1 //===-- NVPTXAsmPrinter.cpp - NVPTX LLVM assembly writer ------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains a printer that converts from our internal representation
11 // of machine-dependent LLVM code to NVPTX assembly language.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "NVPTXAsmPrinter.h"
16 #include "InstPrinter/NVPTXInstPrinter.h"
17 #include "MCTargetDesc/NVPTXMCAsmInfo.h"
18 #include "NVPTX.h"
19 #include "NVPTXInstrInfo.h"
20 #include "NVPTXMCExpr.h"
21 #include "NVPTXMachineFunctionInfo.h"
22 #include "NVPTXRegisterInfo.h"
23 #include "NVPTXTargetMachine.h"
24 #include "NVPTXUtilities.h"
25 #include "cl_common_defines.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/CodeGen/Analysis.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineLoopInfo.h"
31 #include "llvm/CodeGen/MachineModuleInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/IR/DebugInfo.h"
34 #include "llvm/IR/DerivedTypes.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/GlobalVariable.h"
37 #include "llvm/IR/Mangler.h"
38 #include "llvm/IR/Module.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/MC/MCInst.h"
41 #include "llvm/MC/MCStreamer.h"
42 #include "llvm/MC/MCSymbol.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/FormattedStream.h"
46 #include "llvm/Support/Path.h"
47 #include "llvm/Support/TargetRegistry.h"
48 #include "llvm/Support/TimeValue.h"
49 #include "llvm/Target/TargetLoweringObjectFile.h"
50 #include "llvm/Transforms/Utils/UnrollLoop.h"
51 #include <sstream>
52 using namespace llvm;
53
54 #define DEPOTNAME "__local_depot"
55
56 static cl::opt<bool>
57 EmitLineNumbers("nvptx-emit-line-numbers", cl::Hidden,
58 cl::desc("NVPTX Specific: Emit Line numbers even without -G"),
59 cl::init(true));
60
61 static cl::opt<bool>
62 InterleaveSrc("nvptx-emit-src", cl::ZeroOrMore, cl::Hidden,
63 cl::desc("NVPTX Specific: Emit source line in ptx file"),
64 cl::init(false));
65
66 namespace {
67 /// DiscoverDependentGlobals - Return a set of GlobalVariables on which \p V
68 /// depends.
DiscoverDependentGlobals(const Value * V,DenseSet<const GlobalVariable * > & Globals)69 void DiscoverDependentGlobals(const Value *V,
70 DenseSet<const GlobalVariable *> &Globals) {
71 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
72 Globals.insert(GV);
73 else {
74 if (const User *U = dyn_cast<User>(V)) {
75 for (unsigned i = 0, e = U->getNumOperands(); i != e; ++i) {
76 DiscoverDependentGlobals(U->getOperand(i), Globals);
77 }
78 }
79 }
80 }
81
82 /// VisitGlobalVariableForEmission - Add \p GV to the list of GlobalVariable
83 /// instances to be emitted, but only after any dependents have been added
84 /// first.
VisitGlobalVariableForEmission(const GlobalVariable * GV,SmallVectorImpl<const GlobalVariable * > & Order,DenseSet<const GlobalVariable * > & Visited,DenseSet<const GlobalVariable * > & Visiting)85 void VisitGlobalVariableForEmission(
86 const GlobalVariable *GV, SmallVectorImpl<const GlobalVariable *> &Order,
87 DenseSet<const GlobalVariable *> &Visited,
88 DenseSet<const GlobalVariable *> &Visiting) {
89 // Have we already visited this one?
90 if (Visited.count(GV))
91 return;
92
93 // Do we have a circular dependency?
94 if (!Visiting.insert(GV).second)
95 report_fatal_error("Circular dependency found in global variable set");
96
97 // Make sure we visit all dependents first
98 DenseSet<const GlobalVariable *> Others;
99 for (unsigned i = 0, e = GV->getNumOperands(); i != e; ++i)
100 DiscoverDependentGlobals(GV->getOperand(i), Others);
101
102 for (DenseSet<const GlobalVariable *>::iterator I = Others.begin(),
103 E = Others.end();
104 I != E; ++I)
105 VisitGlobalVariableForEmission(*I, Order, Visited, Visiting);
106
107 // Now we can visit ourself
108 Order.push_back(GV);
109 Visited.insert(GV);
110 Visiting.erase(GV);
111 }
112 }
113
emitLineNumberAsDotLoc(const MachineInstr & MI)114 void NVPTXAsmPrinter::emitLineNumberAsDotLoc(const MachineInstr &MI) {
115 if (!EmitLineNumbers)
116 return;
117 if (ignoreLoc(MI))
118 return;
119
120 const DebugLoc &curLoc = MI.getDebugLoc();
121
122 if (!prevDebugLoc && !curLoc)
123 return;
124
125 if (prevDebugLoc == curLoc)
126 return;
127
128 prevDebugLoc = curLoc;
129
130 if (!curLoc)
131 return;
132
133 auto *Scope = cast_or_null<DIScope>(curLoc.getScope());
134 if (!Scope)
135 return;
136
137 StringRef fileName(Scope->getFilename());
138 StringRef dirName(Scope->getDirectory());
139 SmallString<128> FullPathName = dirName;
140 if (!dirName.empty() && !sys::path::is_absolute(fileName)) {
141 sys::path::append(FullPathName, fileName);
142 fileName = FullPathName;
143 }
144
145 if (filenameMap.find(fileName) == filenameMap.end())
146 return;
147
148 // Emit the line from the source file.
149 if (InterleaveSrc)
150 this->emitSrcInText(fileName, curLoc.getLine());
151
152 std::stringstream temp;
153 temp << "\t.loc " << filenameMap[fileName] << " " << curLoc.getLine()
154 << " " << curLoc.getCol();
155 OutStreamer->EmitRawText(temp.str());
156 }
157
EmitInstruction(const MachineInstr * MI)158 void NVPTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
159 SmallString<128> Str;
160 raw_svector_ostream OS(Str);
161 if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == NVPTX::CUDA)
162 emitLineNumberAsDotLoc(*MI);
163
164 MCInst Inst;
165 lowerToMCInst(MI, Inst);
166 EmitToStreamer(*OutStreamer, Inst);
167 }
168
169 // Handle symbol backtracking for targets that do not support image handles
lowerImageHandleOperand(const MachineInstr * MI,unsigned OpNo,MCOperand & MCOp)170 bool NVPTXAsmPrinter::lowerImageHandleOperand(const MachineInstr *MI,
171 unsigned OpNo, MCOperand &MCOp) {
172 const MachineOperand &MO = MI->getOperand(OpNo);
173 const MCInstrDesc &MCID = MI->getDesc();
174
175 if (MCID.TSFlags & NVPTXII::IsTexFlag) {
176 // This is a texture fetch, so operand 4 is a texref and operand 5 is
177 // a samplerref
178 if (OpNo == 4 && MO.isImm()) {
179 lowerImageHandleSymbol(MO.getImm(), MCOp);
180 return true;
181 }
182 if (OpNo == 5 && MO.isImm() && !(MCID.TSFlags & NVPTXII::IsTexModeUnifiedFlag)) {
183 lowerImageHandleSymbol(MO.getImm(), MCOp);
184 return true;
185 }
186
187 return false;
188 } else if (MCID.TSFlags & NVPTXII::IsSuldMask) {
189 unsigned VecSize =
190 1 << (((MCID.TSFlags & NVPTXII::IsSuldMask) >> NVPTXII::IsSuldShift) - 1);
191
192 // For a surface load of vector size N, the Nth operand will be the surfref
193 if (OpNo == VecSize && MO.isImm()) {
194 lowerImageHandleSymbol(MO.getImm(), MCOp);
195 return true;
196 }
197
198 return false;
199 } else if (MCID.TSFlags & NVPTXII::IsSustFlag) {
200 // This is a surface store, so operand 0 is a surfref
201 if (OpNo == 0 && MO.isImm()) {
202 lowerImageHandleSymbol(MO.getImm(), MCOp);
203 return true;
204 }
205
206 return false;
207 } else if (MCID.TSFlags & NVPTXII::IsSurfTexQueryFlag) {
208 // This is a query, so operand 1 is a surfref/texref
209 if (OpNo == 1 && MO.isImm()) {
210 lowerImageHandleSymbol(MO.getImm(), MCOp);
211 return true;
212 }
213
214 return false;
215 }
216
217 return false;
218 }
219
lowerImageHandleSymbol(unsigned Index,MCOperand & MCOp)220 void NVPTXAsmPrinter::lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp) {
221 // Ewwww
222 TargetMachine &TM = const_cast<TargetMachine&>(MF->getTarget());
223 NVPTXTargetMachine &nvTM = static_cast<NVPTXTargetMachine&>(TM);
224 const NVPTXMachineFunctionInfo *MFI = MF->getInfo<NVPTXMachineFunctionInfo>();
225 const char *Sym = MFI->getImageHandleSymbol(Index);
226 std::string *SymNamePtr =
227 nvTM.getManagedStrPool()->getManagedString(Sym);
228 MCOp = GetSymbolRef(OutContext.getOrCreateSymbol(
229 StringRef(SymNamePtr->c_str())));
230 }
231
lowerToMCInst(const MachineInstr * MI,MCInst & OutMI)232 void NVPTXAsmPrinter::lowerToMCInst(const MachineInstr *MI, MCInst &OutMI) {
233 OutMI.setOpcode(MI->getOpcode());
234 // Special: Do not mangle symbol operand of CALL_PROTOTYPE
235 if (MI->getOpcode() == NVPTX::CALL_PROTOTYPE) {
236 const MachineOperand &MO = MI->getOperand(0);
237 OutMI.addOperand(GetSymbolRef(
238 OutContext.getOrCreateSymbol(Twine(MO.getSymbolName()))));
239 return;
240 }
241
242 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
243 const MachineOperand &MO = MI->getOperand(i);
244
245 MCOperand MCOp;
246 if (!nvptxSubtarget->hasImageHandles()) {
247 if (lowerImageHandleOperand(MI, i, MCOp)) {
248 OutMI.addOperand(MCOp);
249 continue;
250 }
251 }
252
253 if (lowerOperand(MO, MCOp))
254 OutMI.addOperand(MCOp);
255 }
256 }
257
lowerOperand(const MachineOperand & MO,MCOperand & MCOp)258 bool NVPTXAsmPrinter::lowerOperand(const MachineOperand &MO,
259 MCOperand &MCOp) {
260 switch (MO.getType()) {
261 default: llvm_unreachable("unknown operand type");
262 case MachineOperand::MO_Register:
263 MCOp = MCOperand::createReg(encodeVirtualRegister(MO.getReg()));
264 break;
265 case MachineOperand::MO_Immediate:
266 MCOp = MCOperand::createImm(MO.getImm());
267 break;
268 case MachineOperand::MO_MachineBasicBlock:
269 MCOp = MCOperand::createExpr(MCSymbolRefExpr::create(
270 MO.getMBB()->getSymbol(), OutContext));
271 break;
272 case MachineOperand::MO_ExternalSymbol:
273 MCOp = GetSymbolRef(GetExternalSymbolSymbol(MO.getSymbolName()));
274 break;
275 case MachineOperand::MO_GlobalAddress:
276 MCOp = GetSymbolRef(getSymbol(MO.getGlobal()));
277 break;
278 case MachineOperand::MO_FPImmediate: {
279 const ConstantFP *Cnt = MO.getFPImm();
280 const APFloat &Val = Cnt->getValueAPF();
281
282 switch (Cnt->getType()->getTypeID()) {
283 default: report_fatal_error("Unsupported FP type"); break;
284 case Type::FloatTyID:
285 MCOp = MCOperand::createExpr(
286 NVPTXFloatMCExpr::createConstantFPSingle(Val, OutContext));
287 break;
288 case Type::DoubleTyID:
289 MCOp = MCOperand::createExpr(
290 NVPTXFloatMCExpr::createConstantFPDouble(Val, OutContext));
291 break;
292 }
293 break;
294 }
295 }
296 return true;
297 }
298
encodeVirtualRegister(unsigned Reg)299 unsigned NVPTXAsmPrinter::encodeVirtualRegister(unsigned Reg) {
300 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
301 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
302
303 DenseMap<unsigned, unsigned> &RegMap = VRegMapping[RC];
304 unsigned RegNum = RegMap[Reg];
305
306 // Encode the register class in the upper 4 bits
307 // Must be kept in sync with NVPTXInstPrinter::printRegName
308 unsigned Ret = 0;
309 if (RC == &NVPTX::Int1RegsRegClass) {
310 Ret = (1 << 28);
311 } else if (RC == &NVPTX::Int16RegsRegClass) {
312 Ret = (2 << 28);
313 } else if (RC == &NVPTX::Int32RegsRegClass) {
314 Ret = (3 << 28);
315 } else if (RC == &NVPTX::Int64RegsRegClass) {
316 Ret = (4 << 28);
317 } else if (RC == &NVPTX::Float32RegsRegClass) {
318 Ret = (5 << 28);
319 } else if (RC == &NVPTX::Float64RegsRegClass) {
320 Ret = (6 << 28);
321 } else {
322 report_fatal_error("Bad register class");
323 }
324
325 // Insert the vreg number
326 Ret |= (RegNum & 0x0FFFFFFF);
327 return Ret;
328 } else {
329 // Some special-use registers are actually physical registers.
330 // Encode this as the register class ID of 0 and the real register ID.
331 return Reg & 0x0FFFFFFF;
332 }
333 }
334
GetSymbolRef(const MCSymbol * Symbol)335 MCOperand NVPTXAsmPrinter::GetSymbolRef(const MCSymbol *Symbol) {
336 const MCExpr *Expr;
337 Expr = MCSymbolRefExpr::create(Symbol, MCSymbolRefExpr::VK_None,
338 OutContext);
339 return MCOperand::createExpr(Expr);
340 }
341
printReturnValStr(const Function * F,raw_ostream & O)342 void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
343 const DataLayout &DL = getDataLayout();
344 const TargetLowering *TLI = nvptxSubtarget->getTargetLowering();
345
346 Type *Ty = F->getReturnType();
347
348 bool isABI = (nvptxSubtarget->getSmVersion() >= 20);
349
350 if (Ty->getTypeID() == Type::VoidTyID)
351 return;
352
353 O << " (";
354
355 if (isABI) {
356 if (Ty->isFloatingPointTy() || Ty->isIntegerTy()) {
357 unsigned size = 0;
358 if (auto *ITy = dyn_cast<IntegerType>(Ty)) {
359 size = ITy->getBitWidth();
360 if (size < 32)
361 size = 32;
362 } else {
363 assert(Ty->isFloatingPointTy() && "Floating point type expected here");
364 size = Ty->getPrimitiveSizeInBits();
365 }
366
367 O << ".param .b" << size << " func_retval0";
368 } else if (isa<PointerType>(Ty)) {
369 O << ".param .b" << TLI->getPointerTy(DL).getSizeInBits()
370 << " func_retval0";
371 } else if ((Ty->getTypeID() == Type::StructTyID) || isa<VectorType>(Ty)) {
372 unsigned totalsz = DL.getTypeAllocSize(Ty);
373 unsigned retAlignment = 0;
374 if (!llvm::getAlign(*F, 0, retAlignment))
375 retAlignment = DL.getABITypeAlignment(Ty);
376 O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
377 << "]";
378 } else
379 llvm_unreachable("Unknown return type");
380 } else {
381 SmallVector<EVT, 16> vtparts;
382 ComputeValueVTs(*TLI, DL, Ty, vtparts);
383 unsigned idx = 0;
384 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
385 unsigned elems = 1;
386 EVT elemtype = vtparts[i];
387 if (vtparts[i].isVector()) {
388 elems = vtparts[i].getVectorNumElements();
389 elemtype = vtparts[i].getVectorElementType();
390 }
391
392 for (unsigned j = 0, je = elems; j != je; ++j) {
393 unsigned sz = elemtype.getSizeInBits();
394 if (elemtype.isInteger() && (sz < 32))
395 sz = 32;
396 O << ".reg .b" << sz << " func_retval" << idx;
397 if (j < je - 1)
398 O << ", ";
399 ++idx;
400 }
401 if (i < e - 1)
402 O << ", ";
403 }
404 }
405 O << ") ";
406 return;
407 }
408
printReturnValStr(const MachineFunction & MF,raw_ostream & O)409 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
410 raw_ostream &O) {
411 const Function *F = MF.getFunction();
412 printReturnValStr(F, O);
413 }
414
415 // Return true if MBB is the header of a loop marked with
416 // llvm.loop.unroll.disable.
417 // TODO: consider "#pragma unroll 1" which is equivalent to "#pragma nounroll".
isLoopHeaderOfNoUnroll(const MachineBasicBlock & MBB) const418 bool NVPTXAsmPrinter::isLoopHeaderOfNoUnroll(
419 const MachineBasicBlock &MBB) const {
420 MachineLoopInfo &LI = getAnalysis<MachineLoopInfo>();
421 // We insert .pragma "nounroll" only to the loop header.
422 if (!LI.isLoopHeader(&MBB))
423 return false;
424
425 // llvm.loop.unroll.disable is marked on the back edges of a loop. Therefore,
426 // we iterate through each back edge of the loop with header MBB, and check
427 // whether its metadata contains llvm.loop.unroll.disable.
428 for (auto I = MBB.pred_begin(); I != MBB.pred_end(); ++I) {
429 const MachineBasicBlock *PMBB = *I;
430 if (LI.getLoopFor(PMBB) != LI.getLoopFor(&MBB)) {
431 // Edges from other loops to MBB are not back edges.
432 continue;
433 }
434 if (const BasicBlock *PBB = PMBB->getBasicBlock()) {
435 if (MDNode *LoopID =
436 PBB->getTerminator()->getMetadata(LLVMContext::MD_loop)) {
437 if (GetUnrollMetadata(LoopID, "llvm.loop.unroll.disable"))
438 return true;
439 }
440 }
441 }
442 return false;
443 }
444
EmitBasicBlockStart(const MachineBasicBlock & MBB) const445 void NVPTXAsmPrinter::EmitBasicBlockStart(const MachineBasicBlock &MBB) const {
446 AsmPrinter::EmitBasicBlockStart(MBB);
447 if (isLoopHeaderOfNoUnroll(MBB))
448 OutStreamer->EmitRawText(StringRef("\t.pragma \"nounroll\";\n"));
449 }
450
EmitFunctionEntryLabel()451 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
452 SmallString<128> Str;
453 raw_svector_ostream O(Str);
454
455 if (!GlobalsEmitted) {
456 emitGlobals(*MF->getFunction()->getParent());
457 GlobalsEmitted = true;
458 }
459
460 // Set up
461 MRI = &MF->getRegInfo();
462 F = MF->getFunction();
463 emitLinkageDirective(F, O);
464 if (llvm::isKernelFunction(*F))
465 O << ".entry ";
466 else {
467 O << ".func ";
468 printReturnValStr(*MF, O);
469 }
470
471 CurrentFnSym->print(O, MAI);
472
473 emitFunctionParamList(*MF, O);
474
475 if (llvm::isKernelFunction(*F))
476 emitKernelFunctionDirectives(*F, O);
477
478 OutStreamer->EmitRawText(O.str());
479
480 prevDebugLoc = DebugLoc();
481 }
482
EmitFunctionBodyStart()483 void NVPTXAsmPrinter::EmitFunctionBodyStart() {
484 VRegMapping.clear();
485 OutStreamer->EmitRawText(StringRef("{\n"));
486 setAndEmitFunctionVirtualRegisters(*MF);
487
488 SmallString<128> Str;
489 raw_svector_ostream O(Str);
490 emitDemotedVars(MF->getFunction(), O);
491 OutStreamer->EmitRawText(O.str());
492 }
493
EmitFunctionBodyEnd()494 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
495 OutStreamer->EmitRawText(StringRef("}\n"));
496 VRegMapping.clear();
497 }
498
emitImplicitDef(const MachineInstr * MI) const499 void NVPTXAsmPrinter::emitImplicitDef(const MachineInstr *MI) const {
500 unsigned RegNo = MI->getOperand(0).getReg();
501 if (TargetRegisterInfo::isVirtualRegister(RegNo)) {
502 OutStreamer->AddComment(Twine("implicit-def: ") +
503 getVirtualRegisterName(RegNo));
504 } else {
505 OutStreamer->AddComment(Twine("implicit-def: ") +
506 nvptxSubtarget->getRegisterInfo()->getName(RegNo));
507 }
508 OutStreamer->AddBlankLine();
509 }
510
emitKernelFunctionDirectives(const Function & F,raw_ostream & O) const511 void NVPTXAsmPrinter::emitKernelFunctionDirectives(const Function &F,
512 raw_ostream &O) const {
513 // If the NVVM IR has some of reqntid* specified, then output
514 // the reqntid directive, and set the unspecified ones to 1.
515 // If none of reqntid* is specified, don't output reqntid directive.
516 unsigned reqntidx, reqntidy, reqntidz;
517 bool specified = false;
518 if (!llvm::getReqNTIDx(F, reqntidx))
519 reqntidx = 1;
520 else
521 specified = true;
522 if (!llvm::getReqNTIDy(F, reqntidy))
523 reqntidy = 1;
524 else
525 specified = true;
526 if (!llvm::getReqNTIDz(F, reqntidz))
527 reqntidz = 1;
528 else
529 specified = true;
530
531 if (specified)
532 O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
533 << "\n";
534
535 // If the NVVM IR has some of maxntid* specified, then output
536 // the maxntid directive, and set the unspecified ones to 1.
537 // If none of maxntid* is specified, don't output maxntid directive.
538 unsigned maxntidx, maxntidy, maxntidz;
539 specified = false;
540 if (!llvm::getMaxNTIDx(F, maxntidx))
541 maxntidx = 1;
542 else
543 specified = true;
544 if (!llvm::getMaxNTIDy(F, maxntidy))
545 maxntidy = 1;
546 else
547 specified = true;
548 if (!llvm::getMaxNTIDz(F, maxntidz))
549 maxntidz = 1;
550 else
551 specified = true;
552
553 if (specified)
554 O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
555 << "\n";
556
557 unsigned mincta;
558 if (llvm::getMinCTASm(F, mincta))
559 O << ".minnctapersm " << mincta << "\n";
560 }
561
562 std::string
getVirtualRegisterName(unsigned Reg) const563 NVPTXAsmPrinter::getVirtualRegisterName(unsigned Reg) const {
564 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
565
566 std::string Name;
567 raw_string_ostream NameStr(Name);
568
569 VRegRCMap::const_iterator I = VRegMapping.find(RC);
570 assert(I != VRegMapping.end() && "Bad register class");
571 const DenseMap<unsigned, unsigned> &RegMap = I->second;
572
573 VRegMap::const_iterator VI = RegMap.find(Reg);
574 assert(VI != RegMap.end() && "Bad virtual register");
575 unsigned MappedVR = VI->second;
576
577 NameStr << getNVPTXRegClassStr(RC) << MappedVR;
578
579 NameStr.flush();
580 return Name;
581 }
582
emitVirtualRegister(unsigned int vr,raw_ostream & O)583 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr,
584 raw_ostream &O) {
585 O << getVirtualRegisterName(vr);
586 }
587
printVecModifiedImmediate(const MachineOperand & MO,const char * Modifier,raw_ostream & O)588 void NVPTXAsmPrinter::printVecModifiedImmediate(
589 const MachineOperand &MO, const char *Modifier, raw_ostream &O) {
590 static const char vecelem[] = { '0', '1', '2', '3', '0', '1', '2', '3' };
591 int Imm = (int) MO.getImm();
592 if (0 == strcmp(Modifier, "vecelem"))
593 O << "_" << vecelem[Imm];
594 else if (0 == strcmp(Modifier, "vecv4comm1")) {
595 if ((Imm < 0) || (Imm > 3))
596 O << "//";
597 } else if (0 == strcmp(Modifier, "vecv4comm2")) {
598 if ((Imm < 4) || (Imm > 7))
599 O << "//";
600 } else if (0 == strcmp(Modifier, "vecv4pos")) {
601 if (Imm < 0)
602 Imm = 0;
603 O << "_" << vecelem[Imm % 4];
604 } else if (0 == strcmp(Modifier, "vecv2comm1")) {
605 if ((Imm < 0) || (Imm > 1))
606 O << "//";
607 } else if (0 == strcmp(Modifier, "vecv2comm2")) {
608 if ((Imm < 2) || (Imm > 3))
609 O << "//";
610 } else if (0 == strcmp(Modifier, "vecv2pos")) {
611 if (Imm < 0)
612 Imm = 0;
613 O << "_" << vecelem[Imm % 2];
614 } else
615 llvm_unreachable("Unknown Modifier on immediate operand");
616 }
617
618
619
emitDeclaration(const Function * F,raw_ostream & O)620 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
621
622 emitLinkageDirective(F, O);
623 if (llvm::isKernelFunction(*F))
624 O << ".entry ";
625 else
626 O << ".func ";
627 printReturnValStr(F, O);
628 getSymbol(F)->print(O, MAI);
629 O << "\n";
630 emitFunctionParamList(F, O);
631 O << ";\n";
632 }
633
usedInGlobalVarDef(const Constant * C)634 static bool usedInGlobalVarDef(const Constant *C) {
635 if (!C)
636 return false;
637
638 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
639 return GV->getName() != "llvm.used";
640 }
641
642 for (const User *U : C->users())
643 if (const Constant *C = dyn_cast<Constant>(U))
644 if (usedInGlobalVarDef(C))
645 return true;
646
647 return false;
648 }
649
usedInOneFunc(const User * U,Function const * & oneFunc)650 static bool usedInOneFunc(const User *U, Function const *&oneFunc) {
651 if (const GlobalVariable *othergv = dyn_cast<GlobalVariable>(U)) {
652 if (othergv->getName() == "llvm.used")
653 return true;
654 }
655
656 if (const Instruction *instr = dyn_cast<Instruction>(U)) {
657 if (instr->getParent() && instr->getParent()->getParent()) {
658 const Function *curFunc = instr->getParent()->getParent();
659 if (oneFunc && (curFunc != oneFunc))
660 return false;
661 oneFunc = curFunc;
662 return true;
663 } else
664 return false;
665 }
666
667 for (const User *UU : U->users())
668 if (!usedInOneFunc(UU, oneFunc))
669 return false;
670
671 return true;
672 }
673
674 /* Find out if a global variable can be demoted to local scope.
675 * Currently, this is valid for CUDA shared variables, which have local
676 * scope and global lifetime. So the conditions to check are :
677 * 1. Is the global variable in shared address space?
678 * 2. Does it have internal linkage?
679 * 3. Is the global variable referenced only in one function?
680 */
canDemoteGlobalVar(const GlobalVariable * gv,Function const * & f)681 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
682 if (!gv->hasInternalLinkage())
683 return false;
684 PointerType *Pty = gv->getType();
685 if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
686 return false;
687
688 const Function *oneFunc = nullptr;
689
690 bool flag = usedInOneFunc(gv, oneFunc);
691 if (!flag)
692 return false;
693 if (!oneFunc)
694 return false;
695 f = oneFunc;
696 return true;
697 }
698
useFuncSeen(const Constant * C,llvm::DenseMap<const Function *,bool> & seenMap)699 static bool useFuncSeen(const Constant *C,
700 llvm::DenseMap<const Function *, bool> &seenMap) {
701 for (const User *U : C->users()) {
702 if (const Constant *cu = dyn_cast<Constant>(U)) {
703 if (useFuncSeen(cu, seenMap))
704 return true;
705 } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
706 const BasicBlock *bb = I->getParent();
707 if (!bb)
708 continue;
709 const Function *caller = bb->getParent();
710 if (!caller)
711 continue;
712 if (seenMap.find(caller) != seenMap.end())
713 return true;
714 }
715 }
716 return false;
717 }
718
emitDeclarations(const Module & M,raw_ostream & O)719 void NVPTXAsmPrinter::emitDeclarations(const Module &M, raw_ostream &O) {
720 llvm::DenseMap<const Function *, bool> seenMap;
721 for (Module::const_iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
722 const Function *F = &*FI;
723
724 if (F->isDeclaration()) {
725 if (F->use_empty())
726 continue;
727 if (F->getIntrinsicID())
728 continue;
729 emitDeclaration(F, O);
730 continue;
731 }
732 for (const User *U : F->users()) {
733 if (const Constant *C = dyn_cast<Constant>(U)) {
734 if (usedInGlobalVarDef(C)) {
735 // The use is in the initialization of a global variable
736 // that is a function pointer, so print a declaration
737 // for the original function
738 emitDeclaration(F, O);
739 break;
740 }
741 // Emit a declaration of this function if the function that
742 // uses this constant expr has already been seen.
743 if (useFuncSeen(C, seenMap)) {
744 emitDeclaration(F, O);
745 break;
746 }
747 }
748
749 if (!isa<Instruction>(U))
750 continue;
751 const Instruction *instr = cast<Instruction>(U);
752 const BasicBlock *bb = instr->getParent();
753 if (!bb)
754 continue;
755 const Function *caller = bb->getParent();
756 if (!caller)
757 continue;
758
759 // If a caller has already been seen, then the caller is
760 // appearing in the module before the callee. so print out
761 // a declaration for the callee.
762 if (seenMap.find(caller) != seenMap.end()) {
763 emitDeclaration(F, O);
764 break;
765 }
766 }
767 seenMap[F] = true;
768 }
769 }
770
recordAndEmitFilenames(Module & M)771 void NVPTXAsmPrinter::recordAndEmitFilenames(Module &M) {
772 DebugInfoFinder DbgFinder;
773 DbgFinder.processModule(M);
774
775 unsigned i = 1;
776 for (const DICompileUnit *DIUnit : DbgFinder.compile_units()) {
777 StringRef Filename = DIUnit->getFilename();
778 StringRef Dirname = DIUnit->getDirectory();
779 SmallString<128> FullPathName = Dirname;
780 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
781 sys::path::append(FullPathName, Filename);
782 Filename = FullPathName;
783 }
784 if (filenameMap.find(Filename) != filenameMap.end())
785 continue;
786 filenameMap[Filename] = i;
787 OutStreamer->EmitDwarfFileDirective(i, "", Filename);
788 ++i;
789 }
790
791 for (DISubprogram *SP : DbgFinder.subprograms()) {
792 StringRef Filename = SP->getFilename();
793 StringRef Dirname = SP->getDirectory();
794 SmallString<128> FullPathName = Dirname;
795 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
796 sys::path::append(FullPathName, Filename);
797 Filename = FullPathName;
798 }
799 if (filenameMap.find(Filename) != filenameMap.end())
800 continue;
801 filenameMap[Filename] = i;
802 OutStreamer->EmitDwarfFileDirective(i, "", Filename);
803 ++i;
804 }
805 }
806
isEmptyXXStructor(GlobalVariable * GV)807 static bool isEmptyXXStructor(GlobalVariable *GV) {
808 if (!GV) return true;
809 const ConstantArray *InitList = dyn_cast<ConstantArray>(GV->getInitializer());
810 if (!InitList) return true; // Not an array; we don't know how to parse.
811 return InitList->getNumOperands() == 0;
812 }
813
doInitialization(Module & M)814 bool NVPTXAsmPrinter::doInitialization(Module &M) {
815 // Construct a default subtarget off of the TargetMachine defaults. The
816 // rest of NVPTX isn't friendly to change subtargets per function and
817 // so the default TargetMachine will have all of the options.
818 const Triple &TT = TM.getTargetTriple();
819 StringRef CPU = TM.getTargetCPU();
820 StringRef FS = TM.getTargetFeatureString();
821 const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
822 const NVPTXSubtarget STI(TT, CPU, FS, NTM);
823
824 if (M.alias_size()) {
825 report_fatal_error("Module has aliases, which NVPTX does not support.");
826 return true; // error
827 }
828 if (!isEmptyXXStructor(M.getNamedGlobal("llvm.global_ctors"))) {
829 report_fatal_error(
830 "Module has a nontrivial global ctor, which NVPTX does not support.");
831 return true; // error
832 }
833 if (!isEmptyXXStructor(M.getNamedGlobal("llvm.global_dtors"))) {
834 report_fatal_error(
835 "Module has a nontrivial global dtor, which NVPTX does not support.");
836 return true; // error
837 }
838
839 SmallString<128> Str1;
840 raw_svector_ostream OS1(Str1);
841
842 MMI = getAnalysisIfAvailable<MachineModuleInfo>();
843
844 // We need to call the parent's one explicitly.
845 //bool Result = AsmPrinter::doInitialization(M);
846
847 // Initialize TargetLoweringObjectFile.
848 const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
849 .Initialize(OutContext, TM);
850
851 Mang = new Mangler();
852
853 // Emit header before any dwarf directives are emitted below.
854 emitHeader(M, OS1, STI);
855 OutStreamer->EmitRawText(OS1.str());
856
857 // Already commented out
858 //bool Result = AsmPrinter::doInitialization(M);
859
860 // Emit module-level inline asm if it exists.
861 if (!M.getModuleInlineAsm().empty()) {
862 OutStreamer->AddComment("Start of file scope inline assembly");
863 OutStreamer->AddBlankLine();
864 OutStreamer->EmitRawText(StringRef(M.getModuleInlineAsm()));
865 OutStreamer->AddBlankLine();
866 OutStreamer->AddComment("End of file scope inline assembly");
867 OutStreamer->AddBlankLine();
868 }
869
870 // If we're not NVCL we're CUDA, go ahead and emit filenames.
871 if (TM.getTargetTriple().getOS() != Triple::NVCL)
872 recordAndEmitFilenames(M);
873
874 GlobalsEmitted = false;
875
876 return false; // success
877 }
878
emitGlobals(const Module & M)879 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
880 SmallString<128> Str2;
881 raw_svector_ostream OS2(Str2);
882
883 emitDeclarations(M, OS2);
884
885 // As ptxas does not support forward references of globals, we need to first
886 // sort the list of module-level globals in def-use order. We visit each
887 // global variable in order, and ensure that we emit it *after* its dependent
888 // globals. We use a little extra memory maintaining both a set and a list to
889 // have fast searches while maintaining a strict ordering.
890 SmallVector<const GlobalVariable *, 8> Globals;
891 DenseSet<const GlobalVariable *> GVVisited;
892 DenseSet<const GlobalVariable *> GVVisiting;
893
894 // Visit each global variable, in order
895 for (const GlobalVariable &I : M.globals())
896 VisitGlobalVariableForEmission(&I, Globals, GVVisited, GVVisiting);
897
898 assert(GVVisited.size() == M.getGlobalList().size() &&
899 "Missed a global variable");
900 assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
901
902 // Print out module-level global variables in proper order
903 for (unsigned i = 0, e = Globals.size(); i != e; ++i)
904 printModuleLevelGV(Globals[i], OS2);
905
906 OS2 << '\n';
907
908 OutStreamer->EmitRawText(OS2.str());
909 }
910
emitHeader(Module & M,raw_ostream & O,const NVPTXSubtarget & STI)911 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O,
912 const NVPTXSubtarget &STI) {
913 O << "//\n";
914 O << "// Generated by LLVM NVPTX Back-End\n";
915 O << "//\n";
916 O << "\n";
917
918 unsigned PTXVersion = STI.getPTXVersion();
919 O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
920
921 O << ".target ";
922 O << STI.getTargetName();
923
924 const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
925 if (NTM.getDrvInterface() == NVPTX::NVCL)
926 O << ", texmode_independent";
927 else {
928 if (!STI.hasDouble())
929 O << ", map_f64_to_f32";
930 }
931
932 if (MAI->doesSupportDebugInformation())
933 O << ", debug";
934
935 O << "\n";
936
937 O << ".address_size ";
938 if (NTM.is64Bit())
939 O << "64";
940 else
941 O << "32";
942 O << "\n";
943
944 O << "\n";
945 }
946
doFinalization(Module & M)947 bool NVPTXAsmPrinter::doFinalization(Module &M) {
948 // If we did not emit any functions, then the global declarations have not
949 // yet been emitted.
950 if (!GlobalsEmitted) {
951 emitGlobals(M);
952 GlobalsEmitted = true;
953 }
954
955 // XXX Temproarily remove global variables so that doFinalization() will not
956 // emit them again (global variables are emitted at beginning).
957
958 Module::GlobalListType &global_list = M.getGlobalList();
959 int i, n = global_list.size();
960 GlobalVariable **gv_array = new GlobalVariable *[n];
961
962 // first, back-up GlobalVariable in gv_array
963 i = 0;
964 for (Module::global_iterator I = global_list.begin(), E = global_list.end();
965 I != E; ++I)
966 gv_array[i++] = &*I;
967
968 // second, empty global_list
969 while (!global_list.empty())
970 global_list.remove(global_list.begin());
971
972 // call doFinalization
973 bool ret = AsmPrinter::doFinalization(M);
974
975 // now we restore global variables
976 for (i = 0; i < n; i++)
977 global_list.insert(global_list.end(), gv_array[i]);
978
979 clearAnnotationCache(&M);
980
981 delete[] gv_array;
982 return ret;
983
984 //bool Result = AsmPrinter::doFinalization(M);
985 // Instead of calling the parents doFinalization, we may
986 // clone parents doFinalization and customize here.
987 // Currently, we if NVISA out the EmitGlobals() in
988 // parent's doFinalization, which is too intrusive.
989 //
990 // Same for the doInitialization.
991 //return Result;
992 }
993
994 // This function emits appropriate linkage directives for
995 // functions and global variables.
996 //
997 // extern function declaration -> .extern
998 // extern function definition -> .visible
999 // external global variable with init -> .visible
1000 // external without init -> .extern
1001 // appending -> not allowed, assert.
1002 // for any linkage other than
1003 // internal, private, linker_private,
1004 // linker_private_weak, linker_private_weak_def_auto,
1005 // we emit -> .weak.
1006
emitLinkageDirective(const GlobalValue * V,raw_ostream & O)1007 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
1008 raw_ostream &O) {
1009 if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == NVPTX::CUDA) {
1010 if (V->hasExternalLinkage()) {
1011 if (isa<GlobalVariable>(V)) {
1012 const GlobalVariable *GVar = cast<GlobalVariable>(V);
1013 if (GVar) {
1014 if (GVar->hasInitializer())
1015 O << ".visible ";
1016 else
1017 O << ".extern ";
1018 }
1019 } else if (V->isDeclaration())
1020 O << ".extern ";
1021 else
1022 O << ".visible ";
1023 } else if (V->hasAppendingLinkage()) {
1024 std::string msg;
1025 msg.append("Error: ");
1026 msg.append("Symbol ");
1027 if (V->hasName())
1028 msg.append(V->getName());
1029 msg.append("has unsupported appending linkage type");
1030 llvm_unreachable(msg.c_str());
1031 } else if (!V->hasInternalLinkage() &&
1032 !V->hasPrivateLinkage()) {
1033 O << ".weak ";
1034 }
1035 }
1036 }
1037
printModuleLevelGV(const GlobalVariable * GVar,raw_ostream & O,bool processDemoted)1038 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
1039 raw_ostream &O,
1040 bool processDemoted) {
1041
1042 // Skip meta data
1043 if (GVar->hasSection()) {
1044 if (GVar->getSection() == "llvm.metadata")
1045 return;
1046 }
1047
1048 // Skip LLVM intrinsic global variables
1049 if (GVar->getName().startswith("llvm.") ||
1050 GVar->getName().startswith("nvvm."))
1051 return;
1052
1053 const DataLayout &DL = getDataLayout();
1054
1055 // GlobalVariables are always constant pointers themselves.
1056 PointerType *PTy = GVar->getType();
1057 Type *ETy = GVar->getValueType();
1058
1059 if (GVar->hasExternalLinkage()) {
1060 if (GVar->hasInitializer())
1061 O << ".visible ";
1062 else
1063 O << ".extern ";
1064 } else if (GVar->hasLinkOnceLinkage() || GVar->hasWeakLinkage() ||
1065 GVar->hasAvailableExternallyLinkage() ||
1066 GVar->hasCommonLinkage()) {
1067 O << ".weak ";
1068 }
1069
1070 if (llvm::isTexture(*GVar)) {
1071 O << ".global .texref " << llvm::getTextureName(*GVar) << ";\n";
1072 return;
1073 }
1074
1075 if (llvm::isSurface(*GVar)) {
1076 O << ".global .surfref " << llvm::getSurfaceName(*GVar) << ";\n";
1077 return;
1078 }
1079
1080 if (GVar->isDeclaration()) {
1081 // (extern) declarations, no definition or initializer
1082 // Currently the only known declaration is for an automatic __local
1083 // (.shared) promoted to global.
1084 emitPTXGlobalVariable(GVar, O);
1085 O << ";\n";
1086 return;
1087 }
1088
1089 if (llvm::isSampler(*GVar)) {
1090 O << ".global .samplerref " << llvm::getSamplerName(*GVar);
1091
1092 const Constant *Initializer = nullptr;
1093 if (GVar->hasInitializer())
1094 Initializer = GVar->getInitializer();
1095 const ConstantInt *CI = nullptr;
1096 if (Initializer)
1097 CI = dyn_cast<ConstantInt>(Initializer);
1098 if (CI) {
1099 unsigned sample = CI->getZExtValue();
1100
1101 O << " = { ";
1102
1103 for (int i = 0,
1104 addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1105 i < 3; i++) {
1106 O << "addr_mode_" << i << " = ";
1107 switch (addr) {
1108 case 0:
1109 O << "wrap";
1110 break;
1111 case 1:
1112 O << "clamp_to_border";
1113 break;
1114 case 2:
1115 O << "clamp_to_edge";
1116 break;
1117 case 3:
1118 O << "wrap";
1119 break;
1120 case 4:
1121 O << "mirror";
1122 break;
1123 }
1124 O << ", ";
1125 }
1126 O << "filter_mode = ";
1127 switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1128 case 0:
1129 O << "nearest";
1130 break;
1131 case 1:
1132 O << "linear";
1133 break;
1134 case 2:
1135 llvm_unreachable("Anisotropic filtering is not supported");
1136 default:
1137 O << "nearest";
1138 break;
1139 }
1140 if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1141 O << ", force_unnormalized_coords = 1";
1142 }
1143 O << " }";
1144 }
1145
1146 O << ";\n";
1147 return;
1148 }
1149
1150 if (GVar->hasPrivateLinkage()) {
1151
1152 if (!strncmp(GVar->getName().data(), "unrollpragma", 12))
1153 return;
1154
1155 // FIXME - need better way (e.g. Metadata) to avoid generating this global
1156 if (!strncmp(GVar->getName().data(), "filename", 8))
1157 return;
1158 if (GVar->use_empty())
1159 return;
1160 }
1161
1162 const Function *demotedFunc = nullptr;
1163 if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
1164 O << "// " << GVar->getName() << " has been demoted\n";
1165 if (localDecls.find(demotedFunc) != localDecls.end())
1166 localDecls[demotedFunc].push_back(GVar);
1167 else {
1168 std::vector<const GlobalVariable *> temp;
1169 temp.push_back(GVar);
1170 localDecls[demotedFunc] = temp;
1171 }
1172 return;
1173 }
1174
1175 O << ".";
1176 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1177
1178 if (isManaged(*GVar)) {
1179 O << " .attribute(.managed)";
1180 }
1181
1182 if (GVar->getAlignment() == 0)
1183 O << " .align " << (int)DL.getPrefTypeAlignment(ETy);
1184 else
1185 O << " .align " << GVar->getAlignment();
1186
1187 if (ETy->isFloatingPointTy() || ETy->isIntegerTy() || ETy->isPointerTy()) {
1188 O << " .";
1189 // Special case: ABI requires that we use .u8 for predicates
1190 if (ETy->isIntegerTy(1))
1191 O << "u8";
1192 else
1193 O << getPTXFundamentalTypeStr(ETy, false);
1194 O << " ";
1195 getSymbol(GVar)->print(O, MAI);
1196
1197 // Ptx allows variable initilization only for constant and global state
1198 // spaces.
1199 if (GVar->hasInitializer()) {
1200 if ((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1201 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) {
1202 const Constant *Initializer = GVar->getInitializer();
1203 // 'undef' is treated as there is no value specified.
1204 if (!Initializer->isNullValue() && !isa<UndefValue>(Initializer)) {
1205 O << " = ";
1206 printScalarConstant(Initializer, O);
1207 }
1208 } else {
1209 // The frontend adds zero-initializer to device and constant variables
1210 // that don't have an initial value, and UndefValue to shared
1211 // variables, so skip warning for this case.
1212 if (!GVar->getInitializer()->isNullValue() &&
1213 !isa<UndefValue>(GVar->getInitializer())) {
1214 report_fatal_error("initial value of '" + GVar->getName() +
1215 "' is not allowed in addrspace(" +
1216 Twine(PTy->getAddressSpace()) + ")");
1217 }
1218 }
1219 }
1220 } else {
1221 unsigned int ElementSize = 0;
1222
1223 // Although PTX has direct support for struct type and array type and
1224 // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1225 // targets that support these high level field accesses. Structs, arrays
1226 // and vectors are lowered into arrays of bytes.
1227 switch (ETy->getTypeID()) {
1228 case Type::StructTyID:
1229 case Type::ArrayTyID:
1230 case Type::VectorTyID:
1231 ElementSize = DL.getTypeStoreSize(ETy);
1232 // Ptx allows variable initilization only for constant and
1233 // global state spaces.
1234 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1235 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1236 GVar->hasInitializer()) {
1237 const Constant *Initializer = GVar->getInitializer();
1238 if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1239 AggBuffer aggBuffer(ElementSize, O, *this);
1240 bufferAggregateConstant(Initializer, &aggBuffer);
1241 if (aggBuffer.numSymbols) {
1242 if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit()) {
1243 O << " .u64 ";
1244 getSymbol(GVar)->print(O, MAI);
1245 O << "[";
1246 O << ElementSize / 8;
1247 } else {
1248 O << " .u32 ";
1249 getSymbol(GVar)->print(O, MAI);
1250 O << "[";
1251 O << ElementSize / 4;
1252 }
1253 O << "]";
1254 } else {
1255 O << " .b8 ";
1256 getSymbol(GVar)->print(O, MAI);
1257 O << "[";
1258 O << ElementSize;
1259 O << "]";
1260 }
1261 O << " = {";
1262 aggBuffer.print();
1263 O << "}";
1264 } else {
1265 O << " .b8 ";
1266 getSymbol(GVar)->print(O, MAI);
1267 if (ElementSize) {
1268 O << "[";
1269 O << ElementSize;
1270 O << "]";
1271 }
1272 }
1273 } else {
1274 O << " .b8 ";
1275 getSymbol(GVar)->print(O, MAI);
1276 if (ElementSize) {
1277 O << "[";
1278 O << ElementSize;
1279 O << "]";
1280 }
1281 }
1282 break;
1283 default:
1284 llvm_unreachable("type not supported yet");
1285 }
1286
1287 }
1288 O << ";\n";
1289 }
1290
emitDemotedVars(const Function * f,raw_ostream & O)1291 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1292 if (localDecls.find(f) == localDecls.end())
1293 return;
1294
1295 std::vector<const GlobalVariable *> &gvars = localDecls[f];
1296
1297 for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1298 O << "\t// demoted variable\n\t";
1299 printModuleLevelGV(gvars[i], O, true);
1300 }
1301 }
1302
emitPTXAddressSpace(unsigned int AddressSpace,raw_ostream & O) const1303 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1304 raw_ostream &O) const {
1305 switch (AddressSpace) {
1306 case llvm::ADDRESS_SPACE_LOCAL:
1307 O << "local";
1308 break;
1309 case llvm::ADDRESS_SPACE_GLOBAL:
1310 O << "global";
1311 break;
1312 case llvm::ADDRESS_SPACE_CONST:
1313 O << "const";
1314 break;
1315 case llvm::ADDRESS_SPACE_SHARED:
1316 O << "shared";
1317 break;
1318 default:
1319 report_fatal_error("Bad address space found while emitting PTX");
1320 break;
1321 }
1322 }
1323
1324 std::string
getPTXFundamentalTypeStr(Type * Ty,bool useB4PTR) const1325 NVPTXAsmPrinter::getPTXFundamentalTypeStr(Type *Ty, bool useB4PTR) const {
1326 switch (Ty->getTypeID()) {
1327 default:
1328 llvm_unreachable("unexpected type");
1329 break;
1330 case Type::IntegerTyID: {
1331 unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1332 if (NumBits == 1)
1333 return "pred";
1334 else if (NumBits <= 64) {
1335 std::string name = "u";
1336 return name + utostr(NumBits);
1337 } else {
1338 llvm_unreachable("Integer too large");
1339 break;
1340 }
1341 break;
1342 }
1343 case Type::FloatTyID:
1344 return "f32";
1345 case Type::DoubleTyID:
1346 return "f64";
1347 case Type::PointerTyID:
1348 if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit())
1349 if (useB4PTR)
1350 return "b64";
1351 else
1352 return "u64";
1353 else if (useB4PTR)
1354 return "b32";
1355 else
1356 return "u32";
1357 }
1358 llvm_unreachable("unexpected type");
1359 return nullptr;
1360 }
1361
emitPTXGlobalVariable(const GlobalVariable * GVar,raw_ostream & O)1362 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1363 raw_ostream &O) {
1364
1365 const DataLayout &DL = getDataLayout();
1366
1367 // GlobalVariables are always constant pointers themselves.
1368 Type *ETy = GVar->getValueType();
1369
1370 O << ".";
1371 emitPTXAddressSpace(GVar->getType()->getAddressSpace(), O);
1372 if (GVar->getAlignment() == 0)
1373 O << " .align " << (int)DL.getPrefTypeAlignment(ETy);
1374 else
1375 O << " .align " << GVar->getAlignment();
1376
1377 if (ETy->isFloatingPointTy() || ETy->isIntegerTy() || ETy->isPointerTy()) {
1378 O << " .";
1379 O << getPTXFundamentalTypeStr(ETy);
1380 O << " ";
1381 getSymbol(GVar)->print(O, MAI);
1382 return;
1383 }
1384
1385 int64_t ElementSize = 0;
1386
1387 // Although PTX has direct support for struct type and array type and LLVM IR
1388 // is very similar to PTX, the LLVM CodeGen does not support for targets that
1389 // support these high level field accesses. Structs and arrays are lowered
1390 // into arrays of bytes.
1391 switch (ETy->getTypeID()) {
1392 case Type::StructTyID:
1393 case Type::ArrayTyID:
1394 case Type::VectorTyID:
1395 ElementSize = DL.getTypeStoreSize(ETy);
1396 O << " .b8 ";
1397 getSymbol(GVar)->print(O, MAI);
1398 O << "[";
1399 if (ElementSize) {
1400 O << ElementSize;
1401 }
1402 O << "]";
1403 break;
1404 default:
1405 llvm_unreachable("type not supported yet");
1406 }
1407 return;
1408 }
1409
getOpenCLAlignment(const DataLayout & DL,Type * Ty)1410 static unsigned int getOpenCLAlignment(const DataLayout &DL, Type *Ty) {
1411 if (Ty->isSingleValueType())
1412 return DL.getPrefTypeAlignment(Ty);
1413
1414 auto *ATy = dyn_cast<ArrayType>(Ty);
1415 if (ATy)
1416 return getOpenCLAlignment(DL, ATy->getElementType());
1417
1418 auto *STy = dyn_cast<StructType>(Ty);
1419 if (STy) {
1420 unsigned int alignStruct = 1;
1421 // Go through each element of the struct and find the
1422 // largest alignment.
1423 for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1424 Type *ETy = STy->getElementType(i);
1425 unsigned int align = getOpenCLAlignment(DL, ETy);
1426 if (align > alignStruct)
1427 alignStruct = align;
1428 }
1429 return alignStruct;
1430 }
1431
1432 auto *FTy = dyn_cast<FunctionType>(Ty);
1433 if (FTy)
1434 return DL.getPointerPrefAlignment();
1435 return DL.getPrefTypeAlignment(Ty);
1436 }
1437
printParamName(Function::const_arg_iterator I,int paramIndex,raw_ostream & O)1438 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1439 int paramIndex, raw_ostream &O) {
1440 getSymbol(I->getParent())->print(O, MAI);
1441 O << "_param_" << paramIndex;
1442 }
1443
emitFunctionParamList(const Function * F,raw_ostream & O)1444 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1445 const DataLayout &DL = getDataLayout();
1446 const AttributeSet &PAL = F->getAttributes();
1447 const TargetLowering *TLI = nvptxSubtarget->getTargetLowering();
1448 Function::const_arg_iterator I, E;
1449 unsigned paramIndex = 0;
1450 bool first = true;
1451 bool isKernelFunc = llvm::isKernelFunction(*F);
1452 bool isABI = (nvptxSubtarget->getSmVersion() >= 20);
1453 MVT thePointerTy = TLI->getPointerTy(DL);
1454
1455 if (F->arg_empty()) {
1456 O << "()\n";
1457 return;
1458 }
1459
1460 O << "(\n";
1461
1462 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1463 Type *Ty = I->getType();
1464
1465 if (!first)
1466 O << ",\n";
1467
1468 first = false;
1469
1470 // Handle image/sampler parameters
1471 if (isKernelFunction(*F)) {
1472 if (isSampler(*I) || isImage(*I)) {
1473 if (isImage(*I)) {
1474 std::string sname = I->getName();
1475 if (isImageWriteOnly(*I) || isImageReadWrite(*I)) {
1476 if (nvptxSubtarget->hasImageHandles())
1477 O << "\t.param .u64 .ptr .surfref ";
1478 else
1479 O << "\t.param .surfref ";
1480 CurrentFnSym->print(O, MAI);
1481 O << "_param_" << paramIndex;
1482 }
1483 else { // Default image is read_only
1484 if (nvptxSubtarget->hasImageHandles())
1485 O << "\t.param .u64 .ptr .texref ";
1486 else
1487 O << "\t.param .texref ";
1488 CurrentFnSym->print(O, MAI);
1489 O << "_param_" << paramIndex;
1490 }
1491 } else {
1492 if (nvptxSubtarget->hasImageHandles())
1493 O << "\t.param .u64 .ptr .samplerref ";
1494 else
1495 O << "\t.param .samplerref ";
1496 CurrentFnSym->print(O, MAI);
1497 O << "_param_" << paramIndex;
1498 }
1499 continue;
1500 }
1501 }
1502
1503 if (!PAL.hasAttribute(paramIndex + 1, Attribute::ByVal)) {
1504 if (Ty->isAggregateType() || Ty->isVectorTy()) {
1505 // Just print .param .align <a> .b8 .param[size];
1506 // <a> = PAL.getparamalignment
1507 // size = typeallocsize of element type
1508 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1509 if (align == 0)
1510 align = DL.getABITypeAlignment(Ty);
1511
1512 unsigned sz = DL.getTypeAllocSize(Ty);
1513 O << "\t.param .align " << align << " .b8 ";
1514 printParamName(I, paramIndex, O);
1515 O << "[" << sz << "]";
1516
1517 continue;
1518 }
1519 // Just a scalar
1520 auto *PTy = dyn_cast<PointerType>(Ty);
1521 if (isKernelFunc) {
1522 if (PTy) {
1523 // Special handling for pointer arguments to kernel
1524 O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1525
1526 if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() !=
1527 NVPTX::CUDA) {
1528 Type *ETy = PTy->getElementType();
1529 int addrSpace = PTy->getAddressSpace();
1530 switch (addrSpace) {
1531 default:
1532 O << ".ptr ";
1533 break;
1534 case llvm::ADDRESS_SPACE_CONST:
1535 O << ".ptr .const ";
1536 break;
1537 case llvm::ADDRESS_SPACE_SHARED:
1538 O << ".ptr .shared ";
1539 break;
1540 case llvm::ADDRESS_SPACE_GLOBAL:
1541 O << ".ptr .global ";
1542 break;
1543 }
1544 O << ".align " << (int)getOpenCLAlignment(DL, ETy) << " ";
1545 }
1546 printParamName(I, paramIndex, O);
1547 continue;
1548 }
1549
1550 // non-pointer scalar to kernel func
1551 O << "\t.param .";
1552 // Special case: predicate operands become .u8 types
1553 if (Ty->isIntegerTy(1))
1554 O << "u8";
1555 else
1556 O << getPTXFundamentalTypeStr(Ty);
1557 O << " ";
1558 printParamName(I, paramIndex, O);
1559 continue;
1560 }
1561 // Non-kernel function, just print .param .b<size> for ABI
1562 // and .reg .b<size> for non-ABI
1563 unsigned sz = 0;
1564 if (isa<IntegerType>(Ty)) {
1565 sz = cast<IntegerType>(Ty)->getBitWidth();
1566 if (sz < 32)
1567 sz = 32;
1568 } else if (isa<PointerType>(Ty))
1569 sz = thePointerTy.getSizeInBits();
1570 else
1571 sz = Ty->getPrimitiveSizeInBits();
1572 if (isABI)
1573 O << "\t.param .b" << sz << " ";
1574 else
1575 O << "\t.reg .b" << sz << " ";
1576 printParamName(I, paramIndex, O);
1577 continue;
1578 }
1579
1580 // param has byVal attribute. So should be a pointer
1581 auto *PTy = dyn_cast<PointerType>(Ty);
1582 assert(PTy && "Param with byval attribute should be a pointer type");
1583 Type *ETy = PTy->getElementType();
1584
1585 if (isABI || isKernelFunc) {
1586 // Just print .param .align <a> .b8 .param[size];
1587 // <a> = PAL.getparamalignment
1588 // size = typeallocsize of element type
1589 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1590 if (align == 0)
1591 align = DL.getABITypeAlignment(ETy);
1592
1593 unsigned sz = DL.getTypeAllocSize(ETy);
1594 O << "\t.param .align " << align << " .b8 ";
1595 printParamName(I, paramIndex, O);
1596 O << "[" << sz << "]";
1597 continue;
1598 } else {
1599 // Split the ETy into constituent parts and
1600 // print .param .b<size> <name> for each part.
1601 // Further, if a part is vector, print the above for
1602 // each vector element.
1603 SmallVector<EVT, 16> vtparts;
1604 ComputeValueVTs(*TLI, DL, ETy, vtparts);
1605 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1606 unsigned elems = 1;
1607 EVT elemtype = vtparts[i];
1608 if (vtparts[i].isVector()) {
1609 elems = vtparts[i].getVectorNumElements();
1610 elemtype = vtparts[i].getVectorElementType();
1611 }
1612
1613 for (unsigned j = 0, je = elems; j != je; ++j) {
1614 unsigned sz = elemtype.getSizeInBits();
1615 if (elemtype.isInteger() && (sz < 32))
1616 sz = 32;
1617 O << "\t.reg .b" << sz << " ";
1618 printParamName(I, paramIndex, O);
1619 if (j < je - 1)
1620 O << ",\n";
1621 ++paramIndex;
1622 }
1623 if (i < e - 1)
1624 O << ",\n";
1625 }
1626 --paramIndex;
1627 continue;
1628 }
1629 }
1630
1631 O << "\n)\n";
1632 }
1633
emitFunctionParamList(const MachineFunction & MF,raw_ostream & O)1634 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1635 raw_ostream &O) {
1636 const Function *F = MF.getFunction();
1637 emitFunctionParamList(F, O);
1638 }
1639
setAndEmitFunctionVirtualRegisters(const MachineFunction & MF)1640 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1641 const MachineFunction &MF) {
1642 SmallString<128> Str;
1643 raw_svector_ostream O(Str);
1644
1645 // Map the global virtual register number to a register class specific
1646 // virtual register number starting from 1 with that class.
1647 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1648 //unsigned numRegClasses = TRI->getNumRegClasses();
1649
1650 // Emit the Fake Stack Object
1651 const MachineFrameInfo *MFI = MF.getFrameInfo();
1652 int NumBytes = (int) MFI->getStackSize();
1653 if (NumBytes) {
1654 O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
1655 << getFunctionNumber() << "[" << NumBytes << "];\n";
1656 if (static_cast<const NVPTXTargetMachine &>(MF.getTarget()).is64Bit()) {
1657 O << "\t.reg .b64 \t%SP;\n";
1658 O << "\t.reg .b64 \t%SPL;\n";
1659 } else {
1660 O << "\t.reg .b32 \t%SP;\n";
1661 O << "\t.reg .b32 \t%SPL;\n";
1662 }
1663 }
1664
1665 // Go through all virtual registers to establish the mapping between the
1666 // global virtual
1667 // register number and the per class virtual register number.
1668 // We use the per class virtual register number in the ptx output.
1669 unsigned int numVRs = MRI->getNumVirtRegs();
1670 for (unsigned i = 0; i < numVRs; i++) {
1671 unsigned int vr = TRI->index2VirtReg(i);
1672 const TargetRegisterClass *RC = MRI->getRegClass(vr);
1673 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1674 int n = regmap.size();
1675 regmap.insert(std::make_pair(vr, n + 1));
1676 }
1677
1678 // Emit register declarations
1679 // @TODO: Extract out the real register usage
1680 // O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
1681 // O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
1682 // O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
1683 // O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
1684 // O << "\t.reg .s64 %rd<" << NVPTXNumRegisters << ">;\n";
1685 // O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
1686 // O << "\t.reg .f64 %fd<" << NVPTXNumRegisters << ">;\n";
1687
1688 // Emit declaration of the virtual registers or 'physical' registers for
1689 // each register class
1690 for (unsigned i=0; i< TRI->getNumRegClasses(); i++) {
1691 const TargetRegisterClass *RC = TRI->getRegClass(i);
1692 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1693 std::string rcname = getNVPTXRegClassName(RC);
1694 std::string rcStr = getNVPTXRegClassStr(RC);
1695 int n = regmap.size();
1696
1697 // Only declare those registers that may be used.
1698 if (n) {
1699 O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
1700 << ">;\n";
1701 }
1702 }
1703
1704 OutStreamer->EmitRawText(O.str());
1705 }
1706
printFPConstant(const ConstantFP * Fp,raw_ostream & O)1707 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
1708 APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
1709 bool ignored;
1710 unsigned int numHex;
1711 const char *lead;
1712
1713 if (Fp->getType()->getTypeID() == Type::FloatTyID) {
1714 numHex = 8;
1715 lead = "0f";
1716 APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &ignored);
1717 } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
1718 numHex = 16;
1719 lead = "0d";
1720 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &ignored);
1721 } else
1722 llvm_unreachable("unsupported fp type");
1723
1724 APInt API = APF.bitcastToAPInt();
1725 std::string hexstr(utohexstr(API.getZExtValue()));
1726 O << lead;
1727 if (hexstr.length() < numHex)
1728 O << std::string(numHex - hexstr.length(), '0');
1729 O << utohexstr(API.getZExtValue());
1730 }
1731
printScalarConstant(const Constant * CPV,raw_ostream & O)1732 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
1733 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
1734 O << CI->getValue();
1735 return;
1736 }
1737 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
1738 printFPConstant(CFP, O);
1739 return;
1740 }
1741 if (isa<ConstantPointerNull>(CPV)) {
1742 O << "0";
1743 return;
1744 }
1745 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1746 bool IsNonGenericPointer = false;
1747 if (GVar->getType()->getAddressSpace() != 0) {
1748 IsNonGenericPointer = true;
1749 }
1750 if (EmitGeneric && !isa<Function>(CPV) && !IsNonGenericPointer) {
1751 O << "generic(";
1752 getSymbol(GVar)->print(O, MAI);
1753 O << ")";
1754 } else {
1755 getSymbol(GVar)->print(O, MAI);
1756 }
1757 return;
1758 }
1759 if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1760 const Value *v = Cexpr->stripPointerCasts();
1761 PointerType *PTy = dyn_cast<PointerType>(Cexpr->getType());
1762 bool IsNonGenericPointer = false;
1763 if (PTy && PTy->getAddressSpace() != 0) {
1764 IsNonGenericPointer = true;
1765 }
1766 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
1767 if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
1768 O << "generic(";
1769 getSymbol(GVar)->print(O, MAI);
1770 O << ")";
1771 } else {
1772 getSymbol(GVar)->print(O, MAI);
1773 }
1774 return;
1775 } else {
1776 lowerConstant(CPV)->print(O, MAI);
1777 return;
1778 }
1779 }
1780 llvm_unreachable("Not scalar type found in printScalarConstant()");
1781 }
1782
1783 // These utility functions assure we get the right sequence of bytes for a given
1784 // type even for big-endian machines
ConvertIntToBytes(unsigned char * p,T val)1785 template <typename T> static void ConvertIntToBytes(unsigned char *p, T val) {
1786 int64_t vp = (int64_t)val;
1787 for (unsigned i = 0; i < sizeof(T); ++i) {
1788 p[i] = (unsigned char)vp;
1789 vp >>= 8;
1790 }
1791 }
ConvertFloatToBytes(unsigned char * p,float val)1792 static void ConvertFloatToBytes(unsigned char *p, float val) {
1793 int32_t *vp = (int32_t *)&val;
1794 for (unsigned i = 0; i < sizeof(int32_t); ++i) {
1795 p[i] = (unsigned char)*vp;
1796 *vp >>= 8;
1797 }
1798 }
ConvertDoubleToBytes(unsigned char * p,double val)1799 static void ConvertDoubleToBytes(unsigned char *p, double val) {
1800 int64_t *vp = (int64_t *)&val;
1801 for (unsigned i = 0; i < sizeof(int64_t); ++i) {
1802 p[i] = (unsigned char)*vp;
1803 *vp >>= 8;
1804 }
1805 }
1806
bufferLEByte(const Constant * CPV,int Bytes,AggBuffer * aggBuffer)1807 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
1808 AggBuffer *aggBuffer) {
1809
1810 const DataLayout &DL = getDataLayout();
1811
1812 if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
1813 int s = DL.getTypeAllocSize(CPV->getType());
1814 if (s < Bytes)
1815 s = Bytes;
1816 aggBuffer->addZeros(s);
1817 return;
1818 }
1819
1820 unsigned char ptr[8];
1821 switch (CPV->getType()->getTypeID()) {
1822
1823 case Type::IntegerTyID: {
1824 Type *ETy = CPV->getType();
1825 if (ETy == Type::getInt8Ty(CPV->getContext())) {
1826 unsigned char c = (unsigned char)cast<ConstantInt>(CPV)->getZExtValue();
1827 ConvertIntToBytes<>(ptr, c);
1828 aggBuffer->addBytes(ptr, 1, Bytes);
1829 } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
1830 short int16 = (short)cast<ConstantInt>(CPV)->getZExtValue();
1831 ConvertIntToBytes<>(ptr, int16);
1832 aggBuffer->addBytes(ptr, 2, Bytes);
1833 } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
1834 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1835 int int32 = (int)(constInt->getZExtValue());
1836 ConvertIntToBytes<>(ptr, int32);
1837 aggBuffer->addBytes(ptr, 4, Bytes);
1838 break;
1839 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1840 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1841 ConstantFoldConstantExpression(Cexpr, DL))) {
1842 int int32 = (int)(constInt->getZExtValue());
1843 ConvertIntToBytes<>(ptr, int32);
1844 aggBuffer->addBytes(ptr, 4, 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(4);
1851 break;
1852 }
1853 }
1854 llvm_unreachable("unsupported integer const type");
1855 } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
1856 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1857 long long int64 = (long long)(constInt->getZExtValue());
1858 ConvertIntToBytes<>(ptr, int64);
1859 aggBuffer->addBytes(ptr, 8, Bytes);
1860 break;
1861 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1862 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1863 ConstantFoldConstantExpression(Cexpr, DL))) {
1864 long long int64 = (long long)(constInt->getZExtValue());
1865 ConvertIntToBytes<>(ptr, int64);
1866 aggBuffer->addBytes(ptr, 8, Bytes);
1867 break;
1868 }
1869 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1870 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1871 aggBuffer->addSymbol(v, Cexpr->getOperand(0));
1872 aggBuffer->addZeros(8);
1873 break;
1874 }
1875 }
1876 llvm_unreachable("unsupported integer const type");
1877 } else
1878 llvm_unreachable("unsupported integer const type");
1879 break;
1880 }
1881 case Type::FloatTyID:
1882 case Type::DoubleTyID: {
1883 const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
1884 Type *Ty = CFP->getType();
1885 if (Ty == Type::getFloatTy(CPV->getContext())) {
1886 float float32 = (float) CFP->getValueAPF().convertToFloat();
1887 ConvertFloatToBytes(ptr, float32);
1888 aggBuffer->addBytes(ptr, 4, Bytes);
1889 } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
1890 double float64 = CFP->getValueAPF().convertToDouble();
1891 ConvertDoubleToBytes(ptr, float64);
1892 aggBuffer->addBytes(ptr, 8, Bytes);
1893 } else {
1894 llvm_unreachable("unsupported fp const type");
1895 }
1896 break;
1897 }
1898 case Type::PointerTyID: {
1899 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1900 aggBuffer->addSymbol(GVar, GVar);
1901 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1902 const Value *v = Cexpr->stripPointerCasts();
1903 aggBuffer->addSymbol(v, Cexpr);
1904 }
1905 unsigned int s = DL.getTypeAllocSize(CPV->getType());
1906 aggBuffer->addZeros(s);
1907 break;
1908 }
1909
1910 case Type::ArrayTyID:
1911 case Type::VectorTyID:
1912 case Type::StructTyID: {
1913 if (isa<ConstantAggregate>(CPV) || isa<ConstantDataSequential>(CPV)) {
1914 int ElementSize = DL.getTypeAllocSize(CPV->getType());
1915 bufferAggregateConstant(CPV, aggBuffer);
1916 if (Bytes > ElementSize)
1917 aggBuffer->addZeros(Bytes - ElementSize);
1918 } else if (isa<ConstantAggregateZero>(CPV))
1919 aggBuffer->addZeros(Bytes);
1920 else
1921 llvm_unreachable("Unexpected Constant type");
1922 break;
1923 }
1924
1925 default:
1926 llvm_unreachable("unsupported type");
1927 }
1928 }
1929
bufferAggregateConstant(const Constant * CPV,AggBuffer * aggBuffer)1930 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
1931 AggBuffer *aggBuffer) {
1932 const DataLayout &DL = getDataLayout();
1933 int Bytes;
1934
1935 // Old constants
1936 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
1937 if (CPV->getNumOperands())
1938 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
1939 bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
1940 return;
1941 }
1942
1943 if (const ConstantDataSequential *CDS =
1944 dyn_cast<ConstantDataSequential>(CPV)) {
1945 if (CDS->getNumElements())
1946 for (unsigned i = 0; i < CDS->getNumElements(); ++i)
1947 bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
1948 aggBuffer);
1949 return;
1950 }
1951
1952 if (isa<ConstantStruct>(CPV)) {
1953 if (CPV->getNumOperands()) {
1954 StructType *ST = cast<StructType>(CPV->getType());
1955 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
1956 if (i == (e - 1))
1957 Bytes = DL.getStructLayout(ST)->getElementOffset(0) +
1958 DL.getTypeAllocSize(ST) -
1959 DL.getStructLayout(ST)->getElementOffset(i);
1960 else
1961 Bytes = DL.getStructLayout(ST)->getElementOffset(i + 1) -
1962 DL.getStructLayout(ST)->getElementOffset(i);
1963 bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
1964 }
1965 }
1966 return;
1967 }
1968 llvm_unreachable("unsupported constant type in printAggregateConstant()");
1969 }
1970
1971 // buildTypeNameMap - Run through symbol table looking for type names.
1972 //
1973
1974
ignoreLoc(const MachineInstr & MI)1975 bool NVPTXAsmPrinter::ignoreLoc(const MachineInstr &MI) {
1976 switch (MI.getOpcode()) {
1977 default:
1978 return false;
1979 case NVPTX::CallArgBeginInst:
1980 case NVPTX::CallArgEndInst0:
1981 case NVPTX::CallArgEndInst1:
1982 case NVPTX::CallArgF32:
1983 case NVPTX::CallArgF64:
1984 case NVPTX::CallArgI16:
1985 case NVPTX::CallArgI32:
1986 case NVPTX::CallArgI32imm:
1987 case NVPTX::CallArgI64:
1988 case NVPTX::CallArgParam:
1989 case NVPTX::CallVoidInst:
1990 case NVPTX::CallVoidInstReg:
1991 case NVPTX::Callseq_End:
1992 case NVPTX::CallVoidInstReg64:
1993 case NVPTX::DeclareParamInst:
1994 case NVPTX::DeclareRetMemInst:
1995 case NVPTX::DeclareRetRegInst:
1996 case NVPTX::DeclareRetScalarInst:
1997 case NVPTX::DeclareScalarParamInst:
1998 case NVPTX::DeclareScalarRegInst:
1999 case NVPTX::StoreParamF32:
2000 case NVPTX::StoreParamF64:
2001 case NVPTX::StoreParamI16:
2002 case NVPTX::StoreParamI32:
2003 case NVPTX::StoreParamI64:
2004 case NVPTX::StoreParamI8:
2005 case NVPTX::StoreRetvalF32:
2006 case NVPTX::StoreRetvalF64:
2007 case NVPTX::StoreRetvalI16:
2008 case NVPTX::StoreRetvalI32:
2009 case NVPTX::StoreRetvalI64:
2010 case NVPTX::StoreRetvalI8:
2011 case NVPTX::LastCallArgF32:
2012 case NVPTX::LastCallArgF64:
2013 case NVPTX::LastCallArgI16:
2014 case NVPTX::LastCallArgI32:
2015 case NVPTX::LastCallArgI32imm:
2016 case NVPTX::LastCallArgI64:
2017 case NVPTX::LastCallArgParam:
2018 case NVPTX::LoadParamMemF32:
2019 case NVPTX::LoadParamMemF64:
2020 case NVPTX::LoadParamMemI16:
2021 case NVPTX::LoadParamMemI32:
2022 case NVPTX::LoadParamMemI64:
2023 case NVPTX::LoadParamMemI8:
2024 case NVPTX::PrototypeInst:
2025 case NVPTX::DBG_VALUE:
2026 return true;
2027 }
2028 return false;
2029 }
2030
2031 /// lowerConstantForGV - Return an MCExpr for the given Constant. This is mostly
2032 /// a copy from AsmPrinter::lowerConstant, except customized to only handle
2033 /// expressions that are representable in PTX and create
2034 /// NVPTXGenericMCSymbolRefExpr nodes for addrspacecast instructions.
2035 const MCExpr *
lowerConstantForGV(const Constant * CV,bool ProcessingGeneric)2036 NVPTXAsmPrinter::lowerConstantForGV(const Constant *CV, bool ProcessingGeneric) {
2037 MCContext &Ctx = OutContext;
2038
2039 if (CV->isNullValue() || isa<UndefValue>(CV))
2040 return MCConstantExpr::create(0, Ctx);
2041
2042 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV))
2043 return MCConstantExpr::create(CI->getZExtValue(), Ctx);
2044
2045 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CV)) {
2046 const MCSymbolRefExpr *Expr =
2047 MCSymbolRefExpr::create(getSymbol(GV), Ctx);
2048 if (ProcessingGeneric) {
2049 return NVPTXGenericMCSymbolRefExpr::create(Expr, Ctx);
2050 } else {
2051 return Expr;
2052 }
2053 }
2054
2055 const ConstantExpr *CE = dyn_cast<ConstantExpr>(CV);
2056 if (!CE) {
2057 llvm_unreachable("Unknown constant value to lower!");
2058 }
2059
2060 switch (CE->getOpcode()) {
2061 default:
2062 // If the code isn't optimized, there may be outstanding folding
2063 // opportunities. Attempt to fold the expression using DataLayout as a
2064 // last resort before giving up.
2065 if (Constant *C = ConstantFoldConstantExpression(CE, getDataLayout()))
2066 if (C != CE)
2067 return lowerConstantForGV(C, ProcessingGeneric);
2068
2069 // Otherwise report the problem to the user.
2070 {
2071 std::string S;
2072 raw_string_ostream OS(S);
2073 OS << "Unsupported expression in static initializer: ";
2074 CE->printAsOperand(OS, /*PrintType=*/false,
2075 !MF ? nullptr : MF->getFunction()->getParent());
2076 report_fatal_error(OS.str());
2077 }
2078
2079 case Instruction::AddrSpaceCast: {
2080 // Strip the addrspacecast and pass along the operand
2081 PointerType *DstTy = cast<PointerType>(CE->getType());
2082 if (DstTy->getAddressSpace() == 0) {
2083 return lowerConstantForGV(cast<const Constant>(CE->getOperand(0)), true);
2084 }
2085 std::string S;
2086 raw_string_ostream OS(S);
2087 OS << "Unsupported expression in static initializer: ";
2088 CE->printAsOperand(OS, /*PrintType=*/ false,
2089 !MF ? 0 : MF->getFunction()->getParent());
2090 report_fatal_error(OS.str());
2091 }
2092
2093 case Instruction::GetElementPtr: {
2094 const DataLayout &DL = getDataLayout();
2095
2096 // Generate a symbolic expression for the byte address
2097 APInt OffsetAI(DL.getPointerTypeSizeInBits(CE->getType()), 0);
2098 cast<GEPOperator>(CE)->accumulateConstantOffset(DL, OffsetAI);
2099
2100 const MCExpr *Base = lowerConstantForGV(CE->getOperand(0),
2101 ProcessingGeneric);
2102 if (!OffsetAI)
2103 return Base;
2104
2105 int64_t Offset = OffsetAI.getSExtValue();
2106 return MCBinaryExpr::createAdd(Base, MCConstantExpr::create(Offset, Ctx),
2107 Ctx);
2108 }
2109
2110 case Instruction::Trunc:
2111 // We emit the value and depend on the assembler to truncate the generated
2112 // expression properly. This is important for differences between
2113 // blockaddress labels. Since the two labels are in the same function, it
2114 // is reasonable to treat their delta as a 32-bit value.
2115 // FALL THROUGH.
2116 case Instruction::BitCast:
2117 return lowerConstantForGV(CE->getOperand(0), ProcessingGeneric);
2118
2119 case Instruction::IntToPtr: {
2120 const DataLayout &DL = getDataLayout();
2121
2122 // Handle casts to pointers by changing them into casts to the appropriate
2123 // integer type. This promotes constant folding and simplifies this code.
2124 Constant *Op = CE->getOperand(0);
2125 Op = ConstantExpr::getIntegerCast(Op, DL.getIntPtrType(CV->getType()),
2126 false/*ZExt*/);
2127 return lowerConstantForGV(Op, ProcessingGeneric);
2128 }
2129
2130 case Instruction::PtrToInt: {
2131 const DataLayout &DL = getDataLayout();
2132
2133 // Support only foldable casts to/from pointers that can be eliminated by
2134 // changing the pointer to the appropriately sized integer type.
2135 Constant *Op = CE->getOperand(0);
2136 Type *Ty = CE->getType();
2137
2138 const MCExpr *OpExpr = lowerConstantForGV(Op, ProcessingGeneric);
2139
2140 // We can emit the pointer value into this slot if the slot is an
2141 // integer slot equal to the size of the pointer.
2142 if (DL.getTypeAllocSize(Ty) == DL.getTypeAllocSize(Op->getType()))
2143 return OpExpr;
2144
2145 // Otherwise the pointer is smaller than the resultant integer, mask off
2146 // the high bits so we are sure to get a proper truncation if the input is
2147 // a constant expr.
2148 unsigned InBits = DL.getTypeAllocSizeInBits(Op->getType());
2149 const MCExpr *MaskExpr = MCConstantExpr::create(~0ULL >> (64-InBits), Ctx);
2150 return MCBinaryExpr::createAnd(OpExpr, MaskExpr, Ctx);
2151 }
2152
2153 // The MC library also has a right-shift operator, but it isn't consistently
2154 // signed or unsigned between different targets.
2155 case Instruction::Add: {
2156 const MCExpr *LHS = lowerConstantForGV(CE->getOperand(0), ProcessingGeneric);
2157 const MCExpr *RHS = lowerConstantForGV(CE->getOperand(1), ProcessingGeneric);
2158 switch (CE->getOpcode()) {
2159 default: llvm_unreachable("Unknown binary operator constant cast expr");
2160 case Instruction::Add: return MCBinaryExpr::createAdd(LHS, RHS, Ctx);
2161 }
2162 }
2163 }
2164 }
2165
2166 // Copy of MCExpr::print customized for NVPTX
printMCExpr(const MCExpr & Expr,raw_ostream & OS)2167 void NVPTXAsmPrinter::printMCExpr(const MCExpr &Expr, raw_ostream &OS) {
2168 switch (Expr.getKind()) {
2169 case MCExpr::Target:
2170 return cast<MCTargetExpr>(&Expr)->printImpl(OS, MAI);
2171 case MCExpr::Constant:
2172 OS << cast<MCConstantExpr>(Expr).getValue();
2173 return;
2174
2175 case MCExpr::SymbolRef: {
2176 const MCSymbolRefExpr &SRE = cast<MCSymbolRefExpr>(Expr);
2177 const MCSymbol &Sym = SRE.getSymbol();
2178 Sym.print(OS, MAI);
2179 return;
2180 }
2181
2182 case MCExpr::Unary: {
2183 const MCUnaryExpr &UE = cast<MCUnaryExpr>(Expr);
2184 switch (UE.getOpcode()) {
2185 case MCUnaryExpr::LNot: OS << '!'; break;
2186 case MCUnaryExpr::Minus: OS << '-'; break;
2187 case MCUnaryExpr::Not: OS << '~'; break;
2188 case MCUnaryExpr::Plus: OS << '+'; break;
2189 }
2190 printMCExpr(*UE.getSubExpr(), OS);
2191 return;
2192 }
2193
2194 case MCExpr::Binary: {
2195 const MCBinaryExpr &BE = cast<MCBinaryExpr>(Expr);
2196
2197 // Only print parens around the LHS if it is non-trivial.
2198 if (isa<MCConstantExpr>(BE.getLHS()) || isa<MCSymbolRefExpr>(BE.getLHS()) ||
2199 isa<NVPTXGenericMCSymbolRefExpr>(BE.getLHS())) {
2200 printMCExpr(*BE.getLHS(), OS);
2201 } else {
2202 OS << '(';
2203 printMCExpr(*BE.getLHS(), OS);
2204 OS<< ')';
2205 }
2206
2207 switch (BE.getOpcode()) {
2208 case MCBinaryExpr::Add:
2209 // Print "X-42" instead of "X+-42".
2210 if (const MCConstantExpr *RHSC = dyn_cast<MCConstantExpr>(BE.getRHS())) {
2211 if (RHSC->getValue() < 0) {
2212 OS << RHSC->getValue();
2213 return;
2214 }
2215 }
2216
2217 OS << '+';
2218 break;
2219 default: llvm_unreachable("Unhandled binary operator");
2220 }
2221
2222 // Only print parens around the LHS if it is non-trivial.
2223 if (isa<MCConstantExpr>(BE.getRHS()) || isa<MCSymbolRefExpr>(BE.getRHS())) {
2224 printMCExpr(*BE.getRHS(), OS);
2225 } else {
2226 OS << '(';
2227 printMCExpr(*BE.getRHS(), OS);
2228 OS << ')';
2229 }
2230 return;
2231 }
2232 }
2233
2234 llvm_unreachable("Invalid expression kind!");
2235 }
2236
2237 /// PrintAsmOperand - Print out an operand for an inline asm expression.
2238 ///
PrintAsmOperand(const MachineInstr * MI,unsigned OpNo,unsigned AsmVariant,const char * ExtraCode,raw_ostream & O)2239 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
2240 unsigned AsmVariant,
2241 const char *ExtraCode, raw_ostream &O) {
2242 if (ExtraCode && ExtraCode[0]) {
2243 if (ExtraCode[1] != 0)
2244 return true; // Unknown modifier.
2245
2246 switch (ExtraCode[0]) {
2247 default:
2248 // See if this is a generic print operand
2249 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
2250 case 'r':
2251 break;
2252 }
2253 }
2254
2255 printOperand(MI, OpNo, O);
2256
2257 return false;
2258 }
2259
PrintAsmMemoryOperand(const MachineInstr * MI,unsigned OpNo,unsigned AsmVariant,const char * ExtraCode,raw_ostream & O)2260 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(
2261 const MachineInstr *MI, unsigned OpNo, unsigned AsmVariant,
2262 const char *ExtraCode, raw_ostream &O) {
2263 if (ExtraCode && ExtraCode[0])
2264 return true; // Unknown modifier
2265
2266 O << '[';
2267 printMemOperand(MI, OpNo, O);
2268 O << ']';
2269
2270 return false;
2271 }
2272
printOperand(const MachineInstr * MI,int opNum,raw_ostream & O,const char * Modifier)2273 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
2274 raw_ostream &O, const char *Modifier) {
2275 const MachineOperand &MO = MI->getOperand(opNum);
2276 switch (MO.getType()) {
2277 case MachineOperand::MO_Register:
2278 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
2279 if (MO.getReg() == NVPTX::VRDepot)
2280 O << DEPOTNAME << getFunctionNumber();
2281 else
2282 O << NVPTXInstPrinter::getRegisterName(MO.getReg());
2283 } else {
2284 emitVirtualRegister(MO.getReg(), O);
2285 }
2286 return;
2287
2288 case MachineOperand::MO_Immediate:
2289 if (!Modifier)
2290 O << MO.getImm();
2291 else if (strstr(Modifier, "vec") == Modifier)
2292 printVecModifiedImmediate(MO, Modifier, O);
2293 else
2294 llvm_unreachable(
2295 "Don't know how to handle modifier on immediate operand");
2296 return;
2297
2298 case MachineOperand::MO_FPImmediate:
2299 printFPConstant(MO.getFPImm(), O);
2300 break;
2301
2302 case MachineOperand::MO_GlobalAddress:
2303 getSymbol(MO.getGlobal())->print(O, MAI);
2304 break;
2305
2306 case MachineOperand::MO_MachineBasicBlock:
2307 MO.getMBB()->getSymbol()->print(O, MAI);
2308 return;
2309
2310 default:
2311 llvm_unreachable("Operand type not supported.");
2312 }
2313 }
2314
printMemOperand(const MachineInstr * MI,int opNum,raw_ostream & O,const char * Modifier)2315 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
2316 raw_ostream &O, const char *Modifier) {
2317 printOperand(MI, opNum, O);
2318
2319 if (Modifier && !strcmp(Modifier, "add")) {
2320 O << ", ";
2321 printOperand(MI, opNum + 1, O);
2322 } else {
2323 if (MI->getOperand(opNum + 1).isImm() &&
2324 MI->getOperand(opNum + 1).getImm() == 0)
2325 return; // don't print ',0' or '+0'
2326 O << "+";
2327 printOperand(MI, opNum + 1, O);
2328 }
2329 }
2330
emitSrcInText(StringRef filename,unsigned line)2331 void NVPTXAsmPrinter::emitSrcInText(StringRef filename, unsigned line) {
2332 std::stringstream temp;
2333 LineReader *reader = this->getReader(filename);
2334 temp << "\n//";
2335 temp << filename.str();
2336 temp << ":";
2337 temp << line;
2338 temp << " ";
2339 temp << reader->readLine(line);
2340 temp << "\n";
2341 this->OutStreamer->EmitRawText(temp.str());
2342 }
2343
getReader(const std::string & filename)2344 LineReader *NVPTXAsmPrinter::getReader(const std::string &filename) {
2345 if (!reader) {
2346 reader = new LineReader(filename);
2347 }
2348
2349 if (reader->fileName() != filename) {
2350 delete reader;
2351 reader = new LineReader(filename);
2352 }
2353
2354 return reader;
2355 }
2356
readLine(unsigned lineNum)2357 std::string LineReader::readLine(unsigned lineNum) {
2358 if (lineNum < theCurLine) {
2359 theCurLine = 0;
2360 fstr.seekg(0, std::ios::beg);
2361 }
2362 while (theCurLine < lineNum) {
2363 fstr.getline(buff, 500);
2364 theCurLine++;
2365 }
2366 return buff;
2367 }
2368
2369 // Force static initialization.
LLVMInitializeNVPTXAsmPrinter()2370 extern "C" void LLVMInitializeNVPTXAsmPrinter() {
2371 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2372 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2373 }
2374