1//===--- PartiallyInlineLibCalls.cpp - Partially inline libcalls ----------===//
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 tries to partially inline the fast path of well-known library
10// functions, such as using square-root instructions for cases where sqrt()
11// does not need to set errno.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/Transforms/Scalar/PartiallyInlineLibCalls.h"
16#include "llvm/Analysis/DomTreeUpdater.h"
17#include "llvm/Analysis/TargetLibraryInfo.h"
18#include "llvm/Analysis/TargetTransformInfo.h"
19#include "llvm/IR/Dominators.h"
20#include "llvm/IR/IRBuilder.h"
21#include "llvm/InitializePasses.h"
22#include "llvm/Support/DebugCounter.h"
23#include "llvm/Transforms/Scalar.h"
24#include "llvm/Transforms/Utils/BasicBlockUtils.h"
25#include <optional>
26
27using namespace llvm;
28
29#define DEBUG_TYPE "partially-inline-libcalls"
30
31DEBUG_COUNTER(PILCounter, "partially-inline-libcalls-transform",
32 "Controls transformations in partially-inline-libcalls");
33
34static bool optimizeSQRT(CallInst *Call, Function *CalledFunc,
35 BasicBlock &CurrBB, Function::iterator &BB,
36 const TargetTransformInfo *TTI, DomTreeUpdater *DTU) {
37 // There is no need to change the IR, since backend will emit sqrt
38 // instruction if the call has already been marked read-only.
39 if (Call->onlyReadsMemory())
40 return false;
41
42 if (!DebugCounter::shouldExecute(CounterName: PILCounter))
43 return false;
44
45 // Do the following transformation:
46 //
47 // (before)
48 // dst = sqrt(src)
49 //
50 // (after)
51 // v0 = sqrt_noreadmem(src) # native sqrt instruction.
52 // [if (v0 is a NaN) || if (src < 0)]
53 // v1 = sqrt(src) # library call.
54 // dst = phi(v0, v1)
55 //
56
57 Type *Ty = Call->getType();
58 IRBuilder<> Builder(Call->getNextNode());
59
60 // Split CurrBB right after the call, create a 'then' block (that branches
61 // back to split-off tail of CurrBB) into which we'll insert a libcall.
62 Instruction *LibCallTerm = SplitBlockAndInsertIfThen(
63 Cond: Builder.getTrue(), SplitBefore: Call->getNextNode(), /*Unreachable=*/false,
64 /*BranchWeights*/ nullptr, DTU);
65
66 auto *CurrBBTerm = cast<BranchInst>(Val: CurrBB.getTerminator());
67 // We want an 'else' block though, not a 'then' block.
68 cast<BranchInst>(Val: CurrBBTerm)->swapSuccessors();
69
70 // Create phi that will merge results of either sqrt and replace all uses.
71 BasicBlock *JoinBB = LibCallTerm->getSuccessor(Idx: 0);
72 JoinBB->setName(CurrBB.getName() + ".split");
73 Builder.SetInsertPoint(TheBB: JoinBB, IP: JoinBB->begin());
74 PHINode *Phi = Builder.CreatePHI(Ty, NumReservedValues: 2);
75 Call->replaceAllUsesWith(V: Phi);
76
77 // Finally, insert the libcall into 'else' block.
78 BasicBlock *LibCallBB = LibCallTerm->getParent();
79 LibCallBB->setName("call.sqrt");
80 Builder.SetInsertPoint(LibCallTerm);
81 Instruction *LibCall = Call->clone();
82 Builder.Insert(I: LibCall);
83
84 // Add memory(none) attribute, so that the backend can use a native sqrt
85 // instruction for this call.
86 Call->setDoesNotAccessMemory();
87
88 // Insert a FP compare instruction and use it as the CurrBB branch condition.
89 Builder.SetInsertPoint(CurrBBTerm);
90 Value *FCmp = TTI->isFCmpOrdCheaperThanFCmpZero(Ty)
91 ? Builder.CreateFCmpORD(LHS: Call, RHS: Call)
92 : Builder.CreateFCmpOGE(LHS: Call->getOperand(i_nocapture: 0),
93 RHS: ConstantFP::get(Ty, V: 0.0));
94 CurrBBTerm->setCondition(FCmp);
95
96 // Add phi operands.
97 Phi->addIncoming(V: Call, BB: &CurrBB);
98 Phi->addIncoming(V: LibCall, BB: LibCallBB);
99
100 BB = JoinBB->getIterator();
101 return true;
102}
103
104static bool runPartiallyInlineLibCalls(Function &F, TargetLibraryInfo *TLI,
105 const TargetTransformInfo *TTI,
106 DominatorTree *DT) {
107 std::optional<DomTreeUpdater> DTU;
108 if (DT)
109 DTU.emplace(args&: DT, args: DomTreeUpdater::UpdateStrategy::Lazy);
110
111 bool Changed = false;
112
113 Function::iterator CurrBB;
114 for (Function::iterator BB = F.begin(), BE = F.end(); BB != BE;) {
115 CurrBB = BB++;
116
117 for (BasicBlock::iterator II = CurrBB->begin(), IE = CurrBB->end();
118 II != IE; ++II) {
119 CallInst *Call = dyn_cast<CallInst>(Val: &*II);
120 Function *CalledFunc;
121
122 if (!Call || !(CalledFunc = Call->getCalledFunction()))
123 continue;
124
125 if (Call->isNoBuiltin() || Call->isStrictFP())
126 continue;
127
128 if (Call->isMustTailCall())
129 continue;
130
131 // Skip if function either has local linkage or is not a known library
132 // function.
133 LibFunc LF;
134 if (CalledFunc->hasLocalLinkage() ||
135 !TLI->getLibFunc(FDecl: *CalledFunc, F&: LF) || !TLI->has(F: LF))
136 continue;
137
138 switch (LF) {
139 case LibFunc_sqrtf:
140 case LibFunc_sqrt:
141 if (TTI->haveFastSqrt(Ty: Call->getType()) &&
142 optimizeSQRT(Call, CalledFunc, CurrBB&: *CurrBB, BB, TTI,
143 DTU: DTU ? &*DTU : nullptr))
144 break;
145 continue;
146 default:
147 continue;
148 }
149
150 Changed = true;
151 break;
152 }
153 }
154
155 return Changed;
156}
157
158PreservedAnalyses
159PartiallyInlineLibCallsPass::run(Function &F, FunctionAnalysisManager &AM) {
160 auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F);
161 auto &TTI = AM.getResult<TargetIRAnalysis>(IR&: F);
162 auto *DT = AM.getCachedResult<DominatorTreeAnalysis>(IR&: F);
163 if (!runPartiallyInlineLibCalls(F, TLI: &TLI, TTI: &TTI, DT))
164 return PreservedAnalyses::all();
165 PreservedAnalyses PA;
166 PA.preserve<DominatorTreeAnalysis>();
167 return PA;
168}
169
170namespace {
171class PartiallyInlineLibCallsLegacyPass : public FunctionPass {
172public:
173 static char ID;
174
175 PartiallyInlineLibCallsLegacyPass() : FunctionPass(ID) {
176 initializePartiallyInlineLibCallsLegacyPassPass(
177 *PassRegistry::getPassRegistry());
178 }
179
180 void getAnalysisUsage(AnalysisUsage &AU) const override {
181 AU.addRequired<TargetLibraryInfoWrapperPass>();
182 AU.addRequired<TargetTransformInfoWrapperPass>();
183 AU.addPreserved<DominatorTreeWrapperPass>();
184 FunctionPass::getAnalysisUsage(AU);
185 }
186
187 bool runOnFunction(Function &F) override {
188 if (skipFunction(F))
189 return false;
190
191 TargetLibraryInfo *TLI =
192 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
193 const TargetTransformInfo *TTI =
194 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
195 DominatorTree *DT = nullptr;
196 if (auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>())
197 DT = &DTWP->getDomTree();
198 return runPartiallyInlineLibCalls(F, TLI, TTI, DT);
199 }
200};
201}
202
203char PartiallyInlineLibCallsLegacyPass::ID = 0;
204INITIALIZE_PASS_BEGIN(PartiallyInlineLibCallsLegacyPass,
205 "partially-inline-libcalls",
206 "Partially inline calls to library functions", false,
207 false)
208INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
209INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
210INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
211INITIALIZE_PASS_END(PartiallyInlineLibCallsLegacyPass,
212 "partially-inline-libcalls",
213 "Partially inline calls to library functions", false, false)
214
215FunctionPass *llvm::createPartiallyInlineLibCallsPass() {
216 return new PartiallyInlineLibCallsLegacyPass();
217}
218