1 //===-- NVPTXAsmPrinter.cpp - NVPTX LLVM assembly writer ------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains a printer that converts from our internal representation
10 // of machine-dependent LLVM code to NVPTX assembly language.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "NVPTXAsmPrinter.h"
15 #include "MCTargetDesc/NVPTXBaseInfo.h"
16 #include "MCTargetDesc/NVPTXInstPrinter.h"
17 #include "MCTargetDesc/NVPTXMCAsmInfo.h"
18 #include "MCTargetDesc/NVPTXTargetStreamer.h"
19 #include "NVPTX.h"
20 #include "NVPTXMCExpr.h"
21 #include "NVPTXMachineFunctionInfo.h"
22 #include "NVPTXRegisterInfo.h"
23 #include "NVPTXSubtarget.h"
24 #include "NVPTXTargetMachine.h"
25 #include "NVPTXUtilities.h"
26 #include "TargetInfo/NVPTXTargetInfo.h"
27 #include "cl_common_defines.h"
28 #include "llvm/ADT/APFloat.h"
29 #include "llvm/ADT/APInt.h"
30 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/DenseSet.h"
32 #include "llvm/ADT/SmallString.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/StringExtras.h"
35 #include "llvm/ADT/StringRef.h"
36 #include "llvm/ADT/Triple.h"
37 #include "llvm/ADT/Twine.h"
38 #include "llvm/Analysis/ConstantFolding.h"
39 #include "llvm/CodeGen/Analysis.h"
40 #include "llvm/CodeGen/MachineBasicBlock.h"
41 #include "llvm/CodeGen/MachineFrameInfo.h"
42 #include "llvm/CodeGen/MachineFunction.h"
43 #include "llvm/CodeGen/MachineInstr.h"
44 #include "llvm/CodeGen/MachineLoopInfo.h"
45 #include "llvm/CodeGen/MachineModuleInfo.h"
46 #include "llvm/CodeGen/MachineOperand.h"
47 #include "llvm/CodeGen/MachineRegisterInfo.h"
48 #include "llvm/CodeGen/TargetLowering.h"
49 #include "llvm/CodeGen/TargetRegisterInfo.h"
50 #include "llvm/CodeGen/ValueTypes.h"
51 #include "llvm/IR/Attributes.h"
52 #include "llvm/IR/BasicBlock.h"
53 #include "llvm/IR/Constant.h"
54 #include "llvm/IR/Constants.h"
55 #include "llvm/IR/DataLayout.h"
56 #include "llvm/IR/DebugInfo.h"
57 #include "llvm/IR/DebugInfoMetadata.h"
58 #include "llvm/IR/DebugLoc.h"
59 #include "llvm/IR/DerivedTypes.h"
60 #include "llvm/IR/Function.h"
61 #include "llvm/IR/GlobalValue.h"
62 #include "llvm/IR/GlobalVariable.h"
63 #include "llvm/IR/Instruction.h"
64 #include "llvm/IR/LLVMContext.h"
65 #include "llvm/IR/Module.h"
66 #include "llvm/IR/Operator.h"
67 #include "llvm/IR/Type.h"
68 #include "llvm/IR/User.h"
69 #include "llvm/MC/MCExpr.h"
70 #include "llvm/MC/MCInst.h"
71 #include "llvm/MC/MCInstrDesc.h"
72 #include "llvm/MC/MCStreamer.h"
73 #include "llvm/MC/MCSymbol.h"
74 #include "llvm/Support/Casting.h"
75 #include "llvm/Support/CommandLine.h"
76 #include "llvm/Support/ErrorHandling.h"
77 #include "llvm/Support/MachineValueType.h"
78 #include "llvm/Support/Path.h"
79 #include "llvm/Support/TargetRegistry.h"
80 #include "llvm/Support/raw_ostream.h"
81 #include "llvm/Target/TargetLoweringObjectFile.h"
82 #include "llvm/Target/TargetMachine.h"
83 #include "llvm/Transforms/Utils/UnrollLoop.h"
84 #include <cassert>
85 #include <cstdint>
86 #include <cstring>
87 #include <new>
88 #include <string>
89 #include <utility>
90 #include <vector>
91 
92 using namespace llvm;
93 
94 #define DEPOTNAME "__local_depot"
95 
96 /// DiscoverDependentGlobals - Return a set of GlobalVariables on which \p V
97 /// depends.
98 static void
99 DiscoverDependentGlobals(const Value *V,
100                          DenseSet<const GlobalVariable *> &Globals) {
101   if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
102     Globals.insert(GV);
103   else {
104     if (const User *U = dyn_cast<User>(V)) {
105       for (unsigned i = 0, e = U->getNumOperands(); i != e; ++i) {
106         DiscoverDependentGlobals(U->getOperand(i), Globals);
107       }
108     }
109   }
110 }
111 
112 /// VisitGlobalVariableForEmission - Add \p GV to the list of GlobalVariable
113 /// instances to be emitted, but only after any dependents have been added
114 /// first.s
115 static void
116 VisitGlobalVariableForEmission(const GlobalVariable *GV,
117                                SmallVectorImpl<const GlobalVariable *> &Order,
118                                DenseSet<const GlobalVariable *> &Visited,
119                                DenseSet<const GlobalVariable *> &Visiting) {
120   // Have we already visited this one?
121   if (Visited.count(GV))
122     return;
123 
124   // Do we have a circular dependency?
125   if (!Visiting.insert(GV).second)
126     report_fatal_error("Circular dependency found in global variable set");
127 
128   // Make sure we visit all dependents first
129   DenseSet<const GlobalVariable *> Others;
130   for (unsigned i = 0, e = GV->getNumOperands(); i != e; ++i)
131     DiscoverDependentGlobals(GV->getOperand(i), Others);
132 
133   for (DenseSet<const GlobalVariable *>::iterator I = Others.begin(),
134                                                   E = Others.end();
135        I != E; ++I)
136     VisitGlobalVariableForEmission(*I, Order, Visited, Visiting);
137 
138   // Now we can visit ourself
139   Order.push_back(GV);
140   Visited.insert(GV);
141   Visiting.erase(GV);
142 }
143 
144 void NVPTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
145   MCInst Inst;
146   lowerToMCInst(MI, Inst);
147   EmitToStreamer(*OutStreamer, Inst);
148 }
149 
150 // Handle symbol backtracking for targets that do not support image handles
151 bool NVPTXAsmPrinter::lowerImageHandleOperand(const MachineInstr *MI,
152                                            unsigned OpNo, MCOperand &MCOp) {
153   const MachineOperand &MO = MI->getOperand(OpNo);
154   const MCInstrDesc &MCID = MI->getDesc();
155 
156   if (MCID.TSFlags & NVPTXII::IsTexFlag) {
157     // This is a texture fetch, so operand 4 is a texref and operand 5 is
158     // a samplerref
159     if (OpNo == 4 && MO.isImm()) {
160       lowerImageHandleSymbol(MO.getImm(), MCOp);
161       return true;
162     }
163     if (OpNo == 5 && MO.isImm() && !(MCID.TSFlags & NVPTXII::IsTexModeUnifiedFlag)) {
164       lowerImageHandleSymbol(MO.getImm(), MCOp);
165       return true;
166     }
167 
168     return false;
169   } else if (MCID.TSFlags & NVPTXII::IsSuldMask) {
170     unsigned VecSize =
171       1 << (((MCID.TSFlags & NVPTXII::IsSuldMask) >> NVPTXII::IsSuldShift) - 1);
172 
173     // For a surface load of vector size N, the Nth operand will be the surfref
174     if (OpNo == VecSize && MO.isImm()) {
175       lowerImageHandleSymbol(MO.getImm(), MCOp);
176       return true;
177     }
178 
179     return false;
180   } else if (MCID.TSFlags & NVPTXII::IsSustFlag) {
181     // This is a surface store, so operand 0 is a surfref
182     if (OpNo == 0 && MO.isImm()) {
183       lowerImageHandleSymbol(MO.getImm(), MCOp);
184       return true;
185     }
186 
187     return false;
188   } else if (MCID.TSFlags & NVPTXII::IsSurfTexQueryFlag) {
189     // This is a query, so operand 1 is a surfref/texref
190     if (OpNo == 1 && MO.isImm()) {
191       lowerImageHandleSymbol(MO.getImm(), MCOp);
192       return true;
193     }
194 
195     return false;
196   }
197 
198   return false;
199 }
200 
201 void NVPTXAsmPrinter::lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp) {
202   // Ewwww
203   LLVMTargetMachine &TM = const_cast<LLVMTargetMachine&>(MF->getTarget());
204   NVPTXTargetMachine &nvTM = static_cast<NVPTXTargetMachine&>(TM);
205   const NVPTXMachineFunctionInfo *MFI = MF->getInfo<NVPTXMachineFunctionInfo>();
206   const char *Sym = MFI->getImageHandleSymbol(Index);
207   std::string *SymNamePtr =
208     nvTM.getManagedStrPool()->getManagedString(Sym);
209   MCOp = GetSymbolRef(OutContext.getOrCreateSymbol(StringRef(*SymNamePtr)));
210 }
211 
212 void NVPTXAsmPrinter::lowerToMCInst(const MachineInstr *MI, MCInst &OutMI) {
213   OutMI.setOpcode(MI->getOpcode());
214   // Special: Do not mangle symbol operand of CALL_PROTOTYPE
215   if (MI->getOpcode() == NVPTX::CALL_PROTOTYPE) {
216     const MachineOperand &MO = MI->getOperand(0);
217     OutMI.addOperand(GetSymbolRef(
218       OutContext.getOrCreateSymbol(Twine(MO.getSymbolName()))));
219     return;
220   }
221 
222   const NVPTXSubtarget &STI = MI->getMF()->getSubtarget<NVPTXSubtarget>();
223   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
224     const MachineOperand &MO = MI->getOperand(i);
225 
226     MCOperand MCOp;
227     if (!STI.hasImageHandles()) {
228       if (lowerImageHandleOperand(MI, i, MCOp)) {
229         OutMI.addOperand(MCOp);
230         continue;
231       }
232     }
233 
234     if (lowerOperand(MO, MCOp))
235       OutMI.addOperand(MCOp);
236   }
237 }
238 
239 bool NVPTXAsmPrinter::lowerOperand(const MachineOperand &MO,
240                                    MCOperand &MCOp) {
241   switch (MO.getType()) {
242   default: llvm_unreachable("unknown operand type");
243   case MachineOperand::MO_Register:
244     MCOp = MCOperand::createReg(encodeVirtualRegister(MO.getReg()));
245     break;
246   case MachineOperand::MO_Immediate:
247     MCOp = MCOperand::createImm(MO.getImm());
248     break;
249   case MachineOperand::MO_MachineBasicBlock:
250     MCOp = MCOperand::createExpr(MCSymbolRefExpr::create(
251         MO.getMBB()->getSymbol(), OutContext));
252     break;
253   case MachineOperand::MO_ExternalSymbol:
254     MCOp = GetSymbolRef(GetExternalSymbolSymbol(MO.getSymbolName()));
255     break;
256   case MachineOperand::MO_GlobalAddress:
257     MCOp = GetSymbolRef(getSymbol(MO.getGlobal()));
258     break;
259   case MachineOperand::MO_FPImmediate: {
260     const ConstantFP *Cnt = MO.getFPImm();
261     const APFloat &Val = Cnt->getValueAPF();
262 
263     switch (Cnt->getType()->getTypeID()) {
264     default: report_fatal_error("Unsupported FP type"); break;
265     case Type::HalfTyID:
266       MCOp = MCOperand::createExpr(
267         NVPTXFloatMCExpr::createConstantFPHalf(Val, OutContext));
268       break;
269     case Type::FloatTyID:
270       MCOp = MCOperand::createExpr(
271         NVPTXFloatMCExpr::createConstantFPSingle(Val, OutContext));
272       break;
273     case Type::DoubleTyID:
274       MCOp = MCOperand::createExpr(
275         NVPTXFloatMCExpr::createConstantFPDouble(Val, OutContext));
276       break;
277     }
278     break;
279   }
280   }
281   return true;
282 }
283 
284 unsigned NVPTXAsmPrinter::encodeVirtualRegister(unsigned Reg) {
285   if (Register::isVirtualRegister(Reg)) {
286     const TargetRegisterClass *RC = MRI->getRegClass(Reg);
287 
288     DenseMap<unsigned, unsigned> &RegMap = VRegMapping[RC];
289     unsigned RegNum = RegMap[Reg];
290 
291     // Encode the register class in the upper 4 bits
292     // Must be kept in sync with NVPTXInstPrinter::printRegName
293     unsigned Ret = 0;
294     if (RC == &NVPTX::Int1RegsRegClass) {
295       Ret = (1 << 28);
296     } else if (RC == &NVPTX::Int16RegsRegClass) {
297       Ret = (2 << 28);
298     } else if (RC == &NVPTX::Int32RegsRegClass) {
299       Ret = (3 << 28);
300     } else if (RC == &NVPTX::Int64RegsRegClass) {
301       Ret = (4 << 28);
302     } else if (RC == &NVPTX::Float32RegsRegClass) {
303       Ret = (5 << 28);
304     } else if (RC == &NVPTX::Float64RegsRegClass) {
305       Ret = (6 << 28);
306     } else if (RC == &NVPTX::Float16RegsRegClass) {
307       Ret = (7 << 28);
308     } else if (RC == &NVPTX::Float16x2RegsRegClass) {
309       Ret = (8 << 28);
310     } else {
311       report_fatal_error("Bad register class");
312     }
313 
314     // Insert the vreg number
315     Ret |= (RegNum & 0x0FFFFFFF);
316     return Ret;
317   } else {
318     // Some special-use registers are actually physical registers.
319     // Encode this as the register class ID of 0 and the real register ID.
320     return Reg & 0x0FFFFFFF;
321   }
322 }
323 
324 MCOperand NVPTXAsmPrinter::GetSymbolRef(const MCSymbol *Symbol) {
325   const MCExpr *Expr;
326   Expr = MCSymbolRefExpr::create(Symbol, MCSymbolRefExpr::VK_None,
327                                  OutContext);
328   return MCOperand::createExpr(Expr);
329 }
330 
331 void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
332   const DataLayout &DL = getDataLayout();
333   const NVPTXSubtarget &STI = TM.getSubtarget<NVPTXSubtarget>(*F);
334   const TargetLowering *TLI = STI.getTargetLowering();
335 
336   Type *Ty = F->getReturnType();
337 
338   bool isABI = (STI.getSmVersion() >= 20);
339 
340   if (Ty->getTypeID() == Type::VoidTyID)
341     return;
342 
343   O << " (";
344 
345   if (isABI) {
346     if (Ty->isFloatingPointTy() || (Ty->isIntegerTy() && !Ty->isIntegerTy(128))) {
347       unsigned size = 0;
348       if (auto *ITy = dyn_cast<IntegerType>(Ty)) {
349         size = ITy->getBitWidth();
350       } else {
351         assert(Ty->isFloatingPointTy() && "Floating point type expected here");
352         size = Ty->getPrimitiveSizeInBits();
353       }
354       // PTX ABI requires all scalar return values to be at least 32
355       // bits in size.  fp16 normally uses .b16 as its storage type in
356       // PTX, so its size must be adjusted here, too.
357       if (size < 32)
358         size = 32;
359 
360       O << ".param .b" << size << " func_retval0";
361     } else if (isa<PointerType>(Ty)) {
362       O << ".param .b" << TLI->getPointerTy(DL).getSizeInBits()
363         << " func_retval0";
364     } else if (Ty->isAggregateType() || Ty->isVectorTy() || Ty->isIntegerTy(128)) {
365       unsigned totalsz = DL.getTypeAllocSize(Ty);
366       unsigned retAlignment = 0;
367       if (!getAlign(*F, 0, retAlignment))
368         retAlignment = DL.getABITypeAlignment(Ty);
369       O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
370         << "]";
371     } else
372       llvm_unreachable("Unknown return type");
373   } else {
374     SmallVector<EVT, 16> vtparts;
375     ComputeValueVTs(*TLI, DL, Ty, vtparts);
376     unsigned idx = 0;
377     for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
378       unsigned elems = 1;
379       EVT elemtype = vtparts[i];
380       if (vtparts[i].isVector()) {
381         elems = vtparts[i].getVectorNumElements();
382         elemtype = vtparts[i].getVectorElementType();
383       }
384 
385       for (unsigned j = 0, je = elems; j != je; ++j) {
386         unsigned sz = elemtype.getSizeInBits();
387         if (elemtype.isInteger() && (sz < 32))
388           sz = 32;
389         O << ".reg .b" << sz << " func_retval" << idx;
390         if (j < je - 1)
391           O << ", ";
392         ++idx;
393       }
394       if (i < e - 1)
395         O << ", ";
396     }
397   }
398   O << ") ";
399 }
400 
401 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
402                                         raw_ostream &O) {
403   const Function &F = MF.getFunction();
404   printReturnValStr(&F, O);
405 }
406 
407 // Return true if MBB is the header of a loop marked with
408 // llvm.loop.unroll.disable.
409 // TODO: consider "#pragma unroll 1" which is equivalent to "#pragma nounroll".
410 bool NVPTXAsmPrinter::isLoopHeaderOfNoUnroll(
411     const MachineBasicBlock &MBB) const {
412   MachineLoopInfo &LI = getAnalysis<MachineLoopInfo>();
413   // We insert .pragma "nounroll" only to the loop header.
414   if (!LI.isLoopHeader(&MBB))
415     return false;
416 
417   // llvm.loop.unroll.disable is marked on the back edges of a loop. Therefore,
418   // we iterate through each back edge of the loop with header MBB, and check
419   // whether its metadata contains llvm.loop.unroll.disable.
420   for (auto I = MBB.pred_begin(); I != MBB.pred_end(); ++I) {
421     const MachineBasicBlock *PMBB = *I;
422     if (LI.getLoopFor(PMBB) != LI.getLoopFor(&MBB)) {
423       // Edges from other loops to MBB are not back edges.
424       continue;
425     }
426     if (const BasicBlock *PBB = PMBB->getBasicBlock()) {
427       if (MDNode *LoopID =
428               PBB->getTerminator()->getMetadata(LLVMContext::MD_loop)) {
429         if (GetUnrollMetadata(LoopID, "llvm.loop.unroll.disable"))
430           return true;
431       }
432     }
433   }
434   return false;
435 }
436 
437 void NVPTXAsmPrinter::EmitBasicBlockStart(const MachineBasicBlock &MBB) {
438   AsmPrinter::EmitBasicBlockStart(MBB);
439   if (isLoopHeaderOfNoUnroll(MBB))
440     OutStreamer->EmitRawText(StringRef("\t.pragma \"nounroll\";\n"));
441 }
442 
443 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
444   SmallString<128> Str;
445   raw_svector_ostream O(Str);
446 
447   if (!GlobalsEmitted) {
448     emitGlobals(*MF->getFunction().getParent());
449     GlobalsEmitted = true;
450   }
451 
452   // Set up
453   MRI = &MF->getRegInfo();
454   F = &MF->getFunction();
455   emitLinkageDirective(F, O);
456   if (isKernelFunction(*F))
457     O << ".entry ";
458   else {
459     O << ".func ";
460     printReturnValStr(*MF, O);
461   }
462 
463   CurrentFnSym->print(O, MAI);
464 
465   emitFunctionParamList(*MF, O);
466 
467   if (isKernelFunction(*F))
468     emitKernelFunctionDirectives(*F, O);
469 
470   OutStreamer->EmitRawText(O.str());
471 
472   VRegMapping.clear();
473   // Emit open brace for function body.
474   OutStreamer->EmitRawText(StringRef("{\n"));
475   setAndEmitFunctionVirtualRegisters(*MF);
476   // Emit initial .loc debug directive for correct relocation symbol data.
477   if (MMI && MMI->hasDebugInfo())
478     emitInitialRawDwarfLocDirective(*MF);
479 }
480 
481 bool NVPTXAsmPrinter::runOnMachineFunction(MachineFunction &F) {
482   bool Result = AsmPrinter::runOnMachineFunction(F);
483   // Emit closing brace for the body of function F.
484   // The closing brace must be emitted here because we need to emit additional
485   // debug labels/data after the last basic block.
486   // We need to emit the closing brace here because we don't have function that
487   // finished emission of the function body.
488   OutStreamer->EmitRawText(StringRef("}\n"));
489   return Result;
490 }
491 
492 void NVPTXAsmPrinter::EmitFunctionBodyStart() {
493   SmallString<128> Str;
494   raw_svector_ostream O(Str);
495   emitDemotedVars(&MF->getFunction(), O);
496   OutStreamer->EmitRawText(O.str());
497 }
498 
499 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
500   VRegMapping.clear();
501 }
502 
503 const MCSymbol *NVPTXAsmPrinter::getFunctionFrameSymbol() const {
504     SmallString<128> Str;
505     raw_svector_ostream(Str) << DEPOTNAME << getFunctionNumber();
506     return OutContext.getOrCreateSymbol(Str);
507 }
508 
509 void NVPTXAsmPrinter::emitImplicitDef(const MachineInstr *MI) const {
510   Register RegNo = MI->getOperand(0).getReg();
511   if (Register::isVirtualRegister(RegNo)) {
512     OutStreamer->AddComment(Twine("implicit-def: ") +
513                             getVirtualRegisterName(RegNo));
514   } else {
515     const NVPTXSubtarget &STI = MI->getMF()->getSubtarget<NVPTXSubtarget>();
516     OutStreamer->AddComment(Twine("implicit-def: ") +
517                             STI.getRegisterInfo()->getName(RegNo));
518   }
519   OutStreamer->AddBlankLine();
520 }
521 
522 void NVPTXAsmPrinter::emitKernelFunctionDirectives(const Function &F,
523                                                    raw_ostream &O) const {
524   // If the NVVM IR has some of reqntid* specified, then output
525   // the reqntid directive, and set the unspecified ones to 1.
526   // If none of reqntid* is specified, don't output reqntid directive.
527   unsigned reqntidx, reqntidy, reqntidz;
528   bool specified = false;
529   if (!getReqNTIDx(F, reqntidx))
530     reqntidx = 1;
531   else
532     specified = true;
533   if (!getReqNTIDy(F, reqntidy))
534     reqntidy = 1;
535   else
536     specified = true;
537   if (!getReqNTIDz(F, reqntidz))
538     reqntidz = 1;
539   else
540     specified = true;
541 
542   if (specified)
543     O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
544       << "\n";
545 
546   // If the NVVM IR has some of maxntid* specified, then output
547   // the maxntid directive, and set the unspecified ones to 1.
548   // If none of maxntid* is specified, don't output maxntid directive.
549   unsigned maxntidx, maxntidy, maxntidz;
550   specified = false;
551   if (!getMaxNTIDx(F, maxntidx))
552     maxntidx = 1;
553   else
554     specified = true;
555   if (!getMaxNTIDy(F, maxntidy))
556     maxntidy = 1;
557   else
558     specified = true;
559   if (!getMaxNTIDz(F, maxntidz))
560     maxntidz = 1;
561   else
562     specified = true;
563 
564   if (specified)
565     O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
566       << "\n";
567 
568   unsigned mincta;
569   if (getMinCTASm(F, mincta))
570     O << ".minnctapersm " << mincta << "\n";
571 
572   unsigned maxnreg;
573   if (getMaxNReg(F, maxnreg))
574     O << ".maxnreg " << maxnreg << "\n";
575 }
576 
577 std::string
578 NVPTXAsmPrinter::getVirtualRegisterName(unsigned Reg) const {
579   const TargetRegisterClass *RC = MRI->getRegClass(Reg);
580 
581   std::string Name;
582   raw_string_ostream NameStr(Name);
583 
584   VRegRCMap::const_iterator I = VRegMapping.find(RC);
585   assert(I != VRegMapping.end() && "Bad register class");
586   const DenseMap<unsigned, unsigned> &RegMap = I->second;
587 
588   VRegMap::const_iterator VI = RegMap.find(Reg);
589   assert(VI != RegMap.end() && "Bad virtual register");
590   unsigned MappedVR = VI->second;
591 
592   NameStr << getNVPTXRegClassStr(RC) << MappedVR;
593 
594   NameStr.flush();
595   return Name;
596 }
597 
598 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr,
599                                           raw_ostream &O) {
600   O << getVirtualRegisterName(vr);
601 }
602 
603 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
604   emitLinkageDirective(F, O);
605   if (isKernelFunction(*F))
606     O << ".entry ";
607   else
608     O << ".func ";
609   printReturnValStr(F, O);
610   getSymbol(F)->print(O, MAI);
611   O << "\n";
612   emitFunctionParamList(F, O);
613   O << ";\n";
614 }
615 
616 static bool usedInGlobalVarDef(const Constant *C) {
617   if (!C)
618     return false;
619 
620   if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
621     return GV->getName() != "llvm.used";
622   }
623 
624   for (const User *U : C->users())
625     if (const Constant *C = dyn_cast<Constant>(U))
626       if (usedInGlobalVarDef(C))
627         return true;
628 
629   return false;
630 }
631 
632 static bool usedInOneFunc(const User *U, Function const *&oneFunc) {
633   if (const GlobalVariable *othergv = dyn_cast<GlobalVariable>(U)) {
634     if (othergv->getName() == "llvm.used")
635       return true;
636   }
637 
638   if (const Instruction *instr = dyn_cast<Instruction>(U)) {
639     if (instr->getParent() && instr->getParent()->getParent()) {
640       const Function *curFunc = instr->getParent()->getParent();
641       if (oneFunc && (curFunc != oneFunc))
642         return false;
643       oneFunc = curFunc;
644       return true;
645     } else
646       return false;
647   }
648 
649   for (const User *UU : U->users())
650     if (!usedInOneFunc(UU, oneFunc))
651       return false;
652 
653   return true;
654 }
655 
656 /* Find out if a global variable can be demoted to local scope.
657  * Currently, this is valid for CUDA shared variables, which have local
658  * scope and global lifetime. So the conditions to check are :
659  * 1. Is the global variable in shared address space?
660  * 2. Does it have internal linkage?
661  * 3. Is the global variable referenced only in one function?
662  */
663 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
664   if (!gv->hasInternalLinkage())
665     return false;
666   PointerType *Pty = gv->getType();
667   if (Pty->getAddressSpace() != ADDRESS_SPACE_SHARED)
668     return false;
669 
670   const Function *oneFunc = nullptr;
671 
672   bool flag = usedInOneFunc(gv, oneFunc);
673   if (!flag)
674     return false;
675   if (!oneFunc)
676     return false;
677   f = oneFunc;
678   return true;
679 }
680 
681 static bool useFuncSeen(const Constant *C,
682                         DenseMap<const Function *, bool> &seenMap) {
683   for (const User *U : C->users()) {
684     if (const Constant *cu = dyn_cast<Constant>(U)) {
685       if (useFuncSeen(cu, seenMap))
686         return true;
687     } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
688       const BasicBlock *bb = I->getParent();
689       if (!bb)
690         continue;
691       const Function *caller = bb->getParent();
692       if (!caller)
693         continue;
694       if (seenMap.find(caller) != seenMap.end())
695         return true;
696     }
697   }
698   return false;
699 }
700 
701 void NVPTXAsmPrinter::emitDeclarations(const Module &M, raw_ostream &O) {
702   DenseMap<const Function *, bool> seenMap;
703   for (Module::const_iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
704     const Function *F = &*FI;
705 
706     if (F->getAttributes().hasFnAttribute("nvptx-libcall-callee")) {
707       emitDeclaration(F, O);
708       continue;
709     }
710 
711     if (F->isDeclaration()) {
712       if (F->use_empty())
713         continue;
714       if (F->getIntrinsicID())
715         continue;
716       emitDeclaration(F, O);
717       continue;
718     }
719     for (const User *U : F->users()) {
720       if (const Constant *C = dyn_cast<Constant>(U)) {
721         if (usedInGlobalVarDef(C)) {
722           // The use is in the initialization of a global variable
723           // that is a function pointer, so print a declaration
724           // for the original function
725           emitDeclaration(F, O);
726           break;
727         }
728         // Emit a declaration of this function if the function that
729         // uses this constant expr has already been seen.
730         if (useFuncSeen(C, seenMap)) {
731           emitDeclaration(F, O);
732           break;
733         }
734       }
735 
736       if (!isa<Instruction>(U))
737         continue;
738       const Instruction *instr = cast<Instruction>(U);
739       const BasicBlock *bb = instr->getParent();
740       if (!bb)
741         continue;
742       const Function *caller = bb->getParent();
743       if (!caller)
744         continue;
745 
746       // If a caller has already been seen, then the caller is
747       // appearing in the module before the callee. so print out
748       // a declaration for the callee.
749       if (seenMap.find(caller) != seenMap.end()) {
750         emitDeclaration(F, O);
751         break;
752       }
753     }
754     seenMap[F] = true;
755   }
756 }
757 
758 static bool isEmptyXXStructor(GlobalVariable *GV) {
759   if (!GV) return true;
760   const ConstantArray *InitList = dyn_cast<ConstantArray>(GV->getInitializer());
761   if (!InitList) return true;  // Not an array; we don't know how to parse.
762   return InitList->getNumOperands() == 0;
763 }
764 
765 bool NVPTXAsmPrinter::doInitialization(Module &M) {
766   // Construct a default subtarget off of the TargetMachine defaults. The
767   // rest of NVPTX isn't friendly to change subtargets per function and
768   // so the default TargetMachine will have all of the options.
769   const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
770   const auto* STI = static_cast<const NVPTXSubtarget*>(NTM.getSubtargetImpl());
771 
772   if (M.alias_size()) {
773     report_fatal_error("Module has aliases, which NVPTX does not support.");
774     return true; // error
775   }
776   if (!isEmptyXXStructor(M.getNamedGlobal("llvm.global_ctors"))) {
777     report_fatal_error(
778         "Module has a nontrivial global ctor, which NVPTX does not support.");
779     return true;  // error
780   }
781   if (!isEmptyXXStructor(M.getNamedGlobal("llvm.global_dtors"))) {
782     report_fatal_error(
783         "Module has a nontrivial global dtor, which NVPTX does not support.");
784     return true;  // error
785   }
786 
787   SmallString<128> Str1;
788   raw_svector_ostream OS1(Str1);
789 
790   // We need to call the parent's one explicitly.
791   bool Result = AsmPrinter::doInitialization(M);
792 
793   // Emit header before any dwarf directives are emitted below.
794   emitHeader(M, OS1, *STI);
795   OutStreamer->EmitRawText(OS1.str());
796 
797   // Emit module-level inline asm if it exists.
798   if (!M.getModuleInlineAsm().empty()) {
799     OutStreamer->AddComment("Start of file scope inline assembly");
800     OutStreamer->AddBlankLine();
801     OutStreamer->EmitRawText(StringRef(M.getModuleInlineAsm()));
802     OutStreamer->AddBlankLine();
803     OutStreamer->AddComment("End of file scope inline assembly");
804     OutStreamer->AddBlankLine();
805   }
806 
807   GlobalsEmitted = false;
808 
809   return Result;
810 }
811 
812 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
813   SmallString<128> Str2;
814   raw_svector_ostream OS2(Str2);
815 
816   emitDeclarations(M, OS2);
817 
818   // As ptxas does not support forward references of globals, we need to first
819   // sort the list of module-level globals in def-use order. We visit each
820   // global variable in order, and ensure that we emit it *after* its dependent
821   // globals. We use a little extra memory maintaining both a set and a list to
822   // have fast searches while maintaining a strict ordering.
823   SmallVector<const GlobalVariable *, 8> Globals;
824   DenseSet<const GlobalVariable *> GVVisited;
825   DenseSet<const GlobalVariable *> GVVisiting;
826 
827   // Visit each global variable, in order
828   for (const GlobalVariable &I : M.globals())
829     VisitGlobalVariableForEmission(&I, Globals, GVVisited, GVVisiting);
830 
831   assert(GVVisited.size() == M.getGlobalList().size() &&
832          "Missed a global variable");
833   assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
834 
835   // Print out module-level global variables in proper order
836   for (unsigned i = 0, e = Globals.size(); i != e; ++i)
837     printModuleLevelGV(Globals[i], OS2);
838 
839   OS2 << '\n';
840 
841   OutStreamer->EmitRawText(OS2.str());
842 }
843 
844 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O,
845                                  const NVPTXSubtarget &STI) {
846   O << "//\n";
847   O << "// Generated by LLVM NVPTX Back-End\n";
848   O << "//\n";
849   O << "\n";
850 
851   unsigned PTXVersion = STI.getPTXVersion();
852   O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
853 
854   O << ".target ";
855   O << STI.getTargetName();
856 
857   const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
858   if (NTM.getDrvInterface() == NVPTX::NVCL)
859     O << ", texmode_independent";
860 
861   bool HasFullDebugInfo = false;
862   for (DICompileUnit *CU : M.debug_compile_units()) {
863     switch(CU->getEmissionKind()) {
864     case DICompileUnit::NoDebug:
865     case DICompileUnit::DebugDirectivesOnly:
866       break;
867     case DICompileUnit::LineTablesOnly:
868     case DICompileUnit::FullDebug:
869       HasFullDebugInfo = true;
870       break;
871     }
872     if (HasFullDebugInfo)
873       break;
874   }
875   if (MMI && MMI->hasDebugInfo() && HasFullDebugInfo)
876     O << ", debug";
877 
878   O << "\n";
879 
880   O << ".address_size ";
881   if (NTM.is64Bit())
882     O << "64";
883   else
884     O << "32";
885   O << "\n";
886 
887   O << "\n";
888 }
889 
890 bool NVPTXAsmPrinter::doFinalization(Module &M) {
891   bool HasDebugInfo = MMI && MMI->hasDebugInfo();
892 
893   // If we did not emit any functions, then the global declarations have not
894   // yet been emitted.
895   if (!GlobalsEmitted) {
896     emitGlobals(M);
897     GlobalsEmitted = true;
898   }
899 
900   // XXX Temproarily remove global variables so that doFinalization() will not
901   // emit them again (global variables are emitted at beginning).
902 
903   Module::GlobalListType &global_list = M.getGlobalList();
904   int i, n = global_list.size();
905   GlobalVariable **gv_array = new GlobalVariable *[n];
906 
907   // first, back-up GlobalVariable in gv_array
908   i = 0;
909   for (Module::global_iterator I = global_list.begin(), E = global_list.end();
910        I != E; ++I)
911     gv_array[i++] = &*I;
912 
913   // second, empty global_list
914   while (!global_list.empty())
915     global_list.remove(global_list.begin());
916 
917   // call doFinalization
918   bool ret = AsmPrinter::doFinalization(M);
919 
920   // now we restore global variables
921   for (i = 0; i < n; i++)
922     global_list.insert(global_list.end(), gv_array[i]);
923 
924   clearAnnotationCache(&M);
925 
926   delete[] gv_array;
927   // Close the last emitted section
928   if (HasDebugInfo) {
929     static_cast<NVPTXTargetStreamer *>(OutStreamer->getTargetStreamer())
930         ->closeLastSection();
931     // Emit empty .debug_loc section for better support of the empty files.
932     OutStreamer->EmitRawText("\t.section\t.debug_loc\t{\t}");
933   }
934 
935   // Output last DWARF .file directives, if any.
936   static_cast<NVPTXTargetStreamer *>(OutStreamer->getTargetStreamer())
937       ->outputDwarfFileDirectives();
938 
939   return ret;
940 
941   //bool Result = AsmPrinter::doFinalization(M);
942   // Instead of calling the parents doFinalization, we may
943   // clone parents doFinalization and customize here.
944   // Currently, we if NVISA out the EmitGlobals() in
945   // parent's doFinalization, which is too intrusive.
946   //
947   // Same for the doInitialization.
948   //return Result;
949 }
950 
951 // This function emits appropriate linkage directives for
952 // functions and global variables.
953 //
954 // extern function declaration            -> .extern
955 // extern function definition             -> .visible
956 // external global variable with init     -> .visible
957 // external without init                  -> .extern
958 // appending                              -> not allowed, assert.
959 // for any linkage other than
960 // internal, private, linker_private,
961 // linker_private_weak, linker_private_weak_def_auto,
962 // we emit                                -> .weak.
963 
964 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
965                                            raw_ostream &O) {
966   if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == NVPTX::CUDA) {
967     if (V->hasExternalLinkage()) {
968       if (isa<GlobalVariable>(V)) {
969         const GlobalVariable *GVar = cast<GlobalVariable>(V);
970         if (GVar) {
971           if (GVar->hasInitializer())
972             O << ".visible ";
973           else
974             O << ".extern ";
975         }
976       } else if (V->isDeclaration())
977         O << ".extern ";
978       else
979         O << ".visible ";
980     } else if (V->hasAppendingLinkage()) {
981       std::string msg;
982       msg.append("Error: ");
983       msg.append("Symbol ");
984       if (V->hasName())
985         msg.append(V->getName());
986       msg.append("has unsupported appending linkage type");
987       llvm_unreachable(msg.c_str());
988     } else if (!V->hasInternalLinkage() &&
989                !V->hasPrivateLinkage()) {
990       O << ".weak ";
991     }
992   }
993 }
994 
995 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
996                                          raw_ostream &O,
997                                          bool processDemoted) {
998   // Skip meta data
999   if (GVar->hasSection()) {
1000     if (GVar->getSection() == "llvm.metadata")
1001       return;
1002   }
1003 
1004   // Skip LLVM intrinsic global variables
1005   if (GVar->getName().startswith("llvm.") ||
1006       GVar->getName().startswith("nvvm."))
1007     return;
1008 
1009   const DataLayout &DL = getDataLayout();
1010 
1011   // GlobalVariables are always constant pointers themselves.
1012   PointerType *PTy = GVar->getType();
1013   Type *ETy = GVar->getValueType();
1014 
1015   if (GVar->hasExternalLinkage()) {
1016     if (GVar->hasInitializer())
1017       O << ".visible ";
1018     else
1019       O << ".extern ";
1020   } else if (GVar->hasLinkOnceLinkage() || GVar->hasWeakLinkage() ||
1021              GVar->hasAvailableExternallyLinkage() ||
1022              GVar->hasCommonLinkage()) {
1023     O << ".weak ";
1024   }
1025 
1026   if (isTexture(*GVar)) {
1027     O << ".global .texref " << getTextureName(*GVar) << ";\n";
1028     return;
1029   }
1030 
1031   if (isSurface(*GVar)) {
1032     O << ".global .surfref " << getSurfaceName(*GVar) << ";\n";
1033     return;
1034   }
1035 
1036   if (GVar->isDeclaration()) {
1037     // (extern) declarations, no definition or initializer
1038     // Currently the only known declaration is for an automatic __local
1039     // (.shared) promoted to global.
1040     emitPTXGlobalVariable(GVar, O);
1041     O << ";\n";
1042     return;
1043   }
1044 
1045   if (isSampler(*GVar)) {
1046     O << ".global .samplerref " << getSamplerName(*GVar);
1047 
1048     const Constant *Initializer = nullptr;
1049     if (GVar->hasInitializer())
1050       Initializer = GVar->getInitializer();
1051     const ConstantInt *CI = nullptr;
1052     if (Initializer)
1053       CI = dyn_cast<ConstantInt>(Initializer);
1054     if (CI) {
1055       unsigned sample = CI->getZExtValue();
1056 
1057       O << " = { ";
1058 
1059       for (int i = 0,
1060                addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1061            i < 3; i++) {
1062         O << "addr_mode_" << i << " = ";
1063         switch (addr) {
1064         case 0:
1065           O << "wrap";
1066           break;
1067         case 1:
1068           O << "clamp_to_border";
1069           break;
1070         case 2:
1071           O << "clamp_to_edge";
1072           break;
1073         case 3:
1074           O << "wrap";
1075           break;
1076         case 4:
1077           O << "mirror";
1078           break;
1079         }
1080         O << ", ";
1081       }
1082       O << "filter_mode = ";
1083       switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1084       case 0:
1085         O << "nearest";
1086         break;
1087       case 1:
1088         O << "linear";
1089         break;
1090       case 2:
1091         llvm_unreachable("Anisotropic filtering is not supported");
1092       default:
1093         O << "nearest";
1094         break;
1095       }
1096       if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1097         O << ", force_unnormalized_coords = 1";
1098       }
1099       O << " }";
1100     }
1101 
1102     O << ";\n";
1103     return;
1104   }
1105 
1106   if (GVar->hasPrivateLinkage()) {
1107     if (strncmp(GVar->getName().data(), "unrollpragma", 12) == 0)
1108       return;
1109 
1110     // FIXME - need better way (e.g. Metadata) to avoid generating this global
1111     if (strncmp(GVar->getName().data(), "filename", 8) == 0)
1112       return;
1113     if (GVar->use_empty())
1114       return;
1115   }
1116 
1117   const Function *demotedFunc = nullptr;
1118   if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
1119     O << "// " << GVar->getName() << " has been demoted\n";
1120     if (localDecls.find(demotedFunc) != localDecls.end())
1121       localDecls[demotedFunc].push_back(GVar);
1122     else {
1123       std::vector<const GlobalVariable *> temp;
1124       temp.push_back(GVar);
1125       localDecls[demotedFunc] = temp;
1126     }
1127     return;
1128   }
1129 
1130   O << ".";
1131   emitPTXAddressSpace(PTy->getAddressSpace(), O);
1132 
1133   if (isManaged(*GVar)) {
1134     O << " .attribute(.managed)";
1135   }
1136 
1137   if (GVar->getAlignment() == 0)
1138     O << " .align " << (int)DL.getPrefTypeAlignment(ETy);
1139   else
1140     O << " .align " << GVar->getAlignment();
1141 
1142   if (ETy->isFloatingPointTy() || ETy->isPointerTy() ||
1143       (ETy->isIntegerTy() && ETy->getScalarSizeInBits() <= 64)) {
1144     O << " .";
1145     // Special case: ABI requires that we use .u8 for predicates
1146     if (ETy->isIntegerTy(1))
1147       O << "u8";
1148     else
1149       O << getPTXFundamentalTypeStr(ETy, false);
1150     O << " ";
1151     getSymbol(GVar)->print(O, MAI);
1152 
1153     // Ptx allows variable initilization only for constant and global state
1154     // spaces.
1155     if (GVar->hasInitializer()) {
1156       if ((PTy->getAddressSpace() == ADDRESS_SPACE_GLOBAL) ||
1157           (PTy->getAddressSpace() == ADDRESS_SPACE_CONST)) {
1158         const Constant *Initializer = GVar->getInitializer();
1159         // 'undef' is treated as there is no value specified.
1160         if (!Initializer->isNullValue() && !isa<UndefValue>(Initializer)) {
1161           O << " = ";
1162           printScalarConstant(Initializer, O);
1163         }
1164       } else {
1165         // The frontend adds zero-initializer to device and constant variables
1166         // that don't have an initial value, and UndefValue to shared
1167         // variables, so skip warning for this case.
1168         if (!GVar->getInitializer()->isNullValue() &&
1169             !isa<UndefValue>(GVar->getInitializer())) {
1170           report_fatal_error("initial value of '" + GVar->getName() +
1171                              "' is not allowed in addrspace(" +
1172                              Twine(PTy->getAddressSpace()) + ")");
1173         }
1174       }
1175     }
1176   } else {
1177     unsigned int ElementSize = 0;
1178 
1179     // Although PTX has direct support for struct type and array type and
1180     // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1181     // targets that support these high level field accesses. Structs, arrays
1182     // and vectors are lowered into arrays of bytes.
1183     switch (ETy->getTypeID()) {
1184     case Type::IntegerTyID: // Integers larger than 64 bits
1185     case Type::StructTyID:
1186     case Type::ArrayTyID:
1187     case Type::VectorTyID:
1188       ElementSize = DL.getTypeStoreSize(ETy);
1189       // Ptx allows variable initilization only for constant and
1190       // global state spaces.
1191       if (((PTy->getAddressSpace() == ADDRESS_SPACE_GLOBAL) ||
1192            (PTy->getAddressSpace() == ADDRESS_SPACE_CONST)) &&
1193           GVar->hasInitializer()) {
1194         const Constant *Initializer = GVar->getInitializer();
1195         if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1196           AggBuffer aggBuffer(ElementSize, O, *this);
1197           bufferAggregateConstant(Initializer, &aggBuffer);
1198           if (aggBuffer.numSymbols) {
1199             if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit()) {
1200               O << " .u64 ";
1201               getSymbol(GVar)->print(O, MAI);
1202               O << "[";
1203               O << ElementSize / 8;
1204             } else {
1205               O << " .u32 ";
1206               getSymbol(GVar)->print(O, MAI);
1207               O << "[";
1208               O << ElementSize / 4;
1209             }
1210             O << "]";
1211           } else {
1212             O << " .b8 ";
1213             getSymbol(GVar)->print(O, MAI);
1214             O << "[";
1215             O << ElementSize;
1216             O << "]";
1217           }
1218           O << " = {";
1219           aggBuffer.print();
1220           O << "}";
1221         } else {
1222           O << " .b8 ";
1223           getSymbol(GVar)->print(O, MAI);
1224           if (ElementSize) {
1225             O << "[";
1226             O << ElementSize;
1227             O << "]";
1228           }
1229         }
1230       } else {
1231         O << " .b8 ";
1232         getSymbol(GVar)->print(O, MAI);
1233         if (ElementSize) {
1234           O << "[";
1235           O << ElementSize;
1236           O << "]";
1237         }
1238       }
1239       break;
1240     default:
1241       llvm_unreachable("type not supported yet");
1242     }
1243   }
1244   O << ";\n";
1245 }
1246 
1247 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1248   if (localDecls.find(f) == localDecls.end())
1249     return;
1250 
1251   std::vector<const GlobalVariable *> &gvars = localDecls[f];
1252 
1253   for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1254     O << "\t// demoted variable\n\t";
1255     printModuleLevelGV(gvars[i], O, true);
1256   }
1257 }
1258 
1259 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1260                                           raw_ostream &O) const {
1261   switch (AddressSpace) {
1262   case ADDRESS_SPACE_LOCAL:
1263     O << "local";
1264     break;
1265   case ADDRESS_SPACE_GLOBAL:
1266     O << "global";
1267     break;
1268   case ADDRESS_SPACE_CONST:
1269     O << "const";
1270     break;
1271   case ADDRESS_SPACE_SHARED:
1272     O << "shared";
1273     break;
1274   default:
1275     report_fatal_error("Bad address space found while emitting PTX: " +
1276                        llvm::Twine(AddressSpace));
1277     break;
1278   }
1279 }
1280 
1281 std::string
1282 NVPTXAsmPrinter::getPTXFundamentalTypeStr(Type *Ty, bool useB4PTR) const {
1283   switch (Ty->getTypeID()) {
1284   default:
1285     llvm_unreachable("unexpected type");
1286     break;
1287   case Type::IntegerTyID: {
1288     unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1289     if (NumBits == 1)
1290       return "pred";
1291     else if (NumBits <= 64) {
1292       std::string name = "u";
1293       return name + utostr(NumBits);
1294     } else {
1295       llvm_unreachable("Integer too large");
1296       break;
1297     }
1298     break;
1299   }
1300   case Type::HalfTyID:
1301     // fp16 is stored as .b16 for compatibility with pre-sm_53 PTX assembly.
1302     return "b16";
1303   case Type::FloatTyID:
1304     return "f32";
1305   case Type::DoubleTyID:
1306     return "f64";
1307   case Type::PointerTyID:
1308     if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit())
1309       if (useB4PTR)
1310         return "b64";
1311       else
1312         return "u64";
1313     else if (useB4PTR)
1314       return "b32";
1315     else
1316       return "u32";
1317   }
1318   llvm_unreachable("unexpected type");
1319   return nullptr;
1320 }
1321 
1322 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1323                                             raw_ostream &O) {
1324   const DataLayout &DL = getDataLayout();
1325 
1326   // GlobalVariables are always constant pointers themselves.
1327   Type *ETy = GVar->getValueType();
1328 
1329   O << ".";
1330   emitPTXAddressSpace(GVar->getType()->getAddressSpace(), O);
1331   if (GVar->getAlignment() == 0)
1332     O << " .align " << (int)DL.getPrefTypeAlignment(ETy);
1333   else
1334     O << " .align " << GVar->getAlignment();
1335 
1336   // Special case for i128
1337   if (ETy->isIntegerTy(128)) {
1338     O << " .b8 ";
1339     getSymbol(GVar)->print(O, MAI);
1340     O << "[16]";
1341     return;
1342   }
1343 
1344   if (ETy->isFloatingPointTy() || ETy->isIntOrPtrTy()) {
1345     O << " .";
1346     O << getPTXFundamentalTypeStr(ETy);
1347     O << " ";
1348     getSymbol(GVar)->print(O, MAI);
1349     return;
1350   }
1351 
1352   int64_t ElementSize = 0;
1353 
1354   // Although PTX has direct support for struct type and array type and LLVM IR
1355   // is very similar to PTX, the LLVM CodeGen does not support for targets that
1356   // support these high level field accesses. Structs and arrays are lowered
1357   // into arrays of bytes.
1358   switch (ETy->getTypeID()) {
1359   case Type::StructTyID:
1360   case Type::ArrayTyID:
1361   case Type::VectorTyID:
1362     ElementSize = DL.getTypeStoreSize(ETy);
1363     O << " .b8 ";
1364     getSymbol(GVar)->print(O, MAI);
1365     O << "[";
1366     if (ElementSize) {
1367       O << ElementSize;
1368     }
1369     O << "]";
1370     break;
1371   default:
1372     llvm_unreachable("type not supported yet");
1373   }
1374 }
1375 
1376 static unsigned int getOpenCLAlignment(const DataLayout &DL, Type *Ty) {
1377   if (Ty->isSingleValueType())
1378     return DL.getPrefTypeAlignment(Ty);
1379 
1380   auto *ATy = dyn_cast<ArrayType>(Ty);
1381   if (ATy)
1382     return getOpenCLAlignment(DL, ATy->getElementType());
1383 
1384   auto *STy = dyn_cast<StructType>(Ty);
1385   if (STy) {
1386     unsigned int alignStruct = 1;
1387     // Go through each element of the struct and find the
1388     // largest alignment.
1389     for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1390       Type *ETy = STy->getElementType(i);
1391       unsigned int align = getOpenCLAlignment(DL, ETy);
1392       if (align > alignStruct)
1393         alignStruct = align;
1394     }
1395     return alignStruct;
1396   }
1397 
1398   auto *FTy = dyn_cast<FunctionType>(Ty);
1399   if (FTy)
1400     return DL.getPointerPrefAlignment().value();
1401   return DL.getPrefTypeAlignment(Ty);
1402 }
1403 
1404 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1405                                      int paramIndex, raw_ostream &O) {
1406   getSymbol(I->getParent())->print(O, MAI);
1407   O << "_param_" << paramIndex;
1408 }
1409 
1410 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1411   const DataLayout &DL = getDataLayout();
1412   const AttributeList &PAL = F->getAttributes();
1413   const NVPTXSubtarget &STI = TM.getSubtarget<NVPTXSubtarget>(*F);
1414   const TargetLowering *TLI = STI.getTargetLowering();
1415   Function::const_arg_iterator I, E;
1416   unsigned paramIndex = 0;
1417   bool first = true;
1418   bool isKernelFunc = isKernelFunction(*F);
1419   bool isABI = (STI.getSmVersion() >= 20);
1420   bool hasImageHandles = STI.hasImageHandles();
1421   MVT thePointerTy = TLI->getPointerTy(DL);
1422 
1423   if (F->arg_empty()) {
1424     O << "()\n";
1425     return;
1426   }
1427 
1428   O << "(\n";
1429 
1430   for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1431     Type *Ty = I->getType();
1432 
1433     if (!first)
1434       O << ",\n";
1435 
1436     first = false;
1437 
1438     // Handle image/sampler parameters
1439     if (isKernelFunction(*F)) {
1440       if (isSampler(*I) || isImage(*I)) {
1441         if (isImage(*I)) {
1442           std::string sname = I->getName();
1443           if (isImageWriteOnly(*I) || isImageReadWrite(*I)) {
1444             if (hasImageHandles)
1445               O << "\t.param .u64 .ptr .surfref ";
1446             else
1447               O << "\t.param .surfref ";
1448             CurrentFnSym->print(O, MAI);
1449             O << "_param_" << paramIndex;
1450           }
1451           else { // Default image is read_only
1452             if (hasImageHandles)
1453               O << "\t.param .u64 .ptr .texref ";
1454             else
1455               O << "\t.param .texref ";
1456             CurrentFnSym->print(O, MAI);
1457             O << "_param_" << paramIndex;
1458           }
1459         } else {
1460           if (hasImageHandles)
1461             O << "\t.param .u64 .ptr .samplerref ";
1462           else
1463             O << "\t.param .samplerref ";
1464           CurrentFnSym->print(O, MAI);
1465           O << "_param_" << paramIndex;
1466         }
1467         continue;
1468       }
1469     }
1470 
1471     if (!PAL.hasParamAttribute(paramIndex, Attribute::ByVal)) {
1472       if (Ty->isAggregateType() || Ty->isVectorTy() || Ty->isIntegerTy(128)) {
1473         // Just print .param .align <a> .b8 .param[size];
1474         // <a> = PAL.getparamalignment
1475         // size = typeallocsize of element type
1476         const Align align = DL.getValueOrABITypeAlignment(
1477             PAL.getParamAlignment(paramIndex), Ty);
1478 
1479         unsigned sz = DL.getTypeAllocSize(Ty);
1480         O << "\t.param .align " << align.value() << " .b8 ";
1481         printParamName(I, paramIndex, O);
1482         O << "[" << sz << "]";
1483 
1484         continue;
1485       }
1486       // Just a scalar
1487       auto *PTy = dyn_cast<PointerType>(Ty);
1488       if (isKernelFunc) {
1489         if (PTy) {
1490           // Special handling for pointer arguments to kernel
1491           O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1492 
1493           if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() !=
1494               NVPTX::CUDA) {
1495             Type *ETy = PTy->getElementType();
1496             int addrSpace = PTy->getAddressSpace();
1497             switch (addrSpace) {
1498             default:
1499               O << ".ptr ";
1500               break;
1501             case ADDRESS_SPACE_CONST:
1502               O << ".ptr .const ";
1503               break;
1504             case ADDRESS_SPACE_SHARED:
1505               O << ".ptr .shared ";
1506               break;
1507             case ADDRESS_SPACE_GLOBAL:
1508               O << ".ptr .global ";
1509               break;
1510             }
1511             O << ".align " << (int)getOpenCLAlignment(DL, ETy) << " ";
1512           }
1513           printParamName(I, paramIndex, O);
1514           continue;
1515         }
1516 
1517         // non-pointer scalar to kernel func
1518         O << "\t.param .";
1519         // Special case: predicate operands become .u8 types
1520         if (Ty->isIntegerTy(1))
1521           O << "u8";
1522         else
1523           O << getPTXFundamentalTypeStr(Ty);
1524         O << " ";
1525         printParamName(I, paramIndex, O);
1526         continue;
1527       }
1528       // Non-kernel function, just print .param .b<size> for ABI
1529       // and .reg .b<size> for non-ABI
1530       unsigned sz = 0;
1531       if (isa<IntegerType>(Ty)) {
1532         sz = cast<IntegerType>(Ty)->getBitWidth();
1533         if (sz < 32)
1534           sz = 32;
1535       } else if (isa<PointerType>(Ty))
1536         sz = thePointerTy.getSizeInBits();
1537       else if (Ty->isHalfTy())
1538         // PTX ABI requires all scalar parameters to be at least 32
1539         // bits in size.  fp16 normally uses .b16 as its storage type
1540         // in PTX, so its size must be adjusted here, too.
1541         sz = 32;
1542       else
1543         sz = Ty->getPrimitiveSizeInBits();
1544       if (isABI)
1545         O << "\t.param .b" << sz << " ";
1546       else
1547         O << "\t.reg .b" << sz << " ";
1548       printParamName(I, paramIndex, O);
1549       continue;
1550     }
1551 
1552     // param has byVal attribute. So should be a pointer
1553     auto *PTy = dyn_cast<PointerType>(Ty);
1554     assert(PTy && "Param with byval attribute should be a pointer type");
1555     Type *ETy = PTy->getElementType();
1556 
1557     if (isABI || isKernelFunc) {
1558       // Just print .param .align <a> .b8 .param[size];
1559       // <a> = PAL.getparamalignment
1560       // size = typeallocsize of element type
1561       Align align =
1562           DL.getValueOrABITypeAlignment(PAL.getParamAlignment(paramIndex), ETy);
1563       // Work around a bug in ptxas. When PTX code takes address of
1564       // byval parameter with alignment < 4, ptxas generates code to
1565       // spill argument into memory. Alas on sm_50+ ptxas generates
1566       // SASS code that fails with misaligned access. To work around
1567       // the problem, make sure that we align byval parameters by at
1568       // least 4. Matching change must be made in LowerCall() where we
1569       // prepare parameters for the call.
1570       //
1571       // TODO: this will need to be undone when we get to support multi-TU
1572       // device-side compilation as it breaks ABI compatibility with nvcc.
1573       // Hopefully ptxas bug is fixed by then.
1574       if (!isKernelFunc && align < Align(4))
1575         align = Align(4);
1576       unsigned sz = DL.getTypeAllocSize(ETy);
1577       O << "\t.param .align " << align.value() << " .b8 ";
1578       printParamName(I, paramIndex, O);
1579       O << "[" << sz << "]";
1580       continue;
1581     } else {
1582       // Split the ETy into constituent parts and
1583       // print .param .b<size> <name> for each part.
1584       // Further, if a part is vector, print the above for
1585       // each vector element.
1586       SmallVector<EVT, 16> vtparts;
1587       ComputeValueVTs(*TLI, DL, ETy, vtparts);
1588       for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1589         unsigned elems = 1;
1590         EVT elemtype = vtparts[i];
1591         if (vtparts[i].isVector()) {
1592           elems = vtparts[i].getVectorNumElements();
1593           elemtype = vtparts[i].getVectorElementType();
1594         }
1595 
1596         for (unsigned j = 0, je = elems; j != je; ++j) {
1597           unsigned sz = elemtype.getSizeInBits();
1598           if (elemtype.isInteger() && (sz < 32))
1599             sz = 32;
1600           O << "\t.reg .b" << sz << " ";
1601           printParamName(I, paramIndex, O);
1602           if (j < je - 1)
1603             O << ",\n";
1604           ++paramIndex;
1605         }
1606         if (i < e - 1)
1607           O << ",\n";
1608       }
1609       --paramIndex;
1610       continue;
1611     }
1612   }
1613 
1614   O << "\n)\n";
1615 }
1616 
1617 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1618                                             raw_ostream &O) {
1619   const Function &F = MF.getFunction();
1620   emitFunctionParamList(&F, O);
1621 }
1622 
1623 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1624     const MachineFunction &MF) {
1625   SmallString<128> Str;
1626   raw_svector_ostream O(Str);
1627 
1628   // Map the global virtual register number to a register class specific
1629   // virtual register number starting from 1 with that class.
1630   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1631   //unsigned numRegClasses = TRI->getNumRegClasses();
1632 
1633   // Emit the Fake Stack Object
1634   const MachineFrameInfo &MFI = MF.getFrameInfo();
1635   int NumBytes = (int) MFI.getStackSize();
1636   if (NumBytes) {
1637     O << "\t.local .align " << MFI.getMaxAlignment() << " .b8 \t" << DEPOTNAME
1638       << getFunctionNumber() << "[" << NumBytes << "];\n";
1639     if (static_cast<const NVPTXTargetMachine &>(MF.getTarget()).is64Bit()) {
1640       O << "\t.reg .b64 \t%SP;\n";
1641       O << "\t.reg .b64 \t%SPL;\n";
1642     } else {
1643       O << "\t.reg .b32 \t%SP;\n";
1644       O << "\t.reg .b32 \t%SPL;\n";
1645     }
1646   }
1647 
1648   // Go through all virtual registers to establish the mapping between the
1649   // global virtual
1650   // register number and the per class virtual register number.
1651   // We use the per class virtual register number in the ptx output.
1652   unsigned int numVRs = MRI->getNumVirtRegs();
1653   for (unsigned i = 0; i < numVRs; i++) {
1654     unsigned int vr = Register::index2VirtReg(i);
1655     const TargetRegisterClass *RC = MRI->getRegClass(vr);
1656     DenseMap<unsigned, unsigned> &regmap = VRegMapping[RC];
1657     int n = regmap.size();
1658     regmap.insert(std::make_pair(vr, n + 1));
1659   }
1660 
1661   // Emit register declarations
1662   // @TODO: Extract out the real register usage
1663   // O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
1664   // O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
1665   // O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
1666   // O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
1667   // O << "\t.reg .s64 %rd<" << NVPTXNumRegisters << ">;\n";
1668   // O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
1669   // O << "\t.reg .f64 %fd<" << NVPTXNumRegisters << ">;\n";
1670 
1671   // Emit declaration of the virtual registers or 'physical' registers for
1672   // each register class
1673   for (unsigned i=0; i< TRI->getNumRegClasses(); i++) {
1674     const TargetRegisterClass *RC = TRI->getRegClass(i);
1675     DenseMap<unsigned, unsigned> &regmap = VRegMapping[RC];
1676     std::string rcname = getNVPTXRegClassName(RC);
1677     std::string rcStr = getNVPTXRegClassStr(RC);
1678     int n = regmap.size();
1679 
1680     // Only declare those registers that may be used.
1681     if (n) {
1682        O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
1683          << ">;\n";
1684     }
1685   }
1686 
1687   OutStreamer->EmitRawText(O.str());
1688 }
1689 
1690 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
1691   APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
1692   bool ignored;
1693   unsigned int numHex;
1694   const char *lead;
1695 
1696   if (Fp->getType()->getTypeID() == Type::FloatTyID) {
1697     numHex = 8;
1698     lead = "0f";
1699     APF.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven, &ignored);
1700   } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
1701     numHex = 16;
1702     lead = "0d";
1703     APF.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, &ignored);
1704   } else
1705     llvm_unreachable("unsupported fp type");
1706 
1707   APInt API = APF.bitcastToAPInt();
1708   O << lead << format_hex_no_prefix(API.getZExtValue(), numHex, /*Upper=*/true);
1709 }
1710 
1711 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
1712   if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
1713     O << CI->getValue();
1714     return;
1715   }
1716   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
1717     printFPConstant(CFP, O);
1718     return;
1719   }
1720   if (isa<ConstantPointerNull>(CPV)) {
1721     O << "0";
1722     return;
1723   }
1724   if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1725     bool IsNonGenericPointer = false;
1726     if (GVar->getType()->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   const DataLayout &DL = getDataLayout();
1789 
1790   if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
1791     int s = DL.getTypeAllocSize(CPV->getType());
1792     if (s < Bytes)
1793       s = Bytes;
1794     aggBuffer->addZeros(s);
1795     return;
1796   }
1797 
1798   unsigned char ptr[8];
1799   switch (CPV->getType()->getTypeID()) {
1800 
1801   case Type::IntegerTyID: {
1802     Type *ETy = CPV->getType();
1803     if (ETy == Type::getInt8Ty(CPV->getContext())) {
1804       unsigned char c = (unsigned char)cast<ConstantInt>(CPV)->getZExtValue();
1805       ConvertIntToBytes<>(ptr, c);
1806       aggBuffer->addBytes(ptr, 1, Bytes);
1807     } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
1808       short int16 = (short)cast<ConstantInt>(CPV)->getZExtValue();
1809       ConvertIntToBytes<>(ptr, int16);
1810       aggBuffer->addBytes(ptr, 2, Bytes);
1811     } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
1812       if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1813         int int32 = (int)(constInt->getZExtValue());
1814         ConvertIntToBytes<>(ptr, int32);
1815         aggBuffer->addBytes(ptr, 4, Bytes);
1816         break;
1817       } else if (const auto *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1818         if (const auto *constInt = dyn_cast_or_null<ConstantInt>(
1819                 ConstantFoldConstant(Cexpr, DL))) {
1820           int int32 = (int)(constInt->getZExtValue());
1821           ConvertIntToBytes<>(ptr, int32);
1822           aggBuffer->addBytes(ptr, 4, Bytes);
1823           break;
1824         }
1825         if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1826           Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1827           aggBuffer->addSymbol(v, Cexpr->getOperand(0));
1828           aggBuffer->addZeros(4);
1829           break;
1830         }
1831       }
1832       llvm_unreachable("unsupported integer const type");
1833     } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
1834       if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1835         long long int64 = (long long)(constInt->getZExtValue());
1836         ConvertIntToBytes<>(ptr, int64);
1837         aggBuffer->addBytes(ptr, 8, Bytes);
1838         break;
1839       } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1840         if (const auto *constInt = dyn_cast_or_null<ConstantInt>(
1841                 ConstantFoldConstant(Cexpr, DL))) {
1842           long long int64 = (long long)(constInt->getZExtValue());
1843           ConvertIntToBytes<>(ptr, int64);
1844           aggBuffer->addBytes(ptr, 8, Bytes);
1845           break;
1846         }
1847         if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1848           Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1849           aggBuffer->addSymbol(v, Cexpr->getOperand(0));
1850           aggBuffer->addZeros(8);
1851           break;
1852         }
1853       }
1854       llvm_unreachable("unsupported integer const type");
1855     } else
1856       llvm_unreachable("unsupported integer const type");
1857     break;
1858   }
1859   case Type::HalfTyID:
1860   case Type::FloatTyID:
1861   case Type::DoubleTyID: {
1862     const auto *CFP = cast<ConstantFP>(CPV);
1863     Type *Ty = CFP->getType();
1864     if (Ty == Type::getHalfTy(CPV->getContext())) {
1865       APInt API = CFP->getValueAPF().bitcastToAPInt();
1866       uint16_t float16 = API.getLoBits(16).getZExtValue();
1867       ConvertIntToBytes<>(ptr, float16);
1868       aggBuffer->addBytes(ptr, 2, Bytes);
1869     } else if (Ty == Type::getFloatTy(CPV->getContext())) {
1870       float float32 = (float) CFP->getValueAPF().convertToFloat();
1871       ConvertFloatToBytes(ptr, float32);
1872       aggBuffer->addBytes(ptr, 4, Bytes);
1873     } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
1874       double float64 = CFP->getValueAPF().convertToDouble();
1875       ConvertDoubleToBytes(ptr, float64);
1876       aggBuffer->addBytes(ptr, 8, Bytes);
1877     } else {
1878       llvm_unreachable("unsupported fp const type");
1879     }
1880     break;
1881   }
1882   case Type::PointerTyID: {
1883     if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1884       aggBuffer->addSymbol(GVar, GVar);
1885     } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1886       const Value *v = Cexpr->stripPointerCasts();
1887       aggBuffer->addSymbol(v, Cexpr);
1888     }
1889     unsigned int s = DL.getTypeAllocSize(CPV->getType());
1890     aggBuffer->addZeros(s);
1891     break;
1892   }
1893 
1894   case Type::ArrayTyID:
1895   case Type::VectorTyID:
1896   case Type::StructTyID: {
1897     if (isa<ConstantAggregate>(CPV) || isa<ConstantDataSequential>(CPV)) {
1898       int ElementSize = DL.getTypeAllocSize(CPV->getType());
1899       bufferAggregateConstant(CPV, aggBuffer);
1900       if (Bytes > ElementSize)
1901         aggBuffer->addZeros(Bytes - ElementSize);
1902     } else if (isa<ConstantAggregateZero>(CPV))
1903       aggBuffer->addZeros(Bytes);
1904     else
1905       llvm_unreachable("Unexpected Constant type");
1906     break;
1907   }
1908 
1909   default:
1910     llvm_unreachable("unsupported type");
1911   }
1912 }
1913 
1914 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
1915                                               AggBuffer *aggBuffer) {
1916   const DataLayout &DL = getDataLayout();
1917   int Bytes;
1918 
1919   // Integers of arbitrary width
1920   if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
1921     APInt Val = CI->getValue();
1922     for (unsigned I = 0, E = DL.getTypeAllocSize(CPV->getType()); I < E; ++I) {
1923       uint8_t Byte = Val.getLoBits(8).getZExtValue();
1924       aggBuffer->addBytes(&Byte, 1, 1);
1925       Val.lshrInPlace(8);
1926     }
1927     return;
1928   }
1929 
1930   // Old constants
1931   if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
1932     if (CPV->getNumOperands())
1933       for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
1934         bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
1935     return;
1936   }
1937 
1938   if (const ConstantDataSequential *CDS =
1939           dyn_cast<ConstantDataSequential>(CPV)) {
1940     if (CDS->getNumElements())
1941       for (unsigned i = 0; i < CDS->getNumElements(); ++i)
1942         bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
1943                      aggBuffer);
1944     return;
1945   }
1946 
1947   if (isa<ConstantStruct>(CPV)) {
1948     if (CPV->getNumOperands()) {
1949       StructType *ST = cast<StructType>(CPV->getType());
1950       for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
1951         if (i == (e - 1))
1952           Bytes = DL.getStructLayout(ST)->getElementOffset(0) +
1953                   DL.getTypeAllocSize(ST) -
1954                   DL.getStructLayout(ST)->getElementOffset(i);
1955         else
1956           Bytes = DL.getStructLayout(ST)->getElementOffset(i + 1) -
1957                   DL.getStructLayout(ST)->getElementOffset(i);
1958         bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
1959       }
1960     }
1961     return;
1962   }
1963   llvm_unreachable("unsupported constant type in printAggregateConstant()");
1964 }
1965 
1966 /// lowerConstantForGV - Return an MCExpr for the given Constant.  This is mostly
1967 /// a copy from AsmPrinter::lowerConstant, except customized to only handle
1968 /// expressions that are representable in PTX and create
1969 /// NVPTXGenericMCSymbolRefExpr nodes for addrspacecast instructions.
1970 const MCExpr *
1971 NVPTXAsmPrinter::lowerConstantForGV(const Constant *CV, bool ProcessingGeneric) {
1972   MCContext &Ctx = OutContext;
1973 
1974   if (CV->isNullValue() || isa<UndefValue>(CV))
1975     return MCConstantExpr::create(0, Ctx);
1976 
1977   if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV))
1978     return MCConstantExpr::create(CI->getZExtValue(), Ctx);
1979 
1980   if (const GlobalValue *GV = dyn_cast<GlobalValue>(CV)) {
1981     const MCSymbolRefExpr *Expr =
1982       MCSymbolRefExpr::create(getSymbol(GV), Ctx);
1983     if (ProcessingGeneric) {
1984       return NVPTXGenericMCSymbolRefExpr::create(Expr, Ctx);
1985     } else {
1986       return Expr;
1987     }
1988   }
1989 
1990   const ConstantExpr *CE = dyn_cast<ConstantExpr>(CV);
1991   if (!CE) {
1992     llvm_unreachable("Unknown constant value to lower!");
1993   }
1994 
1995   switch (CE->getOpcode()) {
1996   default:
1997     // If the code isn't optimized, there may be outstanding folding
1998     // opportunities. Attempt to fold the expression using DataLayout as a
1999     // last resort before giving up.
2000     if (Constant *C = ConstantFoldConstant(CE, getDataLayout()))
2001       if (C && C != CE)
2002         return lowerConstantForGV(C, ProcessingGeneric);
2003 
2004     // Otherwise report the problem to the user.
2005     {
2006       std::string S;
2007       raw_string_ostream OS(S);
2008       OS << "Unsupported expression in static initializer: ";
2009       CE->printAsOperand(OS, /*PrintType=*/false,
2010                      !MF ? nullptr : MF->getFunction().getParent());
2011       report_fatal_error(OS.str());
2012     }
2013 
2014   case Instruction::AddrSpaceCast: {
2015     // Strip the addrspacecast and pass along the operand
2016     PointerType *DstTy = cast<PointerType>(CE->getType());
2017     if (DstTy->getAddressSpace() == 0) {
2018       return lowerConstantForGV(cast<const Constant>(CE->getOperand(0)), true);
2019     }
2020     std::string S;
2021     raw_string_ostream OS(S);
2022     OS << "Unsupported expression in static initializer: ";
2023     CE->printAsOperand(OS, /*PrintType=*/ false,
2024                        !MF ? nullptr : MF->getFunction().getParent());
2025     report_fatal_error(OS.str());
2026   }
2027 
2028   case Instruction::GetElementPtr: {
2029     const DataLayout &DL = getDataLayout();
2030 
2031     // Generate a symbolic expression for the byte address
2032     APInt OffsetAI(DL.getPointerTypeSizeInBits(CE->getType()), 0);
2033     cast<GEPOperator>(CE)->accumulateConstantOffset(DL, OffsetAI);
2034 
2035     const MCExpr *Base = lowerConstantForGV(CE->getOperand(0),
2036                                             ProcessingGeneric);
2037     if (!OffsetAI)
2038       return Base;
2039 
2040     int64_t Offset = OffsetAI.getSExtValue();
2041     return MCBinaryExpr::createAdd(Base, MCConstantExpr::create(Offset, Ctx),
2042                                    Ctx);
2043   }
2044 
2045   case Instruction::Trunc:
2046     // We emit the value and depend on the assembler to truncate the generated
2047     // expression properly.  This is important for differences between
2048     // blockaddress labels.  Since the two labels are in the same function, it
2049     // is reasonable to treat their delta as a 32-bit value.
2050     LLVM_FALLTHROUGH;
2051   case Instruction::BitCast:
2052     return lowerConstantForGV(CE->getOperand(0), ProcessingGeneric);
2053 
2054   case Instruction::IntToPtr: {
2055     const DataLayout &DL = getDataLayout();
2056 
2057     // Handle casts to pointers by changing them into casts to the appropriate
2058     // integer type.  This promotes constant folding and simplifies this code.
2059     Constant *Op = CE->getOperand(0);
2060     Op = ConstantExpr::getIntegerCast(Op, DL.getIntPtrType(CV->getType()),
2061                                       false/*ZExt*/);
2062     return lowerConstantForGV(Op, ProcessingGeneric);
2063   }
2064 
2065   case Instruction::PtrToInt: {
2066     const DataLayout &DL = getDataLayout();
2067 
2068     // Support only foldable casts to/from pointers that can be eliminated by
2069     // changing the pointer to the appropriately sized integer type.
2070     Constant *Op = CE->getOperand(0);
2071     Type *Ty = CE->getType();
2072 
2073     const MCExpr *OpExpr = lowerConstantForGV(Op, ProcessingGeneric);
2074 
2075     // We can emit the pointer value into this slot if the slot is an
2076     // integer slot equal to the size of the pointer.
2077     if (DL.getTypeAllocSize(Ty) == DL.getTypeAllocSize(Op->getType()))
2078       return OpExpr;
2079 
2080     // Otherwise the pointer is smaller than the resultant integer, mask off
2081     // the high bits so we are sure to get a proper truncation if the input is
2082     // a constant expr.
2083     unsigned InBits = DL.getTypeAllocSizeInBits(Op->getType());
2084     const MCExpr *MaskExpr = MCConstantExpr::create(~0ULL >> (64-InBits), Ctx);
2085     return MCBinaryExpr::createAnd(OpExpr, MaskExpr, Ctx);
2086   }
2087 
2088   // The MC library also has a right-shift operator, but it isn't consistently
2089   // signed or unsigned between different targets.
2090   case Instruction::Add: {
2091     const MCExpr *LHS = lowerConstantForGV(CE->getOperand(0), ProcessingGeneric);
2092     const MCExpr *RHS = lowerConstantForGV(CE->getOperand(1), ProcessingGeneric);
2093     switch (CE->getOpcode()) {
2094     default: llvm_unreachable("Unknown binary operator constant cast expr");
2095     case Instruction::Add: return MCBinaryExpr::createAdd(LHS, RHS, Ctx);
2096     }
2097   }
2098   }
2099 }
2100 
2101 // Copy of MCExpr::print customized for NVPTX
2102 void NVPTXAsmPrinter::printMCExpr(const MCExpr &Expr, raw_ostream &OS) {
2103   switch (Expr.getKind()) {
2104   case MCExpr::Target:
2105     return cast<MCTargetExpr>(&Expr)->printImpl(OS, MAI);
2106   case MCExpr::Constant:
2107     OS << cast<MCConstantExpr>(Expr).getValue();
2108     return;
2109 
2110   case MCExpr::SymbolRef: {
2111     const MCSymbolRefExpr &SRE = cast<MCSymbolRefExpr>(Expr);
2112     const MCSymbol &Sym = SRE.getSymbol();
2113     Sym.print(OS, MAI);
2114     return;
2115   }
2116 
2117   case MCExpr::Unary: {
2118     const MCUnaryExpr &UE = cast<MCUnaryExpr>(Expr);
2119     switch (UE.getOpcode()) {
2120     case MCUnaryExpr::LNot:  OS << '!'; break;
2121     case MCUnaryExpr::Minus: OS << '-'; break;
2122     case MCUnaryExpr::Not:   OS << '~'; break;
2123     case MCUnaryExpr::Plus:  OS << '+'; break;
2124     }
2125     printMCExpr(*UE.getSubExpr(), OS);
2126     return;
2127   }
2128 
2129   case MCExpr::Binary: {
2130     const MCBinaryExpr &BE = cast<MCBinaryExpr>(Expr);
2131 
2132     // Only print parens around the LHS if it is non-trivial.
2133     if (isa<MCConstantExpr>(BE.getLHS()) || isa<MCSymbolRefExpr>(BE.getLHS()) ||
2134         isa<NVPTXGenericMCSymbolRefExpr>(BE.getLHS())) {
2135       printMCExpr(*BE.getLHS(), OS);
2136     } else {
2137       OS << '(';
2138       printMCExpr(*BE.getLHS(), OS);
2139       OS<< ')';
2140     }
2141 
2142     switch (BE.getOpcode()) {
2143     case MCBinaryExpr::Add:
2144       // Print "X-42" instead of "X+-42".
2145       if (const MCConstantExpr *RHSC = dyn_cast<MCConstantExpr>(BE.getRHS())) {
2146         if (RHSC->getValue() < 0) {
2147           OS << RHSC->getValue();
2148           return;
2149         }
2150       }
2151 
2152       OS <<  '+';
2153       break;
2154     default: llvm_unreachable("Unhandled binary operator");
2155     }
2156 
2157     // Only print parens around the LHS if it is non-trivial.
2158     if (isa<MCConstantExpr>(BE.getRHS()) || isa<MCSymbolRefExpr>(BE.getRHS())) {
2159       printMCExpr(*BE.getRHS(), OS);
2160     } else {
2161       OS << '(';
2162       printMCExpr(*BE.getRHS(), OS);
2163       OS << ')';
2164     }
2165     return;
2166   }
2167   }
2168 
2169   llvm_unreachable("Invalid expression kind!");
2170 }
2171 
2172 /// PrintAsmOperand - Print out an operand for an inline asm expression.
2173 ///
2174 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
2175                                       const char *ExtraCode, raw_ostream &O) {
2176   if (ExtraCode && ExtraCode[0]) {
2177     if (ExtraCode[1] != 0)
2178       return true; // Unknown modifier.
2179 
2180     switch (ExtraCode[0]) {
2181     default:
2182       // See if this is a generic print operand
2183       return AsmPrinter::PrintAsmOperand(MI, OpNo, ExtraCode, O);
2184     case 'r':
2185       break;
2186     }
2187   }
2188 
2189   printOperand(MI, OpNo, O);
2190 
2191   return false;
2192 }
2193 
2194 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(const MachineInstr *MI,
2195                                             unsigned OpNo,
2196                                             const char *ExtraCode,
2197                                             raw_ostream &O) {
2198   if (ExtraCode && ExtraCode[0])
2199     return true; // Unknown modifier
2200 
2201   O << '[';
2202   printMemOperand(MI, OpNo, O);
2203   O << ']';
2204 
2205   return false;
2206 }
2207 
2208 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
2209                                    raw_ostream &O) {
2210   const MachineOperand &MO = MI->getOperand(opNum);
2211   switch (MO.getType()) {
2212   case MachineOperand::MO_Register:
2213     if (Register::isPhysicalRegister(MO.getReg())) {
2214       if (MO.getReg() == NVPTX::VRDepot)
2215         O << DEPOTNAME << getFunctionNumber();
2216       else
2217         O << NVPTXInstPrinter::getRegisterName(MO.getReg());
2218     } else {
2219       emitVirtualRegister(MO.getReg(), O);
2220     }
2221     break;
2222 
2223   case MachineOperand::MO_Immediate:
2224     O << MO.getImm();
2225     break;
2226 
2227   case MachineOperand::MO_FPImmediate:
2228     printFPConstant(MO.getFPImm(), O);
2229     break;
2230 
2231   case MachineOperand::MO_GlobalAddress:
2232     PrintSymbolOperand(MO, O);
2233     break;
2234 
2235   case MachineOperand::MO_MachineBasicBlock:
2236     MO.getMBB()->getSymbol()->print(O, MAI);
2237     break;
2238 
2239   default:
2240     llvm_unreachable("Operand type not supported.");
2241   }
2242 }
2243 
2244 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
2245                                       raw_ostream &O, const char *Modifier) {
2246   printOperand(MI, opNum, O);
2247 
2248   if (Modifier && strcmp(Modifier, "add") == 0) {
2249     O << ", ";
2250     printOperand(MI, opNum + 1, O);
2251   } else {
2252     if (MI->getOperand(opNum + 1).isImm() &&
2253         MI->getOperand(opNum + 1).getImm() == 0)
2254       return; // don't print ',0' or '+0'
2255     O << "+";
2256     printOperand(MI, opNum + 1, O);
2257   }
2258 }
2259 
2260 // Force static initialization.
2261 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeNVPTXAsmPrinter() {
2262   RegisterAsmPrinter<NVPTXAsmPrinter> X(getTheNVPTXTarget32());
2263   RegisterAsmPrinter<NVPTXAsmPrinter> Y(getTheNVPTXTarget64());
2264 }
2265