1 //===- NVPTXUtilities.cpp - Utility Functions -----------------------------===//
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 miscellaneous utility functions
11 //===----------------------------------------------------------------------===//
12
13 #include "NVPTXUtilities.h"
14 #include "NVPTX.h"
15 #include "llvm/IR/Constants.h"
16 #include "llvm/IR/Function.h"
17 #include "llvm/IR/GlobalVariable.h"
18 #include "llvm/IR/Module.h"
19 #include "llvm/IR/Operator.h"
20 #include <algorithm>
21 #include <cstring>
22 #include <map>
23 #include <string>
24 #include <vector>
25 #include "llvm/Support/ManagedStatic.h"
26 #include "llvm/IR/InstIterator.h"
27 #include "llvm/Support/MutexGuard.h"
28
29 using namespace llvm;
30
31 typedef std::map<std::string, std::vector<unsigned> > key_val_pair_t;
32 typedef std::map<const GlobalValue *, key_val_pair_t> global_val_annot_t;
33 typedef std::map<const Module *, global_val_annot_t> per_module_annot_t;
34
35 ManagedStatic<per_module_annot_t> annotationCache;
36 static sys::Mutex Lock;
37
clearAnnotationCache(const llvm::Module * Mod)38 void llvm::clearAnnotationCache(const llvm::Module *Mod) {
39 MutexGuard Guard(Lock);
40 annotationCache->erase(Mod);
41 }
42
cacheAnnotationFromMD(const MDNode * md,key_val_pair_t & retval)43 static void cacheAnnotationFromMD(const MDNode *md, key_val_pair_t &retval) {
44 MutexGuard Guard(Lock);
45 assert(md && "Invalid mdnode for annotation");
46 assert((md->getNumOperands() % 2) == 1 && "Invalid number of operands");
47 // start index = 1, to skip the global variable key
48 // increment = 2, to skip the value for each property-value pairs
49 for (unsigned i = 1, e = md->getNumOperands(); i != e; i += 2) {
50 // property
51 const MDString *prop = dyn_cast<MDString>(md->getOperand(i));
52 assert(prop && "Annotation property not a string");
53
54 // value
55 ConstantInt *Val = dyn_cast<ConstantInt>(md->getOperand(i + 1));
56 assert(Val && "Value operand not a constant int");
57
58 std::string keyname = prop->getString().str();
59 if (retval.find(keyname) != retval.end())
60 retval[keyname].push_back(Val->getZExtValue());
61 else {
62 std::vector<unsigned> tmp;
63 tmp.push_back(Val->getZExtValue());
64 retval[keyname] = tmp;
65 }
66 }
67 }
68
cacheAnnotationFromMD(const Module * m,const GlobalValue * gv)69 static void cacheAnnotationFromMD(const Module *m, const GlobalValue *gv) {
70 MutexGuard Guard(Lock);
71 NamedMDNode *NMD = m->getNamedMetadata(llvm::NamedMDForAnnotations);
72 if (!NMD)
73 return;
74 key_val_pair_t tmp;
75 for (unsigned i = 0, e = NMD->getNumOperands(); i != e; ++i) {
76 const MDNode *elem = NMD->getOperand(i);
77
78 Value *entity = elem->getOperand(0);
79 // entity may be null due to DCE
80 if (!entity)
81 continue;
82 if (entity != gv)
83 continue;
84
85 // accumulate annotations for entity in tmp
86 cacheAnnotationFromMD(elem, tmp);
87 }
88
89 if (tmp.empty()) // no annotations for this gv
90 return;
91
92 if ((*annotationCache).find(m) != (*annotationCache).end())
93 (*annotationCache)[m][gv] = tmp;
94 else {
95 global_val_annot_t tmp1;
96 tmp1[gv] = tmp;
97 (*annotationCache)[m] = tmp1;
98 }
99 }
100
findOneNVVMAnnotation(const GlobalValue * gv,std::string prop,unsigned & retval)101 bool llvm::findOneNVVMAnnotation(const GlobalValue *gv, std::string prop,
102 unsigned &retval) {
103 MutexGuard Guard(Lock);
104 const Module *m = gv->getParent();
105 if ((*annotationCache).find(m) == (*annotationCache).end())
106 cacheAnnotationFromMD(m, gv);
107 else if ((*annotationCache)[m].find(gv) == (*annotationCache)[m].end())
108 cacheAnnotationFromMD(m, gv);
109 if ((*annotationCache)[m][gv].find(prop) == (*annotationCache)[m][gv].end())
110 return false;
111 retval = (*annotationCache)[m][gv][prop][0];
112 return true;
113 }
114
findAllNVVMAnnotation(const GlobalValue * gv,std::string prop,std::vector<unsigned> & retval)115 bool llvm::findAllNVVMAnnotation(const GlobalValue *gv, std::string prop,
116 std::vector<unsigned> &retval) {
117 MutexGuard Guard(Lock);
118 const Module *m = gv->getParent();
119 if ((*annotationCache).find(m) == (*annotationCache).end())
120 cacheAnnotationFromMD(m, gv);
121 else if ((*annotationCache)[m].find(gv) == (*annotationCache)[m].end())
122 cacheAnnotationFromMD(m, gv);
123 if ((*annotationCache)[m][gv].find(prop) == (*annotationCache)[m][gv].end())
124 return false;
125 retval = (*annotationCache)[m][gv][prop];
126 return true;
127 }
128
isTexture(const llvm::Value & val)129 bool llvm::isTexture(const llvm::Value &val) {
130 if (const GlobalValue *gv = dyn_cast<GlobalValue>(&val)) {
131 unsigned annot;
132 if (llvm::findOneNVVMAnnotation(
133 gv, llvm::PropertyAnnotationNames[llvm::PROPERTY_ISTEXTURE],
134 annot)) {
135 assert((annot == 1) && "Unexpected annotation on a texture symbol");
136 return true;
137 }
138 }
139 return false;
140 }
141
isSurface(const llvm::Value & val)142 bool llvm::isSurface(const llvm::Value &val) {
143 if (const GlobalValue *gv = dyn_cast<GlobalValue>(&val)) {
144 unsigned annot;
145 if (llvm::findOneNVVMAnnotation(
146 gv, llvm::PropertyAnnotationNames[llvm::PROPERTY_ISSURFACE],
147 annot)) {
148 assert((annot == 1) && "Unexpected annotation on a surface symbol");
149 return true;
150 }
151 }
152 return false;
153 }
154
isSampler(const llvm::Value & val)155 bool llvm::isSampler(const llvm::Value &val) {
156 if (const GlobalValue *gv = dyn_cast<GlobalValue>(&val)) {
157 unsigned annot;
158 if (llvm::findOneNVVMAnnotation(
159 gv, llvm::PropertyAnnotationNames[llvm::PROPERTY_ISSAMPLER],
160 annot)) {
161 assert((annot == 1) && "Unexpected annotation on a sampler symbol");
162 return true;
163 }
164 }
165 if (const Argument *arg = dyn_cast<Argument>(&val)) {
166 const Function *func = arg->getParent();
167 std::vector<unsigned> annot;
168 if (llvm::findAllNVVMAnnotation(
169 func, llvm::PropertyAnnotationNames[llvm::PROPERTY_ISSAMPLER],
170 annot)) {
171 if (std::find(annot.begin(), annot.end(), arg->getArgNo()) != annot.end())
172 return true;
173 }
174 }
175 return false;
176 }
177
isImageReadOnly(const llvm::Value & val)178 bool llvm::isImageReadOnly(const llvm::Value &val) {
179 if (const Argument *arg = dyn_cast<Argument>(&val)) {
180 const Function *func = arg->getParent();
181 std::vector<unsigned> annot;
182 if (llvm::findAllNVVMAnnotation(func,
183 llvm::PropertyAnnotationNames[
184 llvm::PROPERTY_ISREADONLY_IMAGE_PARAM],
185 annot)) {
186 if (std::find(annot.begin(), annot.end(), arg->getArgNo()) != annot.end())
187 return true;
188 }
189 }
190 return false;
191 }
192
isImageWriteOnly(const llvm::Value & val)193 bool llvm::isImageWriteOnly(const llvm::Value &val) {
194 if (const Argument *arg = dyn_cast<Argument>(&val)) {
195 const Function *func = arg->getParent();
196 std::vector<unsigned> annot;
197 if (llvm::findAllNVVMAnnotation(func,
198 llvm::PropertyAnnotationNames[
199 llvm::PROPERTY_ISWRITEONLY_IMAGE_PARAM],
200 annot)) {
201 if (std::find(annot.begin(), annot.end(), arg->getArgNo()) != annot.end())
202 return true;
203 }
204 }
205 return false;
206 }
207
isImageReadWrite(const llvm::Value & val)208 bool llvm::isImageReadWrite(const llvm::Value &val) {
209 if (const Argument *arg = dyn_cast<Argument>(&val)) {
210 const Function *func = arg->getParent();
211 std::vector<unsigned> annot;
212 if (llvm::findAllNVVMAnnotation(func,
213 llvm::PropertyAnnotationNames[
214 llvm::PROPERTY_ISREADWRITE_IMAGE_PARAM],
215 annot)) {
216 if (std::find(annot.begin(), annot.end(), arg->getArgNo()) != annot.end())
217 return true;
218 }
219 }
220 return false;
221 }
222
isImage(const llvm::Value & val)223 bool llvm::isImage(const llvm::Value &val) {
224 return llvm::isImageReadOnly(val) || llvm::isImageWriteOnly(val) ||
225 llvm::isImageReadWrite(val);
226 }
227
isManaged(const llvm::Value & val)228 bool llvm::isManaged(const llvm::Value &val) {
229 if(const GlobalValue *gv = dyn_cast<GlobalValue>(&val)) {
230 unsigned annot;
231 if(llvm::findOneNVVMAnnotation(gv,
232 llvm::PropertyAnnotationNames[llvm::PROPERTY_MANAGED],
233 annot)) {
234 assert((annot == 1) && "Unexpected annotation on a managed symbol");
235 return true;
236 }
237 }
238 return false;
239 }
240
getTextureName(const llvm::Value & val)241 std::string llvm::getTextureName(const llvm::Value &val) {
242 assert(val.hasName() && "Found texture variable with no name");
243 return val.getName();
244 }
245
getSurfaceName(const llvm::Value & val)246 std::string llvm::getSurfaceName(const llvm::Value &val) {
247 assert(val.hasName() && "Found surface variable with no name");
248 return val.getName();
249 }
250
getSamplerName(const llvm::Value & val)251 std::string llvm::getSamplerName(const llvm::Value &val) {
252 assert(val.hasName() && "Found sampler variable with no name");
253 return val.getName();
254 }
255
getMaxNTIDx(const Function & F,unsigned & x)256 bool llvm::getMaxNTIDx(const Function &F, unsigned &x) {
257 return (llvm::findOneNVVMAnnotation(
258 &F, llvm::PropertyAnnotationNames[llvm::PROPERTY_MAXNTID_X], x));
259 }
260
getMaxNTIDy(const Function & F,unsigned & y)261 bool llvm::getMaxNTIDy(const Function &F, unsigned &y) {
262 return (llvm::findOneNVVMAnnotation(
263 &F, llvm::PropertyAnnotationNames[llvm::PROPERTY_MAXNTID_Y], y));
264 }
265
getMaxNTIDz(const Function & F,unsigned & z)266 bool llvm::getMaxNTIDz(const Function &F, unsigned &z) {
267 return (llvm::findOneNVVMAnnotation(
268 &F, llvm::PropertyAnnotationNames[llvm::PROPERTY_MAXNTID_Z], z));
269 }
270
getReqNTIDx(const Function & F,unsigned & x)271 bool llvm::getReqNTIDx(const Function &F, unsigned &x) {
272 return (llvm::findOneNVVMAnnotation(
273 &F, llvm::PropertyAnnotationNames[llvm::PROPERTY_REQNTID_X], x));
274 }
275
getReqNTIDy(const Function & F,unsigned & y)276 bool llvm::getReqNTIDy(const Function &F, unsigned &y) {
277 return (llvm::findOneNVVMAnnotation(
278 &F, llvm::PropertyAnnotationNames[llvm::PROPERTY_REQNTID_Y], y));
279 }
280
getReqNTIDz(const Function & F,unsigned & z)281 bool llvm::getReqNTIDz(const Function &F, unsigned &z) {
282 return (llvm::findOneNVVMAnnotation(
283 &F, llvm::PropertyAnnotationNames[llvm::PROPERTY_REQNTID_Z], z));
284 }
285
getMinCTASm(const Function & F,unsigned & x)286 bool llvm::getMinCTASm(const Function &F, unsigned &x) {
287 return (llvm::findOneNVVMAnnotation(
288 &F, llvm::PropertyAnnotationNames[llvm::PROPERTY_MINNCTAPERSM], x));
289 }
290
isKernelFunction(const Function & F)291 bool llvm::isKernelFunction(const Function &F) {
292 unsigned x = 0;
293 bool retval = llvm::findOneNVVMAnnotation(
294 &F, llvm::PropertyAnnotationNames[llvm::PROPERTY_ISKERNEL_FUNCTION], x);
295 if (retval == false) {
296 // There is no NVVM metadata, check the calling convention
297 if (F.getCallingConv() == llvm::CallingConv::PTX_Kernel)
298 return true;
299 else
300 return false;
301 }
302 return (x == 1);
303 }
304
getAlign(const Function & F,unsigned index,unsigned & align)305 bool llvm::getAlign(const Function &F, unsigned index, unsigned &align) {
306 std::vector<unsigned> Vs;
307 bool retval = llvm::findAllNVVMAnnotation(
308 &F, llvm::PropertyAnnotationNames[llvm::PROPERTY_ALIGN], Vs);
309 if (retval == false)
310 return false;
311 for (int i = 0, e = Vs.size(); i < e; i++) {
312 unsigned v = Vs[i];
313 if ((v >> 16) == index) {
314 align = v & 0xFFFF;
315 return true;
316 }
317 }
318 return false;
319 }
320
getAlign(const CallInst & I,unsigned index,unsigned & align)321 bool llvm::getAlign(const CallInst &I, unsigned index, unsigned &align) {
322 if (MDNode *alignNode = I.getMetadata("callalign")) {
323 for (int i = 0, n = alignNode->getNumOperands(); i < n; i++) {
324 if (const ConstantInt *CI =
325 dyn_cast<ConstantInt>(alignNode->getOperand(i))) {
326 unsigned v = CI->getZExtValue();
327 if ((v >> 16) == index) {
328 align = v & 0xFFFF;
329 return true;
330 }
331 if ((v >> 16) > index) {
332 return false;
333 }
334 }
335 }
336 }
337 return false;
338 }
339
isBarrierIntrinsic(Intrinsic::ID id)340 bool llvm::isBarrierIntrinsic(Intrinsic::ID id) {
341 if ((id == Intrinsic::nvvm_barrier0) ||
342 (id == Intrinsic::nvvm_barrier0_popc) ||
343 (id == Intrinsic::nvvm_barrier0_and) ||
344 (id == Intrinsic::nvvm_barrier0_or) ||
345 (id == Intrinsic::cuda_syncthreads))
346 return true;
347 return false;
348 }
349
350 // Interface for checking all memory space transfer related intrinsics
isMemorySpaceTransferIntrinsic(Intrinsic::ID id)351 bool llvm::isMemorySpaceTransferIntrinsic(Intrinsic::ID id) {
352 if (id == Intrinsic::nvvm_ptr_local_to_gen ||
353 id == Intrinsic::nvvm_ptr_shared_to_gen ||
354 id == Intrinsic::nvvm_ptr_global_to_gen ||
355 id == Intrinsic::nvvm_ptr_constant_to_gen ||
356 id == Intrinsic::nvvm_ptr_gen_to_global ||
357 id == Intrinsic::nvvm_ptr_gen_to_shared ||
358 id == Intrinsic::nvvm_ptr_gen_to_local ||
359 id == Intrinsic::nvvm_ptr_gen_to_constant ||
360 id == Intrinsic::nvvm_ptr_gen_to_param) {
361 return true;
362 }
363
364 return false;
365 }
366
367 // consider several special intrinsics in striping pointer casts, and
368 // provide an option to ignore GEP indicies for find out the base address only
369 // which could be used in simple alias disambigurate.
370 const Value *
skipPointerTransfer(const Value * V,bool ignore_GEP_indices)371 llvm::skipPointerTransfer(const Value *V, bool ignore_GEP_indices) {
372 V = V->stripPointerCasts();
373 while (true) {
374 if (const IntrinsicInst *IS = dyn_cast<IntrinsicInst>(V)) {
375 if (isMemorySpaceTransferIntrinsic(IS->getIntrinsicID())) {
376 V = IS->getArgOperand(0)->stripPointerCasts();
377 continue;
378 }
379 } else if (ignore_GEP_indices)
380 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
381 V = GEP->getPointerOperand()->stripPointerCasts();
382 continue;
383 }
384 break;
385 }
386 return V;
387 }
388
389 // consider several special intrinsics in striping pointer casts, and
390 // - ignore GEP indicies for find out the base address only, and
391 // - tracking PHINode
392 // which could be used in simple alias disambigurate.
393 const Value *
skipPointerTransfer(const Value * V,std::set<const Value * > & processed)394 llvm::skipPointerTransfer(const Value *V, std::set<const Value *> &processed) {
395 if (processed.find(V) != processed.end())
396 return nullptr;
397 processed.insert(V);
398
399 const Value *V2 = V->stripPointerCasts();
400 if (V2 != V && processed.find(V2) != processed.end())
401 return nullptr;
402 processed.insert(V2);
403
404 V = V2;
405
406 while (true) {
407 if (const IntrinsicInst *IS = dyn_cast<IntrinsicInst>(V)) {
408 if (isMemorySpaceTransferIntrinsic(IS->getIntrinsicID())) {
409 V = IS->getArgOperand(0)->stripPointerCasts();
410 continue;
411 }
412 } else if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
413 V = GEP->getPointerOperand()->stripPointerCasts();
414 continue;
415 } else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
416 if (V != V2 && processed.find(V) != processed.end())
417 return nullptr;
418 processed.insert(PN);
419 const Value *common = nullptr;
420 for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) {
421 const Value *pv = PN->getIncomingValue(i);
422 const Value *base = skipPointerTransfer(pv, processed);
423 if (base) {
424 if (!common)
425 common = base;
426 else if (common != base)
427 return PN;
428 }
429 }
430 if (!common)
431 return PN;
432 V = common;
433 }
434 break;
435 }
436 return V;
437 }
438
439 // The following are some useful utilities for debuggung
440
getParentBlock(Value * v)441 BasicBlock *llvm::getParentBlock(Value *v) {
442 if (BasicBlock *B = dyn_cast<BasicBlock>(v))
443 return B;
444
445 if (Instruction *I = dyn_cast<Instruction>(v))
446 return I->getParent();
447
448 return nullptr;
449 }
450
getParentFunction(Value * v)451 Function *llvm::getParentFunction(Value *v) {
452 if (Function *F = dyn_cast<Function>(v))
453 return F;
454
455 if (Instruction *I = dyn_cast<Instruction>(v))
456 return I->getParent()->getParent();
457
458 if (BasicBlock *B = dyn_cast<BasicBlock>(v))
459 return B->getParent();
460
461 return nullptr;
462 }
463
464 // Dump a block by name
dumpBlock(Value * v,char * blockName)465 void llvm::dumpBlock(Value *v, char *blockName) {
466 Function *F = getParentFunction(v);
467 if (!F)
468 return;
469
470 for (Function::iterator it = F->begin(), ie = F->end(); it != ie; ++it) {
471 BasicBlock *B = it;
472 if (strcmp(B->getName().data(), blockName) == 0) {
473 B->dump();
474 return;
475 }
476 }
477 }
478
479 // Find an instruction by name
getInst(Value * base,char * instName)480 Instruction *llvm::getInst(Value *base, char *instName) {
481 Function *F = getParentFunction(base);
482 if (!F)
483 return nullptr;
484
485 for (inst_iterator it = inst_begin(F), ie = inst_end(F); it != ie; ++it) {
486 Instruction *I = &*it;
487 if (strcmp(I->getName().data(), instName) == 0) {
488 return I;
489 }
490 }
491
492 return nullptr;
493 }
494
495 // Dump an instruction by nane
dumpInst(Value * base,char * instName)496 void llvm::dumpInst(Value *base, char *instName) {
497 Instruction *I = getInst(base, instName);
498 if (I)
499 I->dump();
500 }
501
502 // Dump an instruction and all dependent instructions
dumpInstRec(Value * v,std::set<Instruction * > * visited)503 void llvm::dumpInstRec(Value *v, std::set<Instruction *> *visited) {
504 if (Instruction *I = dyn_cast<Instruction>(v)) {
505
506 if (visited->find(I) != visited->end())
507 return;
508
509 visited->insert(I);
510
511 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
512 dumpInstRec(I->getOperand(i), visited);
513
514 I->dump();
515 }
516 }
517
518 // Dump an instruction and all dependent instructions
dumpInstRec(Value * v)519 void llvm::dumpInstRec(Value *v) {
520 std::set<Instruction *> visited;
521
522 //BasicBlock *B = getParentBlock(v);
523
524 dumpInstRec(v, &visited);
525 }
526
527 // Dump the parent for Instruction, block or function
dumpParent(Value * v)528 void llvm::dumpParent(Value *v) {
529 if (Instruction *I = dyn_cast<Instruction>(v)) {
530 I->getParent()->dump();
531 return;
532 }
533
534 if (BasicBlock *B = dyn_cast<BasicBlock>(v)) {
535 B->getParent()->dump();
536 return;
537 }
538
539 if (Function *F = dyn_cast<Function>(v)) {
540 F->getParent()->dump();
541 return;
542 }
543 }
544