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() {
440 OutStreamer.EmitRawText(StringRef("{\n"));
441 setAndEmitFunctionVirtualRegisters(*MF);
443 SmallString<128> Str;
444 raw_svector_ostream O(Str);
445 emitDemotedVars(MF->getFunction(), O);
446 OutStreamer.EmitRawText(O.str());
449 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
450 OutStreamer.EmitRawText(StringRef("}\n"));
454 void NVPTXAsmPrinter::emitKernelFunctionDirectives(const Function &F,
455 raw_ostream &O) const {
456 // If the NVVM IR has some of reqntid* specified, then output
457 // the reqntid directive, and set the unspecified ones to 1.
458 // If none of reqntid* is specified, don't output reqntid directive.
459 unsigned reqntidx, reqntidy, reqntidz;
460 bool specified = false;
461 if (llvm::getReqNTIDx(F, reqntidx) == false)
465 if (llvm::getReqNTIDy(F, reqntidy) == false)
469 if (llvm::getReqNTIDz(F, reqntidz) == false)
475 O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
478 // If the NVVM IR has some of maxntid* specified, then output
479 // the maxntid directive, and set the unspecified ones to 1.
480 // If none of maxntid* is specified, don't output maxntid directive.
481 unsigned maxntidx, maxntidy, maxntidz;
483 if (llvm::getMaxNTIDx(F, maxntidx) == false)
487 if (llvm::getMaxNTIDy(F, maxntidy) == false)
491 if (llvm::getMaxNTIDz(F, maxntidz) == false)
497 O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
501 if (llvm::getMinCTASm(F, mincta))
502 O << ".minnctapersm " << mincta << "\n";
505 void NVPTXAsmPrinter::getVirtualRegisterName(unsigned vr, bool isVec,
507 const TargetRegisterClass *RC = MRI->getRegClass(vr);
509 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
510 unsigned mapped_vr = regmap[vr];
513 O << getNVPTXRegClassStr(RC) << mapped_vr;
516 report_fatal_error("Bad register!");
519 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr, bool isVec,
521 getVirtualRegisterName(vr, isVec, O);
524 void NVPTXAsmPrinter::printVecModifiedImmediate(
525 const MachineOperand &MO, const char *Modifier, raw_ostream &O) {
526 static const char vecelem[] = { '0', '1', '2', '3', '0', '1', '2', '3' };
527 int Imm = (int) MO.getImm();
528 if (0 == strcmp(Modifier, "vecelem"))
529 O << "_" << vecelem[Imm];
530 else if (0 == strcmp(Modifier, "vecv4comm1")) {
531 if ((Imm < 0) || (Imm > 3))
533 } else if (0 == strcmp(Modifier, "vecv4comm2")) {
534 if ((Imm < 4) || (Imm > 7))
536 } else if (0 == strcmp(Modifier, "vecv4pos")) {
539 O << "_" << vecelem[Imm % 4];
540 } else if (0 == strcmp(Modifier, "vecv2comm1")) {
541 if ((Imm < 0) || (Imm > 1))
543 } else if (0 == strcmp(Modifier, "vecv2comm2")) {
544 if ((Imm < 2) || (Imm > 3))
546 } else if (0 == strcmp(Modifier, "vecv2pos")) {
549 O << "_" << vecelem[Imm % 2];
551 llvm_unreachable("Unknown Modifier on immediate operand");
554 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
555 raw_ostream &O, const char *Modifier) {
556 const MachineOperand &MO = MI->getOperand(opNum);
557 switch (MO.getType()) {
558 case MachineOperand::MO_Register:
559 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
560 if (MO.getReg() == NVPTX::VRDepot)
561 O << DEPOTNAME << getFunctionNumber();
563 O << getRegisterName(MO.getReg());
566 emitVirtualRegister(MO.getReg(), false, O);
568 if (strcmp(Modifier, "vecfull") == 0)
569 emitVirtualRegister(MO.getReg(), true, O);
572 "Don't know how to handle the modifier on virtual register.");
577 case MachineOperand::MO_Immediate:
580 else if (strstr(Modifier, "vec") == Modifier)
581 printVecModifiedImmediate(MO, Modifier, O);
584 "Don't know how to handle modifier on immediate operand");
587 case MachineOperand::MO_FPImmediate:
588 printFPConstant(MO.getFPImm(), O);
591 case MachineOperand::MO_GlobalAddress:
592 O << *Mang->getSymbol(MO.getGlobal());
595 case MachineOperand::MO_ExternalSymbol: {
596 const char *symbname = MO.getSymbolName();
597 if (strstr(symbname, ".PARAM") == symbname) {
599 sscanf(symbname + 6, "%u[];", &index);
600 printParamName(index, O);
601 } else if (strstr(symbname, ".HLPPARAM") == symbname) {
603 sscanf(symbname + 9, "%u[];", &index);
604 O << *CurrentFnSym << "_param_" << index << "_offset";
610 case MachineOperand::MO_MachineBasicBlock:
611 O << *MO.getMBB()->getSymbol();
615 llvm_unreachable("Operand type not supported.");
619 void NVPTXAsmPrinter::printImplicitDef(const MachineInstr *MI,
620 raw_ostream &O) const {
622 O << "\t// Implicit def :";
623 //printOperand(MI, 0);
628 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
629 raw_ostream &O, const char *Modifier) {
630 printOperand(MI, opNum, O);
632 if (Modifier && !strcmp(Modifier, "add")) {
634 printOperand(MI, opNum + 1, O);
636 if (MI->getOperand(opNum + 1).isImm() &&
637 MI->getOperand(opNum + 1).getImm() == 0)
638 return; // don't print ',0' or '+0'
640 printOperand(MI, opNum + 1, O);
644 void NVPTXAsmPrinter::printLdStCode(const MachineInstr *MI, int opNum,
645 raw_ostream &O, const char *Modifier) {
647 const MachineOperand &MO = MI->getOperand(opNum);
648 int Imm = (int) MO.getImm();
649 if (!strcmp(Modifier, "volatile")) {
652 } else if (!strcmp(Modifier, "addsp")) {
654 case NVPTX::PTXLdStInstCode::GLOBAL:
657 case NVPTX::PTXLdStInstCode::SHARED:
660 case NVPTX::PTXLdStInstCode::LOCAL:
663 case NVPTX::PTXLdStInstCode::PARAM:
666 case NVPTX::PTXLdStInstCode::CONSTANT:
669 case NVPTX::PTXLdStInstCode::GENERIC:
670 if (!nvptxSubtarget.hasGenericLdSt())
674 llvm_unreachable("Wrong Address Space");
676 } else if (!strcmp(Modifier, "sign")) {
677 if (Imm == NVPTX::PTXLdStInstCode::Signed)
679 else if (Imm == NVPTX::PTXLdStInstCode::Unsigned)
683 } else if (!strcmp(Modifier, "vec")) {
684 if (Imm == NVPTX::PTXLdStInstCode::V2)
686 else if (Imm == NVPTX::PTXLdStInstCode::V4)
689 llvm_unreachable("Unknown Modifier");
691 llvm_unreachable("Empty Modifier");
694 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
696 emitLinkageDirective(F, O);
697 if (llvm::isKernelFunction(*F))
701 printReturnValStr(F, O);
702 O << *Mang->getSymbol(F) << "\n";
703 emitFunctionParamList(F, O);
707 static bool usedInGlobalVarDef(const Constant *C) {
711 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
712 if (GV->getName().str() == "llvm.used")
717 for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
719 const Constant *C = dyn_cast<Constant>(*ui);
720 if (usedInGlobalVarDef(C))
726 static bool usedInOneFunc(const User *U, Function const *&oneFunc) {
727 if (const GlobalVariable *othergv = dyn_cast<GlobalVariable>(U)) {
728 if (othergv->getName().str() == "llvm.used")
732 if (const Instruction *instr = dyn_cast<Instruction>(U)) {
733 if (instr->getParent() && instr->getParent()->getParent()) {
734 const Function *curFunc = instr->getParent()->getParent();
735 if (oneFunc && (curFunc != oneFunc))
743 if (const MDNode *md = dyn_cast<MDNode>(U))
744 if (md->hasName() && ((md->getName().str() == "llvm.dbg.gv") ||
745 (md->getName().str() == "llvm.dbg.sp")))
748 for (User::const_use_iterator ui = U->use_begin(), ue = U->use_end();
750 if (usedInOneFunc(*ui, oneFunc) == false)
756 /* Find out if a global variable can be demoted to local scope.
757 * Currently, this is valid for CUDA shared variables, which have local
758 * scope and global lifetime. So the conditions to check are :
759 * 1. Is the global variable in shared address space?
760 * 2. Does it have internal linkage?
761 * 3. Is the global variable referenced only in one function?
763 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
764 if (gv->hasInternalLinkage() == false)
766 const PointerType *Pty = gv->getType();
767 if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
770 const Function *oneFunc = 0;
772 bool flag = usedInOneFunc(gv, oneFunc);
781 static bool useFuncSeen(const Constant *C,
782 llvm::DenseMap<const Function *, bool> &seenMap) {
783 for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
785 if (const Constant *cu = dyn_cast<Constant>(*ui)) {
786 if (useFuncSeen(cu, seenMap))
788 } else if (const Instruction *I = dyn_cast<Instruction>(*ui)) {
789 const BasicBlock *bb = I->getParent();
792 const Function *caller = bb->getParent();
795 if (seenMap.find(caller) != seenMap.end())
802 void NVPTXAsmPrinter::emitDeclarations(const Module &M, raw_ostream &O) {
803 llvm::DenseMap<const Function *, bool> seenMap;
804 for (Module::const_iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
805 const Function *F = FI;
807 if (F->isDeclaration()) {
810 if (F->getIntrinsicID())
812 emitDeclaration(F, O);
815 for (Value::const_use_iterator iter = F->use_begin(),
816 iterEnd = F->use_end();
817 iter != iterEnd; ++iter) {
818 if (const Constant *C = dyn_cast<Constant>(*iter)) {
819 if (usedInGlobalVarDef(C)) {
820 // The use is in the initialization of a global variable
821 // that is a function pointer, so print a declaration
822 // for the original function
823 emitDeclaration(F, O);
826 // Emit a declaration of this function if the function that
827 // uses this constant expr has already been seen.
828 if (useFuncSeen(C, seenMap)) {
829 emitDeclaration(F, O);
834 if (!isa<Instruction>(*iter))
836 const Instruction *instr = cast<Instruction>(*iter);
837 const BasicBlock *bb = instr->getParent();
840 const Function *caller = bb->getParent();
844 // If a caller has already been seen, then the caller is
845 // appearing in the module before the callee. so print out
846 // a declaration for the callee.
847 if (seenMap.find(caller) != seenMap.end()) {
848 emitDeclaration(F, O);
856 void NVPTXAsmPrinter::recordAndEmitFilenames(Module &M) {
857 DebugInfoFinder DbgFinder;
858 DbgFinder.processModule(M);
861 for (DebugInfoFinder::iterator I = DbgFinder.compile_unit_begin(),
862 E = DbgFinder.compile_unit_end();
864 DICompileUnit DIUnit(*I);
865 StringRef Filename(DIUnit.getFilename());
866 StringRef Dirname(DIUnit.getDirectory());
867 SmallString<128> FullPathName = Dirname;
868 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
869 sys::path::append(FullPathName, Filename);
870 Filename = FullPathName.str();
872 if (filenameMap.find(Filename.str()) != filenameMap.end())
874 filenameMap[Filename.str()] = i;
875 OutStreamer.EmitDwarfFileDirective(i, "", Filename.str());
879 for (DebugInfoFinder::iterator I = DbgFinder.subprogram_begin(),
880 E = DbgFinder.subprogram_end();
883 StringRef Filename(SP.getFilename());
884 StringRef Dirname(SP.getDirectory());
885 SmallString<128> FullPathName = Dirname;
886 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
887 sys::path::append(FullPathName, Filename);
888 Filename = FullPathName.str();
890 if (filenameMap.find(Filename.str()) != filenameMap.end())
892 filenameMap[Filename.str()] = i;
897 bool NVPTXAsmPrinter::doInitialization(Module &M) {
899 SmallString<128> Str1;
900 raw_svector_ostream OS1(Str1);
902 MMI = getAnalysisIfAvailable<MachineModuleInfo>();
903 MMI->AnalyzeModule(M);
905 // We need to call the parent's one explicitly.
906 //bool Result = AsmPrinter::doInitialization(M);
908 // Initialize TargetLoweringObjectFile.
909 const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
910 .Initialize(OutContext, TM);
912 Mang = new Mangler(OutContext, &TM);
914 // Emit header before any dwarf directives are emitted below.
916 OutStreamer.EmitRawText(OS1.str());
918 // Already commented out
919 //bool Result = AsmPrinter::doInitialization(M);
921 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
922 recordAndEmitFilenames(M);
924 GlobalsEmitted = false;
926 return false; // success
929 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
930 SmallString<128> Str2;
931 raw_svector_ostream OS2(Str2);
933 emitDeclarations(M, OS2);
935 // As ptxas does not support forward references of globals, we need to first
936 // sort the list of module-level globals in def-use order. We visit each
937 // global variable in order, and ensure that we emit it *after* its dependent
938 // globals. We use a little extra memory maintaining both a set and a list to
939 // have fast searches while maintaining a strict ordering.
940 SmallVector<const GlobalVariable *, 8> Globals;
941 DenseSet<const GlobalVariable *> GVVisited;
942 DenseSet<const GlobalVariable *> GVVisiting;
944 // Visit each global variable, in order
945 for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
947 VisitGlobalVariableForEmission(I, Globals, GVVisited, GVVisiting);
949 assert(GVVisited.size() == M.getGlobalList().size() &&
950 "Missed a global variable");
951 assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
953 // Print out module-level global variables in proper order
954 for (unsigned i = 0, e = Globals.size(); i != e; ++i)
955 printModuleLevelGV(Globals[i], OS2);
959 OutStreamer.EmitRawText(OS2.str());
962 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O) {
964 O << "// Generated by LLVM NVPTX Back-End\n";
968 unsigned PTXVersion = nvptxSubtarget.getPTXVersion();
969 O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
972 O << nvptxSubtarget.getTargetName();
974 if (nvptxSubtarget.getDrvInterface() == NVPTX::NVCL)
975 O << ", texmode_independent";
976 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
977 if (!nvptxSubtarget.hasDouble())
978 O << ", map_f64_to_f32";
981 if (MAI->doesSupportDebugInformation())
986 O << ".address_size ";
987 if (nvptxSubtarget.is64Bit())
996 bool NVPTXAsmPrinter::doFinalization(Module &M) {
998 // If we did not emit any functions, then the global declarations have not
1000 if (!GlobalsEmitted) {
1002 GlobalsEmitted = true;
1005 // XXX Temproarily remove global variables so that doFinalization() will not
1006 // emit them again (global variables are emitted at beginning).
1008 Module::GlobalListType &global_list = M.getGlobalList();
1009 int i, n = global_list.size();
1010 GlobalVariable **gv_array = new GlobalVariable *[n];
1012 // first, back-up GlobalVariable in gv_array
1014 for (Module::global_iterator I = global_list.begin(), E = global_list.end();
1016 gv_array[i++] = &*I;
1018 // second, empty global_list
1019 while (!global_list.empty())
1020 global_list.remove(global_list.begin());
1022 // call doFinalization
1023 bool ret = AsmPrinter::doFinalization(M);
1025 // now we restore global variables
1026 for (i = 0; i < n; i++)
1027 global_list.insert(global_list.end(), gv_array[i]);
1032 //bool Result = AsmPrinter::doFinalization(M);
1033 // Instead of calling the parents doFinalization, we may
1034 // clone parents doFinalization and customize here.
1035 // Currently, we if NVISA out the EmitGlobals() in
1036 // parent's doFinalization, which is too intrusive.
1038 // Same for the doInitialization.
1042 // This function emits appropriate linkage directives for
1043 // functions and global variables.
1045 // extern function declaration -> .extern
1046 // extern function definition -> .visible
1047 // external global variable with init -> .visible
1048 // external without init -> .extern
1049 // appending -> not allowed, assert.
1051 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
1053 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
1054 if (V->hasExternalLinkage()) {
1055 if (isa<GlobalVariable>(V)) {
1056 const GlobalVariable *GVar = cast<GlobalVariable>(V);
1058 if (GVar->hasInitializer())
1063 } else if (V->isDeclaration())
1067 } else if (V->hasAppendingLinkage()) {
1069 msg.append("Error: ");
1070 msg.append("Symbol ");
1072 msg.append(V->getName().str());
1073 msg.append("has unsupported appending linkage type");
1074 llvm_unreachable(msg.c_str());
1079 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
1081 bool processDemoted) {
1084 if (GVar->hasSection()) {
1085 if (GVar->getSection() == "llvm.metadata")
1089 const DataLayout *TD = TM.getDataLayout();
1091 // GlobalVariables are always constant pointers themselves.
1092 const PointerType *PTy = GVar->getType();
1093 Type *ETy = PTy->getElementType();
1095 if (GVar->hasExternalLinkage()) {
1096 if (GVar->hasInitializer())
1102 if (llvm::isTexture(*GVar)) {
1103 O << ".global .texref " << llvm::getTextureName(*GVar) << ";\n";
1107 if (llvm::isSurface(*GVar)) {
1108 O << ".global .surfref " << llvm::getSurfaceName(*GVar) << ";\n";
1112 if (GVar->isDeclaration()) {
1113 // (extern) declarations, no definition or initializer
1114 // Currently the only known declaration is for an automatic __local
1115 // (.shared) promoted to global.
1116 emitPTXGlobalVariable(GVar, O);
1121 if (llvm::isSampler(*GVar)) {
1122 O << ".global .samplerref " << llvm::getSamplerName(*GVar);
1124 const Constant *Initializer = NULL;
1125 if (GVar->hasInitializer())
1126 Initializer = GVar->getInitializer();
1127 const ConstantInt *CI = NULL;
1129 CI = dyn_cast<ConstantInt>(Initializer);
1131 unsigned sample = CI->getZExtValue();
1136 addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1138 O << "addr_mode_" << i << " = ";
1144 O << "clamp_to_border";
1147 O << "clamp_to_edge";
1158 O << "filter_mode = ";
1159 switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1167 assert(0 && "Anisotropic filtering is not supported");
1172 if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1173 O << ", force_unnormalized_coords = 1";
1182 if (GVar->hasPrivateLinkage()) {
1184 if (!strncmp(GVar->getName().data(), "unrollpragma", 12))
1187 // FIXME - need better way (e.g. Metadata) to avoid generating this global
1188 if (!strncmp(GVar->getName().data(), "filename", 8))
1190 if (GVar->use_empty())
1194 const Function *demotedFunc = 0;
1195 if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
1196 O << "// " << GVar->getName().str() << " has been demoted\n";
1197 if (localDecls.find(demotedFunc) != localDecls.end())
1198 localDecls[demotedFunc].push_back(GVar);
1200 std::vector<const GlobalVariable *> temp;
1201 temp.push_back(GVar);
1202 localDecls[demotedFunc] = temp;
1208 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1209 if (GVar->getAlignment() == 0)
1210 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1212 O << " .align " << GVar->getAlignment();
1214 if (ETy->isPrimitiveType() || ETy->isIntegerTy() || isa<PointerType>(ETy)) {
1216 // Special case: ABI requires that we use .u8 for predicates
1217 if (ETy->isIntegerTy(1))
1220 O << getPTXFundamentalTypeStr(ETy, false);
1222 O << *Mang->getSymbol(GVar);
1224 // Ptx allows variable initilization only for constant and global state
1226 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1227 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1228 GVar->hasInitializer()) {
1229 const Constant *Initializer = GVar->getInitializer();
1230 if (!Initializer->isNullValue()) {
1232 printScalarConstant(Initializer, O);
1236 unsigned int ElementSize = 0;
1238 // Although PTX has direct support for struct type and array type and
1239 // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1240 // targets that support these high level field accesses. Structs, arrays
1241 // and vectors are lowered into arrays of bytes.
1242 switch (ETy->getTypeID()) {
1243 case Type::StructTyID:
1244 case Type::ArrayTyID:
1245 case Type::VectorTyID:
1246 ElementSize = TD->getTypeStoreSize(ETy);
1247 // Ptx allows variable initilization only for constant and
1248 // global state spaces.
1249 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1250 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1251 GVar->hasInitializer()) {
1252 const Constant *Initializer = GVar->getInitializer();
1253 if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1254 AggBuffer aggBuffer(ElementSize, O, *this);
1255 bufferAggregateConstant(Initializer, &aggBuffer);
1256 if (aggBuffer.numSymbols) {
1257 if (nvptxSubtarget.is64Bit()) {
1258 O << " .u64 " << *Mang->getSymbol(GVar) << "[";
1259 O << ElementSize / 8;
1261 O << " .u32 " << *Mang->getSymbol(GVar) << "[";
1262 O << ElementSize / 4;
1266 O << " .b8 " << *Mang->getSymbol(GVar) << "[";
1274 O << " .b8 " << *Mang->getSymbol(GVar);
1282 O << " .b8 " << *Mang->getSymbol(GVar);
1291 assert(0 && "type not supported yet");
1298 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1299 if (localDecls.find(f) == localDecls.end())
1302 std::vector<const GlobalVariable *> &gvars = localDecls[f];
1304 for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1305 O << "\t// demoted variable\n\t";
1306 printModuleLevelGV(gvars[i], O, true);
1310 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1311 raw_ostream &O) const {
1312 switch (AddressSpace) {
1313 case llvm::ADDRESS_SPACE_LOCAL:
1316 case llvm::ADDRESS_SPACE_GLOBAL:
1319 case llvm::ADDRESS_SPACE_CONST:
1322 case llvm::ADDRESS_SPACE_SHARED:
1326 report_fatal_error("Bad address space found while emitting PTX");
1332 NVPTXAsmPrinter::getPTXFundamentalTypeStr(const Type *Ty, bool useB4PTR) const {
1333 switch (Ty->getTypeID()) {
1335 llvm_unreachable("unexpected type");
1337 case Type::IntegerTyID: {
1338 unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1341 else if (NumBits <= 64) {
1342 std::string name = "u";
1343 return name + utostr(NumBits);
1345 llvm_unreachable("Integer too large");
1350 case Type::FloatTyID:
1352 case Type::DoubleTyID:
1354 case Type::PointerTyID:
1355 if (nvptxSubtarget.is64Bit())
1365 llvm_unreachable("unexpected type");
1369 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1372 const DataLayout *TD = TM.getDataLayout();
1374 // GlobalVariables are always constant pointers themselves.
1375 const PointerType *PTy = GVar->getType();
1376 Type *ETy = PTy->getElementType();
1379 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1380 if (GVar->getAlignment() == 0)
1381 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1383 O << " .align " << GVar->getAlignment();
1385 if (ETy->isPrimitiveType() || ETy->isIntegerTy() || isa<PointerType>(ETy)) {
1387 O << getPTXFundamentalTypeStr(ETy);
1389 O << *Mang->getSymbol(GVar);
1393 int64_t ElementSize = 0;
1395 // Although PTX has direct support for struct type and array type and LLVM IR
1396 // is very similar to PTX, the LLVM CodeGen does not support for targets that
1397 // support these high level field accesses. Structs and arrays are lowered
1398 // into arrays of bytes.
1399 switch (ETy->getTypeID()) {
1400 case Type::StructTyID:
1401 case Type::ArrayTyID:
1402 case Type::VectorTyID:
1403 ElementSize = TD->getTypeStoreSize(ETy);
1404 O << " .b8 " << *Mang->getSymbol(GVar) << "[";
1406 O << itostr(ElementSize);
1411 assert(0 && "type not supported yet");
1416 static unsigned int getOpenCLAlignment(const DataLayout *TD, Type *Ty) {
1417 if (Ty->isPrimitiveType() || Ty->isIntegerTy() || isa<PointerType>(Ty))
1418 return TD->getPrefTypeAlignment(Ty);
1420 const ArrayType *ATy = dyn_cast<ArrayType>(Ty);
1422 return getOpenCLAlignment(TD, ATy->getElementType());
1424 const VectorType *VTy = dyn_cast<VectorType>(Ty);
1426 Type *ETy = VTy->getElementType();
1427 unsigned int numE = VTy->getNumElements();
1428 unsigned int alignE = TD->getPrefTypeAlignment(ETy);
1432 return numE * alignE;
1435 const StructType *STy = dyn_cast<StructType>(Ty);
1437 unsigned int alignStruct = 1;
1438 // Go through each element of the struct and find the
1439 // largest alignment.
1440 for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1441 Type *ETy = STy->getElementType(i);
1442 unsigned int align = getOpenCLAlignment(TD, ETy);
1443 if (align > alignStruct)
1444 alignStruct = align;
1449 const FunctionType *FTy = dyn_cast<FunctionType>(Ty);
1451 return TD->getPointerPrefAlignment();
1452 return TD->getPrefTypeAlignment(Ty);
1455 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1456 int paramIndex, raw_ostream &O) {
1457 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1458 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA))
1459 O << *Mang->getSymbol(I->getParent()) << "_param_" << paramIndex;
1461 std::string argName = I->getName();
1462 const char *p = argName.c_str();
1473 void NVPTXAsmPrinter::printParamName(int paramIndex, raw_ostream &O) {
1474 Function::const_arg_iterator I, E;
1477 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1478 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)) {
1479 O << *CurrentFnSym << "_param_" << paramIndex;
1483 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, i++) {
1484 if (i == paramIndex) {
1485 printParamName(I, paramIndex, O);
1489 llvm_unreachable("paramIndex out of bound");
1492 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1493 const DataLayout *TD = TM.getDataLayout();
1494 const AttributeSet &PAL = F->getAttributes();
1495 const TargetLowering *TLI = TM.getTargetLowering();
1496 Function::const_arg_iterator I, E;
1497 unsigned paramIndex = 0;
1499 bool isKernelFunc = llvm::isKernelFunction(*F);
1500 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
1501 MVT thePointerTy = TLI->getPointerTy();
1505 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1506 Type *Ty = I->getType();
1513 // Handle image/sampler parameters
1514 if (llvm::isSampler(*I) || llvm::isImage(*I)) {
1515 if (llvm::isImage(*I)) {
1516 std::string sname = I->getName();
1517 if (llvm::isImageWriteOnly(*I))
1518 O << "\t.param .surfref " << *Mang->getSymbol(F) << "_param_"
1520 else // Default image is read_only
1521 O << "\t.param .texref " << *Mang->getSymbol(F) << "_param_"
1523 } else // Should be llvm::isSampler(*I)
1524 O << "\t.param .samplerref " << *Mang->getSymbol(F) << "_param_"
1529 if (PAL.hasAttribute(paramIndex + 1, Attribute::ByVal) == false) {
1530 if (Ty->isVectorTy()) {
1531 // Just print .param .b8 .align <a> .param[size];
1532 // <a> = PAL.getparamalignment
1533 // size = typeallocsize of element type
1534 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1536 align = TD->getABITypeAlignment(Ty);
1538 unsigned sz = TD->getTypeAllocSize(Ty);
1539 O << "\t.param .align " << align << " .b8 ";
1540 printParamName(I, paramIndex, O);
1541 O << "[" << sz << "]";
1546 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1549 // Special handling for pointer arguments to kernel
1550 O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1552 if (nvptxSubtarget.getDrvInterface() != NVPTX::CUDA) {
1553 Type *ETy = PTy->getElementType();
1554 int addrSpace = PTy->getAddressSpace();
1555 switch (addrSpace) {
1559 case llvm::ADDRESS_SPACE_CONST:
1560 O << ".ptr .const ";
1562 case llvm::ADDRESS_SPACE_SHARED:
1563 O << ".ptr .shared ";
1565 case llvm::ADDRESS_SPACE_GLOBAL:
1566 O << ".ptr .global ";
1569 O << ".align " << (int) getOpenCLAlignment(TD, ETy) << " ";
1571 printParamName(I, paramIndex, O);
1575 // non-pointer scalar to kernel func
1577 // Special case: predicate operands become .u8 types
1578 if (Ty->isIntegerTy(1))
1581 O << getPTXFundamentalTypeStr(Ty);
1583 printParamName(I, paramIndex, O);
1586 // Non-kernel function, just print .param .b<size> for ABI
1587 // and .reg .b<size> for non ABY
1589 if (isa<IntegerType>(Ty)) {
1590 sz = cast<IntegerType>(Ty)->getBitWidth();
1593 } else if (isa<PointerType>(Ty))
1594 sz = thePointerTy.getSizeInBits();
1596 sz = Ty->getPrimitiveSizeInBits();
1598 O << "\t.param .b" << sz << " ";
1600 O << "\t.reg .b" << sz << " ";
1601 printParamName(I, paramIndex, O);
1605 // param has byVal attribute. So should be a pointer
1606 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1607 assert(PTy && "Param with byval attribute should be a pointer type");
1608 Type *ETy = PTy->getElementType();
1610 if (isABI || isKernelFunc) {
1611 // Just print .param .b8 .align <a> .param[size];
1612 // <a> = PAL.getparamalignment
1613 // size = typeallocsize of element type
1614 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1616 align = TD->getABITypeAlignment(ETy);
1618 unsigned sz = TD->getTypeAllocSize(ETy);
1619 O << "\t.param .align " << align << " .b8 ";
1620 printParamName(I, paramIndex, O);
1621 O << "[" << sz << "]";
1624 // Split the ETy into constituent parts and
1625 // print .param .b<size> <name> for each part.
1626 // Further, if a part is vector, print the above for
1627 // each vector element.
1628 SmallVector<EVT, 16> vtparts;
1629 ComputeValueVTs(*TLI, ETy, vtparts);
1630 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1632 EVT elemtype = vtparts[i];
1633 if (vtparts[i].isVector()) {
1634 elems = vtparts[i].getVectorNumElements();
1635 elemtype = vtparts[i].getVectorElementType();
1638 for (unsigned j = 0, je = elems; j != je; ++j) {
1639 unsigned sz = elemtype.getSizeInBits();
1640 if (elemtype.isInteger() && (sz < 32))
1642 O << "\t.reg .b" << sz << " ";
1643 printParamName(I, paramIndex, O);
1659 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1661 const Function *F = MF.getFunction();
1662 emitFunctionParamList(F, O);
1665 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1666 const MachineFunction &MF) {
1667 SmallString<128> Str;
1668 raw_svector_ostream O(Str);
1670 // Map the global virtual register number to a register class specific
1671 // virtual register number starting from 1 with that class.
1672 const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo();
1673 //unsigned numRegClasses = TRI->getNumRegClasses();
1675 // Emit the Fake Stack Object
1676 const MachineFrameInfo *MFI = MF.getFrameInfo();
1677 int NumBytes = (int) MFI->getStackSize();
1679 O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
1680 << getFunctionNumber() << "[" << NumBytes << "];\n";
1681 if (nvptxSubtarget.is64Bit()) {
1682 O << "\t.reg .b64 \t%SP;\n";
1683 O << "\t.reg .b64 \t%SPL;\n";
1685 O << "\t.reg .b32 \t%SP;\n";
1686 O << "\t.reg .b32 \t%SPL;\n";
1690 // Go through all virtual registers to establish the mapping between the
1692 // register number and the per class virtual register number.
1693 // We use the per class virtual register number in the ptx output.
1694 unsigned int numVRs = MRI->getNumVirtRegs();
1695 for (unsigned i = 0; i < numVRs; i++) {
1696 unsigned int vr = TRI->index2VirtReg(i);
1697 const TargetRegisterClass *RC = MRI->getRegClass(vr);
1698 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1699 int n = regmap.size();
1700 regmap.insert(std::make_pair(vr, n + 1));
1703 // Emit register declarations
1704 // @TODO: Extract out the real register usage
1705 // O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
1706 // O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
1707 // O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
1708 // O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
1709 // O << "\t.reg .s64 %rl<" << NVPTXNumRegisters << ">;\n";
1710 // O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
1711 // O << "\t.reg .f64 %fl<" << NVPTXNumRegisters << ">;\n";
1713 // Emit declaration of the virtual registers or 'physical' registers for
1714 // each register class
1715 for (unsigned i=0; i< TRI->getNumRegClasses(); i++) {
1716 const TargetRegisterClass *RC = TRI->getRegClass(i);
1717 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1718 std::string rcname = getNVPTXRegClassName(RC);
1719 std::string rcStr = getNVPTXRegClassStr(RC);
1720 int n = regmap.size();
1722 // Only declare those registers that may be used.
1724 O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
1729 OutStreamer.EmitRawText(O.str());
1732 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
1733 APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
1735 unsigned int numHex;
1738 if (Fp->getType()->getTypeID() == Type::FloatTyID) {
1741 APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &ignored);
1742 } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
1745 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &ignored);
1747 llvm_unreachable("unsupported fp type");
1749 APInt API = APF.bitcastToAPInt();
1750 std::string hexstr(utohexstr(API.getZExtValue()));
1752 if (hexstr.length() < numHex)
1753 O << std::string(numHex - hexstr.length(), '0');
1754 O << utohexstr(API.getZExtValue());
1757 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
1758 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
1759 O << CI->getValue();
1762 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
1763 printFPConstant(CFP, O);
1766 if (isa<ConstantPointerNull>(CPV)) {
1770 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1771 O << *Mang->getSymbol(GVar);
1774 if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1775 const Value *v = Cexpr->stripPointerCasts();
1776 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
1777 O << *Mang->getSymbol(GVar);
1780 O << *LowerConstant(CPV, *this);
1784 llvm_unreachable("Not scalar type found in printScalarConstant()");
1787 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
1788 AggBuffer *aggBuffer) {
1790 const DataLayout *TD = TM.getDataLayout();
1792 if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
1793 int s = TD->getTypeAllocSize(CPV->getType());
1796 aggBuffer->addZeros(s);
1801 switch (CPV->getType()->getTypeID()) {
1803 case Type::IntegerTyID: {
1804 const Type *ETy = CPV->getType();
1805 if (ETy == Type::getInt8Ty(CPV->getContext())) {
1807 (unsigned char)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
1809 aggBuffer->addBytes(ptr, 1, Bytes);
1810 } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
1811 short int16 = (short)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
1812 ptr = (unsigned char *)&int16;
1813 aggBuffer->addBytes(ptr, 2, Bytes);
1814 } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
1815 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1816 int int32 = (int)(constInt->getZExtValue());
1817 ptr = (unsigned char *)&int32;
1818 aggBuffer->addBytes(ptr, 4, Bytes);
1820 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1821 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1822 ConstantFoldConstantExpression(Cexpr, TD))) {
1823 int int32 = (int)(constInt->getZExtValue());
1824 ptr = (unsigned char *)&int32;
1825 aggBuffer->addBytes(ptr, 4, Bytes);
1828 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1829 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1830 aggBuffer->addSymbol(v);
1831 aggBuffer->addZeros(4);
1835 llvm_unreachable("unsupported integer const type");
1836 } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
1837 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
1838 long long int64 = (long long)(constInt->getZExtValue());
1839 ptr = (unsigned char *)&int64;
1840 aggBuffer->addBytes(ptr, 8, Bytes);
1842 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1843 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
1844 ConstantFoldConstantExpression(Cexpr, TD))) {
1845 long long int64 = (long long)(constInt->getZExtValue());
1846 ptr = (unsigned char *)&int64;
1847 aggBuffer->addBytes(ptr, 8, Bytes);
1850 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
1851 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
1852 aggBuffer->addSymbol(v);
1853 aggBuffer->addZeros(8);
1857 llvm_unreachable("unsupported integer const type");
1859 llvm_unreachable("unsupported integer const type");
1862 case Type::FloatTyID:
1863 case Type::DoubleTyID: {
1864 const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
1865 const Type *Ty = CFP->getType();
1866 if (Ty == Type::getFloatTy(CPV->getContext())) {
1867 float float32 = (float) CFP->getValueAPF().convertToFloat();
1868 ptr = (unsigned char *)&float32;
1869 aggBuffer->addBytes(ptr, 4, Bytes);
1870 } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
1871 double float64 = CFP->getValueAPF().convertToDouble();
1872 ptr = (unsigned char *)&float64;
1873 aggBuffer->addBytes(ptr, 8, Bytes);
1875 llvm_unreachable("unsupported fp const type");
1879 case Type::PointerTyID: {
1880 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
1881 aggBuffer->addSymbol(GVar);
1882 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
1883 const Value *v = Cexpr->stripPointerCasts();
1884 aggBuffer->addSymbol(v);
1886 unsigned int s = TD->getTypeAllocSize(CPV->getType());
1887 aggBuffer->addZeros(s);
1891 case Type::ArrayTyID:
1892 case Type::VectorTyID:
1893 case Type::StructTyID: {
1894 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV) ||
1895 isa<ConstantStruct>(CPV)) {
1896 int ElementSize = TD->getTypeAllocSize(CPV->getType());
1897 bufferAggregateConstant(CPV, aggBuffer);
1898 if (Bytes > ElementSize)
1899 aggBuffer->addZeros(Bytes - ElementSize);
1900 } else if (isa<ConstantAggregateZero>(CPV))
1901 aggBuffer->addZeros(Bytes);
1903 llvm_unreachable("Unexpected Constant type");
1908 llvm_unreachable("unsupported type");
1912 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
1913 AggBuffer *aggBuffer) {
1914 const DataLayout *TD = TM.getDataLayout();
1918 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
1919 if (CPV->getNumOperands())
1920 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
1921 bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
1925 if (const ConstantDataSequential *CDS =
1926 dyn_cast<ConstantDataSequential>(CPV)) {
1927 if (CDS->getNumElements())
1928 for (unsigned i = 0; i < CDS->getNumElements(); ++i)
1929 bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
1934 if (isa<ConstantStruct>(CPV)) {
1935 if (CPV->getNumOperands()) {
1936 StructType *ST = cast<StructType>(CPV->getType());
1937 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
1939 Bytes = TD->getStructLayout(ST)->getElementOffset(0) +
1940 TD->getTypeAllocSize(ST) -
1941 TD->getStructLayout(ST)->getElementOffset(i);
1943 Bytes = TD->getStructLayout(ST)->getElementOffset(i + 1) -
1944 TD->getStructLayout(ST)->getElementOffset(i);
1945 bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
1950 llvm_unreachable("unsupported constant type in printAggregateConstant()");
1953 // buildTypeNameMap - Run through symbol table looking for type names.
1956 bool NVPTXAsmPrinter::isImageType(const Type *Ty) {
1958 std::map<const Type *, std::string>::iterator PI = TypeNameMap.find(Ty);
1960 if (PI != TypeNameMap.end() && (!PI->second.compare("struct._image1d_t") ||
1961 !PI->second.compare("struct._image2d_t") ||
1962 !PI->second.compare("struct._image3d_t")))
1968 /// PrintAsmOperand - Print out an operand for an inline asm expression.
1970 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
1971 unsigned AsmVariant,
1972 const char *ExtraCode, raw_ostream &O) {
1973 if (ExtraCode && ExtraCode[0]) {
1974 if (ExtraCode[1] != 0)
1975 return true; // Unknown modifier.
1977 switch (ExtraCode[0]) {
1979 // See if this is a generic print operand
1980 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
1986 printOperand(MI, OpNo, O);
1991 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(
1992 const MachineInstr *MI, unsigned OpNo, unsigned AsmVariant,
1993 const char *ExtraCode, raw_ostream &O) {
1994 if (ExtraCode && ExtraCode[0])
1995 return true; // Unknown modifier
1998 printMemOperand(MI, OpNo, O);
2004 bool NVPTXAsmPrinter::ignoreLoc(const MachineInstr &MI) {
2005 switch (MI.getOpcode()) {
2008 case NVPTX::CallArgBeginInst:
2009 case NVPTX::CallArgEndInst0:
2010 case NVPTX::CallArgEndInst1:
2011 case NVPTX::CallArgF32:
2012 case NVPTX::CallArgF64:
2013 case NVPTX::CallArgI16:
2014 case NVPTX::CallArgI32:
2015 case NVPTX::CallArgI32imm:
2016 case NVPTX::CallArgI64:
2017 case NVPTX::CallArgI8:
2018 case NVPTX::CallArgParam:
2019 case NVPTX::CallVoidInst:
2020 case NVPTX::CallVoidInstReg:
2021 case NVPTX::Callseq_End:
2022 case NVPTX::CallVoidInstReg64:
2023 case NVPTX::DeclareParamInst:
2024 case NVPTX::DeclareRetMemInst:
2025 case NVPTX::DeclareRetRegInst:
2026 case NVPTX::DeclareRetScalarInst:
2027 case NVPTX::DeclareScalarParamInst:
2028 case NVPTX::DeclareScalarRegInst:
2029 case NVPTX::StoreParamF32:
2030 case NVPTX::StoreParamF64:
2031 case NVPTX::StoreParamI16:
2032 case NVPTX::StoreParamI32:
2033 case NVPTX::StoreParamI64:
2034 case NVPTX::StoreParamI8:
2035 case NVPTX::StoreParamS32I8:
2036 case NVPTX::StoreParamU32I8:
2037 case NVPTX::StoreParamS32I16:
2038 case NVPTX::StoreParamU32I16:
2039 case NVPTX::StoreRetvalF32:
2040 case NVPTX::StoreRetvalF64:
2041 case NVPTX::StoreRetvalI16:
2042 case NVPTX::StoreRetvalI32:
2043 case NVPTX::StoreRetvalI64:
2044 case NVPTX::StoreRetvalI8:
2045 case NVPTX::LastCallArgF32:
2046 case NVPTX::LastCallArgF64:
2047 case NVPTX::LastCallArgI16:
2048 case NVPTX::LastCallArgI32:
2049 case NVPTX::LastCallArgI32imm:
2050 case NVPTX::LastCallArgI64:
2051 case NVPTX::LastCallArgI8:
2052 case NVPTX::LastCallArgParam:
2053 case NVPTX::LoadParamMemF32:
2054 case NVPTX::LoadParamMemF64:
2055 case NVPTX::LoadParamMemI16:
2056 case NVPTX::LoadParamMemI32:
2057 case NVPTX::LoadParamMemI64:
2058 case NVPTX::LoadParamMemI8:
2059 case NVPTX::LoadParamRegF32:
2060 case NVPTX::LoadParamRegF64:
2061 case NVPTX::LoadParamRegI16:
2062 case NVPTX::LoadParamRegI32:
2063 case NVPTX::LoadParamRegI64:
2064 case NVPTX::LoadParamRegI8:
2065 case NVPTX::PrototypeInst:
2066 case NVPTX::DBG_VALUE:
2072 // Force static initialization.
2073 extern "C" void LLVMInitializeNVPTXBackendAsmPrinter() {
2074 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2075 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2078 void NVPTXAsmPrinter::emitSrcInText(StringRef filename, unsigned line) {
2079 std::stringstream temp;
2080 LineReader *reader = this->getReader(filename.str());
2082 temp << filename.str();
2086 temp << reader->readLine(line);
2088 this->OutStreamer.EmitRawText(Twine(temp.str()));
2091 LineReader *NVPTXAsmPrinter::getReader(std::string filename) {
2092 if (reader == NULL) {
2093 reader = new LineReader(filename);
2096 if (reader->fileName() != filename) {
2098 reader = new LineReader(filename);
2104 std::string LineReader::readLine(unsigned lineNum) {
2105 if (lineNum < theCurLine) {
2107 fstr.seekg(0, std::ios::beg);
2109 while (theCurLine < lineNum) {
2110 fstr.getline(buff, 500);
2116 // Force static initialization.
2117 extern "C" void LLVMInitializeNVPTXAsmPrinter() {
2118 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2119 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);