xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/SpeculativeExecution.cpp (revision a134ebd6e63f658f2d3d04ac0c60d23bcaa86dd7)
1 //===- SpeculativeExecution.cpp ---------------------------------*- C++ -*-===//
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 pass hoists instructions to enable speculative execution on
10 // targets where branches are expensive. This is aimed at GPUs. It
11 // currently works on simple if-then and if-then-else
12 // patterns.
13 //
14 // Removing branches is not the only motivation for this
15 // pass. E.g. consider this code and assume that there is no
16 // addressing mode for multiplying by sizeof(*a):
17 //
18 //   if (b > 0)
19 //     c = a[i + 1]
20 //   if (d > 0)
21 //     e = a[i + 2]
22 //
23 // turns into
24 //
25 //   p = &a[i + 1];
26 //   if (b > 0)
27 //     c = *p;
28 //   q = &a[i + 2];
29 //   if (d > 0)
30 //     e = *q;
31 //
32 // which could later be optimized to
33 //
34 //   r = &a[i];
35 //   if (b > 0)
36 //     c = r[1];
37 //   if (d > 0)
38 //     e = r[2];
39 //
40 // Later passes sink back much of the speculated code that did not enable
41 // further optimization.
42 //
43 // This pass is more aggressive than the function SpeculativeyExecuteBB in
44 // SimplifyCFG. SimplifyCFG will not speculate if no selects are introduced and
45 // it will speculate at most one instruction. It also will not speculate if
46 // there is a value defined in the if-block that is only used in the then-block.
47 // These restrictions make sense since the speculation in SimplifyCFG seems
48 // aimed at introducing cheap selects, while this pass is intended to do more
49 // aggressive speculation while counting on later passes to either capitalize on
50 // that or clean it up.
51 //
52 // If the pass was created by calling
53 // createSpeculativeExecutionIfHasBranchDivergencePass or the
54 // -spec-exec-only-if-divergent-target option is present, this pass only has an
55 // effect on targets where TargetTransformInfo::hasBranchDivergence() is true;
56 // on other targets, it is a nop.
57 //
58 // This lets you include this pass unconditionally in the IR pass pipeline, but
59 // only enable it for relevant targets.
60 //
61 //===----------------------------------------------------------------------===//
62 
63 #include "llvm/Transforms/Scalar/SpeculativeExecution.h"
64 #include "llvm/ADT/SmallPtrSet.h"
65 #include "llvm/Analysis/GlobalsModRef.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/IR/Instructions.h"
68 #include "llvm/IR/Module.h"
69 #include "llvm/IR/Operator.h"
70 #include "llvm/InitializePasses.h"
71 #include "llvm/Support/CommandLine.h"
72 #include "llvm/Support/Debug.h"
73 
74 using namespace llvm;
75 
76 #define DEBUG_TYPE "speculative-execution"
77 
78 // The risk that speculation will not pay off increases with the
79 // number of instructions speculated, so we put a limit on that.
80 static cl::opt<unsigned> SpecExecMaxSpeculationCost(
81     "spec-exec-max-speculation-cost", cl::init(7), cl::Hidden,
82     cl::desc("Speculative execution is not applied to basic blocks where "
83              "the cost of the instructions to speculatively execute "
84              "exceeds this limit."));
85 
86 // Speculating just a few instructions from a larger block tends not
87 // to be profitable and this limit prevents that. A reason for that is
88 // that small basic blocks are more likely to be candidates for
89 // further optimization.
90 static cl::opt<unsigned> SpecExecMaxNotHoisted(
91     "spec-exec-max-not-hoisted", cl::init(5), cl::Hidden,
92     cl::desc("Speculative execution is not applied to basic blocks where the "
93              "number of instructions that would not be speculatively executed "
94              "exceeds this limit."));
95 
96 static cl::opt<bool> SpecExecOnlyIfDivergentTarget(
97     "spec-exec-only-if-divergent-target", cl::init(false), cl::Hidden,
98     cl::desc("Speculative execution is applied only to targets with divergent "
99              "branches, even if the pass was configured to apply only to all "
100              "targets."));
101 
102 namespace {
103 
104 class SpeculativeExecutionLegacyPass : public FunctionPass {
105 public:
106   static char ID;
107   explicit SpeculativeExecutionLegacyPass(bool OnlyIfDivergentTarget = false)
108       : FunctionPass(ID), OnlyIfDivergentTarget(OnlyIfDivergentTarget ||
109                                                 SpecExecOnlyIfDivergentTarget),
110         Impl(OnlyIfDivergentTarget) {}
111 
112   void getAnalysisUsage(AnalysisUsage &AU) const override;
113   bool runOnFunction(Function &F) override;
114 
115   StringRef getPassName() const override {
116     if (OnlyIfDivergentTarget)
117       return "Speculatively execute instructions if target has divergent "
118              "branches";
119     return "Speculatively execute instructions";
120   }
121 
122 private:
123   // Variable preserved purely for correct name printing.
124   const bool OnlyIfDivergentTarget;
125 
126   SpeculativeExecutionPass Impl;
127 };
128 } // namespace
129 
130 char SpeculativeExecutionLegacyPass::ID = 0;
131 INITIALIZE_PASS_BEGIN(SpeculativeExecutionLegacyPass, "speculative-execution",
132                       "Speculatively execute instructions", false, false)
133 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
134 INITIALIZE_PASS_END(SpeculativeExecutionLegacyPass, "speculative-execution",
135                     "Speculatively execute instructions", false, false)
136 
137 void SpeculativeExecutionLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
138   AU.addRequired<TargetTransformInfoWrapperPass>();
139   AU.addPreserved<GlobalsAAWrapperPass>();
140   AU.setPreservesCFG();
141 }
142 
143 bool SpeculativeExecutionLegacyPass::runOnFunction(Function &F) {
144   if (skipFunction(F))
145     return false;
146 
147   auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
148   return Impl.runImpl(F, TTI);
149 }
150 
151 namespace llvm {
152 
153 bool SpeculativeExecutionPass::runImpl(Function &F, TargetTransformInfo *TTI) {
154   if (OnlyIfDivergentTarget && !TTI->hasBranchDivergence()) {
155     LLVM_DEBUG(dbgs() << "Not running SpeculativeExecution because "
156                          "TTI->hasBranchDivergence() is false.\n");
157     return false;
158   }
159 
160   this->TTI = TTI;
161   bool Changed = false;
162   for (auto& B : F) {
163     Changed |= runOnBasicBlock(B);
164   }
165   return Changed;
166 }
167 
168 bool SpeculativeExecutionPass::runOnBasicBlock(BasicBlock &B) {
169   BranchInst *BI = dyn_cast<BranchInst>(B.getTerminator());
170   if (BI == nullptr)
171     return false;
172 
173   if (BI->getNumSuccessors() != 2)
174     return false;
175   BasicBlock &Succ0 = *BI->getSuccessor(0);
176   BasicBlock &Succ1 = *BI->getSuccessor(1);
177 
178   if (&B == &Succ0 || &B == &Succ1 || &Succ0 == &Succ1) {
179     return false;
180   }
181 
182   // Hoist from if-then (triangle).
183   if (Succ0.getSinglePredecessor() != nullptr &&
184       Succ0.getSingleSuccessor() == &Succ1) {
185     return considerHoistingFromTo(Succ0, B);
186   }
187 
188   // Hoist from if-else (triangle).
189   if (Succ1.getSinglePredecessor() != nullptr &&
190       Succ1.getSingleSuccessor() == &Succ0) {
191     return considerHoistingFromTo(Succ1, B);
192   }
193 
194   // Hoist from if-then-else (diamond), but only if it is equivalent to
195   // an if-else or if-then due to one of the branches doing nothing.
196   if (Succ0.getSinglePredecessor() != nullptr &&
197       Succ1.getSinglePredecessor() != nullptr &&
198       Succ1.getSingleSuccessor() != nullptr &&
199       Succ1.getSingleSuccessor() != &B &&
200       Succ1.getSingleSuccessor() == Succ0.getSingleSuccessor()) {
201     // If a block has only one instruction, then that is a terminator
202     // instruction so that the block does nothing. This does happen.
203     if (Succ1.size() == 1) // equivalent to if-then
204       return considerHoistingFromTo(Succ0, B);
205     if (Succ0.size() == 1) // equivalent to if-else
206       return considerHoistingFromTo(Succ1, B);
207   }
208 
209   return false;
210 }
211 
212 static unsigned ComputeSpeculationCost(const Instruction *I,
213                                        const TargetTransformInfo &TTI) {
214   switch (Operator::getOpcode(I)) {
215     case Instruction::GetElementPtr:
216     case Instruction::Add:
217     case Instruction::Mul:
218     case Instruction::And:
219     case Instruction::Or:
220     case Instruction::Select:
221     case Instruction::Shl:
222     case Instruction::Sub:
223     case Instruction::LShr:
224     case Instruction::AShr:
225     case Instruction::Xor:
226     case Instruction::ZExt:
227     case Instruction::SExt:
228     case Instruction::Call:
229     case Instruction::BitCast:
230     case Instruction::PtrToInt:
231     case Instruction::IntToPtr:
232     case Instruction::AddrSpaceCast:
233     case Instruction::FPToUI:
234     case Instruction::FPToSI:
235     case Instruction::UIToFP:
236     case Instruction::SIToFP:
237     case Instruction::FPExt:
238     case Instruction::FPTrunc:
239     case Instruction::FAdd:
240     case Instruction::FSub:
241     case Instruction::FMul:
242     case Instruction::FDiv:
243     case Instruction::FRem:
244     case Instruction::FNeg:
245     case Instruction::ICmp:
246     case Instruction::FCmp:
247       return TTI.getUserCost(I);
248 
249     default:
250       return UINT_MAX; // Disallow anything not whitelisted.
251   }
252 }
253 
254 bool SpeculativeExecutionPass::considerHoistingFromTo(
255     BasicBlock &FromBlock, BasicBlock &ToBlock) {
256   SmallPtrSet<const Instruction *, 8> NotHoisted;
257   const auto AllPrecedingUsesFromBlockHoisted = [&NotHoisted](User *U) {
258     for (Value* V : U->operand_values()) {
259       if (Instruction *I = dyn_cast<Instruction>(V)) {
260         if (NotHoisted.count(I) > 0)
261           return false;
262       }
263     }
264     return true;
265   };
266 
267   unsigned TotalSpeculationCost = 0;
268   for (auto& I : FromBlock) {
269     const unsigned Cost = ComputeSpeculationCost(&I, *TTI);
270     if (Cost != UINT_MAX && isSafeToSpeculativelyExecute(&I) &&
271         AllPrecedingUsesFromBlockHoisted(&I)) {
272       TotalSpeculationCost += Cost;
273       if (TotalSpeculationCost > SpecExecMaxSpeculationCost)
274         return false;  // too much to hoist
275     } else {
276       NotHoisted.insert(&I);
277       if (NotHoisted.size() > SpecExecMaxNotHoisted)
278         return false; // too much left behind
279     }
280   }
281 
282   if (TotalSpeculationCost == 0)
283     return false; // nothing to hoist
284 
285   for (auto I = FromBlock.begin(); I != FromBlock.end();) {
286     // We have to increment I before moving Current as moving Current
287     // changes the list that I is iterating through.
288     auto Current = I;
289     ++I;
290     if (!NotHoisted.count(&*Current)) {
291       Current->moveBefore(ToBlock.getTerminator());
292     }
293   }
294   return true;
295 }
296 
297 FunctionPass *createSpeculativeExecutionPass() {
298   return new SpeculativeExecutionLegacyPass();
299 }
300 
301 FunctionPass *createSpeculativeExecutionIfHasBranchDivergencePass() {
302   return new SpeculativeExecutionLegacyPass(/* OnlyIfDivergentTarget = */ true);
303 }
304 
305 SpeculativeExecutionPass::SpeculativeExecutionPass(bool OnlyIfDivergentTarget)
306     : OnlyIfDivergentTarget(OnlyIfDivergentTarget ||
307                             SpecExecOnlyIfDivergentTarget) {}
308 
309 PreservedAnalyses SpeculativeExecutionPass::run(Function &F,
310                                                 FunctionAnalysisManager &AM) {
311   auto *TTI = &AM.getResult<TargetIRAnalysis>(F);
312 
313   bool Changed = runImpl(F, TTI);
314 
315   if (!Changed)
316     return PreservedAnalyses::all();
317   PreservedAnalyses PA;
318   PA.preserve<GlobalsAA>();
319   PA.preserveSet<CFGAnalyses>();
320   return PA;
321 }
322 }  // namespace llvm
323