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