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 "MCTargetDesc/NVPTXMCAsmInfo.h"
17 #include "NVPTX.h"
18 #include "NVPTXInstrInfo.h"
19 #include "NVPTXMCExpr.h"
20 #include "NVPTXRegisterInfo.h"
21 #include "NVPTXTargetMachine.h"
22 #include "NVPTXUtilities.h"
23 #include "InstPrinter/NVPTXInstPrinter.h"
24 #include "cl_common_defines.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Assembly/Writer.h"
28 #include "llvm/CodeGen/Analysis.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineModuleInfo.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/DebugInfo.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/GlobalVariable.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/Mangler.h"
47 #include "llvm/Target/TargetLoweringObjectFile.h"
48 #include <sstream>
49 using namespace llvm;
50
51 #define DEPOTNAME "__local_depot"
52
53 static cl::opt<bool>
54 EmitLineNumbers("nvptx-emit-line-numbers", cl::Hidden,
55 cl::desc("NVPTX Specific: Emit Line numbers even without -G"),
56 cl::init(true));
57
58 static cl::opt<bool>
59 InterleaveSrc("nvptx-emit-src", cl::ZeroOrMore, cl::Hidden,
60 cl::desc("NVPTX Specific: Emit source line in ptx file"),
61 cl::init(false));
62
63 namespace {
64 /// DiscoverDependentGlobals - Return a set of GlobalVariables on which \p V
65 /// depends.
DiscoverDependentGlobals(const Value * V,DenseSet<const GlobalVariable * > & Globals)66 void DiscoverDependentGlobals(const Value *V,
67 DenseSet<const GlobalVariable *> &Globals) {
68 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
69 Globals.insert(GV);
70 else {
71 if (const User *U = dyn_cast<User>(V)) {
72 for (unsigned i = 0, e = U->getNumOperands(); i != e; ++i) {
73 DiscoverDependentGlobals(U->getOperand(i), Globals);
74 }
75 }
76 }
77 }
78
79 /// VisitGlobalVariableForEmission - Add \p GV to the list of GlobalVariable
80 /// instances to be emitted, but only after any dependents have been added
81 /// first.
VisitGlobalVariableForEmission(const GlobalVariable * GV,SmallVectorImpl<const GlobalVariable * > & Order,DenseSet<const GlobalVariable * > & Visited,DenseSet<const GlobalVariable * > & Visiting)82 void VisitGlobalVariableForEmission(
83 const GlobalVariable *GV, SmallVectorImpl<const GlobalVariable *> &Order,
84 DenseSet<const GlobalVariable *> &Visited,
85 DenseSet<const GlobalVariable *> &Visiting) {
86 // Have we already visited this one?
87 if (Visited.count(GV))
88 return;
89
90 // Do we have a circular dependency?
91 if (Visiting.count(GV))
92 report_fatal_error("Circular dependency found in global variable set");
93
94 // Start visiting this global
95 Visiting.insert(GV);
96
97 // Make sure we visit all dependents first
98 DenseSet<const GlobalVariable *> Others;
99 for (unsigned i = 0, e = GV->getNumOperands(); i != e; ++i)
100 DiscoverDependentGlobals(GV->getOperand(i), Others);
101
102 for (DenseSet<const GlobalVariable *>::iterator I = Others.begin(),
103 E = Others.end();
104 I != E; ++I)
105 VisitGlobalVariableForEmission(*I, Order, Visited, Visiting);
106
107 // Now we can visit ourself
108 Order.push_back(GV);
109 Visited.insert(GV);
110 Visiting.erase(GV);
111 }
112 }
113
114 // @TODO: This is a copy from AsmPrinter.cpp. The function is static, so we
115 // cannot just link to the existing version.
116 /// LowerConstant - Lower the specified LLVM Constant to an MCExpr.
117 ///
118 using namespace nvptx;
LowerConstant(const Constant * CV,AsmPrinter & AP)119 const MCExpr *nvptx::LowerConstant(const Constant *CV, AsmPrinter &AP) {
120 MCContext &Ctx = AP.OutContext;
121
122 if (CV->isNullValue() || isa<UndefValue>(CV))
123 return MCConstantExpr::Create(0, Ctx);
124
125 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV))
126 return MCConstantExpr::Create(CI->getZExtValue(), Ctx);
127
128 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CV))
129 return MCSymbolRefExpr::Create(AP.getSymbol(GV), Ctx);
130
131 if (const BlockAddress *BA = dyn_cast<BlockAddress>(CV))
132 return MCSymbolRefExpr::Create(AP.GetBlockAddressSymbol(BA), Ctx);
133
134 const ConstantExpr *CE = dyn_cast<ConstantExpr>(CV);
135 if (CE == 0)
136 llvm_unreachable("Unknown constant value to lower!");
137
138 switch (CE->getOpcode()) {
139 default:
140 // If the code isn't optimized, there may be outstanding folding
141 // opportunities. Attempt to fold the expression using DataLayout as a
142 // last resort before giving up.
143 if (Constant *C = ConstantFoldConstantExpression(CE, AP.TM.getDataLayout()))
144 if (C != CE)
145 return LowerConstant(C, AP);
146
147 // Otherwise report the problem to the user.
148 {
149 std::string S;
150 raw_string_ostream OS(S);
151 OS << "Unsupported expression in static initializer: ";
152 WriteAsOperand(OS, CE, /*PrintType=*/ false,
153 !AP.MF ? 0 : AP.MF->getFunction()->getParent());
154 report_fatal_error(OS.str());
155 }
156 case Instruction::GetElementPtr: {
157 const DataLayout &TD = *AP.TM.getDataLayout();
158 // Generate a symbolic expression for the byte address
159 APInt OffsetAI(TD.getPointerSizeInBits(), 0);
160 cast<GEPOperator>(CE)->accumulateConstantOffset(TD, OffsetAI);
161
162 const MCExpr *Base = LowerConstant(CE->getOperand(0), AP);
163 if (!OffsetAI)
164 return Base;
165
166 int64_t Offset = OffsetAI.getSExtValue();
167 return MCBinaryExpr::CreateAdd(Base, MCConstantExpr::Create(Offset, Ctx),
168 Ctx);
169 }
170
171 case Instruction::Trunc:
172 // We emit the value and depend on the assembler to truncate the generated
173 // expression properly. This is important for differences between
174 // blockaddress labels. Since the two labels are in the same function, it
175 // is reasonable to treat their delta as a 32-bit value.
176 // FALL THROUGH.
177 case Instruction::BitCast:
178 return LowerConstant(CE->getOperand(0), AP);
179
180 case Instruction::IntToPtr: {
181 const DataLayout &TD = *AP.TM.getDataLayout();
182 // Handle casts to pointers by changing them into casts to the appropriate
183 // integer type. This promotes constant folding and simplifies this code.
184 Constant *Op = CE->getOperand(0);
185 Op = ConstantExpr::getIntegerCast(Op, TD.getIntPtrType(CV->getContext()),
186 false /*ZExt*/);
187 return LowerConstant(Op, AP);
188 }
189
190 case Instruction::PtrToInt: {
191 const DataLayout &TD = *AP.TM.getDataLayout();
192 // Support only foldable casts to/from pointers that can be eliminated by
193 // changing the pointer to the appropriately sized integer type.
194 Constant *Op = CE->getOperand(0);
195 Type *Ty = CE->getType();
196
197 const MCExpr *OpExpr = LowerConstant(Op, AP);
198
199 // We can emit the pointer value into this slot if the slot is an
200 // integer slot equal to the size of the pointer.
201 if (TD.getTypeAllocSize(Ty) == TD.getTypeAllocSize(Op->getType()))
202 return OpExpr;
203
204 // Otherwise the pointer is smaller than the resultant integer, mask off
205 // the high bits so we are sure to get a proper truncation if the input is
206 // a constant expr.
207 unsigned InBits = TD.getTypeAllocSizeInBits(Op->getType());
208 const MCExpr *MaskExpr =
209 MCConstantExpr::Create(~0ULL >> (64 - InBits), Ctx);
210 return MCBinaryExpr::CreateAnd(OpExpr, MaskExpr, Ctx);
211 }
212
213 // The MC library also has a right-shift operator, but it isn't consistently
214 // signed or unsigned between different targets.
215 case Instruction::Add:
216 case Instruction::Sub:
217 case Instruction::Mul:
218 case Instruction::SDiv:
219 case Instruction::SRem:
220 case Instruction::Shl:
221 case Instruction::And:
222 case Instruction::Or:
223 case Instruction::Xor: {
224 const MCExpr *LHS = LowerConstant(CE->getOperand(0), AP);
225 const MCExpr *RHS = LowerConstant(CE->getOperand(1), AP);
226 switch (CE->getOpcode()) {
227 default:
228 llvm_unreachable("Unknown binary operator constant cast expr");
229 case Instruction::Add:
230 return MCBinaryExpr::CreateAdd(LHS, RHS, Ctx);
231 case Instruction::Sub:
232 return MCBinaryExpr::CreateSub(LHS, RHS, Ctx);
233 case Instruction::Mul:
234 return MCBinaryExpr::CreateMul(LHS, RHS, Ctx);
235 case Instruction::SDiv:
236 return MCBinaryExpr::CreateDiv(LHS, RHS, Ctx);
237 case Instruction::SRem:
238 return MCBinaryExpr::CreateMod(LHS, RHS, Ctx);
239 case Instruction::Shl:
240 return MCBinaryExpr::CreateShl(LHS, RHS, Ctx);
241 case Instruction::And:
242 return MCBinaryExpr::CreateAnd(LHS, RHS, Ctx);
243 case Instruction::Or:
244 return MCBinaryExpr::CreateOr(LHS, RHS, Ctx);
245 case Instruction::Xor:
246 return MCBinaryExpr::CreateXor(LHS, RHS, Ctx);
247 }
248 }
249 }
250 }
251
emitLineNumberAsDotLoc(const MachineInstr & MI)252 void NVPTXAsmPrinter::emitLineNumberAsDotLoc(const MachineInstr &MI) {
253 if (!EmitLineNumbers)
254 return;
255 if (ignoreLoc(MI))
256 return;
257
258 DebugLoc curLoc = MI.getDebugLoc();
259
260 if (prevDebugLoc.isUnknown() && curLoc.isUnknown())
261 return;
262
263 if (prevDebugLoc == curLoc)
264 return;
265
266 prevDebugLoc = curLoc;
267
268 if (curLoc.isUnknown())
269 return;
270
271 const MachineFunction *MF = MI.getParent()->getParent();
272 //const TargetMachine &TM = MF->getTarget();
273
274 const LLVMContext &ctx = MF->getFunction()->getContext();
275 DIScope Scope(curLoc.getScope(ctx));
276
277 assert((!Scope || Scope.isScope()) &&
278 "Scope of a DebugLoc should be null or a DIScope.");
279 if (!Scope)
280 return;
281
282 StringRef fileName(Scope.getFilename());
283 StringRef dirName(Scope.getDirectory());
284 SmallString<128> FullPathName = dirName;
285 if (!dirName.empty() && !sys::path::is_absolute(fileName)) {
286 sys::path::append(FullPathName, fileName);
287 fileName = FullPathName.str();
288 }
289
290 if (filenameMap.find(fileName.str()) == filenameMap.end())
291 return;
292
293 // Emit the line from the source file.
294 if (InterleaveSrc)
295 this->emitSrcInText(fileName.str(), curLoc.getLine());
296
297 std::stringstream temp;
298 temp << "\t.loc " << filenameMap[fileName.str()] << " " << curLoc.getLine()
299 << " " << curLoc.getCol();
300 OutStreamer.EmitRawText(Twine(temp.str().c_str()));
301 }
302
EmitInstruction(const MachineInstr * MI)303 void NVPTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
304 SmallString<128> Str;
305 raw_svector_ostream OS(Str);
306 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
307 emitLineNumberAsDotLoc(*MI);
308
309 MCInst Inst;
310 lowerToMCInst(MI, Inst);
311 OutStreamer.EmitInstruction(Inst);
312 }
313
lowerToMCInst(const MachineInstr * MI,MCInst & OutMI)314 void NVPTXAsmPrinter::lowerToMCInst(const MachineInstr *MI, MCInst &OutMI) {
315 OutMI.setOpcode(MI->getOpcode());
316
317 // Special: Do not mangle symbol operand of CALL_PROTOTYPE
318 if (MI->getOpcode() == NVPTX::CALL_PROTOTYPE) {
319 const MachineOperand &MO = MI->getOperand(0);
320 OutMI.addOperand(GetSymbolRef(MO,
321 OutContext.GetOrCreateSymbol(Twine(MO.getSymbolName()))));
322 return;
323 }
324
325 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
326 const MachineOperand &MO = MI->getOperand(i);
327
328 MCOperand MCOp;
329 if (lowerOperand(MO, MCOp))
330 OutMI.addOperand(MCOp);
331 }
332 }
333
lowerOperand(const MachineOperand & MO,MCOperand & MCOp)334 bool NVPTXAsmPrinter::lowerOperand(const MachineOperand &MO,
335 MCOperand &MCOp) {
336 switch (MO.getType()) {
337 default: llvm_unreachable("unknown operand type");
338 case MachineOperand::MO_Register:
339 MCOp = MCOperand::CreateReg(encodeVirtualRegister(MO.getReg()));
340 break;
341 case MachineOperand::MO_Immediate:
342 MCOp = MCOperand::CreateImm(MO.getImm());
343 break;
344 case MachineOperand::MO_MachineBasicBlock:
345 MCOp = MCOperand::CreateExpr(MCSymbolRefExpr::Create(
346 MO.getMBB()->getSymbol(), OutContext));
347 break;
348 case MachineOperand::MO_ExternalSymbol:
349 MCOp = GetSymbolRef(MO, GetExternalSymbolSymbol(MO.getSymbolName()));
350 break;
351 case MachineOperand::MO_GlobalAddress:
352 MCOp = GetSymbolRef(MO, getSymbol(MO.getGlobal()));
353 break;
354 case MachineOperand::MO_FPImmediate: {
355 const ConstantFP *Cnt = MO.getFPImm();
356 APFloat Val = Cnt->getValueAPF();
357
358 switch (Cnt->getType()->getTypeID()) {
359 default: report_fatal_error("Unsupported FP type"); break;
360 case Type::FloatTyID:
361 MCOp = MCOperand::CreateExpr(
362 NVPTXFloatMCExpr::CreateConstantFPSingle(Val, OutContext));
363 break;
364 case Type::DoubleTyID:
365 MCOp = MCOperand::CreateExpr(
366 NVPTXFloatMCExpr::CreateConstantFPDouble(Val, OutContext));
367 break;
368 }
369 break;
370 }
371 }
372 return true;
373 }
374
encodeVirtualRegister(unsigned Reg)375 unsigned NVPTXAsmPrinter::encodeVirtualRegister(unsigned Reg) {
376 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
377 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
378
379 DenseMap<unsigned, unsigned> &RegMap = VRegMapping[RC];
380 unsigned RegNum = RegMap[Reg];
381
382 // Encode the register class in the upper 4 bits
383 // Must be kept in sync with NVPTXInstPrinter::printRegName
384 unsigned Ret = 0;
385 if (RC == &NVPTX::Int1RegsRegClass) {
386 Ret = (1 << 28);
387 } else if (RC == &NVPTX::Int16RegsRegClass) {
388 Ret = (2 << 28);
389 } else if (RC == &NVPTX::Int32RegsRegClass) {
390 Ret = (3 << 28);
391 } else if (RC == &NVPTX::Int64RegsRegClass) {
392 Ret = (4 << 28);
393 } else if (RC == &NVPTX::Float32RegsRegClass) {
394 Ret = (5 << 28);
395 } else if (RC == &NVPTX::Float64RegsRegClass) {
396 Ret = (6 << 28);
397 } else {
398 report_fatal_error("Bad register class");
399 }
400
401 // Insert the vreg number
402 Ret |= (RegNum & 0x0FFFFFFF);
403 return Ret;
404 } else {
405 // Some special-use registers are actually physical registers.
406 // Encode this as the register class ID of 0 and the real register ID.
407 return Reg & 0x0FFFFFFF;
408 }
409 }
410
GetSymbolRef(const MachineOperand & MO,const MCSymbol * Symbol)411 MCOperand NVPTXAsmPrinter::GetSymbolRef(const MachineOperand &MO,
412 const MCSymbol *Symbol) {
413 const MCExpr *Expr;
414 Expr = MCSymbolRefExpr::Create(Symbol, MCSymbolRefExpr::VK_None,
415 OutContext);
416 return MCOperand::CreateExpr(Expr);
417 }
418
printReturnValStr(const Function * F,raw_ostream & O)419 void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
420 const DataLayout *TD = TM.getDataLayout();
421 const TargetLowering *TLI = TM.getTargetLowering();
422
423 Type *Ty = F->getReturnType();
424
425 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
426
427 if (Ty->getTypeID() == Type::VoidTyID)
428 return;
429
430 O << " (";
431
432 if (isABI) {
433 if (Ty->isPrimitiveType() || Ty->isIntegerTy()) {
434 unsigned size = 0;
435 if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty)) {
436 size = ITy->getBitWidth();
437 if (size < 32)
438 size = 32;
439 } else {
440 assert(Ty->isFloatingPointTy() && "Floating point type expected here");
441 size = Ty->getPrimitiveSizeInBits();
442 }
443
444 O << ".param .b" << size << " func_retval0";
445 } else if (isa<PointerType>(Ty)) {
446 O << ".param .b" << TLI->getPointerTy().getSizeInBits()
447 << " func_retval0";
448 } else {
449 if ((Ty->getTypeID() == Type::StructTyID) || isa<VectorType>(Ty)) {
450 SmallVector<EVT, 16> vtparts;
451 ComputeValueVTs(*TLI, Ty, vtparts);
452 unsigned totalsz = 0;
453 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
454 unsigned elems = 1;
455 EVT elemtype = vtparts[i];
456 if (vtparts[i].isVector()) {
457 elems = vtparts[i].getVectorNumElements();
458 elemtype = vtparts[i].getVectorElementType();
459 }
460 for (unsigned j = 0, je = elems; j != je; ++j) {
461 unsigned sz = elemtype.getSizeInBits();
462 if (elemtype.isInteger() && (sz < 8))
463 sz = 8;
464 totalsz += sz / 8;
465 }
466 }
467 unsigned retAlignment = 0;
468 if (!llvm::getAlign(*F, 0, retAlignment))
469 retAlignment = TD->getABITypeAlignment(Ty);
470 O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
471 << "]";
472 } else
473 assert(false && "Unknown return type");
474 }
475 } else {
476 SmallVector<EVT, 16> vtparts;
477 ComputeValueVTs(*TLI, Ty, vtparts);
478 unsigned idx = 0;
479 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
480 unsigned elems = 1;
481 EVT elemtype = vtparts[i];
482 if (vtparts[i].isVector()) {
483 elems = vtparts[i].getVectorNumElements();
484 elemtype = vtparts[i].getVectorElementType();
485 }
486
487 for (unsigned j = 0, je = elems; j != je; ++j) {
488 unsigned sz = elemtype.getSizeInBits();
489 if (elemtype.isInteger() && (sz < 32))
490 sz = 32;
491 O << ".reg .b" << sz << " func_retval" << idx;
492 if (j < je - 1)
493 O << ", ";
494 ++idx;
495 }
496 if (i < e - 1)
497 O << ", ";
498 }
499 }
500 O << ") ";
501 return;
502 }
503
printReturnValStr(const MachineFunction & MF,raw_ostream & O)504 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
505 raw_ostream &O) {
506 const Function *F = MF.getFunction();
507 printReturnValStr(F, O);
508 }
509
EmitFunctionEntryLabel()510 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
511 SmallString<128> Str;
512 raw_svector_ostream O(Str);
513
514 if (!GlobalsEmitted) {
515 emitGlobals(*MF->getFunction()->getParent());
516 GlobalsEmitted = true;
517 }
518
519 // Set up
520 MRI = &MF->getRegInfo();
521 F = MF->getFunction();
522 emitLinkageDirective(F, O);
523 if (llvm::isKernelFunction(*F))
524 O << ".entry ";
525 else {
526 O << ".func ";
527 printReturnValStr(*MF, O);
528 }
529
530 O << *CurrentFnSym;
531
532 emitFunctionParamList(*MF, O);
533
534 if (llvm::isKernelFunction(*F))
535 emitKernelFunctionDirectives(*F, O);
536
537 OutStreamer.EmitRawText(O.str());
538
539 prevDebugLoc = DebugLoc();
540 }
541
EmitFunctionBodyStart()542 void NVPTXAsmPrinter::EmitFunctionBodyStart() {
543 VRegMapping.clear();
544 OutStreamer.EmitRawText(StringRef("{\n"));
545 setAndEmitFunctionVirtualRegisters(*MF);
546
547 SmallString<128> Str;
548 raw_svector_ostream O(Str);
549 emitDemotedVars(MF->getFunction(), O);
550 OutStreamer.EmitRawText(O.str());
551 }
552
EmitFunctionBodyEnd()553 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
554 OutStreamer.EmitRawText(StringRef("}\n"));
555 VRegMapping.clear();
556 }
557
emitImplicitDef(const MachineInstr * MI) const558 void NVPTXAsmPrinter::emitImplicitDef(const MachineInstr *MI) const {
559 unsigned RegNo = MI->getOperand(0).getReg();
560 const TargetRegisterInfo *TRI = TM.getRegisterInfo();
561 if (TRI->isVirtualRegister(RegNo)) {
562 OutStreamer.AddComment(Twine("implicit-def: ") +
563 getVirtualRegisterName(RegNo));
564 } else {
565 OutStreamer.AddComment(Twine("implicit-def: ") +
566 TM.getRegisterInfo()->getName(RegNo));
567 }
568 OutStreamer.AddBlankLine();
569 }
570
emitKernelFunctionDirectives(const Function & F,raw_ostream & O) const571 void NVPTXAsmPrinter::emitKernelFunctionDirectives(const Function &F,
572 raw_ostream &O) const {
573 // If the NVVM IR has some of reqntid* specified, then output
574 // the reqntid directive, and set the unspecified ones to 1.
575 // If none of reqntid* is specified, don't output reqntid directive.
576 unsigned reqntidx, reqntidy, reqntidz;
577 bool specified = false;
578 if (llvm::getReqNTIDx(F, reqntidx) == false)
579 reqntidx = 1;
580 else
581 specified = true;
582 if (llvm::getReqNTIDy(F, reqntidy) == false)
583 reqntidy = 1;
584 else
585 specified = true;
586 if (llvm::getReqNTIDz(F, reqntidz) == false)
587 reqntidz = 1;
588 else
589 specified = true;
590
591 if (specified)
592 O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
593 << "\n";
594
595 // If the NVVM IR has some of maxntid* specified, then output
596 // the maxntid directive, and set the unspecified ones to 1.
597 // If none of maxntid* is specified, don't output maxntid directive.
598 unsigned maxntidx, maxntidy, maxntidz;
599 specified = false;
600 if (llvm::getMaxNTIDx(F, maxntidx) == false)
601 maxntidx = 1;
602 else
603 specified = true;
604 if (llvm::getMaxNTIDy(F, maxntidy) == false)
605 maxntidy = 1;
606 else
607 specified = true;
608 if (llvm::getMaxNTIDz(F, maxntidz) == false)
609 maxntidz = 1;
610 else
611 specified = true;
612
613 if (specified)
614 O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
615 << "\n";
616
617 unsigned mincta;
618 if (llvm::getMinCTASm(F, mincta))
619 O << ".minnctapersm " << mincta << "\n";
620 }
621
622 std::string
getVirtualRegisterName(unsigned Reg) const623 NVPTXAsmPrinter::getVirtualRegisterName(unsigned Reg) const {
624 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
625
626 std::string Name;
627 raw_string_ostream NameStr(Name);
628
629 VRegRCMap::const_iterator I = VRegMapping.find(RC);
630 assert(I != VRegMapping.end() && "Bad register class");
631 const DenseMap<unsigned, unsigned> &RegMap = I->second;
632
633 VRegMap::const_iterator VI = RegMap.find(Reg);
634 assert(VI != RegMap.end() && "Bad virtual register");
635 unsigned MappedVR = VI->second;
636
637 NameStr << getNVPTXRegClassStr(RC) << MappedVR;
638
639 NameStr.flush();
640 return Name;
641 }
642
emitVirtualRegister(unsigned int vr,raw_ostream & O)643 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr,
644 raw_ostream &O) {
645 O << getVirtualRegisterName(vr);
646 }
647
printVecModifiedImmediate(const MachineOperand & MO,const char * Modifier,raw_ostream & O)648 void NVPTXAsmPrinter::printVecModifiedImmediate(
649 const MachineOperand &MO, const char *Modifier, raw_ostream &O) {
650 static const char vecelem[] = { '0', '1', '2', '3', '0', '1', '2', '3' };
651 int Imm = (int) MO.getImm();
652 if (0 == strcmp(Modifier, "vecelem"))
653 O << "_" << vecelem[Imm];
654 else if (0 == strcmp(Modifier, "vecv4comm1")) {
655 if ((Imm < 0) || (Imm > 3))
656 O << "//";
657 } else if (0 == strcmp(Modifier, "vecv4comm2")) {
658 if ((Imm < 4) || (Imm > 7))
659 O << "//";
660 } else if (0 == strcmp(Modifier, "vecv4pos")) {
661 if (Imm < 0)
662 Imm = 0;
663 O << "_" << vecelem[Imm % 4];
664 } else if (0 == strcmp(Modifier, "vecv2comm1")) {
665 if ((Imm < 0) || (Imm > 1))
666 O << "//";
667 } else if (0 == strcmp(Modifier, "vecv2comm2")) {
668 if ((Imm < 2) || (Imm > 3))
669 O << "//";
670 } else if (0 == strcmp(Modifier, "vecv2pos")) {
671 if (Imm < 0)
672 Imm = 0;
673 O << "_" << vecelem[Imm % 2];
674 } else
675 llvm_unreachable("Unknown Modifier on immediate operand");
676 }
677
678
679
emitDeclaration(const Function * F,raw_ostream & O)680 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
681
682 emitLinkageDirective(F, O);
683 if (llvm::isKernelFunction(*F))
684 O << ".entry ";
685 else
686 O << ".func ";
687 printReturnValStr(F, O);
688 O << *getSymbol(F) << "\n";
689 emitFunctionParamList(F, O);
690 O << ";\n";
691 }
692
usedInGlobalVarDef(const Constant * C)693 static bool usedInGlobalVarDef(const Constant *C) {
694 if (!C)
695 return false;
696
697 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
698 if (GV->getName().str() == "llvm.used")
699 return false;
700 return true;
701 }
702
703 for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
704 ui != ue; ++ui) {
705 const Constant *C = dyn_cast<Constant>(*ui);
706 if (usedInGlobalVarDef(C))
707 return true;
708 }
709 return false;
710 }
711
usedInOneFunc(const User * U,Function const * & oneFunc)712 static bool usedInOneFunc(const User *U, Function const *&oneFunc) {
713 if (const GlobalVariable *othergv = dyn_cast<GlobalVariable>(U)) {
714 if (othergv->getName().str() == "llvm.used")
715 return true;
716 }
717
718 if (const Instruction *instr = dyn_cast<Instruction>(U)) {
719 if (instr->getParent() && instr->getParent()->getParent()) {
720 const Function *curFunc = instr->getParent()->getParent();
721 if (oneFunc && (curFunc != oneFunc))
722 return false;
723 oneFunc = curFunc;
724 return true;
725 } else
726 return false;
727 }
728
729 if (const MDNode *md = dyn_cast<MDNode>(U))
730 if (md->hasName() && ((md->getName().str() == "llvm.dbg.gv") ||
731 (md->getName().str() == "llvm.dbg.sp")))
732 return true;
733
734 for (User::const_use_iterator ui = U->use_begin(), ue = U->use_end();
735 ui != ue; ++ui) {
736 if (usedInOneFunc(*ui, oneFunc) == false)
737 return false;
738 }
739 return true;
740 }
741
742 /* Find out if a global variable can be demoted to local scope.
743 * Currently, this is valid for CUDA shared variables, which have local
744 * scope and global lifetime. So the conditions to check are :
745 * 1. Is the global variable in shared address space?
746 * 2. Does it have internal linkage?
747 * 3. Is the global variable referenced only in one function?
748 */
canDemoteGlobalVar(const GlobalVariable * gv,Function const * & f)749 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
750 if (gv->hasInternalLinkage() == false)
751 return false;
752 const PointerType *Pty = gv->getType();
753 if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
754 return false;
755
756 const Function *oneFunc = 0;
757
758 bool flag = usedInOneFunc(gv, oneFunc);
759 if (flag == false)
760 return false;
761 if (!oneFunc)
762 return false;
763 f = oneFunc;
764 return true;
765 }
766
useFuncSeen(const Constant * C,llvm::DenseMap<const Function *,bool> & seenMap)767 static bool useFuncSeen(const Constant *C,
768 llvm::DenseMap<const Function *, bool> &seenMap) {
769 for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
770 ui != ue; ++ui) {
771 if (const Constant *cu = dyn_cast<Constant>(*ui)) {
772 if (useFuncSeen(cu, seenMap))
773 return true;
774 } else if (const Instruction *I = dyn_cast<Instruction>(*ui)) {
775 const BasicBlock *bb = I->getParent();
776 if (!bb)
777 continue;
778 const Function *caller = bb->getParent();
779 if (!caller)
780 continue;
781 if (seenMap.find(caller) != seenMap.end())
782 return true;
783 }
784 }
785 return false;
786 }
787
emitDeclarations(const Module & M,raw_ostream & O)788 void NVPTXAsmPrinter::emitDeclarations(const Module &M, raw_ostream &O) {
789 llvm::DenseMap<const Function *, bool> seenMap;
790 for (Module::const_iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
791 const Function *F = FI;
792
793 if (F->isDeclaration()) {
794 if (F->use_empty())
795 continue;
796 if (F->getIntrinsicID())
797 continue;
798 emitDeclaration(F, O);
799 continue;
800 }
801 for (Value::const_use_iterator iter = F->use_begin(),
802 iterEnd = F->use_end();
803 iter != iterEnd; ++iter) {
804 if (const Constant *C = dyn_cast<Constant>(*iter)) {
805 if (usedInGlobalVarDef(C)) {
806 // The use is in the initialization of a global variable
807 // that is a function pointer, so print a declaration
808 // for the original function
809 emitDeclaration(F, O);
810 break;
811 }
812 // Emit a declaration of this function if the function that
813 // uses this constant expr has already been seen.
814 if (useFuncSeen(C, seenMap)) {
815 emitDeclaration(F, O);
816 break;
817 }
818 }
819
820 if (!isa<Instruction>(*iter))
821 continue;
822 const Instruction *instr = cast<Instruction>(*iter);
823 const BasicBlock *bb = instr->getParent();
824 if (!bb)
825 continue;
826 const Function *caller = bb->getParent();
827 if (!caller)
828 continue;
829
830 // If a caller has already been seen, then the caller is
831 // appearing in the module before the callee. so print out
832 // a declaration for the callee.
833 if (seenMap.find(caller) != seenMap.end()) {
834 emitDeclaration(F, O);
835 break;
836 }
837 }
838 seenMap[F] = true;
839 }
840 }
841
recordAndEmitFilenames(Module & M)842 void NVPTXAsmPrinter::recordAndEmitFilenames(Module &M) {
843 DebugInfoFinder DbgFinder;
844 DbgFinder.processModule(M);
845
846 unsigned i = 1;
847 for (DebugInfoFinder::iterator I = DbgFinder.compile_unit_begin(),
848 E = DbgFinder.compile_unit_end();
849 I != E; ++I) {
850 DICompileUnit DIUnit(*I);
851 StringRef Filename(DIUnit.getFilename());
852 StringRef Dirname(DIUnit.getDirectory());
853 SmallString<128> FullPathName = Dirname;
854 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
855 sys::path::append(FullPathName, Filename);
856 Filename = FullPathName.str();
857 }
858 if (filenameMap.find(Filename.str()) != filenameMap.end())
859 continue;
860 filenameMap[Filename.str()] = i;
861 OutStreamer.EmitDwarfFileDirective(i, "", Filename.str());
862 ++i;
863 }
864
865 for (DebugInfoFinder::iterator I = DbgFinder.subprogram_begin(),
866 E = DbgFinder.subprogram_end();
867 I != E; ++I) {
868 DISubprogram SP(*I);
869 StringRef Filename(SP.getFilename());
870 StringRef Dirname(SP.getDirectory());
871 SmallString<128> FullPathName = Dirname;
872 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
873 sys::path::append(FullPathName, Filename);
874 Filename = FullPathName.str();
875 }
876 if (filenameMap.find(Filename.str()) != filenameMap.end())
877 continue;
878 filenameMap[Filename.str()] = i;
879 ++i;
880 }
881 }
882
doInitialization(Module & M)883 bool NVPTXAsmPrinter::doInitialization(Module &M) {
884
885 SmallString<128> Str1;
886 raw_svector_ostream OS1(Str1);
887
888 MMI = getAnalysisIfAvailable<MachineModuleInfo>();
889 MMI->AnalyzeModule(M);
890
891 // We need to call the parent's one explicitly.
892 //bool Result = AsmPrinter::doInitialization(M);
893
894 // Initialize TargetLoweringObjectFile.
895 const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
896 .Initialize(OutContext, TM);
897
898 Mang = new Mangler(&TM);
899
900 // Emit header before any dwarf directives are emitted below.
901 emitHeader(M, OS1);
902 OutStreamer.EmitRawText(OS1.str());
903
904 // Already commented out
905 //bool Result = AsmPrinter::doInitialization(M);
906
907 // Emit module-level inline asm if it exists.
908 if (!M.getModuleInlineAsm().empty()) {
909 OutStreamer.AddComment("Start of file scope inline assembly");
910 OutStreamer.AddBlankLine();
911 OutStreamer.EmitRawText(StringRef(M.getModuleInlineAsm()));
912 OutStreamer.AddBlankLine();
913 OutStreamer.AddComment("End of file scope inline assembly");
914 OutStreamer.AddBlankLine();
915 }
916
917 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
918 recordAndEmitFilenames(M);
919
920 GlobalsEmitted = false;
921
922 return false; // success
923 }
924
emitGlobals(const Module & M)925 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
926 SmallString<128> Str2;
927 raw_svector_ostream OS2(Str2);
928
929 emitDeclarations(M, OS2);
930
931 // As ptxas does not support forward references of globals, we need to first
932 // sort the list of module-level globals in def-use order. We visit each
933 // global variable in order, and ensure that we emit it *after* its dependent
934 // globals. We use a little extra memory maintaining both a set and a list to
935 // have fast searches while maintaining a strict ordering.
936 SmallVector<const GlobalVariable *, 8> Globals;
937 DenseSet<const GlobalVariable *> GVVisited;
938 DenseSet<const GlobalVariable *> GVVisiting;
939
940 // Visit each global variable, in order
941 for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
942 I != E; ++I)
943 VisitGlobalVariableForEmission(I, Globals, GVVisited, GVVisiting);
944
945 assert(GVVisited.size() == M.getGlobalList().size() &&
946 "Missed a global variable");
947 assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
948
949 // Print out module-level global variables in proper order
950 for (unsigned i = 0, e = Globals.size(); i != e; ++i)
951 printModuleLevelGV(Globals[i], OS2);
952
953 OS2 << '\n';
954
955 OutStreamer.EmitRawText(OS2.str());
956 }
957
emitHeader(Module & M,raw_ostream & O)958 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O) {
959 O << "//\n";
960 O << "// Generated by LLVM NVPTX Back-End\n";
961 O << "//\n";
962 O << "\n";
963
964 unsigned PTXVersion = nvptxSubtarget.getPTXVersion();
965 O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
966
967 O << ".target ";
968 O << nvptxSubtarget.getTargetName();
969
970 if (nvptxSubtarget.getDrvInterface() == NVPTX::NVCL)
971 O << ", texmode_independent";
972 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
973 if (!nvptxSubtarget.hasDouble())
974 O << ", map_f64_to_f32";
975 }
976
977 if (MAI->doesSupportDebugInformation())
978 O << ", debug";
979
980 O << "\n";
981
982 O << ".address_size ";
983 if (nvptxSubtarget.is64Bit())
984 O << "64";
985 else
986 O << "32";
987 O << "\n";
988
989 O << "\n";
990 }
991
doFinalization(Module & M)992 bool NVPTXAsmPrinter::doFinalization(Module &M) {
993
994 // If we did not emit any functions, then the global declarations have not
995 // yet been emitted.
996 if (!GlobalsEmitted) {
997 emitGlobals(M);
998 GlobalsEmitted = true;
999 }
1000
1001 // XXX Temproarily remove global variables so that doFinalization() will not
1002 // emit them again (global variables are emitted at beginning).
1003
1004 Module::GlobalListType &global_list = M.getGlobalList();
1005 int i, n = global_list.size();
1006 GlobalVariable **gv_array = new GlobalVariable *[n];
1007
1008 // first, back-up GlobalVariable in gv_array
1009 i = 0;
1010 for (Module::global_iterator I = global_list.begin(), E = global_list.end();
1011 I != E; ++I)
1012 gv_array[i++] = &*I;
1013
1014 // second, empty global_list
1015 while (!global_list.empty())
1016 global_list.remove(global_list.begin());
1017
1018 // call doFinalization
1019 bool ret = AsmPrinter::doFinalization(M);
1020
1021 // now we restore global variables
1022 for (i = 0; i < n; i++)
1023 global_list.insert(global_list.end(), gv_array[i]);
1024
1025 delete[] gv_array;
1026 return ret;
1027
1028 //bool Result = AsmPrinter::doFinalization(M);
1029 // Instead of calling the parents doFinalization, we may
1030 // clone parents doFinalization and customize here.
1031 // Currently, we if NVISA out the EmitGlobals() in
1032 // parent's doFinalization, which is too intrusive.
1033 //
1034 // Same for the doInitialization.
1035 //return Result;
1036 }
1037
1038 // This function emits appropriate linkage directives for
1039 // functions and global variables.
1040 //
1041 // extern function declaration -> .extern
1042 // extern function definition -> .visible
1043 // external global variable with init -> .visible
1044 // external without init -> .extern
1045 // appending -> not allowed, assert.
1046
emitLinkageDirective(const GlobalValue * V,raw_ostream & O)1047 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
1048 raw_ostream &O) {
1049 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
1050 if (V->hasExternalLinkage()) {
1051 if (isa<GlobalVariable>(V)) {
1052 const GlobalVariable *GVar = cast<GlobalVariable>(V);
1053 if (GVar) {
1054 if (GVar->hasInitializer())
1055 O << ".visible ";
1056 else
1057 O << ".extern ";
1058 }
1059 } else if (V->isDeclaration())
1060 O << ".extern ";
1061 else
1062 O << ".visible ";
1063 } else if (V->hasAppendingLinkage()) {
1064 std::string msg;
1065 msg.append("Error: ");
1066 msg.append("Symbol ");
1067 if (V->hasName())
1068 msg.append(V->getName().str());
1069 msg.append("has unsupported appending linkage type");
1070 llvm_unreachable(msg.c_str());
1071 }
1072 }
1073 }
1074
printModuleLevelGV(const GlobalVariable * GVar,raw_ostream & O,bool processDemoted)1075 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
1076 raw_ostream &O,
1077 bool processDemoted) {
1078
1079 // Skip meta data
1080 if (GVar->hasSection()) {
1081 if (GVar->getSection() == "llvm.metadata")
1082 return;
1083 }
1084
1085 const DataLayout *TD = TM.getDataLayout();
1086
1087 // GlobalVariables are always constant pointers themselves.
1088 const PointerType *PTy = GVar->getType();
1089 Type *ETy = PTy->getElementType();
1090
1091 if (GVar->hasExternalLinkage()) {
1092 if (GVar->hasInitializer())
1093 O << ".visible ";
1094 else
1095 O << ".extern ";
1096 }
1097
1098 if (llvm::isTexture(*GVar)) {
1099 O << ".global .texref " << llvm::getTextureName(*GVar) << ";\n";
1100 return;
1101 }
1102
1103 if (llvm::isSurface(*GVar)) {
1104 O << ".global .surfref " << llvm::getSurfaceName(*GVar) << ";\n";
1105 return;
1106 }
1107
1108 if (GVar->isDeclaration()) {
1109 // (extern) declarations, no definition or initializer
1110 // Currently the only known declaration is for an automatic __local
1111 // (.shared) promoted to global.
1112 emitPTXGlobalVariable(GVar, O);
1113 O << ";\n";
1114 return;
1115 }
1116
1117 if (llvm::isSampler(*GVar)) {
1118 O << ".global .samplerref " << llvm::getSamplerName(*GVar);
1119
1120 const Constant *Initializer = NULL;
1121 if (GVar->hasInitializer())
1122 Initializer = GVar->getInitializer();
1123 const ConstantInt *CI = NULL;
1124 if (Initializer)
1125 CI = dyn_cast<ConstantInt>(Initializer);
1126 if (CI) {
1127 unsigned sample = CI->getZExtValue();
1128
1129 O << " = { ";
1130
1131 for (int i = 0,
1132 addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1133 i < 3; i++) {
1134 O << "addr_mode_" << i << " = ";
1135 switch (addr) {
1136 case 0:
1137 O << "wrap";
1138 break;
1139 case 1:
1140 O << "clamp_to_border";
1141 break;
1142 case 2:
1143 O << "clamp_to_edge";
1144 break;
1145 case 3:
1146 O << "wrap";
1147 break;
1148 case 4:
1149 O << "mirror";
1150 break;
1151 }
1152 O << ", ";
1153 }
1154 O << "filter_mode = ";
1155 switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1156 case 0:
1157 O << "nearest";
1158 break;
1159 case 1:
1160 O << "linear";
1161 break;
1162 case 2:
1163 assert(0 && "Anisotropic filtering is not supported");
1164 default:
1165 O << "nearest";
1166 break;
1167 }
1168 if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1169 O << ", force_unnormalized_coords = 1";
1170 }
1171 O << " }";
1172 }
1173
1174 O << ";\n";
1175 return;
1176 }
1177
1178 if (GVar->hasPrivateLinkage()) {
1179
1180 if (!strncmp(GVar->getName().data(), "unrollpragma", 12))
1181 return;
1182
1183 // FIXME - need better way (e.g. Metadata) to avoid generating this global
1184 if (!strncmp(GVar->getName().data(), "filename", 8))
1185 return;
1186 if (GVar->use_empty())
1187 return;
1188 }
1189
1190 const Function *demotedFunc = 0;
1191 if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
1192 O << "// " << GVar->getName().str() << " has been demoted\n";
1193 if (localDecls.find(demotedFunc) != localDecls.end())
1194 localDecls[demotedFunc].push_back(GVar);
1195 else {
1196 std::vector<const GlobalVariable *> temp;
1197 temp.push_back(GVar);
1198 localDecls[demotedFunc] = temp;
1199 }
1200 return;
1201 }
1202
1203 O << ".";
1204 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1205 if (GVar->getAlignment() == 0)
1206 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1207 else
1208 O << " .align " << GVar->getAlignment();
1209
1210 if (ETy->isPrimitiveType() || ETy->isIntegerTy() || isa<PointerType>(ETy)) {
1211 O << " .";
1212 // Special case: ABI requires that we use .u8 for predicates
1213 if (ETy->isIntegerTy(1))
1214 O << "u8";
1215 else
1216 O << getPTXFundamentalTypeStr(ETy, false);
1217 O << " ";
1218 O << *getSymbol(GVar);
1219
1220 // Ptx allows variable initilization only for constant and global state
1221 // spaces.
1222 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1223 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1224 GVar->hasInitializer()) {
1225 const Constant *Initializer = GVar->getInitializer();
1226 if (!Initializer->isNullValue()) {
1227 O << " = ";
1228 printScalarConstant(Initializer, O);
1229 }
1230 }
1231 } else {
1232 unsigned int ElementSize = 0;
1233
1234 // Although PTX has direct support for struct type and array type and
1235 // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1236 // targets that support these high level field accesses. Structs, arrays
1237 // and vectors are lowered into arrays of bytes.
1238 switch (ETy->getTypeID()) {
1239 case Type::StructTyID:
1240 case Type::ArrayTyID:
1241 case Type::VectorTyID:
1242 ElementSize = TD->getTypeStoreSize(ETy);
1243 // Ptx allows variable initilization only for constant and
1244 // global state spaces.
1245 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1246 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1247 GVar->hasInitializer()) {
1248 const Constant *Initializer = GVar->getInitializer();
1249 if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1250 AggBuffer aggBuffer(ElementSize, O, *this);
1251 bufferAggregateConstant(Initializer, &aggBuffer);
1252 if (aggBuffer.numSymbols) {
1253 if (nvptxSubtarget.is64Bit()) {
1254 O << " .u64 " << *getSymbol(GVar) << "[";
1255 O << ElementSize / 8;
1256 } else {
1257 O << " .u32 " << *getSymbol(GVar) << "[";
1258 O << ElementSize / 4;
1259 }
1260 O << "]";
1261 } else {
1262 O << " .b8 " << *getSymbol(GVar) << "[";
1263 O << ElementSize;
1264 O << "]";
1265 }
1266 O << " = {";
1267 aggBuffer.print();
1268 O << "}";
1269 } else {
1270 O << " .b8 " << *getSymbol(GVar);
1271 if (ElementSize) {
1272 O << "[";
1273 O << ElementSize;
1274 O << "]";
1275 }
1276 }
1277 } else {
1278 O << " .b8 " << *getSymbol(GVar);
1279 if (ElementSize) {
1280 O << "[";
1281 O << ElementSize;
1282 O << "]";
1283 }
1284 }
1285 break;
1286 default:
1287 assert(0 && "type not supported yet");
1288 }
1289
1290 }
1291 O << ";\n";
1292 }
1293
emitDemotedVars(const Function * f,raw_ostream & O)1294 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1295 if (localDecls.find(f) == localDecls.end())
1296 return;
1297
1298 std::vector<const GlobalVariable *> &gvars = localDecls[f];
1299
1300 for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1301 O << "\t// demoted variable\n\t";
1302 printModuleLevelGV(gvars[i], O, true);
1303 }
1304 }
1305
emitPTXAddressSpace(unsigned int AddressSpace,raw_ostream & O) const1306 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1307 raw_ostream &O) const {
1308 switch (AddressSpace) {
1309 case llvm::ADDRESS_SPACE_LOCAL:
1310 O << "local";
1311 break;
1312 case llvm::ADDRESS_SPACE_GLOBAL:
1313 O << "global";
1314 break;
1315 case llvm::ADDRESS_SPACE_CONST:
1316 O << "const";
1317 break;
1318 case llvm::ADDRESS_SPACE_SHARED:
1319 O << "shared";
1320 break;
1321 default:
1322 report_fatal_error("Bad address space found while emitting PTX");
1323 break;
1324 }
1325 }
1326
1327 std::string
getPTXFundamentalTypeStr(const Type * Ty,bool useB4PTR) const1328 NVPTXAsmPrinter::getPTXFundamentalTypeStr(const Type *Ty, bool useB4PTR) const {
1329 switch (Ty->getTypeID()) {
1330 default:
1331 llvm_unreachable("unexpected type");
1332 break;
1333 case Type::IntegerTyID: {
1334 unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1335 if (NumBits == 1)
1336 return "pred";
1337 else if (NumBits <= 64) {
1338 std::string name = "u";
1339 return name + utostr(NumBits);
1340 } else {
1341 llvm_unreachable("Integer too large");
1342 break;
1343 }
1344 break;
1345 }
1346 case Type::FloatTyID:
1347 return "f32";
1348 case Type::DoubleTyID:
1349 return "f64";
1350 case Type::PointerTyID:
1351 if (nvptxSubtarget.is64Bit())
1352 if (useB4PTR)
1353 return "b64";
1354 else
1355 return "u64";
1356 else if (useB4PTR)
1357 return "b32";
1358 else
1359 return "u32";
1360 }
1361 llvm_unreachable("unexpected type");
1362 return NULL;
1363 }
1364
emitPTXGlobalVariable(const GlobalVariable * GVar,raw_ostream & O)1365 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1366 raw_ostream &O) {
1367
1368 const DataLayout *TD = TM.getDataLayout();
1369
1370 // GlobalVariables are always constant pointers themselves.
1371 const PointerType *PTy = GVar->getType();
1372 Type *ETy = PTy->getElementType();
1373
1374 O << ".";
1375 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1376 if (GVar->getAlignment() == 0)
1377 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1378 else
1379 O << " .align " << GVar->getAlignment();
1380
1381 if (ETy->isPrimitiveType() || ETy->isIntegerTy() || isa<PointerType>(ETy)) {
1382 O << " .";
1383 O << getPTXFundamentalTypeStr(ETy);
1384 O << " ";
1385 O << *getSymbol(GVar);
1386 return;
1387 }
1388
1389 int64_t ElementSize = 0;
1390
1391 // Although PTX has direct support for struct type and array type and LLVM IR
1392 // is very similar to PTX, the LLVM CodeGen does not support for targets that
1393 // support these high level field accesses. Structs and arrays are lowered
1394 // into arrays of bytes.
1395 switch (ETy->getTypeID()) {
1396 case Type::StructTyID:
1397 case Type::ArrayTyID:
1398 case Type::VectorTyID:
1399 ElementSize = TD->getTypeStoreSize(ETy);
1400 O << " .b8 " << *getSymbol(GVar) << "[";
1401 if (ElementSize) {
1402 O << itostr(ElementSize);
1403 }
1404 O << "]";
1405 break;
1406 default:
1407 assert(0 && "type not supported yet");
1408 }
1409 return;
1410 }
1411
getOpenCLAlignment(const DataLayout * TD,Type * Ty)1412 static unsigned int getOpenCLAlignment(const DataLayout *TD, Type *Ty) {
1413 if (Ty->isPrimitiveType() || Ty->isIntegerTy() || isa<PointerType>(Ty))
1414 return TD->getPrefTypeAlignment(Ty);
1415
1416 const ArrayType *ATy = dyn_cast<ArrayType>(Ty);
1417 if (ATy)
1418 return getOpenCLAlignment(TD, ATy->getElementType());
1419
1420 const VectorType *VTy = dyn_cast<VectorType>(Ty);
1421 if (VTy) {
1422 Type *ETy = VTy->getElementType();
1423 unsigned int numE = VTy->getNumElements();
1424 unsigned int alignE = TD->getPrefTypeAlignment(ETy);
1425 if (numE == 3)
1426 return 4 * alignE;
1427 else
1428 return numE * alignE;
1429 }
1430
1431 const StructType *STy = dyn_cast<StructType>(Ty);
1432 if (STy) {
1433 unsigned int alignStruct = 1;
1434 // Go through each element of the struct and find the
1435 // largest alignment.
1436 for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1437 Type *ETy = STy->getElementType(i);
1438 unsigned int align = getOpenCLAlignment(TD, ETy);
1439 if (align > alignStruct)
1440 alignStruct = align;
1441 }
1442 return alignStruct;
1443 }
1444
1445 const FunctionType *FTy = dyn_cast<FunctionType>(Ty);
1446 if (FTy)
1447 return TD->getPointerPrefAlignment();
1448 return TD->getPrefTypeAlignment(Ty);
1449 }
1450
printParamName(Function::const_arg_iterator I,int paramIndex,raw_ostream & O)1451 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1452 int paramIndex, raw_ostream &O) {
1453 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1454 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA))
1455 O << *getSymbol(I->getParent()) << "_param_" << paramIndex;
1456 else {
1457 std::string argName = I->getName();
1458 const char *p = argName.c_str();
1459 while (*p) {
1460 if (*p == '.')
1461 O << "_";
1462 else
1463 O << *p;
1464 p++;
1465 }
1466 }
1467 }
1468
printParamName(int paramIndex,raw_ostream & O)1469 void NVPTXAsmPrinter::printParamName(int paramIndex, raw_ostream &O) {
1470 Function::const_arg_iterator I, E;
1471 int i = 0;
1472
1473 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1474 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)) {
1475 O << *CurrentFnSym << "_param_" << paramIndex;
1476 return;
1477 }
1478
1479 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, i++) {
1480 if (i == paramIndex) {
1481 printParamName(I, paramIndex, O);
1482 return;
1483 }
1484 }
1485 llvm_unreachable("paramIndex out of bound");
1486 }
1487
emitFunctionParamList(const Function * F,raw_ostream & O)1488 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1489 const DataLayout *TD = TM.getDataLayout();
1490 const AttributeSet &PAL = F->getAttributes();
1491 const TargetLowering *TLI = TM.getTargetLowering();
1492 Function::const_arg_iterator I, E;
1493 unsigned paramIndex = 0;
1494 bool first = true;
1495 bool isKernelFunc = llvm::isKernelFunction(*F);
1496 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
1497 MVT thePointerTy = TLI->getPointerTy();
1498
1499 O << "(\n";
1500
1501 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1502 Type *Ty = I->getType();
1503
1504 if (!first)
1505 O << ",\n";
1506
1507 first = false;
1508
1509 // Handle image/sampler parameters
1510 if (llvm::isSampler(*I) || llvm::isImage(*I)) {
1511 if (llvm::isImage(*I)) {
1512 std::string sname = I->getName();
1513 if (llvm::isImageWriteOnly(*I))
1514 O << "\t.param .surfref " << *getSymbol(F) << "_param_"
1515 << paramIndex;
1516 else // Default image is read_only
1517 O << "\t.param .texref " << *getSymbol(F) << "_param_"
1518 << paramIndex;
1519 } else // Should be llvm::isSampler(*I)
1520 O << "\t.param .samplerref " << *getSymbol(F) << "_param_"
1521 << paramIndex;
1522 continue;
1523 }
1524
1525 if (PAL.hasAttribute(paramIndex + 1, Attribute::ByVal) == false) {
1526 if (Ty->isVectorTy()) {
1527 // Just print .param .b8 .align <a> .param[size];
1528 // <a> = PAL.getparamalignment
1529 // size = typeallocsize of element type
1530 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1531 if (align == 0)
1532 align = TD->getABITypeAlignment(Ty);
1533
1534 unsigned sz = TD->getTypeAllocSize(Ty);
1535 O << "\t.param .align " << align << " .b8 ";
1536 printParamName(I, paramIndex, O);
1537 O << "[" << sz << "]";
1538
1539 continue;
1540 }
1541 // Just a scalar
1542 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1543 if (isKernelFunc) {
1544 if (PTy) {
1545 // Special handling for pointer arguments to kernel
1546 O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1547
1548 if (nvptxSubtarget.getDrvInterface() != NVPTX::CUDA) {
1549 Type *ETy = PTy->getElementType();
1550 int addrSpace = PTy->getAddressSpace();
1551 switch (addrSpace) {
1552 default:
1553 O << ".ptr ";
1554 break;
1555 case llvm::ADDRESS_SPACE_CONST:
1556 O << ".ptr .const ";
1557 break;
1558 case llvm::ADDRESS_SPACE_SHARED:
1559 O << ".ptr .shared ";
1560 break;
1561 case llvm::ADDRESS_SPACE_GLOBAL:
1562 O << ".ptr .global ";
1563 break;
1564 }
1565 O << ".align " << (int) getOpenCLAlignment(TD, ETy) << " ";
1566 }
1567 printParamName(I, paramIndex, O);
1568 continue;
1569 }
1570
1571 // non-pointer scalar to kernel func
1572 O << "\t.param .";
1573 // Special case: predicate operands become .u8 types
1574 if (Ty->isIntegerTy(1))
1575 O << "u8";
1576 else
1577 O << getPTXFundamentalTypeStr(Ty);
1578 O << " ";
1579 printParamName(I, paramIndex, O);
1580 continue;
1581 }
1582 // Non-kernel function, just print .param .b<size> for ABI
1583 // and .reg .b<size> for non ABY
1584 unsigned sz = 0;
1585 if (isa<IntegerType>(Ty)) {
1586 sz = cast<IntegerType>(Ty)->getBitWidth();
1587 if (sz < 32)
1588 sz = 32;
1589 } else if (isa<PointerType>(Ty))
1590 sz = thePointerTy.getSizeInBits();
1591 else
1592 sz = Ty->getPrimitiveSizeInBits();
1593 if (isABI)
1594 O << "\t.param .b" << sz << " ";
1595 else
1596 O << "\t.reg .b" << sz << " ";
1597 printParamName(I, paramIndex, O);
1598 continue;
1599 }
1600
1601 // param has byVal attribute. So should be a pointer
1602 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1603 assert(PTy && "Param with byval attribute should be a pointer type");
1604 Type *ETy = PTy->getElementType();
1605
1606 if (isABI || isKernelFunc) {
1607 // Just print .param .b8 .align <a> .param[size];
1608 // <a> = PAL.getparamalignment
1609 // size = typeallocsize of element type
1610 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1611 if (align == 0)
1612 align = TD->getABITypeAlignment(ETy);
1613
1614 unsigned sz = TD->getTypeAllocSize(ETy);
1615 O << "\t.param .align " << align << " .b8 ";
1616 printParamName(I, paramIndex, O);
1617 O << "[" << sz << "]";
1618 continue;
1619 } else {
1620 // Split the ETy into constituent parts and
1621 // print .param .b<size> <name> for each part.
1622 // Further, if a part is vector, print the above for
1623 // each vector element.
1624 SmallVector<EVT, 16> vtparts;
1625 ComputeValueVTs(*TLI, ETy, vtparts);
1626 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1627 unsigned elems = 1;
1628 EVT elemtype = vtparts[i];
1629 if (vtparts[i].isVector()) {
1630 elems = vtparts[i].getVectorNumElements();
1631 elemtype = vtparts[i].getVectorElementType();
1632 }
1633
1634 for (unsigned j = 0, je = elems; j != je; ++j) {
1635 unsigned sz = elemtype.getSizeInBits();
1636 if (elemtype.isInteger() && (sz < 32))
1637 sz = 32;
1638 O << "\t.reg .b" << sz << " ";
1639 printParamName(I, paramIndex, O);
1640 if (j < je - 1)
1641 O << ",\n";
1642 ++paramIndex;
1643 }
1644 if (i < e - 1)
1645 O << ",\n";
1646 }
1647 --paramIndex;
1648 continue;
1649 }
1650 }
1651
1652 O << "\n)\n";
1653 }
1654
emitFunctionParamList(const MachineFunction & MF,raw_ostream & O)1655 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1656 raw_ostream &O) {
1657 const Function *F = MF.getFunction();
1658 emitFunctionParamList(F, O);
1659 }
1660
setAndEmitFunctionVirtualRegisters(const MachineFunction & MF)1661 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1662 const MachineFunction &MF) {
1663 SmallString<128> Str;
1664 raw_svector_ostream O(Str);
1665
1666 // Map the global virtual register number to a register class specific
1667 // virtual register number starting from 1 with that class.
1668 const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo();
1669 //unsigned numRegClasses = TRI->getNumRegClasses();
1670
1671 // Emit the Fake Stack Object
1672 const MachineFrameInfo *MFI = MF.getFrameInfo();
1673 int NumBytes = (int) MFI->getStackSize();
1674 if (NumBytes) {
1675 O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
1676 << getFunctionNumber() << "[" << NumBytes << "];\n";
1677 if (nvptxSubtarget.is64Bit()) {
1678 O << "\t.reg .b64 \t%SP;\n";
1679 O << "\t.reg .b64 \t%SPL;\n";
1680 } else {
1681 O << "\t.reg .b32 \t%SP;\n";
1682 O << "\t.reg .b32 \t%SPL;\n";
1683 }
1684 }
1685
1686 // Go through all virtual registers to establish the mapping between the
1687 // global virtual
1688 // register number and the per class virtual register number.
1689 // We use the per class virtual register number in the ptx output.
1690 unsigned int numVRs = MRI->getNumVirtRegs();
1691 for (unsigned i = 0; i < numVRs; i++) {
1692 unsigned int vr = TRI->index2VirtReg(i);
1693 const TargetRegisterClass *RC = MRI->getRegClass(vr);
1694 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1695 int n = regmap.size();
1696 regmap.insert(std::make_pair(vr, n + 1));
1697 }
1698
1699 // Emit register declarations
1700 // @TODO: Extract out the real register usage
1701 // O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
1702 // O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
1703 // O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
1704 // O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
1705 // O << "\t.reg .s64 %rl<" << NVPTXNumRegisters << ">;\n";
1706 // O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
1707 // O << "\t.reg .f64 %fl<" << NVPTXNumRegisters << ">;\n";
1708
1709 // Emit declaration of the virtual registers or 'physical' registers for
1710 // each register class
1711 for (unsigned i=0; i< TRI->getNumRegClasses(); i++) {
1712 const TargetRegisterClass *RC = TRI->getRegClass(i);
1713 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1714 std::string rcname = getNVPTXRegClassName(RC);
1715 std::string rcStr = getNVPTXRegClassStr(RC);
1716 int n = regmap.size();
1717
1718 // Only declare those registers that may be used.
1719 if (n) {
1720 O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
1721 << ">;\n";
1722 }
1723 }
1724
1725 OutStreamer.EmitRawText(O.str());
1726 }
1727
printFPConstant(const ConstantFP * Fp,raw_ostream & O)1728 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
1729 APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
1730 bool ignored;
1731 unsigned int numHex;
1732 const char *lead;
1733
1734 if (Fp->getType()->getTypeID() == Type::FloatTyID) {
1735 numHex = 8;
1736 lead = "0f";
1737 APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &ignored);
1738 } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
1739 numHex = 16;
1740 lead = "0d";
1741 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &ignored);
1742 } else
1743 llvm_unreachable("unsupported fp type");
1744
1745 APInt API = APF.bitcastToAPInt();
1746 std::string hexstr(utohexstr(API.getZExtValue()));
1747 O << lead;
1748 if (hexstr.length() < numHex)
1749 O << std::string(numHex - hexstr.length(), '0');
1750 O << utohexstr(API.getZExtValue());
1751 }
1752
printScalarConstant(const Constant * CPV,raw_ostream & O)1753 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
1754 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
1755 O << CI->getValue();
1756 return;
1757 }
1758 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
1759 printFPConstant(CFP, O);
1760 return;
1761 }
1762 if (isa<ConstantPointerNull>(CPV)) {
1763 O << "0";
1764 return;
1765 }
1766 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1767 O << *getSymbol(GVar);
1768 return;
1769 }
1770 if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1771 const Value *v = Cexpr->stripPointerCasts();
1772 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
1773 O << *getSymbol(GVar);
1774 return;
1775 } else {
1776 O << *LowerConstant(CPV, *this);
1777 return;
1778 }
1779 }
1780 llvm_unreachable("Not scalar type found in printScalarConstant()");
1781 }
1782
bufferLEByte(const Constant * CPV,int Bytes,AggBuffer * aggBuffer)1783 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
1784 AggBuffer *aggBuffer) {
1785
1786 const DataLayout *TD = TM.getDataLayout();
1787
1788 if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
1789 int s = TD->getTypeAllocSize(CPV->getType());
1790 if (s < Bytes)
1791 s = Bytes;
1792 aggBuffer->addZeros(s);
1793 return;
1794 }
1795
1796 unsigned char *ptr;
1797 switch (CPV->getType()->getTypeID()) {
1798
1799 case Type::IntegerTyID: {
1800 const Type *ETy = CPV->getType();
1801 if (ETy == Type::getInt8Ty(CPV->getContext())) {
1802 unsigned char c =
1803 (unsigned char)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
1804 ptr = &c;
1805 aggBuffer->addBytes(ptr, 1, Bytes);
1806 } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
1807 short int16 = (short)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
1808 ptr = (unsigned char *)&int16;
1809 aggBuffer->addBytes(ptr, 2, Bytes);
1810 } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
1811 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1812 int int32 = (int)(constInt->getZExtValue());
1813 ptr = (unsigned char *)&int32;
1814 aggBuffer->addBytes(ptr, 4, Bytes);
1815 break;
1816 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1817 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1818 ConstantFoldConstantExpression(Cexpr, TD))) {
1819 int int32 = (int)(constInt->getZExtValue());
1820 ptr = (unsigned char *)&int32;
1821 aggBuffer->addBytes(ptr, 4, Bytes);
1822 break;
1823 }
1824 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1825 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1826 aggBuffer->addSymbol(v);
1827 aggBuffer->addZeros(4);
1828 break;
1829 }
1830 }
1831 llvm_unreachable("unsupported integer const type");
1832 } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
1833 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1834 long long int64 = (long long)(constInt->getZExtValue());
1835 ptr = (unsigned char *)&int64;
1836 aggBuffer->addBytes(ptr, 8, Bytes);
1837 break;
1838 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1839 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1840 ConstantFoldConstantExpression(Cexpr, TD))) {
1841 long long int64 = (long long)(constInt->getZExtValue());
1842 ptr = (unsigned char *)&int64;
1843 aggBuffer->addBytes(ptr, 8, Bytes);
1844 break;
1845 }
1846 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1847 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1848 aggBuffer->addSymbol(v);
1849 aggBuffer->addZeros(8);
1850 break;
1851 }
1852 }
1853 llvm_unreachable("unsupported integer const type");
1854 } else
1855 llvm_unreachable("unsupported integer const type");
1856 break;
1857 }
1858 case Type::FloatTyID:
1859 case Type::DoubleTyID: {
1860 const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
1861 const Type *Ty = CFP->getType();
1862 if (Ty == Type::getFloatTy(CPV->getContext())) {
1863 float float32 = (float) CFP->getValueAPF().convertToFloat();
1864 ptr = (unsigned char *)&float32;
1865 aggBuffer->addBytes(ptr, 4, Bytes);
1866 } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
1867 double float64 = CFP->getValueAPF().convertToDouble();
1868 ptr = (unsigned char *)&float64;
1869 aggBuffer->addBytes(ptr, 8, Bytes);
1870 } else {
1871 llvm_unreachable("unsupported fp const type");
1872 }
1873 break;
1874 }
1875 case Type::PointerTyID: {
1876 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1877 aggBuffer->addSymbol(GVar);
1878 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1879 const Value *v = Cexpr->stripPointerCasts();
1880 aggBuffer->addSymbol(v);
1881 }
1882 unsigned int s = TD->getTypeAllocSize(CPV->getType());
1883 aggBuffer->addZeros(s);
1884 break;
1885 }
1886
1887 case Type::ArrayTyID:
1888 case Type::VectorTyID:
1889 case Type::StructTyID: {
1890 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV) ||
1891 isa<ConstantStruct>(CPV) || isa<ConstantDataSequential>(CPV)) {
1892 int ElementSize = TD->getTypeAllocSize(CPV->getType());
1893 bufferAggregateConstant(CPV, aggBuffer);
1894 if (Bytes > ElementSize)
1895 aggBuffer->addZeros(Bytes - ElementSize);
1896 } else if (isa<ConstantAggregateZero>(CPV))
1897 aggBuffer->addZeros(Bytes);
1898 else
1899 llvm_unreachable("Unexpected Constant type");
1900 break;
1901 }
1902
1903 default:
1904 llvm_unreachable("unsupported type");
1905 }
1906 }
1907
bufferAggregateConstant(const Constant * CPV,AggBuffer * aggBuffer)1908 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
1909 AggBuffer *aggBuffer) {
1910 const DataLayout *TD = TM.getDataLayout();
1911 int Bytes;
1912
1913 // Old constants
1914 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
1915 if (CPV->getNumOperands())
1916 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
1917 bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
1918 return;
1919 }
1920
1921 if (const ConstantDataSequential *CDS =
1922 dyn_cast<ConstantDataSequential>(CPV)) {
1923 if (CDS->getNumElements())
1924 for (unsigned i = 0; i < CDS->getNumElements(); ++i)
1925 bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
1926 aggBuffer);
1927 return;
1928 }
1929
1930 if (isa<ConstantStruct>(CPV)) {
1931 if (CPV->getNumOperands()) {
1932 StructType *ST = cast<StructType>(CPV->getType());
1933 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
1934 if (i == (e - 1))
1935 Bytes = TD->getStructLayout(ST)->getElementOffset(0) +
1936 TD->getTypeAllocSize(ST) -
1937 TD->getStructLayout(ST)->getElementOffset(i);
1938 else
1939 Bytes = TD->getStructLayout(ST)->getElementOffset(i + 1) -
1940 TD->getStructLayout(ST)->getElementOffset(i);
1941 bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
1942 }
1943 }
1944 return;
1945 }
1946 llvm_unreachable("unsupported constant type in printAggregateConstant()");
1947 }
1948
1949 // buildTypeNameMap - Run through symbol table looking for type names.
1950 //
1951
isImageType(const Type * Ty)1952 bool NVPTXAsmPrinter::isImageType(const Type *Ty) {
1953
1954 std::map<const Type *, std::string>::iterator PI = TypeNameMap.find(Ty);
1955
1956 if (PI != TypeNameMap.end() && (!PI->second.compare("struct._image1d_t") ||
1957 !PI->second.compare("struct._image2d_t") ||
1958 !PI->second.compare("struct._image3d_t")))
1959 return true;
1960
1961 return false;
1962 }
1963
1964
ignoreLoc(const MachineInstr & MI)1965 bool NVPTXAsmPrinter::ignoreLoc(const MachineInstr &MI) {
1966 switch (MI.getOpcode()) {
1967 default:
1968 return false;
1969 case NVPTX::CallArgBeginInst:
1970 case NVPTX::CallArgEndInst0:
1971 case NVPTX::CallArgEndInst1:
1972 case NVPTX::CallArgF32:
1973 case NVPTX::CallArgF64:
1974 case NVPTX::CallArgI16:
1975 case NVPTX::CallArgI32:
1976 case NVPTX::CallArgI32imm:
1977 case NVPTX::CallArgI64:
1978 case NVPTX::CallArgParam:
1979 case NVPTX::CallVoidInst:
1980 case NVPTX::CallVoidInstReg:
1981 case NVPTX::Callseq_End:
1982 case NVPTX::CallVoidInstReg64:
1983 case NVPTX::DeclareParamInst:
1984 case NVPTX::DeclareRetMemInst:
1985 case NVPTX::DeclareRetRegInst:
1986 case NVPTX::DeclareRetScalarInst:
1987 case NVPTX::DeclareScalarParamInst:
1988 case NVPTX::DeclareScalarRegInst:
1989 case NVPTX::StoreParamF32:
1990 case NVPTX::StoreParamF64:
1991 case NVPTX::StoreParamI16:
1992 case NVPTX::StoreParamI32:
1993 case NVPTX::StoreParamI64:
1994 case NVPTX::StoreParamI8:
1995 case NVPTX::StoreRetvalF32:
1996 case NVPTX::StoreRetvalF64:
1997 case NVPTX::StoreRetvalI16:
1998 case NVPTX::StoreRetvalI32:
1999 case NVPTX::StoreRetvalI64:
2000 case NVPTX::StoreRetvalI8:
2001 case NVPTX::LastCallArgF32:
2002 case NVPTX::LastCallArgF64:
2003 case NVPTX::LastCallArgI16:
2004 case NVPTX::LastCallArgI32:
2005 case NVPTX::LastCallArgI32imm:
2006 case NVPTX::LastCallArgI64:
2007 case NVPTX::LastCallArgParam:
2008 case NVPTX::LoadParamMemF32:
2009 case NVPTX::LoadParamMemF64:
2010 case NVPTX::LoadParamMemI16:
2011 case NVPTX::LoadParamMemI32:
2012 case NVPTX::LoadParamMemI64:
2013 case NVPTX::LoadParamMemI8:
2014 case NVPTX::PrototypeInst:
2015 case NVPTX::DBG_VALUE:
2016 return true;
2017 }
2018 return false;
2019 }
2020
2021 /// PrintAsmOperand - Print out an operand for an inline asm expression.
2022 ///
PrintAsmOperand(const MachineInstr * MI,unsigned OpNo,unsigned AsmVariant,const char * ExtraCode,raw_ostream & O)2023 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
2024 unsigned AsmVariant,
2025 const char *ExtraCode, raw_ostream &O) {
2026 if (ExtraCode && ExtraCode[0]) {
2027 if (ExtraCode[1] != 0)
2028 return true; // Unknown modifier.
2029
2030 switch (ExtraCode[0]) {
2031 default:
2032 // See if this is a generic print operand
2033 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
2034 case 'r':
2035 break;
2036 }
2037 }
2038
2039 printOperand(MI, OpNo, O);
2040
2041 return false;
2042 }
2043
PrintAsmMemoryOperand(const MachineInstr * MI,unsigned OpNo,unsigned AsmVariant,const char * ExtraCode,raw_ostream & O)2044 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(
2045 const MachineInstr *MI, unsigned OpNo, unsigned AsmVariant,
2046 const char *ExtraCode, raw_ostream &O) {
2047 if (ExtraCode && ExtraCode[0])
2048 return true; // Unknown modifier
2049
2050 O << '[';
2051 printMemOperand(MI, OpNo, O);
2052 O << ']';
2053
2054 return false;
2055 }
2056
printOperand(const MachineInstr * MI,int opNum,raw_ostream & O,const char * Modifier)2057 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
2058 raw_ostream &O, const char *Modifier) {
2059 const MachineOperand &MO = MI->getOperand(opNum);
2060 switch (MO.getType()) {
2061 case MachineOperand::MO_Register:
2062 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
2063 if (MO.getReg() == NVPTX::VRDepot)
2064 O << DEPOTNAME << getFunctionNumber();
2065 else
2066 O << NVPTXInstPrinter::getRegisterName(MO.getReg());
2067 } else {
2068 emitVirtualRegister(MO.getReg(), O);
2069 }
2070 return;
2071
2072 case MachineOperand::MO_Immediate:
2073 if (!Modifier)
2074 O << MO.getImm();
2075 else if (strstr(Modifier, "vec") == Modifier)
2076 printVecModifiedImmediate(MO, Modifier, O);
2077 else
2078 llvm_unreachable(
2079 "Don't know how to handle modifier on immediate operand");
2080 return;
2081
2082 case MachineOperand::MO_FPImmediate:
2083 printFPConstant(MO.getFPImm(), O);
2084 break;
2085
2086 case MachineOperand::MO_GlobalAddress:
2087 O << *getSymbol(MO.getGlobal());
2088 break;
2089
2090 case MachineOperand::MO_ExternalSymbol: {
2091 const char *symbname = MO.getSymbolName();
2092 if (strstr(symbname, ".PARAM") == symbname) {
2093 unsigned index;
2094 sscanf(symbname + 6, "%u[];", &index);
2095 printParamName(index, O);
2096 } else if (strstr(symbname, ".HLPPARAM") == symbname) {
2097 unsigned index;
2098 sscanf(symbname + 9, "%u[];", &index);
2099 O << *CurrentFnSym << "_param_" << index << "_offset";
2100 } else
2101 O << symbname;
2102 break;
2103 }
2104
2105 case MachineOperand::MO_MachineBasicBlock:
2106 O << *MO.getMBB()->getSymbol();
2107 return;
2108
2109 default:
2110 llvm_unreachable("Operand type not supported.");
2111 }
2112 }
2113
printMemOperand(const MachineInstr * MI,int opNum,raw_ostream & O,const char * Modifier)2114 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
2115 raw_ostream &O, const char *Modifier) {
2116 printOperand(MI, opNum, O);
2117
2118 if (Modifier && !strcmp(Modifier, "add")) {
2119 O << ", ";
2120 printOperand(MI, opNum + 1, O);
2121 } else {
2122 if (MI->getOperand(opNum + 1).isImm() &&
2123 MI->getOperand(opNum + 1).getImm() == 0)
2124 return; // don't print ',0' or '+0'
2125 O << "+";
2126 printOperand(MI, opNum + 1, O);
2127 }
2128 }
2129
2130
2131 // Force static initialization.
LLVMInitializeNVPTXBackendAsmPrinter()2132 extern "C" void LLVMInitializeNVPTXBackendAsmPrinter() {
2133 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2134 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2135 }
2136
emitSrcInText(StringRef filename,unsigned line)2137 void NVPTXAsmPrinter::emitSrcInText(StringRef filename, unsigned line) {
2138 std::stringstream temp;
2139 LineReader *reader = this->getReader(filename.str());
2140 temp << "\n//";
2141 temp << filename.str();
2142 temp << ":";
2143 temp << line;
2144 temp << " ";
2145 temp << reader->readLine(line);
2146 temp << "\n";
2147 this->OutStreamer.EmitRawText(Twine(temp.str()));
2148 }
2149
getReader(std::string filename)2150 LineReader *NVPTXAsmPrinter::getReader(std::string filename) {
2151 if (reader == NULL) {
2152 reader = new LineReader(filename);
2153 }
2154
2155 if (reader->fileName() != filename) {
2156 delete reader;
2157 reader = new LineReader(filename);
2158 }
2159
2160 return reader;
2161 }
2162
readLine(unsigned lineNum)2163 std::string LineReader::readLine(unsigned lineNum) {
2164 if (lineNum < theCurLine) {
2165 theCurLine = 0;
2166 fstr.seekg(0, std::ios::beg);
2167 }
2168 while (theCurLine < lineNum) {
2169 fstr.getline(buff, 500);
2170 theCurLine++;
2171 }
2172 return buff;
2173 }
2174
2175 // Force static initialization.
LLVMInitializeNVPTXAsmPrinter()2176 extern "C" void LLVMInitializeNVPTXAsmPrinter() {
2177 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2178 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2179 }
2180