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 "InstPrinter/NVPTXInstPrinter.h"
17 #include "MCTargetDesc/NVPTXMCAsmInfo.h"
19 #include "NVPTXInstrInfo.h"
20 #include "NVPTXMachineFunctionInfo.h"
21 #include "NVPTXMCExpr.h"
22 #include "NVPTXRegisterInfo.h"
23 #include "NVPTXTargetMachine.h"
24 #include "NVPTXUtilities.h"
25 #include "cl_common_defines.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/CodeGen/Analysis.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineModuleInfo.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/IR/DebugInfo.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/GlobalVariable.h"
36 #include "llvm/IR/Mangler.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/IR/Operator.h"
39 #include "llvm/MC/MCStreamer.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/FormattedStream.h"
44 #include "llvm/Support/Path.h"
45 #include "llvm/Support/TargetRegistry.h"
46 #include "llvm/Support/TimeValue.h"
47 #include "llvm/Target/TargetLoweringObjectFile.h"
51 #define DEPOTNAME "__local_depot"
54 EmitLineNumbers("nvptx-emit-line-numbers", cl::Hidden,
55 cl::desc("NVPTX Specific: Emit Line numbers even without -G"),
59 InterleaveSrc("nvptx-emit-src", cl::ZeroOrMore, cl::Hidden,
60 cl::desc("NVPTX Specific: Emit source line in ptx file"),
64 /// DiscoverDependentGlobals - Return a set of GlobalVariables on which \p V
66 void DiscoverDependentGlobals(const Value *V,
67 DenseSet<const GlobalVariable *> &Globals) {
68 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
71 if (const User *U = dyn_cast<User>(V)) {
72 for (unsigned i = 0, e = U->getNumOperands(); i != e; ++i) {
73 DiscoverDependentGlobals(U->getOperand(i), Globals);
79 /// VisitGlobalVariableForEmission - Add \p GV to the list of GlobalVariable
80 /// instances to be emitted, but only after any dependents have been added
82 void VisitGlobalVariableForEmission(
83 const GlobalVariable *GV, SmallVectorImpl<const GlobalVariable *> &Order,
84 DenseSet<const GlobalVariable *> &Visited,
85 DenseSet<const GlobalVariable *> &Visiting) {
86 // Have we already visited this one?
87 if (Visited.count(GV))
90 // Do we have a circular dependency?
91 if (Visiting.count(GV))
92 report_fatal_error("Circular dependency found in global variable set");
94 // Start visiting this global
97 // Make sure we visit all dependents first
98 DenseSet<const GlobalVariable *> Others;
99 for (unsigned i = 0, e = GV->getNumOperands(); i != e; ++i)
100 DiscoverDependentGlobals(GV->getOperand(i), Others);
102 for (DenseSet<const GlobalVariable *>::iterator I = Others.begin(),
105 VisitGlobalVariableForEmission(*I, Order, Visited, Visiting);
107 // Now we can visit ourself
114 // @TODO: This is a copy from AsmPrinter.cpp. The function is static, so we
115 // cannot just link to the existing version.
116 /// LowerConstant - Lower the specified LLVM Constant to an MCExpr.
118 using namespace nvptx;
119 const MCExpr *nvptx::LowerConstant(const Constant *CV, AsmPrinter &AP) {
120 MCContext &Ctx = AP.OutContext;
122 if (CV->isNullValue() || isa<UndefValue>(CV))
123 return MCConstantExpr::Create(0, Ctx);
125 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV))
126 return MCConstantExpr::Create(CI->getZExtValue(), Ctx);
128 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CV))
129 return MCSymbolRefExpr::Create(AP.getSymbol(GV), Ctx);
131 if (const BlockAddress *BA = dyn_cast<BlockAddress>(CV))
132 return MCSymbolRefExpr::Create(AP.GetBlockAddressSymbol(BA), Ctx);
134 const ConstantExpr *CE = dyn_cast<ConstantExpr>(CV);
136 llvm_unreachable("Unknown constant value to lower!");
138 switch (CE->getOpcode()) {
140 // If the code isn't optimized, there may be outstanding folding
141 // opportunities. Attempt to fold the expression using DataLayout as a
142 // last resort before giving up.
143 if (Constant *C = ConstantFoldConstantExpression(CE, AP.TM.getDataLayout()))
145 return LowerConstant(C, AP);
147 // Otherwise report the problem to the user.
150 raw_string_ostream OS(S);
151 OS << "Unsupported expression in static initializer: ";
152 CE->printAsOperand(OS, /*PrintType=*/ false,
153 !AP.MF ? nullptr : AP.MF->getFunction()->getParent());
154 report_fatal_error(OS.str());
156 case Instruction::AddrSpaceCast: {
157 // Strip any addrspace(1)->addrspace(0) addrspace casts. These will be
158 // handled by the generic() logic in the MCExpr printer
159 PointerType *DstTy = cast<PointerType>(CE->getType());
160 PointerType *SrcTy = cast<PointerType>(CE->getOperand(0)->getType());
161 if (SrcTy->getAddressSpace() == 1 && DstTy->getAddressSpace() == 0) {
162 return LowerConstant(cast<const Constant>(CE->getOperand(0)), AP);
165 raw_string_ostream OS(S);
166 OS << "Unsupported expression in static initializer: ";
167 CE->printAsOperand(OS, /*PrintType=*/ false,
168 !AP.MF ? nullptr : AP.MF->getFunction()->getParent());
169 report_fatal_error(OS.str());
171 case Instruction::GetElementPtr: {
172 const DataLayout &TD = *AP.TM.getDataLayout();
173 // Generate a symbolic expression for the byte address
174 APInt OffsetAI(TD.getPointerSizeInBits(), 0);
175 cast<GEPOperator>(CE)->accumulateConstantOffset(TD, OffsetAI);
177 const MCExpr *Base = LowerConstant(CE->getOperand(0), AP);
181 int64_t Offset = OffsetAI.getSExtValue();
182 return MCBinaryExpr::CreateAdd(Base, MCConstantExpr::Create(Offset, Ctx),
186 case Instruction::Trunc:
187 // We emit the value and depend on the assembler to truncate the generated
188 // expression properly. This is important for differences between
189 // blockaddress labels. Since the two labels are in the same function, it
190 // is reasonable to treat their delta as a 32-bit value.
192 case Instruction::BitCast:
193 return LowerConstant(CE->getOperand(0), AP);
195 case Instruction::IntToPtr: {
196 const DataLayout &TD = *AP.TM.getDataLayout();
197 // Handle casts to pointers by changing them into casts to the appropriate
198 // integer type. This promotes constant folding and simplifies this code.
199 Constant *Op = CE->getOperand(0);
200 Op = ConstantExpr::getIntegerCast(Op, TD.getIntPtrType(CV->getContext()),
202 return LowerConstant(Op, AP);
205 case Instruction::PtrToInt: {
206 const DataLayout &TD = *AP.TM.getDataLayout();
207 // Support only foldable casts to/from pointers that can be eliminated by
208 // changing the pointer to the appropriately sized integer type.
209 Constant *Op = CE->getOperand(0);
210 Type *Ty = CE->getType();
212 const MCExpr *OpExpr = LowerConstant(Op, AP);
214 // We can emit the pointer value into this slot if the slot is an
215 // integer slot equal to the size of the pointer.
216 if (TD.getTypeAllocSize(Ty) == TD.getTypeAllocSize(Op->getType()))
219 // Otherwise the pointer is smaller than the resultant integer, mask off
220 // the high bits so we are sure to get a proper truncation if the input is
222 unsigned InBits = TD.getTypeAllocSizeInBits(Op->getType());
223 const MCExpr *MaskExpr =
224 MCConstantExpr::Create(~0ULL >> (64 - InBits), Ctx);
225 return MCBinaryExpr::CreateAnd(OpExpr, MaskExpr, Ctx);
228 // The MC library also has a right-shift operator, but it isn't consistently
229 // signed or unsigned between different targets.
230 case Instruction::Add:
231 case Instruction::Sub:
232 case Instruction::Mul:
233 case Instruction::SDiv:
234 case Instruction::SRem:
235 case Instruction::Shl:
236 case Instruction::And:
237 case Instruction::Or:
238 case Instruction::Xor: {
239 const MCExpr *LHS = LowerConstant(CE->getOperand(0), AP);
240 const MCExpr *RHS = LowerConstant(CE->getOperand(1), AP);
241 switch (CE->getOpcode()) {
243 llvm_unreachable("Unknown binary operator constant cast expr");
244 case Instruction::Add:
245 return MCBinaryExpr::CreateAdd(LHS, RHS, Ctx);
246 case Instruction::Sub:
247 return MCBinaryExpr::CreateSub(LHS, RHS, Ctx);
248 case Instruction::Mul:
249 return MCBinaryExpr::CreateMul(LHS, RHS, Ctx);
250 case Instruction::SDiv:
251 return MCBinaryExpr::CreateDiv(LHS, RHS, Ctx);
252 case Instruction::SRem:
253 return MCBinaryExpr::CreateMod(LHS, RHS, Ctx);
254 case Instruction::Shl:
255 return MCBinaryExpr::CreateShl(LHS, RHS, Ctx);
256 case Instruction::And:
257 return MCBinaryExpr::CreateAnd(LHS, RHS, Ctx);
258 case Instruction::Or:
259 return MCBinaryExpr::CreateOr(LHS, RHS, Ctx);
260 case Instruction::Xor:
261 return MCBinaryExpr::CreateXor(LHS, RHS, Ctx);
267 void NVPTXAsmPrinter::emitLineNumberAsDotLoc(const MachineInstr &MI) {
268 if (!EmitLineNumbers)
273 DebugLoc curLoc = MI.getDebugLoc();
275 if (prevDebugLoc.isUnknown() && curLoc.isUnknown())
278 if (prevDebugLoc == curLoc)
281 prevDebugLoc = curLoc;
283 if (curLoc.isUnknown())
286 const MachineFunction *MF = MI.getParent()->getParent();
287 //const TargetMachine &TM = MF->getTarget();
289 const LLVMContext &ctx = MF->getFunction()->getContext();
290 DIScope Scope(curLoc.getScope(ctx));
292 assert((!Scope || Scope.isScope()) &&
293 "Scope of a DebugLoc should be null or a DIScope.");
297 StringRef fileName(Scope.getFilename());
298 StringRef dirName(Scope.getDirectory());
299 SmallString<128> FullPathName = dirName;
300 if (!dirName.empty() && !sys::path::is_absolute(fileName)) {
301 sys::path::append(FullPathName, fileName);
302 fileName = FullPathName.str();
305 if (filenameMap.find(fileName.str()) == filenameMap.end())
308 // Emit the line from the source file.
310 this->emitSrcInText(fileName.str(), curLoc.getLine());
312 std::stringstream temp;
313 temp << "\t.loc " << filenameMap[fileName.str()] << " " << curLoc.getLine()
314 << " " << curLoc.getCol();
315 OutStreamer.EmitRawText(Twine(temp.str().c_str()));
318 void NVPTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
319 SmallString<128> Str;
320 raw_svector_ostream OS(Str);
321 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
322 emitLineNumberAsDotLoc(*MI);
325 lowerToMCInst(MI, Inst);
326 EmitToStreamer(OutStreamer, Inst);
329 // Handle symbol backtracking for targets that do not support image handles
330 bool NVPTXAsmPrinter::lowerImageHandleOperand(const MachineInstr *MI,
331 unsigned OpNo, MCOperand &MCOp) {
332 const MachineOperand &MO = MI->getOperand(OpNo);
334 switch (MI->getOpcode()) {
335 default: return false;
336 case NVPTX::TEX_1D_F32_I32:
337 case NVPTX::TEX_1D_F32_F32:
338 case NVPTX::TEX_1D_F32_F32_LEVEL:
339 case NVPTX::TEX_1D_F32_F32_GRAD:
340 case NVPTX::TEX_1D_I32_I32:
341 case NVPTX::TEX_1D_I32_F32:
342 case NVPTX::TEX_1D_I32_F32_LEVEL:
343 case NVPTX::TEX_1D_I32_F32_GRAD:
344 case NVPTX::TEX_1D_ARRAY_F32_I32:
345 case NVPTX::TEX_1D_ARRAY_F32_F32:
346 case NVPTX::TEX_1D_ARRAY_F32_F32_LEVEL:
347 case NVPTX::TEX_1D_ARRAY_F32_F32_GRAD:
348 case NVPTX::TEX_1D_ARRAY_I32_I32:
349 case NVPTX::TEX_1D_ARRAY_I32_F32:
350 case NVPTX::TEX_1D_ARRAY_I32_F32_LEVEL:
351 case NVPTX::TEX_1D_ARRAY_I32_F32_GRAD:
352 case NVPTX::TEX_2D_F32_I32:
353 case NVPTX::TEX_2D_F32_F32:
354 case NVPTX::TEX_2D_F32_F32_LEVEL:
355 case NVPTX::TEX_2D_F32_F32_GRAD:
356 case NVPTX::TEX_2D_I32_I32:
357 case NVPTX::TEX_2D_I32_F32:
358 case NVPTX::TEX_2D_I32_F32_LEVEL:
359 case NVPTX::TEX_2D_I32_F32_GRAD:
360 case NVPTX::TEX_2D_ARRAY_F32_I32:
361 case NVPTX::TEX_2D_ARRAY_F32_F32:
362 case NVPTX::TEX_2D_ARRAY_F32_F32_LEVEL:
363 case NVPTX::TEX_2D_ARRAY_F32_F32_GRAD:
364 case NVPTX::TEX_2D_ARRAY_I32_I32:
365 case NVPTX::TEX_2D_ARRAY_I32_F32:
366 case NVPTX::TEX_2D_ARRAY_I32_F32_LEVEL:
367 case NVPTX::TEX_2D_ARRAY_I32_F32_GRAD:
368 case NVPTX::TEX_3D_F32_I32:
369 case NVPTX::TEX_3D_F32_F32:
370 case NVPTX::TEX_3D_F32_F32_LEVEL:
371 case NVPTX::TEX_3D_F32_F32_GRAD:
372 case NVPTX::TEX_3D_I32_I32:
373 case NVPTX::TEX_3D_I32_F32:
374 case NVPTX::TEX_3D_I32_F32_LEVEL:
375 case NVPTX::TEX_3D_I32_F32_GRAD:
377 // This is a texture fetch, so operand 4 is a texref and operand 5 is
380 lowerImageHandleSymbol(MO.getImm(), MCOp);
384 lowerImageHandleSymbol(MO.getImm(), MCOp);
390 case NVPTX::SULD_1D_I8_TRAP:
391 case NVPTX::SULD_1D_I16_TRAP:
392 case NVPTX::SULD_1D_I32_TRAP:
393 case NVPTX::SULD_1D_ARRAY_I8_TRAP:
394 case NVPTX::SULD_1D_ARRAY_I16_TRAP:
395 case NVPTX::SULD_1D_ARRAY_I32_TRAP:
396 case NVPTX::SULD_2D_I8_TRAP:
397 case NVPTX::SULD_2D_I16_TRAP:
398 case NVPTX::SULD_2D_I32_TRAP:
399 case NVPTX::SULD_2D_ARRAY_I8_TRAP:
400 case NVPTX::SULD_2D_ARRAY_I16_TRAP:
401 case NVPTX::SULD_2D_ARRAY_I32_TRAP:
402 case NVPTX::SULD_3D_I8_TRAP:
403 case NVPTX::SULD_3D_I16_TRAP:
404 case NVPTX::SULD_3D_I32_TRAP: {
405 // This is a V1 surface load, so operand 1 is a surfref
407 lowerImageHandleSymbol(MO.getImm(), MCOp);
413 case NVPTX::SULD_1D_V2I8_TRAP:
414 case NVPTX::SULD_1D_V2I16_TRAP:
415 case NVPTX::SULD_1D_V2I32_TRAP:
416 case NVPTX::SULD_1D_ARRAY_V2I8_TRAP:
417 case NVPTX::SULD_1D_ARRAY_V2I16_TRAP:
418 case NVPTX::SULD_1D_ARRAY_V2I32_TRAP:
419 case NVPTX::SULD_2D_V2I8_TRAP:
420 case NVPTX::SULD_2D_V2I16_TRAP:
421 case NVPTX::SULD_2D_V2I32_TRAP:
422 case NVPTX::SULD_2D_ARRAY_V2I8_TRAP:
423 case NVPTX::SULD_2D_ARRAY_V2I16_TRAP:
424 case NVPTX::SULD_2D_ARRAY_V2I32_TRAP:
425 case NVPTX::SULD_3D_V2I8_TRAP:
426 case NVPTX::SULD_3D_V2I16_TRAP:
427 case NVPTX::SULD_3D_V2I32_TRAP: {
428 // This is a V2 surface load, so operand 2 is a surfref
430 lowerImageHandleSymbol(MO.getImm(), MCOp);
436 case NVPTX::SULD_1D_V4I8_TRAP:
437 case NVPTX::SULD_1D_V4I16_TRAP:
438 case NVPTX::SULD_1D_V4I32_TRAP:
439 case NVPTX::SULD_1D_ARRAY_V4I8_TRAP:
440 case NVPTX::SULD_1D_ARRAY_V4I16_TRAP:
441 case NVPTX::SULD_1D_ARRAY_V4I32_TRAP:
442 case NVPTX::SULD_2D_V4I8_TRAP:
443 case NVPTX::SULD_2D_V4I16_TRAP:
444 case NVPTX::SULD_2D_V4I32_TRAP:
445 case NVPTX::SULD_2D_ARRAY_V4I8_TRAP:
446 case NVPTX::SULD_2D_ARRAY_V4I16_TRAP:
447 case NVPTX::SULD_2D_ARRAY_V4I32_TRAP:
448 case NVPTX::SULD_3D_V4I8_TRAP:
449 case NVPTX::SULD_3D_V4I16_TRAP:
450 case NVPTX::SULD_3D_V4I32_TRAP: {
451 // This is a V4 surface load, so operand 4 is a surfref
453 lowerImageHandleSymbol(MO.getImm(), MCOp);
459 case NVPTX::SUST_B_1D_B8_TRAP:
460 case NVPTX::SUST_B_1D_B16_TRAP:
461 case NVPTX::SUST_B_1D_B32_TRAP:
462 case NVPTX::SUST_B_1D_V2B8_TRAP:
463 case NVPTX::SUST_B_1D_V2B16_TRAP:
464 case NVPTX::SUST_B_1D_V2B32_TRAP:
465 case NVPTX::SUST_B_1D_V4B8_TRAP:
466 case NVPTX::SUST_B_1D_V4B16_TRAP:
467 case NVPTX::SUST_B_1D_V4B32_TRAP:
468 case NVPTX::SUST_B_1D_ARRAY_B8_TRAP:
469 case NVPTX::SUST_B_1D_ARRAY_B16_TRAP:
470 case NVPTX::SUST_B_1D_ARRAY_B32_TRAP:
471 case NVPTX::SUST_B_1D_ARRAY_V2B8_TRAP:
472 case NVPTX::SUST_B_1D_ARRAY_V2B16_TRAP:
473 case NVPTX::SUST_B_1D_ARRAY_V2B32_TRAP:
474 case NVPTX::SUST_B_1D_ARRAY_V4B8_TRAP:
475 case NVPTX::SUST_B_1D_ARRAY_V4B16_TRAP:
476 case NVPTX::SUST_B_1D_ARRAY_V4B32_TRAP:
477 case NVPTX::SUST_B_2D_B8_TRAP:
478 case NVPTX::SUST_B_2D_B16_TRAP:
479 case NVPTX::SUST_B_2D_B32_TRAP:
480 case NVPTX::SUST_B_2D_V2B8_TRAP:
481 case NVPTX::SUST_B_2D_V2B16_TRAP:
482 case NVPTX::SUST_B_2D_V2B32_TRAP:
483 case NVPTX::SUST_B_2D_V4B8_TRAP:
484 case NVPTX::SUST_B_2D_V4B16_TRAP:
485 case NVPTX::SUST_B_2D_V4B32_TRAP:
486 case NVPTX::SUST_B_2D_ARRAY_B8_TRAP:
487 case NVPTX::SUST_B_2D_ARRAY_B16_TRAP:
488 case NVPTX::SUST_B_2D_ARRAY_B32_TRAP:
489 case NVPTX::SUST_B_2D_ARRAY_V2B8_TRAP:
490 case NVPTX::SUST_B_2D_ARRAY_V2B16_TRAP:
491 case NVPTX::SUST_B_2D_ARRAY_V2B32_TRAP:
492 case NVPTX::SUST_B_2D_ARRAY_V4B8_TRAP:
493 case NVPTX::SUST_B_2D_ARRAY_V4B16_TRAP:
494 case NVPTX::SUST_B_2D_ARRAY_V4B32_TRAP:
495 case NVPTX::SUST_B_3D_B8_TRAP:
496 case NVPTX::SUST_B_3D_B16_TRAP:
497 case NVPTX::SUST_B_3D_B32_TRAP:
498 case NVPTX::SUST_B_3D_V2B8_TRAP:
499 case NVPTX::SUST_B_3D_V2B16_TRAP:
500 case NVPTX::SUST_B_3D_V2B32_TRAP:
501 case NVPTX::SUST_B_3D_V4B8_TRAP:
502 case NVPTX::SUST_B_3D_V4B16_TRAP:
503 case NVPTX::SUST_B_3D_V4B32_TRAP:
504 case NVPTX::SUST_P_1D_B8_TRAP:
505 case NVPTX::SUST_P_1D_B16_TRAP:
506 case NVPTX::SUST_P_1D_B32_TRAP:
507 case NVPTX::SUST_P_1D_V2B8_TRAP:
508 case NVPTX::SUST_P_1D_V2B16_TRAP:
509 case NVPTX::SUST_P_1D_V2B32_TRAP:
510 case NVPTX::SUST_P_1D_V4B8_TRAP:
511 case NVPTX::SUST_P_1D_V4B16_TRAP:
512 case NVPTX::SUST_P_1D_V4B32_TRAP:
513 case NVPTX::SUST_P_1D_ARRAY_B8_TRAP:
514 case NVPTX::SUST_P_1D_ARRAY_B16_TRAP:
515 case NVPTX::SUST_P_1D_ARRAY_B32_TRAP:
516 case NVPTX::SUST_P_1D_ARRAY_V2B8_TRAP:
517 case NVPTX::SUST_P_1D_ARRAY_V2B16_TRAP:
518 case NVPTX::SUST_P_1D_ARRAY_V2B32_TRAP:
519 case NVPTX::SUST_P_1D_ARRAY_V4B8_TRAP:
520 case NVPTX::SUST_P_1D_ARRAY_V4B16_TRAP:
521 case NVPTX::SUST_P_1D_ARRAY_V4B32_TRAP:
522 case NVPTX::SUST_P_2D_B8_TRAP:
523 case NVPTX::SUST_P_2D_B16_TRAP:
524 case NVPTX::SUST_P_2D_B32_TRAP:
525 case NVPTX::SUST_P_2D_V2B8_TRAP:
526 case NVPTX::SUST_P_2D_V2B16_TRAP:
527 case NVPTX::SUST_P_2D_V2B32_TRAP:
528 case NVPTX::SUST_P_2D_V4B8_TRAP:
529 case NVPTX::SUST_P_2D_V4B16_TRAP:
530 case NVPTX::SUST_P_2D_V4B32_TRAP:
531 case NVPTX::SUST_P_2D_ARRAY_B8_TRAP:
532 case NVPTX::SUST_P_2D_ARRAY_B16_TRAP:
533 case NVPTX::SUST_P_2D_ARRAY_B32_TRAP:
534 case NVPTX::SUST_P_2D_ARRAY_V2B8_TRAP:
535 case NVPTX::SUST_P_2D_ARRAY_V2B16_TRAP:
536 case NVPTX::SUST_P_2D_ARRAY_V2B32_TRAP:
537 case NVPTX::SUST_P_2D_ARRAY_V4B8_TRAP:
538 case NVPTX::SUST_P_2D_ARRAY_V4B16_TRAP:
539 case NVPTX::SUST_P_2D_ARRAY_V4B32_TRAP:
540 case NVPTX::SUST_P_3D_B8_TRAP:
541 case NVPTX::SUST_P_3D_B16_TRAP:
542 case NVPTX::SUST_P_3D_B32_TRAP:
543 case NVPTX::SUST_P_3D_V2B8_TRAP:
544 case NVPTX::SUST_P_3D_V2B16_TRAP:
545 case NVPTX::SUST_P_3D_V2B32_TRAP:
546 case NVPTX::SUST_P_3D_V4B8_TRAP:
547 case NVPTX::SUST_P_3D_V4B16_TRAP:
548 case NVPTX::SUST_P_3D_V4B32_TRAP: {
549 // This is a surface store, so operand 0 is a surfref
551 lowerImageHandleSymbol(MO.getImm(), MCOp);
557 case NVPTX::TXQ_CHANNEL_ORDER:
558 case NVPTX::TXQ_CHANNEL_DATA_TYPE:
559 case NVPTX::TXQ_WIDTH:
560 case NVPTX::TXQ_HEIGHT:
561 case NVPTX::TXQ_DEPTH:
562 case NVPTX::TXQ_ARRAY_SIZE:
563 case NVPTX::TXQ_NUM_SAMPLES:
564 case NVPTX::TXQ_NUM_MIPMAP_LEVELS:
565 case NVPTX::SUQ_CHANNEL_ORDER:
566 case NVPTX::SUQ_CHANNEL_DATA_TYPE:
567 case NVPTX::SUQ_WIDTH:
568 case NVPTX::SUQ_HEIGHT:
569 case NVPTX::SUQ_DEPTH:
570 case NVPTX::SUQ_ARRAY_SIZE: {
571 // This is a query, so operand 1 is a surfref/texref
573 lowerImageHandleSymbol(MO.getImm(), MCOp);
582 void NVPTXAsmPrinter::lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp) {
584 TargetMachine &TM = const_cast<TargetMachine&>(MF->getTarget());
585 NVPTXTargetMachine &nvTM = static_cast<NVPTXTargetMachine&>(TM);
586 const NVPTXMachineFunctionInfo *MFI = MF->getInfo<NVPTXMachineFunctionInfo>();
587 const char *Sym = MFI->getImageHandleSymbol(Index);
588 std::string *SymNamePtr =
589 nvTM.getManagedStrPool()->getManagedString(Sym);
590 MCOp = GetSymbolRef(OutContext.GetOrCreateSymbol(
591 StringRef(SymNamePtr->c_str())));
594 void NVPTXAsmPrinter::lowerToMCInst(const MachineInstr *MI, MCInst &OutMI) {
595 OutMI.setOpcode(MI->getOpcode());
596 const NVPTXSubtarget &ST = TM.getSubtarget<NVPTXSubtarget>();
598 // Special: Do not mangle symbol operand of CALL_PROTOTYPE
599 if (MI->getOpcode() == NVPTX::CALL_PROTOTYPE) {
600 const MachineOperand &MO = MI->getOperand(0);
601 OutMI.addOperand(GetSymbolRef(
602 OutContext.GetOrCreateSymbol(Twine(MO.getSymbolName()))));
606 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
607 const MachineOperand &MO = MI->getOperand(i);
610 if (!ST.hasImageHandles()) {
611 if (lowerImageHandleOperand(MI, i, MCOp)) {
612 OutMI.addOperand(MCOp);
617 if (lowerOperand(MO, MCOp))
618 OutMI.addOperand(MCOp);
622 bool NVPTXAsmPrinter::lowerOperand(const MachineOperand &MO,
624 switch (MO.getType()) {
625 default: llvm_unreachable("unknown operand type");
626 case MachineOperand::MO_Register:
627 MCOp = MCOperand::CreateReg(encodeVirtualRegister(MO.getReg()));
629 case MachineOperand::MO_Immediate:
630 MCOp = MCOperand::CreateImm(MO.getImm());
632 case MachineOperand::MO_MachineBasicBlock:
633 MCOp = MCOperand::CreateExpr(MCSymbolRefExpr::Create(
634 MO.getMBB()->getSymbol(), OutContext));
636 case MachineOperand::MO_ExternalSymbol:
637 MCOp = GetSymbolRef(GetExternalSymbolSymbol(MO.getSymbolName()));
639 case MachineOperand::MO_GlobalAddress:
640 MCOp = GetSymbolRef(getSymbol(MO.getGlobal()));
642 case MachineOperand::MO_FPImmediate: {
643 const ConstantFP *Cnt = MO.getFPImm();
644 APFloat Val = Cnt->getValueAPF();
646 switch (Cnt->getType()->getTypeID()) {
647 default: report_fatal_error("Unsupported FP type"); break;
648 case Type::FloatTyID:
649 MCOp = MCOperand::CreateExpr(
650 NVPTXFloatMCExpr::CreateConstantFPSingle(Val, OutContext));
652 case Type::DoubleTyID:
653 MCOp = MCOperand::CreateExpr(
654 NVPTXFloatMCExpr::CreateConstantFPDouble(Val, OutContext));
663 unsigned NVPTXAsmPrinter::encodeVirtualRegister(unsigned Reg) {
664 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
665 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
667 DenseMap<unsigned, unsigned> &RegMap = VRegMapping[RC];
668 unsigned RegNum = RegMap[Reg];
670 // Encode the register class in the upper 4 bits
671 // Must be kept in sync with NVPTXInstPrinter::printRegName
673 if (RC == &NVPTX::Int1RegsRegClass) {
675 } else if (RC == &NVPTX::Int16RegsRegClass) {
677 } else if (RC == &NVPTX::Int32RegsRegClass) {
679 } else if (RC == &NVPTX::Int64RegsRegClass) {
681 } else if (RC == &NVPTX::Float32RegsRegClass) {
683 } else if (RC == &NVPTX::Float64RegsRegClass) {
686 report_fatal_error("Bad register class");
689 // Insert the vreg number
690 Ret |= (RegNum & 0x0FFFFFFF);
693 // Some special-use registers are actually physical registers.
694 // Encode this as the register class ID of 0 and the real register ID.
695 return Reg & 0x0FFFFFFF;
699 MCOperand NVPTXAsmPrinter::GetSymbolRef(const MCSymbol *Symbol) {
701 Expr = MCSymbolRefExpr::Create(Symbol, MCSymbolRefExpr::VK_None,
703 return MCOperand::CreateExpr(Expr);
706 void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
707 const DataLayout *TD = TM.getDataLayout();
708 const TargetLowering *TLI = TM.getTargetLowering();
710 Type *Ty = F->getReturnType();
712 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
714 if (Ty->getTypeID() == Type::VoidTyID)
720 if (Ty->isFloatingPointTy() || Ty->isIntegerTy()) {
722 if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty)) {
723 size = ITy->getBitWidth();
727 assert(Ty->isFloatingPointTy() && "Floating point type expected here");
728 size = Ty->getPrimitiveSizeInBits();
731 O << ".param .b" << size << " func_retval0";
732 } else if (isa<PointerType>(Ty)) {
733 O << ".param .b" << TLI->getPointerTy().getSizeInBits()
736 if ((Ty->getTypeID() == Type::StructTyID) || isa<VectorType>(Ty)) {
737 unsigned totalsz = TD->getTypeAllocSize(Ty);
738 unsigned retAlignment = 0;
739 if (!llvm::getAlign(*F, 0, retAlignment))
740 retAlignment = TD->getABITypeAlignment(Ty);
741 O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
744 assert(false && "Unknown return type");
747 SmallVector<EVT, 16> vtparts;
748 ComputeValueVTs(*TLI, Ty, vtparts);
750 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
752 EVT elemtype = vtparts[i];
753 if (vtparts[i].isVector()) {
754 elems = vtparts[i].getVectorNumElements();
755 elemtype = vtparts[i].getVectorElementType();
758 for (unsigned j = 0, je = elems; j != je; ++j) {
759 unsigned sz = elemtype.getSizeInBits();
760 if (elemtype.isInteger() && (sz < 32))
762 O << ".reg .b" << sz << " func_retval" << idx;
775 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
777 const Function *F = MF.getFunction();
778 printReturnValStr(F, O);
781 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
782 SmallString<128> Str;
783 raw_svector_ostream O(Str);
785 if (!GlobalsEmitted) {
786 emitGlobals(*MF->getFunction()->getParent());
787 GlobalsEmitted = true;
791 MRI = &MF->getRegInfo();
792 F = MF->getFunction();
793 emitLinkageDirective(F, O);
794 if (llvm::isKernelFunction(*F))
798 printReturnValStr(*MF, O);
803 emitFunctionParamList(*MF, O);
805 if (llvm::isKernelFunction(*F))
806 emitKernelFunctionDirectives(*F, O);
808 OutStreamer.EmitRawText(O.str());
810 prevDebugLoc = DebugLoc();
813 void NVPTXAsmPrinter::EmitFunctionBodyStart() {
815 OutStreamer.EmitRawText(StringRef("{\n"));
816 setAndEmitFunctionVirtualRegisters(*MF);
818 SmallString<128> Str;
819 raw_svector_ostream O(Str);
820 emitDemotedVars(MF->getFunction(), O);
821 OutStreamer.EmitRawText(O.str());
824 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
825 OutStreamer.EmitRawText(StringRef("}\n"));
829 void NVPTXAsmPrinter::emitImplicitDef(const MachineInstr *MI) const {
830 unsigned RegNo = MI->getOperand(0).getReg();
831 const TargetRegisterInfo *TRI = TM.getRegisterInfo();
832 if (TRI->isVirtualRegister(RegNo)) {
833 OutStreamer.AddComment(Twine("implicit-def: ") +
834 getVirtualRegisterName(RegNo));
836 OutStreamer.AddComment(Twine("implicit-def: ") +
837 TM.getRegisterInfo()->getName(RegNo));
839 OutStreamer.AddBlankLine();
842 void NVPTXAsmPrinter::emitKernelFunctionDirectives(const Function &F,
843 raw_ostream &O) const {
844 // If the NVVM IR has some of reqntid* specified, then output
845 // the reqntid directive, and set the unspecified ones to 1.
846 // If none of reqntid* is specified, don't output reqntid directive.
847 unsigned reqntidx, reqntidy, reqntidz;
848 bool specified = false;
849 if (llvm::getReqNTIDx(F, reqntidx) == false)
853 if (llvm::getReqNTIDy(F, reqntidy) == false)
857 if (llvm::getReqNTIDz(F, reqntidz) == false)
863 O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
866 // If the NVVM IR has some of maxntid* specified, then output
867 // the maxntid directive, and set the unspecified ones to 1.
868 // If none of maxntid* is specified, don't output maxntid directive.
869 unsigned maxntidx, maxntidy, maxntidz;
871 if (llvm::getMaxNTIDx(F, maxntidx) == false)
875 if (llvm::getMaxNTIDy(F, maxntidy) == false)
879 if (llvm::getMaxNTIDz(F, maxntidz) == false)
885 O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
889 if (llvm::getMinCTASm(F, mincta))
890 O << ".minnctapersm " << mincta << "\n";
894 NVPTXAsmPrinter::getVirtualRegisterName(unsigned Reg) const {
895 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
898 raw_string_ostream NameStr(Name);
900 VRegRCMap::const_iterator I = VRegMapping.find(RC);
901 assert(I != VRegMapping.end() && "Bad register class");
902 const DenseMap<unsigned, unsigned> &RegMap = I->second;
904 VRegMap::const_iterator VI = RegMap.find(Reg);
905 assert(VI != RegMap.end() && "Bad virtual register");
906 unsigned MappedVR = VI->second;
908 NameStr << getNVPTXRegClassStr(RC) << MappedVR;
914 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr,
916 O << getVirtualRegisterName(vr);
919 void NVPTXAsmPrinter::printVecModifiedImmediate(
920 const MachineOperand &MO, const char *Modifier, raw_ostream &O) {
921 static const char vecelem[] = { '0', '1', '2', '3', '0', '1', '2', '3' };
922 int Imm = (int) MO.getImm();
923 if (0 == strcmp(Modifier, "vecelem"))
924 O << "_" << vecelem[Imm];
925 else if (0 == strcmp(Modifier, "vecv4comm1")) {
926 if ((Imm < 0) || (Imm > 3))
928 } else if (0 == strcmp(Modifier, "vecv4comm2")) {
929 if ((Imm < 4) || (Imm > 7))
931 } else if (0 == strcmp(Modifier, "vecv4pos")) {
934 O << "_" << vecelem[Imm % 4];
935 } else if (0 == strcmp(Modifier, "vecv2comm1")) {
936 if ((Imm < 0) || (Imm > 1))
938 } else if (0 == strcmp(Modifier, "vecv2comm2")) {
939 if ((Imm < 2) || (Imm > 3))
941 } else if (0 == strcmp(Modifier, "vecv2pos")) {
944 O << "_" << vecelem[Imm % 2];
946 llvm_unreachable("Unknown Modifier on immediate operand");
951 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
953 emitLinkageDirective(F, O);
954 if (llvm::isKernelFunction(*F))
958 printReturnValStr(F, O);
959 O << *getSymbol(F) << "\n";
960 emitFunctionParamList(F, O);
964 static bool usedInGlobalVarDef(const Constant *C) {
968 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
969 if (GV->getName().str() == "llvm.used")
974 for (const User *U : C->users())
975 if (const Constant *C = dyn_cast<Constant>(U))
976 if (usedInGlobalVarDef(C))
982 static bool usedInOneFunc(const User *U, Function const *&oneFunc) {
983 if (const GlobalVariable *othergv = dyn_cast<GlobalVariable>(U)) {
984 if (othergv->getName().str() == "llvm.used")
988 if (const Instruction *instr = dyn_cast<Instruction>(U)) {
989 if (instr->getParent() && instr->getParent()->getParent()) {
990 const Function *curFunc = instr->getParent()->getParent();
991 if (oneFunc && (curFunc != oneFunc))
999 if (const MDNode *md = dyn_cast<MDNode>(U))
1000 if (md->hasName() && ((md->getName().str() == "llvm.dbg.gv") ||
1001 (md->getName().str() == "llvm.dbg.sp")))
1004 for (const User *UU : U->users())
1005 if (usedInOneFunc(UU, oneFunc) == false)
1011 /* Find out if a global variable can be demoted to local scope.
1012 * Currently, this is valid for CUDA shared variables, which have local
1013 * scope and global lifetime. So the conditions to check are :
1014 * 1. Is the global variable in shared address space?
1015 * 2. Does it have internal linkage?
1016 * 3. Is the global variable referenced only in one function?
1018 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
1019 if (gv->hasInternalLinkage() == false)
1021 const PointerType *Pty = gv->getType();
1022 if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
1025 const Function *oneFunc = nullptr;
1027 bool flag = usedInOneFunc(gv, oneFunc);
1036 static bool useFuncSeen(const Constant *C,
1037 llvm::DenseMap<const Function *, bool> &seenMap) {
1038 for (const User *U : C->users()) {
1039 if (const Constant *cu = dyn_cast<Constant>(U)) {
1040 if (useFuncSeen(cu, seenMap))
1042 } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
1043 const BasicBlock *bb = I->getParent();
1046 const Function *caller = bb->getParent();
1049 if (seenMap.find(caller) != seenMap.end())
1056 void NVPTXAsmPrinter::emitDeclarations(const Module &M, raw_ostream &O) {
1057 llvm::DenseMap<const Function *, bool> seenMap;
1058 for (Module::const_iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI) {
1059 const Function *F = FI;
1061 if (F->isDeclaration()) {
1064 if (F->getIntrinsicID())
1066 emitDeclaration(F, O);
1069 for (const User *U : F->users()) {
1070 if (const Constant *C = dyn_cast<Constant>(U)) {
1071 if (usedInGlobalVarDef(C)) {
1072 // The use is in the initialization of a global variable
1073 // that is a function pointer, so print a declaration
1074 // for the original function
1075 emitDeclaration(F, O);
1078 // Emit a declaration of this function if the function that
1079 // uses this constant expr has already been seen.
1080 if (useFuncSeen(C, seenMap)) {
1081 emitDeclaration(F, O);
1086 if (!isa<Instruction>(U))
1088 const Instruction *instr = cast<Instruction>(U);
1089 const BasicBlock *bb = instr->getParent();
1092 const Function *caller = bb->getParent();
1096 // If a caller has already been seen, then the caller is
1097 // appearing in the module before the callee. so print out
1098 // a declaration for the callee.
1099 if (seenMap.find(caller) != seenMap.end()) {
1100 emitDeclaration(F, O);
1108 void NVPTXAsmPrinter::recordAndEmitFilenames(Module &M) {
1109 DebugInfoFinder DbgFinder;
1110 DbgFinder.processModule(M);
1113 for (DICompileUnit DIUnit : DbgFinder.compile_units()) {
1114 StringRef Filename(DIUnit.getFilename());
1115 StringRef Dirname(DIUnit.getDirectory());
1116 SmallString<128> FullPathName = Dirname;
1117 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
1118 sys::path::append(FullPathName, Filename);
1119 Filename = FullPathName.str();
1121 if (filenameMap.find(Filename.str()) != filenameMap.end())
1123 filenameMap[Filename.str()] = i;
1124 OutStreamer.EmitDwarfFileDirective(i, "", Filename.str());
1128 for (DISubprogram SP : DbgFinder.subprograms()) {
1129 StringRef Filename(SP.getFilename());
1130 StringRef Dirname(SP.getDirectory());
1131 SmallString<128> FullPathName = Dirname;
1132 if (!Dirname.empty() && !sys::path::is_absolute(Filename)) {
1133 sys::path::append(FullPathName, Filename);
1134 Filename = FullPathName.str();
1136 if (filenameMap.find(Filename.str()) != filenameMap.end())
1138 filenameMap[Filename.str()] = i;
1143 bool NVPTXAsmPrinter::doInitialization(Module &M) {
1145 SmallString<128> Str1;
1146 raw_svector_ostream OS1(Str1);
1148 MMI = getAnalysisIfAvailable<MachineModuleInfo>();
1149 MMI->AnalyzeModule(M);
1151 // We need to call the parent's one explicitly.
1152 //bool Result = AsmPrinter::doInitialization(M);
1154 // Initialize TargetLoweringObjectFile.
1155 const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
1156 .Initialize(OutContext, TM);
1158 Mang = new Mangler(TM.getDataLayout());
1160 // Emit header before any dwarf directives are emitted below.
1162 OutStreamer.EmitRawText(OS1.str());
1164 // Already commented out
1165 //bool Result = AsmPrinter::doInitialization(M);
1167 // Emit module-level inline asm if it exists.
1168 if (!M.getModuleInlineAsm().empty()) {
1169 OutStreamer.AddComment("Start of file scope inline assembly");
1170 OutStreamer.AddBlankLine();
1171 OutStreamer.EmitRawText(StringRef(M.getModuleInlineAsm()));
1172 OutStreamer.AddBlankLine();
1173 OutStreamer.AddComment("End of file scope inline assembly");
1174 OutStreamer.AddBlankLine();
1177 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
1178 recordAndEmitFilenames(M);
1180 GlobalsEmitted = false;
1182 return false; // success
1185 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
1186 SmallString<128> Str2;
1187 raw_svector_ostream OS2(Str2);
1189 emitDeclarations(M, OS2);
1191 // As ptxas does not support forward references of globals, we need to first
1192 // sort the list of module-level globals in def-use order. We visit each
1193 // global variable in order, and ensure that we emit it *after* its dependent
1194 // globals. We use a little extra memory maintaining both a set and a list to
1195 // have fast searches while maintaining a strict ordering.
1196 SmallVector<const GlobalVariable *, 8> Globals;
1197 DenseSet<const GlobalVariable *> GVVisited;
1198 DenseSet<const GlobalVariable *> GVVisiting;
1200 // Visit each global variable, in order
1201 for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
1203 VisitGlobalVariableForEmission(I, Globals, GVVisited, GVVisiting);
1205 assert(GVVisited.size() == M.getGlobalList().size() &&
1206 "Missed a global variable");
1207 assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
1209 // Print out module-level global variables in proper order
1210 for (unsigned i = 0, e = Globals.size(); i != e; ++i)
1211 printModuleLevelGV(Globals[i], OS2);
1215 OutStreamer.EmitRawText(OS2.str());
1218 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O) {
1220 O << "// Generated by LLVM NVPTX Back-End\n";
1224 unsigned PTXVersion = nvptxSubtarget.getPTXVersion();
1225 O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
1228 O << nvptxSubtarget.getTargetName();
1230 if (nvptxSubtarget.getDrvInterface() == NVPTX::NVCL)
1231 O << ", texmode_independent";
1232 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
1233 if (!nvptxSubtarget.hasDouble())
1234 O << ", map_f64_to_f32";
1237 if (MAI->doesSupportDebugInformation())
1242 O << ".address_size ";
1243 if (nvptxSubtarget.is64Bit())
1252 bool NVPTXAsmPrinter::doFinalization(Module &M) {
1254 // If we did not emit any functions, then the global declarations have not
1255 // yet been emitted.
1256 if (!GlobalsEmitted) {
1258 GlobalsEmitted = true;
1261 // XXX Temproarily remove global variables so that doFinalization() will not
1262 // emit them again (global variables are emitted at beginning).
1264 Module::GlobalListType &global_list = M.getGlobalList();
1265 int i, n = global_list.size();
1266 GlobalVariable **gv_array = new GlobalVariable *[n];
1268 // first, back-up GlobalVariable in gv_array
1270 for (Module::global_iterator I = global_list.begin(), E = global_list.end();
1272 gv_array[i++] = &*I;
1274 // second, empty global_list
1275 while (!global_list.empty())
1276 global_list.remove(global_list.begin());
1278 // call doFinalization
1279 bool ret = AsmPrinter::doFinalization(M);
1281 // now we restore global variables
1282 for (i = 0; i < n; i++)
1283 global_list.insert(global_list.end(), gv_array[i]);
1285 clearAnnotationCache(&M);
1290 //bool Result = AsmPrinter::doFinalization(M);
1291 // Instead of calling the parents doFinalization, we may
1292 // clone parents doFinalization and customize here.
1293 // Currently, we if NVISA out the EmitGlobals() in
1294 // parent's doFinalization, which is too intrusive.
1296 // Same for the doInitialization.
1300 // This function emits appropriate linkage directives for
1301 // functions and global variables.
1303 // extern function declaration -> .extern
1304 // extern function definition -> .visible
1305 // external global variable with init -> .visible
1306 // external without init -> .extern
1307 // appending -> not allowed, assert.
1308 // for any linkage other than
1309 // internal, private, linker_private,
1310 // linker_private_weak, linker_private_weak_def_auto,
1311 // we emit -> .weak.
1313 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
1315 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
1316 if (V->hasExternalLinkage()) {
1317 if (isa<GlobalVariable>(V)) {
1318 const GlobalVariable *GVar = cast<GlobalVariable>(V);
1320 if (GVar->hasInitializer())
1325 } else if (V->isDeclaration())
1329 } else if (V->hasAppendingLinkage()) {
1331 msg.append("Error: ");
1332 msg.append("Symbol ");
1334 msg.append(V->getName().str());
1335 msg.append("has unsupported appending linkage type");
1336 llvm_unreachable(msg.c_str());
1337 } else if (!V->hasInternalLinkage() &&
1338 !V->hasPrivateLinkage()) {
1344 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
1346 bool processDemoted) {
1349 if (GVar->hasSection()) {
1350 if (GVar->getSection() == StringRef("llvm.metadata"))
1354 const DataLayout *TD = TM.getDataLayout();
1356 // GlobalVariables are always constant pointers themselves.
1357 const PointerType *PTy = GVar->getType();
1358 Type *ETy = PTy->getElementType();
1360 if (GVar->hasExternalLinkage()) {
1361 if (GVar->hasInitializer())
1367 if (llvm::isTexture(*GVar)) {
1368 O << ".global .texref " << llvm::getTextureName(*GVar) << ";\n";
1372 if (llvm::isSurface(*GVar)) {
1373 O << ".global .surfref " << llvm::getSurfaceName(*GVar) << ";\n";
1377 if (GVar->isDeclaration()) {
1378 // (extern) declarations, no definition or initializer
1379 // Currently the only known declaration is for an automatic __local
1380 // (.shared) promoted to global.
1381 emitPTXGlobalVariable(GVar, O);
1386 if (llvm::isSampler(*GVar)) {
1387 O << ".global .samplerref " << llvm::getSamplerName(*GVar);
1389 const Constant *Initializer = nullptr;
1390 if (GVar->hasInitializer())
1391 Initializer = GVar->getInitializer();
1392 const ConstantInt *CI = nullptr;
1394 CI = dyn_cast<ConstantInt>(Initializer);
1396 unsigned sample = CI->getZExtValue();
1401 addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1403 O << "addr_mode_" << i << " = ";
1409 O << "clamp_to_border";
1412 O << "clamp_to_edge";
1423 O << "filter_mode = ";
1424 switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1432 llvm_unreachable("Anisotropic filtering is not supported");
1437 if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1438 O << ", force_unnormalized_coords = 1";
1447 if (GVar->hasPrivateLinkage()) {
1449 if (!strncmp(GVar->getName().data(), "unrollpragma", 12))
1452 // FIXME - need better way (e.g. Metadata) to avoid generating this global
1453 if (!strncmp(GVar->getName().data(), "filename", 8))
1455 if (GVar->use_empty())
1459 const Function *demotedFunc = nullptr;
1460 if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
1461 O << "// " << GVar->getName().str() << " has been demoted\n";
1462 if (localDecls.find(demotedFunc) != localDecls.end())
1463 localDecls[demotedFunc].push_back(GVar);
1465 std::vector<const GlobalVariable *> temp;
1466 temp.push_back(GVar);
1467 localDecls[demotedFunc] = temp;
1473 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1474 if (GVar->getAlignment() == 0)
1475 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1477 O << " .align " << GVar->getAlignment();
1479 if (ETy->isSingleValueType()) {
1481 // Special case: ABI requires that we use .u8 for predicates
1482 if (ETy->isIntegerTy(1))
1485 O << getPTXFundamentalTypeStr(ETy, false);
1487 O << *getSymbol(GVar);
1489 // Ptx allows variable initilization only for constant and global state
1491 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1492 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1493 GVar->hasInitializer()) {
1494 const Constant *Initializer = GVar->getInitializer();
1495 if (!Initializer->isNullValue()) {
1497 printScalarConstant(Initializer, O);
1501 unsigned int ElementSize = 0;
1503 // Although PTX has direct support for struct type and array type and
1504 // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1505 // targets that support these high level field accesses. Structs, arrays
1506 // and vectors are lowered into arrays of bytes.
1507 switch (ETy->getTypeID()) {
1508 case Type::StructTyID:
1509 case Type::ArrayTyID:
1510 case Type::VectorTyID:
1511 ElementSize = TD->getTypeStoreSize(ETy);
1512 // Ptx allows variable initilization only for constant and
1513 // global state spaces.
1514 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1515 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1516 GVar->hasInitializer()) {
1517 const Constant *Initializer = GVar->getInitializer();
1518 if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1519 AggBuffer aggBuffer(ElementSize, O, *this);
1520 bufferAggregateConstant(Initializer, &aggBuffer);
1521 if (aggBuffer.numSymbols) {
1522 if (nvptxSubtarget.is64Bit()) {
1523 O << " .u64 " << *getSymbol(GVar) << "[";
1524 O << ElementSize / 8;
1526 O << " .u32 " << *getSymbol(GVar) << "[";
1527 O << ElementSize / 4;
1531 O << " .b8 " << *getSymbol(GVar) << "[";
1539 O << " .b8 " << *getSymbol(GVar);
1547 O << " .b8 " << *getSymbol(GVar);
1556 llvm_unreachable("type not supported yet");
1563 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1564 if (localDecls.find(f) == localDecls.end())
1567 std::vector<const GlobalVariable *> &gvars = localDecls[f];
1569 for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1570 O << "\t// demoted variable\n\t";
1571 printModuleLevelGV(gvars[i], O, true);
1575 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1576 raw_ostream &O) const {
1577 switch (AddressSpace) {
1578 case llvm::ADDRESS_SPACE_LOCAL:
1581 case llvm::ADDRESS_SPACE_GLOBAL:
1584 case llvm::ADDRESS_SPACE_CONST:
1587 case llvm::ADDRESS_SPACE_SHARED:
1591 report_fatal_error("Bad address space found while emitting PTX");
1597 NVPTXAsmPrinter::getPTXFundamentalTypeStr(const Type *Ty, bool useB4PTR) const {
1598 switch (Ty->getTypeID()) {
1600 llvm_unreachable("unexpected type");
1602 case Type::IntegerTyID: {
1603 unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1606 else if (NumBits <= 64) {
1607 std::string name = "u";
1608 return name + utostr(NumBits);
1610 llvm_unreachable("Integer too large");
1615 case Type::FloatTyID:
1617 case Type::DoubleTyID:
1619 case Type::PointerTyID:
1620 if (nvptxSubtarget.is64Bit())
1630 llvm_unreachable("unexpected type");
1634 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1637 const DataLayout *TD = TM.getDataLayout();
1639 // GlobalVariables are always constant pointers themselves.
1640 const PointerType *PTy = GVar->getType();
1641 Type *ETy = PTy->getElementType();
1644 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1645 if (GVar->getAlignment() == 0)
1646 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1648 O << " .align " << GVar->getAlignment();
1650 if (ETy->isSingleValueType()) {
1652 O << getPTXFundamentalTypeStr(ETy);
1654 O << *getSymbol(GVar);
1658 int64_t ElementSize = 0;
1660 // Although PTX has direct support for struct type and array type and LLVM IR
1661 // is very similar to PTX, the LLVM CodeGen does not support for targets that
1662 // support these high level field accesses. Structs and arrays are lowered
1663 // into arrays of bytes.
1664 switch (ETy->getTypeID()) {
1665 case Type::StructTyID:
1666 case Type::ArrayTyID:
1667 case Type::VectorTyID:
1668 ElementSize = TD->getTypeStoreSize(ETy);
1669 O << " .b8 " << *getSymbol(GVar) << "[";
1671 O << itostr(ElementSize);
1676 llvm_unreachable("type not supported yet");
1681 static unsigned int getOpenCLAlignment(const DataLayout *TD, Type *Ty) {
1682 if (Ty->isSingleValueType())
1683 return TD->getPrefTypeAlignment(Ty);
1685 const ArrayType *ATy = dyn_cast<ArrayType>(Ty);
1687 return getOpenCLAlignment(TD, ATy->getElementType());
1689 const VectorType *VTy = dyn_cast<VectorType>(Ty);
1691 Type *ETy = VTy->getElementType();
1692 unsigned int numE = VTy->getNumElements();
1693 unsigned int alignE = TD->getPrefTypeAlignment(ETy);
1697 return numE * alignE;
1700 const StructType *STy = dyn_cast<StructType>(Ty);
1702 unsigned int alignStruct = 1;
1703 // Go through each element of the struct and find the
1704 // largest alignment.
1705 for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1706 Type *ETy = STy->getElementType(i);
1707 unsigned int align = getOpenCLAlignment(TD, ETy);
1708 if (align > alignStruct)
1709 alignStruct = align;
1714 const FunctionType *FTy = dyn_cast<FunctionType>(Ty);
1716 return TD->getPointerPrefAlignment();
1717 return TD->getPrefTypeAlignment(Ty);
1720 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1721 int paramIndex, raw_ostream &O) {
1722 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1723 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA))
1724 O << *getSymbol(I->getParent()) << "_param_" << paramIndex;
1726 std::string argName = I->getName();
1727 const char *p = argName.c_str();
1738 void NVPTXAsmPrinter::printParamName(int paramIndex, raw_ostream &O) {
1739 Function::const_arg_iterator I, E;
1742 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1743 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)) {
1744 O << *CurrentFnSym << "_param_" << paramIndex;
1748 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, i++) {
1749 if (i == paramIndex) {
1750 printParamName(I, paramIndex, O);
1754 llvm_unreachable("paramIndex out of bound");
1757 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1758 const DataLayout *TD = TM.getDataLayout();
1759 const AttributeSet &PAL = F->getAttributes();
1760 const TargetLowering *TLI = TM.getTargetLowering();
1761 Function::const_arg_iterator I, E;
1762 unsigned paramIndex = 0;
1764 bool isKernelFunc = llvm::isKernelFunction(*F);
1765 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
1766 MVT thePointerTy = TLI->getPointerTy();
1770 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1771 Type *Ty = I->getType();
1778 // Handle image/sampler parameters
1779 if (isKernelFunction(*F)) {
1780 if (isSampler(*I) || isImage(*I)) {
1782 std::string sname = I->getName();
1783 if (isImageWriteOnly(*I) || isImageReadWrite(*I)) {
1784 if (nvptxSubtarget.hasImageHandles())
1785 O << "\t.param .u64 .ptr .surfref ";
1787 O << "\t.param .surfref ";
1788 O << *CurrentFnSym << "_param_" << paramIndex;
1790 else { // Default image is read_only
1791 if (nvptxSubtarget.hasImageHandles())
1792 O << "\t.param .u64 .ptr .texref ";
1794 O << "\t.param .texref ";
1795 O << *CurrentFnSym << "_param_" << paramIndex;
1798 if (nvptxSubtarget.hasImageHandles())
1799 O << "\t.param .u64 .ptr .samplerref ";
1801 O << "\t.param .samplerref ";
1802 O << *CurrentFnSym << "_param_" << paramIndex;
1808 if (PAL.hasAttribute(paramIndex + 1, Attribute::ByVal) == false) {
1809 if (Ty->isAggregateType() || Ty->isVectorTy()) {
1810 // Just print .param .align <a> .b8 .param[size];
1811 // <a> = PAL.getparamalignment
1812 // size = typeallocsize of element type
1813 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1815 align = TD->getABITypeAlignment(Ty);
1817 unsigned sz = TD->getTypeAllocSize(Ty);
1818 O << "\t.param .align " << align << " .b8 ";
1819 printParamName(I, paramIndex, O);
1820 O << "[" << sz << "]";
1825 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1828 // Special handling for pointer arguments to kernel
1829 O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1831 if (nvptxSubtarget.getDrvInterface() != NVPTX::CUDA) {
1832 Type *ETy = PTy->getElementType();
1833 int addrSpace = PTy->getAddressSpace();
1834 switch (addrSpace) {
1838 case llvm::ADDRESS_SPACE_CONST:
1839 O << ".ptr .const ";
1841 case llvm::ADDRESS_SPACE_SHARED:
1842 O << ".ptr .shared ";
1844 case llvm::ADDRESS_SPACE_GLOBAL:
1845 O << ".ptr .global ";
1848 O << ".align " << (int) getOpenCLAlignment(TD, ETy) << " ";
1850 printParamName(I, paramIndex, O);
1854 // non-pointer scalar to kernel func
1856 // Special case: predicate operands become .u8 types
1857 if (Ty->isIntegerTy(1))
1860 O << getPTXFundamentalTypeStr(Ty);
1862 printParamName(I, paramIndex, O);
1865 // Non-kernel function, just print .param .b<size> for ABI
1866 // and .reg .b<size> for non-ABI
1868 if (isa<IntegerType>(Ty)) {
1869 sz = cast<IntegerType>(Ty)->getBitWidth();
1872 } else if (isa<PointerType>(Ty))
1873 sz = thePointerTy.getSizeInBits();
1875 sz = Ty->getPrimitiveSizeInBits();
1877 O << "\t.param .b" << sz << " ";
1879 O << "\t.reg .b" << sz << " ";
1880 printParamName(I, paramIndex, O);
1884 // param has byVal attribute. So should be a pointer
1885 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1886 assert(PTy && "Param with byval attribute should be a pointer type");
1887 Type *ETy = PTy->getElementType();
1889 if (isABI || isKernelFunc) {
1890 // Just print .param .align <a> .b8 .param[size];
1891 // <a> = PAL.getparamalignment
1892 // size = typeallocsize of element type
1893 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1895 align = TD->getABITypeAlignment(ETy);
1897 unsigned sz = TD->getTypeAllocSize(ETy);
1898 O << "\t.param .align " << align << " .b8 ";
1899 printParamName(I, paramIndex, O);
1900 O << "[" << sz << "]";
1903 // Split the ETy into constituent parts and
1904 // print .param .b<size> <name> for each part.
1905 // Further, if a part is vector, print the above for
1906 // each vector element.
1907 SmallVector<EVT, 16> vtparts;
1908 ComputeValueVTs(*TLI, ETy, vtparts);
1909 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1911 EVT elemtype = vtparts[i];
1912 if (vtparts[i].isVector()) {
1913 elems = vtparts[i].getVectorNumElements();
1914 elemtype = vtparts[i].getVectorElementType();
1917 for (unsigned j = 0, je = elems; j != je; ++j) {
1918 unsigned sz = elemtype.getSizeInBits();
1919 if (elemtype.isInteger() && (sz < 32))
1921 O << "\t.reg .b" << sz << " ";
1922 printParamName(I, paramIndex, O);
1938 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1940 const Function *F = MF.getFunction();
1941 emitFunctionParamList(F, O);
1944 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1945 const MachineFunction &MF) {
1946 SmallString<128> Str;
1947 raw_svector_ostream O(Str);
1949 // Map the global virtual register number to a register class specific
1950 // virtual register number starting from 1 with that class.
1951 const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo();
1952 //unsigned numRegClasses = TRI->getNumRegClasses();
1954 // Emit the Fake Stack Object
1955 const MachineFrameInfo *MFI = MF.getFrameInfo();
1956 int NumBytes = (int) MFI->getStackSize();
1958 O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
1959 << getFunctionNumber() << "[" << NumBytes << "];\n";
1960 if (nvptxSubtarget.is64Bit()) {
1961 O << "\t.reg .b64 \t%SP;\n";
1962 O << "\t.reg .b64 \t%SPL;\n";
1964 O << "\t.reg .b32 \t%SP;\n";
1965 O << "\t.reg .b32 \t%SPL;\n";
1969 // Go through all virtual registers to establish the mapping between the
1971 // register number and the per class virtual register number.
1972 // We use the per class virtual register number in the ptx output.
1973 unsigned int numVRs = MRI->getNumVirtRegs();
1974 for (unsigned i = 0; i < numVRs; i++) {
1975 unsigned int vr = TRI->index2VirtReg(i);
1976 const TargetRegisterClass *RC = MRI->getRegClass(vr);
1977 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1978 int n = regmap.size();
1979 regmap.insert(std::make_pair(vr, n + 1));
1982 // Emit register declarations
1983 // @TODO: Extract out the real register usage
1984 // O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
1985 // O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
1986 // O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
1987 // O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
1988 // O << "\t.reg .s64 %rl<" << NVPTXNumRegisters << ">;\n";
1989 // O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
1990 // O << "\t.reg .f64 %fl<" << NVPTXNumRegisters << ">;\n";
1992 // Emit declaration of the virtual registers or 'physical' registers for
1993 // each register class
1994 for (unsigned i=0; i< TRI->getNumRegClasses(); i++) {
1995 const TargetRegisterClass *RC = TRI->getRegClass(i);
1996 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
1997 std::string rcname = getNVPTXRegClassName(RC);
1998 std::string rcStr = getNVPTXRegClassStr(RC);
1999 int n = regmap.size();
2001 // Only declare those registers that may be used.
2003 O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
2008 OutStreamer.EmitRawText(O.str());
2011 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
2012 APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
2014 unsigned int numHex;
2017 if (Fp->getType()->getTypeID() == Type::FloatTyID) {
2020 APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &ignored);
2021 } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
2024 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &ignored);
2026 llvm_unreachable("unsupported fp type");
2028 APInt API = APF.bitcastToAPInt();
2029 std::string hexstr(utohexstr(API.getZExtValue()));
2031 if (hexstr.length() < numHex)
2032 O << std::string(numHex - hexstr.length(), '0');
2033 O << utohexstr(API.getZExtValue());
2036 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
2037 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
2038 O << CI->getValue();
2041 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
2042 printFPConstant(CFP, O);
2045 if (isa<ConstantPointerNull>(CPV)) {
2049 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
2050 PointerType *PTy = dyn_cast<PointerType>(GVar->getType());
2051 bool IsNonGenericPointer = false;
2052 if (PTy && PTy->getAddressSpace() != 0) {
2053 IsNonGenericPointer = true;
2055 if (EmitGeneric && !isa<Function>(CPV) && !IsNonGenericPointer) {
2057 O << *getSymbol(GVar);
2060 O << *getSymbol(GVar);
2064 if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
2065 const Value *v = Cexpr->stripPointerCasts();
2066 PointerType *PTy = dyn_cast<PointerType>(Cexpr->getType());
2067 bool IsNonGenericPointer = false;
2068 if (PTy && PTy->getAddressSpace() != 0) {
2069 IsNonGenericPointer = true;
2071 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
2072 if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
2074 O << *getSymbol(GVar);
2077 O << *getSymbol(GVar);
2081 O << *LowerConstant(CPV, *this);
2085 llvm_unreachable("Not scalar type found in printScalarConstant()");
2088 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
2089 AggBuffer *aggBuffer) {
2091 const DataLayout *TD = TM.getDataLayout();
2093 if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
2094 int s = TD->getTypeAllocSize(CPV->getType());
2097 aggBuffer->addZeros(s);
2102 switch (CPV->getType()->getTypeID()) {
2104 case Type::IntegerTyID: {
2105 const Type *ETy = CPV->getType();
2106 if (ETy == Type::getInt8Ty(CPV->getContext())) {
2108 (unsigned char)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
2110 aggBuffer->addBytes(ptr, 1, Bytes);
2111 } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
2112 short int16 = (short)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
2113 ptr = (unsigned char *)&int16;
2114 aggBuffer->addBytes(ptr, 2, Bytes);
2115 } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
2116 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
2117 int int32 = (int)(constInt->getZExtValue());
2118 ptr = (unsigned char *)&int32;
2119 aggBuffer->addBytes(ptr, 4, Bytes);
2121 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
2122 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
2123 ConstantFoldConstantExpression(Cexpr, TD))) {
2124 int int32 = (int)(constInt->getZExtValue());
2125 ptr = (unsigned char *)&int32;
2126 aggBuffer->addBytes(ptr, 4, Bytes);
2129 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
2130 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
2131 aggBuffer->addSymbol(v);
2132 aggBuffer->addZeros(4);
2136 llvm_unreachable("unsupported integer const type");
2137 } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
2138 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
2139 long long int64 = (long long)(constInt->getZExtValue());
2140 ptr = (unsigned char *)&int64;
2141 aggBuffer->addBytes(ptr, 8, Bytes);
2143 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
2144 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
2145 ConstantFoldConstantExpression(Cexpr, TD))) {
2146 long long int64 = (long long)(constInt->getZExtValue());
2147 ptr = (unsigned char *)&int64;
2148 aggBuffer->addBytes(ptr, 8, Bytes);
2151 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
2152 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
2153 aggBuffer->addSymbol(v);
2154 aggBuffer->addZeros(8);
2158 llvm_unreachable("unsupported integer const type");
2160 llvm_unreachable("unsupported integer const type");
2163 case Type::FloatTyID:
2164 case Type::DoubleTyID: {
2165 const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
2166 const Type *Ty = CFP->getType();
2167 if (Ty == Type::getFloatTy(CPV->getContext())) {
2168 float float32 = (float) CFP->getValueAPF().convertToFloat();
2169 ptr = (unsigned char *)&float32;
2170 aggBuffer->addBytes(ptr, 4, Bytes);
2171 } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
2172 double float64 = CFP->getValueAPF().convertToDouble();
2173 ptr = (unsigned char *)&float64;
2174 aggBuffer->addBytes(ptr, 8, Bytes);
2176 llvm_unreachable("unsupported fp const type");
2180 case Type::PointerTyID: {
2181 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
2182 aggBuffer->addSymbol(GVar);
2183 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
2184 const Value *v = Cexpr->stripPointerCasts();
2185 aggBuffer->addSymbol(v);
2187 unsigned int s = TD->getTypeAllocSize(CPV->getType());
2188 aggBuffer->addZeros(s);
2192 case Type::ArrayTyID:
2193 case Type::VectorTyID:
2194 case Type::StructTyID: {
2195 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV) ||
2196 isa<ConstantStruct>(CPV) || isa<ConstantDataSequential>(CPV)) {
2197 int ElementSize = TD->getTypeAllocSize(CPV->getType());
2198 bufferAggregateConstant(CPV, aggBuffer);
2199 if (Bytes > ElementSize)
2200 aggBuffer->addZeros(Bytes - ElementSize);
2201 } else if (isa<ConstantAggregateZero>(CPV))
2202 aggBuffer->addZeros(Bytes);
2204 llvm_unreachable("Unexpected Constant type");
2209 llvm_unreachable("unsupported type");
2213 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
2214 AggBuffer *aggBuffer) {
2215 const DataLayout *TD = TM.getDataLayout();
2219 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
2220 if (CPV->getNumOperands())
2221 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
2222 bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
2226 if (const ConstantDataSequential *CDS =
2227 dyn_cast<ConstantDataSequential>(CPV)) {
2228 if (CDS->getNumElements())
2229 for (unsigned i = 0; i < CDS->getNumElements(); ++i)
2230 bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
2235 if (isa<ConstantStruct>(CPV)) {
2236 if (CPV->getNumOperands()) {
2237 StructType *ST = cast<StructType>(CPV->getType());
2238 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
2240 Bytes = TD->getStructLayout(ST)->getElementOffset(0) +
2241 TD->getTypeAllocSize(ST) -
2242 TD->getStructLayout(ST)->getElementOffset(i);
2244 Bytes = TD->getStructLayout(ST)->getElementOffset(i + 1) -
2245 TD->getStructLayout(ST)->getElementOffset(i);
2246 bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
2251 llvm_unreachable("unsupported constant type in printAggregateConstant()");
2254 // buildTypeNameMap - Run through symbol table looking for type names.
2257 bool NVPTXAsmPrinter::isImageType(const Type *Ty) {
2259 std::map<const Type *, std::string>::iterator PI = TypeNameMap.find(Ty);
2261 if (PI != TypeNameMap.end() && (!PI->second.compare("struct._image1d_t") ||
2262 !PI->second.compare("struct._image2d_t") ||
2263 !PI->second.compare("struct._image3d_t")))
2270 bool NVPTXAsmPrinter::ignoreLoc(const MachineInstr &MI) {
2271 switch (MI.getOpcode()) {
2274 case NVPTX::CallArgBeginInst:
2275 case NVPTX::CallArgEndInst0:
2276 case NVPTX::CallArgEndInst1:
2277 case NVPTX::CallArgF32:
2278 case NVPTX::CallArgF64:
2279 case NVPTX::CallArgI16:
2280 case NVPTX::CallArgI32:
2281 case NVPTX::CallArgI32imm:
2282 case NVPTX::CallArgI64:
2283 case NVPTX::CallArgParam:
2284 case NVPTX::CallVoidInst:
2285 case NVPTX::CallVoidInstReg:
2286 case NVPTX::Callseq_End:
2287 case NVPTX::CallVoidInstReg64:
2288 case NVPTX::DeclareParamInst:
2289 case NVPTX::DeclareRetMemInst:
2290 case NVPTX::DeclareRetRegInst:
2291 case NVPTX::DeclareRetScalarInst:
2292 case NVPTX::DeclareScalarParamInst:
2293 case NVPTX::DeclareScalarRegInst:
2294 case NVPTX::StoreParamF32:
2295 case NVPTX::StoreParamF64:
2296 case NVPTX::StoreParamI16:
2297 case NVPTX::StoreParamI32:
2298 case NVPTX::StoreParamI64:
2299 case NVPTX::StoreParamI8:
2300 case NVPTX::StoreRetvalF32:
2301 case NVPTX::StoreRetvalF64:
2302 case NVPTX::StoreRetvalI16:
2303 case NVPTX::StoreRetvalI32:
2304 case NVPTX::StoreRetvalI64:
2305 case NVPTX::StoreRetvalI8:
2306 case NVPTX::LastCallArgF32:
2307 case NVPTX::LastCallArgF64:
2308 case NVPTX::LastCallArgI16:
2309 case NVPTX::LastCallArgI32:
2310 case NVPTX::LastCallArgI32imm:
2311 case NVPTX::LastCallArgI64:
2312 case NVPTX::LastCallArgParam:
2313 case NVPTX::LoadParamMemF32:
2314 case NVPTX::LoadParamMemF64:
2315 case NVPTX::LoadParamMemI16:
2316 case NVPTX::LoadParamMemI32:
2317 case NVPTX::LoadParamMemI64:
2318 case NVPTX::LoadParamMemI8:
2319 case NVPTX::PrototypeInst:
2320 case NVPTX::DBG_VALUE:
2326 /// PrintAsmOperand - Print out an operand for an inline asm expression.
2328 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
2329 unsigned AsmVariant,
2330 const char *ExtraCode, raw_ostream &O) {
2331 if (ExtraCode && ExtraCode[0]) {
2332 if (ExtraCode[1] != 0)
2333 return true; // Unknown modifier.
2335 switch (ExtraCode[0]) {
2337 // See if this is a generic print operand
2338 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
2344 printOperand(MI, OpNo, O);
2349 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(
2350 const MachineInstr *MI, unsigned OpNo, unsigned AsmVariant,
2351 const char *ExtraCode, raw_ostream &O) {
2352 if (ExtraCode && ExtraCode[0])
2353 return true; // Unknown modifier
2356 printMemOperand(MI, OpNo, O);
2362 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
2363 raw_ostream &O, const char *Modifier) {
2364 const MachineOperand &MO = MI->getOperand(opNum);
2365 switch (MO.getType()) {
2366 case MachineOperand::MO_Register:
2367 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
2368 if (MO.getReg() == NVPTX::VRDepot)
2369 O << DEPOTNAME << getFunctionNumber();
2371 O << NVPTXInstPrinter::getRegisterName(MO.getReg());
2373 emitVirtualRegister(MO.getReg(), O);
2377 case MachineOperand::MO_Immediate:
2380 else if (strstr(Modifier, "vec") == Modifier)
2381 printVecModifiedImmediate(MO, Modifier, O);
2384 "Don't know how to handle modifier on immediate operand");
2387 case MachineOperand::MO_FPImmediate:
2388 printFPConstant(MO.getFPImm(), O);
2391 case MachineOperand::MO_GlobalAddress:
2392 O << *getSymbol(MO.getGlobal());
2395 case MachineOperand::MO_MachineBasicBlock:
2396 O << *MO.getMBB()->getSymbol();
2400 llvm_unreachable("Operand type not supported.");
2404 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
2405 raw_ostream &O, const char *Modifier) {
2406 printOperand(MI, opNum, O);
2408 if (Modifier && !strcmp(Modifier, "add")) {
2410 printOperand(MI, opNum + 1, O);
2412 if (MI->getOperand(opNum + 1).isImm() &&
2413 MI->getOperand(opNum + 1).getImm() == 0)
2414 return; // don't print ',0' or '+0'
2416 printOperand(MI, opNum + 1, O);
2421 // Force static initialization.
2422 extern "C" void LLVMInitializeNVPTXBackendAsmPrinter() {
2423 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2424 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2427 void NVPTXAsmPrinter::emitSrcInText(StringRef filename, unsigned line) {
2428 std::stringstream temp;
2429 LineReader *reader = this->getReader(filename.str());
2431 temp << filename.str();
2435 temp << reader->readLine(line);
2437 this->OutStreamer.EmitRawText(Twine(temp.str()));
2440 LineReader *NVPTXAsmPrinter::getReader(std::string filename) {
2442 reader = new LineReader(filename);
2445 if (reader->fileName() != filename) {
2447 reader = new LineReader(filename);
2453 std::string LineReader::readLine(unsigned lineNum) {
2454 if (lineNum < theCurLine) {
2456 fstr.seekg(0, std::ios::beg);
2458 while (theCurLine < lineNum) {
2459 fstr.getline(buff, 500);
2465 // Force static initialization.
2466 extern "C" void LLVMInitializeNVPTXAsmPrinter() {
2467 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2468 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);