargetLowering();
1431   Function::const_arg_iterator I, E;
1432   unsigned paramIndex = 0;
1433   bool first = true;
1434   bool isKernelFunc = llvm::isKernelFunction(*F);
1435   bool isABI = (nvptxSubtarget->getSmVersion() >= 20);
1436   MVT thePointerTy = TLI->getPointerTy(*TD);
1437 
1438   O << "(\n";
1439 
1440   for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1441     Type *Ty = I->getType();
1442 
1443     if (!first)
1444       O << ",\n";
1445 
1446     first = false;
1447 
1448     // Handle image/sampler parameters
1449     if (isKernelFunction(*F)) {
1450       if (isSampler(*I) || isImage(*I)) {
1451         if (isImage(*I)) {
1452           std::string sname = I->getName();
1453           if (isImageWriteOnly(*I) || isImageReadWrite(*I)) {
1454             if (nvptxSubtarget->hasImageHandles())
1455               O << "\t.param .u64 .ptr .surfref ";
1456             else
1457               O << "\t.param .surfref ";
1458             CurrentFnSym->print(O, MAI);
1459             O << "_param_" << paramIndex;
1460           }
1461           else { // Default image is read_only
1462             if (nvptxSubtarget->hasImageHandles())
1463               O << "\t.param .u64 .ptr .texref ";
1464             else
1465               O << "\t.param .texref ";
1466             CurrentFnSym->print(O, MAI);
1467             O << "_param_" << paramIndex;
1468           }
1469         } else {
1470           if (nvptxSubtarget->hasImageHandles())
1471             O << "\t.param .u64 .ptr .samplerref ";
1472           else
1473             O << "\t.param .samplerref ";
1474           CurrentFnSym->print(O, MAI);
1475           O << "_param_" << paramIndex;
1476         }
1477         continue;
1478       }
1479     }
1480 
1481     if (!PAL.hasAttribute(paramIndex + 1, Attribute::ByVal)) {
1482       if (Ty->isAggregateType() || Ty->isVectorTy()) {
1483         // Just print .param .align <a> .b8 .param[size];
1484         // <a> = PAL.getparamalignment
1485         // size = typeallocsize of element type
1486         unsigned align = PAL.getParamAlignment(paramIndex + 1);
1487         if (align == 0)
1488           align = TD->getABITypeAlignment(Ty);
1489 
1490         unsigned sz = TD->getTypeAllocSize(Ty);
1491         O << "\t.param .align " << align << " .b8 ";
1492         printParamName(I, paramIndex, O);
1493         O << "[" << sz << "]";
1494 
1495         continue;
1496       }
1497       // Just a scalar
1498       const PointerType *PTy = dyn_cast<PointerType>(Ty);
1499       if (isKernelFunc) {
1500         if (PTy) {
1501           // Special handling for pointer arguments to kernel
1502           O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1503 
1504           if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() !=
1505               NVPTX::CUDA) {
1506             Type *ETy = PTy->getElementType();
1507             int addrSpace = PTy->getAddressSpace();
1508             switch (addrSpace) {
1509             default:
1510               O << ".ptr ";
1511               break;
1512             case llvm::ADDRESS_SPACE_CONST:
1513               O << ".ptr .const ";
1514               break;
1515             case llvm::ADDRESS_SPACE_SHARED:
1516               O << ".ptr .shared ";
1517               break;
1518             case llvm::ADDRESS_SPACE_GLOBAL:
1519               O << ".ptr .global ";
1520               break;
1521             }
1522             O << ".align " << (int) getOpenCLAlignment(TD, ETy) << " ";
1523           }
1524           printParamName(I, paramIndex, O);
1525           continue;
1526         }
1527 
1528         // non-pointer scalar to kernel func
1529         O << "\t.param .";
1530         // Special case: predicate operands become .u8 types
1531         if (Ty->isIntegerTy(1))
1532           O << "u8";
1533         else
1534           O << getPTXFundamentalTypeStr(Ty);
1535         O << " ";
1536         printParamName(I, paramIndex, O);
1537         continue;
1538       }
1539       // Non-kernel function, just print .param .b<size> for ABI
1540       // and .reg .b<size> for non-ABI
1541       unsigned sz = 0;
1542       if (isa<IntegerType>(Ty)) {
1543         sz = cast<IntegerType>(Ty)->getBitWidth();
1544         if (sz < 32)
1545           sz = 32;
1546       } else if (isa<PointerType>(Ty))
1547         sz = thePointerTy.getSizeInBits();
1548       else
1549         sz = Ty->getPrimitiveSizeInBits();
1550       if (isABI)
1551         O << "\t.param .b" << sz << " ";
1552       else
1553         O << "\t.reg .b" << sz << " ";
1554       printParamName(I, paramIndex, O);
1555       continue;
1556     }
1557 
1558     // param has byVal attribute. So should be a pointer
1559     const PointerType *PTy = dyn_cast<PointerType>(Ty);
1560     assert(PTy && "Param with byval attribute should be a pointer type");
1561     Type *ETy = PTy->getElementType();
1562 
1563     if (isABI || isKernelFunc) {
1564       // Just print .param .align <a> .b8 .param[size];
1565       // <a> = PAL.getparamalignment
1566       // size = typeallocsize of element type
1567       unsigned align = PAL.getParamAlignment(paramIndex + 1);
1568       if (align == 0)
1569         align = TD->getABITypeAlignment(ETy);
1570 
1571       unsigned sz = TD->getTypeAllocSize(ETy);
1572       O << "\t.param .align " << align << " .b8 ";
1573       printParamName(I, paramIndex, O);
1574       O << "[" << sz << "]";
1575       continue;
1576     } else {
1577       // Split the ETy into constituent parts and
1578       // print .param .b<size> <name> for each part.
1579       // Further, if a part is vector, print the above for
1580       // each vector element.
1581       SmallVector<EVT, 16> vtparts;
1582       ComputeValueVTs(*TLI, getDataLayout(), ETy, vtparts);
1583       for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1584         unsigned elems = 1;
1585         EVT elemtype = vtparts[i];
1586         if (vtparts[i].isVector()) {
1587           elems = vtparts[i].getVectorNumElements();
1588           elemtype = vtparts[i].getVectorElementType();
1589         }
1590 
1591         for (unsigned j = 0, je = elems; j != je; ++j) {
1592           unsigned sz = elemtype.getSizeInBits();
1593           if (elemtype.isInteger() && (sz < 32))
1594             sz = 32;
1595           O << "\t.reg .b" << sz << " ";
1596           printParamName(I, paramIndex, O);
1597           if (j < je - 1)
1598             O << ",\n";
1599           ++paramIndex;
1600         }
1601         if (i < e - 1)
1602           O << ",\n";
1603       }
1604       --paramIndex;
1605       continue;
1606     }
1607   }
1608 
1609   O << "\n)\n";
1610 }
1611 
1612 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1613                                             raw_ostream &O) {
1614   const Function *F = MF.getFunction();
1615   emitFunctionParamList(F, O);
1616 }
1617 
1618 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1619     const MachineFunction &MF) {
1620   SmallString<128> Str;
1621   raw_svector_ostream O(Str);
1622 
1623   // Map the global virtual register number to a register class specific
1624   // virtual register number starting from 1 with that class.
1625   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1626   //unsigned numRegClasses = TRI->getNumRegClasses();
1627 
1628   // Emit the Fake Stack Object
1629   const MachineFrameInfo *MFI = MF.getFrameInfo();
1630   int NumBytes = (int) MFI->getStackSize();
1631   if (NumBytes) {
1632     O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
1633       << getFunctionNumber() << "[" << NumBytes << "];\n";
1634     if (static_cast<const NVPTXTargetMachine &>(MF.getTarget()).is64Bit()) {
1635       O << "\t.reg .b64 \t%SP;\n";
1636       O << "\t.reg .b64 \t%SPL;\n";
1637     } else {
1638       O << "\t.reg .b32 \t%SP;\n";
1639       O << "\t.reg .b32 \t%SPL;\n";
1640     }
1641   }
1642 
1643   // Go through all virtual registers to establish the mapping between the
1644   // global virtual
1645   // register number and the per class virtual register number.
1646   // We use the per class virtual register number in the ptx output.
1647   unsigned int numVRs = MRI->getNumVirtRegs();
1648   for (unsigned i = 0; i < numVRs; i++) {
1649     unsigned int vr = TRI->index2VirtReg(i);
1650     const TargetRegisterClass *RC = MRI->getRegClass(vr);
1651     DenseMap<unsigned, unsigned> &regmap = VRegMapping[RC];
1652     int n = regmap.size();
1653     regmap.insert(std::make_pair(vr, n + 1));
1654   }
1655 
1656   // Emit register declarations
1657   // @TODO: Extract out the real register usage
1658   // O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
1659   // O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
1660   // O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
1661   // O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
1662   // O << "\t.reg .s64 %rd<" << NVPTXNumRegisters << ">;\n";
1663   // O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
1664   // O << "\t.reg .f64 %fd<" << NVPTXNumRegisters << ">;\n";
1665 
1666   // Emit declaration of the virtual registers or 'physical' registers for
1667   // each register class
1668   for (unsigned i=0; i< TRI->getNumRegClasses(); i++) {
1669     const TargetRegisterClass *RC = TRI->getRegClass(i);
1670     DenseMap<unsigned, unsigned> &regmap = VRegMapping[RC];
1671     std::string rcname = getNVPTXRegClassName(RC);
1672     std::string rcStr = getNVPTXRegClassStr(RC);
1673     int n = regmap.size();
1674 
1675     // Only declare those registers that may be used.
1676     if (n) {
1677        O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
1678          << ">;\n";
1679     }
1680   }
1681 
1682   OutStreamer->EmitRawText(O.str());
1683 }
1684 
1685 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
1686   APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
1687   bool ignored;
1688   unsigned int numHex;
1689   const char *lead;
1690 
1691   if (Fp->getType()->getTypeID() == Type::FloatTyID) {
1692     numHex = 8;
1693     lead = "0f";
1694     APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &ignored);
1695   } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
1696     numHex = 16;
1697     lead = "0d";
1698     APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &ignored);
1699   } else
1700     llvm_unreachable("unsupported fp type");
1701 
1702   APInt API = APF.bitcastToAPInt();
1703   std::string hexstr(utohexstr(API.getZExtValue()));
1704   O << lead;
1705   if (hexstr.length() < numHex)
1706     O << std::string(numHex - hexstr.length(), '0');
1707   O << utohexstr(API.getZExtValue());
1708 }
1709 
1710 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
1711   if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
1712     O << CI->getValue();
1713     return;
1714   }
1715   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
1716     printFPConstant(CFP, O);
1717     return;
1718   }
1719   if (isa<ConstantPointerNull>(CPV)) {
1720     O << "0";
1721     return;
1722   }
1723   if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1724     PointerType *PTy = dyn_cast<PointerType>(GVar->getType());
1725     bool IsNonGenericPointer = false;
1726     if (PTy && PTy->getAddressSpace() != 0) {
1727       IsNonGenericPointer = true;
1728     }
1729     if (EmitGeneric && !isa<Function>(CPV) && !IsNonGenericPointer) {
1730       O << "generic(";
1731       getSymbol(GVar)->print(O, MAI);
1732       O << ")";
1733     } else {
1734       getSymbol(GVar)->print(O, MAI);
1735     }
1736     return;
1737   }
1738   if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1739     const Value *v = Cexpr->stripPointerCasts();
1740     PointerType *PTy = dyn_cast<PointerType>(Cexpr->getType());
1741     bool IsNonGenericPointer = false;
1742     if (PTy && PTy->getAddressSpace() != 0) {
1743       IsNonGenericPointer = true;
1744     }
1745     if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
1746       if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
1747         O << "generic(";
1748         getSymbol(GVar)->print(O, MAI);
1749         O << ")";
1750       } else {
1751         getSymbol(GVar)->print(O, MAI);
1752       }
1753       return;
1754     } else {
1755       lowerConstant(CPV)->print(O, MAI);
1756       return;
1757     }
1758   }
1759   llvm_unreachable("Not scalar type found in printScalarConstant()");
1760 }
1761 
1762 // These utility functions assure we get the right sequence of bytes for a given
1763 // type even for big-endian machines
1764 template <typename T> static void ConvertIntToBytes(unsigned char *p, T val) {
1765   int64_t vp = (int64_t)val;
1766   for (unsigned i = 0; i < sizeof(T); ++i) {
1767     p[i] = (unsigned char)vp;
1768     vp >>= 8;
1769   }
1770 }
1771 static void ConvertFloatToBytes(unsigned char *p, float val) {
1772   int32_t *vp = (int32_t *)&val;
1773   for (unsigned i = 0; i < sizeof(int32_t); ++i) {
1774     p[i] = (unsigned char)*vp;
1775     *vp >>= 8;
1776   }
1777 }
1778 static void ConvertDoubleToBytes(unsigned char *p, double val) {
1779   int64_t *vp = (int64_t *)&val;
1780   for (unsigned i = 0; i < sizeof(int64_t); ++i) {
1781     p[i] = (unsigned char)*vp;
1782     *vp >>= 8;
1783   }
1784 }
1785 
1786 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
1787                                    AggBuffer *aggBuffer) {
1788 
1789   const DataLayout *TD = TM.getDataLayout();
1790 
1791   if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
1792     int s = TD->getTypeAllocSize(CPV->getType());
1793     if (s < Bytes)
1794       s = Bytes;
1795     aggBuffer->addZeros(s);
1796     return;
1797   }
1798 
1799   unsigned char ptr[8];
1800   switch (CPV->getType()->getTypeID()) {
1801 
1802   case Type::IntegerTyID: {
1803     const Type *ETy = CPV->getType();
1804     if (ETy == Type::getInt8Ty(CPV->getContext())) {
1805       unsigned char c = (unsigned char)cast<ConstantInt>(CPV)->getZExtValue();
1806       ConvertIntToBytes<>(ptr, c);
1807       aggBuffer->addBytes(ptr, 1, Bytes);
1808     } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
1809       short int16 = (short)cast<ConstantInt>(CPV)->getZExtValue();
1810       ConvertIntToBytes<>(ptr, int16);
1811       aggBuffer->addBytes(ptr, 2, Bytes);
1812     } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
1813       if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1814         int int32 = (int)(constInt->getZExtValue());
1815         ConvertIntToBytes<>(ptr, int32);
1816         aggBuffer->addBytes(ptr, 4, Bytes);
1817         break;
1818       } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1819         if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1820                 ConstantFoldConstantExpression(Cexpr, *TD))) {
1821           int int32 = (int)(constInt->getZExtValue());
1822           ConvertIntToBytes<>(ptr, int32);
1823           aggBuffer->addBytes(ptr, 4, Bytes);
1824           break;
1825         }
1826         if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1827           Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1828           aggBuffer->addSymbol(v, Cexpr->getOperand(0));
1829           aggBuffer->addZeros(4);
1830           break;
1831         }
1832       }
1833       llvm_unreachable("unsupported integer const type");
1834     } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
1835       if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1836         long long int64 = (long long)(constInt->getZExtValue());
1837         ConvertIntToBytes<>(ptr, int64);
1838         aggBuffer->addBytes(ptr, 8, Bytes);
1839         break;
1840       } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1841         if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1842                 ConstantFoldConstantExpression(Cexpr, *TD))) {
1843           long long int64 = (long long)(constInt->getZExtValue());
1844           ConvertIntToBytes<>(ptr, int64);
1845           aggBuffer->addBytes(ptr, 8, Bytes);
1846           break;
1847         }
1848         if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1849           Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1850           aggBuffer->addSymbol(v, Cexpr->getOperand(0));
1851           aggBuffer->addZeros(8);
1852           break;
1853         }
1854       }
1855       llvm_unreachable("unsupported integer const type");
1856     } else
1857       llvm_unreachable("unsupported integer const type");
1858     break;
1859   }
1860   case Type::FloatTyID:
1861   case Type::DoubleTyID: {
1862     const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
1863     const Type *Ty = CFP->getType();
1864     if (Ty == Type::getFloatTy(CPV->getContext())) {
1865       float float32 = (float) CFP->getValueAPF().convertToFloat();
1866       ConvertFloatToBytes(ptr, float32);
1867       aggBuffer->addBytes(ptr, 4, Bytes);
1868     } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
1869       double float64 = CFP->getValueAPF().convertToDouble();
1870       ConvertDoubleToBytes(ptr, float64);
1871       aggBuffer->addBytes(ptr, 8, Bytes);
1872     } else {
1873       llvm_unreachable("unsupported fp const type");
1874     }
1875     break;
1876   }
1877   case Type::PointerTyID: {
1878     if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1879       aggBuffer->addSymbol(GVar, GVar);
1880     } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1881       const Value *v = Cexpr->stripPointerCasts();
1882       aggBuffer->addSymbol(v, Cexpr);
1883     }
1884     unsigned int s = TD->getTypeAllocSize(CPV->getType());
1885     aggBuffer->addZeros(s);
1886     break;
1887   }
1888 
1889   case Type::ArrayTyID:
1890   case Type::VectorTyID:
1891   case Type::StructTyID: {
1892     if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV) ||
1893         isa<ConstantStruct>(CPV) || isa<ConstantDataSequential>(CPV)) {
1894       int ElementSize = TD->getTypeAllocSize(CPV->getType());
1895       bufferAggregateConstant(CPV, aggBuffer);
1896       if (Bytes > ElementSize)
1897         aggBuffer->addZeros(Bytes - ElementSize);
1898     } else if (isa<ConstantAggregateZero>(CPV))
1899       aggBuffer->addZeros(Bytes);
1900     else
1901       llvm_unreachable("Unexpected Constant type");
1902     break;
1903   }
1904 
1905   default:
1906     llvm_unreachable("unsupported type");
1907   }
1908 }
1909 
1910 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
1911                                               AggBuffer *aggBuffer) {
1912   const DataLayout *TD = TM.getDataLayout();
1913   int Bytes;
1914 
1915   // Old constants
1916   if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
1917     if (CPV->getNumOperands())
1918       for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
1919         bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
1920     return;
1921   }
1922 
1923   if (const ConstantDataSequential *CDS =
1924           dyn_cast<ConstantDataSequential>(CPV)) {
1925     if (CDS->getNumElements())
1926       for (unsigned i = 0; i < CDS->getNumElements(); ++i)
1927         bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
1928                      aggBuffer);
1929     return;
1930   }
1931 
1932   if (isa<ConstantStruct>(CPV)) {
1933     if (CPV->getNumOperands()) {
1934       StructType *ST = cast<StructType>(CPV->getType());
1935       for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
1936         if (i == (e - 1))
1937           Bytes = TD->getStructLayout(ST)->getElementOffset(0) +
1938                   TD->getTypeAllocSize(ST) -
1939                   TD->getStructLayout(ST)->getElementOffset(i);
1940         else
1941           Bytes = TD->getStructLayout(ST)->getElementOffset(i + 1) -
1942                   TD->getStructLayout(ST)->getElementOffset(i);
1943         bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
1944       }
1945     }
1946     return;
1947   }
1948   llvm_unreachable("unsupported constant type in printAggregateConstant()");
1949 }
1950 
1951 // buildTypeNameMap - Run through symbol table looking for type names.
1952 //
1953 
1954 bool NVPTXAsmPrinter::isImageType(const Type *Ty) {
1955 
1956   std::map<const Type *, std::string>::iterator PI = TypeNameMap.find(Ty);
1957 
1958   if (PI != TypeNameMap.end() && (!PI->second.compare("struct._image1d_t") ||
1959                                   !PI->second.compare("struct._image2d_t") ||
1960                                   !PI->second.compare("struct._image3d_t")))
1961     return true;
1962 
1963   return false;
1964 }
1965 
1966 
1967 bool NVPTXAsmPrinter::ignoreLoc(const MachineInstr &MI) {
1968   switch (MI.getOpcode()) {
1969   default:
1970     return false;
1971   case NVPTX::CallArgBeginInst:
1972   case NVPTX::CallArgEndInst0:
1973   case NVPTX::CallArgEndInst1:
1974   case NVPTX::CallArgF32:
1975   case NVPTX::CallArgF64:
1976   case NVPTX::CallArgI16:
1977   case NVPTX::CallArgI32:
1978   case NVPTX::CallArgI32imm:
1979   case NVPTX::CallArgI64:
1980   case NVPTX::CallArgParam:
1981   case NVPTX::CallVoidInst:
1982   case NVPTX::CallVoidInstReg:
1983   case NVPTX::Callseq_End:
1984   case NVPTX::CallVoidInstReg64:
1985   case NVPTX::DeclareParamInst:
1986   case NVPTX::DeclareRetMemInst:
1987   case NVPTX::DeclareRetRegInst:
1988   case NVPTX::DeclareRetScalarInst:
1989   case NVPTX::DeclareScalarParamInst:
1990   case NVPTX::DeclareScalarRegInst:
1991   case NVPTX::StoreParamF32:
1992   case NVPTX::StoreParamF64:
1993   case NVPTX::StoreParamI16:
1994   case NVPTX::StoreParamI32:
1995   case NVPTX::StoreParamI64:
1996   case NVPTX::StoreParamI8:
1997   case NVPTX::StoreRetvalF32:
1998   case NVPTX::StoreRetvalF64:
1999   case NVPTX::StoreRetvalI16:
2000   case NVPTX::StoreRetvalI32:
2001   case NVPTX::StoreRetvalI64:
2002   case NVPTX::StoreRetvalI8:
2003   case NVPTX::LastCallArgF32:
2004   case NVPTX::LastCallArgF64:
2005   case NVPTX::LastCallArgI16:
2006   case NVPTX::LastCallArgI32:
2007   case NVPTX::LastCallArgI32imm:
2008   case NVPTX::LastCallArgI64:
2009   case NVPTX::LastCallArgParam:
2010   case NVPTX::LoadParamMemF32:
2011   case NVPTX::LoadParamMemF64:
2012   case NVPTX::LoadParamMemI16:
2013   case NVPTX::LoadParamMemI32:
2014   case NVPTX::LoadParamMemI64:
2015   case NVPTX::LoadParamMemI8:
2016   case NVPTX::PrototypeInst:
2017   case NVPTX::DBG_VALUE:
2018     return true;
2019   }
2020   return false;
2021 }
2022 
2023 /// lowerConstantForGV - Return an MCExpr for the given Constant.  This is mostly
2024 /// a copy from AsmPrinter::lowerConstant, except customized to only handle
2025 /// expressions that are representable in PTX and create
2026 /// NVPTXGenericMCSymbolRefExpr nodes for addrspacecast instructions.
2027 const MCExpr *
2028 NVPTXAsmPrinter::lowerConstantForGV(const Constant *CV, bool ProcessingGeneric) {
2029   MCContext &Ctx = OutContext;
2030 
2031   if (CV->isNullValue() || isa<UndefValue>(CV))
2032     return MCConstantExpr::create(0, Ctx);
2033 
2034   if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV))
2035     return MCConstantExpr::create(CI->getZExtValue(), Ctx);
2036 
2037   if (const GlobalValue *GV = dyn_cast<GlobalValue>(CV)) {
2038     const MCSymbolRefExpr *Expr =
2039       MCSymbolRefExpr::create(getSymbol(GV), Ctx);
2040     if (ProcessingGeneric) {
2041       return NVPTXGenericMCSymbolRefExpr::create(Expr, Ctx);
2042     } else {
2043       return Expr;
2044     }
2045   }
2046 
2047   const ConstantExpr *CE = dyn_cast<ConstantExpr>(CV);
2048   if (!CE) {
2049     llvm_unreachable("Unknown constant value to lower!");
2050   }
2051 
2052   switch (CE->getOpcode()) {
2053   default:
2054     // If the code isn't optimized, there may be outstanding folding
2055     // opportunities. Attempt to fold the expression using DataLayout as a
2056     // last resort before giving up.
2057     if (Constant *C = ConstantFoldConstantExpression(CE, *TM.getDataLayout()))
2058       if (C != CE)
2059         return lowerConstantForGV(C, ProcessingGeneric);
2060 
2061     // Otherwise report the problem to the user.
2062     {
2063       std::string S;
2064       raw_string_ostream OS(S);
2065       OS << "Unsupported expression in static initializer: ";
2066       CE->printAsOperand(OS, /*PrintType=*/false,
2067                      !MF ? nullptr : MF->getFunction()->getParent());
2068       report_fatal_error(OS.str());
2069     }
2070 
2071   case Instruction::AddrSpaceCast: {
2072     // Strip the addrspacecast and pass along the operand
2073     PointerType *DstTy = cast<PointerType>(CE->getType());
2074     if (DstTy->getAddressSpace() == 0) {
2075       return lowerConstantForGV(cast<const Constant>(CE->getOperand(0)), true);
2076     }
2077     std::string S;
2078     raw_string_ostream OS(S);
2079     OS << "Unsupported expression in static initializer: ";
2080     CE->printAsOperand(OS, /*PrintType=*/ false,
2081                        !MF ? 0 : MF->getFunction()->getParent());
2082     report_fatal_error(OS.str());
2083   }
2084 
2085   case Instruction::GetElementPtr: {
2086     const DataLayout &DL = *TM.getDataLayout();
2087 
2088     // Generate a symbolic expression for the byte address
2089     APInt OffsetAI(DL.getPointerTypeSizeInBits(CE->getType()), 0);
2090     cast<GEPOperator>(CE)->accumulateConstantOffset(DL, OffsetAI);
2091 
2092     const MCExpr *Base = lowerConstantForGV(CE->getOperand(0),
2093                                             ProcessingGeneric);
2094     if (!OffsetAI)
2095       return Base;
2096 
2097     int64_t Offset = OffsetAI.getSExtValue();
2098     return MCBinaryExpr::createAdd(Base, MCConstantExpr::create(Offset, Ctx),
2099                                    Ctx);
2100   }
2101 
2102   case Instruction::Trunc:
2103     // We emit the value and depend on the assembler to truncate the generated
2104     // expression properly.  This is important for differences between
2105     // blockaddress labels.  Since the two labels are in the same function, it
2106     // is reasonable to treat their delta as a 32-bit value.
2107     // FALL THROUGH.
2108   case Instruction::BitCast:
2109     return lowerConstantForGV(CE->getOperand(0), ProcessingGeneric);
2110 
2111   case Instruction::IntToPtr: {
2112     const DataLayout &DL = *TM.getDataLayout();
2113 
2114     // Handle casts to pointers by changing them into casts to the appropriate
2115     // integer type.  This promotes constant folding and simplifies this code.
2116     Constant *Op = CE->getOperand(0);
2117     Op = ConstantExpr::getIntegerCast(Op, DL.getIntPtrType(CV->getType()),
2118                                       false/*ZExt*/);
2119     return lowerConstantForGV(Op, ProcessingGeneric);
2120   }
2121 
2122   case Instruction::PtrToInt: {
2123     const DataLayout &DL = *TM.getDataLayout();
2124 
2125     // Support only foldable casts to/from pointers that can be eliminated by
2126     // changing the pointer to the appropriately sized integer type.
2127     Constant *Op = CE->getOperand(0);
2128     Type *Ty = CE->getType();
2129 
2130     const MCExpr *OpExpr = lowerConstantForGV(Op, ProcessingGeneric);
2131 
2132     // We can emit the pointer value into this slot if the slot is an
2133     // integer slot equal to the size of the pointer.
2134     if (DL.getTypeAllocSize(Ty) == DL.getTypeAllocSize(Op->getType()))
2135       return OpExpr;
2136 
2137     // Otherwise the pointer is smaller than the resultant integer, mask off
2138     // the high bits so we are sure to get a proper truncation if the input is
2139     // a constant expr.
2140     unsigned InBits = DL.getTypeAllocSizeInBits(Op->getType());
2141     const MCExpr *MaskExpr = MCConstantExpr::create(~0ULL >> (64-InBits), Ctx);
2142     return MCBinaryExpr::createAnd(OpExpr, MaskExpr, Ctx);
2143   }
2144 
2145   // The MC library also has a right-shift operator, but it isn't consistently
2146   // signed or unsigned between different targets.
2147   case Instruction::Add: {
2148     const MCExpr *LHS = lowerConstantForGV(CE->getOperand(0), ProcessingGeneric);
2149     const MCExpr *RHS = lowerConstantForGV(CE->getOperand(1), ProcessingGeneric);
2150     switch (CE->getOpcode()) {
2151     default: llvm_unreachable("Unknown binary operator constant cast expr");
2152     case Instruction::Add: return MCBinaryExpr::createAdd(LHS, RHS, Ctx);
2153     }
2154   }
2155   }
2156 }
2157 
2158 // Copy of MCExpr::print customized for NVPTX
2159 void NVPTXAsmPrinter::printMCExpr(const MCExpr &Expr, raw_ostream &OS) {
2160   switch (Expr.getKind()) {
2161   case MCExpr::Target:
2162     return cast<MCTargetExpr>(&Expr)->printImpl(OS, MAI);
2163   case MCExpr::Constant:
2164     OS << cast<MCConstantExpr>(Expr).getValue();
2165     return;
2166 
2167   case MCExpr::SymbolRef: {
2168     const MCSymbolRefExpr &SRE = cast<MCSymbolRefExpr>(Expr);
2169     const MCSymbol &Sym = SRE.getSymbol();
2170     Sym.print(OS, MAI);
2171     return;
2172   }
2173 
2174   case MCExpr::Unary: {
2175     const MCUnaryExpr &UE = cast<MCUnaryExpr>(Expr);
2176     switch (UE.getOpcode()) {
2177     case MCUnaryExpr::LNot:  OS << '!'; break;
2178     case MCUnaryExpr::Minus: OS << '-'; break;
2179     case MCUnaryExpr::Not:   OS << '~'; break;
2180     case MCUnaryExpr::Plus:  OS << '+'; break;
2181     }
2182     printMCExpr(*UE.getSubExpr(), OS);
2183     return;
2184   }
2185 
2186   case MCExpr::Binary: {
2187     const MCBinaryExpr &BE = cast<MCBinaryExpr>(Expr);
2188 
2189     // Only print parens around the LHS if it is non-trivial.
2190     if (isa<MCConstantExpr>(BE.getLHS()) || isa<MCSymbolRefExpr>(BE.getLHS()) ||
2191         isa<NVPTXGenericMCSymbolRefExpr>(BE.getLHS())) {
2192       printMCExpr(*BE.getLHS(), OS);
2193     } else {
2194       OS << '(';
2195       printMCExpr(*BE.getLHS(), OS);
2196       OS<< ')';
2197     }
2198 
2199     switch (BE.getOpcode()) {
2200     case MCBinaryExpr::Add:
2201       // Print "X-42" instead of "X+-42".
2202       if (const MCConstantExpr *RHSC = dyn_cast<MCConstantExpr>(BE.getRHS())) {
2203         if (RHSC->getValue() < 0) {
2204           OS << RHSC->getValue();
2205           return;
2206         }
2207       }
2208 
2209       OS <<  '+';
2210       break;
2211     default: llvm_unreachable("Unhandled binary operator");
2212     }
2213 
2214     // Only print parens around the LHS if it is non-trivial.
2215     if (isa<MCConstantExpr>(BE.getRHS()) || isa<MCSymbolRefExpr>(BE.getRHS())) {
2216       printMCExpr(*BE.getRHS(), OS);
2217     } else {
2218       OS << '(';
2219       printMCExpr(*BE.getRHS(), OS);
2220       OS << ')';
2221     }
2222     return;
2223   }
2224   }
2225 
2226   llvm_unreachable("Invalid expression kind!");
2227 }
2228 
2229 /// PrintAsmOperand - Print out an operand for an inline asm expression.
2230 ///
2231 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
2232                                       unsigned AsmVariant,
2233                                       const char *ExtraCode, raw_ostream &O) {
2234   if (ExtraCode && ExtraCode[0]) {
2235     if (ExtraCode[1] != 0)
2236       return true; // Unknown modifier.
2237 
2238     switch (ExtraCode[0]) {
2239     default:
2240       // See if this is a generic print operand
2241       return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
2242     case 'r':
2243       break;
2244     }
2245   }
2246 
2247   printOperand(MI, OpNo, O);
2248 
2249   return false;
2250 }
2251 
2252 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(
2253     const MachineInstr *MI, unsigned OpNo, unsigned AsmVariant,
2254     const char *ExtraCode, raw_ostream &O) {
2255   if (ExtraCode && ExtraCode[0])
2256     return true; // Unknown modifier
2257 
2258   O << '[';
2259   printMemOperand(MI, OpNo, O);
2260   O << ']';
2261 
2262   return false;
2263 }
2264 
2265 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
2266                                    raw_ostream &O, const char *Modifier) {
2267   const MachineOperand &MO = MI->getOperand(opNum);
2268   switch (MO.getType()) {
2269   case MachineOperand::MO_Register:
2270     if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
2271       if (MO.getReg() == NVPTX::VRDepot)
2272         O << DEPOTNAME << getFunctionNumber();
2273       else
2274         O << NVPTXInstPrinter::getRegisterName(MO.getReg());
2275     } else {
2276       emitVirtualRegister(MO.getReg(), O);
2277     }
2278     return;
2279 
2280   case MachineOperand::MO_Immediate:
2281     if (!Modifier)
2282       O << MO.getImm();
2283     else if (strstr(Modifier, "vec") == Modifier)
2284       printVecModifiedImmediate(MO, Modifier, O);
2285     else
2286       llvm_unreachable(
2287           "Don't know how to handle modifier on immediate operand");
2288     return;
2289 
2290   case MachineOperand::MO_FPImmediate:
2291     printFPConstant(MO.getFPImm(), O);
2292     break;
2293 
2294   case MachineOperand::MO_GlobalAddress:
2295     getSymbol(MO.getGlobal())->print(O, MAI);
2296     break;
2297 
2298   case MachineOperand::MO_MachineBasicBlock:
2299     MO.getMBB()->getSymbol()->print(O, MAI);
2300     return;
2301 
2302   default:
2303     llvm_unreachable("Operand type not supported.");
2304   }
2305 }
2306 
2307 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
2308                                       raw_ostream &O, const char *Modifier) {
2309   printOperand(MI, opNum, O);
2310 
2311   if (Modifier && !strcmp(Modifier, "add")) {
2312     O << ", ";
2313     printOperand(MI, opNum + 1, O);
2314   } else {
2315     if (MI->getOperand(opNum + 1).isImm() &&
2316         MI->getOperand(opNum + 1).getImm() == 0)
2317       return; // don't print ',0' or '+0'
2318     O << "+";
2319     printOperand(MI, opNum + 1, O);
2320   }
2321 }
2322 
2323 void NVPTXAsmPrinter::emitSrcInText(StringRef filename, unsigned line) {
2324   std::stringstream temp;
2325   LineReader *reader = this->getReader(filename);
2326   temp << "\n//";
2327   temp << filename.str();
2328   temp << ":";
2329   temp << line;
2330   temp << " ";
2331   temp << reader->readLine(line);
2332   temp << "\n";
2333   this->OutStreamer->EmitRawText(temp.str());
2334 }
2335 
2336 LineReader *NVPTXAsmPrinter::getReader(std::string filename) {
2337   if (!reader) {
2338     reader = new LineReader(filename);
2339   }
2340 
2341   if (reader->fileName() != filename) {
2342     delete reader;
2343     reader = new LineReader(filename);
2344   }
2345 
2346   return reader;
2347 }
2348 
2349 std::string LineReader::readLine(unsigned lineNum) {
2350   if (lineNum < theCurLine) {
2351     theCurLine = 0;
2352     fstr.seekg(0, std::ios::beg);
2353   }
2354   while (theCurLine < lineNum) {
2355     fstr.getline(buff, 500);
2356     theCurLine++;
2357   }
2358   return buff;
2359 }
2360 
2361 // Force static initialization.
2362 extern "C" void LLVMInitializeNVPTXAsmPrinter() {
2363   RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2364   RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2365 }
2366