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 "NVPTXRegisterInfo.h"
22 #include "NVPTXTargetMachine.h"
23 #include "NVPTXUtilities.h"
24 #include "cl_common_defines.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/CodeGen/Analysis.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineModuleInfo.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/DebugInfo.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/GlobalVariable.h"
35 #include "llvm/IR/Mangler.h"
36 #include "llvm/IR/Module.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/MC/MCStreamer.h"
39 #include "llvm/MC/MCSymbol.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/FormattedStream.h"
43 #include "llvm/Support/Path.h"
44 #include "llvm/Support/TargetRegistry.h"
45 #include "llvm/Support/TimeValue.h"
46 #include "llvm/Target/TargetLoweringObjectFile.h"
47 #include <sstream>
48 using namespace llvm;
49 
50 #define DEPOTNAME "__local_depot"
51 
52 static cl::opt<bool>
53 EmitLineNumbers("nvptx-emit-line-numbers", cl::Hidden,
54                 cl::desc("NVPTX Specific: Emit Line numbers even without -G"),
55                 cl::init(true));
56 
57 static cl::opt<bool>
58 InterleaveSrc("nvptx-emit-src", cl::ZeroOrMore, cl::Hidden,
59               cl::desc("NVPTX Specific: Emit source line in ptx file"),
60               cl::init(false));
61 
62 namespace {
63 /// DiscoverDependentGlobals - Return a set of GlobalVariables on which \p V
64 /// depends.
65 void DiscoverDependentGlobals(const Value *V,
66                               DenseSet<const GlobalVariable *> &Globals) {
67   if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
68     Globals.insert(GV);
69   else {
70     if (const User *U = dyn_cast<User>(V)) {
71       for (unsigned i = 0, e = U->getNumOperands(); i != e; ++i) {
72         DiscoverDependentGlobals(U->getOperand(i), Globals);
73       }
74     }
75   }
76 }
77 
78 /// VisitGlobalVariableForEmission - Add \p GV to the list of GlobalVariable
79 /// instances to be emitted, but only after any dependents have been added
80 /// first.
81 void VisitGlobalVariableForEmission(
82     const GlobalVariable *GV, SmallVectorImpl<const GlobalVariable *> &Order,
83     DenseSet<const GlobalVariable *> &Visited,
84     DenseSet<const GlobalVariable *> &Visiting) {
85   // Have we already visited this one?
86   if (Visited.count(GV))
87     return;
88 
89   // Do we have a circular dependency?
90   if (Visiting.count(GV))
91     report_fatal_error("Circular dependency found in global variable set");
92 
93   // Start visiting this global
94   Visiting.insert(GV);
95 
96   // Make sure we visit all dependents first
97   DenseSet<const GlobalVariable *> Others;
98   for (unsigned i = 0, e = GV->getNumOperands(); i != e; ++i)
99     DiscoverDependentGlobals(GV->getOperand(i), Others);
100 
101   for (DenseSet<const GlobalVariable *>::iterator I = Others.begin(),
102                                                   E = Others.end();
103        I != E; ++I)
104     VisitGlobalVariableForEmission(*I, Order, Visited, Visiting);
105 
106   // Now we can visit ourself
107   Order.push_back(GV);
108   Visited.insert(GV);
109   Visiting.erase(GV);
110 }
111 }
112 
113 // @TODO: This is a copy from AsmPrinter.cpp.  The function is static, so we
114 // cannot just link to the existing version.
115 /// LowerConstant - Lower the specified LLVM Constant to an MCExpr.
116 ///
117 using namespace nvptx;
118 const MCExpr *nvptx::LowerConstant(const Constant *CV, AsmPrinter &AP) {
119   MCContext &Ctx = AP.OutContext;
120 
121   if (CV->isNullValue() || isa<UndefValue>(CV))
122     return MCConstantExpr::Create(0, Ctx);
123 
124   if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV))
125     return MCConstantExpr::Create(CI->getZExtValue(), Ctx);
126 
127   if (const GlobalValue *GV = dyn_cast<GlobalValue>(CV))
128     return MCSymbolRefExpr::Create(AP.getSymbol(GV), Ctx);
129 
130   if (const BlockAddress *BA = dyn_cast<BlockAddress>(CV))
131     return MCSymbolRefExpr::Create(AP.GetBlockAddressSymbol(BA), Ctx);
132 
133   const ConstantExpr *CE = dyn_cast<ConstantExpr>(CV);
134   if (CE == 0)
135     llvm_unreachable("Unknown constant value to lower!");
136 
137   switch (CE->getOpcode()) {
138   default:
139     // If the code isn't optimized, there may be outstanding folding
140     // opportunities. Attempt to fold the expression using DataLayout as a
141     // last resort before giving up.
142     if (Constant *C = ConstantFoldConstantExpression(CE, AP.TM.getDataLayout()))
143       if (C != CE)
144         return LowerConstant(C, AP);
145 
146     // Otherwise report the problem to the user.
147     {
148       std::string S;
149       raw_string_ostream OS(S);
150       OS << "Unsupported expression in static initializer: ";
151       CE->printAsOperand(OS, /*PrintType=*/ false,
152                      !AP.MF ? 0 : AP.MF->getFunction()->getParent());
153       report_fatal_error(OS.str());
154     }
155   case Instruction::GetElementPtr: {
156     const DataLayout &TD = *AP.TM.getDataLayout();
157     // Generate a symbolic expression for the byte address
158     APInt OffsetAI(TD.getPointerSizeInBits(), 0);
159     cast<GEPOperator>(CE)->accumulateConstantOffset(TD, OffsetAI);
160 
161     const MCExpr *Base = LowerConstant(CE->getOperand(0), AP);
162     if (!OffsetAI)
163       return Base;
164 
165     int64_t Offset = OffsetAI.getSExtValue();
166     return MCBinaryExpr::CreateAdd(Base, MCConstantExpr::Create(Offset, Ctx),
167                                    Ctx);
168   }
169 
170   case Instruction::Trunc:
171     // We emit the value and depend on the assembler to truncate the generated
172     // expression properly.  This is important for differences between
173     // blockaddress labels.  Since the two labels are in the same function, it
174     // is reasonable to treat their delta as a 32-bit value.
175   // FALL THROUGH.
176   case Instruction::BitCast:
177     return LowerConstant(CE->getOperand(0), AP);
178 
179   case Instruction::IntToPtr: {
180     const DataLayout &TD = *AP.TM.getDataLayout();
181     // Handle casts to pointers by changing them into casts to the appropriate
182     // integer type.  This promotes constant folding and simplifies this code.
183     Constant *Op = CE->getOperand(0);
184     Op = ConstantExpr::getIntegerCast(Op, TD.getIntPtrType(CV->getContext()),
185                                       false /*ZExt*/);
186     return LowerConstant(Op, AP);
187   }
188 
189   case Instruction::PtrToInt: {
190     const DataLayout &TD = *AP.TM.getDataLayout();
191     // Support only foldable casts to/from pointers that can be eliminated by
192     // changing the pointer to the appropriately sized integer type.
193     Constant *Op = CE->getOperand(0);
194     Type *Ty = CE->getType();
195 
196     const MCExpr *OpExpr = LowerConstant(Op, AP);
197 
198     // We can emit the pointer value into this slot if the slot is an
199     // integer slot equal to the size of the pointer.
200     if (TD.getTypeAllocSize(Ty) == TD.getTypeAllocSize(Op->getType()))
201       return OpExpr;
202 
203     // Otherwise the pointer is smaller than the resultant integer, mask off
204     // the high bits so we are sure to get a proper truncation if the input is
205     // a constant expr.
206     unsigned InBits = TD.getTypeAllocSizeInBits(Op->getType());
207     const MCExpr *MaskExpr =
208         MCConstantExpr::Create(~0ULL >> (64 - InBits), Ctx);
209     return MCBinaryExpr::CreateAnd(OpExpr, MaskExpr, Ctx);
210   }
211 
212     // The MC library also has a right-shift operator, but it isn't consistently
213   // signed or unsigned between different targets.
214   case Instruction::Add:
215   case Instruction::Sub:
216   case Instruction::Mul:
217   case Instruction::SDiv:
218   case Instruction::SRem:
219   case Instruction::Shl:
220   case Instruction::And:
221   case Instruction::Or:
222   case Instruction::Xor: {
223     const MCExpr *LHS = LowerConstant(CE->getOperand(0), AP);
224     const MCExpr *RHS = LowerConstant(CE->getOperand(1), AP);
225     switch (CE->getOpcode()) {
226     default:
227       llvm_unreachable("Unknown binary operator constant cast expr");
228     case Instruction::Add:
229       return MCBinaryExpr::CreateAdd(LHS, RHS, Ctx);
230     case Instruction::Sub:
231       return MCBinaryExpr::CreateSub(LHS, RHS, Ctx);
232     case Instruction::Mul:
233       return MCBinaryExpr::CreateMul(LHS, RHS, Ctx);
234     case Instruction::SDiv:
235       return MCBinaryExpr::CreateDiv(LHS, RHS, Ctx);
236     case Instruction::SRem:
237       return MCBinaryExpr::CreateMod(LHS, RHS, Ctx);
238     case Instruction::Shl:
239       return MCBinaryExpr::CreateShl(LHS, RHS, Ctx);
240     case Instruction::And:
241       return MCBinaryExpr::CreateAnd(LHS, RHS, Ctx);
242     case Instruction::Or:
243       return MCBinaryExpr::CreateOr(LHS, RHS, Ctx);
244     case Instruction::Xor:
245       return MCBinaryExpr::CreateXor(LHS, RHS, Ctx);
246     }
247   }
248   }
249 }
250 
251 void NVPTXAsmPrinter::emitLineNumberAsDotLoc(const MachineInstr &MI) {
252   if (!EmitLineNumbers)
253     return;
254   if (ignoreLoc(MI))
255     return;
256 
257   DebugLoc curLoc = MI.getDebugLoc();
258 
259   if (prevDebugLoc.isUnknown() && curLoc.isUnknown())
260     return;
261 
262   if (prevDebugLoc == curLoc)
263     return;
264 
265   prevDebugLoc = curLoc;
266 
267   if (curLoc.isUnknown())
268     return;
269 
270   const MachineFunction *MF = MI.getParent()->getParent();
271   //const TargetMachine &TM = MF->getTarget();
272 
273   const LLVMContext &ctx = MF->getFunction()->getContext();
274   DIScope Scope(curLoc.getScope(ctx));
275 
276   assert((!Scope || Scope.isScope()) &&
277     "Scope of a DebugLoc should be null or a DIScope.");
278   if (!Scope)
279      return;
280 
281   StringRef fileName(Scope.getFilename());
282   StringRef dirName(Scope.getDirectory());
283   SmallString<128> FullPathName = dirName;
284   if (!dirName.empty() && !sys::path::is_absolute(fileName)) {
285     sys::path::append(FullPathName, fileName);
286     fileName = FullPathName.str();
287   }
288 
289   if (filenameMap.find(fileName.str()) == filenameMap.end())
290     return;
291 
292   // Emit the line from the source file.
293   if (InterleaveSrc)
294     this->emitSrcInText(fileName.str(), curLoc.getLine());
295 
296   std::stringstream temp;
297   temp << "\t.loc " << filenameMap[fileName.str()] << " " << curLoc.getLine()
298        << " " << curLoc.getCol();
299   OutStreamer.EmitRawText(Twine(temp.str().c_str()));
300 }
301 
302 void NVPTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
303   SmallString<128> Str;
304   raw_svector_ostream OS(Str);
305   if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
306     emitLineNumberAsDotLoc(*MI);
307 
308   MCInst Inst;
309   lowerToMCInst(MI, Inst);
310   EmitToStreamer(OutStreamer, Inst);
311 }
312 
313 void NVPTXAsmPrinter::lowerToMCInst(const MachineInstr *MI, MCInst &OutMI) {
314   OutMI.setOpcode(MI->getOpcode());
315 
316   // Special: Do not mangle symbol operand of CALL_PROTOTYPE
317   if (MI->getOpcode() == NVPTX::CALL_PROTOTYPE) {
318     const MachineOperand &MO = MI->getOperand(0);
319     OutMI.addOperand(GetSymbolRef(MO,
320       OutContext.GetOrCreateSymbol(Twine(MO.getSymbolName()))));
321     return;
322   }
323 
324   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
325     const MachineOperand &MO = MI->getOperand(i);
326 
327     MCOperand MCOp;
328     if (lowerOperand(MO, MCOp))
329       OutMI.addOperand(MCOp);
330   }
331 }
332 
333 bool NVPTXAsmPrinter::lowerOperand(const MachineOperand &MO,
334                                    MCOperand &MCOp) {
335   switch (MO.getType()) {
336   default: llvm_unreachable("unknown operand type");
337   case MachineOperand::MO_Register:
338     MCOp = MCOperand::CreateReg(encodeVirtualRegister(MO.getReg()));
339     break;
340   case MachineOperand::MO_Immediate:
341     MCOp = MCOperand::CreateImm(MO.getImm());
342     break;
343   case MachineOperand::MO_MachineBasicBlock:
344     MCOp = MCOperand::CreateExpr(MCSymbolRefExpr::Create(
345         MO.getMBB()->getSymbol(), OutContext));
346     break;
347   case MachineOperand::MO_ExternalSymbol:
348     MCOp = GetSymbolRef(MO, GetExternalSymbolSymbol(MO.getSymbolName()));
349     break;
350   case MachineOperand::MO_GlobalAddress:
351     MCOp = GetSymbolRef(MO, getSymbol(MO.getGlobal()));
352     break;
353   case MachineOperand::MO_FPImmediate: {
354     const ConstantFP *Cnt = MO.getFPImm();
355     APFloat Val = Cnt->getValueAPF();
356 
357     switch (Cnt->getType()->getTypeID()) {
358     default: report_fatal_error("Unsupported FP type"); break;
359     case Type::FloatTyID:
360       MCOp = MCOperand::CreateExpr(
361         NVPTXFloatMCExpr::CreateConstantFPSingle(Val, OutContext));
362       break;
363     case Type::DoubleTyID:
364       MCOp = MCOperand::CreateExpr(
365         NVPTXFloatMCExpr::CreateConstantFPDouble(Val, OutContext));
366       break;
367     }
368     break;
369   }
370   }
371   return true;
372 }
373 
374 unsigned NVPTXAsmPrinter::encodeVirtualRegister(unsigned Reg) {
375   if (TargetRegisterInfo::isVirtualRegister(Reg)) {
376     const TargetRegisterClass *RC = MRI->getRegClass(Reg);
377 
378     DenseMap<unsigned, unsigned> &RegMap = VRegMapping[RC];
379     unsigned RegNum = RegMap[Reg];
380 
381     // Encode the register class in the upper 4 bits
382     // Must be kept in sync with NVPTXInstPrinter::printRegName
383     unsigned Ret = 0;
384     if (RC == &NVPTX::Int1RegsRegClass) {
385       Ret = (1 << 28);
386     } else if (RC == &NVPTX::Int16RegsRegClass) {
387       Ret = (2 << 28);
388     } else if (RC == &NVPTX::Int32RegsRegClass) {
389       Ret = (3 << 28);
390     } else if (RC == &NVPTX::Int64RegsRegClass) {
391       Ret = (4 << 28);
392     } else if (RC == &NVPTX::Float32RegsRegClass) {
393       Ret = (5 << 28);
394     } else if (RC == &NVPTX::Float64RegsRegClass) {
395       Ret = (6 << 28);
396     } else {
397       report_fatal_error("Bad register class");
398     }
399 
400     // Insert the vreg number
401     Ret |= (RegNum & 0x0FFFFFFF);
402     return Ret;
403   } else {
404     // Some special-use registers are actually physical registers.
405     // Encode this as the register class ID of 0 and the real register ID.
406     return Reg & 0x0FFFFFFF;
407   }
408 }
409 
410 MCOperand NVPTXAsmPrinter::GetSymbolRef(const MachineOperand &MO,
411                                         const MCSymbol *Symbol) {
412   const MCExpr *Expr;
413   Expr = MCSymbolRefExpr::Create(Symbol, MCSymbolRefExpr::VK_None,
414                                  OutContext);
415   return MCOperand::CreateExpr(Expr);
416 }
417 
418 void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
419   const DataLayout *TD = TM.getDataLayout();
420   const TargetLowering *TLI = TM.getTargetLowering();
421 
422   Type *Ty = F->getReturnType();
423 
424   bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
425 
426   if (Ty->getTypeID() == Type::VoidTyID)
427     return;
428 
429   O << " (";
430 
431   if (isABI) {
432     if (Ty->isFloatingPointTy() || Ty->isIntegerTy()) {
433       unsigned size = 0;
434       if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty)) {
435         size = ITy->getBitWidth();
436         if (size < 32)
437           size = 32;
438       } else {
439         assert(Ty->isFloatingPointTy() && "Floating point type expected here");
440         size = Ty->getPrimitiveSizeInBits();
441       }
442 
443       O << ".param .b" << size << " func_retval0";
444     } else if (isa<PointerType>(Ty)) {
445       O << ".param .b" << TLI->getPointerTy().getSizeInBits()
446         << " func_retval0";
447     } else {
448       if ((Ty->getTypeID() == Type::StructTyID) || isa<VectorType>(Ty)) {
449         SmallVector<EVT, 16> vtparts;
450         ComputeValueVTs(*TLI, Ty, vtparts);
451         unsigned totalsz = 0;
452         for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
453           unsigned elems = 1;
454           EVT elemtype = vtparts[i];
455           if (vtparts[i].isVector()) {
456             elems = vtparts[i].getVectorNumElements();
457             elemtype = vtparts[i].getVectorElementType();
458           }
459           for (unsigned j = 0, je = elems; j != je; ++j) {
460             unsigned sz = elemtype.getSizeInBits();
461             if (elemtype.isInteger() && (sz < 8))
462               sz = 8;
463             totalsz += sz / 8;
464           }
465         }
466         unsigned retAlignment = 0;
467         if (!llvm::getAlign(*F, 0, retAlignment))
468           retAlignment = TD->getABITypeAlignment(Ty);
469         O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
470           << "]";
471       } else
472         assert(false && "Unknown return type");
473     }
474   } else {
475     SmallVector<EVT, 16> vtparts;
476     ComputeValueVTs(*TLI, Ty, vtparts);
477     unsigned idx = 0;
478     for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
479       unsigned elems = 1;
480       EVT elemtype = vtparts[i];
481       if (vtparts[i].isVector()) {
482         elems = vtparts[i].getVectorNumElements();
483         elemtype = vtparts[i].getVectorElementType();
484       }
485 
486       for (unsigned j = 0, je = elems; j != je; ++j) {
487         unsigned sz = elemtype.getSizeInBits();
488         if (elemtype.isInteger() && (sz < 32))
489           sz = 32;
490         O << ".reg .b" << sz << " func_retval" << idx;
491         if (j < je - 1)
492           O << ", ";
493         ++idx;
494       }
495       if (i < e - 1)
496         O << ", ";
497     }
498   }
499   O << ") ";
500   return;
501 }
502 
503 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
504                                         raw_ostream &O) {
505   const Function *F = MF.getFunction();
506   printReturnValStr(F, O);
507 }
508 
509 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
510   SmallString<128> Str;
511   raw_svector_ostream O(Str);
512 
513   if (!GlobalsEmitted) {
514     emitGlobals(*MF->getFunction()->getParent());
515     GlobalsEmitted = true;
516   }
517 
518   // Set up
519   MRI = &MF->getRegInfo();
520   F = MF->getFunction();
521   emitLinkageDirective(F, O);
522   if (llvm::isKernelFunction(*F))
523     O << ".entry ";
524   else {
525     O << ".func ";
526     printReturnValStr(*MF, O);
527   }
528 
529   O << *CurrentFnSym;
530 
531   emitFunctionParamList(*MF, O);
532 
533   if (llvm::isKernelFunction(*F))
534     emitKernelFunctionDirectives(*F, O);
535 
536   OutStreamer.EmitRawText(O.str());
537 
538   prevDebugLoc = DebugLoc();
539 }
540 
541 void NVPTXAsmPrinter::EmitFunctionBodyStart() {
542   VRegMapping.clear();
543   OutStreamer.EmitRawText(StringRef("{\n"));
544   setAndEmitFunctionVirtualRegisters(*MF);
545 
546   SmallString<128> Str;
547   raw_svector_ostream O(Str);
548   emitDemotedVars(MF->getFunction(), O);
549   OutStreamer.EmitRawText(O.str());
550 }
551 
552 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
553   OutStreamer.EmitRawText(StringRef("}\n"));
554   VRegMapping.clear();
555 }
556 
557 void NVPTXAsmPrinter::emitImplicitDef(const MachineInstr *MI) const {
558   unsigned RegNo = MI->getOperand(0).getReg();
559   const TargetRegisterInfo *TRI = TM.getRegisterInfo();
560   if (TRI->isVirtualRegister(RegNo)) {
561     OutStreamer.AddComment(Twine("implicit-def: ") +
562                            getVirtualRegisterName(RegNo));
563   } else {
564     OutStreamer.AddComment(Twine("implicit-def: ") +
565                            TM.getRegisterInfo()->getName(RegNo));
566   }
567   OutStreamer.AddBlankLine();
568 }
569 
570 void NVPTXAsmPrinter::emitKernelFunctionDirectives(const Function &F,
571                                                    raw_ostream &O) const {
572   // If the NVVM IR has some of reqntid* specified, then output
573   // the reqntid directive, and set the unspecified ones to 1.
574   // If none of reqntid* is specified, don't output reqntid directive.
575   unsigned reqntidx, reqntidy, reqntidz;
576   bool specified = false;
577   if (llvm::getReqNTIDx(F, reqntidx) == false)
578     reqntidx = 1;
579   else
580     specified = true;
581   if (llvm::getReqNTIDy(F, reqntidy) == false)
582     reqntidy = 1;
583   else
584     specified = true;
585   if (llvm::getReqNTIDz(F, reqntidz) == false)
586     reqntidz = 1;
587   else
588     specified = true;
589 
590   if (specified)
591     O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
592       << "\n";
593 
594   // If the NVVM IR has some of maxntid* specified, then output
595   // the maxntid directive, and set the unspecified ones to 1.
596   // If none of maxntid* is specified, don't output maxntid directive.
597   unsigned maxntidx, maxntidy, maxntidz;
598   specified = false;
599   if (llvm::getMaxNTIDx(F, maxntidx) == false)
600     maxntidx = 1;
601   else
602     specified = true;
603   if (llvm::getMaxNTIDy(F, maxntidy) == false)
604     maxntidy = 1;
605   else
606     specified = true;
607   if (llvm::getMaxNTIDz(F, maxntidz) == false)
608     maxntidz = 1;
609   else
610     specified = true;
611 
612   if (specified)
613     O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
614       << "\n";
615 
616   unsigned mincta;
617   if (llvm::getMinCTASm(F, mincta))
618     O << ".minnctapersm " << mincta << "\n";
619 }
620 
621 std::string
622 NVPTXAsmPrinter::getVirtualRegisterName(unsigned Reg) const {
623   const TargetRegisterClass *RC = MRI->getRegClass(Reg);
624 
625   std::string Name;
626   raw_string_ostream NameStr(Name);
627 
628   VRegRCMap::const_iterator I = VRegMapping.find(RC);
629   assert(I != VRegMapping.end() && "Bad register class");
630   const DenseMap<unsigned, unsigned> &RegMap = I->second;
631 
632   VRegMap::const_iterator VI = RegMap.find(Reg);
633   assert(VI != RegMap.end() && "Bad virtual register");
634   unsigned MappedVR = VI->second;
635 
636   NameStr << getNVPTXRegClassStr(RC) << MappedVR;
637 
638   NameStr.flush();
639   return Name;
640 }
641 
642 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr,
643                                           raw_ostream &O) {
644   O << getVirtualRegisterName(vr);
645 }
646 
647 void NVPTXAsmPrinter::printVecModifiedImmediate(
648     const MachineOperand &MO, const char *Modifier, raw_ostream &O) {
649   static const char vecelem[] = { '0', '1', '2', '3', '0', '1', '2', '3' };
650   int Imm = (int) MO.getImm();
651   if (0 == strcmp(Modifier, "vecelem"))
652     O << "_" << vecelem[Imm];
653   else if (0 == strcmp(Modifier, "vecv4comm1")) {
654     if ((Imm < 0) || (Imm > 3))
655       O << "//";
656   } else if (0 == strcmp(Modifier, "vecv4comm2")) {
657     if ((Imm < 4) || (Imm > 7))
658       O << "//";
659   } else if (0 == strcmp(Modifier, "vecv4pos")) {
660     if (Imm < 0)
661       Imm = 0;
662     O << "_" << vecelem[Imm % 4];
663   } else if (0 == strcmp(Modifier, "vecv2comm1")) {
664     if ((Imm < 0) || (Imm > 1))
665       O << "//";
666   } else if (0 == strcmp(Modifier, "vecv2comm2")) {
667     if ((Imm < 2) || (Imm > 3))
668       O << "//";
669   } else if (0 == strcmp(Modifier, "vecv2pos")) {
670     if (Imm < 0)
671       Imm = 0;
672     O << "_" << vecelem[Imm % 2];
673   } else
674     llvm_unreachable("Unknown Modifier on immediate operand");
675 }
676 
677 
678 
679 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
680 
681   emitLinkageDirective(F, O);
682   if (llvm::isKernelFunction(*F))
683     O << ".entry ";
684   else
685     O << ".func ";
686   printReturnValStr(F, O);
687   O << *getSymbol(F) << "\n";
688   emitFunctionParamList(F, O);
689   O << ";\n";
690 }
691 
692 static bool usedInGlobalVarDef(const Constant *C) {
693   if (!C)
694     return false;
695 
696   if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
697     if (GV->getName().str() == "llvm.used")
698       return false;
699     return true;
700   }
701 
702   for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
703        ui != ue; ++ui) {
704     const Constant *C = dyn_cast<Constant>(*ui);
705     if (usedInGlobalVarDef(C))
706       return true;
707   }
708   return false;
709 }
710 
711 static bool usedInOneFunc(const User *U, Function const *&oneFunc) {
712   if (const GlobalVariable *othergv = dyn_cast<GlobalVariable>(U)) {
713     if (othergv->getName().str() == "llvm.used")
714       return true;
715   }
716 
717   if (const Instruction *instr = dyn_cast<Instruction>(U)) {
718     if (instr->getParent() && instr->getParent()->getParent()) {
719       const Function *curFunc = instr->getParent()->getParent();
720       if (oneFunc && (curFunc != oneFunc))
721         return false;
722       oneFunc = curFunc;
723       return true;
724     } else
725       return false;
726   }
727 
728   if (const MDNode *md = dyn_cast<MDNode>(U))
729     if (md->hasName() && ((md->getName().str() == "llvm.dbg.gv") ||
730                           (md->getName().str() == "llvm.dbg.sp")))
731       return true;
732 
733   for (User::const_use_iterator ui = U->use_begin(), ue = U->use_end();
734        ui != ue; ++ui) {
735     if (usedInOneFunc(*ui, oneFunc) == false)
736       return false;
737   }
738   return true;
739 }
740 
741 /* Find out if a global variable can be demoted to local scope.
742  * Currently, this is valid for CUDA shared variables, which have local
743  * scope and global lifetime. So the conditions to check are :
744  * 1. Is the global variable in shared address space?
745  * 2. Does it have internal linkage?
746  * 3. Is the global variable referenced only in one function?
747  */
748 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
749   if (gv->hasInternalLinkage() == false)
750     return false;
751   const PointerType *Pty = gv->getType();
752   if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
753     return false;
754 
755   const Function *oneFunc = 0;
756 
757   bool flag = usedInOneFunc(gv, oneFunc);
758   if (flag == false)
759     return false;
760   if (!oneFunc)
761     return false;
762   f = oneFunc;
763   return true;
764 }
765 
766 static bool useFuncSeen(const Constant *C,
767                         llvm::DenseMap<const Function *, bool> &seenMap) {
768   for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
769        ui != ue; ++ui) {
770     if (const Constant *cu = dyn_cast<Constant>(*ui)) {
771       if (useFuncSeen(cu, seenMap))
772         return true;
773     } else if (const Instruction *I = dyn_cast<Instruction>(*ui)) {
774       const BasicBlock *bb = I->getParent();
775       if (!bb)
776         continue;
777       const Function *caller = bb->getParent();
778       if (!caller)
779         continue;
780       if (seenMap.find(caller) != seenMap.end())
781         return true;
782     }
783   }
784   return false;
785 }
786 
787 void NVPTXAsmPrinter::emitDeclarations(const Module &M, raw_ostream &O) {
788   llvm::DenseMap<const Function *, bool> seenMap;
789   for (Module::const_iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
790     const Function *F = FI;
791 
792     if (F->isDeclaration()) {
793       if (F->use_empty())
794         continue;
795       if (F->getIntrinsicID())
796         continue;
797       emitDeclaration(F, O);
798       continue;
799     }
800     for (Value::const_use_iterator iter = F->use_begin(),
801                                    iterEnd = F->use_end();
802          iter != iterEnd; ++iter) {
803       if (const Constant *C = dyn_cast<Constant>(*iter)) {
804         if (usedInGlobalVarDef(C)) {
805           // The use is in the initialization of a global variable
806           // that is a function pointer, so print a declaration
807           // for the original function
808           emitDeclaration(F, O);
809           break;
810         }
811         // Emit a declaration of this function if the function that
812         // uses this constant expr has already been seen.
813         if (useFuncSeen(C, seenMap)) {
814           emitDeclaration(F, O);
815           break;
816         }
817       }
818 
819       if (!isa<Instruction>(*iter))
820         continue;
821       const Instruction *instr = cast<Instruction>(*iter);
822       const BasicBlock *bb = instr->getParent();
823       if (!bb)
824         continue;
825       const Function *caller = bb->getParent();
826       if (!caller)
827         continue;
828 
829       // If a caller has already been seen, then the caller is
830       // appearing in the module before the callee. so print out
831       // a declaration for the callee.
832       if (seenMap.find(caller) != seenMap.end()) {
833         emitDeclaration(F, O);
834         break;
835       }
836     }
837     seenMap[F] = true;
838   }
839 }
840 
841 void NVPTXAsmPrinter::recordAndEmitFilenames(Module &M) {
842   DebugInfoFinder DbgFinder;
843   DbgFinder.processModule(M);
844 
845   unsigned i = 1;
846   for (DebugInfoFinder::iterator I = DbgFinder.compile_unit_begin(),
847                                  E = DbgFinder.compile_unit_end();
848        I != E; ++I) {
849     DICompileUnit DIUnit(*I);
850     StringRef Filename(DIUnit.getFilename());
851     StringRef Dirname(DIUnit.getDirectory());
852     SmallString<128> FullPathName = Dirname;
853     if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
854       sys::path::append(FullPathName, Filename);
855       Filename = FullPathName.str();
856     }
857     if (filenameMap.find(Filename.str()) != filenameMap.end())
858       continue;
859     filenameMap[Filename.str()] = i;
860     OutStreamer.EmitDwarfFileDirective(i, "", Filename.str());
861     ++i;
862   }
863 
864   for (DebugInfoFinder::iterator I = DbgFinder.subprogram_begin(),
865                                  E = DbgFinder.subprogram_end();
866        I != E; ++I) {
867     DISubprogram SP(*I);
868     StringRef Filename(SP.getFilename());
869     StringRef Dirname(SP.getDirectory());
870     SmallString<128> FullPathName = Dirname;
871     if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
872       sys::path::append(FullPathName, Filename);
873       Filename = FullPathName.str();
874     }
875     if (filenameMap.find(Filename.str()) != filenameMap.end())
876       continue;
877     filenameMap[Filename.str()] = i;
878     ++i;
879   }
880 }
881 
882 bool NVPTXAsmPrinter::doInitialization(Module &M) {
883 
884   SmallString<128> Str1;
885   raw_svector_ostream OS1(Str1);
886 
887   MMI = getAnalysisIfAvailable<MachineModuleInfo>();
888   MMI->AnalyzeModule(M);
889 
890   // We need to call the parent's one explicitly.
891   //bool Result = AsmPrinter::doInitialization(M);
892 
893   // Initialize TargetLoweringObjectFile.
894   const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
895       .Initialize(OutContext, TM);
896 
897   Mang = new Mangler(TM.getDataLayout());
898 
899   // Emit header before any dwarf directives are emitted below.
900   emitHeader(M, OS1);
901   OutStreamer.EmitRawText(OS1.str());
902 
903   // Already commented out
904   //bool Result = AsmPrinter::doInitialization(M);
905 
906   // Emit module-level inline asm if it exists.
907   if (!M.getModuleInlineAsm().empty()) {
908     OutStreamer.AddComment("Start of file scope inline assembly");
909     OutStreamer.AddBlankLine();
910     OutStreamer.EmitRawText(StringRef(M.getModuleInlineAsm()));
911     OutStreamer.AddBlankLine();
912     OutStreamer.AddComment("End of file scope inline assembly");
913     OutStreamer.AddBlankLine();
914   }
915 
916   if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
917     recordAndEmitFilenames(M);
918 
919   GlobalsEmitted = false;
920 
921   return false; // success
922 }
923 
924 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
925   SmallString<128> Str2;
926   raw_svector_ostream OS2(Str2);
927 
928   emitDeclarations(M, OS2);
929 
930   // As ptxas does not support forward references of globals, we need to first
931   // sort the list of module-level globals in def-use order. We visit each
932   // global variable in order, and ensure that we emit it *after* its dependent
933   // globals. We use a little extra memory maintaining both a set and a list to
934   // have fast searches while maintaining a strict ordering.
935   SmallVector<const GlobalVariable *, 8> Globals;
936   DenseSet<const GlobalVariable *> GVVisited;
937   DenseSet<const GlobalVariable *> GVVisiting;
938 
939   // Visit each global variable, in order
940   for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
941        I != E; ++I)
942     VisitGlobalVariableForEmission(I, Globals, GVVisited, GVVisiting);
943 
944   assert(GVVisited.size() == M.getGlobalList().size() &&
945          "Missed a global variable");
946   assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
947 
948   // Print out module-level global variables in proper order
949   for (unsigned i = 0, e = Globals.size(); i != e; ++i)
950     printModuleLevelGV(Globals[i], OS2);
951 
952   OS2 << '\n';
953 
954   OutStreamer.EmitRawText(OS2.str());
955 }
956 
957 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O) {
958   O << "//\n";
959   O << "// Generated by LLVM NVPTX Back-End\n";
960   O << "//\n";
961   O << "\n";
962 
963   unsigned PTXVersion = nvptxSubtarget.getPTXVersion();
964   O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
965 
966   O << ".target ";
967   O << nvptxSubtarget.getTargetName();
968 
969   if (nvptxSubtarget.getDrvInterface() == NVPTX::NVCL)
970     O << ", texmode_independent";
971   if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
972     if (!nvptxSubtarget.hasDouble())
973       O << ", map_f64_to_f32";
974   }
975 
976   if (MAI->doesSupportDebugInformation())
977     O << ", debug";
978 
979   O << "\n";
980 
981   O << ".address_size ";
982   if (nvptxSubtarget.is64Bit())
983     O << "64";
984   else
985     O << "32";
986   O << "\n";
987 
988   O << "\n";
989 }
990 
991 bool NVPTXAsmPrinter::doFinalization(Module &M) {
992 
993   // If we did not emit any functions, then the global declarations have not
994   // yet been emitted.
995   if (!GlobalsEmitted) {
996     emitGlobals(M);
997     GlobalsEmitted = true;
998   }
999 
1000   // XXX Temproarily remove global variables so that doFinalization() will not
1001   // emit them again (global variables are emitted at beginning).
1002 
1003   Module::GlobalListType &global_list = M.getGlobalList();
1004   int i, n = global_list.size();
1005   GlobalVariable **gv_array = new GlobalVariable *[n];
1006 
1007   // first, back-up GlobalVariable in gv_array
1008   i = 0;
1009   for (Module::global_iterator I = global_list.begin(), E = global_list.end();
1010        I != E; ++I)
1011     gv_array[i++] = &*I;
1012 
1013   // second, empty global_list
1014   while (!global_list.empty())
1015     global_list.remove(global_list.begin());
1016 
1017   // call doFinalization
1018   bool ret = AsmPrinter::doFinalization(M);
1019 
1020   // now we restore global variables
1021   for (i = 0; i < n; i++)
1022     global_list.insert(global_list.end(), gv_array[i]);
1023 
1024   delete[] gv_array;
1025   return ret;
1026 
1027   //bool Result = AsmPrinter::doFinalization(M);
1028   // Instead of calling the parents doFinalization, we may
1029   // clone parents doFinalization and customize here.
1030   // Currently, we if NVISA out the EmitGlobals() in
1031   // parent's doFinalization, which is too intrusive.
1032   //
1033   // Same for the doInitialization.
1034   //return Result;
1035 }
1036 
1037 // This function emits appropriate linkage directives for
1038 // functions and global variables.
1039 //
1040 // extern function declaration            -> .extern
1041 // extern function definition             -> .visible
1042 // external global variable with init     -> .visible
1043 // external without init                  -> .extern
1044 // appending                              -> not allowed, assert.
1045 
1046 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
1047                                            raw_ostream &O) {
1048   if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
1049     if (V->hasExternalLinkage()) {
1050       if (isa<GlobalVariable>(V)) {
1051         const GlobalVariable *GVar = cast<GlobalVariable>(V);
1052         if (GVar) {
1053           if (GVar->hasInitializer())
1054             O << ".visible ";
1055           else
1056             O << ".extern ";
1057         }
1058       } else if (V->isDeclaration())
1059         O << ".extern ";
1060       else
1061         O << ".visible ";
1062     } else if (V->hasAppendingLinkage()) {
1063       std::string msg;
1064       msg.append("Error: ");
1065       msg.append("Symbol ");
1066       if (V->hasName())
1067         msg.append(V->getName().str());
1068       msg.append("has unsupported appending linkage type");
1069       llvm_unreachable(msg.c_str());
1070     }
1071   }
1072 }
1073 
1074 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
1075                                          raw_ostream &O,
1076                                          bool processDemoted) {
1077 
1078   // Skip meta data
1079   if (GVar->hasSection()) {
1080     if (GVar->getSection() == "llvm.metadata")
1081       return;
1082   }
1083 
1084   const DataLayout *TD = TM.getDataLayout();
1085 
1086   // GlobalVariables are always constant pointers themselves.
1087   const PointerType *PTy = GVar->getType();
1088   Type *ETy = PTy->getElementType();
1089 
1090   if (GVar->hasExternalLinkage()) {
1091     if (GVar->hasInitializer())
1092       O << ".visible ";
1093     else
1094       O << ".extern ";
1095   }
1096 
1097   if (llvm::isTexture(*GVar)) {
1098     O << ".global .texref " << llvm::getTextureName(*GVar) << ";\n";
1099     return;
1100   }
1101 
1102   if (llvm::isSurface(*GVar)) {
1103     O << ".global .surfref " << llvm::getSurfaceName(*GVar) << ";\n";
1104     return;
1105   }
1106 
1107   if (GVar->isDeclaration()) {
1108     // (extern) declarations, no definition or initializer
1109     // Currently the only known declaration is for an automatic __local
1110     // (.shared) promoted to global.
1111     emitPTXGlobalVariable(GVar, O);
1112     O << ";\n";
1113     return;
1114   }
1115 
1116   if (llvm::isSampler(*GVar)) {
1117     O << ".global .samplerref " << llvm::getSamplerName(*GVar);
1118 
1119     const Constant *Initializer = NULL;
1120     if (GVar->hasInitializer())
1121       Initializer = GVar->getInitializer();
1122     const ConstantInt *CI = NULL;
1123     if (Initializer)
1124       CI = dyn_cast<ConstantInt>(Initializer);
1125     if (CI) {
1126       unsigned sample = CI->getZExtValue();
1127 
1128       O << " = { ";
1129 
1130       for (int i = 0,
1131                addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1132            i < 3; i++) {
1133         O << "addr_mode_" << i << " = ";
1134         switch (addr) {
1135         case 0:
1136           O << "wrap";
1137           break;
1138         case 1:
1139           O << "clamp_to_border";
1140           break;
1141         case 2:
1142           O << "clamp_to_edge";
1143           break;
1144         case 3:
1145           O << "wrap";
1146           break;
1147         case 4:
1148           O << "mirror";
1149           break;
1150         }
1151         O << ", ";
1152       }
1153       O << "filter_mode = ";
1154       switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1155       case 0:
1156         O << "nearest";
1157         break;
1158       case 1:
1159         O << "linear";
1160         break;
1161       case 2:
1162         assert(0 && "Anisotropic filtering is not supported");
1163       default:
1164         O << "nearest";
1165         break;
1166       }
1167       if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1168         O << ", force_unnormalized_coords = 1";
1169       }
1170       O << " }";
1171     }
1172 
1173     O << ";\n";
1174     return;
1175   }
1176 
1177   if (GVar->hasPrivateLinkage()) {
1178 
1179     if (!strncmp(GVar->getName().data(), "unrollpragma", 12))
1180       return;
1181 
1182     // FIXME - need better way (e.g. Metadata) to avoid generating this global
1183     if (!strncmp(GVar->getName().data(), "filename", 8))
1184       return;
1185     if (GVar->use_empty())
1186       return;
1187   }
1188 
1189   const Function *demotedFunc = 0;
1190   if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
1191     O << "// " << GVar->getName().str() << " has been demoted\n";
1192     if (localDecls.find(demotedFunc) != localDecls.end())
1193       localDecls[demotedFunc].push_back(GVar);
1194     else {
1195       std::vector<const GlobalVariable *> temp;
1196       temp.push_back(GVar);
1197       localDecls[demotedFunc] = temp;
1198     }
1199     return;
1200   }
1201 
1202   O << ".";
1203   emitPTXAddressSpace(PTy->getAddressSpace(), O);
1204   if (GVar->getAlignment() == 0)
1205     O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1206   else
1207     O << " .align " << GVar->getAlignment();
1208 
1209   if (ETy->isSingleValueType()) {
1210     O << " .";
1211     // Special case: ABI requires that we use .u8 for predicates
1212     if (ETy->isIntegerTy(1))
1213       O << "u8";
1214     else
1215       O << getPTXFundamentalTypeStr(ETy, false);
1216     O << " ";
1217     O << *getSymbol(GVar);
1218 
1219     // Ptx allows variable initilization only for constant and global state
1220     // spaces.
1221     if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1222          (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1223         GVar->hasInitializer()) {
1224       const Constant *Initializer = GVar->getInitializer();
1225       if (!Initializer->isNullValue()) {
1226         O << " = ";
1227         printScalarConstant(Initializer, O);
1228       }
1229     }
1230   } else {
1231     unsigned int ElementSize = 0;
1232 
1233     // Although PTX has direct support for struct type and array type and
1234     // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1235     // targets that support these high level field accesses. Structs, arrays
1236     // and vectors are lowered into arrays of bytes.
1237     switch (ETy->getTypeID()) {
1238     case Type::StructTyID:
1239     case Type::ArrayTyID:
1240     case Type::VectorTyID:
1241       ElementSize = TD->getTypeStoreSize(ETy);
1242       // Ptx allows variable initilization only for constant and
1243       // global state spaces.
1244       if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1245            (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1246           GVar->hasInitializer()) {
1247         const Constant *Initializer = GVar->getInitializer();
1248         if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1249           AggBuffer aggBuffer(ElementSize, O, *this);
1250           bufferAggregateConstant(Initializer, &aggBuffer);
1251           if (aggBuffer.numSymbols) {
1252             if (nvptxSubtarget.is64Bit()) {
1253               O << " .u64 " << *getSymbol(GVar) << "[";
1254               O << ElementSize / 8;
1255             } else {
1256               O << " .u32 " << *getSymbol(GVar) << "[";
1257               O << ElementSize / 4;
1258             }
1259             O << "]";
1260           } else {
1261             O << " .b8 " << *getSymbol(GVar) << "[";
1262             O << ElementSize;
1263             O << "]";
1264           }
1265           O << " = {";
1266           aggBuffer.print();
1267           O << "}";
1268         } else {
1269           O << " .b8 " << *getSymbol(GVar);
1270           if (ElementSize) {
1271             O << "[";
1272             O << ElementSize;
1273             O << "]";
1274           }
1275         }
1276       } else {
1277         O << " .b8 " << *getSymbol(GVar);
1278         if (ElementSize) {
1279           O << "[";
1280           O << ElementSize;
1281           O << "]";
1282         }
1283       }
1284       break;
1285     default:
1286       assert(0 && "type not supported yet");
1287     }
1288 
1289   }
1290   O << ";\n";
1291 }
1292 
1293 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1294   if (localDecls.find(f) == localDecls.end())
1295     return;
1296 
1297   std::vector<const GlobalVariable *> &gvars = localDecls[f];
1298 
1299   for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1300     O << "\t// demoted variable\n\t";
1301     printModuleLevelGV(gvars[i], O, true);
1302   }
1303 }
1304 
1305 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1306                                           raw_ostream &O) const {
1307   switch (AddressSpace) {
1308   case llvm::ADDRESS_SPACE_LOCAL:
1309     O << "local";
1310     break;
1311   case llvm::ADDRESS_SPACE_GLOBAL:
1312     O << "global";
1313     break;
1314   case llvm::ADDRESS_SPACE_CONST:
1315     O << "const";
1316     break;
1317   case llvm::ADDRESS_SPACE_SHARED:
1318     O << "shared";
1319     break;
1320   default:
1321     report_fatal_error("Bad address space found while emitting PTX");
1322     break;
1323   }
1324 }
1325 
1326 std::string
1327 NVPTXAsmPrinter::getPTXFundamentalTypeStr(const Type *Ty, bool useB4PTR) const {
1328   switch (Ty->getTypeID()) {
1329   default:
1330     llvm_unreachable("unexpected type");
1331     break;
1332   case Type::IntegerTyID: {
1333     unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1334     if (NumBits == 1)
1335       return "pred";
1336     else if (NumBits <= 64) {
1337       std::string name = "u";
1338       return name + utostr(NumBits);
1339     } else {
1340       llvm_unreachable("Integer too large");
1341       break;
1342     }
1343     break;
1344   }
1345   case Type::FloatTyID:
1346     return "f32";
1347   case Type::DoubleTyID:
1348     return "f64";
1349   case Type::PointerTyID:
1350     if (nvptxSubtarget.is64Bit())
1351       if (useB4PTR)
1352         return "b64";
1353       else
1354         return "u64";
1355     else if (useB4PTR)
1356       return "b32";
1357     else
1358       return "u32";
1359   }
1360   llvm_unreachable("unexpected type");
1361   return NULL;
1362 }
1363 
1364 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1365                                             raw_ostream &O) {
1366 
1367   const DataLayout *TD = TM.getDataLayout();
1368 
1369   // GlobalVariables are always constant pointers themselves.
1370   const PointerType *PTy = GVar->getType();
1371   Type *ETy = PTy->getElementType();
1372 
1373   O << ".";
1374   emitPTXAddressSpace(PTy->getAddressSpace(), O);
1375   if (GVar->getAlignment() == 0)
1376     O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1377   else
1378     O << " .align " << GVar->getAlignment();
1379 
1380   if (ETy->isSingleValueType()) {
1381     O << " .";
1382     O << getPTXFundamentalTypeStr(ETy);
1383     O << " ";
1384     O << *getSymbol(GVar);
1385     return;
1386   }
1387 
1388   int64_t ElementSize = 0;
1389 
1390   // Although PTX has direct support for struct type and array type and LLVM IR
1391   // is very similar to PTX, the LLVM CodeGen does not support for targets that
1392   // support these high level field accesses. Structs and arrays are lowered
1393   // into arrays of bytes.
1394   switch (ETy->getTypeID()) {
1395   case Type::StructTyID:
1396   case Type::ArrayTyID:
1397   case Type::VectorTyID:
1398     ElementSize = TD->getTypeStoreSize(ETy);
1399     O << " .b8 " << *getSymbol(GVar) << "[";
1400     if (ElementSize) {
1401       O << itostr(ElementSize);
1402     }
1403     O << "]";
1404     break;
1405   default:
1406     assert(0 && "type not supported yet");
1407   }
1408   return;
1409 }
1410 
1411 static unsigned int getOpenCLAlignment(const DataLayout *TD, Type *Ty) {
1412   if (Ty->isSingleValueType())
1413     return TD->getPrefTypeAlignment(Ty);
1414 
1415   const ArrayType *ATy = dyn_cast<ArrayType>(Ty);
1416   if (ATy)
1417     return getOpenCLAlignment(TD, ATy->getElementType());
1418 
1419   const VectorType *VTy = dyn_cast<VectorType>(Ty);
1420   if (VTy) {
1421     Type *ETy = VTy->getElementType();
1422     unsigned int numE = VTy->getNumElements();
1423     unsigned int alignE = TD->getPrefTypeAlignment(ETy);
1424     if (numE == 3)
1425       return 4 * alignE;
1426     else
1427       return numE * alignE;
1428   }
1429 
1430   const StructType *STy = dyn_cast<StructType>(Ty);
1431   if (STy) {
1432     unsigned int alignStruct = 1;
1433     // Go through each element of the struct and find the
1434     // largest alignment.
1435     for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1436       Type *ETy = STy->getElementType(i);
1437       unsigned int align = getOpenCLAlignment(TD, ETy);
1438       if (align > alignStruct)
1439         alignStruct = align;
1440     }
1441     return alignStruct;
1442   }
1443 
1444   const FunctionType *FTy = dyn_cast<FunctionType>(Ty);
1445   if (FTy)
1446     return TD->getPointerPrefAlignment();
1447   return TD->getPrefTypeAlignment(Ty);
1448 }
1449 
1450 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1451                                      int paramIndex, raw_ostream &O) {
1452   if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1453       (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA))
1454     O << *getSymbol(I->getParent()) << "_param_" << paramIndex;
1455   else {
1456     std::string argName = I->getName();
1457     const char *p = argName.c_str();
1458     while (*p) {
1459       if (*p == '.')
1460         O << "_";
1461       else
1462         O << *p;
1463       p++;
1464     }
1465   }
1466 }
1467 
1468 void NVPTXAsmPrinter::printParamName(int paramIndex, raw_ostream &O) {
1469   Function::const_arg_iterator I, E;
1470   int i = 0;
1471 
1472   if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1473       (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)) {
1474     O << *CurrentFnSym << "_param_" << paramIndex;
1475     return;
1476   }
1477 
1478   for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, i++) {
1479     if (i == paramIndex) {
1480       printParamName(I, paramIndex, O);
1481       return;
1482     }
1483   }
1484   llvm_unreachable("paramIndex out of bound");
1485 }
1486 
1487 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1488   const DataLayout *TD = TM.getDataLayout();
1489   const AttributeSet &PAL = F->getAttributes();
1490   const TargetLowering *TLI = TM.getTargetLowering();
1491   Function::const_arg_iterator I, E;
1492   unsigned paramIndex = 0;
1493   bool first = true;
1494   bool isKernelFunc = llvm::isKernelFunction(*F);
1495   bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
1496   MVT thePointerTy = TLI->getPointerTy();
1497 
1498   O << "(\n";
1499 
1500   for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1501     Type *Ty = I->getType();
1502 
1503     if (!first)
1504       O << ",\n";
1505 
1506     first = false;
1507 
1508     // Handle image/sampler parameters
1509     if (llvm::isSampler(*I) || llvm::isImage(*I)) {
1510       if (llvm::isImage(*I)) {
1511         std::string sname = I->getName();
1512         if (llvm::isImageWriteOnly(*I))
1513           O << "\t.param .surfref " << *getSymbol(F) << "_param_"
1514             << paramIndex;
1515         else // Default image is read_only
1516           O << "\t.param .texref " << *getSymbol(F) << "_param_"
1517             << paramIndex;
1518       } else // Should be llvm::isSampler(*I)
1519         O << "\t.param .samplerref " << *getSymbol(F) << "_param_"
1520           << paramIndex;
1521       continue;
1522     }
1523 
1524     if (PAL.hasAttribute(paramIndex + 1, Attribute::ByVal) == false) {
1525       if (Ty->isAggregateType() || Ty->isVectorTy()) {
1526         // Just print .param .align <a> .b8 .param[size];
1527         // <a> = PAL.getparamalignment
1528         // size = typeallocsize of element type
1529         unsigned align = PAL.getParamAlignment(paramIndex + 1);
1530         if (align == 0)
1531           align = TD->getABITypeAlignment(Ty);
1532 
1533         unsigned sz = TD->getTypeAllocSize(Ty);
1534         O << "\t.param .align " << align << " .b8 ";
1535         printParamName(I, paramIndex, O);
1536         O << "[" << sz << "]";
1537 
1538         continue;
1539       }
1540       // Just a scalar
1541       const PointerType *PTy = dyn_cast<PointerType>(Ty);
1542       if (isKernelFunc) {
1543         if (PTy) {
1544           // Special handling for pointer arguments to kernel
1545           O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1546 
1547           if (nvptxSubtarget.getDrvInterface() != NVPTX::CUDA) {
1548             Type *ETy = PTy->getElementType();
1549             int addrSpace = PTy->getAddressSpace();
1550             switch (addrSpace) {
1551             default:
1552               O << ".ptr ";
1553               break;
1554             case llvm::ADDRESS_SPACE_CONST:
1555               O << ".ptr .const ";
1556               break;
1557             case llvm::ADDRESS_SPACE_SHARED:
1558               O << ".ptr .shared ";
1559               break;
1560             case llvm::ADDRESS_SPACE_GLOBAL:
1561               O << ".ptr .global ";
1562               break;
1563             }
1564             O << ".align " << (int) getOpenCLAlignment(TD, ETy) << " ";
1565           }
1566           printParamName(I, paramIndex, O);
1567           continue;
1568         }
1569 
1570         // non-pointer scalar to kernel func
1571         O << "\t.param .";
1572         // Special case: predicate operands become .u8 types
1573         if (Ty->isIntegerTy(1))
1574           O << "u8";
1575         else
1576           O << getPTXFundamentalTypeStr(Ty);
1577         O << " ";
1578         printParamName(I, paramIndex, O);
1579         continue;
1580       }
1581       // Non-kernel function, just print .param .b<size> for ABI
1582       // and .reg .b<size> for non-ABI
1583       unsigned sz = 0;
1584       if (isa<IntegerType>(Ty)) {
1585         sz = cast<IntegerType>(Ty)->getBitWidth();
1586         if (sz < 32)
1587           sz = 32;
1588       } else if (isa<PointerType>(Ty))
1589         sz = thePointerTy.getSizeInBits();
1590       else
1591         sz = Ty->getPrimitiveSizeInBits();
1592       if (isABI)
1593         O << "\t.param .b" << sz << " ";
1594       else
1595         O << "\t.reg .b" << sz << " ";
1596       printParamName(I, paramIndex, O);
1597       continue;
1598     }
1599 
1600     // param has byVal attribute. So should be a pointer
1601     const PointerType *PTy = dyn_cast<PointerType>(Ty);
1602     assert(PTy && "Param with byval attribute should be a pointer type");
1603     Type *ETy = PTy->getElementType();
1604 
1605     if (isABI || isKernelFunc) {
1606       // Just print .param .align <a> .b8 .param[size];
1607       // <a> = PAL.getparamalignment
1608       // size = typeallocsize of element type
1609       unsigned align = PAL.getParamAlignment(paramIndex + 1);
1610       if (align == 0)
1611         align = TD->getABITypeAlignment(ETy);
1612 
1613       unsigned sz = TD->getTypeAllocSize(ETy);
1614       O << "\t.param .align " << align << " .b8 ";
1615       printParamName(I, paramIndex, O);
1616       O << "[" << sz << "]";
1617       continue;
1618     } else {
1619       // Split the ETy into constituent parts and
1620       // print .param .b<size> <name> for each part.
1621       // Further, if a part is vector, print the above for
1622       // each vector element.
1623       SmallVector<EVT, 16> vtparts;
1624       ComputeValueVTs(*TLI, ETy, vtparts);
1625       for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1626         unsigned elems = 1;
1627         EVT elemtype = vtparts[i];
1628         if (vtparts[i].isVector()) {
1629           elems = vtparts[i].getVectorNumElements();
1630           elemtype = vtparts[i].getVectorElementType();
1631         }
1632 
1633         for (unsigned j = 0, je = elems; j != je; ++j) {
1634           unsigned sz = elemtype.getSizeInBits();
1635           if (elemtype.isInteger() && (sz < 32))
1636             sz = 32;
1637           O << "\t.reg .b" << sz << " ";
1638           printParamName(I, paramIndex, O);
1639           if (j < je - 1)
1640             O << ",\n";
1641           ++paramIndex;
1642         }
1643         if (i < e - 1)
1644           O << ",\n";
1645       }
1646       --paramIndex;
1647       continue;
1648     }
1649   }
1650 
1651   O << "\n)\n";
1652 }
1653 
1654 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1655                                             raw_ostream &O) {
1656   const Function *F = MF.getFunction();
1657   emitFunctionParamList(F, O);
1658 }
1659 
1660 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1661     const MachineFunction &MF) {
1662   SmallString<128> Str;
1663   raw_svector_ostream O(Str);
1664 
1665   // Map the global virtual register number to a register class specific
1666   // virtual register number starting from 1 with that class.
1667   const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo();
1668   //unsigned numRegClasses = TRI->getNumRegClasses();
1669 
1670   // Emit the Fake Stack Object
1671   const MachineFrameInfo *MFI = MF.getFrameInfo();
1672   int NumBytes = (int) MFI->getStackSize();
1673   if (NumBytes) {
1674     O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
1675       << getFunctionNumber() << "[" << NumBytes << "];\n";
1676     if (nvptxSubtarget.is64Bit()) {
1677       O << "\t.reg .b64 \t%SP;\n";
1678       O << "\t.reg .b64 \t%SPL;\n";
1679     } else {
1680       O << "\t.reg .b32 \t%SP;\n";
1681       O << "\t.reg .b32 \t%SPL;\n";
1682     }
1683   }
1684 
1685   // Go through all virtual registers to establish the mapping between the
1686   // global virtual
1687   // register number and the per class virtual register number.
1688   // We use the per class virtual register number in the ptx output.
1689   unsigned int numVRs = MRI->getNumVirtRegs();
1690   for (unsigned i = 0; i < numVRs; i++) {
1691     unsigned int vr = TRI->index2VirtReg(i);
1692     const TargetRegisterClass *RC = MRI->getRegClass(vr);
1693     DenseMap<unsigned, unsigned> &regmap = VRegMapping[RC];
1694     int n = regmap.size();
1695     regmap.insert(std::make_pair(vr, n + 1));
1696   }
1697 
1698   // Emit register declarations
1699   // @TODO: Extract out the real register usage
1700   // O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
1701   // O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
1702   // O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
1703   // O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
1704   // O << "\t.reg .s64 %rl<" << NVPTXNumRegisters << ">;\n";
1705   // O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
1706   // O << "\t.reg .f64 %fl<" << NVPTXNumRegisters << ">;\n";
1707 
1708   // Emit declaration of the virtual registers or 'physical' registers for
1709   // each register class
1710   for (unsigned i=0; i< TRI->getNumRegClasses(); i++) {
1711     const TargetRegisterClass *RC = TRI->getRegClass(i);
1712     DenseMap<unsigned, unsigned> &regmap = VRegMapping[RC];
1713     std::string rcname = getNVPTXRegClassName(RC);
1714     std::string rcStr = getNVPTXRegClassStr(RC);
1715     int n = regmap.size();
1716 
1717     // Only declare those registers that may be used.
1718     if (n) {
1719        O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
1720          << ">;\n";
1721     }
1722   }
1723 
1724   OutStreamer.EmitRawText(O.str());
1725 }
1726 
1727 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
1728   APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
1729   bool ignored;
1730   unsigned int numHex;
1731   const char *lead;
1732 
1733   if (Fp->getType()->getTypeID() == Type::FloatTyID) {
1734     numHex = 8;
1735     lead = "0f";
1736     APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &ignored);
1737   } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
1738     numHex = 16;
1739     lead = "0d";
1740     APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &ignored);
1741   } else
1742     llvm_unreachable("unsupported fp type");
1743 
1744   APInt API = APF.bitcastToAPInt();
1745   std::string hexstr(utohexstr(API.getZExtValue()));
1746   O << lead;
1747   if (hexstr.length() < numHex)
1748     O << std::string(numHex - hexstr.length(), '0');
1749   O << utohexstr(API.getZExtValue());
1750 }
1751 
1752 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
1753   if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
1754     O << CI->getValue();
1755     return;
1756   }
1757   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
1758     printFPConstant(CFP, O);
1759     return;
1760   }
1761   if (isa<ConstantPointerNull>(CPV)) {
1762     O << "0";
1763     return;
1764   }
1765   if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1766     O << *getSymbol(GVar);
1767     return;
1768   }
1769   if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1770     const Value *v = Cexpr->stripPointerCasts();
1771     if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
1772       O << *getSymbol(GVar);
1773       return;
1774     } else {
1775       O << *LowerConstant(CPV, *this);
1776       return;
1777     }
1778   }
1779   llvm_unreachable("Not scalar type found in printScalarConstant()");
1780 }
1781 
1782 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
1783                                    AggBuffer *aggBuffer) {
1784 
1785   const DataLayout *TD = TM.getDataLayout();
1786 
1787   if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
1788     int s = TD->getTypeAllocSize(CPV->getType());
1789     if (s < Bytes)
1790       s = Bytes;
1791     aggBuffer->addZeros(s);
1792     return;
1793   }
1794 
1795   unsigned char *ptr;
1796   switch (CPV->getType()->getTypeID()) {
1797 
1798   case Type::IntegerTyID: {
1799     const Type *ETy = CPV->getType();
1800     if (ETy == Type::getInt8Ty(CPV->getContext())) {
1801       unsigned char c =
1802           (unsigned char)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
1803       ptr = &c;
1804       aggBuffer->addBytes(ptr, 1, Bytes);
1805     } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
1806       short int16 = (short)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
1807       ptr = (unsigned char *)&int16;
1808       aggBuffer->addBytes(ptr, 2, Bytes);
1809     } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
1810       if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1811         int int32 = (int)(constInt->getZExtValue());
1812         ptr = (unsigned char *)&int32;
1813         aggBuffer->addBytes(ptr, 4, Bytes);
1814         break;
1815       } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1816         if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1817                 ConstantFoldConstantExpression(Cexpr, TD))) {
1818           int int32 = (int)(constInt->getZExtValue());
1819           ptr = (unsigned char *)&int32;
1820           aggBuffer->addBytes(ptr, 4, Bytes);
1821           break;
1822         }
1823         if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1824           Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1825           aggBuffer->addSymbol(v);
1826           aggBuffer->addZeros(4);
1827           break;
1828         }
1829       }
1830       llvm_unreachable("unsupported integer const type");
1831     } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
1832       if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1833         long long int64 = (long long)(constInt->getZExtValue());
1834         ptr = (unsigned char *)&int64;
1835         aggBuffer->addBytes(ptr, 8, Bytes);
1836         break;
1837       } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1838         if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1839                 ConstantFoldConstantExpression(Cexpr, TD))) {
1840           long long int64 = (long long)(constInt->getZExtValue());
1841           ptr = (unsigned char *)&int64;
1842           aggBuffer->addBytes(ptr, 8, Bytes);
1843           break;
1844         }
1845         if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1846           Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1847           aggBuffer->addSymbol(v);
1848           aggBuffer->addZeros(8);
1849           break;
1850         }
1851       }
1852       llvm_unreachable("unsupported integer const type");
1853     } else
1854       llvm_unreachable("unsupported integer const type");
1855     break;
1856   }
1857   case Type::FloatTyID:
1858   case Type::DoubleTyID: {
1859     const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
1860     const Type *Ty = CFP->getType();
1861     if (Ty == Type::getFloatTy(CPV->getContext())) {
1862       float float32 = (float) CFP->getValueAPF().convertToFloat();
1863       ptr = (unsigned char *)&float32;
1864       aggBuffer->addBytes(ptr, 4, Bytes);
1865     } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
1866       double float64 = CFP->getValueAPF().convertToDouble();
1867       ptr = (unsigned char *)&float64;
1868       aggBuffer->addBytes(ptr, 8, Bytes);
1869     } else {
1870       llvm_unreachable("unsupported fp const type");
1871     }
1872     break;
1873   }
1874   case Type::PointerTyID: {
1875     if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1876       aggBuffer->addSymbol(GVar);
1877     } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1878       const Value *v = Cexpr->stripPointerCasts();
1879       aggBuffer->addSymbol(v);
1880     }
1881     unsigned int s = TD->getTypeAllocSize(CPV->getType());
1882     aggBuffer->addZeros(s);
1883     break;
1884   }
1885 
1886   case Type::ArrayTyID:
1887   case Type::VectorTyID:
1888   case Type::StructTyID: {
1889     if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV) ||
1890         isa<ConstantStruct>(CPV) || isa<ConstantDataSequential>(CPV)) {
1891       int ElementSize = TD->getTypeAllocSize(CPV->getType());
1892       bufferAggregateConstant(CPV, aggBuffer);
1893       if (Bytes > ElementSize)
1894         aggBuffer->addZeros(Bytes - ElementSize);
1895     } else if (isa<ConstantAggregateZero>(CPV))
1896       aggBuffer->addZeros(Bytes);
1897     else
1898       llvm_unreachable("Unexpected Constant type");
1899     break;
1900   }
1901 
1902   default:
1903     llvm_unreachable("unsupported type");
1904   }
1905 }
1906 
1907 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
1908                                               AggBuffer *aggBuffer) {
1909   const DataLayout *TD = TM.getDataLayout();
1910   int Bytes;
1911 
1912   // Old constants
1913   if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
1914     if (CPV->getNumOperands())
1915       for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
1916         bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
1917     return;
1918   }
1919 
1920   if (const ConstantDataSequential *CDS =
1921           dyn_cast<ConstantDataSequential>(CPV)) {
1922     if (CDS->getNumElements())
1923       for (unsigned i = 0; i < CDS->getNumElements(); ++i)
1924         bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
1925                      aggBuffer);
1926     return;
1927   }
1928 
1929   if (isa<ConstantStruct>(CPV)) {
1930     if (CPV->getNumOperands()) {
1931       StructType *ST = cast<StructType>(CPV->getType());
1932       for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
1933         if (i == (e - 1))
1934           Bytes = TD->getStructLayout(ST)->getElementOffset(0) +
1935                   TD->getTypeAllocSize(ST) -
1936                   TD->getStructLayout(ST)->getElementOffset(i);
1937         else
1938           Bytes = TD->getStructLayout(ST)->getElementOffset(i + 1) -
1939                   TD->getStructLayout(ST)->getElementOffset(i);
1940         bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
1941       }
1942     }
1943     return;
1944   }
1945   llvm_unreachable("unsupported constant type in printAggregateConstant()");
1946 }
1947 
1948 // buildTypeNameMap - Run through symbol table looking for type names.
1949 //
1950 
1951 bool NVPTXAsmPrinter::isImageType(const Type *Ty) {
1952 
1953   std::map<const Type *, std::string>::iterator PI = TypeNameMap.find(Ty);
1954 
1955   if (PI != TypeNameMap.end() && (!PI->second.compare("struct._image1d_t") ||
1956                                   !PI->second.compare("struct._image2d_t") ||
1957                                   !PI->second.compare("struct._image3d_t")))
1958     return true;
1959 
1960   return false;
1961 }
1962 
1963 
1964 bool NVPTXAsmPrinter::ignoreLoc(const MachineInstr &MI) {
1965   switch (MI.getOpcode()) {
1966   default:
1967     return false;
1968   case NVPTX::CallArgBeginInst:
1969   case NVPTX::CallArgEndInst0:
1970   case NVPTX::CallArgEndInst1:
1971   case NVPTX::CallArgF32:
1972   case NVPTX::CallArgF64:
1973   case NVPTX::CallArgI16:
1974   case NVPTX::CallArgI32:
1975   case NVPTX::CallArgI32imm:
1976   case NVPTX::CallArgI64:
1977   case NVPTX::CallArgParam:
1978   case NVPTX::CallVoidInst:
1979   case NVPTX::CallVoidInstReg:
1980   case NVPTX::Callseq_End:
1981   case NVPTX::CallVoidInstReg64:
1982   case NVPTX::DeclareParamInst:
1983   case NVPTX::DeclareRetMemInst:
1984   case NVPTX::DeclareRetRegInst:
1985   case NVPTX::DeclareRetScalarInst:
1986   case NVPTX::DeclareScalarParamInst:
1987   case NVPTX::DeclareScalarRegInst:
1988   case NVPTX::StoreParamF32:
1989   case NVPTX::StoreParamF64:
1990   case NVPTX::StoreParamI16:
1991   case NVPTX::StoreParamI32:
1992   case NVPTX::StoreParamI64:
1993   case NVPTX::StoreParamI8:
1994   case NVPTX::StoreRetvalF32:
1995   case NVPTX::StoreRetvalF64:
1996   case NVPTX::StoreRetvalI16:
1997   case NVPTX::StoreRetvalI32:
1998   case NVPTX::StoreRetvalI64:
1999   case NVPTX::StoreRetvalI8:
2000   case NVPTX::LastCallArgF32:
2001   case NVPTX::LastCallArgF64:
2002   case NVPTX::LastCallArgI16:
2003   case NVPTX::LastCallArgI32:
2004   case NVPTX::LastCallArgI32imm:
2005   case NVPTX::LastCallArgI64:
2006   case NVPTX::LastCallArgParam:
2007   case NVPTX::LoadParamMemF32:
2008   case NVPTX::LoadParamMemF64:
2009   case NVPTX::LoadParamMemI16:
2010   case NVPTX::LoadParamMemI32:
2011   case NVPTX::LoadParamMemI64:
2012   case NVPTX::LoadParamMemI8:
2013   case NVPTX::PrototypeInst:
2014   case NVPTX::DBG_VALUE:
2015     return true;
2016   }
2017   return false;
2018 }
2019 
2020 /// PrintAsmOperand - Print out an operand for an inline asm expression.
2021 ///
2022 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
2023                                       unsigned AsmVariant,
2024                                       const char *ExtraCode, raw_ostream &O) {
2025   if (ExtraCode && ExtraCode[0]) {
2026     if (ExtraCode[1] != 0)
2027       return true; // Unknown modifier.
2028 
2029     switch (ExtraCode[0]) {
2030     default:
2031       // See if this is a generic print operand
2032       return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
2033     case 'r':
2034       break;
2035     }
2036   }
2037 
2038   printOperand(MI, OpNo, O);
2039 
2040   return false;
2041 }
2042 
2043 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(
2044     const MachineInstr *MI, unsigned OpNo, unsigned AsmVariant,
2045     const char *ExtraCode, raw_ostream &O) {
2046   if (ExtraCode && ExtraCode[0])
2047     return true; // Unknown modifier
2048 
2049   O << '[';
2050   printMemOperand(MI, OpNo, O);
2051   O << ']';
2052 
2053   return false;
2054 }
2055 
2056 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
2057                                    raw_ostream &O, const char *Modifier) {
2058   const MachineOperand &MO = MI->getOperand(opNum);
2059   switch (MO.getType()) {
2060   case MachineOperand::MO_Register:
2061     if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
2062       if (MO.getReg() == NVPTX::VRDepot)
2063         O << DEPOTNAME << getFunctionNumber();
2064       else
2065         O << NVPTXInstPrinter::getRegisterName(MO.getReg());
2066     } else {
2067       emitVirtualRegister(MO.getReg(), O);
2068     }
2069     return;
2070 
2071   case MachineOperand::MO_Immediate:
2072     if (!Modifier)
2073       O << MO.getImm();
2074     else if (strstr(Modifier, "vec") == Modifier)
2075       printVecModifiedImmediate(MO, Modifier, O);
2076     else
2077       llvm_unreachable(
2078           "Don't know how to handle modifier on immediate operand");
2079     return;
2080 
2081   case MachineOperand::MO_FPImmediate:
2082     printFPConstant(MO.getFPImm(), O);
2083     break;
2084 
2085   case MachineOperand::MO_GlobalAddress:
2086     O << *getSymbol(MO.getGlobal());
2087     break;
2088 
2089   case MachineOperand::MO_MachineBasicBlock:
2090     O << *MO.getMBB()->getSymbol();
2091     return;
2092 
2093   default:
2094     llvm_unreachable("Operand type not supported.");
2095   }
2096 }
2097 
2098 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
2099                                       raw_ostream &O, const char *Modifier) {
2100   printOperand(MI, opNum, O);
2101 
2102   if (Modifier && !strcmp(Modifier, "add")) {
2103     O << ", ";
2104     printOperand(MI, opNum + 1, O);
2105   } else {
2106     if (MI->getOperand(opNum + 1).isImm() &&
2107         MI->getOperand(opNum + 1).getImm() == 0)
2108       return; // don't print ',0' or '+0'
2109     O << "+";
2110     printOperand(MI, opNum + 1, O);
2111   }
2112 }
2113 
2114 
2115 // Force static initialization.
2116 extern "C" void LLVMInitializeNVPTXBackendAsmPrinter() {
2117   RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2118   RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2119 }
2120 
2121 void NVPTXAsmPrinter::emitSrcInText(StringRef filename, unsigned line) {
2122   std::stringstream temp;
2123   LineReader *reader = this->getReader(filename.str());
2124   temp << "\n//";
2125   temp << filename.str();
2126   temp << ":";
2127   temp << line;
2128   temp << " ";
2129   temp << reader->readLine(line);
2130   temp << "\n";
2131   this->OutStreamer.EmitRawText(Twine(temp.str()));
2132 }
2133 
2134 LineReader *NVPTXAsmPrinter::getReader(std::string filename) {
2135   if (reader == NULL) {
2136     reader = new LineReader(filename);
2137   }
2138 
2139   if (reader->fileName() != filename) {
2140     delete reader;
2141     reader = new LineReader(filename);
2142   }
2143 
2144   return reader;
2145 }
2146 
2147 std::string LineReader::readLine(unsigned lineNum) {
2148   if (lineNum < theCurLine) {
2149     theCurLine = 0;
2150     fstr.seekg(0, std::ios::beg);
2151   }
2152   while (theCurLine < lineNum) {
2153     fstr.getline(buff, 500);
2154     theCurLine++;
2155   }
2156   return buff;
2157 }
2158 
2159 // Force static initialization.
2160 extern "C" void LLVMInitializeNVPTXAsmPrinter() {
2161   RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2162   RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2163 }
2164