1 //===-- NVPTXAsmPrinter.cpp - NVPTX LLVM assembly writer ------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains a printer that converts from our internal representation
11 // of machine-dependent LLVM code to NVPTX assembly language.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "NVPTXAsmPrinter.h"
16 #include "InstPrinter/NVPTXInstPrinter.h"
17 #include "MCTargetDesc/NVPTXMCAsmInfo.h"
18 #include "NVPTX.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"
48 #include <sstream>
49 using namespace llvm;
50
51 #define DEPOTNAME "__local_depot"
52
53 static cl::opt<bool>
54 EmitLineNumbers("nvptx-emit-line-numbers", cl::Hidden,
55 cl::desc("NVPTX Specific: Emit Line numbers even without -G"),
56 cl::init(true));
57
58 static cl::opt<bool>
59 InterleaveSrc("nvptx-emit-src", cl::ZeroOrMore, cl::Hidden,
60 cl::desc("NVPTX Specific: Emit source line in ptx file"),
61 cl::init(false));
62
63 namespace {
64 /// DiscoverDependentGlobals - Return a set of GlobalVariables on which \p V
65 /// depends.
DiscoverDependentGlobals(const Value * V,DenseSet<const GlobalVariable * > & Globals)66 void DiscoverDependentGlobals(const Value *V,
67 DenseSet<const GlobalVariable *> &Globals) {
68 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
69 Globals.insert(GV);
70 else {
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);
74 }
75 }
76 }
77 }
78
79 /// VisitGlobalVariableForEmission - Add \p GV to the list of GlobalVariable
80 /// instances to be emitted, but only after any dependents have been added
81 /// first.
VisitGlobalVariableForEmission(const GlobalVariable * GV,SmallVectorImpl<const GlobalVariable * > & Order,DenseSet<const GlobalVariable * > & Visited,DenseSet<const GlobalVariable * > & Visiting)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))
88 return;
89
90 // Do we have a circular dependency?
91 if (Visiting.count(GV))
92 report_fatal_error("Circular dependency found in global variable set");
93
94 // Start visiting this global
95 Visiting.insert(GV);
96
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);
101
102 for (DenseSet<const GlobalVariable *>::iterator I = Others.begin(),
103 E = Others.end();
104 I != E; ++I)
105 VisitGlobalVariableForEmission(*I, Order, Visited, Visiting);
106
107 // Now we can visit ourself
108 Order.push_back(GV);
109 Visited.insert(GV);
110 Visiting.erase(GV);
111 }
112 }
113
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.
117 ///
118 using namespace nvptx;
LowerConstant(const Constant * CV,AsmPrinter & AP)119 const MCExpr *nvptx::LowerConstant(const Constant *CV, AsmPrinter &AP) {
120 MCContext &Ctx = AP.OutContext;
121
122 if (CV->isNullValue() || isa<UndefValue>(CV))
123 return MCConstantExpr::Create(0, Ctx);
124
125 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV))
126 return MCConstantExpr::Create(CI->getZExtValue(), Ctx);
127
128 if (const GlobalValue *GV = dyn_cast<GlobalValue>(CV))
129 return MCSymbolRefExpr::Create(AP.getSymbol(GV), Ctx);
130
131 if (const BlockAddress *BA = dyn_cast<BlockAddress>(CV))
132 return MCSymbolRefExpr::Create(AP.GetBlockAddressSymbol(BA), Ctx);
133
134 const ConstantExpr *CE = dyn_cast<ConstantExpr>(CV);
135 if (!CE)
136 llvm_unreachable("Unknown constant value to lower!");
137
138 switch (CE->getOpcode()) {
139 default:
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()))
144 if (C != CE)
145 return LowerConstant(C, AP);
146
147 // Otherwise report the problem to the user.
148 {
149 std::string S;
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());
155 }
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);
163 }
164 std::string S;
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());
170 }
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);
176
177 const MCExpr *Base = LowerConstant(CE->getOperand(0), AP);
178 if (!OffsetAI)
179 return Base;
180
181 int64_t Offset = OffsetAI.getSExtValue();
182 return MCBinaryExpr::CreateAdd(Base, MCConstantExpr::Create(Offset, Ctx),
183 Ctx);
184 }
185
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.
191 // FALL THROUGH.
192 case Instruction::BitCast:
193 return LowerConstant(CE->getOperand(0), AP);
194
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()),
201 false /*ZExt*/);
202 return LowerConstant(Op, AP);
203 }
204
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();
211
212 const MCExpr *OpExpr = LowerConstant(Op, AP);
213
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()))
217 return OpExpr;
218
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
221 // a constant expr.
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);
226 }
227
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()) {
242 default:
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);
262 }
263 }
264 }
265 }
266
emitLineNumberAsDotLoc(const MachineInstr & MI)267 void NVPTXAsmPrinter::emitLineNumberAsDotLoc(const MachineInstr &MI) {
268 if (!EmitLineNumbers)
269 return;
270 if (ignoreLoc(MI))
271 return;
272
273 DebugLoc curLoc = MI.getDebugLoc();
274
275 if (prevDebugLoc.isUnknown() && curLoc.isUnknown())
276 return;
277
278 if (prevDebugLoc == curLoc)
279 return;
280
281 prevDebugLoc = curLoc;
282
283 if (curLoc.isUnknown())
284 return;
285
286 const MachineFunction *MF = MI.getParent()->getParent();
287 //const TargetMachine &TM = MF->getTarget();
288
289 const LLVMContext &ctx = MF->getFunction()->getContext();
290 DIScope Scope(curLoc.getScope(ctx));
291
292 assert((!Scope || Scope.isScope()) &&
293 "Scope of a DebugLoc should be null or a DIScope.");
294 if (!Scope)
295 return;
296
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();
303 }
304
305 if (filenameMap.find(fileName.str()) == filenameMap.end())
306 return;
307
308 // Emit the line from the source file.
309 if (InterleaveSrc)
310 this->emitSrcInText(fileName.str(), curLoc.getLine());
311
312 std::stringstream temp;
313 temp << "\t.loc " << filenameMap[fileName.str()] << " " << curLoc.getLine()
314 << " " << curLoc.getCol();
315 OutStreamer.EmitRawText(Twine(temp.str().c_str()));
316 }
317
EmitInstruction(const MachineInstr * MI)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);
323
324 MCInst Inst;
325 lowerToMCInst(MI, Inst);
326 EmitToStreamer(OutStreamer, Inst);
327 }
328
329 // Handle symbol backtracking for targets that do not support image handles
lowerImageHandleOperand(const MachineInstr * MI,unsigned OpNo,MCOperand & MCOp)330 bool NVPTXAsmPrinter::lowerImageHandleOperand(const MachineInstr *MI,
331 unsigned OpNo, MCOperand &MCOp) {
332 const MachineOperand &MO = MI->getOperand(OpNo);
333
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:
376 {
377 // This is a texture fetch, so operand 4 is a texref and operand 5 is
378 // a samplerref
379 if (OpNo == 4) {
380 lowerImageHandleSymbol(MO.getImm(), MCOp);
381 return true;
382 }
383 if (OpNo == 5) {
384 lowerImageHandleSymbol(MO.getImm(), MCOp);
385 return true;
386 }
387
388 return false;
389 }
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
406 if (OpNo == 1) {
407 lowerImageHandleSymbol(MO.getImm(), MCOp);
408 return true;
409 }
410
411 return false;
412 }
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
429 if (OpNo == 2) {
430 lowerImageHandleSymbol(MO.getImm(), MCOp);
431 return true;
432 }
433
434 return false;
435 }
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
452 if (OpNo == 4) {
453 lowerImageHandleSymbol(MO.getImm(), MCOp);
454 return true;
455 }
456
457 return false;
458 }
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
550 if (OpNo == 0) {
551 lowerImageHandleSymbol(MO.getImm(), MCOp);
552 return true;
553 }
554
555 return false;
556 }
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
572 if (OpNo == 1) {
573 lowerImageHandleSymbol(MO.getImm(), MCOp);
574 return true;
575 }
576
577 return false;
578 }
579 }
580 }
581
lowerImageHandleSymbol(unsigned Index,MCOperand & MCOp)582 void NVPTXAsmPrinter::lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp) {
583 // Ewwww
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())));
592 }
593
lowerToMCInst(const MachineInstr * MI,MCInst & OutMI)594 void NVPTXAsmPrinter::lowerToMCInst(const MachineInstr *MI, MCInst &OutMI) {
595 OutMI.setOpcode(MI->getOpcode());
596 const NVPTXSubtarget &ST = TM.getSubtarget<NVPTXSubtarget>();
597
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()))));
603 return;
604 }
605
606 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
607 const MachineOperand &MO = MI->getOperand(i);
608
609 MCOperand MCOp;
610 if (!ST.hasImageHandles()) {
611 if (lowerImageHandleOperand(MI, i, MCOp)) {
612 OutMI.addOperand(MCOp);
613 continue;
614 }
615 }
616
617 if (lowerOperand(MO, MCOp))
618 OutMI.addOperand(MCOp);
619 }
620 }
621
lowerOperand(const MachineOperand & MO,MCOperand & MCOp)622 bool NVPTXAsmPrinter::lowerOperand(const MachineOperand &MO,
623 MCOperand &MCOp) {
624 switch (MO.getType()) {
625 default: llvm_unreachable("unknown operand type");
626 case MachineOperand::MO_Register:
627 MCOp = MCOperand::CreateReg(encodeVirtualRegister(MO.getReg()));
628 break;
629 case MachineOperand::MO_Immediate:
630 MCOp = MCOperand::CreateImm(MO.getImm());
631 break;
632 case MachineOperand::MO_MachineBasicBlock:
633 MCOp = MCOperand::CreateExpr(MCSymbolRefExpr::Create(
634 MO.getMBB()->getSymbol(), OutContext));
635 break;
636 case MachineOperand::MO_ExternalSymbol:
637 MCOp = GetSymbolRef(GetExternalSymbolSymbol(MO.getSymbolName()));
638 break;
639 case MachineOperand::MO_GlobalAddress:
640 MCOp = GetSymbolRef(getSymbol(MO.getGlobal()));
641 break;
642 case MachineOperand::MO_FPImmediate: {
643 const ConstantFP *Cnt = MO.getFPImm();
644 APFloat Val = Cnt->getValueAPF();
645
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));
651 break;
652 case Type::DoubleTyID:
653 MCOp = MCOperand::CreateExpr(
654 NVPTXFloatMCExpr::CreateConstantFPDouble(Val, OutContext));
655 break;
656 }
657 break;
658 }
659 }
660 return true;
661 }
662
encodeVirtualRegister(unsigned Reg)663 unsigned NVPTXAsmPrinter::encodeVirtualRegister(unsigned Reg) {
664 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
665 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
666
667 DenseMap<unsigned, unsigned> &RegMap = VRegMapping[RC];
668 unsigned RegNum = RegMap[Reg];
669
670 // Encode the register class in the upper 4 bits
671 // Must be kept in sync with NVPTXInstPrinter::printRegName
672 unsigned Ret = 0;
673 if (RC == &NVPTX::Int1RegsRegClass) {
674 Ret = (1 << 28);
675 } else if (RC == &NVPTX::Int16RegsRegClass) {
676 Ret = (2 << 28);
677 } else if (RC == &NVPTX::Int32RegsRegClass) {
678 Ret = (3 << 28);
679 } else if (RC == &NVPTX::Int64RegsRegClass) {
680 Ret = (4 << 28);
681 } else if (RC == &NVPTX::Float32RegsRegClass) {
682 Ret = (5 << 28);
683 } else if (RC == &NVPTX::Float64RegsRegClass) {
684 Ret = (6 << 28);
685 } else {
686 report_fatal_error("Bad register class");
687 }
688
689 // Insert the vreg number
690 Ret |= (RegNum & 0x0FFFFFFF);
691 return Ret;
692 } else {
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;
696 }
697 }
698
GetSymbolRef(const MCSymbol * Symbol)699 MCOperand NVPTXAsmPrinter::GetSymbolRef(const MCSymbol *Symbol) {
700 const MCExpr *Expr;
701 Expr = MCSymbolRefExpr::Create(Symbol, MCSymbolRefExpr::VK_None,
702 OutContext);
703 return MCOperand::CreateExpr(Expr);
704 }
705
printReturnValStr(const Function * F,raw_ostream & O)706 void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
707 const DataLayout *TD = TM.getDataLayout();
708 const TargetLowering *TLI = TM.getTargetLowering();
709
710 Type *Ty = F->getReturnType();
711
712 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
713
714 if (Ty->getTypeID() == Type::VoidTyID)
715 return;
716
717 O << " (";
718
719 if (isABI) {
720 if (Ty->isFloatingPointTy() || Ty->isIntegerTy()) {
721 unsigned size = 0;
722 if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty)) {
723 size = ITy->getBitWidth();
724 if (size < 32)
725 size = 32;
726 } else {
727 assert(Ty->isFloatingPointTy() && "Floating point type expected here");
728 size = Ty->getPrimitiveSizeInBits();
729 }
730
731 O << ".param .b" << size << " func_retval0";
732 } else if (isa<PointerType>(Ty)) {
733 O << ".param .b" << TLI->getPointerTy().getSizeInBits()
734 << " func_retval0";
735 } else {
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
742 << "]";
743 } else
744 assert(false && "Unknown return type");
745 }
746 } else {
747 SmallVector<EVT, 16> vtparts;
748 ComputeValueVTs(*TLI, Ty, vtparts);
749 unsigned idx = 0;
750 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
751 unsigned elems = 1;
752 EVT elemtype = vtparts[i];
753 if (vtparts[i].isVector()) {
754 elems = vtparts[i].getVectorNumElements();
755 elemtype = vtparts[i].getVectorElementType();
756 }
757
758 for (unsigned j = 0, je = elems; j != je; ++j) {
759 unsigned sz = elemtype.getSizeInBits();
760 if (elemtype.isInteger() && (sz < 32))
761 sz = 32;
762 O << ".reg .b" << sz << " func_retval" << idx;
763 if (j < je - 1)
764 O << ", ";
765 ++idx;
766 }
767 if (i < e - 1)
768 O << ", ";
769 }
770 }
771 O << ") ";
772 return;
773 }
774
printReturnValStr(const MachineFunction & MF,raw_ostream & O)775 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
776 raw_ostream &O) {
777 const Function *F = MF.getFunction();
778 printReturnValStr(F, O);
779 }
780
EmitFunctionEntryLabel()781 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
782 SmallString<128> Str;
783 raw_svector_ostream O(Str);
784
785 if (!GlobalsEmitted) {
786 emitGlobals(*MF->getFunction()->getParent());
787 GlobalsEmitted = true;
788 }
789
790 // Set up
791 MRI = &MF->getRegInfo();
792 F = MF->getFunction();
793 emitLinkageDirective(F, O);
794 if (llvm::isKernelFunction(*F))
795 O << ".entry ";
796 else {
797 O << ".func ";
798 printReturnValStr(*MF, O);
799 }
800
801 O << *CurrentFnSym;
802
803 emitFunctionParamList(*MF, O);
804
805 if (llvm::isKernelFunction(*F))
806 emitKernelFunctionDirectives(*F, O);
807
808 OutStreamer.EmitRawText(O.str());
809
810 prevDebugLoc = DebugLoc();
811 }
812
EmitFunctionBodyStart()813 void NVPTXAsmPrinter::EmitFunctionBodyStart() {
814 VRegMapping.clear();
815 OutStreamer.EmitRawText(StringRef("{\n"));
816 setAndEmitFunctionVirtualRegisters(*MF);
817
818 SmallString<128> Str;
819 raw_svector_ostream O(Str);
820 emitDemotedVars(MF->getFunction(), O);
821 OutStreamer.EmitRawText(O.str());
822 }
823
EmitFunctionBodyEnd()824 void NVPTXAsmPrinter::EmitFunctionBodyEnd() {
825 OutStreamer.EmitRawText(StringRef("}\n"));
826 VRegMapping.clear();
827 }
828
emitImplicitDef(const MachineInstr * MI) const829 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));
835 } else {
836 OutStreamer.AddComment(Twine("implicit-def: ") +
837 TM.getRegisterInfo()->getName(RegNo));
838 }
839 OutStreamer.AddBlankLine();
840 }
841
emitKernelFunctionDirectives(const Function & F,raw_ostream & O) const842 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)
850 reqntidx = 1;
851 else
852 specified = true;
853 if (llvm::getReqNTIDy(F, reqntidy) == false)
854 reqntidy = 1;
855 else
856 specified = true;
857 if (llvm::getReqNTIDz(F, reqntidz) == false)
858 reqntidz = 1;
859 else
860 specified = true;
861
862 if (specified)
863 O << ".reqntid " << reqntidx << ", " << reqntidy << ", " << reqntidz
864 << "\n";
865
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;
870 specified = false;
871 if (llvm::getMaxNTIDx(F, maxntidx) == false)
872 maxntidx = 1;
873 else
874 specified = true;
875 if (llvm::getMaxNTIDy(F, maxntidy) == false)
876 maxntidy = 1;
877 else
878 specified = true;
879 if (llvm::getMaxNTIDz(F, maxntidz) == false)
880 maxntidz = 1;
881 else
882 specified = true;
883
884 if (specified)
885 O << ".maxntid " << maxntidx << ", " << maxntidy << ", " << maxntidz
886 << "\n";
887
888 unsigned mincta;
889 if (llvm::getMinCTASm(F, mincta))
890 O << ".minnctapersm " << mincta << "\n";
891 }
892
893 std::string
getVirtualRegisterName(unsigned Reg) const894 NVPTXAsmPrinter::getVirtualRegisterName(unsigned Reg) const {
895 const TargetRegisterClass *RC = MRI->getRegClass(Reg);
896
897 std::string Name;
898 raw_string_ostream NameStr(Name);
899
900 VRegRCMap::const_iterator I = VRegMapping.find(RC);
901 assert(I != VRegMapping.end() && "Bad register class");
902 const DenseMap<unsigned, unsigned> &RegMap = I->second;
903
904 VRegMap::const_iterator VI = RegMap.find(Reg);
905 assert(VI != RegMap.end() && "Bad virtual register");
906 unsigned MappedVR = VI->second;
907
908 NameStr << getNVPTXRegClassStr(RC) << MappedVR;
909
910 NameStr.flush();
911 return Name;
912 }
913
emitVirtualRegister(unsigned int vr,raw_ostream & O)914 void NVPTXAsmPrinter::emitVirtualRegister(unsigned int vr,
915 raw_ostream &O) {
916 O << getVirtualRegisterName(vr);
917 }
918
printVecModifiedImmediate(const MachineOperand & MO,const char * Modifier,raw_ostream & O)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))
927 O << "//";
928 } else if (0 == strcmp(Modifier, "vecv4comm2")) {
929 if ((Imm < 4) || (Imm > 7))
930 O << "//";
931 } else if (0 == strcmp(Modifier, "vecv4pos")) {
932 if (Imm < 0)
933 Imm = 0;
934 O << "_" << vecelem[Imm % 4];
935 } else if (0 == strcmp(Modifier, "vecv2comm1")) {
936 if ((Imm < 0) || (Imm > 1))
937 O << "//";
938 } else if (0 == strcmp(Modifier, "vecv2comm2")) {
939 if ((Imm < 2) || (Imm > 3))
940 O << "//";
941 } else if (0 == strcmp(Modifier, "vecv2pos")) {
942 if (Imm < 0)
943 Imm = 0;
944 O << "_" << vecelem[Imm % 2];
945 } else
946 llvm_unreachable("Unknown Modifier on immediate operand");
947 }
948
949
950
emitDeclaration(const Function * F,raw_ostream & O)951 void NVPTXAsmPrinter::emitDeclaration(const Function *F, raw_ostream &O) {
952
953 emitLinkageDirective(F, O);
954 if (llvm::isKernelFunction(*F))
955 O << ".entry ";
956 else
957 O << ".func ";
958 printReturnValStr(F, O);
959 O << *getSymbol(F) << "\n";
960 emitFunctionParamList(F, O);
961 O << ";\n";
962 }
963
usedInGlobalVarDef(const Constant * C)964 static bool usedInGlobalVarDef(const Constant *C) {
965 if (!C)
966 return false;
967
968 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) {
969 if (GV->getName().str() == "llvm.used")
970 return false;
971 return true;
972 }
973
974 for (const User *U : C->users())
975 if (const Constant *C = dyn_cast<Constant>(U))
976 if (usedInGlobalVarDef(C))
977 return true;
978
979 return false;
980 }
981
usedInOneFunc(const User * U,Function const * & oneFunc)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")
985 return true;
986 }
987
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))
992 return false;
993 oneFunc = curFunc;
994 return true;
995 } else
996 return false;
997 }
998
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")))
1002 return true;
1003
1004 for (const User *UU : U->users())
1005 if (usedInOneFunc(UU, oneFunc) == false)
1006 return false;
1007
1008 return true;
1009 }
1010
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?
1017 */
canDemoteGlobalVar(const GlobalVariable * gv,Function const * & f)1018 static bool canDemoteGlobalVar(const GlobalVariable *gv, Function const *&f) {
1019 if (gv->hasInternalLinkage() == false)
1020 return false;
1021 const PointerType *Pty = gv->getType();
1022 if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
1023 return false;
1024
1025 const Function *oneFunc = nullptr;
1026
1027 bool flag = usedInOneFunc(gv, oneFunc);
1028 if (flag == false)
1029 return false;
1030 if (!oneFunc)
1031 return false;
1032 f = oneFunc;
1033 return true;
1034 }
1035
useFuncSeen(const Constant * C,llvm::DenseMap<const Function *,bool> & seenMap)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))
1041 return true;
1042 } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
1043 const BasicBlock *bb = I->getParent();
1044 if (!bb)
1045 continue;
1046 const Function *caller = bb->getParent();
1047 if (!caller)
1048 continue;
1049 if (seenMap.find(caller) != seenMap.end())
1050 return true;
1051 }
1052 }
1053 return false;
1054 }
1055
emitDeclarations(const Module & M,raw_ostream & O)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;
1060
1061 if (F->isDeclaration()) {
1062 if (F->use_empty())
1063 continue;
1064 if (F->getIntrinsicID())
1065 continue;
1066 emitDeclaration(F, O);
1067 continue;
1068 }
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);
1076 break;
1077 }
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);
1082 break;
1083 }
1084 }
1085
1086 if (!isa<Instruction>(U))
1087 continue;
1088 const Instruction *instr = cast<Instruction>(U);
1089 const BasicBlock *bb = instr->getParent();
1090 if (!bb)
1091 continue;
1092 const Function *caller = bb->getParent();
1093 if (!caller)
1094 continue;
1095
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);
1101 break;
1102 }
1103 }
1104 seenMap[F] = true;
1105 }
1106 }
1107
recordAndEmitFilenames(Module & M)1108 void NVPTXAsmPrinter::recordAndEmitFilenames(Module &M) {
1109 DebugInfoFinder DbgFinder;
1110 DbgFinder.processModule(M);
1111
1112 unsigned i = 1;
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();
1120 }
1121 if (filenameMap.find(Filename.str()) != filenameMap.end())
1122 continue;
1123 filenameMap[Filename.str()] = i;
1124 OutStreamer.EmitDwarfFileDirective(i, "", Filename.str());
1125 ++i;
1126 }
1127
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();
1135 }
1136 if (filenameMap.find(Filename.str()) != filenameMap.end())
1137 continue;
1138 filenameMap[Filename.str()] = i;
1139 ++i;
1140 }
1141 }
1142
doInitialization(Module & M)1143 bool NVPTXAsmPrinter::doInitialization(Module &M) {
1144
1145 SmallString<128> Str1;
1146 raw_svector_ostream OS1(Str1);
1147
1148 MMI = getAnalysisIfAvailable<MachineModuleInfo>();
1149 MMI->AnalyzeModule(M);
1150
1151 // We need to call the parent's one explicitly.
1152 //bool Result = AsmPrinter::doInitialization(M);
1153
1154 // Initialize TargetLoweringObjectFile.
1155 const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
1156 .Initialize(OutContext, TM);
1157
1158 Mang = new Mangler(TM.getDataLayout());
1159
1160 // Emit header before any dwarf directives are emitted below.
1161 emitHeader(M, OS1);
1162 OutStreamer.EmitRawText(OS1.str());
1163
1164 // Already commented out
1165 //bool Result = AsmPrinter::doInitialization(M);
1166
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();
1175 }
1176
1177 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
1178 recordAndEmitFilenames(M);
1179
1180 GlobalsEmitted = false;
1181
1182 return false; // success
1183 }
1184
emitGlobals(const Module & M)1185 void NVPTXAsmPrinter::emitGlobals(const Module &M) {
1186 SmallString<128> Str2;
1187 raw_svector_ostream OS2(Str2);
1188
1189 emitDeclarations(M, OS2);
1190
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;
1199
1200 // Visit each global variable, in order
1201 for (Module::const_global_iterator I = M.global_begin(), E = M.global_end();
1202 I != E; ++I)
1203 VisitGlobalVariableForEmission(I, Globals, GVVisited, GVVisiting);
1204
1205 assert(GVVisited.size() == M.getGlobalList().size() &&
1206 "Missed a global variable");
1207 assert(GVVisiting.size() == 0 && "Did not fully process a global variable");
1208
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);
1212
1213 OS2 << '\n';
1214
1215 OutStreamer.EmitRawText(OS2.str());
1216 }
1217
emitHeader(Module & M,raw_ostream & O)1218 void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O) {
1219 O << "//\n";
1220 O << "// Generated by LLVM NVPTX Back-End\n";
1221 O << "//\n";
1222 O << "\n";
1223
1224 unsigned PTXVersion = nvptxSubtarget.getPTXVersion();
1225 O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
1226
1227 O << ".target ";
1228 O << nvptxSubtarget.getTargetName();
1229
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";
1235 }
1236
1237 if (MAI->doesSupportDebugInformation())
1238 O << ", debug";
1239
1240 O << "\n";
1241
1242 O << ".address_size ";
1243 if (nvptxSubtarget.is64Bit())
1244 O << "64";
1245 else
1246 O << "32";
1247 O << "\n";
1248
1249 O << "\n";
1250 }
1251
doFinalization(Module & M)1252 bool NVPTXAsmPrinter::doFinalization(Module &M) {
1253
1254 // If we did not emit any functions, then the global declarations have not
1255 // yet been emitted.
1256 if (!GlobalsEmitted) {
1257 emitGlobals(M);
1258 GlobalsEmitted = true;
1259 }
1260
1261 // XXX Temproarily remove global variables so that doFinalization() will not
1262 // emit them again (global variables are emitted at beginning).
1263
1264 Module::GlobalListType &global_list = M.getGlobalList();
1265 int i, n = global_list.size();
1266 GlobalVariable **gv_array = new GlobalVariable *[n];
1267
1268 // first, back-up GlobalVariable in gv_array
1269 i = 0;
1270 for (Module::global_iterator I = global_list.begin(), E = global_list.end();
1271 I != E; ++I)
1272 gv_array[i++] = &*I;
1273
1274 // second, empty global_list
1275 while (!global_list.empty())
1276 global_list.remove(global_list.begin());
1277
1278 // call doFinalization
1279 bool ret = AsmPrinter::doFinalization(M);
1280
1281 // now we restore global variables
1282 for (i = 0; i < n; i++)
1283 global_list.insert(global_list.end(), gv_array[i]);
1284
1285 clearAnnotationCache(&M);
1286
1287 delete[] gv_array;
1288 return ret;
1289
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.
1295 //
1296 // Same for the doInitialization.
1297 //return Result;
1298 }
1299
1300 // This function emits appropriate linkage directives for
1301 // functions and global variables.
1302 //
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.
1312
emitLinkageDirective(const GlobalValue * V,raw_ostream & O)1313 void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
1314 raw_ostream &O) {
1315 if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
1316 if (V->hasExternalLinkage()) {
1317 if (isa<GlobalVariable>(V)) {
1318 const GlobalVariable *GVar = cast<GlobalVariable>(V);
1319 if (GVar) {
1320 if (GVar->hasInitializer())
1321 O << ".visible ";
1322 else
1323 O << ".extern ";
1324 }
1325 } else if (V->isDeclaration())
1326 O << ".extern ";
1327 else
1328 O << ".visible ";
1329 } else if (V->hasAppendingLinkage()) {
1330 std::string msg;
1331 msg.append("Error: ");
1332 msg.append("Symbol ");
1333 if (V->hasName())
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()) {
1339 O << ".weak ";
1340 }
1341 }
1342 }
1343
printModuleLevelGV(const GlobalVariable * GVar,raw_ostream & O,bool processDemoted)1344 void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
1345 raw_ostream &O,
1346 bool processDemoted) {
1347
1348 // Skip meta data
1349 if (GVar->hasSection()) {
1350 if (GVar->getSection() == StringRef("llvm.metadata"))
1351 return;
1352 }
1353
1354 // Skip LLVM intrinsic global variables
1355 if (GVar->getName().startswith("llvm.") ||
1356 GVar->getName().startswith("nvvm."))
1357 return;
1358
1359 const DataLayout *TD = TM.getDataLayout();
1360
1361 // GlobalVariables are always constant pointers themselves.
1362 const PointerType *PTy = GVar->getType();
1363 Type *ETy = PTy->getElementType();
1364
1365 if (GVar->hasExternalLinkage()) {
1366 if (GVar->hasInitializer())
1367 O << ".visible ";
1368 else
1369 O << ".extern ";
1370 } else if (GVar->hasLinkOnceLinkage() || GVar->hasWeakLinkage() ||
1371 GVar->hasAvailableExternallyLinkage() ||
1372 GVar->hasCommonLinkage()) {
1373 O << ".weak ";
1374 }
1375
1376 if (llvm::isTexture(*GVar)) {
1377 O << ".global .texref " << llvm::getTextureName(*GVar) << ";\n";
1378 return;
1379 }
1380
1381 if (llvm::isSurface(*GVar)) {
1382 O << ".global .surfref " << llvm::getSurfaceName(*GVar) << ";\n";
1383 return;
1384 }
1385
1386 if (GVar->isDeclaration()) {
1387 // (extern) declarations, no definition or initializer
1388 // Currently the only known declaration is for an automatic __local
1389 // (.shared) promoted to global.
1390 emitPTXGlobalVariable(GVar, O);
1391 O << ";\n";
1392 return;
1393 }
1394
1395 if (llvm::isSampler(*GVar)) {
1396 O << ".global .samplerref " << llvm::getSamplerName(*GVar);
1397
1398 const Constant *Initializer = nullptr;
1399 if (GVar->hasInitializer())
1400 Initializer = GVar->getInitializer();
1401 const ConstantInt *CI = nullptr;
1402 if (Initializer)
1403 CI = dyn_cast<ConstantInt>(Initializer);
1404 if (CI) {
1405 unsigned sample = CI->getZExtValue();
1406
1407 O << " = { ";
1408
1409 for (int i = 0,
1410 addr = ((sample & __CLK_ADDRESS_MASK) >> __CLK_ADDRESS_BASE);
1411 i < 3; i++) {
1412 O << "addr_mode_" << i << " = ";
1413 switch (addr) {
1414 case 0:
1415 O << "wrap";
1416 break;
1417 case 1:
1418 O << "clamp_to_border";
1419 break;
1420 case 2:
1421 O << "clamp_to_edge";
1422 break;
1423 case 3:
1424 O << "wrap";
1425 break;
1426 case 4:
1427 O << "mirror";
1428 break;
1429 }
1430 O << ", ";
1431 }
1432 O << "filter_mode = ";
1433 switch ((sample & __CLK_FILTER_MASK) >> __CLK_FILTER_BASE) {
1434 case 0:
1435 O << "nearest";
1436 break;
1437 case 1:
1438 O << "linear";
1439 break;
1440 case 2:
1441 llvm_unreachable("Anisotropic filtering is not supported");
1442 default:
1443 O << "nearest";
1444 break;
1445 }
1446 if (!((sample & __CLK_NORMALIZED_MASK) >> __CLK_NORMALIZED_BASE)) {
1447 O << ", force_unnormalized_coords = 1";
1448 }
1449 O << " }";
1450 }
1451
1452 O << ";\n";
1453 return;
1454 }
1455
1456 if (GVar->hasPrivateLinkage()) {
1457
1458 if (!strncmp(GVar->getName().data(), "unrollpragma", 12))
1459 return;
1460
1461 // FIXME - need better way (e.g. Metadata) to avoid generating this global
1462 if (!strncmp(GVar->getName().data(), "filename", 8))
1463 return;
1464 if (GVar->use_empty())
1465 return;
1466 }
1467
1468 const Function *demotedFunc = nullptr;
1469 if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
1470 O << "// " << GVar->getName().str() << " has been demoted\n";
1471 if (localDecls.find(demotedFunc) != localDecls.end())
1472 localDecls[demotedFunc].push_back(GVar);
1473 else {
1474 std::vector<const GlobalVariable *> temp;
1475 temp.push_back(GVar);
1476 localDecls[demotedFunc] = temp;
1477 }
1478 return;
1479 }
1480
1481 O << ".";
1482 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1483
1484 if (isManaged(*GVar)) {
1485 O << " .attribute(.managed)";
1486 }
1487
1488 if (GVar->getAlignment() == 0)
1489 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1490 else
1491 O << " .align " << GVar->getAlignment();
1492
1493 if (ETy->isSingleValueType()) {
1494 O << " .";
1495 // Special case: ABI requires that we use .u8 for predicates
1496 if (ETy->isIntegerTy(1))
1497 O << "u8";
1498 else
1499 O << getPTXFundamentalTypeStr(ETy, false);
1500 O << " ";
1501 O << *getSymbol(GVar);
1502
1503 // Ptx allows variable initilization only for constant and global state
1504 // spaces.
1505 if (GVar->hasInitializer()) {
1506 if ((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1507 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) {
1508 const Constant *Initializer = GVar->getInitializer();
1509 // 'undef' is treated as there is no value spefied.
1510 if (!Initializer->isNullValue() && !isa<UndefValue>(Initializer)) {
1511 O << " = ";
1512 printScalarConstant(Initializer, O);
1513 }
1514 } else {
1515 // The frontend adds zero-initializer to variables that don't have an
1516 // initial value, so skip warning for this case.
1517 if (!GVar->getInitializer()->isNullValue()) {
1518 std::string warnMsg = "initial value of '" + GVar->getName().str() +
1519 "' is not allowed in addrspace(" +
1520 llvm::utostr_32(PTy->getAddressSpace()) + ")";
1521 report_fatal_error(warnMsg.c_str());
1522 }
1523 }
1524 }
1525 } else {
1526 unsigned int ElementSize = 0;
1527
1528 // Although PTX has direct support for struct type and array type and
1529 // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
1530 // targets that support these high level field accesses. Structs, arrays
1531 // and vectors are lowered into arrays of bytes.
1532 switch (ETy->getTypeID()) {
1533 case Type::StructTyID:
1534 case Type::ArrayTyID:
1535 case Type::VectorTyID:
1536 ElementSize = TD->getTypeStoreSize(ETy);
1537 // Ptx allows variable initilization only for constant and
1538 // global state spaces.
1539 if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
1540 (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
1541 GVar->hasInitializer()) {
1542 const Constant *Initializer = GVar->getInitializer();
1543 if (!isa<UndefValue>(Initializer) && !Initializer->isNullValue()) {
1544 AggBuffer aggBuffer(ElementSize, O, *this);
1545 bufferAggregateConstant(Initializer, &aggBuffer);
1546 if (aggBuffer.numSymbols) {
1547 if (nvptxSubtarget.is64Bit()) {
1548 O << " .u64 " << *getSymbol(GVar) << "[";
1549 O << ElementSize / 8;
1550 } else {
1551 O << " .u32 " << *getSymbol(GVar) << "[";
1552 O << ElementSize / 4;
1553 }
1554 O << "]";
1555 } else {
1556 O << " .b8 " << *getSymbol(GVar) << "[";
1557 O << ElementSize;
1558 O << "]";
1559 }
1560 O << " = {";
1561 aggBuffer.print();
1562 O << "}";
1563 } else {
1564 O << " .b8 " << *getSymbol(GVar);
1565 if (ElementSize) {
1566 O << "[";
1567 O << ElementSize;
1568 O << "]";
1569 }
1570 }
1571 } else {
1572 O << " .b8 " << *getSymbol(GVar);
1573 if (ElementSize) {
1574 O << "[";
1575 O << ElementSize;
1576 O << "]";
1577 }
1578 }
1579 break;
1580 default:
1581 llvm_unreachable("type not supported yet");
1582 }
1583
1584 }
1585 O << ";\n";
1586 }
1587
emitDemotedVars(const Function * f,raw_ostream & O)1588 void NVPTXAsmPrinter::emitDemotedVars(const Function *f, raw_ostream &O) {
1589 if (localDecls.find(f) == localDecls.end())
1590 return;
1591
1592 std::vector<const GlobalVariable *> &gvars = localDecls[f];
1593
1594 for (unsigned i = 0, e = gvars.size(); i != e; ++i) {
1595 O << "\t// demoted variable\n\t";
1596 printModuleLevelGV(gvars[i], O, true);
1597 }
1598 }
1599
emitPTXAddressSpace(unsigned int AddressSpace,raw_ostream & O) const1600 void NVPTXAsmPrinter::emitPTXAddressSpace(unsigned int AddressSpace,
1601 raw_ostream &O) const {
1602 switch (AddressSpace) {
1603 case llvm::ADDRESS_SPACE_LOCAL:
1604 O << "local";
1605 break;
1606 case llvm::ADDRESS_SPACE_GLOBAL:
1607 O << "global";
1608 break;
1609 case llvm::ADDRESS_SPACE_CONST:
1610 O << "const";
1611 break;
1612 case llvm::ADDRESS_SPACE_SHARED:
1613 O << "shared";
1614 break;
1615 default:
1616 report_fatal_error("Bad address space found while emitting PTX");
1617 break;
1618 }
1619 }
1620
1621 std::string
getPTXFundamentalTypeStr(const Type * Ty,bool useB4PTR) const1622 NVPTXAsmPrinter::getPTXFundamentalTypeStr(const Type *Ty, bool useB4PTR) const {
1623 switch (Ty->getTypeID()) {
1624 default:
1625 llvm_unreachable("unexpected type");
1626 break;
1627 case Type::IntegerTyID: {
1628 unsigned NumBits = cast<IntegerType>(Ty)->getBitWidth();
1629 if (NumBits == 1)
1630 return "pred";
1631 else if (NumBits <= 64) {
1632 std::string name = "u";
1633 return name + utostr(NumBits);
1634 } else {
1635 llvm_unreachable("Integer too large");
1636 break;
1637 }
1638 break;
1639 }
1640 case Type::FloatTyID:
1641 return "f32";
1642 case Type::DoubleTyID:
1643 return "f64";
1644 case Type::PointerTyID:
1645 if (nvptxSubtarget.is64Bit())
1646 if (useB4PTR)
1647 return "b64";
1648 else
1649 return "u64";
1650 else if (useB4PTR)
1651 return "b32";
1652 else
1653 return "u32";
1654 }
1655 llvm_unreachable("unexpected type");
1656 return nullptr;
1657 }
1658
emitPTXGlobalVariable(const GlobalVariable * GVar,raw_ostream & O)1659 void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
1660 raw_ostream &O) {
1661
1662 const DataLayout *TD = TM.getDataLayout();
1663
1664 // GlobalVariables are always constant pointers themselves.
1665 const PointerType *PTy = GVar->getType();
1666 Type *ETy = PTy->getElementType();
1667
1668 O << ".";
1669 emitPTXAddressSpace(PTy->getAddressSpace(), O);
1670 if (GVar->getAlignment() == 0)
1671 O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
1672 else
1673 O << " .align " << GVar->getAlignment();
1674
1675 if (ETy->isSingleValueType()) {
1676 O << " .";
1677 O << getPTXFundamentalTypeStr(ETy);
1678 O << " ";
1679 O << *getSymbol(GVar);
1680 return;
1681 }
1682
1683 int64_t ElementSize = 0;
1684
1685 // Although PTX has direct support for struct type and array type and LLVM IR
1686 // is very similar to PTX, the LLVM CodeGen does not support for targets that
1687 // support these high level field accesses. Structs and arrays are lowered
1688 // into arrays of bytes.
1689 switch (ETy->getTypeID()) {
1690 case Type::StructTyID:
1691 case Type::ArrayTyID:
1692 case Type::VectorTyID:
1693 ElementSize = TD->getTypeStoreSize(ETy);
1694 O << " .b8 " << *getSymbol(GVar) << "[";
1695 if (ElementSize) {
1696 O << itostr(ElementSize);
1697 }
1698 O << "]";
1699 break;
1700 default:
1701 llvm_unreachable("type not supported yet");
1702 }
1703 return;
1704 }
1705
getOpenCLAlignment(const DataLayout * TD,Type * Ty)1706 static unsigned int getOpenCLAlignment(const DataLayout *TD, Type *Ty) {
1707 if (Ty->isSingleValueType())
1708 return TD->getPrefTypeAlignment(Ty);
1709
1710 const ArrayType *ATy = dyn_cast<ArrayType>(Ty);
1711 if (ATy)
1712 return getOpenCLAlignment(TD, ATy->getElementType());
1713
1714 const VectorType *VTy = dyn_cast<VectorType>(Ty);
1715 if (VTy) {
1716 Type *ETy = VTy->getElementType();
1717 unsigned int numE = VTy->getNumElements();
1718 unsigned int alignE = TD->getPrefTypeAlignment(ETy);
1719 if (numE == 3)
1720 return 4 * alignE;
1721 else
1722 return numE * alignE;
1723 }
1724
1725 const StructType *STy = dyn_cast<StructType>(Ty);
1726 if (STy) {
1727 unsigned int alignStruct = 1;
1728 // Go through each element of the struct and find the
1729 // largest alignment.
1730 for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) {
1731 Type *ETy = STy->getElementType(i);
1732 unsigned int align = getOpenCLAlignment(TD, ETy);
1733 if (align > alignStruct)
1734 alignStruct = align;
1735 }
1736 return alignStruct;
1737 }
1738
1739 const FunctionType *FTy = dyn_cast<FunctionType>(Ty);
1740 if (FTy)
1741 return TD->getPointerPrefAlignment();
1742 return TD->getPrefTypeAlignment(Ty);
1743 }
1744
printParamName(Function::const_arg_iterator I,int paramIndex,raw_ostream & O)1745 void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
1746 int paramIndex, raw_ostream &O) {
1747 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1748 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA))
1749 O << *getSymbol(I->getParent()) << "_param_" << paramIndex;
1750 else {
1751 std::string argName = I->getName();
1752 const char *p = argName.c_str();
1753 while (*p) {
1754 if (*p == '.')
1755 O << "_";
1756 else
1757 O << *p;
1758 p++;
1759 }
1760 }
1761 }
1762
printParamName(int paramIndex,raw_ostream & O)1763 void NVPTXAsmPrinter::printParamName(int paramIndex, raw_ostream &O) {
1764 Function::const_arg_iterator I, E;
1765 int i = 0;
1766
1767 if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
1768 (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)) {
1769 O << *CurrentFnSym << "_param_" << paramIndex;
1770 return;
1771 }
1772
1773 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, i++) {
1774 if (i == paramIndex) {
1775 printParamName(I, paramIndex, O);
1776 return;
1777 }
1778 }
1779 llvm_unreachable("paramIndex out of bound");
1780 }
1781
emitFunctionParamList(const Function * F,raw_ostream & O)1782 void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
1783 const DataLayout *TD = TM.getDataLayout();
1784 const AttributeSet &PAL = F->getAttributes();
1785 const TargetLowering *TLI = TM.getTargetLowering();
1786 Function::const_arg_iterator I, E;
1787 unsigned paramIndex = 0;
1788 bool first = true;
1789 bool isKernelFunc = llvm::isKernelFunction(*F);
1790 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
1791 MVT thePointerTy = TLI->getPointerTy();
1792
1793 O << "(\n";
1794
1795 for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, paramIndex++) {
1796 Type *Ty = I->getType();
1797
1798 if (!first)
1799 O << ",\n";
1800
1801 first = false;
1802
1803 // Handle image/sampler parameters
1804 if (isKernelFunction(*F)) {
1805 if (isSampler(*I) || isImage(*I)) {
1806 if (isImage(*I)) {
1807 std::string sname = I->getName();
1808 if (isImageWriteOnly(*I) || isImageReadWrite(*I)) {
1809 if (nvptxSubtarget.hasImageHandles())
1810 O << "\t.param .u64 .ptr .surfref ";
1811 else
1812 O << "\t.param .surfref ";
1813 O << *CurrentFnSym << "_param_" << paramIndex;
1814 }
1815 else { // Default image is read_only
1816 if (nvptxSubtarget.hasImageHandles())
1817 O << "\t.param .u64 .ptr .texref ";
1818 else
1819 O << "\t.param .texref ";
1820 O << *CurrentFnSym << "_param_" << paramIndex;
1821 }
1822 } else {
1823 if (nvptxSubtarget.hasImageHandles())
1824 O << "\t.param .u64 .ptr .samplerref ";
1825 else
1826 O << "\t.param .samplerref ";
1827 O << *CurrentFnSym << "_param_" << paramIndex;
1828 }
1829 continue;
1830 }
1831 }
1832
1833 if (PAL.hasAttribute(paramIndex + 1, Attribute::ByVal) == false) {
1834 if (Ty->isAggregateType() || Ty->isVectorTy()) {
1835 // Just print .param .align <a> .b8 .param[size];
1836 // <a> = PAL.getparamalignment
1837 // size = typeallocsize of element type
1838 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1839 if (align == 0)
1840 align = TD->getABITypeAlignment(Ty);
1841
1842 unsigned sz = TD->getTypeAllocSize(Ty);
1843 O << "\t.param .align " << align << " .b8 ";
1844 printParamName(I, paramIndex, O);
1845 O << "[" << sz << "]";
1846
1847 continue;
1848 }
1849 // Just a scalar
1850 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1851 if (isKernelFunc) {
1852 if (PTy) {
1853 // Special handling for pointer arguments to kernel
1854 O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
1855
1856 if (nvptxSubtarget.getDrvInterface() != NVPTX::CUDA) {
1857 Type *ETy = PTy->getElementType();
1858 int addrSpace = PTy->getAddressSpace();
1859 switch (addrSpace) {
1860 default:
1861 O << ".ptr ";
1862 break;
1863 case llvm::ADDRESS_SPACE_CONST:
1864 O << ".ptr .const ";
1865 break;
1866 case llvm::ADDRESS_SPACE_SHARED:
1867 O << ".ptr .shared ";
1868 break;
1869 case llvm::ADDRESS_SPACE_GLOBAL:
1870 O << ".ptr .global ";
1871 break;
1872 }
1873 O << ".align " << (int) getOpenCLAlignment(TD, ETy) << " ";
1874 }
1875 printParamName(I, paramIndex, O);
1876 continue;
1877 }
1878
1879 // non-pointer scalar to kernel func
1880 O << "\t.param .";
1881 // Special case: predicate operands become .u8 types
1882 if (Ty->isIntegerTy(1))
1883 O << "u8";
1884 else
1885 O << getPTXFundamentalTypeStr(Ty);
1886 O << " ";
1887 printParamName(I, paramIndex, O);
1888 continue;
1889 }
1890 // Non-kernel function, just print .param .b<size> for ABI
1891 // and .reg .b<size> for non-ABI
1892 unsigned sz = 0;
1893 if (isa<IntegerType>(Ty)) {
1894 sz = cast<IntegerType>(Ty)->getBitWidth();
1895 if (sz < 32)
1896 sz = 32;
1897 } else if (isa<PointerType>(Ty))
1898 sz = thePointerTy.getSizeInBits();
1899 else
1900 sz = Ty->getPrimitiveSizeInBits();
1901 if (isABI)
1902 O << "\t.param .b" << sz << " ";
1903 else
1904 O << "\t.reg .b" << sz << " ";
1905 printParamName(I, paramIndex, O);
1906 continue;
1907 }
1908
1909 // param has byVal attribute. So should be a pointer
1910 const PointerType *PTy = dyn_cast<PointerType>(Ty);
1911 assert(PTy && "Param with byval attribute should be a pointer type");
1912 Type *ETy = PTy->getElementType();
1913
1914 if (isABI || isKernelFunc) {
1915 // Just print .param .align <a> .b8 .param[size];
1916 // <a> = PAL.getparamalignment
1917 // size = typeallocsize of element type
1918 unsigned align = PAL.getParamAlignment(paramIndex + 1);
1919 if (align == 0)
1920 align = TD->getABITypeAlignment(ETy);
1921
1922 unsigned sz = TD->getTypeAllocSize(ETy);
1923 O << "\t.param .align " << align << " .b8 ";
1924 printParamName(I, paramIndex, O);
1925 O << "[" << sz << "]";
1926 continue;
1927 } else {
1928 // Split the ETy into constituent parts and
1929 // print .param .b<size> <name> for each part.
1930 // Further, if a part is vector, print the above for
1931 // each vector element.
1932 SmallVector<EVT, 16> vtparts;
1933 ComputeValueVTs(*TLI, ETy, vtparts);
1934 for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
1935 unsigned elems = 1;
1936 EVT elemtype = vtparts[i];
1937 if (vtparts[i].isVector()) {
1938 elems = vtparts[i].getVectorNumElements();
1939 elemtype = vtparts[i].getVectorElementType();
1940 }
1941
1942 for (unsigned j = 0, je = elems; j != je; ++j) {
1943 unsigned sz = elemtype.getSizeInBits();
1944 if (elemtype.isInteger() && (sz < 32))
1945 sz = 32;
1946 O << "\t.reg .b" << sz << " ";
1947 printParamName(I, paramIndex, O);
1948 if (j < je - 1)
1949 O << ",\n";
1950 ++paramIndex;
1951 }
1952 if (i < e - 1)
1953 O << ",\n";
1954 }
1955 --paramIndex;
1956 continue;
1957 }
1958 }
1959
1960 O << "\n)\n";
1961 }
1962
emitFunctionParamList(const MachineFunction & MF,raw_ostream & O)1963 void NVPTXAsmPrinter::emitFunctionParamList(const MachineFunction &MF,
1964 raw_ostream &O) {
1965 const Function *F = MF.getFunction();
1966 emitFunctionParamList(F, O);
1967 }
1968
setAndEmitFunctionVirtualRegisters(const MachineFunction & MF)1969 void NVPTXAsmPrinter::setAndEmitFunctionVirtualRegisters(
1970 const MachineFunction &MF) {
1971 SmallString<128> Str;
1972 raw_svector_ostream O(Str);
1973
1974 // Map the global virtual register number to a register class specific
1975 // virtual register number starting from 1 with that class.
1976 const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo();
1977 //unsigned numRegClasses = TRI->getNumRegClasses();
1978
1979 // Emit the Fake Stack Object
1980 const MachineFrameInfo *MFI = MF.getFrameInfo();
1981 int NumBytes = (int) MFI->getStackSize();
1982 if (NumBytes) {
1983 O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
1984 << getFunctionNumber() << "[" << NumBytes << "];\n";
1985 if (nvptxSubtarget.is64Bit()) {
1986 O << "\t.reg .b64 \t%SP;\n";
1987 O << "\t.reg .b64 \t%SPL;\n";
1988 } else {
1989 O << "\t.reg .b32 \t%SP;\n";
1990 O << "\t.reg .b32 \t%SPL;\n";
1991 }
1992 }
1993
1994 // Go through all virtual registers to establish the mapping between the
1995 // global virtual
1996 // register number and the per class virtual register number.
1997 // We use the per class virtual register number in the ptx output.
1998 unsigned int numVRs = MRI->getNumVirtRegs();
1999 for (unsigned i = 0; i < numVRs; i++) {
2000 unsigned int vr = TRI->index2VirtReg(i);
2001 const TargetRegisterClass *RC = MRI->getRegClass(vr);
2002 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
2003 int n = regmap.size();
2004 regmap.insert(std::make_pair(vr, n + 1));
2005 }
2006
2007 // Emit register declarations
2008 // @TODO: Extract out the real register usage
2009 // O << "\t.reg .pred %p<" << NVPTXNumRegisters << ">;\n";
2010 // O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
2011 // O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
2012 // O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
2013 // O << "\t.reg .s64 %rl<" << NVPTXNumRegisters << ">;\n";
2014 // O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
2015 // O << "\t.reg .f64 %fl<" << NVPTXNumRegisters << ">;\n";
2016
2017 // Emit declaration of the virtual registers or 'physical' registers for
2018 // each register class
2019 for (unsigned i=0; i< TRI->getNumRegClasses(); i++) {
2020 const TargetRegisterClass *RC = TRI->getRegClass(i);
2021 DenseMap<unsigned, unsigned> ®map = VRegMapping[RC];
2022 std::string rcname = getNVPTXRegClassName(RC);
2023 std::string rcStr = getNVPTXRegClassStr(RC);
2024 int n = regmap.size();
2025
2026 // Only declare those registers that may be used.
2027 if (n) {
2028 O << "\t.reg " << rcname << " \t" << rcStr << "<" << (n+1)
2029 << ">;\n";
2030 }
2031 }
2032
2033 OutStreamer.EmitRawText(O.str());
2034 }
2035
printFPConstant(const ConstantFP * Fp,raw_ostream & O)2036 void NVPTXAsmPrinter::printFPConstant(const ConstantFP *Fp, raw_ostream &O) {
2037 APFloat APF = APFloat(Fp->getValueAPF()); // make a copy
2038 bool ignored;
2039 unsigned int numHex;
2040 const char *lead;
2041
2042 if (Fp->getType()->getTypeID() == Type::FloatTyID) {
2043 numHex = 8;
2044 lead = "0f";
2045 APF.convert(APFloat::IEEEsingle, APFloat::rmNearestTiesToEven, &ignored);
2046 } else if (Fp->getType()->getTypeID() == Type::DoubleTyID) {
2047 numHex = 16;
2048 lead = "0d";
2049 APF.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &ignored);
2050 } else
2051 llvm_unreachable("unsupported fp type");
2052
2053 APInt API = APF.bitcastToAPInt();
2054 std::string hexstr(utohexstr(API.getZExtValue()));
2055 O << lead;
2056 if (hexstr.length() < numHex)
2057 O << std::string(numHex - hexstr.length(), '0');
2058 O << utohexstr(API.getZExtValue());
2059 }
2060
printScalarConstant(const Constant * CPV,raw_ostream & O)2061 void NVPTXAsmPrinter::printScalarConstant(const Constant *CPV, raw_ostream &O) {
2062 if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
2063 O << CI->getValue();
2064 return;
2065 }
2066 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV)) {
2067 printFPConstant(CFP, O);
2068 return;
2069 }
2070 if (isa<ConstantPointerNull>(CPV)) {
2071 O << "0";
2072 return;
2073 }
2074 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
2075 PointerType *PTy = dyn_cast<PointerType>(GVar->getType());
2076 bool IsNonGenericPointer = false;
2077 if (PTy && PTy->getAddressSpace() != 0) {
2078 IsNonGenericPointer = true;
2079 }
2080 if (EmitGeneric && !isa<Function>(CPV) && !IsNonGenericPointer) {
2081 O << "generic(";
2082 O << *getSymbol(GVar);
2083 O << ")";
2084 } else {
2085 O << *getSymbol(GVar);
2086 }
2087 return;
2088 }
2089 if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
2090 const Value *v = Cexpr->stripPointerCasts();
2091 PointerType *PTy = dyn_cast<PointerType>(Cexpr->getType());
2092 bool IsNonGenericPointer = false;
2093 if (PTy && PTy->getAddressSpace() != 0) {
2094 IsNonGenericPointer = true;
2095 }
2096 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
2097 if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
2098 O << "generic(";
2099 O << *getSymbol(GVar);
2100 O << ")";
2101 } else {
2102 O << *getSymbol(GVar);
2103 }
2104 return;
2105 } else {
2106 O << *LowerConstant(CPV, *this);
2107 return;
2108 }
2109 }
2110 llvm_unreachable("Not scalar type found in printScalarConstant()");
2111 }
2112
bufferLEByte(const Constant * CPV,int Bytes,AggBuffer * aggBuffer)2113 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
2114 AggBuffer *aggBuffer) {
2115
2116 const DataLayout *TD = TM.getDataLayout();
2117
2118 if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
2119 int s = TD->getTypeAllocSize(CPV->getType());
2120 if (s < Bytes)
2121 s = Bytes;
2122 aggBuffer->addZeros(s);
2123 return;
2124 }
2125
2126 unsigned char *ptr;
2127 switch (CPV->getType()->getTypeID()) {
2128
2129 case Type::IntegerTyID: {
2130 const Type *ETy = CPV->getType();
2131 if (ETy == Type::getInt8Ty(CPV->getContext())) {
2132 unsigned char c =
2133 (unsigned char)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
2134 ptr = &c;
2135 aggBuffer->addBytes(ptr, 1, Bytes);
2136 } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
2137 short int16 = (short)(dyn_cast<ConstantInt>(CPV))->getZExtValue();
2138 ptr = (unsigned char *)&int16;
2139 aggBuffer->addBytes(ptr, 2, Bytes);
2140 } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
2141 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
2142 int int32 = (int)(constInt->getZExtValue());
2143 ptr = (unsigned char *)&int32;
2144 aggBuffer->addBytes(ptr, 4, Bytes);
2145 break;
2146 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
2147 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
2148 ConstantFoldConstantExpression(Cexpr, TD))) {
2149 int int32 = (int)(constInt->getZExtValue());
2150 ptr = (unsigned char *)&int32;
2151 aggBuffer->addBytes(ptr, 4, Bytes);
2152 break;
2153 }
2154 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
2155 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
2156 aggBuffer->addSymbol(v);
2157 aggBuffer->addZeros(4);
2158 break;
2159 }
2160 }
2161 llvm_unreachable("unsupported integer const type");
2162 } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
2163 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
2164 long long int64 = (long long)(constInt->getZExtValue());
2165 ptr = (unsigned char *)&int64;
2166 aggBuffer->addBytes(ptr, 8, Bytes);
2167 break;
2168 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
2169 if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
2170 ConstantFoldConstantExpression(Cexpr, TD))) {
2171 long long int64 = (long long)(constInt->getZExtValue());
2172 ptr = (unsigned char *)&int64;
2173 aggBuffer->addBytes(ptr, 8, Bytes);
2174 break;
2175 }
2176 if (Cexpr->getOpcode() == Instruction::PtrToInt) {
2177 Value *v = Cexpr->getOperand(0)->stripPointerCasts();
2178 aggBuffer->addSymbol(v);
2179 aggBuffer->addZeros(8);
2180 break;
2181 }
2182 }
2183 llvm_unreachable("unsupported integer const type");
2184 } else
2185 llvm_unreachable("unsupported integer const type");
2186 break;
2187 }
2188 case Type::FloatTyID:
2189 case Type::DoubleTyID: {
2190 const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
2191 const Type *Ty = CFP->getType();
2192 if (Ty == Type::getFloatTy(CPV->getContext())) {
2193 float float32 = (float) CFP->getValueAPF().convertToFloat();
2194 ptr = (unsigned char *)&float32;
2195 aggBuffer->addBytes(ptr, 4, Bytes);
2196 } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
2197 double float64 = CFP->getValueAPF().convertToDouble();
2198 ptr = (unsigned char *)&float64;
2199 aggBuffer->addBytes(ptr, 8, Bytes);
2200 } else {
2201 llvm_unreachable("unsupported fp const type");
2202 }
2203 break;
2204 }
2205 case Type::PointerTyID: {
2206 if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
2207 aggBuffer->addSymbol(GVar);
2208 } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
2209 const Value *v = Cexpr->stripPointerCasts();
2210 aggBuffer->addSymbol(v);
2211 }
2212 unsigned int s = TD->getTypeAllocSize(CPV->getType());
2213 aggBuffer->addZeros(s);
2214 break;
2215 }
2216
2217 case Type::ArrayTyID:
2218 case Type::VectorTyID:
2219 case Type::StructTyID: {
2220 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV) ||
2221 isa<ConstantStruct>(CPV) || isa<ConstantDataSequential>(CPV)) {
2222 int ElementSize = TD->getTypeAllocSize(CPV->getType());
2223 bufferAggregateConstant(CPV, aggBuffer);
2224 if (Bytes > ElementSize)
2225 aggBuffer->addZeros(Bytes - ElementSize);
2226 } else if (isa<ConstantAggregateZero>(CPV))
2227 aggBuffer->addZeros(Bytes);
2228 else
2229 llvm_unreachable("Unexpected Constant type");
2230 break;
2231 }
2232
2233 default:
2234 llvm_unreachable("unsupported type");
2235 }
2236 }
2237
bufferAggregateConstant(const Constant * CPV,AggBuffer * aggBuffer)2238 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
2239 AggBuffer *aggBuffer) {
2240 const DataLayout *TD = TM.getDataLayout();
2241 int Bytes;
2242
2243 // Old constants
2244 if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
2245 if (CPV->getNumOperands())
2246 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
2247 bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
2248 return;
2249 }
2250
2251 if (const ConstantDataSequential *CDS =
2252 dyn_cast<ConstantDataSequential>(CPV)) {
2253 if (CDS->getNumElements())
2254 for (unsigned i = 0; i < CDS->getNumElements(); ++i)
2255 bufferLEByte(cast<Constant>(CDS->getElementAsConstant(i)), 0,
2256 aggBuffer);
2257 return;
2258 }
2259
2260 if (isa<ConstantStruct>(CPV)) {
2261 if (CPV->getNumOperands()) {
2262 StructType *ST = cast<StructType>(CPV->getType());
2263 for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i) {
2264 if (i == (e - 1))
2265 Bytes = TD->getStructLayout(ST)->getElementOffset(0) +
2266 TD->getTypeAllocSize(ST) -
2267 TD->getStructLayout(ST)->getElementOffset(i);
2268 else
2269 Bytes = TD->getStructLayout(ST)->getElementOffset(i + 1) -
2270 TD->getStructLayout(ST)->getElementOffset(i);
2271 bufferLEByte(cast<Constant>(CPV->getOperand(i)), Bytes, aggBuffer);
2272 }
2273 }
2274 return;
2275 }
2276 llvm_unreachable("unsupported constant type in printAggregateConstant()");
2277 }
2278
2279 // buildTypeNameMap - Run through symbol table looking for type names.
2280 //
2281
isImageType(const Type * Ty)2282 bool NVPTXAsmPrinter::isImageType(const Type *Ty) {
2283
2284 std::map<const Type *, std::string>::iterator PI = TypeNameMap.find(Ty);
2285
2286 if (PI != TypeNameMap.end() && (!PI->second.compare("struct._image1d_t") ||
2287 !PI->second.compare("struct._image2d_t") ||
2288 !PI->second.compare("struct._image3d_t")))
2289 return true;
2290
2291 return false;
2292 }
2293
2294
ignoreLoc(const MachineInstr & MI)2295 bool NVPTXAsmPrinter::ignoreLoc(const MachineInstr &MI) {
2296 switch (MI.getOpcode()) {
2297 default:
2298 return false;
2299 case NVPTX::CallArgBeginInst:
2300 case NVPTX::CallArgEndInst0:
2301 case NVPTX::CallArgEndInst1:
2302 case NVPTX::CallArgF32:
2303 case NVPTX::CallArgF64:
2304 case NVPTX::CallArgI16:
2305 case NVPTX::CallArgI32:
2306 case NVPTX::CallArgI32imm:
2307 case NVPTX::CallArgI64:
2308 case NVPTX::CallArgParam:
2309 case NVPTX::CallVoidInst:
2310 case NVPTX::CallVoidInstReg:
2311 case NVPTX::Callseq_End:
2312 case NVPTX::CallVoidInstReg64:
2313 case NVPTX::DeclareParamInst:
2314 case NVPTX::DeclareRetMemInst:
2315 case NVPTX::DeclareRetRegInst:
2316 case NVPTX::DeclareRetScalarInst:
2317 case NVPTX::DeclareScalarParamInst:
2318 case NVPTX::DeclareScalarRegInst:
2319 case NVPTX::StoreParamF32:
2320 case NVPTX::StoreParamF64:
2321 case NVPTX::StoreParamI16:
2322 case NVPTX::StoreParamI32:
2323 case NVPTX::StoreParamI64:
2324 case NVPTX::StoreParamI8:
2325 case NVPTX::StoreRetvalF32:
2326 case NVPTX::StoreRetvalF64:
2327 case NVPTX::StoreRetvalI16:
2328 case NVPTX::StoreRetvalI32:
2329 case NVPTX::StoreRetvalI64:
2330 case NVPTX::StoreRetvalI8:
2331 case NVPTX::LastCallArgF32:
2332 case NVPTX::LastCallArgF64:
2333 case NVPTX::LastCallArgI16:
2334 case NVPTX::LastCallArgI32:
2335 case NVPTX::LastCallArgI32imm:
2336 case NVPTX::LastCallArgI64:
2337 case NVPTX::LastCallArgParam:
2338 case NVPTX::LoadParamMemF32:
2339 case NVPTX::LoadParamMemF64:
2340 case NVPTX::LoadParamMemI16:
2341 case NVPTX::LoadParamMemI32:
2342 case NVPTX::LoadParamMemI64:
2343 case NVPTX::LoadParamMemI8:
2344 case NVPTX::PrototypeInst:
2345 case NVPTX::DBG_VALUE:
2346 return true;
2347 }
2348 return false;
2349 }
2350
2351 /// PrintAsmOperand - Print out an operand for an inline asm expression.
2352 ///
PrintAsmOperand(const MachineInstr * MI,unsigned OpNo,unsigned AsmVariant,const char * ExtraCode,raw_ostream & O)2353 bool NVPTXAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
2354 unsigned AsmVariant,
2355 const char *ExtraCode, raw_ostream &O) {
2356 if (ExtraCode && ExtraCode[0]) {
2357 if (ExtraCode[1] != 0)
2358 return true; // Unknown modifier.
2359
2360 switch (ExtraCode[0]) {
2361 default:
2362 // See if this is a generic print operand
2363 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O);
2364 case 'r':
2365 break;
2366 }
2367 }
2368
2369 printOperand(MI, OpNo, O);
2370
2371 return false;
2372 }
2373
PrintAsmMemoryOperand(const MachineInstr * MI,unsigned OpNo,unsigned AsmVariant,const char * ExtraCode,raw_ostream & O)2374 bool NVPTXAsmPrinter::PrintAsmMemoryOperand(
2375 const MachineInstr *MI, unsigned OpNo, unsigned AsmVariant,
2376 const char *ExtraCode, raw_ostream &O) {
2377 if (ExtraCode && ExtraCode[0])
2378 return true; // Unknown modifier
2379
2380 O << '[';
2381 printMemOperand(MI, OpNo, O);
2382 O << ']';
2383
2384 return false;
2385 }
2386
printOperand(const MachineInstr * MI,int opNum,raw_ostream & O,const char * Modifier)2387 void NVPTXAsmPrinter::printOperand(const MachineInstr *MI, int opNum,
2388 raw_ostream &O, const char *Modifier) {
2389 const MachineOperand &MO = MI->getOperand(opNum);
2390 switch (MO.getType()) {
2391 case MachineOperand::MO_Register:
2392 if (TargetRegisterInfo::isPhysicalRegister(MO.getReg())) {
2393 if (MO.getReg() == NVPTX::VRDepot)
2394 O << DEPOTNAME << getFunctionNumber();
2395 else
2396 O << NVPTXInstPrinter::getRegisterName(MO.getReg());
2397 } else {
2398 emitVirtualRegister(MO.getReg(), O);
2399 }
2400 return;
2401
2402 case MachineOperand::MO_Immediate:
2403 if (!Modifier)
2404 O << MO.getImm();
2405 else if (strstr(Modifier, "vec") == Modifier)
2406 printVecModifiedImmediate(MO, Modifier, O);
2407 else
2408 llvm_unreachable(
2409 "Don't know how to handle modifier on immediate operand");
2410 return;
2411
2412 case MachineOperand::MO_FPImmediate:
2413 printFPConstant(MO.getFPImm(), O);
2414 break;
2415
2416 case MachineOperand::MO_GlobalAddress:
2417 O << *getSymbol(MO.getGlobal());
2418 break;
2419
2420 case MachineOperand::MO_MachineBasicBlock:
2421 O << *MO.getMBB()->getSymbol();
2422 return;
2423
2424 default:
2425 llvm_unreachable("Operand type not supported.");
2426 }
2427 }
2428
printMemOperand(const MachineInstr * MI,int opNum,raw_ostream & O,const char * Modifier)2429 void NVPTXAsmPrinter::printMemOperand(const MachineInstr *MI, int opNum,
2430 raw_ostream &O, const char *Modifier) {
2431 printOperand(MI, opNum, O);
2432
2433 if (Modifier && !strcmp(Modifier, "add")) {
2434 O << ", ";
2435 printOperand(MI, opNum + 1, O);
2436 } else {
2437 if (MI->getOperand(opNum + 1).isImm() &&
2438 MI->getOperand(opNum + 1).getImm() == 0)
2439 return; // don't print ',0' or '+0'
2440 O << "+";
2441 printOperand(MI, opNum + 1, O);
2442 }
2443 }
2444
2445
2446 // Force static initialization.
LLVMInitializeNVPTXBackendAsmPrinter()2447 extern "C" void LLVMInitializeNVPTXBackendAsmPrinter() {
2448 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2449 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2450 }
2451
emitSrcInText(StringRef filename,unsigned line)2452 void NVPTXAsmPrinter::emitSrcInText(StringRef filename, unsigned line) {
2453 std::stringstream temp;
2454 LineReader *reader = this->getReader(filename.str());
2455 temp << "\n//";
2456 temp << filename.str();
2457 temp << ":";
2458 temp << line;
2459 temp << " ";
2460 temp << reader->readLine(line);
2461 temp << "\n";
2462 this->OutStreamer.EmitRawText(Twine(temp.str()));
2463 }
2464
getReader(std::string filename)2465 LineReader *NVPTXAsmPrinter::getReader(std::string filename) {
2466 if (!reader) {
2467 reader = new LineReader(filename);
2468 }
2469
2470 if (reader->fileName() != filename) {
2471 delete reader;
2472 reader = new LineReader(filename);
2473 }
2474
2475 return reader;
2476 }
2477
readLine(unsigned lineNum)2478 std::string LineReader::readLine(unsigned lineNum) {
2479 if (lineNum < theCurLine) {
2480 theCurLine = 0;
2481 fstr.seekg(0, std::ios::beg);
2482 }
2483 while (theCurLine < lineNum) {
2484 fstr.getline(buff, 500);
2485 theCurLine++;
2486 }
2487 return buff;
2488 }
2489
2490 // Force static initialization.
LLVMInitializeNVPTXAsmPrinter()2491 extern "C" void LLVMInitializeNVPTXAsmPrinter() {
2492 RegisterAsmPrinter<NVPTXAsmPrinter> X(TheNVPTXTarget32);
2493 RegisterAsmPrinter<NVPTXAsmPrinter> Y(TheNVPTXTarget64);
2494 }
2495