[RISCV] Fix mgather -> riscv.masked.strided.load combine not extending indices (...
[llvm-project.git] / llvm / lib / Analysis / BlockFrequencyInfo.cpp
blob96c9bfa0e372cedcd1a6b2cf336c17d22961eef8
1 //===- BlockFrequencyInfo.cpp - Block Frequency Analysis ------------------===//
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 // Loops should be simplified before this analysis.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/Analysis/BlockFrequencyInfo.h"
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/iterator.h"
16 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
17 #include "llvm/Analysis/BranchProbabilityInfo.h"
18 #include "llvm/Analysis/LoopInfo.h"
19 #include "llvm/IR/CFG.h"
20 #include "llvm/IR/Function.h"
21 #include "llvm/IR/PassManager.h"
22 #include "llvm/InitializePasses.h"
23 #include "llvm/Pass.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/GraphWriter.h"
26 #include "llvm/Support/raw_ostream.h"
27 #include <cassert>
28 #include <optional>
29 #include <string>
31 using namespace llvm;
33 #define DEBUG_TYPE "block-freq"
35 static cl::opt<GVDAGType> ViewBlockFreqPropagationDAG(
36 "view-block-freq-propagation-dags", cl::Hidden,
37 cl::desc("Pop up a window to show a dag displaying how block "
38 "frequencies propagation through the CFG."),
39 cl::values(clEnumValN(GVDT_None, "none", "do not display graphs."),
40 clEnumValN(GVDT_Fraction, "fraction",
41 "display a graph using the "
42 "fractional block frequency representation."),
43 clEnumValN(GVDT_Integer, "integer",
44 "display a graph using the raw "
45 "integer fractional block frequency representation."),
46 clEnumValN(GVDT_Count, "count", "display a graph using the real "
47 "profile count if available.")));
49 namespace llvm {
50 cl::opt<std::string>
51 ViewBlockFreqFuncName("view-bfi-func-name", cl::Hidden,
52 cl::desc("The option to specify "
53 "the name of the function "
54 "whose CFG will be displayed."));
56 cl::opt<unsigned>
57 ViewHotFreqPercent("view-hot-freq-percent", cl::init(10), cl::Hidden,
58 cl::desc("An integer in percent used to specify "
59 "the hot blocks/edges to be displayed "
60 "in red: a block or edge whose frequency "
61 "is no less than the max frequency of the "
62 "function multiplied by this percent."));
64 // Command line option to turn on CFG dot or text dump after profile annotation.
65 cl::opt<PGOViewCountsType> PGOViewCounts(
66 "pgo-view-counts", cl::Hidden,
67 cl::desc("A boolean option to show CFG dag or text with "
68 "block profile counts and branch probabilities "
69 "right after PGO profile annotation step. The "
70 "profile counts are computed using branch "
71 "probabilities from the runtime profile data and "
72 "block frequency propagation algorithm. To view "
73 "the raw counts from the profile, use option "
74 "-pgo-view-raw-counts instead. To limit graph "
75 "display to only one function, use filtering option "
76 "-view-bfi-func-name."),
77 cl::values(clEnumValN(PGOVCT_None, "none", "do not show."),
78 clEnumValN(PGOVCT_Graph, "graph", "show a graph."),
79 clEnumValN(PGOVCT_Text, "text", "show in text.")));
81 static cl::opt<bool> PrintBFI("print-bfi", cl::init(false), cl::Hidden,
82 cl::desc("Print the block frequency info."));
84 cl::opt<std::string>
85 PrintBFIFuncName("print-bfi-func-name", cl::Hidden,
86 cl::desc("The option to specify the name of the function "
87 "whose block frequency info is printed."));
88 } // namespace llvm
90 namespace llvm {
92 static GVDAGType getGVDT() {
93 if (PGOViewCounts == PGOVCT_Graph)
94 return GVDT_Count;
95 return ViewBlockFreqPropagationDAG;
98 template <>
99 struct GraphTraits<BlockFrequencyInfo *> {
100 using NodeRef = const BasicBlock *;
101 using ChildIteratorType = const_succ_iterator;
102 using nodes_iterator = pointer_iterator<Function::const_iterator>;
104 static NodeRef getEntryNode(const BlockFrequencyInfo *G) {
105 return &G->getFunction()->front();
108 static ChildIteratorType child_begin(const NodeRef N) {
109 return succ_begin(N);
112 static ChildIteratorType child_end(const NodeRef N) { return succ_end(N); }
114 static nodes_iterator nodes_begin(const BlockFrequencyInfo *G) {
115 return nodes_iterator(G->getFunction()->begin());
118 static nodes_iterator nodes_end(const BlockFrequencyInfo *G) {
119 return nodes_iterator(G->getFunction()->end());
123 using BFIDOTGTraitsBase =
124 BFIDOTGraphTraitsBase<BlockFrequencyInfo, BranchProbabilityInfo>;
126 template <>
127 struct DOTGraphTraits<BlockFrequencyInfo *> : public BFIDOTGTraitsBase {
128 explicit DOTGraphTraits(bool isSimple = false)
129 : BFIDOTGTraitsBase(isSimple) {}
131 std::string getNodeLabel(const BasicBlock *Node,
132 const BlockFrequencyInfo *Graph) {
134 return BFIDOTGTraitsBase::getNodeLabel(Node, Graph, getGVDT());
137 std::string getNodeAttributes(const BasicBlock *Node,
138 const BlockFrequencyInfo *Graph) {
139 return BFIDOTGTraitsBase::getNodeAttributes(Node, Graph,
140 ViewHotFreqPercent);
143 std::string getEdgeAttributes(const BasicBlock *Node, EdgeIter EI,
144 const BlockFrequencyInfo *BFI) {
145 return BFIDOTGTraitsBase::getEdgeAttributes(Node, EI, BFI, BFI->getBPI(),
146 ViewHotFreqPercent);
150 } // end namespace llvm
152 BlockFrequencyInfo::BlockFrequencyInfo() = default;
154 BlockFrequencyInfo::BlockFrequencyInfo(const Function &F,
155 const BranchProbabilityInfo &BPI,
156 const LoopInfo &LI) {
157 calculate(F, BPI, LI);
160 BlockFrequencyInfo::BlockFrequencyInfo(BlockFrequencyInfo &&Arg)
161 : BFI(std::move(Arg.BFI)) {}
163 BlockFrequencyInfo &BlockFrequencyInfo::operator=(BlockFrequencyInfo &&RHS) {
164 releaseMemory();
165 BFI = std::move(RHS.BFI);
166 return *this;
169 // Explicitly define the default constructor otherwise it would be implicitly
170 // defined at the first ODR-use which is the BFI member in the
171 // LazyBlockFrequencyInfo header. The dtor needs the BlockFrequencyInfoImpl
172 // template instantiated which is not available in the header.
173 BlockFrequencyInfo::~BlockFrequencyInfo() = default;
175 bool BlockFrequencyInfo::invalidate(Function &F, const PreservedAnalyses &PA,
176 FunctionAnalysisManager::Invalidator &) {
177 // Check whether the analysis, all analyses on functions, or the function's
178 // CFG have been preserved.
179 auto PAC = PA.getChecker<BlockFrequencyAnalysis>();
180 return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() ||
181 PAC.preservedSet<CFGAnalyses>());
184 void BlockFrequencyInfo::calculate(const Function &F,
185 const BranchProbabilityInfo &BPI,
186 const LoopInfo &LI) {
187 if (!BFI)
188 BFI.reset(new ImplType);
189 BFI->calculate(F, BPI, LI);
190 if (ViewBlockFreqPropagationDAG != GVDT_None &&
191 (ViewBlockFreqFuncName.empty() ||
192 F.getName().equals(ViewBlockFreqFuncName))) {
193 view();
195 if (PrintBFI &&
196 (PrintBFIFuncName.empty() || F.getName().equals(PrintBFIFuncName))) {
197 print(dbgs());
201 BlockFrequency BlockFrequencyInfo::getBlockFreq(const BasicBlock *BB) const {
202 return BFI ? BFI->getBlockFreq(BB) : BlockFrequency(0);
205 std::optional<uint64_t>
206 BlockFrequencyInfo::getBlockProfileCount(const BasicBlock *BB,
207 bool AllowSynthetic) const {
208 if (!BFI)
209 return std::nullopt;
211 return BFI->getBlockProfileCount(*getFunction(), BB, AllowSynthetic);
214 std::optional<uint64_t>
215 BlockFrequencyInfo::getProfileCountFromFreq(BlockFrequency Freq) const {
216 if (!BFI)
217 return std::nullopt;
218 return BFI->getProfileCountFromFreq(*getFunction(), Freq);
221 bool BlockFrequencyInfo::isIrrLoopHeader(const BasicBlock *BB) {
222 assert(BFI && "Expected analysis to be available");
223 return BFI->isIrrLoopHeader(BB);
226 void BlockFrequencyInfo::setBlockFreq(const BasicBlock *BB,
227 BlockFrequency Freq) {
228 assert(BFI && "Expected analysis to be available");
229 BFI->setBlockFreq(BB, Freq);
232 void BlockFrequencyInfo::setBlockFreqAndScale(
233 const BasicBlock *ReferenceBB, BlockFrequency Freq,
234 SmallPtrSetImpl<BasicBlock *> &BlocksToScale) {
235 assert(BFI && "Expected analysis to be available");
236 // Use 128 bits APInt to avoid overflow.
237 APInt NewFreq(128, Freq.getFrequency());
238 APInt OldFreq(128, BFI->getBlockFreq(ReferenceBB).getFrequency());
239 APInt BBFreq(128, 0);
240 for (auto *BB : BlocksToScale) {
241 BBFreq = BFI->getBlockFreq(BB).getFrequency();
242 // Multiply first by NewFreq and then divide by OldFreq
243 // to minimize loss of precision.
244 BBFreq *= NewFreq;
245 // udiv is an expensive operation in the general case. If this ends up being
246 // a hot spot, one of the options proposed in
247 // https://reviews.llvm.org/D28535#650071 could be used to avoid this.
248 BBFreq = BBFreq.udiv(OldFreq);
249 BFI->setBlockFreq(BB, BlockFrequency(BBFreq.getLimitedValue()));
251 BFI->setBlockFreq(ReferenceBB, Freq);
254 /// Pop up a ghostview window with the current block frequency propagation
255 /// rendered using dot.
256 void BlockFrequencyInfo::view(StringRef title) const {
257 ViewGraph(const_cast<BlockFrequencyInfo *>(this), title);
260 const Function *BlockFrequencyInfo::getFunction() const {
261 return BFI ? BFI->getFunction() : nullptr;
264 const BranchProbabilityInfo *BlockFrequencyInfo::getBPI() const {
265 return BFI ? &BFI->getBPI() : nullptr;
268 BlockFrequency BlockFrequencyInfo::getEntryFreq() const {
269 return BFI ? BFI->getEntryFreq() : BlockFrequency(0);
272 void BlockFrequencyInfo::releaseMemory() { BFI.reset(); }
274 void BlockFrequencyInfo::print(raw_ostream &OS) const {
275 if (BFI)
276 BFI->print(OS);
279 void BlockFrequencyInfo::verifyMatch(BlockFrequencyInfo &Other) const {
280 if (BFI)
281 BFI->verifyMatch(*Other.BFI);
284 Printable llvm::printBlockFreq(const BlockFrequencyInfo &BFI,
285 BlockFrequency Freq) {
286 return Printable([&BFI, Freq](raw_ostream &OS) {
287 printBlockFreqImpl(OS, BFI.getEntryFreq(), Freq);
291 Printable llvm::printBlockFreq(const BlockFrequencyInfo &BFI,
292 const BasicBlock &BB) {
293 return printBlockFreq(BFI, BFI.getBlockFreq(&BB));
296 INITIALIZE_PASS_BEGIN(BlockFrequencyInfoWrapperPass, "block-freq",
297 "Block Frequency Analysis", true, true)
298 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
299 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
300 INITIALIZE_PASS_END(BlockFrequencyInfoWrapperPass, "block-freq",
301 "Block Frequency Analysis", true, true)
303 char BlockFrequencyInfoWrapperPass::ID = 0;
305 BlockFrequencyInfoWrapperPass::BlockFrequencyInfoWrapperPass()
306 : FunctionPass(ID) {
307 initializeBlockFrequencyInfoWrapperPassPass(*PassRegistry::getPassRegistry());
310 BlockFrequencyInfoWrapperPass::~BlockFrequencyInfoWrapperPass() = default;
312 void BlockFrequencyInfoWrapperPass::print(raw_ostream &OS,
313 const Module *) const {
314 BFI.print(OS);
317 void BlockFrequencyInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
318 AU.addRequired<BranchProbabilityInfoWrapperPass>();
319 AU.addRequired<LoopInfoWrapperPass>();
320 AU.setPreservesAll();
323 void BlockFrequencyInfoWrapperPass::releaseMemory() { BFI.releaseMemory(); }
325 bool BlockFrequencyInfoWrapperPass::runOnFunction(Function &F) {
326 BranchProbabilityInfo &BPI =
327 getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
328 LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
329 BFI.calculate(F, BPI, LI);
330 return false;
333 AnalysisKey BlockFrequencyAnalysis::Key;
334 BlockFrequencyInfo BlockFrequencyAnalysis::run(Function &F,
335 FunctionAnalysisManager &AM) {
336 auto &BP = AM.getResult<BranchProbabilityAnalysis>(F);
337 auto &LI = AM.getResult<LoopAnalysis>(F);
338 BlockFrequencyInfo BFI;
339 BFI.calculate(F, BP, LI);
340 return BFI;
343 PreservedAnalyses
344 BlockFrequencyPrinterPass::run(Function &F, FunctionAnalysisManager &AM) {
345 OS << "Printing analysis results of BFI for function "
346 << "'" << F.getName() << "':"
347 << "\n";
348 AM.getResult<BlockFrequencyAnalysis>(F).print(OS);
349 return PreservedAnalyses::all();