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/TargetTransformInfo.h"
67#include "llvm/Analysis/ValueTracking.h"
68#include "llvm/IR/Instructions.h"
69#include "llvm/IR/IntrinsicInst.h"
70#include "llvm/IR/Operator.h"
71#include "llvm/InitializePasses.h"
72#include "llvm/Support/CommandLine.h"
73#include "llvm/Support/Debug.h"
74
75using namespace llvm;
76
77#define DEBUG_TYPE "speculative-execution"
78
79// The risk that speculation will not pay off increases with the
80// number of instructions speculated, so we put a limit on that.
81static cl::opt<unsigned> SpecExecMaxSpeculationCost(
82 "spec-exec-max-speculation-cost", cl::init(Val: 7), cl::Hidden,
83 cl::desc("Speculative execution is not applied to basic blocks where "
84 "the cost of the instructions to speculatively execute "
85 "exceeds this limit."));
86
87// Speculating just a few instructions from a larger block tends not
88// to be profitable and this limit prevents that. A reason for that is
89// that small basic blocks are more likely to be candidates for
90// further optimization.
91static cl::opt<unsigned> SpecExecMaxNotHoisted(
92 "spec-exec-max-not-hoisted", cl::init(Val: 5), cl::Hidden,
93 cl::desc("Speculative execution is not applied to basic blocks where the "
94 "number of instructions that would not be speculatively executed "
95 "exceeds this limit."));
96
97static cl::opt<bool> SpecExecOnlyIfDivergentTarget(
98 "spec-exec-only-if-divergent-target", cl::init(Val: false), cl::Hidden,
99 cl::desc("Speculative execution is applied only to targets with divergent "
100 "branches, even if the pass was configured to apply only to all "
101 "targets."));
102
103namespace {
104
105class SpeculativeExecutionLegacyPass : public FunctionPass {
106public:
107 static char ID;
108 explicit SpeculativeExecutionLegacyPass(bool OnlyIfDivergentTarget = false)
109 : FunctionPass(ID), OnlyIfDivergentTarget(OnlyIfDivergentTarget ||
110 SpecExecOnlyIfDivergentTarget),
111 Impl(OnlyIfDivergentTarget) {}
112
113 void getAnalysisUsage(AnalysisUsage &AU) const override;
114 bool runOnFunction(Function &F) override;
115
116 StringRef getPassName() const override {
117 if (OnlyIfDivergentTarget)
118 return "Speculatively execute instructions if target has divergent "
119 "branches";
120 return "Speculatively execute instructions";
121 }
122
123private:
124 // Variable preserved purely for correct name printing.
125 const bool OnlyIfDivergentTarget;
126
127 SpeculativeExecutionPass Impl;
128};
129} // namespace
130
131char SpeculativeExecutionLegacyPass::ID = 0;
132INITIALIZE_PASS_BEGIN(SpeculativeExecutionLegacyPass, "speculative-execution",
133 "Speculatively execute instructions", false, false)
134INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
135INITIALIZE_PASS_END(SpeculativeExecutionLegacyPass, "speculative-execution",
136 "Speculatively execute instructions", false, false)
137
138void SpeculativeExecutionLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
139 AU.addRequired<TargetTransformInfoWrapperPass>();
140 AU.addPreserved<GlobalsAAWrapperPass>();
141 AU.setPreservesCFG();
142}
143
144bool SpeculativeExecutionLegacyPass::runOnFunction(Function &F) {
145 if (skipFunction(F))
146 return false;
147
148 auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
149 return Impl.runImpl(F, TTI);
150}
151
152namespace llvm {
153
154bool SpeculativeExecutionPass::runImpl(Function &F, TargetTransformInfo *TTI) {
155 if (OnlyIfDivergentTarget && !TTI->hasBranchDivergence(F: &F)) {
156 LLVM_DEBUG(dbgs() << "Not running SpeculativeExecution because "
157 "TTI->hasBranchDivergence() is false.\n");
158 return false;
159 }
160
161 this->TTI = TTI;
162 bool Changed = false;
163 for (auto& B : F) {
164 Changed |= runOnBasicBlock(B);
165 }
166 return Changed;
167}
168
169bool SpeculativeExecutionPass::runOnBasicBlock(BasicBlock &B) {
170 BranchInst *BI = dyn_cast<BranchInst>(Val: B.getTerminator());
171 if (BI == nullptr)
172 return false;
173
174 if (BI->getNumSuccessors() != 2)
175 return false;
176 BasicBlock &Succ0 = *BI->getSuccessor(i: 0);
177 BasicBlock &Succ1 = *BI->getSuccessor(i: 1);
178
179 if (&B == &Succ0 || &B == &Succ1 || &Succ0 == &Succ1) {
180 return false;
181 }
182
183 // Hoist from if-then (triangle).
184 if (Succ0.getSinglePredecessor() != nullptr &&
185 Succ0.getSingleSuccessor() == &Succ1) {
186 return considerHoistingFromTo(FromBlock&: Succ0, ToBlock&: B);
187 }
188
189 // Hoist from if-else (triangle).
190 if (Succ1.getSinglePredecessor() != nullptr &&
191 Succ1.getSingleSuccessor() == &Succ0) {
192 return considerHoistingFromTo(FromBlock&: Succ1, ToBlock&: B);
193 }
194
195 // Hoist from if-then-else (diamond), but only if it is equivalent to
196 // an if-else or if-then due to one of the branches doing nothing.
197 if (Succ0.getSinglePredecessor() != nullptr &&
198 Succ1.getSinglePredecessor() != nullptr &&
199 Succ1.getSingleSuccessor() != nullptr &&
200 Succ1.getSingleSuccessor() != &B &&
201 Succ1.getSingleSuccessor() == Succ0.getSingleSuccessor()) {
202 // If a block has only one instruction, then that is a terminator
203 // instruction so that the block does nothing. This does happen.
204 if (Succ1.size() == 1) // equivalent to if-then
205 return considerHoistingFromTo(FromBlock&: Succ0, ToBlock&: B);
206 if (Succ0.size() == 1) // equivalent to if-else
207 return considerHoistingFromTo(FromBlock&: Succ1, ToBlock&: B);
208 }
209
210 return false;
211}
212
213static InstructionCost ComputeSpeculationCost(const Instruction *I,
214 const TargetTransformInfo &TTI) {
215 switch (Operator::getOpcode(V: I)) {
216 case Instruction::GetElementPtr:
217 case Instruction::Add:
218 case Instruction::Mul:
219 case Instruction::And:
220 case Instruction::Or:
221 case Instruction::Select:
222 case Instruction::Shl:
223 case Instruction::Sub:
224 case Instruction::LShr:
225 case Instruction::AShr:
226 case Instruction::Xor:
227 case Instruction::ZExt:
228 case Instruction::SExt:
229 case Instruction::Call:
230 case Instruction::BitCast:
231 case Instruction::PtrToInt:
232 case Instruction::IntToPtr:
233 case Instruction::AddrSpaceCast:
234 case Instruction::FPToUI:
235 case Instruction::FPToSI:
236 case Instruction::UIToFP:
237 case Instruction::SIToFP:
238 case Instruction::FPExt:
239 case Instruction::FPTrunc:
240 case Instruction::FAdd:
241 case Instruction::FSub:
242 case Instruction::FMul:
243 case Instruction::FDiv:
244 case Instruction::FRem:
245 case Instruction::FNeg:
246 case Instruction::ICmp:
247 case Instruction::FCmp:
248 case Instruction::Trunc:
249 case Instruction::Freeze:
250 case Instruction::ExtractElement:
251 case Instruction::InsertElement:
252 case Instruction::ShuffleVector:
253 case Instruction::ExtractValue:
254 case Instruction::InsertValue:
255 return TTI.getInstructionCost(U: I, CostKind: TargetTransformInfo::TCK_SizeAndLatency);
256
257 default:
258 return InstructionCost::getInvalid(); // Disallow anything not explicitly
259 // listed.
260 }
261}
262
263// Do not hoist any debug info intrinsics.
264// ...
265// if (cond) {
266// x = y * z;
267// foo();
268// }
269// ...
270// -------- Which then becomes:
271// ...
272// if.then:
273// %x = mul i32 %y, %z
274// call void @llvm.dbg.value(%x, !"x", !DIExpression())
275// call void foo()
276//
277// SpeculativeExecution might decide to hoist the 'y * z' calculation
278// out of the 'if' block, because it is more efficient that way, so the
279// '%x = mul i32 %y, %z' moves to the block above. But it might also
280// decide to hoist the 'llvm.dbg.value' call.
281// This is incorrect, because even if we've moved the calculation of
282// 'y * z', we should not see the value of 'x' change unless we
283// actually go inside the 'if' block.
284
285bool SpeculativeExecutionPass::considerHoistingFromTo(
286 BasicBlock &FromBlock, BasicBlock &ToBlock) {
287 SmallPtrSet<const Instruction *, 8> NotHoisted;
288 auto HasNoUnhoistedInstr = [&NotHoisted](auto Values) {
289 for (const Value *V : Values) {
290 if (const auto *I = dyn_cast_or_null<Instruction>(Val: V))
291 if (NotHoisted.contains(Ptr: I))
292 return false;
293 }
294 return true;
295 };
296 auto AllPrecedingUsesFromBlockHoisted =
297 [&HasNoUnhoistedInstr](const User *U) {
298 // Do not hoist any debug info intrinsics.
299 if (isa<DbgInfoIntrinsic>(Val: U))
300 return false;
301
302 return HasNoUnhoistedInstr(U->operand_values());
303 };
304
305 InstructionCost TotalSpeculationCost = 0;
306 unsigned NotHoistedInstCount = 0;
307 for (const auto &I : FromBlock) {
308 const InstructionCost Cost = ComputeSpeculationCost(I: &I, TTI: *TTI);
309 if (Cost.isValid() && isSafeToSpeculativelyExecute(I: &I) &&
310 AllPrecedingUsesFromBlockHoisted(&I)) {
311 TotalSpeculationCost += Cost;
312 if (TotalSpeculationCost > SpecExecMaxSpeculationCost)
313 return false; // too much to hoist
314 } else {
315 // Debug info intrinsics should not be counted for threshold.
316 if (!isa<DbgInfoIntrinsic>(Val: I))
317 NotHoistedInstCount++;
318 if (NotHoistedInstCount > SpecExecMaxNotHoisted)
319 return false; // too much left behind
320 NotHoisted.insert(Ptr: &I);
321 }
322 }
323
324 for (auto I = FromBlock.begin(); I != FromBlock.end();) {
325 // We have to increment I before moving Current as moving Current
326 // changes the list that I is iterating through.
327 auto Current = I;
328 ++I;
329 if (!NotHoisted.count(Ptr: &*Current)) {
330 Current->moveBefore(MovePos: ToBlock.getTerminator());
331 Current->dropLocation();
332 }
333 }
334 return true;
335}
336
337FunctionPass *createSpeculativeExecutionPass() {
338 return new SpeculativeExecutionLegacyPass();
339}
340
341FunctionPass *createSpeculativeExecutionIfHasBranchDivergencePass() {
342 return new SpeculativeExecutionLegacyPass(/* OnlyIfDivergentTarget = */ true);
343}
344
345SpeculativeExecutionPass::SpeculativeExecutionPass(bool OnlyIfDivergentTarget)
346 : OnlyIfDivergentTarget(OnlyIfDivergentTarget ||
347 SpecExecOnlyIfDivergentTarget) {}
348
349PreservedAnalyses SpeculativeExecutionPass::run(Function &F,
350 FunctionAnalysisManager &AM) {
351 auto *TTI = &AM.getResult<TargetIRAnalysis>(IR&: F);
352
353 bool Changed = runImpl(F, TTI);
354
355 if (!Changed)
356 return PreservedAnalyses::all();
357 PreservedAnalyses PA;
358 PA.preserveSet<CFGAnalyses>();
359 return PA;
360}
361
362void SpeculativeExecutionPass::printPipeline(
363 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
364 static_cast<PassInfoMixin<SpeculativeExecutionPass> *>(this)->printPipeline(
365 OS, MapClassName2PassName);
366 OS << '<';
367 if (OnlyIfDivergentTarget)
368 OS << "only-if-divergent-target";
369 OS << '>';
370}
371} // namespace llvm
372