|
150
|
1 //===- ScalarEvolutionsTest.cpp - ScalarEvolution unit tests --------------===//
|
|
|
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 #include "llvm/ADT/SmallVector.h"
|
|
|
10 #include "llvm/Analysis/AssumptionCache.h"
|
|
|
11 #include "llvm/Analysis/LoopInfo.h"
|
|
|
12 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
|
|
173
|
13 #include "llvm/Analysis/ScalarEvolutionNormalization.h"
|
|
150
|
14 #include "llvm/Analysis/TargetLibraryInfo.h"
|
|
|
15 #include "llvm/AsmParser/Parser.h"
|
|
|
16 #include "llvm/IR/Constants.h"
|
|
|
17 #include "llvm/IR/Dominators.h"
|
|
|
18 #include "llvm/IR/GlobalVariable.h"
|
|
|
19 #include "llvm/IR/IRBuilder.h"
|
|
|
20 #include "llvm/IR/InstIterator.h"
|
|
|
21 #include "llvm/IR/LLVMContext.h"
|
|
|
22 #include "llvm/IR/LegacyPassManager.h"
|
|
|
23 #include "llvm/IR/Module.h"
|
|
|
24 #include "llvm/IR/Verifier.h"
|
|
|
25 #include "llvm/Support/SourceMgr.h"
|
|
|
26 #include "gtest/gtest.h"
|
|
|
27
|
|
|
28 namespace llvm {
|
|
|
29
|
|
|
30 // We use this fixture to ensure that we clean up ScalarEvolution before
|
|
|
31 // deleting the PassManager.
|
|
|
32 class ScalarEvolutionsTest : public testing::Test {
|
|
|
33 protected:
|
|
|
34 LLVMContext Context;
|
|
|
35 Module M;
|
|
|
36 TargetLibraryInfoImpl TLII;
|
|
|
37 TargetLibraryInfo TLI;
|
|
|
38
|
|
|
39 std::unique_ptr<AssumptionCache> AC;
|
|
|
40 std::unique_ptr<DominatorTree> DT;
|
|
|
41 std::unique_ptr<LoopInfo> LI;
|
|
|
42
|
|
|
43 ScalarEvolutionsTest() : M("", Context), TLII(), TLI(TLII) {}
|
|
|
44
|
|
|
45 ScalarEvolution buildSE(Function &F) {
|
|
|
46 AC.reset(new AssumptionCache(F));
|
|
|
47 DT.reset(new DominatorTree(F));
|
|
|
48 LI.reset(new LoopInfo(*DT));
|
|
|
49 return ScalarEvolution(F, TLI, *AC, *DT, *LI);
|
|
|
50 }
|
|
|
51
|
|
|
52 void runWithSE(
|
|
|
53 Module &M, StringRef FuncName,
|
|
|
54 function_ref<void(Function &F, LoopInfo &LI, ScalarEvolution &SE)> Test) {
|
|
|
55 auto *F = M.getFunction(FuncName);
|
|
|
56 ASSERT_NE(F, nullptr) << "Could not find " << FuncName;
|
|
|
57 ScalarEvolution SE = buildSE(*F);
|
|
|
58 Test(*F, *LI, SE);
|
|
|
59 }
|
|
|
60
|
|
|
61 static Optional<APInt> computeConstantDifference(ScalarEvolution &SE,
|
|
|
62 const SCEV *LHS,
|
|
|
63 const SCEV *RHS) {
|
|
|
64 return SE.computeConstantDifference(LHS, RHS);
|
|
|
65 }
|
|
221
|
66
|
|
|
67 static bool matchURem(ScalarEvolution &SE, const SCEV *Expr, const SCEV *&LHS,
|
|
|
68 const SCEV *&RHS) {
|
|
|
69 return SE.matchURem(Expr, LHS, RHS);
|
|
|
70 }
|
|
|
71
|
|
|
72 static bool isImpliedCond(
|
|
|
73 ScalarEvolution &SE, ICmpInst::Predicate Pred, const SCEV *LHS,
|
|
|
74 const SCEV *RHS, ICmpInst::Predicate FoundPred, const SCEV *FoundLHS,
|
|
|
75 const SCEV *FoundRHS) {
|
|
|
76 return SE.isImpliedCond(Pred, LHS, RHS, FoundPred, FoundLHS, FoundRHS);
|
|
|
77 }
|
|
150
|
78 };
|
|
|
79
|
|
|
80 TEST_F(ScalarEvolutionsTest, SCEVUnknownRAUW) {
|
|
|
81 FunctionType *FTy = FunctionType::get(Type::getVoidTy(Context),
|
|
|
82 std::vector<Type *>(), false);
|
|
|
83 Function *F = Function::Create(FTy, Function::ExternalLinkage, "f", M);
|
|
|
84 BasicBlock *BB = BasicBlock::Create(Context, "entry", F);
|
|
|
85 ReturnInst::Create(Context, nullptr, BB);
|
|
|
86
|
|
|
87 Type *Ty = Type::getInt1Ty(Context);
|
|
|
88 Constant *Init = Constant::getNullValue(Ty);
|
|
|
89 Value *V0 = new GlobalVariable(M, Ty, false, GlobalValue::ExternalLinkage, Init, "V0");
|
|
|
90 Value *V1 = new GlobalVariable(M, Ty, false, GlobalValue::ExternalLinkage, Init, "V1");
|
|
|
91 Value *V2 = new GlobalVariable(M, Ty, false, GlobalValue::ExternalLinkage, Init, "V2");
|
|
|
92
|
|
|
93 ScalarEvolution SE = buildSE(*F);
|
|
|
94
|
|
|
95 const SCEV *S0 = SE.getSCEV(V0);
|
|
|
96 const SCEV *S1 = SE.getSCEV(V1);
|
|
|
97 const SCEV *S2 = SE.getSCEV(V2);
|
|
|
98
|
|
|
99 const SCEV *P0 = SE.getAddExpr(S0, S0);
|
|
|
100 const SCEV *P1 = SE.getAddExpr(S1, S1);
|
|
|
101 const SCEV *P2 = SE.getAddExpr(S2, S2);
|
|
|
102
|
|
|
103 const SCEVMulExpr *M0 = cast<SCEVMulExpr>(P0);
|
|
|
104 const SCEVMulExpr *M1 = cast<SCEVMulExpr>(P1);
|
|
|
105 const SCEVMulExpr *M2 = cast<SCEVMulExpr>(P2);
|
|
|
106
|
|
|
107 EXPECT_EQ(cast<SCEVConstant>(M0->getOperand(0))->getValue()->getZExtValue(),
|
|
|
108 2u);
|
|
|
109 EXPECT_EQ(cast<SCEVConstant>(M1->getOperand(0))->getValue()->getZExtValue(),
|
|
|
110 2u);
|
|
|
111 EXPECT_EQ(cast<SCEVConstant>(M2->getOperand(0))->getValue()->getZExtValue(),
|
|
|
112 2u);
|
|
|
113
|
|
|
114 // Before the RAUWs, these are all pointing to separate values.
|
|
|
115 EXPECT_EQ(cast<SCEVUnknown>(M0->getOperand(1))->getValue(), V0);
|
|
|
116 EXPECT_EQ(cast<SCEVUnknown>(M1->getOperand(1))->getValue(), V1);
|
|
|
117 EXPECT_EQ(cast<SCEVUnknown>(M2->getOperand(1))->getValue(), V2);
|
|
|
118
|
|
|
119 // Do some RAUWs.
|
|
|
120 V2->replaceAllUsesWith(V1);
|
|
|
121 V1->replaceAllUsesWith(V0);
|
|
|
122
|
|
|
123 // After the RAUWs, these should all be pointing to V0.
|
|
|
124 EXPECT_EQ(cast<SCEVUnknown>(M0->getOperand(1))->getValue(), V0);
|
|
|
125 EXPECT_EQ(cast<SCEVUnknown>(M1->getOperand(1))->getValue(), V0);
|
|
|
126 EXPECT_EQ(cast<SCEVUnknown>(M2->getOperand(1))->getValue(), V0);
|
|
|
127 }
|
|
|
128
|
|
|
129 TEST_F(ScalarEvolutionsTest, SimplifiedPHI) {
|
|
|
130 FunctionType *FTy = FunctionType::get(Type::getVoidTy(Context),
|
|
|
131 std::vector<Type *>(), false);
|
|
|
132 Function *F = Function::Create(FTy, Function::ExternalLinkage, "f", M);
|
|
|
133 BasicBlock *EntryBB = BasicBlock::Create(Context, "entry", F);
|
|
|
134 BasicBlock *LoopBB = BasicBlock::Create(Context, "loop", F);
|
|
|
135 BasicBlock *ExitBB = BasicBlock::Create(Context, "exit", F);
|
|
|
136 BranchInst::Create(LoopBB, EntryBB);
|
|
|
137 BranchInst::Create(LoopBB, ExitBB, UndefValue::get(Type::getInt1Ty(Context)),
|
|
|
138 LoopBB);
|
|
|
139 ReturnInst::Create(Context, nullptr, ExitBB);
|
|
|
140 auto *Ty = Type::getInt32Ty(Context);
|
|
|
141 auto *PN = PHINode::Create(Ty, 2, "", &*LoopBB->begin());
|
|
|
142 PN->addIncoming(Constant::getNullValue(Ty), EntryBB);
|
|
|
143 PN->addIncoming(UndefValue::get(Ty), LoopBB);
|
|
|
144 ScalarEvolution SE = buildSE(*F);
|
|
|
145 auto *S1 = SE.getSCEV(PN);
|
|
|
146 auto *S2 = SE.getSCEV(PN);
|
|
|
147 auto *ZeroConst = SE.getConstant(Ty, 0);
|
|
|
148
|
|
|
149 // At some point, only the first call to getSCEV returned the simplified
|
|
|
150 // SCEVConstant and later calls just returned a SCEVUnknown referencing the
|
|
|
151 // PHI node.
|
|
|
152 EXPECT_EQ(S1, ZeroConst);
|
|
|
153 EXPECT_EQ(S1, S2);
|
|
|
154 }
|
|
|
155
|
|
|
156
|
|
|
157 static Instruction *getInstructionByName(Function &F, StringRef Name) {
|
|
|
158 for (auto &I : instructions(F))
|
|
|
159 if (I.getName() == Name)
|
|
|
160 return &I;
|
|
|
161 llvm_unreachable("Expected to find instruction!");
|
|
|
162 }
|
|
|
163
|
|
221
|
164 static Value *getArgByName(Function &F, StringRef Name) {
|
|
|
165 for (auto &Arg : F.args())
|
|
|
166 if (Arg.getName() == Name)
|
|
|
167 return &Arg;
|
|
|
168 llvm_unreachable("Expected to find instruction!");
|
|
|
169 }
|
|
150
|
170 TEST_F(ScalarEvolutionsTest, CommutativeExprOperandOrder) {
|
|
|
171 LLVMContext C;
|
|
|
172 SMDiagnostic Err;
|
|
|
173 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
174 "target datalayout = \"e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128\" "
|
|
|
175 " "
|
|
|
176 "@var_0 = external global i32, align 4"
|
|
|
177 "@var_1 = external global i32, align 4"
|
|
|
178 "@var_2 = external global i32, align 4"
|
|
|
179 " "
|
|
|
180 "declare i32 @unknown(i32, i32, i32)"
|
|
|
181 " "
|
|
|
182 "define void @f_1(i8* nocapture %arr, i32 %n, i32* %A, i32* %B) "
|
|
|
183 " local_unnamed_addr { "
|
|
|
184 "entry: "
|
|
|
185 " %entrycond = icmp sgt i32 %n, 0 "
|
|
|
186 " br i1 %entrycond, label %loop.ph, label %for.end "
|
|
|
187 " "
|
|
|
188 "loop.ph: "
|
|
|
189 " %a = load i32, i32* %A, align 4 "
|
|
|
190 " %b = load i32, i32* %B, align 4 "
|
|
|
191 " %mul = mul nsw i32 %b, %a "
|
|
|
192 " %iv0.init = getelementptr inbounds i8, i8* %arr, i32 %mul "
|
|
|
193 " br label %loop "
|
|
|
194 " "
|
|
|
195 "loop: "
|
|
|
196 " %iv0 = phi i8* [ %iv0.inc, %loop ], [ %iv0.init, %loop.ph ] "
|
|
|
197 " %iv1 = phi i32 [ %iv1.inc, %loop ], [ 0, %loop.ph ] "
|
|
|
198 " %conv = trunc i32 %iv1 to i8 "
|
|
|
199 " store i8 %conv, i8* %iv0, align 1 "
|
|
|
200 " %iv0.inc = getelementptr inbounds i8, i8* %iv0, i32 %b "
|
|
|
201 " %iv1.inc = add nuw nsw i32 %iv1, 1 "
|
|
|
202 " %exitcond = icmp eq i32 %iv1.inc, %n "
|
|
|
203 " br i1 %exitcond, label %for.end.loopexit, label %loop "
|
|
|
204 " "
|
|
|
205 "for.end.loopexit: "
|
|
|
206 " br label %for.end "
|
|
|
207 " "
|
|
|
208 "for.end: "
|
|
|
209 " ret void "
|
|
|
210 "} "
|
|
|
211 " "
|
|
|
212 "define void @f_2(i32* %X, i32* %Y, i32* %Z) { "
|
|
|
213 " %x = load i32, i32* %X "
|
|
|
214 " %y = load i32, i32* %Y "
|
|
|
215 " %z = load i32, i32* %Z "
|
|
|
216 " ret void "
|
|
|
217 "} "
|
|
|
218 " "
|
|
|
219 "define void @f_3() { "
|
|
|
220 " %x = load i32, i32* @var_0"
|
|
|
221 " %y = load i32, i32* @var_1"
|
|
|
222 " %z = load i32, i32* @var_2"
|
|
|
223 " ret void"
|
|
|
224 "} "
|
|
|
225 " "
|
|
|
226 "define void @f_4(i32 %a, i32 %b, i32 %c) { "
|
|
|
227 " %x = call i32 @unknown(i32 %a, i32 %b, i32 %c)"
|
|
|
228 " %y = call i32 @unknown(i32 %b, i32 %c, i32 %a)"
|
|
|
229 " %z = call i32 @unknown(i32 %c, i32 %a, i32 %b)"
|
|
|
230 " ret void"
|
|
|
231 "} "
|
|
|
232 ,
|
|
|
233 Err, C);
|
|
|
234
|
|
|
235 assert(M && "Could not parse module?");
|
|
|
236 assert(!verifyModule(*M) && "Must have been well formed!");
|
|
|
237
|
|
|
238 runWithSE(*M, "f_1", [&](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
239 auto *IV0 = getInstructionByName(F, "iv0");
|
|
|
240 auto *IV0Inc = getInstructionByName(F, "iv0.inc");
|
|
|
241
|
|
|
242 auto *FirstExprForIV0 = SE.getSCEV(IV0);
|
|
|
243 auto *FirstExprForIV0Inc = SE.getSCEV(IV0Inc);
|
|
|
244 auto *SecondExprForIV0 = SE.getSCEV(IV0);
|
|
|
245
|
|
|
246 EXPECT_TRUE(isa<SCEVAddRecExpr>(FirstExprForIV0));
|
|
|
247 EXPECT_TRUE(isa<SCEVAddRecExpr>(FirstExprForIV0Inc));
|
|
|
248 EXPECT_TRUE(isa<SCEVAddRecExpr>(SecondExprForIV0));
|
|
|
249 });
|
|
|
250
|
|
|
251 auto CheckCommutativeMulExprs = [&](ScalarEvolution &SE, const SCEV *A,
|
|
|
252 const SCEV *B, const SCEV *C) {
|
|
|
253 EXPECT_EQ(SE.getMulExpr(A, B), SE.getMulExpr(B, A));
|
|
|
254 EXPECT_EQ(SE.getMulExpr(B, C), SE.getMulExpr(C, B));
|
|
|
255 EXPECT_EQ(SE.getMulExpr(A, C), SE.getMulExpr(C, A));
|
|
|
256
|
|
|
257 SmallVector<const SCEV *, 3> Ops0 = {A, B, C};
|
|
|
258 SmallVector<const SCEV *, 3> Ops1 = {A, C, B};
|
|
|
259 SmallVector<const SCEV *, 3> Ops2 = {B, A, C};
|
|
|
260 SmallVector<const SCEV *, 3> Ops3 = {B, C, A};
|
|
|
261 SmallVector<const SCEV *, 3> Ops4 = {C, B, A};
|
|
|
262 SmallVector<const SCEV *, 3> Ops5 = {C, A, B};
|
|
|
263
|
|
|
264 auto *Mul0 = SE.getMulExpr(Ops0);
|
|
|
265 auto *Mul1 = SE.getMulExpr(Ops1);
|
|
|
266 auto *Mul2 = SE.getMulExpr(Ops2);
|
|
|
267 auto *Mul3 = SE.getMulExpr(Ops3);
|
|
|
268 auto *Mul4 = SE.getMulExpr(Ops4);
|
|
|
269 auto *Mul5 = SE.getMulExpr(Ops5);
|
|
|
270
|
|
|
271 EXPECT_EQ(Mul0, Mul1) << "Expected " << *Mul0 << " == " << *Mul1;
|
|
|
272 EXPECT_EQ(Mul1, Mul2) << "Expected " << *Mul1 << " == " << *Mul2;
|
|
|
273 EXPECT_EQ(Mul2, Mul3) << "Expected " << *Mul2 << " == " << *Mul3;
|
|
|
274 EXPECT_EQ(Mul3, Mul4) << "Expected " << *Mul3 << " == " << *Mul4;
|
|
|
275 EXPECT_EQ(Mul4, Mul5) << "Expected " << *Mul4 << " == " << *Mul5;
|
|
|
276 };
|
|
|
277
|
|
|
278 for (StringRef FuncName : {"f_2", "f_3", "f_4"})
|
|
|
279 runWithSE(
|
|
|
280 *M, FuncName, [&](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
281 CheckCommutativeMulExprs(SE, SE.getSCEV(getInstructionByName(F, "x")),
|
|
|
282 SE.getSCEV(getInstructionByName(F, "y")),
|
|
|
283 SE.getSCEV(getInstructionByName(F, "z")));
|
|
|
284 });
|
|
|
285 }
|
|
|
286
|
|
|
287 TEST_F(ScalarEvolutionsTest, CompareSCEVComplexity) {
|
|
|
288 FunctionType *FTy =
|
|
|
289 FunctionType::get(Type::getVoidTy(Context), std::vector<Type *>(), false);
|
|
|
290 Function *F = Function::Create(FTy, Function::ExternalLinkage, "f", M);
|
|
|
291 BasicBlock *EntryBB = BasicBlock::Create(Context, "entry", F);
|
|
|
292 BasicBlock *LoopBB = BasicBlock::Create(Context, "bb1", F);
|
|
|
293 BranchInst::Create(LoopBB, EntryBB);
|
|
|
294
|
|
|
295 auto *Ty = Type::getInt32Ty(Context);
|
|
|
296 SmallVector<Instruction*, 8> Muls(8), Acc(8), NextAcc(8);
|
|
|
297
|
|
|
298 Acc[0] = PHINode::Create(Ty, 2, "", LoopBB);
|
|
|
299 Acc[1] = PHINode::Create(Ty, 2, "", LoopBB);
|
|
|
300 Acc[2] = PHINode::Create(Ty, 2, "", LoopBB);
|
|
|
301 Acc[3] = PHINode::Create(Ty, 2, "", LoopBB);
|
|
|
302 Acc[4] = PHINode::Create(Ty, 2, "", LoopBB);
|
|
|
303 Acc[5] = PHINode::Create(Ty, 2, "", LoopBB);
|
|
|
304 Acc[6] = PHINode::Create(Ty, 2, "", LoopBB);
|
|
|
305 Acc[7] = PHINode::Create(Ty, 2, "", LoopBB);
|
|
|
306
|
|
|
307 for (int i = 0; i < 20; i++) {
|
|
|
308 Muls[0] = BinaryOperator::CreateMul(Acc[0], Acc[0], "", LoopBB);
|
|
|
309 NextAcc[0] = BinaryOperator::CreateAdd(Muls[0], Acc[4], "", LoopBB);
|
|
|
310 Muls[1] = BinaryOperator::CreateMul(Acc[1], Acc[1], "", LoopBB);
|
|
|
311 NextAcc[1] = BinaryOperator::CreateAdd(Muls[1], Acc[5], "", LoopBB);
|
|
|
312 Muls[2] = BinaryOperator::CreateMul(Acc[2], Acc[2], "", LoopBB);
|
|
|
313 NextAcc[2] = BinaryOperator::CreateAdd(Muls[2], Acc[6], "", LoopBB);
|
|
|
314 Muls[3] = BinaryOperator::CreateMul(Acc[3], Acc[3], "", LoopBB);
|
|
|
315 NextAcc[3] = BinaryOperator::CreateAdd(Muls[3], Acc[7], "", LoopBB);
|
|
|
316
|
|
|
317 Muls[4] = BinaryOperator::CreateMul(Acc[4], Acc[4], "", LoopBB);
|
|
|
318 NextAcc[4] = BinaryOperator::CreateAdd(Muls[4], Acc[0], "", LoopBB);
|
|
|
319 Muls[5] = BinaryOperator::CreateMul(Acc[5], Acc[5], "", LoopBB);
|
|
|
320 NextAcc[5] = BinaryOperator::CreateAdd(Muls[5], Acc[1], "", LoopBB);
|
|
|
321 Muls[6] = BinaryOperator::CreateMul(Acc[6], Acc[6], "", LoopBB);
|
|
|
322 NextAcc[6] = BinaryOperator::CreateAdd(Muls[6], Acc[2], "", LoopBB);
|
|
|
323 Muls[7] = BinaryOperator::CreateMul(Acc[7], Acc[7], "", LoopBB);
|
|
|
324 NextAcc[7] = BinaryOperator::CreateAdd(Muls[7], Acc[3], "", LoopBB);
|
|
|
325 Acc = NextAcc;
|
|
|
326 }
|
|
|
327
|
|
|
328 auto II = LoopBB->begin();
|
|
|
329 for (int i = 0; i < 8; i++) {
|
|
|
330 PHINode *Phi = cast<PHINode>(&*II++);
|
|
|
331 Phi->addIncoming(Acc[i], LoopBB);
|
|
|
332 Phi->addIncoming(UndefValue::get(Ty), EntryBB);
|
|
|
333 }
|
|
|
334
|
|
|
335 BasicBlock *ExitBB = BasicBlock::Create(Context, "bb2", F);
|
|
|
336 BranchInst::Create(LoopBB, ExitBB, UndefValue::get(Type::getInt1Ty(Context)),
|
|
|
337 LoopBB);
|
|
|
338
|
|
|
339 Acc[0] = BinaryOperator::CreateAdd(Acc[0], Acc[1], "", ExitBB);
|
|
|
340 Acc[1] = BinaryOperator::CreateAdd(Acc[2], Acc[3], "", ExitBB);
|
|
|
341 Acc[2] = BinaryOperator::CreateAdd(Acc[4], Acc[5], "", ExitBB);
|
|
|
342 Acc[3] = BinaryOperator::CreateAdd(Acc[6], Acc[7], "", ExitBB);
|
|
|
343 Acc[0] = BinaryOperator::CreateAdd(Acc[0], Acc[1], "", ExitBB);
|
|
|
344 Acc[1] = BinaryOperator::CreateAdd(Acc[2], Acc[3], "", ExitBB);
|
|
|
345 Acc[0] = BinaryOperator::CreateAdd(Acc[0], Acc[1], "", ExitBB);
|
|
|
346
|
|
|
347 ReturnInst::Create(Context, nullptr, ExitBB);
|
|
|
348
|
|
|
349 ScalarEvolution SE = buildSE(*F);
|
|
|
350
|
|
|
351 EXPECT_NE(nullptr, SE.getSCEV(Acc[0]));
|
|
|
352 }
|
|
|
353
|
|
|
354 TEST_F(ScalarEvolutionsTest, CompareValueComplexity) {
|
|
|
355 IntegerType *IntPtrTy = M.getDataLayout().getIntPtrType(Context);
|
|
|
356 PointerType *IntPtrPtrTy = IntPtrTy->getPointerTo();
|
|
|
357
|
|
|
358 FunctionType *FTy =
|
|
|
359 FunctionType::get(Type::getVoidTy(Context), {IntPtrTy, IntPtrTy}, false);
|
|
|
360 Function *F = Function::Create(FTy, Function::ExternalLinkage, "f", M);
|
|
|
361 BasicBlock *EntryBB = BasicBlock::Create(Context, "entry", F);
|
|
|
362
|
|
|
363 Value *X = &*F->arg_begin();
|
|
|
364 Value *Y = &*std::next(F->arg_begin());
|
|
|
365
|
|
|
366 const int ValueDepth = 10;
|
|
|
367 for (int i = 0; i < ValueDepth; i++) {
|
|
|
368 X = new LoadInst(IntPtrTy, new IntToPtrInst(X, IntPtrPtrTy, "", EntryBB),
|
|
|
369 "",
|
|
|
370 /*isVolatile*/ false, EntryBB);
|
|
|
371 Y = new LoadInst(IntPtrTy, new IntToPtrInst(Y, IntPtrPtrTy, "", EntryBB),
|
|
|
372 "",
|
|
|
373 /*isVolatile*/ false, EntryBB);
|
|
|
374 }
|
|
|
375
|
|
|
376 auto *MulA = BinaryOperator::CreateMul(X, Y, "", EntryBB);
|
|
|
377 auto *MulB = BinaryOperator::CreateMul(Y, X, "", EntryBB);
|
|
|
378 ReturnInst::Create(Context, nullptr, EntryBB);
|
|
|
379
|
|
|
380 // This test isn't checking for correctness. Today making A and B resolve to
|
|
|
381 // the same SCEV would require deeper searching in CompareValueComplexity,
|
|
|
382 // which will slow down compilation. However, this test can fail (with LLVM's
|
|
|
383 // behavior still being correct) if we ever have a smarter
|
|
|
384 // CompareValueComplexity that is both fast and more accurate.
|
|
|
385
|
|
|
386 ScalarEvolution SE = buildSE(*F);
|
|
|
387 auto *A = SE.getSCEV(MulA);
|
|
|
388 auto *B = SE.getSCEV(MulB);
|
|
|
389 EXPECT_NE(A, B);
|
|
|
390 }
|
|
|
391
|
|
|
392 TEST_F(ScalarEvolutionsTest, SCEVAddExpr) {
|
|
|
393 Type *Ty32 = Type::getInt32Ty(Context);
|
|
|
394 Type *ArgTys[] = {Type::getInt64Ty(Context), Ty32};
|
|
|
395
|
|
|
396 FunctionType *FTy =
|
|
|
397 FunctionType::get(Type::getVoidTy(Context), ArgTys, false);
|
|
|
398 Function *F = Function::Create(FTy, Function::ExternalLinkage, "f", M);
|
|
|
399
|
|
|
400 Argument *A1 = &*F->arg_begin();
|
|
|
401 Argument *A2 = &*(std::next(F->arg_begin()));
|
|
|
402 BasicBlock *EntryBB = BasicBlock::Create(Context, "entry", F);
|
|
|
403
|
|
|
404 Instruction *Trunc = CastInst::CreateTruncOrBitCast(A1, Ty32, "", EntryBB);
|
|
|
405 Instruction *Mul1 = BinaryOperator::CreateMul(Trunc, A2, "", EntryBB);
|
|
|
406 Instruction *Add1 = BinaryOperator::CreateAdd(Mul1, Trunc, "", EntryBB);
|
|
|
407 Mul1 = BinaryOperator::CreateMul(Add1, Trunc, "", EntryBB);
|
|
|
408 Instruction *Add2 = BinaryOperator::CreateAdd(Mul1, Add1, "", EntryBB);
|
|
|
409 // FIXME: The size of this is arbitrary and doesn't seem to change the
|
|
|
410 // result, but SCEV will do quadratic work for these so a large number here
|
|
|
411 // will be extremely slow. We should revisit what and how this is testing
|
|
|
412 // SCEV.
|
|
|
413 for (int i = 0; i < 10; i++) {
|
|
|
414 Mul1 = BinaryOperator::CreateMul(Add2, Add1, "", EntryBB);
|
|
|
415 Add1 = Add2;
|
|
|
416 Add2 = BinaryOperator::CreateAdd(Mul1, Add1, "", EntryBB);
|
|
|
417 }
|
|
|
418
|
|
|
419 ReturnInst::Create(Context, nullptr, EntryBB);
|
|
|
420 ScalarEvolution SE = buildSE(*F);
|
|
|
421 EXPECT_NE(nullptr, SE.getSCEV(Mul1));
|
|
|
422 }
|
|
|
423
|
|
|
424 static Instruction &GetInstByName(Function &F, StringRef Name) {
|
|
|
425 for (auto &I : instructions(F))
|
|
|
426 if (I.getName() == Name)
|
|
|
427 return I;
|
|
|
428 llvm_unreachable("Could not find instructions!");
|
|
|
429 }
|
|
|
430
|
|
|
431 TEST_F(ScalarEvolutionsTest, SCEVNormalization) {
|
|
|
432 LLVMContext C;
|
|
|
433 SMDiagnostic Err;
|
|
|
434 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
435 "target datalayout = \"e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128\" "
|
|
|
436 " "
|
|
|
437 "@var_0 = external global i32, align 4"
|
|
|
438 "@var_1 = external global i32, align 4"
|
|
|
439 "@var_2 = external global i32, align 4"
|
|
|
440 " "
|
|
|
441 "declare i32 @unknown(i32, i32, i32)"
|
|
|
442 " "
|
|
|
443 "define void @f_1(i8* nocapture %arr, i32 %n, i32* %A, i32* %B) "
|
|
|
444 " local_unnamed_addr { "
|
|
|
445 "entry: "
|
|
|
446 " br label %loop.ph "
|
|
|
447 " "
|
|
|
448 "loop.ph: "
|
|
|
449 " br label %loop "
|
|
|
450 " "
|
|
|
451 "loop: "
|
|
|
452 " %iv0 = phi i32 [ %iv0.inc, %loop ], [ 0, %loop.ph ] "
|
|
|
453 " %iv1 = phi i32 [ %iv1.inc, %loop ], [ -2147483648, %loop.ph ] "
|
|
|
454 " %iv0.inc = add i32 %iv0, 1 "
|
|
|
455 " %iv1.inc = add i32 %iv1, 3 "
|
|
|
456 " br i1 undef, label %for.end.loopexit, label %loop "
|
|
|
457 " "
|
|
|
458 "for.end.loopexit: "
|
|
|
459 " ret void "
|
|
|
460 "} "
|
|
|
461 " "
|
|
|
462 "define void @f_2(i32 %a, i32 %b, i32 %c, i32 %d) "
|
|
|
463 " local_unnamed_addr { "
|
|
|
464 "entry: "
|
|
|
465 " br label %loop_0 "
|
|
|
466 " "
|
|
|
467 "loop_0: "
|
|
|
468 " br i1 undef, label %loop_0, label %loop_1 "
|
|
|
469 " "
|
|
|
470 "loop_1: "
|
|
|
471 " br i1 undef, label %loop_2, label %loop_1 "
|
|
|
472 " "
|
|
|
473 " "
|
|
|
474 "loop_2: "
|
|
|
475 " br i1 undef, label %end, label %loop_2 "
|
|
|
476 " "
|
|
|
477 "end: "
|
|
|
478 " ret void "
|
|
|
479 "} "
|
|
|
480 ,
|
|
|
481 Err, C);
|
|
|
482
|
|
|
483 assert(M && "Could not parse module?");
|
|
|
484 assert(!verifyModule(*M) && "Must have been well formed!");
|
|
|
485
|
|
|
486 runWithSE(*M, "f_1", [&](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
487 auto &I0 = GetInstByName(F, "iv0");
|
|
|
488 auto &I1 = *I0.getNextNode();
|
|
|
489
|
|
|
490 auto *S0 = cast<SCEVAddRecExpr>(SE.getSCEV(&I0));
|
|
|
491 PostIncLoopSet Loops;
|
|
|
492 Loops.insert(S0->getLoop());
|
|
|
493 auto *N0 = normalizeForPostIncUse(S0, Loops, SE);
|
|
|
494 auto *D0 = denormalizeForPostIncUse(N0, Loops, SE);
|
|
|
495 EXPECT_EQ(S0, D0) << *S0 << " " << *D0;
|
|
|
496
|
|
|
497 auto *S1 = cast<SCEVAddRecExpr>(SE.getSCEV(&I1));
|
|
|
498 Loops.clear();
|
|
|
499 Loops.insert(S1->getLoop());
|
|
|
500 auto *N1 = normalizeForPostIncUse(S1, Loops, SE);
|
|
|
501 auto *D1 = denormalizeForPostIncUse(N1, Loops, SE);
|
|
|
502 EXPECT_EQ(S1, D1) << *S1 << " " << *D1;
|
|
|
503 });
|
|
|
504
|
|
|
505 runWithSE(*M, "f_2", [&](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
506 auto *L2 = *LI.begin();
|
|
|
507 auto *L1 = *std::next(LI.begin());
|
|
|
508 auto *L0 = *std::next(LI.begin(), 2);
|
|
|
509
|
|
|
510 auto GetAddRec = [&SE](const Loop *L, std::initializer_list<const SCEV *> Ops) {
|
|
|
511 SmallVector<const SCEV *, 4> OpsCopy(Ops);
|
|
|
512 return SE.getAddRecExpr(OpsCopy, L, SCEV::FlagAnyWrap);
|
|
|
513 };
|
|
|
514
|
|
|
515 auto GetAdd = [&SE](std::initializer_list<const SCEV *> Ops) {
|
|
|
516 SmallVector<const SCEV *, 4> OpsCopy(Ops);
|
|
|
517 return SE.getAddExpr(OpsCopy, SCEV::FlagAnyWrap);
|
|
|
518 };
|
|
|
519
|
|
|
520 // We first populate the AddRecs vector with a few "interesting" SCEV
|
|
|
521 // expressions, and then we go through the list and assert that each
|
|
|
522 // expression in it has an invertible normalization.
|
|
|
523
|
|
|
524 std::vector<const SCEV *> Exprs;
|
|
|
525 {
|
|
|
526 const SCEV *V0 = SE.getSCEV(&*F.arg_begin());
|
|
|
527 const SCEV *V1 = SE.getSCEV(&*std::next(F.arg_begin(), 1));
|
|
|
528 const SCEV *V2 = SE.getSCEV(&*std::next(F.arg_begin(), 2));
|
|
|
529 const SCEV *V3 = SE.getSCEV(&*std::next(F.arg_begin(), 3));
|
|
|
530
|
|
|
531 Exprs.push_back(GetAddRec(L0, {V0})); // 0
|
|
|
532 Exprs.push_back(GetAddRec(L0, {V0, V1})); // 1
|
|
|
533 Exprs.push_back(GetAddRec(L0, {V0, V1, V2})); // 2
|
|
|
534 Exprs.push_back(GetAddRec(L0, {V0, V1, V2, V3})); // 3
|
|
|
535
|
|
|
536 Exprs.push_back(
|
|
|
537 GetAddRec(L1, {Exprs[1], Exprs[2], Exprs[3], Exprs[0]})); // 4
|
|
|
538 Exprs.push_back(
|
|
|
539 GetAddRec(L1, {Exprs[1], Exprs[2], Exprs[0], Exprs[3]})); // 5
|
|
|
540 Exprs.push_back(
|
|
|
541 GetAddRec(L1, {Exprs[1], Exprs[3], Exprs[3], Exprs[1]})); // 6
|
|
|
542
|
|
|
543 Exprs.push_back(GetAdd({Exprs[6], Exprs[3], V2})); // 7
|
|
|
544
|
|
|
545 Exprs.push_back(
|
|
|
546 GetAddRec(L2, {Exprs[4], Exprs[3], Exprs[3], Exprs[5]})); // 8
|
|
|
547
|
|
|
548 Exprs.push_back(
|
|
|
549 GetAddRec(L2, {Exprs[4], Exprs[6], Exprs[7], Exprs[3], V0})); // 9
|
|
|
550 }
|
|
|
551
|
|
|
552 std::vector<PostIncLoopSet> LoopSets;
|
|
|
553 for (int i = 0; i < 8; i++) {
|
|
|
554 LoopSets.emplace_back();
|
|
|
555 if (i & 1)
|
|
|
556 LoopSets.back().insert(L0);
|
|
|
557 if (i & 2)
|
|
|
558 LoopSets.back().insert(L1);
|
|
|
559 if (i & 4)
|
|
|
560 LoopSets.back().insert(L2);
|
|
|
561 }
|
|
|
562
|
|
|
563 for (const auto &LoopSet : LoopSets)
|
|
|
564 for (auto *S : Exprs) {
|
|
|
565 {
|
|
|
566 auto *N = llvm::normalizeForPostIncUse(S, LoopSet, SE);
|
|
|
567 auto *D = llvm::denormalizeForPostIncUse(N, LoopSet, SE);
|
|
|
568
|
|
|
569 // Normalization and then denormalizing better give us back the same
|
|
|
570 // value.
|
|
|
571 EXPECT_EQ(S, D) << "S = " << *S << " D = " << *D << " N = " << *N;
|
|
|
572 }
|
|
|
573 {
|
|
|
574 auto *D = llvm::denormalizeForPostIncUse(S, LoopSet, SE);
|
|
|
575 auto *N = llvm::normalizeForPostIncUse(D, LoopSet, SE);
|
|
|
576
|
|
|
577 // Denormalization and then normalizing better give us back the same
|
|
|
578 // value.
|
|
|
579 EXPECT_EQ(S, N) << "S = " << *S << " N = " << *N;
|
|
|
580 }
|
|
|
581 }
|
|
|
582 });
|
|
|
583 }
|
|
|
584
|
|
|
585 // Expect the call of getZeroExtendExpr will not cost exponential time.
|
|
|
586 TEST_F(ScalarEvolutionsTest, SCEVZeroExtendExpr) {
|
|
|
587 LLVMContext C;
|
|
|
588 SMDiagnostic Err;
|
|
|
589
|
|
|
590 // Generate a function like below:
|
|
|
591 // define void @foo() {
|
|
|
592 // entry:
|
|
|
593 // br label %for.cond
|
|
|
594 //
|
|
|
595 // for.cond:
|
|
|
596 // %0 = phi i64 [ 100, %entry ], [ %dec, %for.inc ]
|
|
|
597 // %cmp = icmp sgt i64 %0, 90
|
|
|
598 // br i1 %cmp, label %for.inc, label %for.cond1
|
|
|
599 //
|
|
|
600 // for.inc:
|
|
|
601 // %dec = add nsw i64 %0, -1
|
|
|
602 // br label %for.cond
|
|
|
603 //
|
|
|
604 // for.cond1:
|
|
|
605 // %1 = phi i64 [ 100, %for.cond ], [ %dec5, %for.inc2 ]
|
|
|
606 // %cmp3 = icmp sgt i64 %1, 90
|
|
|
607 // br i1 %cmp3, label %for.inc2, label %for.cond4
|
|
|
608 //
|
|
|
609 // for.inc2:
|
|
|
610 // %dec5 = add nsw i64 %1, -1
|
|
|
611 // br label %for.cond1
|
|
|
612 //
|
|
|
613 // ......
|
|
|
614 //
|
|
|
615 // for.cond89:
|
|
|
616 // %19 = phi i64 [ 100, %for.cond84 ], [ %dec94, %for.inc92 ]
|
|
|
617 // %cmp93 = icmp sgt i64 %19, 90
|
|
|
618 // br i1 %cmp93, label %for.inc92, label %for.end
|
|
|
619 //
|
|
|
620 // for.inc92:
|
|
|
621 // %dec94 = add nsw i64 %19, -1
|
|
|
622 // br label %for.cond89
|
|
|
623 //
|
|
|
624 // for.end:
|
|
|
625 // %gep = getelementptr i8, i8* null, i64 %dec
|
|
|
626 // %gep6 = getelementptr i8, i8* %gep, i64 %dec5
|
|
|
627 // ......
|
|
|
628 // %gep95 = getelementptr i8, i8* %gep91, i64 %dec94
|
|
|
629 // ret void
|
|
|
630 // }
|
|
|
631 FunctionType *FTy = FunctionType::get(Type::getVoidTy(Context), {}, false);
|
|
|
632 Function *F = Function::Create(FTy, Function::ExternalLinkage, "foo", M);
|
|
|
633
|
|
|
634 BasicBlock *EntryBB = BasicBlock::Create(Context, "entry", F);
|
|
|
635 BasicBlock *CondBB = BasicBlock::Create(Context, "for.cond", F);
|
|
|
636 BasicBlock *EndBB = BasicBlock::Create(Context, "for.end", F);
|
|
|
637 BranchInst::Create(CondBB, EntryBB);
|
|
|
638 BasicBlock *PrevBB = EntryBB;
|
|
|
639
|
|
|
640 Type *I64Ty = Type::getInt64Ty(Context);
|
|
|
641 Type *I8Ty = Type::getInt8Ty(Context);
|
|
|
642 Type *I8PtrTy = Type::getInt8PtrTy(Context);
|
|
|
643 Value *Accum = Constant::getNullValue(I8PtrTy);
|
|
|
644 int Iters = 20;
|
|
|
645 for (int i = 0; i < Iters; i++) {
|
|
|
646 BasicBlock *IncBB = BasicBlock::Create(Context, "for.inc", F, EndBB);
|
|
|
647 auto *PN = PHINode::Create(I64Ty, 2, "", CondBB);
|
|
|
648 PN->addIncoming(ConstantInt::get(Context, APInt(64, 100)), PrevBB);
|
|
|
649 auto *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_SGT, PN,
|
|
|
650 ConstantInt::get(Context, APInt(64, 90)), "cmp",
|
|
|
651 CondBB);
|
|
|
652 BasicBlock *NextBB;
|
|
|
653 if (i != Iters - 1)
|
|
|
654 NextBB = BasicBlock::Create(Context, "for.cond", F, EndBB);
|
|
|
655 else
|
|
|
656 NextBB = EndBB;
|
|
|
657 BranchInst::Create(IncBB, NextBB, Cmp, CondBB);
|
|
|
658 auto *Dec = BinaryOperator::CreateNSWAdd(
|
|
|
659 PN, ConstantInt::get(Context, APInt(64, -1)), "dec", IncBB);
|
|
|
660 PN->addIncoming(Dec, IncBB);
|
|
|
661 BranchInst::Create(CondBB, IncBB);
|
|
|
662
|
|
|
663 Accum = GetElementPtrInst::Create(I8Ty, Accum, PN, "gep", EndBB);
|
|
|
664
|
|
|
665 PrevBB = CondBB;
|
|
|
666 CondBB = NextBB;
|
|
|
667 }
|
|
|
668 ReturnInst::Create(Context, nullptr, EndBB);
|
|
|
669 ScalarEvolution SE = buildSE(*F);
|
|
|
670 const SCEV *S = SE.getSCEV(Accum);
|
|
|
671 Type *I128Ty = Type::getInt128Ty(Context);
|
|
|
672 SE.getZeroExtendExpr(S, I128Ty);
|
|
|
673 }
|
|
|
674
|
|
|
675 // Make sure that SCEV invalidates exit limits after invalidating the values it
|
|
|
676 // depends on when we forget a loop.
|
|
|
677 TEST_F(ScalarEvolutionsTest, SCEVExitLimitForgetLoop) {
|
|
|
678 /*
|
|
|
679 * Create the following code:
|
|
|
680 * func(i64 addrspace(10)* %arg)
|
|
|
681 * top:
|
|
|
682 * br label %L.ph
|
|
|
683 * L.ph:
|
|
|
684 * br label %L
|
|
|
685 * L:
|
|
|
686 * %phi = phi i64 [i64 0, %L.ph], [ %add, %L2 ]
|
|
|
687 * %add = add i64 %phi2, 1
|
|
|
688 * %cond = icmp slt i64 %add, 1000; then becomes 2000.
|
|
|
689 * br i1 %cond, label %post, label %L2
|
|
|
690 * post:
|
|
|
691 * ret void
|
|
|
692 *
|
|
|
693 */
|
|
|
694
|
|
|
695 // Create a module with non-integral pointers in it's datalayout
|
|
|
696 Module NIM("nonintegral", Context);
|
|
|
697 std::string DataLayout = M.getDataLayoutStr();
|
|
|
698 if (!DataLayout.empty())
|
|
|
699 DataLayout += "-";
|
|
|
700 DataLayout += "ni:10";
|
|
|
701 NIM.setDataLayout(DataLayout);
|
|
|
702
|
|
|
703 Type *T_int64 = Type::getInt64Ty(Context);
|
|
|
704 Type *T_pint64 = T_int64->getPointerTo(10);
|
|
|
705
|
|
|
706 FunctionType *FTy =
|
|
|
707 FunctionType::get(Type::getVoidTy(Context), {T_pint64}, false);
|
|
|
708 Function *F = Function::Create(FTy, Function::ExternalLinkage, "foo", NIM);
|
|
|
709
|
|
|
710 BasicBlock *Top = BasicBlock::Create(Context, "top", F);
|
|
|
711 BasicBlock *LPh = BasicBlock::Create(Context, "L.ph", F);
|
|
|
712 BasicBlock *L = BasicBlock::Create(Context, "L", F);
|
|
|
713 BasicBlock *Post = BasicBlock::Create(Context, "post", F);
|
|
|
714
|
|
|
715 IRBuilder<> Builder(Top);
|
|
|
716 Builder.CreateBr(LPh);
|
|
|
717
|
|
|
718 Builder.SetInsertPoint(LPh);
|
|
|
719 Builder.CreateBr(L);
|
|
|
720
|
|
|
721 Builder.SetInsertPoint(L);
|
|
|
722 PHINode *Phi = Builder.CreatePHI(T_int64, 2);
|
|
|
723 auto *Add = cast<Instruction>(
|
|
|
724 Builder.CreateAdd(Phi, ConstantInt::get(T_int64, 1), "add"));
|
|
|
725 auto *Limit = ConstantInt::get(T_int64, 1000);
|
|
|
726 auto *Cond = cast<Instruction>(
|
|
|
727 Builder.CreateICmp(ICmpInst::ICMP_SLT, Add, Limit, "cond"));
|
|
|
728 auto *Br = cast<Instruction>(Builder.CreateCondBr(Cond, L, Post));
|
|
|
729 Phi->addIncoming(ConstantInt::get(T_int64, 0), LPh);
|
|
|
730 Phi->addIncoming(Add, L);
|
|
|
731
|
|
|
732 Builder.SetInsertPoint(Post);
|
|
|
733 Builder.CreateRetVoid();
|
|
|
734
|
|
|
735 ScalarEvolution SE = buildSE(*F);
|
|
|
736 auto *Loop = LI->getLoopFor(L);
|
|
|
737 const SCEV *EC = SE.getBackedgeTakenCount(Loop);
|
|
|
738 EXPECT_FALSE(isa<SCEVCouldNotCompute>(EC));
|
|
|
739 EXPECT_TRUE(isa<SCEVConstant>(EC));
|
|
|
740 EXPECT_EQ(cast<SCEVConstant>(EC)->getAPInt().getLimitedValue(), 999u);
|
|
|
741
|
|
|
742 // The add recurrence {5,+,1} does not correspond to any PHI in the IR, and
|
|
|
743 // that is relevant to this test.
|
|
|
744 auto *Five = SE.getConstant(APInt(/*numBits=*/64, 5));
|
|
|
745 auto *AR =
|
|
|
746 SE.getAddRecExpr(Five, SE.getOne(T_int64), Loop, SCEV::FlagAnyWrap);
|
|
|
747 const SCEV *ARAtLoopExit = SE.getSCEVAtScope(AR, nullptr);
|
|
|
748 EXPECT_FALSE(isa<SCEVCouldNotCompute>(ARAtLoopExit));
|
|
|
749 EXPECT_TRUE(isa<SCEVConstant>(ARAtLoopExit));
|
|
|
750 EXPECT_EQ(cast<SCEVConstant>(ARAtLoopExit)->getAPInt().getLimitedValue(),
|
|
|
751 1004u);
|
|
|
752
|
|
|
753 SE.forgetLoop(Loop);
|
|
|
754 Br->eraseFromParent();
|
|
|
755 Cond->eraseFromParent();
|
|
|
756
|
|
|
757 Builder.SetInsertPoint(L);
|
|
|
758 auto *NewCond = Builder.CreateICmp(
|
|
|
759 ICmpInst::ICMP_SLT, Add, ConstantInt::get(T_int64, 2000), "new.cond");
|
|
|
760 Builder.CreateCondBr(NewCond, L, Post);
|
|
|
761 const SCEV *NewEC = SE.getBackedgeTakenCount(Loop);
|
|
|
762 EXPECT_FALSE(isa<SCEVCouldNotCompute>(NewEC));
|
|
|
763 EXPECT_TRUE(isa<SCEVConstant>(NewEC));
|
|
|
764 EXPECT_EQ(cast<SCEVConstant>(NewEC)->getAPInt().getLimitedValue(), 1999u);
|
|
|
765 const SCEV *NewARAtLoopExit = SE.getSCEVAtScope(AR, nullptr);
|
|
|
766 EXPECT_FALSE(isa<SCEVCouldNotCompute>(NewARAtLoopExit));
|
|
|
767 EXPECT_TRUE(isa<SCEVConstant>(NewARAtLoopExit));
|
|
|
768 EXPECT_EQ(cast<SCEVConstant>(NewARAtLoopExit)->getAPInt().getLimitedValue(),
|
|
|
769 2004u);
|
|
|
770 }
|
|
|
771
|
|
|
772 // Make sure that SCEV invalidates exit limits after invalidating the values it
|
|
|
773 // depends on when we forget a value.
|
|
|
774 TEST_F(ScalarEvolutionsTest, SCEVExitLimitForgetValue) {
|
|
|
775 /*
|
|
|
776 * Create the following code:
|
|
|
777 * func(i64 addrspace(10)* %arg)
|
|
|
778 * top:
|
|
|
779 * br label %L.ph
|
|
|
780 * L.ph:
|
|
|
781 * %load = load i64 addrspace(10)* %arg
|
|
|
782 * br label %L
|
|
|
783 * L:
|
|
|
784 * %phi = phi i64 [i64 0, %L.ph], [ %add, %L2 ]
|
|
|
785 * %add = add i64 %phi2, 1
|
|
|
786 * %cond = icmp slt i64 %add, %load ; then becomes 2000.
|
|
|
787 * br i1 %cond, label %post, label %L2
|
|
|
788 * post:
|
|
|
789 * ret void
|
|
|
790 *
|
|
|
791 */
|
|
|
792
|
|
|
793 // Create a module with non-integral pointers in it's datalayout
|
|
|
794 Module NIM("nonintegral", Context);
|
|
|
795 std::string DataLayout = M.getDataLayoutStr();
|
|
|
796 if (!DataLayout.empty())
|
|
|
797 DataLayout += "-";
|
|
|
798 DataLayout += "ni:10";
|
|
|
799 NIM.setDataLayout(DataLayout);
|
|
|
800
|
|
|
801 Type *T_int64 = Type::getInt64Ty(Context);
|
|
|
802 Type *T_pint64 = T_int64->getPointerTo(10);
|
|
|
803
|
|
|
804 FunctionType *FTy =
|
|
|
805 FunctionType::get(Type::getVoidTy(Context), {T_pint64}, false);
|
|
|
806 Function *F = Function::Create(FTy, Function::ExternalLinkage, "foo", NIM);
|
|
|
807
|
|
|
808 Argument *Arg = &*F->arg_begin();
|
|
|
809
|
|
|
810 BasicBlock *Top = BasicBlock::Create(Context, "top", F);
|
|
|
811 BasicBlock *LPh = BasicBlock::Create(Context, "L.ph", F);
|
|
|
812 BasicBlock *L = BasicBlock::Create(Context, "L", F);
|
|
|
813 BasicBlock *Post = BasicBlock::Create(Context, "post", F);
|
|
|
814
|
|
|
815 IRBuilder<> Builder(Top);
|
|
|
816 Builder.CreateBr(LPh);
|
|
|
817
|
|
|
818 Builder.SetInsertPoint(LPh);
|
|
|
819 auto *Load = cast<Instruction>(Builder.CreateLoad(T_int64, Arg, "load"));
|
|
|
820 Builder.CreateBr(L);
|
|
|
821
|
|
|
822 Builder.SetInsertPoint(L);
|
|
|
823 PHINode *Phi = Builder.CreatePHI(T_int64, 2);
|
|
|
824 auto *Add = cast<Instruction>(
|
|
|
825 Builder.CreateAdd(Phi, ConstantInt::get(T_int64, 1), "add"));
|
|
|
826 auto *Cond = cast<Instruction>(
|
|
|
827 Builder.CreateICmp(ICmpInst::ICMP_SLT, Add, Load, "cond"));
|
|
|
828 auto *Br = cast<Instruction>(Builder.CreateCondBr(Cond, L, Post));
|
|
|
829 Phi->addIncoming(ConstantInt::get(T_int64, 0), LPh);
|
|
|
830 Phi->addIncoming(Add, L);
|
|
|
831
|
|
|
832 Builder.SetInsertPoint(Post);
|
|
|
833 Builder.CreateRetVoid();
|
|
|
834
|
|
|
835 ScalarEvolution SE = buildSE(*F);
|
|
|
836 auto *Loop = LI->getLoopFor(L);
|
|
|
837 const SCEV *EC = SE.getBackedgeTakenCount(Loop);
|
|
|
838 EXPECT_FALSE(isa<SCEVCouldNotCompute>(EC));
|
|
|
839 EXPECT_FALSE(isa<SCEVConstant>(EC));
|
|
|
840
|
|
|
841 SE.forgetValue(Load);
|
|
|
842 Br->eraseFromParent();
|
|
|
843 Cond->eraseFromParent();
|
|
|
844 Load->eraseFromParent();
|
|
|
845
|
|
|
846 Builder.SetInsertPoint(L);
|
|
|
847 auto *NewCond = Builder.CreateICmp(
|
|
|
848 ICmpInst::ICMP_SLT, Add, ConstantInt::get(T_int64, 2000), "new.cond");
|
|
|
849 Builder.CreateCondBr(NewCond, L, Post);
|
|
|
850 const SCEV *NewEC = SE.getBackedgeTakenCount(Loop);
|
|
|
851 EXPECT_FALSE(isa<SCEVCouldNotCompute>(NewEC));
|
|
|
852 EXPECT_TRUE(isa<SCEVConstant>(NewEC));
|
|
|
853 EXPECT_EQ(cast<SCEVConstant>(NewEC)->getAPInt().getLimitedValue(), 1999u);
|
|
|
854 }
|
|
|
855
|
|
|
856 TEST_F(ScalarEvolutionsTest, SCEVAddRecFromPHIwithLargeConstants) {
|
|
|
857 // Reference: https://reviews.llvm.org/D37265
|
|
|
858 // Make sure that SCEV does not blow up when constructing an AddRec
|
|
|
859 // with predicates for a phi with the update pattern:
|
|
|
860 // (SExt/ZExt ix (Trunc iy (%SymbolicPHI) to ix) to iy) + InvariantAccum
|
|
|
861 // when either the initial value of the Phi or the InvariantAccum are
|
|
|
862 // constants that are too large to fit in an ix but are zero when truncated to
|
|
|
863 // ix.
|
|
|
864 FunctionType *FTy =
|
|
|
865 FunctionType::get(Type::getVoidTy(Context), std::vector<Type *>(), false);
|
|
|
866 Function *F =
|
|
|
867 Function::Create(FTy, Function::ExternalLinkage, "addrecphitest", M);
|
|
|
868
|
|
|
869 /*
|
|
|
870 Create IR:
|
|
|
871 entry:
|
|
|
872 br label %loop
|
|
|
873 loop:
|
|
|
874 %0 = phi i64 [-9223372036854775808, %entry], [%3, %loop]
|
|
|
875 %1 = shl i64 %0, 32
|
|
|
876 %2 = ashr exact i64 %1, 32
|
|
|
877 %3 = add i64 %2, -9223372036854775808
|
|
|
878 br i1 undef, label %exit, label %loop
|
|
|
879 exit:
|
|
|
880 ret void
|
|
|
881 */
|
|
|
882 BasicBlock *EntryBB = BasicBlock::Create(Context, "entry", F);
|
|
|
883 BasicBlock *LoopBB = BasicBlock::Create(Context, "loop", F);
|
|
|
884 BasicBlock *ExitBB = BasicBlock::Create(Context, "exit", F);
|
|
|
885
|
|
|
886 // entry:
|
|
|
887 BranchInst::Create(LoopBB, EntryBB);
|
|
|
888 // loop:
|
|
|
889 auto *MinInt64 =
|
|
|
890 ConstantInt::get(Context, APInt(64, 0x8000000000000000U, true));
|
|
|
891 auto *Int64_32 = ConstantInt::get(Context, APInt(64, 32));
|
|
|
892 auto *Br = BranchInst::Create(
|
|
|
893 LoopBB, ExitBB, UndefValue::get(Type::getInt1Ty(Context)), LoopBB);
|
|
|
894 auto *Phi = PHINode::Create(Type::getInt64Ty(Context), 2, "", Br);
|
|
|
895 auto *Shl = BinaryOperator::CreateShl(Phi, Int64_32, "", Br);
|
|
|
896 auto *AShr = BinaryOperator::CreateExactAShr(Shl, Int64_32, "", Br);
|
|
|
897 auto *Add = BinaryOperator::CreateAdd(AShr, MinInt64, "", Br);
|
|
|
898 Phi->addIncoming(MinInt64, EntryBB);
|
|
|
899 Phi->addIncoming(Add, LoopBB);
|
|
|
900 // exit:
|
|
|
901 ReturnInst::Create(Context, nullptr, ExitBB);
|
|
|
902
|
|
|
903 // Make sure that SCEV doesn't blow up
|
|
|
904 ScalarEvolution SE = buildSE(*F);
|
|
|
905 SCEVUnionPredicate Preds;
|
|
|
906 const SCEV *Expr = SE.getSCEV(Phi);
|
|
|
907 EXPECT_NE(nullptr, Expr);
|
|
|
908 EXPECT_TRUE(isa<SCEVUnknown>(Expr));
|
|
|
909 auto Result = SE.createAddRecFromPHIWithCasts(cast<SCEVUnknown>(Expr));
|
|
|
910 }
|
|
|
911
|
|
|
912 TEST_F(ScalarEvolutionsTest, SCEVAddRecFromPHIwithLargeConstantAccum) {
|
|
|
913 // Make sure that SCEV does not blow up when constructing an AddRec
|
|
|
914 // with predicates for a phi with the update pattern:
|
|
|
915 // (SExt/ZExt ix (Trunc iy (%SymbolicPHI) to ix) to iy) + InvariantAccum
|
|
|
916 // when the InvariantAccum is a constant that is too large to fit in an
|
|
|
917 // ix but are zero when truncated to ix, and the initial value of the
|
|
|
918 // phi is not a constant.
|
|
|
919 Type *Int32Ty = Type::getInt32Ty(Context);
|
|
|
920 SmallVector<Type *, 1> Types;
|
|
|
921 Types.push_back(Int32Ty);
|
|
|
922 FunctionType *FTy = FunctionType::get(Type::getVoidTy(Context), Types, false);
|
|
|
923 Function *F =
|
|
|
924 Function::Create(FTy, Function::ExternalLinkage, "addrecphitest", M);
|
|
|
925
|
|
|
926 /*
|
|
|
927 Create IR:
|
|
|
928 define @addrecphitest(i32)
|
|
|
929 entry:
|
|
|
930 br label %loop
|
|
|
931 loop:
|
|
|
932 %1 = phi i32 [%0, %entry], [%4, %loop]
|
|
|
933 %2 = shl i32 %1, 16
|
|
|
934 %3 = ashr exact i32 %2, 16
|
|
|
935 %4 = add i32 %3, -2147483648
|
|
|
936 br i1 undef, label %exit, label %loop
|
|
|
937 exit:
|
|
|
938 ret void
|
|
|
939 */
|
|
|
940 BasicBlock *EntryBB = BasicBlock::Create(Context, "entry", F);
|
|
|
941 BasicBlock *LoopBB = BasicBlock::Create(Context, "loop", F);
|
|
|
942 BasicBlock *ExitBB = BasicBlock::Create(Context, "exit", F);
|
|
|
943
|
|
|
944 // entry:
|
|
|
945 BranchInst::Create(LoopBB, EntryBB);
|
|
|
946 // loop:
|
|
|
947 auto *MinInt32 = ConstantInt::get(Context, APInt(32, 0x80000000U, true));
|
|
|
948 auto *Int32_16 = ConstantInt::get(Context, APInt(32, 16));
|
|
|
949 auto *Br = BranchInst::Create(
|
|
|
950 LoopBB, ExitBB, UndefValue::get(Type::getInt1Ty(Context)), LoopBB);
|
|
|
951 auto *Phi = PHINode::Create(Int32Ty, 2, "", Br);
|
|
|
952 auto *Shl = BinaryOperator::CreateShl(Phi, Int32_16, "", Br);
|
|
|
953 auto *AShr = BinaryOperator::CreateExactAShr(Shl, Int32_16, "", Br);
|
|
|
954 auto *Add = BinaryOperator::CreateAdd(AShr, MinInt32, "", Br);
|
|
|
955 auto *Arg = &*(F->arg_begin());
|
|
|
956 Phi->addIncoming(Arg, EntryBB);
|
|
|
957 Phi->addIncoming(Add, LoopBB);
|
|
|
958 // exit:
|
|
|
959 ReturnInst::Create(Context, nullptr, ExitBB);
|
|
|
960
|
|
|
961 // Make sure that SCEV doesn't blow up
|
|
|
962 ScalarEvolution SE = buildSE(*F);
|
|
|
963 SCEVUnionPredicate Preds;
|
|
|
964 const SCEV *Expr = SE.getSCEV(Phi);
|
|
|
965 EXPECT_NE(nullptr, Expr);
|
|
|
966 EXPECT_TRUE(isa<SCEVUnknown>(Expr));
|
|
|
967 auto Result = SE.createAddRecFromPHIWithCasts(cast<SCEVUnknown>(Expr));
|
|
|
968 }
|
|
|
969
|
|
|
970 TEST_F(ScalarEvolutionsTest, SCEVFoldSumOfTruncs) {
|
|
|
971 // Verify that the following SCEV gets folded to a zero:
|
|
|
972 // (-1 * (trunc i64 (-1 * %0) to i32)) + (-1 * (trunc i64 %0 to i32)
|
|
|
973 Type *ArgTy = Type::getInt64Ty(Context);
|
|
|
974 Type *Int32Ty = Type::getInt32Ty(Context);
|
|
|
975 SmallVector<Type *, 1> Types;
|
|
|
976 Types.push_back(ArgTy);
|
|
|
977 FunctionType *FTy = FunctionType::get(Type::getVoidTy(Context), Types, false);
|
|
|
978 Function *F = Function::Create(FTy, Function::ExternalLinkage, "f", M);
|
|
|
979 BasicBlock *BB = BasicBlock::Create(Context, "entry", F);
|
|
|
980 ReturnInst::Create(Context, nullptr, BB);
|
|
|
981
|
|
|
982 ScalarEvolution SE = buildSE(*F);
|
|
|
983
|
|
|
984 auto *Arg = &*(F->arg_begin());
|
|
|
985 const auto *ArgSCEV = SE.getSCEV(Arg);
|
|
|
986
|
|
|
987 // Build the SCEV
|
|
|
988 const auto *A0 = SE.getNegativeSCEV(ArgSCEV);
|
|
|
989 const auto *A1 = SE.getTruncateExpr(A0, Int32Ty);
|
|
|
990 const auto *A = SE.getNegativeSCEV(A1);
|
|
|
991
|
|
|
992 const auto *B0 = SE.getTruncateExpr(ArgSCEV, Int32Ty);
|
|
|
993 const auto *B = SE.getNegativeSCEV(B0);
|
|
|
994
|
|
|
995 const auto *Expr = SE.getAddExpr(A, B);
|
|
|
996 // Verify that the SCEV was folded to 0
|
|
|
997 const auto *ZeroConst = SE.getConstant(Int32Ty, 0);
|
|
|
998 EXPECT_EQ(Expr, ZeroConst);
|
|
|
999 }
|
|
|
1000
|
|
|
1001 // Check logic of SCEV expression size computation.
|
|
|
1002 TEST_F(ScalarEvolutionsTest, SCEVComputeExpressionSize) {
|
|
|
1003 /*
|
|
|
1004 * Create the following code:
|
|
|
1005 * void func(i64 %a, i64 %b)
|
|
|
1006 * entry:
|
|
|
1007 * %s1 = add i64 %a, 1
|
|
|
1008 * %s2 = udiv i64 %s1, %b
|
|
|
1009 * br label %exit
|
|
|
1010 * exit:
|
|
|
1011 * ret
|
|
|
1012 */
|
|
|
1013
|
|
|
1014 // Create a module.
|
|
|
1015 Module M("SCEVComputeExpressionSize", Context);
|
|
|
1016
|
|
|
1017 Type *T_int64 = Type::getInt64Ty(Context);
|
|
|
1018
|
|
|
1019 FunctionType *FTy =
|
|
|
1020 FunctionType::get(Type::getVoidTy(Context), { T_int64, T_int64 }, false);
|
|
|
1021 Function *F = Function::Create(FTy, Function::ExternalLinkage, "func", M);
|
|
|
1022 Argument *A = &*F->arg_begin();
|
|
|
1023 Argument *B = &*std::next(F->arg_begin());
|
|
|
1024 ConstantInt *C = ConstantInt::get(Context, APInt(64, 1));
|
|
|
1025
|
|
|
1026 BasicBlock *Entry = BasicBlock::Create(Context, "entry", F);
|
|
|
1027 BasicBlock *Exit = BasicBlock::Create(Context, "exit", F);
|
|
|
1028
|
|
|
1029 IRBuilder<> Builder(Entry);
|
|
|
1030 auto *S1 = cast<Instruction>(Builder.CreateAdd(A, C, "s1"));
|
|
|
1031 auto *S2 = cast<Instruction>(Builder.CreateUDiv(S1, B, "s2"));
|
|
|
1032 Builder.CreateBr(Exit);
|
|
|
1033
|
|
|
1034 Builder.SetInsertPoint(Exit);
|
|
|
1035 Builder.CreateRetVoid();
|
|
|
1036
|
|
|
1037 ScalarEvolution SE = buildSE(*F);
|
|
|
1038 // Get S2 first to move it to cache.
|
|
|
1039 const SCEV *AS = SE.getSCEV(A);
|
|
|
1040 const SCEV *BS = SE.getSCEV(B);
|
|
|
1041 const SCEV *CS = SE.getSCEV(C);
|
|
|
1042 const SCEV *S1S = SE.getSCEV(S1);
|
|
|
1043 const SCEV *S2S = SE.getSCEV(S2);
|
|
|
1044 EXPECT_EQ(AS->getExpressionSize(), 1u);
|
|
|
1045 EXPECT_EQ(BS->getExpressionSize(), 1u);
|
|
|
1046 EXPECT_EQ(CS->getExpressionSize(), 1u);
|
|
|
1047 EXPECT_EQ(S1S->getExpressionSize(), 3u);
|
|
|
1048 EXPECT_EQ(S2S->getExpressionSize(), 5u);
|
|
|
1049 }
|
|
|
1050
|
|
|
1051 TEST_F(ScalarEvolutionsTest, SCEVLoopDecIntrinsic) {
|
|
|
1052 LLVMContext C;
|
|
|
1053 SMDiagnostic Err;
|
|
|
1054 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
1055 "define void @foo(i32 %N) { "
|
|
|
1056 "entry: "
|
|
|
1057 " %cmp3 = icmp sgt i32 %N, 0 "
|
|
|
1058 " br i1 %cmp3, label %for.body, label %for.cond.cleanup "
|
|
|
1059 "for.cond.cleanup: "
|
|
|
1060 " ret void "
|
|
|
1061 "for.body: "
|
|
|
1062 " %i.04 = phi i32 [ %inc, %for.body ], [ 100, %entry ] "
|
|
|
1063 " %inc = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 %i.04, i32 1) "
|
|
|
1064 " %exitcond = icmp ne i32 %inc, 0 "
|
|
|
1065 " br i1 %exitcond, label %for.cond.cleanup, label %for.body "
|
|
|
1066 "} "
|
|
|
1067 "declare i32 @llvm.loop.decrement.reg.i32.i32.i32(i32, i32) ",
|
|
|
1068 Err, C);
|
|
|
1069
|
|
|
1070 ASSERT_TRUE(M && "Could not parse module?");
|
|
|
1071 ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1072
|
|
|
1073 runWithSE(*M, "foo", [&](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1074 auto *ScevInc = SE.getSCEV(getInstructionByName(F, "inc"));
|
|
|
1075 EXPECT_TRUE(isa<SCEVAddRecExpr>(ScevInc));
|
|
|
1076 });
|
|
|
1077 }
|
|
|
1078
|
|
|
1079 TEST_F(ScalarEvolutionsTest, SCEVComputeConstantDifference) {
|
|
|
1080 LLVMContext C;
|
|
|
1081 SMDiagnostic Err;
|
|
|
1082 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
1083 "define void @foo(i32 %sz, i32 %pp) { "
|
|
|
1084 "entry: "
|
|
|
1085 " %v0 = add i32 %pp, 0 "
|
|
|
1086 " %v3 = add i32 %pp, 3 "
|
|
|
1087 " br label %loop.body "
|
|
|
1088 "loop.body: "
|
|
|
1089 " %iv = phi i32 [ %iv.next, %loop.body ], [ 0, %entry ] "
|
|
|
1090 " %xa = add nsw i32 %iv, %v0 "
|
|
|
1091 " %yy = add nsw i32 %iv, %v3 "
|
|
|
1092 " %xb = sub nsw i32 %yy, 3 "
|
|
|
1093 " %iv.next = add nsw i32 %iv, 1 "
|
|
|
1094 " %cmp = icmp sle i32 %iv.next, %sz "
|
|
|
1095 " br i1 %cmp, label %loop.body, label %exit "
|
|
|
1096 "exit: "
|
|
|
1097 " ret void "
|
|
|
1098 "} ",
|
|
|
1099 Err, C);
|
|
|
1100
|
|
|
1101 ASSERT_TRUE(M && "Could not parse module?");
|
|
|
1102 ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1103
|
|
|
1104 runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1105 auto *ScevV0 = SE.getSCEV(getInstructionByName(F, "v0")); // %pp
|
|
|
1106 auto *ScevV3 = SE.getSCEV(getInstructionByName(F, "v3")); // (3 + %pp)
|
|
|
1107 auto *ScevIV = SE.getSCEV(getInstructionByName(F, "iv")); // {0,+,1}
|
|
|
1108 auto *ScevXA = SE.getSCEV(getInstructionByName(F, "xa")); // {%pp,+,1}
|
|
|
1109 auto *ScevYY = SE.getSCEV(getInstructionByName(F, "yy")); // {(3 + %pp),+,1}
|
|
|
1110 auto *ScevXB = SE.getSCEV(getInstructionByName(F, "xb")); // {%pp,+,1}
|
|
|
1111 auto *ScevIVNext = SE.getSCEV(getInstructionByName(F, "iv.next")); // {1,+,1}
|
|
|
1112
|
|
|
1113 auto diff = [&SE](const SCEV *LHS, const SCEV *RHS) -> Optional<int> {
|
|
|
1114 auto ConstantDiffOrNone = computeConstantDifference(SE, LHS, RHS);
|
|
|
1115 if (!ConstantDiffOrNone)
|
|
|
1116 return None;
|
|
|
1117
|
|
|
1118 auto ExtDiff = ConstantDiffOrNone->getSExtValue();
|
|
|
1119 int Diff = ExtDiff;
|
|
|
1120 assert(Diff == ExtDiff && "Integer overflow");
|
|
|
1121 return Diff;
|
|
|
1122 };
|
|
|
1123
|
|
|
1124 EXPECT_EQ(diff(ScevV3, ScevV0), 3);
|
|
|
1125 EXPECT_EQ(diff(ScevV0, ScevV3), -3);
|
|
|
1126 EXPECT_EQ(diff(ScevV0, ScevV0), 0);
|
|
|
1127 EXPECT_EQ(diff(ScevV3, ScevV3), 0);
|
|
|
1128 EXPECT_EQ(diff(ScevIV, ScevIV), 0);
|
|
|
1129 EXPECT_EQ(diff(ScevXA, ScevXB), 0);
|
|
|
1130 EXPECT_EQ(diff(ScevXA, ScevYY), -3);
|
|
|
1131 EXPECT_EQ(diff(ScevYY, ScevXB), 3);
|
|
|
1132 EXPECT_EQ(diff(ScevIV, ScevIVNext), -1);
|
|
|
1133 EXPECT_EQ(diff(ScevIVNext, ScevIV), 1);
|
|
|
1134 EXPECT_EQ(diff(ScevIVNext, ScevIVNext), 0);
|
|
|
1135 EXPECT_EQ(diff(ScevV0, ScevIV), None);
|
|
|
1136 EXPECT_EQ(diff(ScevIVNext, ScevV3), None);
|
|
|
1137 EXPECT_EQ(diff(ScevYY, ScevV3), None);
|
|
|
1138 });
|
|
|
1139 }
|
|
|
1140
|
|
221
|
1141 TEST_F(ScalarEvolutionsTest, SCEVrewriteUnknowns) {
|
|
|
1142 LLVMContext C;
|
|
|
1143 SMDiagnostic Err;
|
|
|
1144 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
1145 "define void @foo(i32 %i) { "
|
|
|
1146 "entry: "
|
|
|
1147 " %cmp3 = icmp ult i32 %i, 16 "
|
|
|
1148 " br i1 %cmp3, label %loop.body, label %exit "
|
|
|
1149 "loop.body: "
|
|
|
1150 " %iv = phi i32 [ %iv.next, %loop.body ], [ %i, %entry ] "
|
|
|
1151 " %iv.next = add nsw i32 %iv, 1 "
|
|
|
1152 " %cmp = icmp eq i32 %iv.next, 16 "
|
|
|
1153 " br i1 %cmp, label %exit, label %loop.body "
|
|
|
1154 "exit: "
|
|
|
1155 " ret void "
|
|
|
1156 "} ",
|
|
|
1157 Err, C);
|
|
|
1158
|
|
|
1159 ASSERT_TRUE(M && "Could not parse module?");
|
|
|
1160 ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1161
|
|
|
1162 runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1163 auto *ScevIV = SE.getSCEV(getInstructionByName(F, "iv")); // {0,+,1}
|
|
|
1164 auto *ScevI = SE.getSCEV(getArgByName(F, "i")); // {0,+,1}
|
|
|
1165
|
|
|
1166 ValueToSCEVMapTy RewriteMap;
|
|
|
1167 RewriteMap[cast<SCEVUnknown>(ScevI)->getValue()] =
|
|
|
1168 SE.getUMinExpr(ScevI, SE.getConstant(ScevI->getType(), 17));
|
|
|
1169 auto *WithUMin = SCEVParameterRewriter::rewrite(ScevIV, SE, RewriteMap);
|
|
|
1170
|
|
|
1171 EXPECT_NE(WithUMin, ScevIV);
|
|
|
1172 auto *AR = dyn_cast<SCEVAddRecExpr>(WithUMin);
|
|
|
1173 EXPECT_TRUE(AR);
|
|
|
1174 EXPECT_EQ(AR->getStart(),
|
|
|
1175 SE.getUMinExpr(ScevI, SE.getConstant(ScevI->getType(), 17)));
|
|
|
1176 EXPECT_EQ(AR->getStepRecurrence(SE),
|
|
|
1177 cast<SCEVAddRecExpr>(ScevIV)->getStepRecurrence(SE));
|
|
|
1178 });
|
|
|
1179 }
|
|
|
1180
|
|
|
1181 TEST_F(ScalarEvolutionsTest, SCEVAddNUW) {
|
|
|
1182 LLVMContext C;
|
|
|
1183 SMDiagnostic Err;
|
|
|
1184 std::unique_ptr<Module> M = parseAssemblyString("define void @foo(i32 %x) { "
|
|
|
1185 " ret void "
|
|
|
1186 "} ",
|
|
|
1187 Err, C);
|
|
|
1188
|
|
|
1189 ASSERT_TRUE(M && "Could not parse module?");
|
|
|
1190 ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1191
|
|
|
1192 runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1193 auto *X = SE.getSCEV(getArgByName(F, "x"));
|
|
|
1194 auto *One = SE.getOne(X->getType());
|
|
|
1195 auto *Sum = SE.getAddExpr(X, One, SCEV::FlagNUW);
|
|
|
1196 EXPECT_TRUE(SE.isKnownPredicate(ICmpInst::ICMP_UGE, Sum, X));
|
|
|
1197 EXPECT_TRUE(SE.isKnownPredicate(ICmpInst::ICMP_UGT, Sum, X));
|
|
|
1198 });
|
|
|
1199 }
|
|
|
1200
|
|
|
1201 TEST_F(ScalarEvolutionsTest, SCEVgetRanges) {
|
|
|
1202 LLVMContext C;
|
|
|
1203 SMDiagnostic Err;
|
|
|
1204 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
1205 "define void @foo(i32 %i) { "
|
|
|
1206 "entry: "
|
|
|
1207 " br label %loop.body "
|
|
|
1208 "loop.body: "
|
|
|
1209 " %iv = phi i32 [ %iv.next, %loop.body ], [ 0, %entry ] "
|
|
|
1210 " %iv.next = add nsw i32 %iv, 1 "
|
|
|
1211 " %cmp = icmp eq i32 %iv.next, 16 "
|
|
|
1212 " br i1 %cmp, label %exit, label %loop.body "
|
|
|
1213 "exit: "
|
|
|
1214 " ret void "
|
|
|
1215 "} ",
|
|
|
1216 Err, C);
|
|
|
1217
|
|
|
1218 runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1219 auto *ScevIV = SE.getSCEV(getInstructionByName(F, "iv")); // {0,+,1}
|
|
|
1220 auto *ScevI = SE.getSCEV(getArgByName(F, "i"));
|
|
|
1221 EXPECT_EQ(SE.getUnsignedRange(ScevIV).getLower(), 0);
|
|
|
1222 EXPECT_EQ(SE.getUnsignedRange(ScevIV).getUpper(), 16);
|
|
|
1223
|
|
|
1224 auto *Add = SE.getAddExpr(ScevI, ScevIV);
|
|
|
1225 ValueToSCEVMapTy RewriteMap;
|
|
|
1226 RewriteMap[cast<SCEVUnknown>(ScevI)->getValue()] =
|
|
|
1227 SE.getUMinExpr(ScevI, SE.getConstant(ScevI->getType(), 17));
|
|
|
1228 auto *AddWithUMin = SCEVParameterRewriter::rewrite(Add, SE, RewriteMap);
|
|
|
1229 EXPECT_EQ(SE.getUnsignedRange(AddWithUMin).getLower(), 0);
|
|
|
1230 EXPECT_EQ(SE.getUnsignedRange(AddWithUMin).getUpper(), 33);
|
|
|
1231 });
|
|
|
1232 }
|
|
|
1233
|
|
|
1234 TEST_F(ScalarEvolutionsTest, SCEVgetExitLimitForGuardedLoop) {
|
|
|
1235 LLVMContext C;
|
|
|
1236 SMDiagnostic Err;
|
|
|
1237 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
1238 "define void @foo(i32 %i) { "
|
|
|
1239 "entry: "
|
|
|
1240 " %cmp3 = icmp ult i32 %i, 16 "
|
|
|
1241 " br i1 %cmp3, label %loop.body, label %exit "
|
|
|
1242 "loop.body: "
|
|
|
1243 " %iv = phi i32 [ %iv.next, %loop.body ], [ %i, %entry ] "
|
|
|
1244 " %iv.next = add nsw i32 %iv, 1 "
|
|
|
1245 " %cmp = icmp eq i32 %iv.next, 16 "
|
|
|
1246 " br i1 %cmp, label %exit, label %loop.body "
|
|
|
1247 "exit: "
|
|
|
1248 " ret void "
|
|
|
1249 "} ",
|
|
|
1250 Err, C);
|
|
|
1251
|
|
|
1252 ASSERT_TRUE(M && "Could not parse module?");
|
|
|
1253 ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1254
|
|
|
1255 runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1256 auto *ScevIV = SE.getSCEV(getInstructionByName(F, "iv")); // {0,+,1}
|
|
|
1257 const Loop *L = cast<SCEVAddRecExpr>(ScevIV)->getLoop();
|
|
|
1258
|
|
|
1259 const SCEV *BTC = SE.getBackedgeTakenCount(L);
|
|
|
1260 EXPECT_FALSE(isa<SCEVConstant>(BTC));
|
|
|
1261 const SCEV *MaxBTC = SE.getConstantMaxBackedgeTakenCount(L);
|
|
|
1262 EXPECT_EQ(cast<SCEVConstant>(MaxBTC)->getAPInt(), 15);
|
|
|
1263 });
|
|
|
1264 }
|
|
|
1265
|
|
|
1266 TEST_F(ScalarEvolutionsTest, ImpliedViaAddRecStart) {
|
|
|
1267 LLVMContext C;
|
|
|
1268 SMDiagnostic Err;
|
|
|
1269 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
1270 "define void @foo(i32* %p) { "
|
|
|
1271 "entry: "
|
|
|
1272 " %x = load i32, i32* %p, !range !0 "
|
|
|
1273 " br label %loop "
|
|
|
1274 "loop: "
|
|
|
1275 " %iv = phi i32 [ %x, %entry], [%iv.next, %backedge] "
|
|
|
1276 " %ne.check = icmp ne i32 %iv, 0 "
|
|
|
1277 " br i1 %ne.check, label %backedge, label %exit "
|
|
|
1278 "backedge: "
|
|
|
1279 " %iv.next = add i32 %iv, -1 "
|
|
|
1280 " br label %loop "
|
|
|
1281 "exit:"
|
|
|
1282 " ret void "
|
|
|
1283 "} "
|
|
|
1284 "!0 = !{i32 0, i32 2147483647}",
|
|
|
1285 Err, C);
|
|
|
1286
|
|
|
1287 ASSERT_TRUE(M && "Could not parse module?");
|
|
|
1288 ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1289
|
|
|
1290 runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1291 auto *X = SE.getSCEV(getInstructionByName(F, "x"));
|
|
|
1292 auto *Context = getInstructionByName(F, "iv.next");
|
|
|
1293 EXPECT_TRUE(SE.isKnownPredicateAt(ICmpInst::ICMP_NE, X,
|
|
|
1294 SE.getZero(X->getType()), Context));
|
|
|
1295 });
|
|
|
1296 }
|
|
|
1297
|
|
|
1298 TEST_F(ScalarEvolutionsTest, UnsignedIsImpliedViaOperations) {
|
|
|
1299 LLVMContext C;
|
|
|
1300 SMDiagnostic Err;
|
|
|
1301 std::unique_ptr<Module> M =
|
|
|
1302 parseAssemblyString("define void @foo(i32* %p1, i32* %p2) { "
|
|
|
1303 "entry: "
|
|
|
1304 " %x = load i32, i32* %p1, !range !0 "
|
|
|
1305 " %cond = icmp ne i32 %x, 0 "
|
|
|
1306 " br i1 %cond, label %guarded, label %exit "
|
|
|
1307 "guarded: "
|
|
|
1308 " %y = add i32 %x, -1 "
|
|
|
1309 " ret void "
|
|
|
1310 "exit: "
|
|
|
1311 " ret void "
|
|
|
1312 "} "
|
|
|
1313 "!0 = !{i32 0, i32 2147483647}",
|
|
|
1314 Err, C);
|
|
|
1315
|
|
|
1316 ASSERT_TRUE(M && "Could not parse module?");
|
|
|
1317 ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1318
|
|
|
1319 runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1320 auto *X = SE.getSCEV(getInstructionByName(F, "x"));
|
|
|
1321 auto *Y = SE.getSCEV(getInstructionByName(F, "y"));
|
|
|
1322 auto *Guarded = getInstructionByName(F, "y")->getParent();
|
|
|
1323 ASSERT_TRUE(Guarded);
|
|
|
1324 EXPECT_TRUE(
|
|
|
1325 SE.isBasicBlockEntryGuardedByCond(Guarded, ICmpInst::ICMP_ULT, Y, X));
|
|
|
1326 EXPECT_TRUE(
|
|
|
1327 SE.isBasicBlockEntryGuardedByCond(Guarded, ICmpInst::ICMP_UGT, X, Y));
|
|
|
1328 });
|
|
|
1329 }
|
|
|
1330
|
|
|
1331 TEST_F(ScalarEvolutionsTest, ProveImplicationViaNarrowing) {
|
|
|
1332 LLVMContext C;
|
|
|
1333 SMDiagnostic Err;
|
|
|
1334 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
1335 "define i32 @foo(i32 %start, i32* %q) { "
|
|
|
1336 "entry: "
|
|
|
1337 " %wide.start = zext i32 %start to i64 "
|
|
|
1338 " br label %loop "
|
|
|
1339 "loop: "
|
|
|
1340 " %wide.iv = phi i64 [%wide.start, %entry], [%wide.iv.next, %backedge] "
|
|
|
1341 " %iv = phi i32 [%start, %entry], [%iv.next, %backedge] "
|
|
|
1342 " %cond = icmp eq i64 %wide.iv, 0 "
|
|
|
1343 " br i1 %cond, label %exit, label %backedge "
|
|
|
1344 "backedge: "
|
|
|
1345 " %iv.next = add i32 %iv, -1 "
|
|
|
1346 " %index = zext i32 %iv.next to i64 "
|
|
|
1347 " %load.addr = getelementptr i32, i32* %q, i64 %index "
|
|
|
1348 " %stop = load i32, i32* %load.addr "
|
|
|
1349 " %loop.cond = icmp eq i32 %stop, 0 "
|
|
|
1350 " %wide.iv.next = add nsw i64 %wide.iv, -1 "
|
|
|
1351 " br i1 %loop.cond, label %loop, label %failure "
|
|
|
1352 "exit: "
|
|
|
1353 " ret i32 0 "
|
|
|
1354 "failure: "
|
|
|
1355 " unreachable "
|
|
|
1356 "} ",
|
|
|
1357 Err, C);
|
|
|
1358
|
|
|
1359 ASSERT_TRUE(M && "Could not parse module?");
|
|
|
1360 ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1361
|
|
|
1362 runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1363 auto *IV = SE.getSCEV(getInstructionByName(F, "iv"));
|
|
|
1364 auto *Zero = SE.getZero(IV->getType());
|
|
|
1365 auto *Backedge = getInstructionByName(F, "iv.next")->getParent();
|
|
|
1366 ASSERT_TRUE(Backedge);
|
|
|
1367 (void)IV;
|
|
|
1368 (void)Zero;
|
|
|
1369 // FIXME: This can only be proved with turned on option
|
|
|
1370 // scalar-evolution-use-expensive-range-sharpening which is currently off.
|
|
|
1371 // Enable the check once it's switched true by default.
|
|
|
1372 // EXPECT_TRUE(SE.isBasicBlockEntryGuardedByCond(Backedge,
|
|
|
1373 // ICmpInst::ICMP_UGT,
|
|
|
1374 // IV, Zero));
|
|
|
1375 });
|
|
|
1376 }
|
|
|
1377
|
|
|
1378 TEST_F(ScalarEvolutionsTest, ImpliedCond) {
|
|
|
1379 LLVMContext C;
|
|
|
1380 SMDiagnostic Err;
|
|
|
1381 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
1382 "define void @foo(i32 %len) { "
|
|
|
1383 "entry: "
|
|
|
1384 " br label %loop "
|
|
|
1385 "loop: "
|
|
|
1386 " %iv = phi i32 [ 0, %entry], [%iv.next, %loop] "
|
|
|
1387 " %iv.next = add nsw i32 %iv, 1 "
|
|
|
1388 " %cmp = icmp slt i32 %iv, %len "
|
|
|
1389 " br i1 %cmp, label %loop, label %exit "
|
|
|
1390 "exit:"
|
|
|
1391 " ret void "
|
|
|
1392 "}",
|
|
|
1393 Err, C);
|
|
|
1394
|
|
|
1395 ASSERT_TRUE(M && "Could not parse module?");
|
|
|
1396 ASSERT_TRUE(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1397
|
|
|
1398 runWithSE(*M, "foo", [](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1399 Instruction *IV = getInstructionByName(F, "iv");
|
|
|
1400 Type *Ty = IV->getType();
|
|
|
1401 const SCEV *Zero = SE.getZero(Ty);
|
|
|
1402 const SCEV *MinusOne = SE.getMinusOne(Ty);
|
|
|
1403 // {0,+,1}<nuw><nsw>
|
|
|
1404 const SCEV *AddRec_0_1 = SE.getSCEV(IV);
|
|
|
1405 // {0,+,-1}<nw>
|
|
|
1406 const SCEV *AddRec_0_N1 = SE.getNegativeSCEV(AddRec_0_1);
|
|
|
1407
|
|
|
1408 // {0,+,1}<nuw><nsw> > 0 -> {0,+,-1}<nw> < 0
|
|
|
1409 EXPECT_TRUE(isImpliedCond(SE, ICmpInst::ICMP_SLT, AddRec_0_N1, Zero,
|
|
|
1410 ICmpInst::ICMP_SGT, AddRec_0_1, Zero));
|
|
|
1411 // {0,+,-1}<nw> < -1 -> {0,+,1}<nuw><nsw> > 0
|
|
|
1412 EXPECT_TRUE(isImpliedCond(SE, ICmpInst::ICMP_SGT, AddRec_0_1, Zero,
|
|
|
1413 ICmpInst::ICMP_SLT, AddRec_0_N1, MinusOne));
|
|
|
1414 });
|
|
|
1415 }
|
|
|
1416
|
|
|
1417 TEST_F(ScalarEvolutionsTest, MatchURem) {
|
|
|
1418 LLVMContext C;
|
|
|
1419 SMDiagnostic Err;
|
|
|
1420 std::unique_ptr<Module> M = parseAssemblyString(
|
|
|
1421 "target datalayout = \"e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128\" "
|
|
|
1422 " "
|
|
|
1423 "define void @test(i32 %a, i32 %b, i16 %c, i64 %d) {"
|
|
|
1424 "entry: "
|
|
|
1425 " %rem1 = urem i32 %a, 2"
|
|
|
1426 " %rem2 = urem i32 %a, 5"
|
|
|
1427 " %rem3 = urem i32 %a, %b"
|
|
|
1428 " %c.ext = zext i16 %c to i32"
|
|
|
1429 " %rem4 = urem i32 %c.ext, 2"
|
|
|
1430 " %ext = zext i32 %rem4 to i64"
|
|
|
1431 " %rem5 = urem i64 %d, 17179869184"
|
|
|
1432 " ret void "
|
|
|
1433 "} ",
|
|
|
1434 Err, C);
|
|
|
1435
|
|
|
1436 assert(M && "Could not parse module?");
|
|
|
1437 assert(!verifyModule(*M) && "Must have been well formed!");
|
|
|
1438
|
|
|
1439 runWithSE(*M, "test", [&](Function &F, LoopInfo &LI, ScalarEvolution &SE) {
|
|
|
1440 for (auto *N : {"rem1", "rem2", "rem3", "rem5"}) {
|
|
|
1441 auto *URemI = getInstructionByName(F, N);
|
|
|
1442 auto *S = SE.getSCEV(URemI);
|
|
|
1443 const SCEV *LHS, *RHS;
|
|
|
1444 EXPECT_TRUE(matchURem(SE, S, LHS, RHS));
|
|
|
1445 EXPECT_EQ(LHS, SE.getSCEV(URemI->getOperand(0)));
|
|
|
1446 EXPECT_EQ(RHS, SE.getSCEV(URemI->getOperand(1)));
|
|
|
1447 EXPECT_EQ(LHS->getType(), S->getType());
|
|
|
1448 EXPECT_EQ(RHS->getType(), S->getType());
|
|
|
1449 }
|
|
|
1450
|
|
|
1451 // Check the case where the urem operand is zero-extended. Make sure the
|
|
|
1452 // match results are extended to the size of the input expression.
|
|
|
1453 auto *Ext = getInstructionByName(F, "ext");
|
|
|
1454 auto *URem1 = getInstructionByName(F, "rem4");
|
|
|
1455 auto *S = SE.getSCEV(Ext);
|
|
|
1456 const SCEV *LHS, *RHS;
|
|
|
1457 EXPECT_TRUE(matchURem(SE, S, LHS, RHS));
|
|
|
1458 EXPECT_NE(LHS, SE.getSCEV(URem1->getOperand(0)));
|
|
|
1459 // RHS and URem1->getOperand(1) have different widths, so compare the
|
|
|
1460 // integer values.
|
|
|
1461 EXPECT_EQ(cast<SCEVConstant>(RHS)->getValue()->getZExtValue(),
|
|
|
1462 cast<SCEVConstant>(SE.getSCEV(URem1->getOperand(1)))
|
|
|
1463 ->getValue()
|
|
|
1464 ->getZExtValue());
|
|
|
1465 EXPECT_EQ(LHS->getType(), S->getType());
|
|
|
1466 EXPECT_EQ(RHS->getType(), S->getType());
|
|
|
1467 });
|
|
|
1468 }
|
|
|
1469
|
|
150
|
1470 } // end namespace llvm
|
