1 //===-- PTXAsmPrinter.cpp - PTX LLVM assembly writer ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains a printer that converts from our internal representation
11 // of machine-dependent LLVM code to PTX assembly language.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "ptx-asm-printer"
18 #include "PTXAsmPrinter.h"
19 #include "PTXMachineFunctionInfo.h"
20 #include "PTXParamManager.h"
21 #include "PTXRegisterInfo.h"
22 #include "PTXTargetMachine.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Module.h"
25 #include "llvm/ADT/SmallString.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/Twine.h"
28 #include "llvm/Analysis/DebugInfo.h"
29 #include "llvm/CodeGen/AsmPrinter.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineInstr.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/MC/MCContext.h"
34 #include "llvm/MC/MCExpr.h"
35 #include "llvm/MC/MCInst.h"
36 #include "llvm/MC/MCStreamer.h"
37 #include "llvm/MC/MCSymbol.h"
38 #include "llvm/Target/Mangler.h"
39 #include "llvm/Target/TargetLoweringObjectFile.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/MathExtras.h"
44 #include "llvm/Support/Path.h"
45 #include "llvm/Support/TargetRegistry.h"
46 #include "llvm/Support/raw_ostream.h"
50 static const char PARAM_PREFIX[] = "__param_";
51 static const char RETURN_PREFIX[] = "__ret_";
53 static const char *getRegisterTypeName(unsigned RegNo,
54 const MachineRegisterInfo& MRI) {
55 const TargetRegisterClass *TRC = MRI.getRegClass(RegNo);
57 #define TEST_REGCLS(cls, clsstr) \
58 if (PTX::cls ## RegisterClass == TRC) return # clsstr;
60 TEST_REGCLS(RegPred, pred);
61 TEST_REGCLS(RegI16, b16);
62 TEST_REGCLS(RegI32, b32);
63 TEST_REGCLS(RegI64, b64);
64 TEST_REGCLS(RegF32, b32);
65 TEST_REGCLS(RegF64, b64);
68 llvm_unreachable("Not in any register class!");
72 static const char *getStateSpaceName(unsigned addressSpace) {
73 switch (addressSpace) {
74 default: llvm_unreachable("Unknown state space");
75 case PTX::GLOBAL: return "global";
76 case PTX::CONSTANT: return "const";
77 case PTX::LOCAL: return "local";
78 case PTX::PARAMETER: return "param";
79 case PTX::SHARED: return "shared";
84 static const char *getTypeName(Type* type) {
86 switch (type->getTypeID()) {
87 default: llvm_unreachable("Unknown type");
88 case Type::FloatTyID: return ".f32";
89 case Type::DoubleTyID: return ".f64";
90 case Type::IntegerTyID:
91 switch (type->getPrimitiveSizeInBits()) {
92 default: llvm_unreachable("Unknown integer bit-width");
93 case 16: return ".u16";
94 case 32: return ".u32";
95 case 64: return ".u64";
98 case Type::PointerTyID:
99 type = dyn_cast<SequentialType>(type)->getElementType();
106 bool PTXAsmPrinter::doFinalization(Module &M) {
107 // XXX Temproarily remove global variables so that doFinalization() will not
108 // emit them again (global variables are emitted at beginning).
110 Module::GlobalListType &global_list = M.getGlobalList();
111 int i, n = global_list.size();
112 GlobalVariable **gv_array = new GlobalVariable* [n];
114 // first, back-up GlobalVariable in gv_array
116 for (Module::global_iterator I = global_list.begin(), E = global_list.end();
120 // second, empty global_list
121 while (!global_list.empty())
122 global_list.remove(global_list.begin());
124 // call doFinalization
125 bool ret = AsmPrinter::doFinalization(M);
127 // now we restore global variables
128 for (i = 0; i < n; i ++)
129 global_list.insert(global_list.end(), gv_array[i]);
135 void PTXAsmPrinter::EmitStartOfAsmFile(Module &M)
137 const PTXSubtarget& ST = TM.getSubtarget<PTXSubtarget>();
139 // Emit the PTX .version and .target attributes
140 OutStreamer.EmitRawText(Twine("\t.version " + ST.getPTXVersionString()));
141 OutStreamer.EmitRawText(Twine("\t.target " + ST.getTargetString() +
142 (ST.supportsDouble() ? ""
143 : ", map_f64_to_f32")));
144 // .address_size directive is optional, but it must immediately follow
145 // the .target directive if present within a module
146 if (ST.supportsPTX23()) {
147 std::string addrSize = ST.is64Bit() ? "64" : "32";
148 OutStreamer.EmitRawText(Twine("\t.address_size " + addrSize));
151 OutStreamer.AddBlankLine();
153 // Define any .file directives
154 DebugInfoFinder DbgFinder;
155 DbgFinder.processModule(M);
157 for (DebugInfoFinder::iterator I = DbgFinder.compile_unit_begin(),
158 E = DbgFinder.compile_unit_end(); I != E; ++I) {
159 DICompileUnit DIUnit(*I);
160 StringRef FN = DIUnit.getFilename();
161 StringRef Dir = DIUnit.getDirectory();
162 GetOrCreateSourceID(FN, Dir);
165 OutStreamer.AddBlankLine();
167 // declare global variables
168 for (Module::const_global_iterator i = M.global_begin(), e = M.global_end();
170 EmitVariableDeclaration(i);
173 void PTXAsmPrinter::EmitFunctionBodyStart() {
174 OutStreamer.EmitRawText(Twine("{"));
176 const PTXMachineFunctionInfo *MFI = MF->getInfo<PTXMachineFunctionInfo>();
177 const PTXParamManager &PM = MFI->getParamManager();
179 // Print register definitions
184 numRegs = MFI->getNumRegistersForClass(PTX::RegPredRegisterClass);
186 regDefs += "\t.reg .pred %p<";
187 regDefs += utostr(numRegs);
192 numRegs = MFI->getNumRegistersForClass(PTX::RegI16RegisterClass);
194 regDefs += "\t.reg .b16 %rh<";
195 regDefs += utostr(numRegs);
200 numRegs = MFI->getNumRegistersForClass(PTX::RegI32RegisterClass);
202 regDefs += "\t.reg .b32 %r<";
203 regDefs += utostr(numRegs);
208 numRegs = MFI->getNumRegistersForClass(PTX::RegI64RegisterClass);
210 regDefs += "\t.reg .b64 %rd<";
211 regDefs += utostr(numRegs);
216 numRegs = MFI->getNumRegistersForClass(PTX::RegF32RegisterClass);
218 regDefs += "\t.reg .f32 %f<";
219 regDefs += utostr(numRegs);
224 numRegs = MFI->getNumRegistersForClass(PTX::RegF64RegisterClass);
226 regDefs += "\t.reg .f64 %fd<";
227 regDefs += utostr(numRegs);
232 for (PTXParamManager::param_iterator i = PM.local_begin(), e = PM.local_end();
234 regDefs += "\t.param .b";
235 regDefs += utostr(PM.getParamSize(*i));
237 regDefs += PM.getParamName(*i);
241 OutStreamer.EmitRawText(Twine(regDefs));
244 const MachineFrameInfo* FrameInfo = MF->getFrameInfo();
245 DEBUG(dbgs() << "Have " << FrameInfo->getNumObjects()
246 << " frame object(s)\n");
247 for (unsigned i = 0, e = FrameInfo->getNumObjects(); i != e; ++i) {
248 DEBUG(dbgs() << "Size of object: " << FrameInfo->getObjectSize(i) << "\n");
249 if (FrameInfo->getObjectSize(i) > 0) {
250 std::string def = "\t.local .align ";
251 def += utostr(FrameInfo->getObjectAlignment(i));
256 def += utostr(FrameInfo->getObjectSize(i)); // Convert to bits
259 OutStreamer.EmitRawText(Twine(def));
263 //unsigned Index = 1;
264 // Print parameter passing params
265 //for (PTXMachineFunctionInfo::param_iterator
266 // i = MFI->paramBegin(), e = MFI->paramEnd(); i != e; ++i) {
267 // std::string def = "\t.param .b";
268 // def += utostr(*i);
270 // def += utostr(Index);
273 // OutStreamer.EmitRawText(Twine(def));
277 void PTXAsmPrinter::EmitFunctionBodyEnd() {
278 OutStreamer.EmitRawText(Twine("}"));
281 void PTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
286 raw_string_ostream OS(str);
288 DebugLoc DL = MI->getDebugLoc();
289 if (!DL.isUnknown()) {
291 const MDNode *S = DL.getScope(MF->getFunction()->getContext());
293 // This is taken from DwarfDebug.cpp, which is conveniently not a public
299 DIDescriptor Scope(S);
300 if (Scope.isCompileUnit()) {
302 Fn = CU.getFilename();
303 Dir = CU.getDirectory();
304 } else if (Scope.isFile()) {
306 Fn = F.getFilename();
307 Dir = F.getDirectory();
308 } else if (Scope.isSubprogram()) {
310 Fn = SP.getFilename();
311 Dir = SP.getDirectory();
312 } else if (Scope.isLexicalBlock()) {
313 DILexicalBlock DB(S);
314 Fn = DB.getFilename();
315 Dir = DB.getDirectory();
317 assert(0 && "Unexpected scope info");
319 Src = GetOrCreateSourceID(Fn, Dir);
321 OutStreamer.EmitDwarfLocDirective(Src, DL.getLine(), DL.getCol(),
324 const MCDwarfLoc& MDL = OutContext.getCurrentDwarfLoc();
327 OS << utostr(MDL.getFileNum());
329 OS << utostr(MDL.getLine());
331 OS << utostr(MDL.getColumn());
337 printPredicateOperand(MI, OS);
339 // Write instruction to str
340 if (MI->getOpcode() == PTX::CALL) {
343 printInstruction(MI, OS);
348 StringRef strref = StringRef(str);
349 OutStreamer.EmitRawText(strref);
353 LowerPTXMachineInstrToMCInst(MI, TmpInst, *this);
354 OutStreamer.EmitInstruction(TmpInst);
357 void PTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
359 const MachineOperand &MO = MI->getOperand(opNum);
360 const PTXMachineFunctionInfo *MFI = MF->getInfo<PTXMachineFunctionInfo>();
362 switch (MO.getType()) {
364 llvm_unreachable("<unknown operand type>");
366 case MachineOperand::MO_GlobalAddress:
367 OS << *Mang->getSymbol(MO.getGlobal());
369 case MachineOperand::MO_Immediate:
370 OS << (long) MO.getImm();
372 case MachineOperand::MO_MachineBasicBlock:
373 OS << *MO.getMBB()->getSymbol();
375 case MachineOperand::MO_Register:
376 OS << MFI->getRegisterName(MO.getReg());
378 case MachineOperand::MO_ExternalSymbol:
379 OS << MO.getSymbolName();
381 case MachineOperand::MO_FPImmediate:
382 APInt constFP = MO.getFPImm()->getValueAPF().bitcastToAPInt();
383 bool isFloat = MO.getFPImm()->getType()->getTypeID() == Type::FloatTyID;
384 // Emit 0F for 32-bit floats and 0D for 64-bit doubles.
391 // Emit the encoded floating-point value.
392 if (constFP.getZExtValue() > 0) {
393 OS << constFP.toString(16, false);
397 // If We have a double-precision zero, pad to 8-bytes.
406 void PTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
407 raw_ostream &OS, const char *Modifier) {
408 printOperand(MI, opNum, OS);
410 if (MI->getOperand(opNum+1).isImm() && MI->getOperand(opNum+1).getImm() == 0)
411 return; // don't print "+0"
414 printOperand(MI, opNum+1, OS);
417 void PTXAsmPrinter::printReturnOperand(const MachineInstr *MI, int opNum,
418 raw_ostream &OS, const char *Modifier) {
419 //OS << RETURN_PREFIX << (int) MI->getOperand(opNum).getImm() + 1;
423 void PTXAsmPrinter::EmitVariableDeclaration(const GlobalVariable *gv) {
424 // Check to see if this is a special global used by LLVM, if so, emit it.
425 if (EmitSpecialLLVMGlobal(gv))
428 MCSymbol *gvsym = Mang->getSymbol(gv);
430 assert(gvsym->isUndefined() && "Cannot define a symbol twice!");
434 // check if it is defined in some other translation unit
435 if (gv->isDeclaration())
438 // state space: e.g., .global
440 decl += getStateSpaceName(gv->getType()->getAddressSpace());
443 // alignment (optional)
444 unsigned alignment = gv->getAlignment();
445 if (alignment != 0) {
447 decl += utostr(std::max(1U, Log2_32(gv->getAlignment())));
452 if (PointerType::classof(gv->getType())) {
453 PointerType* pointerTy = dyn_cast<PointerType>(gv->getType());
454 Type* elementTy = pointerTy->getElementType();
457 decl += gvsym->getName();
460 if (elementTy->isArrayTy())
462 assert(elementTy->isArrayTy() && "Only pointers to arrays are supported");
464 ArrayType* arrayTy = dyn_cast<ArrayType>(elementTy);
465 elementTy = arrayTy->getElementType();
467 unsigned numElements = arrayTy->getNumElements();
469 while (elementTy->isArrayTy()) {
471 arrayTy = dyn_cast<ArrayType>(elementTy);
472 elementTy = arrayTy->getElementType();
474 numElements *= arrayTy->getNumElements();
477 // FIXME: isPrimitiveType() == false for i16?
478 assert(elementTy->isSingleValueType() &&
479 "Non-primitive types are not handled");
481 // Compute the size of the array, in bytes.
482 uint64_t arraySize = (elementTy->getPrimitiveSizeInBits() >> 3)
485 decl += utostr(arraySize);
490 // handle string constants (assume ConstantArray means string)
492 if (gv->hasInitializer())
494 const Constant *C = gv->getInitializer();
495 if (const ConstantArray *CA = dyn_cast<ConstantArray>(C))
499 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i)
501 if (i > 0) decl += ",";
504 utohexstr(cast<ConstantInt>(CA->getOperand(i))->getZExtValue());
512 // Note: this is currently the fall-through case and most likely generates
514 decl += getTypeName(gv->getType());
517 decl += gvsym->getName();
519 if (ArrayType::classof(gv->getType()) ||
520 PointerType::classof(gv->getType()))
526 OutStreamer.EmitRawText(Twine(decl));
528 OutStreamer.AddBlankLine();
531 void PTXAsmPrinter::EmitFunctionEntryLabel() {
532 // The function label could have already been emitted if two symbols end up
533 // conflicting due to asm renaming. Detect this and emit an error.
534 if (!CurrentFnSym->isUndefined()) {
535 report_fatal_error("'" + Twine(CurrentFnSym->getName()) +
536 "' label emitted multiple times to assembly file");
540 const PTXMachineFunctionInfo *MFI = MF->getInfo<PTXMachineFunctionInfo>();
541 const PTXParamManager &PM = MFI->getParamManager();
542 const bool isKernel = MFI->isKernel();
543 const PTXSubtarget& ST = TM.getSubtarget<PTXSubtarget>();
544 const MachineRegisterInfo& MRI = MF->getRegInfo();
546 std::string decl = isKernel ? ".entry" : ".func";
552 if (ST.useParamSpaceForDeviceArgs()) {
553 for (PTXParamManager::param_iterator i = PM.ret_begin(), e = PM.ret_end(),
554 b = i; i != e; ++i) {
560 decl += utostr(PM.getParamSize(*i));
562 decl += PM.getParamName(*i);
565 for (PTXMachineFunctionInfo::reg_iterator
566 i = MFI->retreg_begin(), e = MFI->retreg_end(), b = i;
572 decl += getRegisterTypeName(*i, MRI);
574 decl += MFI->getRegisterName(*i);
580 // Print function name
582 decl += CurrentFnSym->getName().str();
589 if (isKernel || ST.useParamSpaceForDeviceArgs()) {
590 for (PTXParamManager::param_iterator i = PM.arg_begin(), e = PM.arg_end(),
591 b = i; i != e; ++i) {
597 decl += utostr(PM.getParamSize(*i));
599 decl += PM.getParamName(*i);
602 for (PTXMachineFunctionInfo::reg_iterator
603 i = MFI->argreg_begin(), e = MFI->argreg_end(), b = i;
610 decl += getRegisterTypeName(*i, MRI);
612 decl += MFI->getRegisterName(*i);
617 OutStreamer.EmitRawText(Twine(decl));
621 printPredicateOperand(const MachineInstr *MI, raw_ostream &O) {
622 int i = MI->findFirstPredOperandIdx();
624 llvm_unreachable("missing predicate operand");
626 unsigned reg = MI->getOperand(i).getReg();
627 int predOp = MI->getOperand(i+1).getImm();
628 const PTXMachineFunctionInfo *MFI = MF->getInfo<PTXMachineFunctionInfo>();
630 DEBUG(dbgs() << "predicate: (" << reg << ", " << predOp << ")\n");
632 if (reg != PTX::NoRegister) {
634 if (predOp == PTX::PRED_NEGATE)
636 O << MFI->getRegisterName(reg);
641 printCall(const MachineInstr *MI, raw_ostream &O) {
643 // The first two operands are the predicate slot
645 while (!MI->getOperand(Index).isGlobal()) {
651 printOperand(MI, Index, O);
659 assert(MI->getOperand(Index).isGlobal() &&
660 "A GlobalAddress must follow the return arguments");
662 const GlobalValue *Address = MI->getOperand(Index).getGlobal();
663 O << Address->getName() << ", (";
666 while (Index < MI->getNumOperands()) {
667 printOperand(MI, Index, O);
668 if (Index < MI->getNumOperands()-1) {
677 unsigned PTXAsmPrinter::GetOrCreateSourceID(StringRef FileName,
679 // If FE did not provide a file name, then assume stdin.
680 if (FileName.empty())
681 return GetOrCreateSourceID("<stdin>", StringRef());
683 // MCStream expects full path name as filename.
684 if (!DirName.empty() && !sys::path::is_absolute(FileName)) {
685 SmallString<128> FullPathName = DirName;
686 sys::path::append(FullPathName, FileName);
687 // Here FullPathName will be copied into StringMap by GetOrCreateSourceID.
688 return GetOrCreateSourceID(StringRef(FullPathName), StringRef());
691 StringMapEntry<unsigned> &Entry = SourceIdMap.GetOrCreateValue(FileName);
692 if (Entry.getValue())
693 return Entry.getValue();
695 unsigned SrcId = SourceIdMap.size();
696 Entry.setValue(SrcId);
698 // Print out a .file directive to specify files for .loc directives.
699 OutStreamer.EmitDwarfFileDirective(SrcId, Entry.getKey());
704 MCOperand PTXAsmPrinter::GetSymbolRef(const MachineOperand &MO,
705 const MCSymbol *Symbol) {
707 Expr = MCSymbolRefExpr::Create(Symbol, MCSymbolRefExpr::VK_None, OutContext);
708 return MCOperand::CreateExpr(Expr);
711 bool PTXAsmPrinter::lowerOperand(const MachineOperand &MO, MCOperand &MCOp) {
712 const PTXMachineFunctionInfo *MFI = MF->getInfo<PTXMachineFunctionInfo>();
714 const char *RegSymbolName;
715 switch (MO.getType()) {
717 llvm_unreachable("Unknown operand type");
718 case MachineOperand::MO_Register:
719 // We create register operands as symbols, since the PTXInstPrinter class
720 // has no way to map virtual registers back to a name without some ugly
722 // FIXME: Figure out a better way to handle virtual register naming.
723 RegSymbolName = MFI->getRegisterName(MO.getReg());
724 Expr = MCSymbolRefExpr::Create(RegSymbolName, MCSymbolRefExpr::VK_None,
726 MCOp = MCOperand::CreateExpr(Expr);
728 case MachineOperand::MO_Immediate:
729 MCOp = MCOperand::CreateImm(MO.getImm());
731 case MachineOperand::MO_MachineBasicBlock:
732 MCOp = MCOperand::CreateExpr(MCSymbolRefExpr::Create(
733 MO.getMBB()->getSymbol(), OutContext));
735 case MachineOperand::MO_GlobalAddress:
736 MCOp = GetSymbolRef(MO, Mang->getSymbol(MO.getGlobal()));
738 case MachineOperand::MO_ExternalSymbol:
739 MCOp = GetSymbolRef(MO, GetExternalSymbolSymbol(MO.getSymbolName()));
741 case MachineOperand::MO_FPImmediate:
742 APFloat Val = MO.getFPImm()->getValueAPF();
744 Val.convert(APFloat::IEEEdouble, APFloat::rmTowardZero, &ignored);
745 MCOp = MCOperand::CreateFPImm(Val.convertToDouble());
752 // Force static initialization.
753 extern "C" void LLVMInitializePTXAsmPrinter() {
754 RegisterAsmPrinter<PTXAsmPrinter> X(ThePTX32Target);
755 RegisterAsmPrinter<PTXAsmPrinter> Y(ThePTX64Target);