| 1 | //===- NVPTXUtilities.cpp - Utility Functions -----------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file contains miscellaneous utility functions |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "NVPTXUtilities.h" |
| 14 | #include "NVPTX.h" |
| 15 | #include "NVPTXTargetMachine.h" |
| 16 | #include "llvm/ADT/ArrayRef.h" |
| 17 | #include "llvm/ADT/SmallVector.h" |
| 18 | #include "llvm/ADT/StringRef.h" |
| 19 | #include "llvm/IR/Argument.h" |
| 20 | #include "llvm/IR/Constants.h" |
| 21 | #include "llvm/IR/Function.h" |
| 22 | #include "llvm/IR/GlobalVariable.h" |
| 23 | #include "llvm/IR/Module.h" |
| 24 | #include "llvm/Support/Alignment.h" |
| 25 | #include "llvm/Support/ModRef.h" |
| 26 | #include "llvm/Support/Mutex.h" |
| 27 | #include <cstdint> |
| 28 | #include <cstring> |
| 29 | #include <map> |
| 30 | #include <mutex> |
| 31 | #include <optional> |
| 32 | #include <string> |
| 33 | #include <vector> |
| 34 | |
| 35 | namespace llvm { |
| 36 | |
| 37 | namespace { |
| 38 | typedef std::map<std::string, std::vector<unsigned>> key_val_pair_t; |
| 39 | typedef std::map<const GlobalValue *, key_val_pair_t> global_val_annot_t; |
| 40 | |
| 41 | struct AnnotationCache { |
| 42 | sys::Mutex Lock; |
| 43 | std::map<const Module *, global_val_annot_t> Cache; |
| 44 | }; |
| 45 | |
| 46 | AnnotationCache &getAnnotationCache() { |
| 47 | static AnnotationCache AC; |
| 48 | return AC; |
| 49 | } |
| 50 | } // anonymous namespace |
| 51 | |
| 52 | void clearAnnotationCache(const Module *Mod) { |
| 53 | auto &AC = getAnnotationCache(); |
| 54 | std::lock_guard<sys::Mutex> Guard(AC.Lock); |
| 55 | AC.Cache.erase(x: Mod); |
| 56 | } |
| 57 | |
| 58 | static void cacheAnnotationFromMD(const MDNode *MetadataNode, |
| 59 | key_val_pair_t &retval) { |
| 60 | auto &AC = getAnnotationCache(); |
| 61 | std::lock_guard<sys::Mutex> Guard(AC.Lock); |
| 62 | assert(MetadataNode && "Invalid mdnode for annotation" ); |
| 63 | assert((MetadataNode->getNumOperands() % 2) == 1 && |
| 64 | "Invalid number of operands" ); |
| 65 | // start index = 1, to skip the global variable key |
| 66 | // increment = 2, to skip the value for each property-value pairs |
| 67 | for (unsigned i = 1, e = MetadataNode->getNumOperands(); i != e; i += 2) { |
| 68 | // property |
| 69 | const MDString *prop = dyn_cast<MDString>(Val: MetadataNode->getOperand(I: i)); |
| 70 | assert(prop && "Annotation property not a string" ); |
| 71 | std::string Key = prop->getString().str(); |
| 72 | |
| 73 | // value |
| 74 | if (ConstantInt *Val = mdconst::dyn_extract<ConstantInt>( |
| 75 | MD: MetadataNode->getOperand(I: i + 1))) { |
| 76 | retval[Key].push_back(x: Val->getZExtValue()); |
| 77 | } else { |
| 78 | llvm_unreachable("Value operand not a constant int" ); |
| 79 | } |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | static void cacheAnnotationFromMD(const Module *m, const GlobalValue *gv) { |
| 84 | auto &AC = getAnnotationCache(); |
| 85 | std::lock_guard<sys::Mutex> Guard(AC.Lock); |
| 86 | NamedMDNode *NMD = m->getNamedMetadata(Name: "nvvm.annotations" ); |
| 87 | if (!NMD) |
| 88 | return; |
| 89 | key_val_pair_t tmp; |
| 90 | for (unsigned i = 0, e = NMD->getNumOperands(); i != e; ++i) { |
| 91 | const MDNode *elem = NMD->getOperand(i); |
| 92 | |
| 93 | GlobalValue *entity = |
| 94 | mdconst::dyn_extract_or_null<GlobalValue>(MD: elem->getOperand(I: 0)); |
| 95 | // entity may be null due to DCE |
| 96 | if (!entity) |
| 97 | continue; |
| 98 | if (entity != gv) |
| 99 | continue; |
| 100 | |
| 101 | // accumulate annotations for entity in tmp |
| 102 | cacheAnnotationFromMD(MetadataNode: elem, retval&: tmp); |
| 103 | } |
| 104 | |
| 105 | if (tmp.empty()) // no annotations for this gv |
| 106 | return; |
| 107 | |
| 108 | AC.Cache[m][gv] = std::move(tmp); |
| 109 | } |
| 110 | |
| 111 | static std::optional<unsigned> findOneNVVMAnnotation(const GlobalValue *gv, |
| 112 | const std::string &prop) { |
| 113 | auto &AC = getAnnotationCache(); |
| 114 | std::lock_guard<sys::Mutex> Guard(AC.Lock); |
| 115 | const Module *m = gv->getParent(); |
| 116 | auto ACIt = AC.Cache.find(x: m); |
| 117 | if (ACIt == AC.Cache.end()) |
| 118 | cacheAnnotationFromMD(m, gv); |
| 119 | else if (ACIt->second.find(x: gv) == ACIt->second.end()) |
| 120 | cacheAnnotationFromMD(m, gv); |
| 121 | // Look up AC.Cache[m][gv] again because cacheAnnotationFromMD may have |
| 122 | // inserted the entry. |
| 123 | auto &KVP = AC.Cache[m][gv]; |
| 124 | auto It = KVP.find(x: prop); |
| 125 | if (It == KVP.end()) |
| 126 | return std::nullopt; |
| 127 | return It->second[0]; |
| 128 | } |
| 129 | |
| 130 | static bool findAllNVVMAnnotation(const GlobalValue *gv, |
| 131 | const std::string &prop, |
| 132 | std::vector<unsigned> &retval) { |
| 133 | auto &AC = getAnnotationCache(); |
| 134 | std::lock_guard<sys::Mutex> Guard(AC.Lock); |
| 135 | const Module *m = gv->getParent(); |
| 136 | auto ACIt = AC.Cache.find(x: m); |
| 137 | if (ACIt == AC.Cache.end()) |
| 138 | cacheAnnotationFromMD(m, gv); |
| 139 | else if (ACIt->second.find(x: gv) == ACIt->second.end()) |
| 140 | cacheAnnotationFromMD(m, gv); |
| 141 | // Look up AC.Cache[m][gv] again because cacheAnnotationFromMD may have |
| 142 | // inserted the entry. |
| 143 | auto &KVP = AC.Cache[m][gv]; |
| 144 | auto It = KVP.find(x: prop); |
| 145 | if (It == KVP.end()) |
| 146 | return false; |
| 147 | retval = It->second; |
| 148 | return true; |
| 149 | } |
| 150 | |
| 151 | static bool globalHasNVVMAnnotation(const Value &V, const std::string &Prop) { |
| 152 | if (const auto *GV = dyn_cast<GlobalValue>(Val: &V)) |
| 153 | if (const auto Annot = findOneNVVMAnnotation(gv: GV, prop: Prop)) { |
| 154 | assert((*Annot == 1) && "Unexpected annotation on a symbol" ); |
| 155 | return true; |
| 156 | } |
| 157 | |
| 158 | return false; |
| 159 | } |
| 160 | |
| 161 | static bool argHasNVVMAnnotation(const Value &Val, |
| 162 | const std::string &Annotation) { |
| 163 | if (const Argument *Arg = dyn_cast<Argument>(Val: &Val)) { |
| 164 | const Function *Func = Arg->getParent(); |
| 165 | std::vector<unsigned> Annot; |
| 166 | if (findAllNVVMAnnotation(gv: Func, prop: Annotation, retval&: Annot)) { |
| 167 | if (is_contained(Range&: Annot, Element: Arg->getArgNo())) |
| 168 | return true; |
| 169 | } |
| 170 | } |
| 171 | return false; |
| 172 | } |
| 173 | |
| 174 | static std::optional<unsigned> getFnAttrParsedInt(const Function &F, |
| 175 | StringRef Attr) { |
| 176 | return F.hasFnAttribute(Kind: Attr) |
| 177 | ? std::optional(F.getFnAttributeAsParsedInteger(Kind: Attr)) |
| 178 | : std::nullopt; |
| 179 | } |
| 180 | |
| 181 | static SmallVector<unsigned, 3> getFnAttrParsedVector(const Function &F, |
| 182 | StringRef Attr) { |
| 183 | SmallVector<unsigned, 3> V; |
| 184 | auto &Ctx = F.getContext(); |
| 185 | |
| 186 | if (F.hasFnAttribute(Kind: Attr)) { |
| 187 | // We expect the attribute value to be of the form "x[,y[,z]]", where x, y, |
| 188 | // and z are unsigned values. |
| 189 | StringRef S = F.getFnAttribute(Kind: Attr).getValueAsString(); |
| 190 | for (unsigned I = 0; I < 3 && !S.empty(); I++) { |
| 191 | auto [First, Rest] = S.split(Separator: "," ); |
| 192 | unsigned IntVal; |
| 193 | if (First.trim().getAsInteger(Radix: 0, Result&: IntVal)) |
| 194 | Ctx.emitError(ErrorStr: "can't parse integer attribute " + First + " in " + Attr); |
| 195 | |
| 196 | V.push_back(Elt: IntVal); |
| 197 | S = Rest; |
| 198 | } |
| 199 | } |
| 200 | return V; |
| 201 | } |
| 202 | |
| 203 | static std::optional<uint64_t> getVectorProduct(ArrayRef<unsigned> V) { |
| 204 | if (V.empty()) |
| 205 | return std::nullopt; |
| 206 | |
| 207 | return std::accumulate(first: V.begin(), last: V.end(), init: 1, binary_op: std::multiplies<uint64_t>{}); |
| 208 | } |
| 209 | |
| 210 | bool isParamGridConstant(const Argument &Arg) { |
| 211 | assert(isKernelFunction(*Arg.getParent()) && |
| 212 | "only kernel arguments can be grid_constant" ); |
| 213 | |
| 214 | if (!Arg.hasByValAttr()) |
| 215 | return false; |
| 216 | |
| 217 | // Lowering an argument as a grid_constant violates the byval semantics (and |
| 218 | // the C++ API) by reusing the same memory location for the argument across |
| 219 | // multiple threads. If an argument doesn't read memory and its address is not |
| 220 | // captured (its address is not compared with any value), then the tweak of |
| 221 | // the C++ API and byval semantics is unobservable by the program and we can |
| 222 | // lower the arg as a grid_constant. |
| 223 | if (Arg.onlyReadsMemory()) { |
| 224 | const auto CI = Arg.getAttributes().getCaptureInfo(); |
| 225 | if (!capturesAddress(CC: CI) && !capturesFullProvenance(CC: CI)) |
| 226 | return true; |
| 227 | } |
| 228 | |
| 229 | // "grid_constant" counts argument indices starting from 1 |
| 230 | if (Arg.hasAttribute(Kind: "nvvm.grid_constant" )) |
| 231 | return true; |
| 232 | |
| 233 | return false; |
| 234 | } |
| 235 | |
| 236 | bool isTexture(const Value &V) { return globalHasNVVMAnnotation(V, Prop: "texture" ); } |
| 237 | |
| 238 | bool isSurface(const Value &V) { return globalHasNVVMAnnotation(V, Prop: "surface" ); } |
| 239 | |
| 240 | bool isSampler(const Value &V) { |
| 241 | const char *AnnotationName = "sampler" ; |
| 242 | |
| 243 | return globalHasNVVMAnnotation(V, Prop: AnnotationName) || |
| 244 | argHasNVVMAnnotation(Val: V, Annotation: AnnotationName); |
| 245 | } |
| 246 | |
| 247 | bool isImageReadOnly(const Value &V) { |
| 248 | return argHasNVVMAnnotation(Val: V, Annotation: "rdoimage" ); |
| 249 | } |
| 250 | |
| 251 | bool isImageWriteOnly(const Value &V) { |
| 252 | return argHasNVVMAnnotation(Val: V, Annotation: "wroimage" ); |
| 253 | } |
| 254 | |
| 255 | bool isImageReadWrite(const Value &V) { |
| 256 | return argHasNVVMAnnotation(Val: V, Annotation: "rdwrimage" ); |
| 257 | } |
| 258 | |
| 259 | bool isImage(const Value &V) { |
| 260 | return isImageReadOnly(V) || isImageWriteOnly(V) || isImageReadWrite(V); |
| 261 | } |
| 262 | |
| 263 | bool isManaged(const Value &V) { return globalHasNVVMAnnotation(V, Prop: "managed" ); } |
| 264 | |
| 265 | StringRef getTextureName(const Value &V) { |
| 266 | assert(V.hasName() && "Found texture variable with no name" ); |
| 267 | return V.getName(); |
| 268 | } |
| 269 | |
| 270 | StringRef getSurfaceName(const Value &V) { |
| 271 | assert(V.hasName() && "Found surface variable with no name" ); |
| 272 | return V.getName(); |
| 273 | } |
| 274 | |
| 275 | StringRef getSamplerName(const Value &V) { |
| 276 | assert(V.hasName() && "Found sampler variable with no name" ); |
| 277 | return V.getName(); |
| 278 | } |
| 279 | |
| 280 | SmallVector<unsigned, 3> getMaxNTID(const Function &F) { |
| 281 | return getFnAttrParsedVector(F, Attr: "nvvm.maxntid" ); |
| 282 | } |
| 283 | |
| 284 | SmallVector<unsigned, 3> getReqNTID(const Function &F) { |
| 285 | return getFnAttrParsedVector(F, Attr: "nvvm.reqntid" ); |
| 286 | } |
| 287 | |
| 288 | SmallVector<unsigned, 3> getClusterDim(const Function &F) { |
| 289 | return getFnAttrParsedVector(F, Attr: "nvvm.cluster_dim" ); |
| 290 | } |
| 291 | |
| 292 | std::optional<uint64_t> getOverallMaxNTID(const Function &F) { |
| 293 | // Note: The semantics here are a bit strange. The PTX ISA states the |
| 294 | // following (11.4.2. Performance-Tuning Directives: .maxntid): |
| 295 | // |
| 296 | // Note that this directive guarantees that the total number of threads does |
| 297 | // not exceed the maximum, but does not guarantee that the limit in any |
| 298 | // particular dimension is not exceeded. |
| 299 | const auto MaxNTID = getMaxNTID(F); |
| 300 | return getVectorProduct(V: MaxNTID); |
| 301 | } |
| 302 | |
| 303 | std::optional<uint64_t> getOverallReqNTID(const Function &F) { |
| 304 | // Note: The semantics here are a bit strange. See getMaxNTID. |
| 305 | const auto ReqNTID = getReqNTID(F); |
| 306 | return getVectorProduct(V: ReqNTID); |
| 307 | } |
| 308 | |
| 309 | std::optional<uint64_t> getOverallClusterRank(const Function &F) { |
| 310 | // maxclusterrank and cluster_dim are mutually exclusive. |
| 311 | if (const auto ClusterRank = getMaxClusterRank(F)) |
| 312 | return ClusterRank; |
| 313 | |
| 314 | // Note: The semantics here are a bit strange. See getMaxNTID. |
| 315 | const auto ClusterDim = getClusterDim(F); |
| 316 | return getVectorProduct(V: ClusterDim); |
| 317 | } |
| 318 | |
| 319 | std::optional<unsigned> getMaxClusterRank(const Function &F) { |
| 320 | return getFnAttrParsedInt(F, Attr: "nvvm.maxclusterrank" ); |
| 321 | } |
| 322 | |
| 323 | std::optional<unsigned> getMinCTASm(const Function &F) { |
| 324 | return getFnAttrParsedInt(F, Attr: "nvvm.minctasm" ); |
| 325 | } |
| 326 | |
| 327 | std::optional<unsigned> getMaxNReg(const Function &F) { |
| 328 | return getFnAttrParsedInt(F, Attr: "nvvm.maxnreg" ); |
| 329 | } |
| 330 | |
| 331 | bool hasBlocksAreClusters(const Function &F) { |
| 332 | return F.hasFnAttribute(Kind: "nvvm.blocksareclusters" ); |
| 333 | } |
| 334 | |
| 335 | MaybeAlign getAlign(const CallInst &I, unsigned Index) { |
| 336 | // First check the alignstack metadata |
| 337 | if (MaybeAlign StackAlign = |
| 338 | I.getAttributes().getAttributes(Index).getStackAlignment()) |
| 339 | return StackAlign; |
| 340 | |
| 341 | // If that is missing, check the legacy nvvm metadata |
| 342 | if (MDNode *alignNode = I.getMetadata(Kind: "callalign" )) { |
| 343 | for (int i = 0, n = alignNode->getNumOperands(); i < n; i++) { |
| 344 | if (const ConstantInt *CI = |
| 345 | mdconst::dyn_extract<ConstantInt>(MD: alignNode->getOperand(I: i))) { |
| 346 | unsigned V = CI->getZExtValue(); |
| 347 | if ((V >> 16) == Index) |
| 348 | return Align(V & 0xFFFF); |
| 349 | if ((V >> 16) > Index) |
| 350 | return std::nullopt; |
| 351 | } |
| 352 | } |
| 353 | } |
| 354 | return std::nullopt; |
| 355 | } |
| 356 | |
| 357 | Function *getMaybeBitcastedCallee(const CallBase *CB) { |
| 358 | return dyn_cast<Function>(Val: CB->getCalledOperand()->stripPointerCasts()); |
| 359 | } |
| 360 | |
| 361 | bool shouldEmitPTXNoReturn(const Value *V, const TargetMachine &TM) { |
| 362 | const auto &ST = |
| 363 | *static_cast<const NVPTXTargetMachine &>(TM).getSubtargetImpl(); |
| 364 | if (!ST.hasNoReturn()) |
| 365 | return false; |
| 366 | |
| 367 | assert((isa<Function>(V) || isa<CallInst>(V)) && |
| 368 | "Expect either a call instruction or a function" ); |
| 369 | |
| 370 | if (const CallInst *CallI = dyn_cast<CallInst>(Val: V)) |
| 371 | return CallI->doesNotReturn() && |
| 372 | CallI->getFunctionType()->getReturnType()->isVoidTy(); |
| 373 | |
| 374 | const Function *F = cast<Function>(Val: V); |
| 375 | return F->doesNotReturn() && |
| 376 | F->getFunctionType()->getReturnType()->isVoidTy() && |
| 377 | !isKernelFunction(F: *F); |
| 378 | } |
| 379 | |
| 380 | } // namespace llvm |
| 381 | |