| 1 |
//===- DDGTest.cpp - DDGAnalysis 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/Analysis/DDG.h" |
| 10 |
#include "llvm/Analysis/AliasAnalysis.h" |
| 11 |
#include "llvm/Analysis/AssumptionCache.h" |
| 12 |
#include "llvm/Analysis/BasicAliasAnalysis.h" |
| 13 |
#include "llvm/Analysis/LoopInfo.h" |
| 14 |
#include "llvm/Analysis/ScalarEvolution.h" |
| 15 |
#include "llvm/Analysis/TargetLibraryInfo.h" |
| 16 |
#include "llvm/AsmParser/Parser.h" |
| 17 |
#include "llvm/IR/Dominators.h" |
| 18 |
#include "llvm/Support/SourceMgr.h" |
| 19 |
#include "gtest/gtest.h" |
| 20 |
|
| 21 |
using namespace llvm; |
| 22 |
|
| 23 |
/// Build the DDG analysis for a loop and run the given test \p Test. |
| 24 |
static void runTest(Module &M, StringRef FuncName, |
| 25 |
function_ref<void(Function &F, LoopInfo &LI, |
| 26 |
DependenceInfo &DI, ScalarEvolution &SE)> |
| 27 |
Test) { |
| 28 |
auto *F = M.getFunction(FuncName); |
| 29 |
ASSERT_NE(F, nullptr) << "Could not find " << FuncName; |
| 30 |
|
| 31 |
TargetLibraryInfoImpl TLII; |
| 32 |
TargetLibraryInfo TLI(TLII); |
| 33 |
AssumptionCache AC(*F); |
| 34 |
DominatorTree DT(*F); |
| 35 |
LoopInfo LI(DT); |
| 36 |
ScalarEvolution SE(*F, TLI, AC, DT, LI); |
| 37 |
AAResults AA(TLI); |
| 38 |
DependenceInfo DI(F, &AA, &SE, &LI); |
| 39 |
Test(*F, LI, DI, SE); |
| 40 |
} |
| 41 |
|
| 42 |
static std::unique_ptr<Module> makeLLVMModule(LLVMContext &Context, |
| 43 |
const char *ModuleStr) { |
| 44 |
SMDiagnostic Err; |
| 45 |
return parseAssemblyString(ModuleStr, Err, Context); |
| 46 |
} |
| 47 |
|
| 48 |
TEST(DDGTest, getDependencies) { |
| 49 |
const char *ModuleStr = |
| 50 |
"target datalayout = \"e-m:e-i64:64-n32:64\"\n" |
| 51 |
"target triple = \"powerpc64le-unknown-linux-gnu\"\n" |
| 52 |
"\n" |
| 53 |
"define dso_local void @foo(i32 signext %n, i32* noalias %A, i32* " |
| 54 |
"noalias %B) {\n" |
| 55 |
"entry:\n" |
| 56 |
" %cmp1 = icmp sgt i32 %n, 0\n" |
| 57 |
" br i1 %cmp1, label %for.body.preheader, label %for.end\n" |
| 58 |
"\n" |
| 59 |
"for.body.preheader:\n" |
| 60 |
" %wide.trip.count = zext i32 %n to i64\n" |
| 61 |
" br label %for.body\n" |
| 62 |
" \n" |
| 63 |
" for.body:\n" |
| 64 |
" %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ " |
| 65 |
"%indvars.iv.next, %for.body ]\n" |
| 66 |
" %arrayidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv\n" |
| 67 |
" %0 = trunc i64 %indvars.iv to i32\n" |
| 68 |
" store i32 %0, i32* %arrayidx, align 4\n" |
| 69 |
" %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1\n" |
| 70 |
" %arrayidx2 = getelementptr inbounds i32, i32* %A, i64 " |
| 71 |
"%indvars.iv.next\n" |
| 72 |
" %1 = load i32, i32* %arrayidx2, align 4\n" |
| 73 |
" %add3 = add nsw i32 %1, 1\n" |
| 74 |
" %arrayidx5 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv\n" |
| 75 |
" store i32 %add3, i32* %arrayidx5, align 4\n" |
| 76 |
" %exitcond = icmp ne i64 %indvars.iv.next, %wide.trip.count\n" |
| 77 |
" br i1 %exitcond, label %for.body, label %for.end.loopexit\n" |
| 78 |
"\n" |
| 79 |
"for.end.loopexit:\n" |
| 80 |
" br label %for.end\n" |
| 81 |
"\n" |
| 82 |
"for.end:\n" |
| 83 |
" ret void\n" |
| 84 |
"}\n"; |
| 85 |
|
| 86 |
LLVMContext Context; |
| 87 |
std::unique_ptr<Module> M = makeLLVMModule(Context, ModuleStr); |
| 88 |
|
| 89 |
runTest( |
| 90 |
*M, "foo", |
| 91 |
[&](Function &F, LoopInfo &LI, DependenceInfo &DI, ScalarEvolution &SE) { |
| 92 |
Loop *L = *LI.begin(); |
| 93 |
assert(L && "expected the loop to be identified."); |
| 94 |
|
| 95 |
DataDependenceGraph DDG(*L, LI, DI); |
| 96 |
|
| 97 |
// Collect all the nodes that have an outgoing memory edge |
| 98 |
// while collecting all memory edges as well. There should |
| 99 |
// only be one node with an outgoing memory edge and there |
| 100 |
// should only be one memory edge in the entire graph. |
| 101 |
std::vector<DDGNode *> DependenceSourceNodes; |
| 102 |
std::vector<DDGEdge *> MemoryEdges; |
| 103 |
for (DDGNode *N : DDG) { |
| 104 |
for (DDGEdge *E : *N) { |
| 105 |
bool SourceAdded = false; |
| 106 |
if (E->isMemoryDependence()) { |
| 107 |
MemoryEdges.push_back(E); |
| 108 |
if (!SourceAdded) { |
| 109 |
DependenceSourceNodes.push_back(N); |
| 110 |
SourceAdded = true; |
| 111 |
} |
| 112 |
} |
| 113 |
} |
| 114 |
} |
| 115 |
|
| 116 |
EXPECT_EQ(DependenceSourceNodes.size(), 1ull); |
| 117 |
EXPECT_EQ(MemoryEdges.size(), 1ull); |
| 118 |
|
| 119 |
DataDependenceGraph::DependenceList DL; |
| 120 |
DDG.getDependencies(*DependenceSourceNodes.back(), |
| 121 |
MemoryEdges.back()->getTargetNode(), DL); |
| 122 |
|
| 123 |
EXPECT_EQ(DL.size(), 1ull); |
| 124 |
EXPECT_TRUE(DL.back()->isAnti()); |
| 125 |
EXPECT_EQ(DL.back()->getLevels(), 1u); |
| 126 |
EXPECT_NE(DL.back()->getDistance(1), nullptr); |
| 127 |
EXPECT_EQ(DL.back()->getDistance(1), |
| 128 |
SE.getOne(DL.back()->getDistance(1)->getType())); |
| 129 |
}); |
| 130 |
} |
| 131 |
|
| 132 |
/// Test to make sure that when pi-blocks are formed, multiple edges of the same |
| 133 |
/// kind and direction are collapsed into a single edge. |
| 134 |
/// In the test below, %loadASubI belongs to an outside node, which has input |
| 135 |
/// dependency with multiple load instructions in the pi-block containing |
| 136 |
/// %loadBSubI. We expect a single memory dependence edge from the outside node |
| 137 |
/// to this pi-block. The pi-block also contains %add and %add7 both of which |
| 138 |
/// feed a phi in an outside node. We expect a single def-use edge from the |
| 139 |
/// pi-block to the node containing that phi. |
| 140 |
TEST(DDGTest, avoidDuplicateEdgesToFromPiBlocks) { |
| 141 |
const char *ModuleStr = |
| 142 |
"target datalayout = \"e-m:e-i64:64-n32:64-v256:256:256-v512:512:512\"\n" |
| 143 |
"\n" |
| 144 |
"define void @foo(float* noalias %A, float* noalias %B, float* noalias " |
| 145 |
"%C, float* noalias %D, i32 signext %n) {\n" |
| 146 |
"entry:\n" |
| 147 |
" %cmp1 = icmp sgt i32 %n, 0\n" |
| 148 |
" br i1 %cmp1, label %for.body.preheader, label %for.end\n" |
| 149 |
"\n" |
| 150 |
"for.body.preheader: ; preds = %entry\n" |
| 151 |
" %wide.trip.count = zext i32 %n to i64\n" |
| 152 |
" br label %for.body\n" |
| 153 |
"\n" |
| 154 |
"for.body: ; preds = " |
| 155 |
"%for.body.preheader, %if.end\n" |
| 156 |
" %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, " |
| 157 |
"%if.end ]\n" |
| 158 |
" %arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv\n" |
| 159 |
" %loadASubI = load float, float* %arrayidx, align 4\n" |
| 160 |
" %arrayidx2 = getelementptr inbounds float, float* %B, i64 " |
| 161 |
"%indvars.iv\n" |
| 162 |
" %loadBSubI = load float, float* %arrayidx2, align 4\n" |
| 163 |
" %add = fadd fast float %loadASubI, %loadBSubI\n" |
| 164 |
" %arrayidx4 = getelementptr inbounds float, float* %A, i64 " |
| 165 |
"%indvars.iv\n" |
| 166 |
" store float %add, float* %arrayidx4, align 4\n" |
| 167 |
" %arrayidx6 = getelementptr inbounds float, float* %A, i64 " |
| 168 |
"%indvars.iv\n" |
| 169 |
" %0 = load float, float* %arrayidx6, align 4\n" |
| 170 |
" %add7 = fadd fast float %0, 1.000000e+00\n" |
| 171 |
" %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1\n" |
| 172 |
" %arrayidx10 = getelementptr inbounds float, float* %B, i64 " |
| 173 |
"%indvars.iv.next\n" |
| 174 |
" store float %add7, float* %arrayidx10, align 4\n" |
| 175 |
" %arrayidx12 = getelementptr inbounds float, float* %A, i64 " |
| 176 |
"%indvars.iv\n" |
| 177 |
" %1 = load float, float* %arrayidx12, align 4\n" |
| 178 |
" %cmp13 = fcmp fast ogt float %1, 1.000000e+02\n" |
| 179 |
" br i1 %cmp13, label %if.then, label %if.else\n" |
| 180 |
"\n" |
| 181 |
"if.then: ; preds = %for.body\n" |
| 182 |
" br label %if.end\n" |
| 183 |
"\n" |
| 184 |
"if.else: ; preds = %for.body\n" |
| 185 |
" br label %if.end\n" |
| 186 |
"\n" |
| 187 |
"if.end: ; preds = %if.else, " |
| 188 |
"%if.then\n" |
| 189 |
" %ff.0 = phi float [ %add, %if.then ], [ %add7, %if.else ]\n" |
| 190 |
" store float %ff.0, float* %C, align 4\n" |
| 191 |
" %exitcond = icmp ne i64 %indvars.iv.next, %wide.trip.count\n" |
| 192 |
" br i1 %exitcond, label %for.body, label %for.end.loopexit\n" |
| 193 |
"\n" |
| 194 |
"for.end.loopexit: ; preds = %if.end\n" |
| 195 |
" br label %for.end\n" |
| 196 |
"\n" |
| 197 |
"for.end: ; preds = " |
| 198 |
"%for.end.loopexit, %entry\n" |
| 199 |
" ret void\n" |
| 200 |
"}\n"; |
| 201 |
|
| 202 |
LLVMContext Context; |
| 203 |
std::unique_ptr<Module> M = makeLLVMModule(Context, ModuleStr); |
| 204 |
|
| 205 |
runTest( |
| 206 |
*M, "foo", |
| 207 |
[&](Function &F, LoopInfo &LI, DependenceInfo &DI, ScalarEvolution &SE) { |
| 208 |
Loop *L = *LI.begin(); |
| 209 |
assert(L && "expected the loop to be identified."); |
| 210 |
|
| 211 |
DataDependenceGraph DDG(*L, LI, DI); |
| 212 |
|
| 213 |
const DDGNode *LoadASubI = nullptr; |
| 214 |
for (DDGNode *N : DDG) { |
| 215 |
if (!isa<SimpleDDGNode>(N)) |
| 216 |
continue; |
| 217 |
SmallVector<Instruction *, 8> IList; |
| 218 |
N->collectInstructions([](const Instruction *I) { return true; }, |
| 219 |
IList); |
| 220 |
if (llvm::any_of(IList, [](Instruction *I) { |
| 221 |
return I->getName() == "loadASubI"; |
| 222 |
})) { |
| 223 |
LoadASubI = N; |
| 224 |
break; |
| 225 |
} |
| 226 |
} |
| 227 |
assert(LoadASubI && "Did not find load of A[i]"); |
| 228 |
|
| 229 |
const PiBlockDDGNode *PiBlockWithBSubI = nullptr; |
| 230 |
for (DDGNode *N : DDG) { |
| 231 |
if (!isa<PiBlockDDGNode>(N)) |
| 232 |
continue; |
| 233 |
for (DDGNode *M : cast<PiBlockDDGNode>(N)->getNodes()) { |
| 234 |
SmallVector<Instruction *, 8> IList; |
| 235 |
M->collectInstructions([](const Instruction *I) { return true; }, |
| 236 |
IList); |
| 237 |
if (llvm::any_of(IList, [](Instruction *I) { |
| 238 |
return I->getName() == "loadBSubI"; |
| 239 |
})) { |
| 240 |
PiBlockWithBSubI = static_cast<PiBlockDDGNode *>(N); |
| 241 |
break; |
| 242 |
} |
| 243 |
} |
| 244 |
if (PiBlockWithBSubI) |
| 245 |
break; |
| 246 |
} |
| 247 |
assert(PiBlockWithBSubI && |
| 248 |
"Did not find pi-block containing load of B[i]"); |
| 249 |
|
| 250 |
const DDGNode *FFPhi = nullptr; |
| 251 |
for (DDGNode *N : DDG) { |
| 252 |
if (!isa<SimpleDDGNode>(N)) |
| 253 |
continue; |
| 254 |
SmallVector<Instruction *, 8> IList; |
| 255 |
N->collectInstructions([](const Instruction *I) { return true; }, |
| 256 |
IList); |
| 257 |
if (llvm::any_of(IList, [](Instruction *I) { |
| 258 |
return I->getName() == "ff.0"; |
| 259 |
})) { |
| 260 |
FFPhi = N; |
| 261 |
break; |
| 262 |
} |
| 263 |
} |
| 264 |
assert(FFPhi && "Did not find ff.0 phi instruction"); |
| 265 |
|
| 266 |
// Expect a single memory edge from '%0 = A[i]' to the pi-block. This |
| 267 |
// means the duplicate incoming memory edges are removed during pi-block |
| 268 |
// formation. |
| 269 |
SmallVector<DDGEdge *, 4> EL; |
| 270 |
LoadASubI->findEdgesTo(*PiBlockWithBSubI, EL); |
| 271 |
unsigned NumMemoryEdges = llvm::count_if( |
| 272 |
EL, [](DDGEdge *Edge) { return Edge->isMemoryDependence(); }); |
| 273 |
EXPECT_EQ(NumMemoryEdges, 1ull); |
| 274 |
|
| 275 |
/// Expect a single def-use edge from the pi-block to '%ff.0 = phi...`. |
| 276 |
/// This means the duplicate outgoing def-use edges are removed during |
| 277 |
/// pi-block formation. |
| 278 |
EL.clear(); |
| 279 |
PiBlockWithBSubI->findEdgesTo(*FFPhi, EL); |
| 280 |
NumMemoryEdges = |
| 281 |
llvm::count_if(EL, [](DDGEdge *Edge) { return Edge->isDefUse(); }); |
| 282 |
EXPECT_EQ(NumMemoryEdges, 1ull); |
| 283 |
}); |
| 284 |
} |