1 //===-- NVPTXAsmPrinter.cpp - NVPTX 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 NVPTX assembly language.
13 //===----------------------------------------------------------------------===//
15 #include "NVPTXAsmPrinter.h"
16 #include "MCTargetDesc/NVPTXMCAsmInfo.h"
18 #include "NVPTXInstrInfo.h"
19 #include "NVPTXNumRegisters.h"
20 #include "NVPTXRegisterInfo.h"
21 #include "NVPTXTargetMachine.h"
22 #include "NVPTXUtilities.h"
23 #include "cl_common_defines.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Assembly/Writer.h"
27 #include "llvm/CodeGen/Analysis.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineModuleInfo.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/DebugInfo.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/GlobalVariable.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/MC/MCStreamer.h"
38 #include "llvm/MC/MCSymbol.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/FormattedStream.h"
42 #include "llvm/Support/Path.h"
43 #include "llvm/Support/TargetRegistry.h"
44 #include "llvm/Support/TimeValue.h"
45 #include "llvm/Target/Mangler.h"
46 #include "llvm/Target/TargetLoweringObjectFile.h"
50 #include "NVPTXGenAsmWriter.inc"
52 bool RegAllocNilUsed = true;
54 #define DEPOTNAME "__local_depot"
57 EmitLineNumbers("nvptx-emit-line-numbers",
58 cl::desc("NVPTX Specific: Emit Line numbers even without -G"),
61 namespace llvm { bool InterleaveSrcInPtx = false; }
63 static cl::opt<bool, true>
64 InterleaveSrc("nvptx-emit-src", cl::ZeroOrMore,
65 cl::desc("NVPTX Specific: Emit source line in ptx file"),
66 cl::location(llvm::InterleaveSrcInPtx));
69 /// DiscoverDependentGlobals - Return a set of GlobalVariables on which \p V
71 void DiscoverDependentGlobals(const Value *V,
72 DenseSet<const GlobalVariable *> &Globals) {
73 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
76 if (const User *U = dyn_cast<User>(V)) {
77 for (unsigned i = 0, e = U->getNumOperands(); i != e; ++i) {
78 DiscoverDependentGlobals(U->getOperand(i), Globals);
84 /// VisitGlobalVariableForEmission - Add \p GV to the list of GlobalVariable
85 /// instances to be emitted, but only after any dependents have been added
87 void VisitGlobalVariableForEmission(
88 const GlobalVariable *GV, SmallVectorImpl<const GlobalVariable *> &Order,
89 DenseSet<const GlobalVariable *> &Visited,
90 DenseSet<const GlobalVariable *> &Visiting) {
91 // Have we already visited this one?
92 if (Visited.count(GV))
95 // Do we have a circular dependency?
96 if (Visiting.count(GV))
97 report_fatal_error("Circular dependency found in global variable set");
99 // Start visiting this global
102 // Make sure we visit all dependents first
103 DenseSet<const GlobalVariable *> Others;
104 for (unsigned i = 0, e = GV->getNumOperands(); i != e; ++i)
105 DiscoverDependentGlobals(GV->getOperand(i), Others);
107 for (DenseSet<const GlobalVariable *>::iterator I = Others.begin(),
110 VisitGlobalVariableForEmission(*I, Order, Visited, Visiting);
112 // Now we can visit ourself
119 // @TODO: This is a copy from AsmPrinter.cpp. The function is static, so we
120 // cannot just link to the existing version.
121 /// LowerConstant - Lower the specified LLVM Constant to an MCExpr.
123 using namespace nvptx;
124 const MCExpr *nvptx::LowerConstant(const Constant *CV, AsmPrinter &AP) {
125 MCContext &Ctx = AP.OutContext;
127 if (CV->isNullValue() || isa<UndefValue>(CV))
128 return MCConstantExpr::Create(0, Ctx);
130 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV))
131 return MCConstantExpr::Create(CI->getZExtValue(), Ctx);
133 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CV))
134 return MCSymbolRefExpr::Create(AP.Mang->getSymbol(GV), Ctx);
136 if (const BlockAddress *BA = dyn_cast<BlockAddress>(CV))
137 return MCSymbolRefExpr::Create(AP.GetBlockAddressSymbol(BA), Ctx);
139 const ConstantExpr *CE = dyn_cast<ConstantExpr>(CV);
141 llvm_unreachable("Unknown constant value to lower!");
143 switch (CE->getOpcode()) {
145 // If the code isn't optimized, there may be outstanding folding
146 // opportunities. Attempt to fold the expression using DataLayout as a
147 // last resort before giving up.
148 if (Constant *C = ConstantFoldConstantExpression(CE, AP.TM.getDataLayout()))
150 return LowerConstant(C, AP);
152 // Otherwise report the problem to the user.
155 raw_string_ostream OS(S);
156 OS << "Unsupported expression in static initializer: ";
157 WriteAsOperand(OS, CE, /*PrintType=*/ false,
158 !AP.MF ? 0 : AP.MF->getFunction()->getParent());
159 report_fatal_error(OS.str());
161 case Instruction::GetElementPtr: {
162 const DataLayout &TD = *AP.TM.getDataLayout();
163 // Generate a symbolic expression for the byte address
164 APInt OffsetAI(TD.getPointerSizeInBits(), 0);
165 cast<GEPOperator>(CE)->accumulateConstantOffset(TD, OffsetAI);
167 const MCExpr *Base = LowerConstant(CE->getOperand(0), AP);
171 int64_t Offset = OffsetAI.getSExtValue();
172 return MCBinaryExpr::CreateAdd(Base, MCConstantExpr::Create(Offset, Ctx),
176 case Instruction::Trunc:
177 // We emit the value and depend on the assembler to truncate the generated
178 // expression properly. This is important for differences between
179 // blockaddress labels. Since the two labels are in the same function, it
180 // is reasonable to treat their delta as a 32-bit value.
182 case Instruction::BitCast:
183 return LowerConstant(CE->getOperand(0), AP);
185 case Instruction::IntToPtr: {
186 const DataLayout &TD = *AP.TM.getDataLayout();
187 // Handle casts to pointers by changing them into casts to the appropriate
188 // integer type. This promotes constant folding and simplifies this code.
189 Constant *Op = CE->getOperand(0);
190 Op = ConstantExpr::getIntegerCast(Op, TD.getIntPtrType(CV->getContext()),
192 return LowerConstant(Op, AP);
195 case Instruction::PtrToInt: {
196 const DataLayout &TD = *AP.TM.getDataLayout();
197 // Support only foldable casts to/from pointers that can be eliminated by
198 // changing the pointer to the appropriately sized integer type.
199 Constant *Op = CE->getOperand(0);
200 Type *Ty = CE->getType();
202 const MCExpr *OpExpr = LowerConstant(Op, AP);
204 // We can emit the pointer value into this slot if the slot is an
205 // integer slot equal to the size of the pointer.
206 if (TD.getTypeAllocSize(Ty) == TD.getTypeAllocSize(Op->getType()))
209 // Otherwise the pointer is smaller than the resultant integer, mask off
210 // the high bits so we are sure to get a proper truncation if the input is
212 unsigned InBits = TD.getTypeAllocSizeInBits(Op->getType());
213 const MCExpr *MaskExpr =
214 MCConstantExpr::Create(~0ULL >> (64 - InBits), Ctx);
215 return MCBinaryExpr::CreateAnd(OpExpr, MaskExpr, Ctx);
218 // The MC library also has a right-shift operator, but it isn't consistently
219 // signed or unsigned between different targets.
220 case Instruction::Add:
221 case Instruction::Sub:
222 case Instruction::Mul:
223 case Instruction::SDiv:
224 case Instruction::SRem:
225 case Instruction::Shl:
226 case Instruction::And:
227 case Instruction::Or:
228 case Instruction::Xor: {
229 const MCExpr *LHS = LowerConstant(CE->getOperand(0), AP);
230 const MCExpr *RHS = LowerConstant(CE->getOperand(1), AP);
231 switch (CE->getOpcode()) {
233 llvm_unreachable("Unknown binary operator constant cast expr");
234 case Instruction::Add:
235 return MCBinaryExpr::CreateAdd(LHS, RHS, Ctx);
236 case Instruction::Sub:
237 return MCBinaryExpr::CreateSub(LHS, RHS, Ctx);
238 case Instruction::Mul:
239 return MCBinaryExpr::CreateMul(LHS, RHS, Ctx);
240 case Instruction::SDiv:
241 return MCBinaryExpr::CreateDiv(LHS, RHS, Ctx);
242 case Instruction::SRem:
243 return MCBinaryExpr::CreateMod(LHS, RHS, Ctx);
244 case Instruction::Shl:
245 return MCBinaryExpr::CreateShl(LHS, RHS, Ctx);
246 case Instruction::And:
247 return MCBinaryExpr::CreateAnd(LHS, RHS, Ctx);
248 case Instruction::Or:
249 return MCBinaryExpr::CreateOr(LHS, RHS, Ctx);
250 case Instruction::Xor:
251 return MCBinaryExpr::CreateXor(LHS, RHS, Ctx);
257 void NVPTXAsmPrinter::emitLineNumberAsDotLoc(const MachineInstr &MI) {
258 if (!EmitLineNumbers)
263 DebugLoc curLoc = MI.getDebugLoc();
265 if (prevDebugLoc.isUnknown() && curLoc.isUnknown())
268 if (prevDebugLoc == curLoc)
271 prevDebugLoc = curLoc;
273 if (curLoc.isUnknown())
276 const MachineFunction *MF = MI.getParent()->getParent();
277 //const TargetMachine &TM = MF->getTarget();
279 const LLVMContext &ctx = MF->getFunction()->getContext();
280 DIScope Scope(curLoc.getScope(ctx));
285 StringRef fileName(Scope.getFilename());
286 StringRef dirName(Scope.getDirectory());
287 SmallString<128> FullPathName = dirName;
288 if (!dirName.empty() && !sys::path::is_absolute(fileName)) {
289 sys::path::append(FullPathName, fileName);
290 fileName = FullPathName.str();
293 if (filenameMap.find(fileName.str()) == filenameMap.end())
296 // Emit the line from the source file.
297 if (llvm::InterleaveSrcInPtx)
298 this->emitSrcInText(fileName.str(), curLoc.getLine());
300 std::stringstream temp;
301 temp << "\t.loc " << filenameMap[fileName.str()] << " " << curLoc.getLine()
302 << " " << curLoc.getCol();
303 OutStreamer.EmitRawText(Twine(temp.str().c_str()));
306 void NVPTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
307 SmallString<128> Str;
308 raw_svector_ostream OS(Str);
309 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
310 emitLineNumberAsDotLoc(*MI);
311 printInstruction(MI, OS);
312 OutStreamer.EmitRawText(OS.str());
315 void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
316 const DataLayout *TD = TM.getDataLayout();
317 const TargetLowering *TLI = TM.getTargetLowering();
319 Type *Ty = F->getReturnType();
321 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
323 if (Ty->getTypeID() == Type::VoidTyID)
329 if (Ty->isPrimitiveType() || Ty->isIntegerTy()) {
331 if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty)) {
332 size = ITy->getBitWidth();
336 assert(Ty->isFloatingPointTy() && "Floating point type expected here");
337 size = Ty->getPrimitiveSizeInBits();
340 O << ".param .b" << size << " func_retval0";
341 } else if (isa<PointerType>(Ty)) {
342 O << ".param .b" << TLI->getPointerTy().getSizeInBits()
345 if ((Ty->getTypeID() == Type::StructTyID) || isa<VectorType>(Ty)) {
346 SmallVector<EVT, 16> vtparts;
347 ComputeValueVTs(*TLI, Ty, vtparts);
348 unsigned totalsz = 0;
349 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
351 EVT elemtype = vtparts[i];
352 if (vtparts[i].isVector()) {
353 elems = vtparts[i].getVectorNumElements();
354 elemtype = vtparts[i].getVectorElementType();
356 for (unsigned j = 0, je = elems; j != je; ++j) {
357 unsigned sz = elemtype.getSizeInBits();
358 if (elemtype.isInteger() && (sz < 8))
363 unsigned retAlignment = 0;
364 if (!llvm::getAlign(*F, 0, retAlignment))
365 retAlignment = TD->getABITypeAlignment(Ty);
366 O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
369 assert(false && "Unknown return type");
372 SmallVector<EVT, 16> vtparts;
373 ComputeValueVTs(*TLI, Ty, vtparts);
375 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
377 EVT elemtype = vtparts[i];
378 if (vtparts[i].isVector()) {
379 elems = vtparts[i].getVectorNumElements();
380 elemtype = vtparts[i].getVectorElementType();
383 for (unsigned j = 0, je = elems; j != je; ++j) {
384 unsigned sz = elemtype.getSizeInBits();
385 if (elemtype.isInteger() && (sz < 32))
387 O << ".reg .b" << sz << " func_retval" << idx;
400 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
402 const Function *F = MF.getFunction();
403 printReturnValStr(F, O);
406 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
407 SmallString<128> Str;
408 raw_svector_ostream O(Str);
410 if (!GlobalsEmitted) {
411 emitGlobals(*MF->getFunction()->getParent());
412 GlobalsEmitted = true;
416 MRI = &MF->getRegInfo();
417 F = MF->getFunction();
418 emitLinkageDirective(F, O);
419 if (llvm::isKernelFunction(*F))
423 printReturnValStr(*MF, O);
428 emitFunctionParamList(*MF, O);
430 if (llvm::isKernelFunction(*F))
431 emitKernelFunctionDirectives(*F, O);
433 OutStreamer.EmitRawText(O.str());
435 prevDebugLoc = DebugLoc();
438 void NVPTXAsmPrinter::EmitFunctionBodyStart() {
439 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
440 unsigned numRegClasses = TRI.getNumRegClasses();
441 VRidGlobal2LocalMap = new std::map<unsigned, unsigned>[numRegClasses + 1];
442 OutStreamer.EmitRawText(StringRef("{\n"));
443 setAndEmitFunctionVirtualRegisters(*MF);
445 SmallString<128> Str;
446 raw_svector_ostream O(Str);
447 emitDemotedVars(MF->getFunction(), O);
448 OutStreamer.EmitRawText(O.str());
451 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
452 OutStreamer.EmitRawText(StringRef("}\n"));
453 delete[] VRidGlobal2LocalMap;
456 void NVPTXAsmPrinter::emitKernelFunctionDirectives(const Function &F,
457 raw_ostream &O) const {
458 // If the NVVM IR has some of reqntid* specified, then output
459 // the reqntid directive, and set the unspecified ones to 1.
460 // If none of reqntid* is specified, don't output reqntid directive.
461 unsigned reqntidx, reqntidy, reqntidz;
462 bool specified = false;
463 if (llvm::getReqNTIDx(F, reqntidx) == false)
467 if (llvm::getReqNTIDy(F, reqntidy) == false)
471 if (llvm::getReqNTIDz(F, reqntidz) == false)
477 O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
480 // If the NVVM IR has some of maxntid* specified, then output
481 // the maxntid directive, and set the unspecified ones to 1.
482 // If none of maxntid* is specified, don't output maxntid directive.
483 unsigned maxntidx, maxntidy, maxntidz;
485 if (llvm::getMaxNTIDx(F, maxntidx) == false)
489 if (llvm::getMaxNTIDy(F, maxntidy) == false)
493 if (llvm::getMaxNTIDz(F, maxntidz) == false)
499 O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
503 if (llvm::getMinCTASm(F, mincta))
504 O << ".minnctapersm " << mincta << "\n";
507 void NVPTXAsmPrinter::getVirtualRegisterName(unsigned vr, bool isVec,
509 const TargetRegisterClass *RC = MRI->getRegClass(vr);
510 unsigned id = RC->getID();
512 std::map<unsigned, unsigned> ®map = VRidGlobal2LocalMap[id];
513 unsigned mapped_vr = regmap[vr];
516 O << getNVPTXRegClassStr(RC) << mapped_vr;
519 report_fatal_error("Bad register!");
522 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr, bool isVec,
524 getVirtualRegisterName(vr, isVec, O);
527 void NVPTXAsmPrinter::printVecModifiedImmediate(
528 const MachineOperand &MO, const char *Modifier, raw_ostream &O) {
529 static const char vecelem[] = { '0', '1', '2', '3', '0', '1', '2', '3' };
530 int Imm = (int) MO.getImm();
531 if (0 == strcmp(Modifier, "vecelem"))
532 O << "_" << vecelem[Imm];
533 else if (0 == strcmp(Modifier, "vecv4comm1")) {
534 if ((Imm < 0) || (Imm > 3))
536 } else if (0 == strcmp(Modifier, "vecv4comm2")) {
537 if ((Imm < 4) || (Imm > 7))
539 } else if (0 == strcmp(Modifier, "vecv4pos")) {
542 O << "_" << vecelem[Imm % 4];
543 } else if (0 == strcmp(Modifier, "vecv2comm1")) {
544 if ((Imm < 0) || (Imm > 1))
546 } else if (0 == strcmp(Modifier, "vecv2comm2")) {
547 if ((Imm < 2) || (Imm > 3))
549 } else if (0 == strcmp(Modifier, "vecv2pos")) {
552 O << "_" << vecelem[Imm % 2];
554 llvm_unreachable("Unknown Modifier on immediate operand");
557 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
558 raw_ostream &O, const char *Modifier) {
559 const MachineOperand &MO = MI->getOperand(opNum);
560 switch (MO.getType()) {
561 case MachineOperand::MO_Register:
562 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
563 if (MO.getReg() == NVPTX::VRDepot)
564 O << DEPOTNAME << getFunctionNumber();
566 O << getRegisterName(MO.getReg());
569 emitVirtualRegister(MO.getReg(), false, O);
571 if (strcmp(Modifier, "vecfull") == 0)
572 emitVirtualRegister(MO.getReg(), true, O);
575 "Don't know how to handle the modifier on virtual register.");
580 case MachineOperand::MO_Immediate:
583 else if (strstr(Modifier, "vec") == Modifier)
584 printVecModifiedImmediate(MO, Modifier, O);
587 "Don't know how to handle modifier on immediate operand");
590 case MachineOperand::MO_FPImmediate:
591 printFPConstant(MO.getFPImm(), O);
594 case MachineOperand::MO_GlobalAddress:
595 O << *Mang->getSymbol(MO.getGlobal());
598 case MachineOperand::MO_ExternalSymbol: {
599 const char *symbname = MO.getSymbolName();
600 if (strstr(symbname, ".PARAM") == symbname) {
602 sscanf(symbname + 6, "%u[];", &index);
603 printParamName(index, O);
604 } else if (strstr(symbname, ".HLPPARAM") == symbname) {
606 sscanf(symbname + 9, "%u[];", &index);
607 O << *CurrentFnSym << "_param_" << index << "_offset";
613 case MachineOperand::MO_MachineBasicBlock:
614 O << *MO.getMBB()->getSymbol();
618 llvm_unreachable("Operand type not supported.");
622 void NVPTXAsmPrinter::printImplicitDef(const MachineInstr *MI,
623 raw_ostream &O) const {
625 O << "\t// Implicit def :";
626 //printOperand(MI, 0);
631 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
632 raw_ostream &O, const char *Modifier) {
633 printOperand(MI, opNum, O);
635 if (Modifier && !strcmp(Modifier, "add")) {
637 printOperand(MI, opNum + 1, O);
639 if (MI->getOperand(opNum + 1).isImm() &&
640 MI->getOperand(opNum + 1).getImm() == 0)
641 return; // don't print ',0' or '+0'
643 printOperand(MI, opNum + 1, O);
647 void NVPTXAsmPrinter::printLdStCode(const MachineInstr *MI, int opNum,
648 raw_ostream &O, const char *Modifier) {
650 const MachineOperand &MO = MI->getOperand(opNum);
651 int Imm = (int) MO.getImm();
652 if (!strcmp(Modifier, "volatile")) {
655 } else if (!strcmp(Modifier, "addsp")) {
657 case NVPTX::PTXLdStInstCode::GLOBAL:
660 case NVPTX::PTXLdStInstCode::SHARED:
663 case NVPTX::PTXLdStInstCode::LOCAL:
666 case NVPTX::PTXLdStInstCode::PARAM:
669 case NVPTX::PTXLdStInstCode::CONSTANT:
672 case NVPTX::PTXLdStInstCode::GENERIC:
673 if (!nvptxSubtarget.hasGenericLdSt())
677 llvm_unreachable("Wrong Address Space");
679 } else if (!strcmp(Modifier, "sign")) {
680 if (Imm == NVPTX::PTXLdStInstCode::Signed)
682 else if (Imm == NVPTX::PTXLdStInstCode::Unsigned)
686 } else if (!strcmp(Modifier, "vec")) {
687 if (Imm == NVPTX::PTXLdStInstCode::V2)
689 else if (Imm == NVPTX::PTXLdStInstCode::V4)
692 llvm_unreachable("Unknown Modifier");
694 llvm_unreachable("Empty Modifier");
697 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
699 emitLinkageDirective(F, O);
700 if (llvm::isKernelFunction(*F))
704 printReturnValStr(F, O);
705 O << *Mang->getSymbol(F) << "\n";
706 emitFunctionParamList(F, O);
710 static bool usedInGlobalVarDef(const Constant *C) {
714 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
715 if (GV->getName().str() == "llvm.used")
720 for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
722 const Constant *C = dyn_cast<Constant>(*ui);
723 if (usedInGlobalVarDef(C))
729 static bool usedInOneFunc(const User *U, Function const *&oneFunc) {
730 if (const GlobalVariable *othergv = dyn_cast<GlobalVariable>(U)) {
731 if (othergv->getName().str() == "llvm.used")
735 if (const Instruction *instr = dyn_cast<Instruction>(U)) {
736 if (instr->getParent() && instr->getParent()->getParent()) {
737 const Function *curFunc = instr->getParent()->getParent();
738 if (oneFunc && (curFunc != oneFunc))
746 if (const MDNode *md = dyn_cast<MDNode>(U))
747 if (md->hasName() && ((md->getName().str() == "llvm.dbg.gv") ||
748 (md->getName().str() == "llvm.dbg.sp")))
751 for (User::const_use_iterator ui = U->use_begin(), ue = U->use_end();
753 if (usedInOneFunc(*ui, oneFunc) == false)
759 /* Find out if a global variable can be demoted to local scope.
760 * Currently, this is valid for CUDA shared variables, which have local
761 * scope and global lifetime. So the conditions to check are :
762 * 1. Is the global variable in shared address space?
763 * 2. Does it have internal linkage?
764 * 3. Is the global variable referenced only in one function?
766 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
767 if (gv->hasInternalLinkage() == false)
769 const PointerType *Pty = gv->getType();
770 if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
773 const Function *oneFunc = 0;
775 bool flag = usedInOneFunc(gv, oneFunc);
784 static bool useFuncSeen(const Constant *C,
785 llvm::DenseMap<const Function *, bool> &seenMap) {
786 for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
788 if (const Constant *cu = dyn_cast<Constant>(*ui)) {
789 if (useFuncSeen(cu, seenMap))
791 } else if (const Instruction *I = dyn_cast<Instruction>(*ui)) {
792 const BasicBlock *bb = I->getParent();
795 const Function *caller = bb->getParent();
798 if (seenMap.find(caller) != seenMap.end())
805 void NVPTXAsmPrinter::emitDeclarations(const Module &M, raw_ostream &O) {
806 llvm::DenseMap<const Function *, bool> seenMap;
807 for (Module::const_iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
808 const Function *F = FI;
810 if (F->isDeclaration()) {
813 if (F->getIntrinsicID())
815 emitDeclaration(F, O);
818 for (Value::const_use_iterator iter = F->use_begin(),
819 iterEnd = F->use_end();
820 iter != iterEnd; ++iter) {
821 if (const Constant *C = dyn_cast<Constant>(*iter)) {
822 if (usedInGlobalVarDef(C)) {
823 // The use is in the initialization of a global variable
824 // that is a function pointer, so print a declaration
825 // for the original function
826 emitDeclaration(F, O);
829 // Emit a declaration of this function if the function that
830 // uses this constant expr has already been seen.
831 if (useFuncSeen(C, seenMap)) {
832 emitDeclaration(F, O);
837 if (!isa<Instruction>(*iter))
839 const Instruction *instr = cast<Instruction>(*iter);
840 const BasicBlock *bb = instr->getParent();
843 const Function *caller = bb->getParent();
847 // If a caller has already been seen, then the caller is
848 // appearing in the module before the callee. so print out
849 // a declaration for the callee.
850 if (seenMap.find(caller) != seenMap.end()) {
851 emitDeclaration(F, O);
859 void NVPTXAsmPrinter::recordAndEmitFilenames(Module &M) {
860 DebugInfoFinder DbgFinder;
861 DbgFinder.processModule(M);
864 for (DebugInfoFinder::iterator I = DbgFinder.compile_unit_begin(),
865 E = DbgFinder.compile_unit_end();
867 DICompileUnit DIUnit(*I);
868 StringRef Filename(DIUnit.getFilename());
869 StringRef Dirname(DIUnit.getDirectory());
870 SmallString<128> FullPathName = Dirname;
871 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
872 sys::path::append(FullPathName, Filename);
873 Filename = FullPathName.str();
875 if (filenameMap.find(Filename.str()) != filenameMap.end())
877 filenameMap[Filename.str()] = i;
878 OutStreamer.EmitDwarfFileDirective(i, "", Filename.str());
882 for (DebugInfoFinder::iterator I = DbgFinder.subprogram_begin(),
883 E = DbgFinder.subprogram_end();
886 StringRef Filename(SP.getFilename());
887 StringRef Dirname(SP.getDirectory());
888 SmallString<128> FullPathName = Dirname;
889 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
890 sys::path::append(FullPathName, Filename);
891 Filename = FullPathName.str();
893 if (filenameMap.find(Filename.str()) != filenameMap.end())
895 filenameMap[Filename.str()] = i;
900 bool NVPTXAsmPrinter::doInitialization(Module &M) {
902 SmallString<128> Str1;
903 raw_svector_ostream OS1(Str1);
905 MMI = getAnalysisIfAvailable<MachineModuleInfo>();
906 MMI->AnalyzeModule(M);
908 // We need to call the parent's one explicitly.
909 //bool Result = AsmPrinter::doInitialization(M);
911 // Initialize TargetLoweringObjectFile.
912 const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
913 .Initialize(OutContext, TM);
915 Mang = new Mangler(OutContext, *TM.getDataLayout());
917 // Emit header before any dwarf directives are emitted below.
919 OutStreamer.EmitRawText(OS1.str());
921 // Already commented out
922 //bool Result = AsmPrinter::doInitialization(M);
924 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
925 recordAndEmitFilenames(M);
927 GlobalsEmitted = false;
929 return false; // success
932 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
933 SmallString<128> Str2;
934 raw_svector_ostream OS2(Str2);
936 emitDeclarations(M, OS2);
938 // As ptxas does not support forward references of globals, we need to first
939 // sort the list of module-level globals in def-use order. We visit each
940 // global variable in order, and ensure that we emit it *after* its dependent
941 // globals. We use a little extra memory maintaining both a set and a list to
942 // have fast searches while maintaining a strict ordering.
943 SmallVector<const GlobalVariable *, 8> Globals;
944 DenseSet<const GlobalVariable *> GVVisited;
945 DenseSet<const GlobalVariable *> GVVisiting;
947 // Visit each global variable, in order
948 for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
950 VisitGlobalVariableForEmission(I, Globals, GVVisited, GVVisiting);
952 assert(GVVisited.size() == M.getGlobalList().size() &&
953 "Missed a global variable");
954 assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
956 // Print out module-level global variables in proper order
957 for (unsigned i = 0, e = Globals.size(); i != e; ++i)
958 printModuleLevelGV(Globals[i], OS2);
962 OutStreamer.EmitRawText(OS2.str());
965 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O) {
967 O << "// Generated by LLVM NVPTX Back-End\n";
971 unsigned PTXVersion = nvptxSubtarget.getPTXVersion();
972 O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
975 O << nvptxSubtarget.getTargetName();
977 if (nvptxSubtarget.getDrvInterface() == NVPTX::NVCL)
978 O << ", texmode_independent";
979 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
980 if (!nvptxSubtarget.hasDouble())
981 O << ", map_f64_to_f32";
984 if (MAI->doesSupportDebugInformation())
989 O << ".address_size ";
990 if (nvptxSubtarget.is64Bit())
999 bool NVPTXAsmPrinter::doFinalization(Module &M) {
1001 // If we did not emit any functions, then the global declarations have not
1002 // yet been emitted.
1003 if (!GlobalsEmitted) {
1005 GlobalsEmitted = true;
1008 // XXX Temproarily remove global variables so that doFinalization() will not
1009 // emit them again (global variables are emitted at beginning).
1011 Module::GlobalListType &global_list = M.getGlobalList();
1012 int i, n = global_list.size();
1013 GlobalVariable **gv_array = new GlobalVariable *[n];
1015 // first, back-up GlobalVariable in gv_array
1017 for (Module::global_iterator I = global_list.begin(), E = global_list.end();
1019 gv_array[i++] = &*I;
1021 // second, empty global_list
1022 while (!global_list.empty())
1023 global_list.remove(global_list.begin());
1025 // call doFinalization
1026 bool ret = AsmPrinter::doFinalization(M);
1028 // now we restore global variables
1029 for (i = 0; i < n; i++)
1030 global_list.insert(global_list.end(), gv_array[i]);
1035 //bool Result = AsmPrinter::doFinalization(M);
1036 // Instead of calling the parents doFinalization, we may
1037 // clone parents doFinalization and customize here.
1038 // Currently, we if NVISA out the EmitGlobals() in
1039 // parent's doFinalization, which is too intrusive.
1041 // Same for the doInitialization.
1045 // This function emits appropriate linkage directives for
1046 // functions and global variables.
1048 // extern function declaration -> .extern
1049 // extern function definition -> .visible
1050 // external global variable with init -> .visible
1051 // external without init -> .extern
1052 // appending -> not allowed, assert.
1054 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
1056 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
1057 if (V->hasExternalLinkage()) {
1058 if (isa<GlobalVariable>(V)) {
1059 const GlobalVariable *GVar = cast<GlobalVariable>(V);
1061 if (GVar->hasInitializer())
1066 } else if (V->isDeclaration())
1070 } else if (V->hasAppendingLinkage()) {
1072 msg.append("Error: ");
1073 msg.append("Symbol ");
1075 msg.append(V->getName().str());
1076 msg.append("has unsupported appending linkage type");
1077 llvm_unreachable(msg.c_str());
1082 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
1084 bool processDemoted) {
1087 if (GVar->hasSection()) {
1088 if (GVar->getSection() == "llvm.metadata")
1092 const DataLayout *TD = TM.getDataLayout();
1094 // GlobalVariables are always constant pointers themselves.
1095 const PointerType *PTy = GVar->getType();
1096 Type *ETy = PTy->getElementType();
1098 if (GVar->hasExternalLinkage()) {
1099 if (GVar->hasInitializer())
1105 if (llvm::isTexture(*GVar)) {
1106 O << ".global .texref " << llvm::getTextureName(*GVar) << ";\n";
1110 if (llvm::isSurface(*GVar)) {
1111 O << ".global .surfref " << llvm::getSurfaceName(*GVar) << ";\n";
1115 if (GVar->isDeclaration()) {
1116 // (extern) declarations, no definition or initializer
1117 // Currently the only known declaration is for an automatic __local
1118 // (.shared) promoted to global.
1119 emitPTXGlobalVariable(GVar, O);
1124 if (llvm::isSampler(*GVar)) {
1125 O << ".global .samplerref " << llvm::getSamplerName(*GVar);
1127 const Constant *Initializer = NULL;
1128 if (GVar->hasInitializer())
1129 Initializer = GVar->getInitializer();
1130 const ConstantInt *CI = NULL;
1132 CI = dyn_cast<ConstantInt>(Initializer);
1134 unsigned sample = CI->getZExtValue();
1139 addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1141 O << "addr_mode_" << i << " = ";
1147 O << "clamp_to_border";
1150 O << "clamp_to_edge";
1161 O << "filter_mode = ";
1162 switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1170 assert(0 && "Anisotropic filtering is not supported");
1175 if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1176 O << ", force_unnormalized_coords = 1";
1185 if (GVar->hasPrivateLinkage()) {
1187 if (!strncmp(GVar->getName().data(), "unrollpragma", 12))
1190 // FIXME - need better way (e.g. Metadata) to avoid generating this global
1191 if (!strncmp(GVar->getName().data(), "filename", 8))
1193 if (GVar->use_empty())
1197 const Function *demotedFunc = 0;
1198 if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
1199 O << "// " << GVar->getName().str() << " has been demoted\n";
1200 if (localDecls.find(demotedFunc) != localDecls.end())
1201 localDecls[demotedFunc].push_back(GVar);
1203 std::vector<const GlobalVariable *> temp;
1204 temp.push_back(GVar);
1205 localDecls[demotedFunc] = temp;
1211 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1212 if (GVar->getAlignment() == 0)
1213 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1215 O << " .align " << GVar->getAlignment();
1217 if (ETy->isPrimitiveType() || ETy->isIntegerTy() || isa<PointerType>(ETy)) {
1219 // Special case: ABI requires that we use .u8 for predicates
1220 if (ETy->isIntegerTy(1))
1223 O << getPTXFundamentalTypeStr(ETy, false);
1225 O << *Mang->getSymbol(GVar);
1227 // Ptx allows variable initilization only for constant and global state
1229 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1230 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST_NOT_GEN) ||
1231 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1232 GVar->hasInitializer()) {
1233 const Constant *Initializer = GVar->getInitializer();
1234 if (!Initializer->isNullValue()) {
1236 printScalarConstant(Initializer, O);
1240 unsigned int ElementSize = 0;
1242 // Although PTX has direct support for struct type and array type and
1243 // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1244 // targets that support these high level field accesses. Structs, arrays
1245 // and vectors are lowered into arrays of bytes.
1246 switch (ETy->getTypeID()) {
1247 case Type::StructTyID:
1248 case Type::ArrayTyID:
1249 case Type::VectorTyID:
1250 ElementSize = TD->getTypeStoreSize(ETy);
1251 // Ptx allows variable initilization only for constant and
1252 // global state spaces.
1253 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1254 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST_NOT_GEN) ||
1255 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1256 GVar->hasInitializer()) {
1257 const Constant *Initializer = GVar->getInitializer();
1258 if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1259 AggBuffer aggBuffer(ElementSize, O, *this);
1260 bufferAggregateConstant(Initializer, &aggBuffer);
1261 if (aggBuffer.numSymbols) {
1262 if (nvptxSubtarget.is64Bit()) {
1263 O << " .u64 " << *Mang->getSymbol(GVar) << "[";
1264 O << ElementSize / 8;
1266 O << " .u32 " << *Mang->getSymbol(GVar) << "[";
1267 O << ElementSize / 4;
1271 O << " .b8 " << *Mang->getSymbol(GVar) << "[";
1279 O << " .b8 " << *Mang->getSymbol(GVar);
1287 O << " .b8 " << *Mang->getSymbol(GVar);
1296 assert(0 && "type not supported yet");
1303 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1304 if (localDecls.find(f) == localDecls.end())
1307 std::vector<const GlobalVariable *> &gvars = localDecls[f];
1309 for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1310 O << "\t// demoted variable\n\t";
1311 printModuleLevelGV(gvars[i], O, true);
1315 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1316 raw_ostream &O) const {
1317 switch (AddressSpace) {
1318 case llvm::ADDRESS_SPACE_LOCAL:
1321 case llvm::ADDRESS_SPACE_GLOBAL:
1324 case llvm::ADDRESS_SPACE_CONST:
1325 // This logic should be consistent with that in
1326 // getCodeAddrSpace() (NVPTXISelDATToDAT.cpp)
1327 if (nvptxSubtarget.hasGenericLdSt())
1332 case llvm::ADDRESS_SPACE_CONST_NOT_GEN:
1335 case llvm::ADDRESS_SPACE_SHARED:
1339 report_fatal_error("Bad address space found while emitting PTX");
1345 NVPTXAsmPrinter::getPTXFundamentalTypeStr(const Type *Ty, bool useB4PTR) const {
1346 switch (Ty->getTypeID()) {
1348 llvm_unreachable("unexpected type");
1350 case Type::IntegerTyID: {
1351 unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1354 else if (NumBits <= 64) {
1355 std::string name = "u";
1356 return name + utostr(NumBits);
1358 llvm_unreachable("Integer too large");
1363 case Type::FloatTyID:
1365 case Type::DoubleTyID:
1367 case Type::PointerTyID:
1368 if (nvptxSubtarget.is64Bit())
1378 llvm_unreachable("unexpected type");
1382 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1385 const DataLayout *TD = TM.getDataLayout();
1387 // GlobalVariables are always constant pointers themselves.
1388 const PointerType *PTy = GVar->getType();
1389 Type *ETy = PTy->getElementType();
1392 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1393 if (GVar->getAlignment() == 0)
1394 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1396 O << " .align " << GVar->getAlignment();
1398 if (ETy->isPrimitiveType() || ETy->isIntegerTy() || isa<PointerType>(ETy)) {
1400 O << getPTXFundamentalTypeStr(ETy);
1402 O << *Mang->getSymbol(GVar);
1406 int64_t ElementSize = 0;
1408 // Although PTX has direct support for struct type and array type and LLVM IR
1409 // is very similar to PTX, the LLVM CodeGen does not support for targets that
1410 // support these high level field accesses. Structs and arrays are lowered
1411 // into arrays of bytes.
1412 switch (ETy->getTypeID()) {
1413 case Type::StructTyID:
1414 case Type::ArrayTyID:
1415 case Type::VectorTyID:
1416 ElementSize = TD->getTypeStoreSize(ETy);
1417 O << " .b8 " << *Mang->getSymbol(GVar) << "[";
1419 O << itostr(ElementSize);
1424 assert(0 && "type not supported yet");
1429 static unsigned int getOpenCLAlignment(const DataLayout *TD, Type *Ty) {
1430 if (Ty->isPrimitiveType() || Ty->isIntegerTy() || isa<PointerType>(Ty))
1431 return TD->getPrefTypeAlignment(Ty);
1433 const ArrayType *ATy = dyn_cast<ArrayType>(Ty);
1435 return getOpenCLAlignment(TD, ATy->getElementType());
1437 const VectorType *VTy = dyn_cast<VectorType>(Ty);
1439 Type *ETy = VTy->getElementType();
1440 unsigned int numE = VTy->getNumElements();
1441 unsigned int alignE = TD->getPrefTypeAlignment(ETy);
1445 return numE * alignE;
1448 const StructType *STy = dyn_cast<StructType>(Ty);
1450 unsigned int alignStruct = 1;
1451 // Go through each element of the struct and find the
1452 // largest alignment.
1453 for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1454 Type *ETy = STy->getElementType(i);
1455 unsigned int align = getOpenCLAlignment(TD, ETy);
1456 if (align > alignStruct)
1457 alignStruct = align;
1462 const FunctionType *FTy = dyn_cast<FunctionType>(Ty);
1464 return TD->getPointerPrefAlignment();
1465 return TD->getPrefTypeAlignment(Ty);
1468 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1469 int paramIndex, raw_ostream &O) {
1470 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1471 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA))
1472 O << *Mang->getSymbol(I->getParent()) << "_param_" << paramIndex;
1474 std::string argName = I->getName();
1475 const char *p = argName.c_str();
1486 void NVPTXAsmPrinter::printParamName(int paramIndex, raw_ostream &O) {
1487 Function::const_arg_iterator I, E;
1490 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1491 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)) {
1492 O << *CurrentFnSym << "_param_" << paramIndex;
1496 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, i++) {
1497 if (i == paramIndex) {
1498 printParamName(I, paramIndex, O);
1502 llvm_unreachable("paramIndex out of bound");
1505 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1506 const DataLayout *TD = TM.getDataLayout();
1507 const AttributeSet &PAL = F->getAttributes();
1508 const TargetLowering *TLI = TM.getTargetLowering();
1509 Function::const_arg_iterator I, E;
1510 unsigned paramIndex = 0;
1512 bool isKernelFunc = llvm::isKernelFunction(*F);
1513 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
1514 MVT thePointerTy = TLI->getPointerTy();
1518 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1519 Type *Ty = I->getType();
1526 // Handle image/sampler parameters
1527 if (llvm::isSampler(*I) || llvm::isImage(*I)) {
1528 if (llvm::isImage(*I)) {
1529 std::string sname = I->getName();
1530 if (llvm::isImageWriteOnly(*I))
1531 O << "\t.param .surfref " << *Mang->getSymbol(F) << "_param_"
1533 else // Default image is read_only
1534 O << "\t.param .texref " << *Mang->getSymbol(F) << "_param_"
1536 } else // Should be llvm::isSampler(*I)
1537 O << "\t.param .samplerref " << *Mang->getSymbol(F) << "_param_"
1542 if (PAL.hasAttribute(paramIndex + 1, Attribute::ByVal) == false) {
1543 if (Ty->isVectorTy()) {
1544 // Just print .param .b8 .align <a> .param[size];
1545 // <a> = PAL.getparamalignment
1546 // size = typeallocsize of element type
1547 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1549 align = TD->getABITypeAlignment(Ty);
1551 unsigned sz = TD->getTypeAllocSize(Ty);
1552 O << "\t.param .align " << align << " .b8 ";
1553 printParamName(I, paramIndex, O);
1554 O << "[" << sz << "]";
1559 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1562 // Special handling for pointer arguments to kernel
1563 O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1565 if (nvptxSubtarget.getDrvInterface() != NVPTX::CUDA) {
1566 Type *ETy = PTy->getElementType();
1567 int addrSpace = PTy->getAddressSpace();
1568 switch (addrSpace) {
1572 case llvm::ADDRESS_SPACE_CONST_NOT_GEN:
1573 O << ".ptr .const ";
1575 case llvm::ADDRESS_SPACE_SHARED:
1576 O << ".ptr .shared ";
1578 case llvm::ADDRESS_SPACE_GLOBAL:
1579 case llvm::ADDRESS_SPACE_CONST:
1580 O << ".ptr .global ";
1583 O << ".align " << (int) getOpenCLAlignment(TD, ETy) << " ";
1585 printParamName(I, paramIndex, O);
1589 // non-pointer scalar to kernel func
1591 // Special case: predicate operands become .u8 types
1592 if (Ty->isIntegerTy(1))
1595 O << getPTXFundamentalTypeStr(Ty);
1597 printParamName(I, paramIndex, O);
1600 // Non-kernel function, just print .param .b<size> for ABI
1601 // and .reg .b<size> for non ABY
1603 if (isa<IntegerType>(Ty)) {
1604 sz = cast<IntegerType>(Ty)->getBitWidth();
1607 } else if (isa<PointerType>(Ty))
1608 sz = thePointerTy.getSizeInBits();
1610 sz = Ty->getPrimitiveSizeInBits();
1612 O << "\t.param .b" << sz << " ";
1614 O << "\t.reg .b" << sz << " ";
1615 printParamName(I, paramIndex, O);
1619 // param has byVal attribute. So should be a pointer
1620 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1621 assert(PTy && "Param with byval attribute should be a pointer type");
1622 Type *ETy = PTy->getElementType();
1624 if (isABI || isKernelFunc) {
1625 // Just print .param .b8 .align <a> .param[size];
1626 // <a> = PAL.getparamalignment
1627 // size = typeallocsize of element type
1628 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1630 align = TD->getABITypeAlignment(ETy);
1632 unsigned sz = TD->getTypeAllocSize(ETy);
1633 O << "\t.param .align " << align << " .b8 ";
1634 printParamName(I, paramIndex, O);
1635 O << "[" << sz << "]";
1638 // Split the ETy into constituent parts and
1639 // print .param .b<size> <name> for each part.
1640 // Further, if a part is vector, print the above for
1641 // each vector element.
1642 SmallVector<EVT, 16> vtparts;
1643 ComputeValueVTs(*TLI, ETy, vtparts);
1644 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1646 EVT elemtype = vtparts[i];
1647 if (vtparts[i].isVector()) {
1648 elems = vtparts[i].getVectorNumElements();
1649 elemtype = vtparts[i].getVectorElementType();
1652 for (unsigned j = 0, je = elems; j != je; ++j) {
1653 unsigned sz = elemtype.getSizeInBits();
1654 if (elemtype.isInteger() && (sz < 32))
1656 O << "\t.reg .b" << sz << " ";
1657 printParamName(I, paramIndex, O);
1673 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1675 const Function *F = MF.getFunction();
1676 emitFunctionParamList(F, O);
1679 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1680 const MachineFunction &MF) {
1681 SmallString<128> Str;
1682 raw_svector_ostream O(Str);
1684 // Map the global virtual register number to a register class specific
1685 // virtual register number starting from 1 with that class.
1686 const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo();
1687 //unsigned numRegClasses = TRI->getNumRegClasses();
1689 // Emit the Fake Stack Object
1690 const MachineFrameInfo *MFI = MF.getFrameInfo();
1691 int NumBytes = (int) MFI->getStackSize();
1693 O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
1694 << getFunctionNumber() << "[" << NumBytes << "];\n";
1695 if (nvptxSubtarget.is64Bit()) {
1696 O << "\t.reg .b64 \t%SP;\n";
1697 O << "\t.reg .b64 \t%SPL;\n";
1699 O << "\t.reg .b32 \t%SP;\n";
1700 O << "\t.reg .b32 \t%SPL;\n";
1704 // Go through all virtual registers to establish the mapping between the
1706 // register number and the per class virtual register number.
1707 // We use the per class virtual register number in the ptx output.
1708 unsigned int numVRs = MRI->getNumVirtRegs();
1709 for (unsigned i = 0; i < numVRs; i++) {
1710 unsigned int vr = TRI->index2VirtReg(i);
1711 const TargetRegisterClass *RC = MRI->getRegClass(vr);
1712 std::map<unsigned, unsigned> ®map = VRidGlobal2LocalMap[RC->getID()];
1713 int n = regmap.size();
1714 regmap.insert(std::make_pair(vr, n + 1));
1717 // Emit register declarations
1718 // @TODO: Extract out the real register usage
1719 O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
1720 O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
1721 O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
1722 O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
1723 O << "\t.reg .s64 %rl<" << NVPTXNumRegisters << ">;\n";
1724 O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
1725 O << "\t.reg .f64 %fl<" << NVPTXNumRegisters << ">;\n";
1727 // Emit declaration of the virtual registers or 'physical' registers for
1728 // each register class
1729 //for (unsigned i=0; i< numRegClasses; i++) {
1730 // std::map<unsigned, unsigned> ®map = VRidGlobal2LocalMap[i];
1731 // const TargetRegisterClass *RC = TRI->getRegClass(i);
1732 // std::string rcname = getNVPTXRegClassName(RC);
1733 // std::string rcStr = getNVPTXRegClassStr(RC);
1734 // //int n = regmap.size();
1735 // if (!isNVPTXVectorRegClass(RC)) {
1736 // O << "\t.reg " << rcname << " \t" << rcStr << "<"
1737 // << NVPTXNumRegisters << ">;\n";
1740 // Only declare those registers that may be used. And do not emit vector
1742 // they are all elementized to scalar registers.
1743 //if (n && !isNVPTXVectorRegClass(RC)) {
1744 // if (RegAllocNilUsed) {
1745 // O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
1749 // O << "\t.reg " << rcname << " \t" << StrToUpper(rcStr)
1750 // << "<" << 32 << ">;\n";
1755 OutStreamer.EmitRawText(O.str());
1758 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
1759 APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
1761 unsigned int numHex;
1764 if (Fp->getType()->getTypeID() == Type::FloatTyID) {
1767 APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &ignored);
1768 } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
1771 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &ignored);
1773 llvm_unreachable("unsupported fp type");
1775 APInt API = APF.bitcastToAPInt();
1776 std::string hexstr(utohexstr(API.getZExtValue()));
1778 if (hexstr.length() < numHex)
1779 O << std::string(numHex - hexstr.length(), '0');
1780 O << utohexstr(API.getZExtValue());
1783 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
1784 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
1785 O << CI->getValue();
1788 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
1789 printFPConstant(CFP, O);
1792 if (isa<ConstantPointerNull>(CPV)) {
1796 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1797 O << *Mang->getSymbol(GVar);
1800 if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1801 const Value *v = Cexpr->stripPointerCasts();
1802 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
1803 O << *Mang->getSymbol(GVar);
1806 O << *LowerConstant(CPV, *this);
1810 llvm_unreachable("Not scalar type found in printScalarConstant()");
1813 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
1814 AggBuffer *aggBuffer) {
1816 const DataLayout *TD = TM.getDataLayout();
1818 if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
1819 int s = TD->getTypeAllocSize(CPV->getType());
1822 aggBuffer->addZeros(s);
1827 switch (CPV->getType()->getTypeID()) {
1829 case Type::IntegerTyID: {
1830 const Type *ETy = CPV->getType();
1831 if (ETy == Type::getInt8Ty(CPV->getContext())) {
1833 (unsigned char)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
1835 aggBuffer->addBytes(ptr, 1, Bytes);
1836 } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
1837 short int16 = (short)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
1838 ptr = (unsigned char *)&int16;
1839 aggBuffer->addBytes(ptr, 2, Bytes);
1840 } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
1841 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1842 int int32 = (int)(constInt->getZExtValue());
1843 ptr = (unsigned char *)&int32;
1844 aggBuffer->addBytes(ptr, 4, Bytes);
1846 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1847 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1848 ConstantFoldConstantExpression(Cexpr, TD))) {
1849 int int32 = (int)(constInt->getZExtValue());
1850 ptr = (unsigned char *)&int32;
1851 aggBuffer->addBytes(ptr, 4, Bytes);
1854 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1855 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1856 aggBuffer->addSymbol(v);
1857 aggBuffer->addZeros(4);
1861 llvm_unreachable("unsupported integer const type");
1862 } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
1863 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1864 long long int64 = (long long)(constInt->getZExtValue());
1865 ptr = (unsigned char *)&int64;
1866 aggBuffer->addBytes(ptr, 8, Bytes);
1868 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1869 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1870 ConstantFoldConstantExpression(Cexpr, TD))) {
1871 long long int64 = (long long)(constInt->getZExtValue());
1872 ptr = (unsigned char *)&int64;
1873 aggBuffer->addBytes(ptr, 8, Bytes);
1876 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1877 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1878 aggBuffer->addSymbol(v);
1879 aggBuffer->addZeros(8);
1883 llvm_unreachable("unsupported integer const type");
1885 llvm_unreachable("unsupported integer const type");
1888 case Type::FloatTyID:
1889 case Type::DoubleTyID: {
1890 const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
1891 const Type *Ty = CFP->getType();
1892 if (Ty == Type::getFloatTy(CPV->getContext())) {
1893 float float32 = (float) CFP->getValueAPF().convertToFloat();
1894 ptr = (unsigned char *)&float32;
1895 aggBuffer->addBytes(ptr, 4, Bytes);
1896 } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
1897 double float64 = CFP->getValueAPF().convertToDouble();
1898 ptr = (unsigned char *)&float64;
1899 aggBuffer->addBytes(ptr, 8, Bytes);
1901 llvm_unreachable("unsupported fp const type");
1905 case Type::PointerTyID: {
1906 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1907 aggBuffer->addSymbol(GVar);
1908 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1909 const Value *v = Cexpr->stripPointerCasts();
1910 aggBuffer->addSymbol(v);
1912 unsigned int s = TD->getTypeAllocSize(CPV->getType());
1913 aggBuffer->addZeros(s);
1917 case Type::ArrayTyID:
1918 case Type::VectorTyID:
1919 case Type::StructTyID: {
1920 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV) ||
1921 isa<ConstantStruct>(CPV)) {
1922 int ElementSize = TD->getTypeAllocSize(CPV->getType());
1923 bufferAggregateConstant(CPV, aggBuffer);
1924 if (Bytes > ElementSize)
1925 aggBuffer->addZeros(Bytes - ElementSize);
1926 } else if (isa<ConstantAggregateZero>(CPV))
1927 aggBuffer->addZeros(Bytes);
1929 llvm_unreachable("Unexpected Constant type");
1934 llvm_unreachable("unsupported type");
1938 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
1939 AggBuffer *aggBuffer) {
1940 const DataLayout *TD = TM.getDataLayout();
1944 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
1945 if (CPV->getNumOperands())
1946 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
1947 bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
1951 if (const ConstantDataSequential *CDS =
1952 dyn_cast<ConstantDataSequential>(CPV)) {
1953 if (CDS->getNumElements())
1954 for (unsigned i = 0; i < CDS->getNumElements(); ++i)
1955 bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
1960 if (isa<ConstantStruct>(CPV)) {
1961 if (CPV->getNumOperands()) {
1962 StructType *ST = cast<StructType>(CPV->getType());
1963 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
1965 Bytes = TD->getStructLayout(ST)->getElementOffset(0) +
1966 TD->getTypeAllocSize(ST) -
1967 TD->getStructLayout(ST)->getElementOffset(i);
1969 Bytes = TD->getStructLayout(ST)->getElementOffset(i + 1) -
1970 TD->getStructLayout(ST)->getElementOffset(i);
1971 bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
1976 llvm_unreachable("unsupported constant type in printAggregateConstant()");
1979 // buildTypeNameMap - Run through symbol table looking for type names.
1982 bool NVPTXAsmPrinter::isImageType(const Type *Ty) {
1984 std::map<const Type *, std::string>::iterator PI = TypeNameMap.find(Ty);
1986 if (PI != TypeNameMap.end() && (!PI->second.compare("struct._image1d_t") ||
1987 !PI->second.compare("struct._image2d_t") ||
1988 !PI->second.compare("struct._image3d_t")))
1994 /// PrintAsmOperand - Print out an operand for an inline asm expression.
1996 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
1997 unsigned AsmVariant,
1998 const char *ExtraCode, raw_ostream &O) {
1999 if (ExtraCode && ExtraCode[0]) {
2000 if (ExtraCode[1] != 0)
2001 return true; // Unknown modifier.
2003 switch (ExtraCode[0]) {
2005 // See if this is a generic print operand
2006 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
2012 printOperand(MI, OpNo, O);
2017 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(
2018 const MachineInstr *MI, unsigned OpNo, unsigned AsmVariant,
2019 const char *ExtraCode, raw_ostream &O) {
2020 if (ExtraCode && ExtraCode[0])
2021 return true; // Unknown modifier
2024 printMemOperand(MI, OpNo, O);
2030 bool NVPTXAsmPrinter::ignoreLoc(const MachineInstr &MI) {
2031 switch (MI.getOpcode()) {
2034 case NVPTX::CallArgBeginInst:
2035 case NVPTX::CallArgEndInst0:
2036 case NVPTX::CallArgEndInst1:
2037 case NVPTX::CallArgF32:
2038 case NVPTX::CallArgF64:
2039 case NVPTX::CallArgI16:
2040 case NVPTX::CallArgI32:
2041 case NVPTX::CallArgI32imm:
2042 case NVPTX::CallArgI64:
2043 case NVPTX::CallArgI8:
2044 case NVPTX::CallArgParam:
2045 case NVPTX::CallVoidInst:
2046 case NVPTX::CallVoidInstReg:
2047 case NVPTX::Callseq_End:
2048 case NVPTX::CallVoidInstReg64:
2049 case NVPTX::DeclareParamInst:
2050 case NVPTX::DeclareRetMemInst:
2051 case NVPTX::DeclareRetRegInst:
2052 case NVPTX::DeclareRetScalarInst:
2053 case NVPTX::DeclareScalarParamInst:
2054 case NVPTX::DeclareScalarRegInst:
2055 case NVPTX::StoreParamF32:
2056 case NVPTX::StoreParamF64:
2057 case NVPTX::StoreParamI16:
2058 case NVPTX::StoreParamI32:
2059 case NVPTX::StoreParamI64:
2060 case NVPTX::StoreParamI8:
2061 case NVPTX::StoreParamS32I8:
2062 case NVPTX::StoreParamU32I8:
2063 case NVPTX::StoreParamS32I16:
2064 case NVPTX::StoreParamU32I16:
2065 case NVPTX::StoreRetvalF32:
2066 case NVPTX::StoreRetvalF64:
2067 case NVPTX::StoreRetvalI16:
2068 case NVPTX::StoreRetvalI32:
2069 case NVPTX::StoreRetvalI64:
2070 case NVPTX::StoreRetvalI8:
2071 case NVPTX::LastCallArgF32:
2072 case NVPTX::LastCallArgF64:
2073 case NVPTX::LastCallArgI16:
2074 case NVPTX::LastCallArgI32:
2075 case NVPTX::LastCallArgI32imm:
2076 case NVPTX::LastCallArgI64:
2077 case NVPTX::LastCallArgI8:
2078 case NVPTX::LastCallArgParam:
2079 case NVPTX::LoadParamMemF32:
2080 case NVPTX::LoadParamMemF64:
2081 case NVPTX::LoadParamMemI16:
2082 case NVPTX::LoadParamMemI32:
2083 case NVPTX::LoadParamMemI64:
2084 case NVPTX::LoadParamMemI8:
2085 case NVPTX::LoadParamRegF32:
2086 case NVPTX::LoadParamRegF64:
2087 case NVPTX::LoadParamRegI16:
2088 case NVPTX::LoadParamRegI32:
2089 case NVPTX::LoadParamRegI64:
2090 case NVPTX::LoadParamRegI8:
2091 case NVPTX::PrototypeInst:
2092 case NVPTX::DBG_VALUE:
2098 // Force static initialization.
2099 extern "C" void LLVMInitializeNVPTXBackendAsmPrinter() {
2100 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2101 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2104 void NVPTXAsmPrinter::emitSrcInText(StringRef filename, unsigned line) {
2105 std::stringstream temp;
2106 LineReader *reader = this->getReader(filename.str());
2108 temp << filename.str();
2112 temp << reader->readLine(line);
2114 this->OutStreamer.EmitRawText(Twine(temp.str()));
2117 LineReader *NVPTXAsmPrinter::getReader(std::string filename) {
2118 if (reader == NULL) {
2119 reader = new LineReader(filename);
2122 if (reader->fileName() != filename) {
2124 reader = new LineReader(filename);
2130 std::string LineReader::readLine(unsigned lineNum) {
2131 if (lineNum < theCurLine) {
2133 fstr.seekg(0, std::ios::beg);
2135 while (theCurLine < lineNum) {
2136 fstr.getline(buff, 500);
2142 // Force static initialization.
2143 extern "C" void LLVMInitializeNVPTXAsmPrinter() {
2144 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2145 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);