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/Argument.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Function.h"
26 #include "llvm/Module.h"
27 #include "llvm/ADT/SmallString.h"
28 #include "llvm/ADT/Twine.h"
29 #include "llvm/Analysis/DebugInfo.h"
30 #include "llvm/CodeGen/AsmPrinter.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineInstr.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/MC/MCContext.h"
35 #include "llvm/MC/MCExpr.h"
36 #include "llvm/MC/MCInst.h"
37 #include "llvm/MC/MCStreamer.h"
38 #include "llvm/MC/MCSymbol.h"
39 #include "llvm/Target/Mangler.h"
40 #include "llvm/Target/TargetLoweringObjectFile.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/Path.h"
46 #include "llvm/Support/TargetRegistry.h"
47 #include "llvm/Support/raw_ostream.h"
51 static const char PARAM_PREFIX[] = "__param_";
52 static const char RETURN_PREFIX[] = "__ret_";
54 static const char *getRegisterTypeName(unsigned RegType) {
57 llvm_unreachable("Unknown register type");
58 case PTXRegisterType::Pred:
60 case PTXRegisterType::B16:
62 case PTXRegisterType::B32:
64 case PTXRegisterType::B64:
66 case PTXRegisterType::F32:
68 case PTXRegisterType::F64:
73 static const char *getStateSpaceName(unsigned addressSpace) {
74 switch (addressSpace) {
75 default: llvm_unreachable("Unknown state space");
76 case PTXStateSpace::Global: return "global";
77 case PTXStateSpace::Constant: return "const";
78 case PTXStateSpace::Local: return "local";
79 case PTXStateSpace::Parameter: return "param";
80 case PTXStateSpace::Shared: return "shared";
85 static const char *getTypeName(Type* type) {
87 switch (type->getTypeID()) {
88 default: llvm_unreachable("Unknown type");
89 case Type::FloatTyID: return ".f32";
90 case Type::DoubleTyID: return ".f64";
91 case Type::IntegerTyID:
92 switch (type->getPrimitiveSizeInBits()) {
93 default: llvm_unreachable("Unknown integer bit-width");
94 case 16: return ".u16";
95 case 32: return ".u32";
96 case 64: return ".u64";
99 case Type::PointerTyID:
100 type = dyn_cast<SequentialType>(type)->getElementType();
107 bool PTXAsmPrinter::doFinalization(Module &M) {
108 // XXX Temproarily remove global variables so that doFinalization() will not
109 // emit them again (global variables are emitted at beginning).
111 Module::GlobalListType &global_list = M.getGlobalList();
112 int i, n = global_list.size();
113 GlobalVariable **gv_array = new GlobalVariable* [n];
115 // first, back-up GlobalVariable in gv_array
117 for (Module::global_iterator I = global_list.begin(), E = global_list.end();
121 // second, empty global_list
122 while (!global_list.empty())
123 global_list.remove(global_list.begin());
125 // call doFinalization
126 bool ret = AsmPrinter::doFinalization(M);
128 // now we restore global variables
129 for (i = 0; i < n; i ++)
130 global_list.insert(global_list.end(), gv_array[i]);
136 void PTXAsmPrinter::EmitStartOfAsmFile(Module &M)
138 const PTXSubtarget& ST = TM.getSubtarget<PTXSubtarget>();
140 // Emit the PTX .version and .target attributes
141 OutStreamer.EmitRawText(Twine("\t.version ") + ST.getPTXVersionString());
142 OutStreamer.EmitRawText(Twine("\t.target ") + ST.getTargetString() +
143 (ST.supportsDouble() ? ""
144 : ", map_f64_to_f32"));
145 // .address_size directive is optional, but it must immediately follow
146 // the .target directive if present within a module
147 if (ST.supportsPTX23()) {
148 const char *addrSize = ST.is64Bit() ? "64" : "32";
149 OutStreamer.EmitRawText(Twine("\t.address_size ") + addrSize);
152 OutStreamer.AddBlankLine();
154 // Define any .file directives
155 DebugInfoFinder DbgFinder;
156 DbgFinder.processModule(M);
158 for (DebugInfoFinder::iterator I = DbgFinder.compile_unit_begin(),
159 E = DbgFinder.compile_unit_end(); I != E; ++I) {
160 DICompileUnit DIUnit(*I);
161 StringRef FN = DIUnit.getFilename();
162 StringRef Dir = DIUnit.getDirectory();
163 GetOrCreateSourceID(FN, Dir);
166 OutStreamer.AddBlankLine();
168 // declare external functions
169 for (Module::const_iterator i = M.begin(), e = M.end();
171 EmitFunctionDeclaration(i);
173 // declare global variables
174 for (Module::const_global_iterator i = M.global_begin(), e = M.global_end();
176 EmitVariableDeclaration(i);
179 void PTXAsmPrinter::EmitFunctionBodyStart() {
180 OutStreamer.EmitRawText(Twine("{"));
182 const PTXMachineFunctionInfo *MFI = MF->getInfo<PTXMachineFunctionInfo>();
183 const PTXParamManager &PM = MFI->getParamManager();
185 // Print register definitions
186 SmallString<128> regDefs;
187 raw_svector_ostream os(regDefs);
191 numRegs = MFI->countRegisters(PTXRegisterType::Pred, PTXRegisterSpace::Reg);
193 os << "\t.reg .pred %p<" << numRegs << ">;\n";
196 numRegs = MFI->countRegisters(PTXRegisterType::B16, PTXRegisterSpace::Reg);
198 os << "\t.reg .b16 %rh<" << numRegs << ">;\n";
201 numRegs = MFI->countRegisters(PTXRegisterType::B32, PTXRegisterSpace::Reg);
203 os << "\t.reg .b32 %r<" << numRegs << ">;\n";
206 numRegs = MFI->countRegisters(PTXRegisterType::B64, PTXRegisterSpace::Reg);
208 os << "\t.reg .b64 %rd<" << numRegs << ">;\n";
211 numRegs = MFI->countRegisters(PTXRegisterType::F32, PTXRegisterSpace::Reg);
213 os << "\t.reg .f32 %f<" << numRegs << ">;\n";
216 numRegs = MFI->countRegisters(PTXRegisterType::F64, PTXRegisterSpace::Reg);
218 os << "\t.reg .f64 %fd<" << numRegs << ">;\n";
221 for (PTXParamManager::param_iterator i = PM.local_begin(), e = PM.local_end();
223 os << "\t.param .b" << PM.getParamSize(*i) << ' ' << PM.getParamName(*i)
226 OutStreamer.EmitRawText(os.str());
229 const MachineFrameInfo* FrameInfo = MF->getFrameInfo();
230 DEBUG(dbgs() << "Have " << FrameInfo->getNumObjects()
231 << " frame object(s)\n");
232 for (unsigned i = 0, e = FrameInfo->getNumObjects(); i != e; ++i) {
233 DEBUG(dbgs() << "Size of object: " << FrameInfo->getObjectSize(i) << "\n");
234 if (FrameInfo->getObjectSize(i) > 0) {
235 OutStreamer.EmitRawText("\t.local .align " +
236 Twine(FrameInfo->getObjectAlignment(i)) +
240 Twine(FrameInfo->getObjectSize(i)) +
245 //unsigned Index = 1;
246 // Print parameter passing params
247 //for (PTXMachineFunctionInfo::param_iterator
248 // i = MFI->paramBegin(), e = MFI->paramEnd(); i != e; ++i) {
249 // std::string def = "\t.param .b";
250 // def += utostr(*i);
252 // def += utostr(Index);
255 // OutStreamer.EmitRawText(Twine(def));
259 void PTXAsmPrinter::EmitFunctionBodyEnd() {
260 OutStreamer.EmitRawText(Twine("}"));
263 void PTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
265 LowerPTXMachineInstrToMCInst(MI, TmpInst, *this);
266 OutStreamer.EmitInstruction(TmpInst);
269 void PTXAsmPrinter::EmitVariableDeclaration(const GlobalVariable *gv) {
270 // Check to see if this is a special global used by LLVM, if so, emit it.
271 if (EmitSpecialLLVMGlobal(gv))
274 MCSymbol *gvsym = Mang->getSymbol(gv);
276 assert(gvsym->isUndefined() && "Cannot define a symbol twice!");
278 SmallString<128> decl;
279 raw_svector_ostream os(decl);
281 // check if it is defined in some other translation unit
282 if (gv->isDeclaration())
285 // state space: e.g., .global
286 os << '.' << getStateSpaceName(gv->getType()->getAddressSpace()) << ' ';
288 // alignment (optional)
289 unsigned alignment = gv->getAlignment();
291 os << ".align " << gv->getAlignment() << ' ';
294 if (PointerType::classof(gv->getType())) {
295 PointerType* pointerTy = dyn_cast<PointerType>(gv->getType());
296 Type* elementTy = pointerTy->getElementType();
298 if (elementTy->isArrayTy()) {
299 assert(elementTy->isArrayTy() && "Only pointers to arrays are supported");
301 ArrayType* arrayTy = dyn_cast<ArrayType>(elementTy);
302 elementTy = arrayTy->getElementType();
304 unsigned numElements = arrayTy->getNumElements();
306 while (elementTy->isArrayTy()) {
307 arrayTy = dyn_cast<ArrayType>(elementTy);
308 elementTy = arrayTy->getElementType();
310 numElements *= arrayTy->getNumElements();
313 // FIXME: isPrimitiveType() == false for i16?
314 assert(elementTy->isSingleValueType() &&
315 "Non-primitive types are not handled");
317 // Find the size of the element in bits
318 unsigned elementSize = elementTy->getPrimitiveSizeInBits();
320 os << ".b" << elementSize << ' ' << gvsym->getName()
321 << '[' << numElements << ']';
323 os << ".b8" << gvsym->getName() << "[]";
326 // handle string constants (assume ConstantArray means string)
327 if (gv->hasInitializer()) {
328 const Constant *C = gv->getInitializer();
329 if (const ConstantArray *CA = dyn_cast<ConstantArray>(C)) {
332 for (unsigned i = 0, e = C->getNumOperands(); i != e; ++i) {
337 os.write_hex(cast<ConstantInt>(CA->getOperand(i))->getZExtValue());
344 // Note: this is currently the fall-through case and most likely generates
346 os << getTypeName(gv->getType()) << ' ' << gvsym->getName();
348 if (isa<ArrayType>(gv->getType()) || isa<PointerType>(gv->getType()))
354 OutStreamer.EmitRawText(os.str());
355 OutStreamer.AddBlankLine();
358 void PTXAsmPrinter::EmitFunctionEntryLabel() {
359 // The function label could have already been emitted if two symbols end up
360 // conflicting due to asm renaming. Detect this and emit an error.
361 if (!CurrentFnSym->isUndefined()) {
362 report_fatal_error("'" + Twine(CurrentFnSym->getName()) +
363 "' label emitted multiple times to assembly file");
367 const PTXMachineFunctionInfo *MFI = MF->getInfo<PTXMachineFunctionInfo>();
368 const PTXParamManager &PM = MFI->getParamManager();
369 const bool isKernel = MFI->isKernel();
370 const PTXSubtarget& ST = TM.getSubtarget<PTXSubtarget>();
372 SmallString<128> decl;
373 raw_svector_ostream os(decl);
374 os << (isKernel ? ".entry" : ".func");
378 if (ST.useParamSpaceForDeviceArgs()) {
379 for (PTXParamManager::param_iterator i = PM.ret_begin(), e = PM.ret_end(),
380 b = i; i != e; ++i) {
384 os << ".param .b" << PM.getParamSize(*i) << ' ' << PM.getParamName(*i);
387 for (PTXMachineFunctionInfo::reg_iterator
388 i = MFI->retreg_begin(), e = MFI->retreg_end(), b = i;
393 os << ".reg " << getRegisterTypeName(MFI->getRegisterType(*i)) << ' '
394 << MFI->getRegisterName(*i);
400 // Print function name
401 os << ' ' << CurrentFnSym->getName() << " (";
403 const Function *F = MF->getFunction();
406 if (isKernel || ST.useParamSpaceForDeviceArgs()) {
407 /*for (PTXParamManager::param_iterator i = PM.arg_begin(), e = PM.arg_end(),
408 b = i; i != e; ++i) {
412 os << ".param .b" << PM.getParamSize(*i) << ' ' << PM.getParamName(*i);
415 for (Function::const_arg_iterator i = F->arg_begin(), e = F->arg_end(),
416 b = i; i != e; ++i) {
419 const Type *ArgType = (*i).getType();
421 if (ArgType->isPointerTy()) {
427 os << ArgType->getPrimitiveSizeInBits();
429 if (ArgType->isPointerTy() && ST.emitPtrAttribute()) {
430 const PointerType *PtrType = dyn_cast<const PointerType>(ArgType);
432 switch (PtrType->getAddressSpace()) {
434 llvm_unreachable("Unknown address space in argument");
435 case PTXStateSpace::Global:
438 case PTXStateSpace::Shared:
443 os << " __param_" << Counter++;
446 for (PTXMachineFunctionInfo::reg_iterator
447 i = MFI->argreg_begin(), e = MFI->argreg_end(), b = i;
452 os << ".reg " << getRegisterTypeName(MFI->getRegisterType(*i)) << ' '
453 << MFI->getRegisterName(*i);
458 OutStreamer.EmitRawText(os.str());
461 void PTXAsmPrinter::EmitFunctionDeclaration(const Function* func)
463 const PTXSubtarget& ST = TM.getSubtarget<PTXSubtarget>();
465 std::string decl = "";
467 // hard-coded emission of extern vprintf function
469 if (func->getName() == "printf" || func->getName() == "puts") {
470 decl += ".extern .func (.param .b32 __param_1) vprintf (.param .b";
475 decl += " __param_2, .param .b";
480 decl += " __param_3)\n";
483 OutStreamer.EmitRawText(Twine(decl));
486 unsigned PTXAsmPrinter::GetOrCreateSourceID(StringRef FileName,
488 // If FE did not provide a file name, then assume stdin.
489 if (FileName.empty())
490 return GetOrCreateSourceID("<stdin>", StringRef());
492 // MCStream expects full path name as filename.
493 if (!DirName.empty() && !sys::path::is_absolute(FileName)) {
494 SmallString<128> FullPathName = DirName;
495 sys::path::append(FullPathName, FileName);
496 // Here FullPathName will be copied into StringMap by GetOrCreateSourceID.
497 return GetOrCreateSourceID(StringRef(FullPathName), StringRef());
500 StringMapEntry<unsigned> &Entry = SourceIdMap.GetOrCreateValue(FileName);
501 if (Entry.getValue())
502 return Entry.getValue();
504 unsigned SrcId = SourceIdMap.size();
505 Entry.setValue(SrcId);
507 // Print out a .file directive to specify files for .loc directives.
508 OutStreamer.EmitDwarfFileDirective(SrcId, "", Entry.getKey());
513 MCOperand PTXAsmPrinter::GetSymbolRef(const MachineOperand &MO,
514 const MCSymbol *Symbol) {
516 Expr = MCSymbolRefExpr::Create(Symbol, MCSymbolRefExpr::VK_None, OutContext);
517 return MCOperand::CreateExpr(Expr);
520 MCOperand PTXAsmPrinter::lowerOperand(const MachineOperand &MO) {
522 const PTXMachineFunctionInfo *MFI = MF->getInfo<PTXMachineFunctionInfo>();
524 switch (MO.getType()) {
526 llvm_unreachable("Unknown operand type");
527 case MachineOperand::MO_Register:
528 if (MO.getReg() > 0) {
529 // Encode the register
530 EncodedReg = MFI->getEncodedRegister(MO.getReg());
534 MCOp = MCOperand::CreateReg(EncodedReg);
536 case MachineOperand::MO_Immediate:
537 MCOp = MCOperand::CreateImm(MO.getImm());
539 case MachineOperand::MO_MachineBasicBlock:
540 MCOp = MCOperand::CreateExpr(MCSymbolRefExpr::Create(
541 MO.getMBB()->getSymbol(), OutContext));
543 case MachineOperand::MO_GlobalAddress:
544 MCOp = GetSymbolRef(MO, Mang->getSymbol(MO.getGlobal()));
546 case MachineOperand::MO_ExternalSymbol:
547 MCOp = GetSymbolRef(MO, GetExternalSymbolSymbol(MO.getSymbolName()));
549 case MachineOperand::MO_FPImmediate:
550 APFloat Val = MO.getFPImm()->getValueAPF();
552 Val.convert(APFloat::IEEEdouble, APFloat::rmTowardZero, &ignored);
553 MCOp = MCOperand::CreateFPImm(Val.convertToDouble());
560 // Force static initialization.
561 extern "C" void LLVMInitializePTXAsmPrinter() {
562 RegisterAsmPrinter<PTXAsmPrinter> X(ThePTX32Target);
563 RegisterAsmPrinter<PTXAsmPrinter> Y(ThePTX64Target);