| 1 |
//===- MLRegAllocDevelopmentFeatures.cpp - test dev MLRegalloc features ---===// |
| 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 "../../lib/CodeGen/MLRegallocEvictAdvisor.h" |
| 10 |
#include "llvm/ADT/Triple.h" |
| 11 |
#include "llvm/Analysis/NoInferenceModelRunner.h" |
| 12 |
#include "llvm/CodeGen/MachineBasicBlock.h" |
| 13 |
#include "llvm/CodeGen/MachineFunction.h" |
| 14 |
#include "llvm/CodeGen/MachineModuleInfo.h" |
| 15 |
#include "llvm/CodeGen/SlotIndexes.h" |
| 16 |
#include "llvm/CodeGen/TargetFrameLowering.h" |
| 17 |
#include "llvm/CodeGen/TargetInstrInfo.h" |
| 18 |
#include "llvm/CodeGen/TargetLowering.h" |
| 19 |
#include "llvm/IR/LLVMContext.h" |
| 20 |
#include "llvm/MC/TargetRegistry.h" |
| 21 |
#include "llvm/Support/Allocator.h" |
| 22 |
#include "llvm/Support/CodeGen.h" |
| 23 |
#include "llvm/Support/TargetSelect.h" |
| 24 |
#include "llvm/Target/TargetMachine.h" |
| 25 |
#include "llvm/Target/TargetOptions.h" |
| 26 |
#include "gmock/gmock.h" |
| 27 |
#include "gtest/gtest.h" |
| 28 |
|
| 29 |
#include <string> |
| 30 |
#include <vector> |
| 31 |
|
| 32 |
using testing::ContainerEq; |
| 33 |
using testing::Test; |
| 34 |
|
| 35 |
namespace { |
| 36 |
|
| 37 |
#include "MFCommon.inc" |
| 38 |
|
| 39 |
struct LRPosInfoIndexes { |
| 40 |
size_t StartIndex; |
| 41 |
size_t EndIndex; |
| 42 |
size_t PhysReg; |
| 43 |
}; |
| 44 |
|
| 45 |
class RegallocDevelopmentFeaturesTest : public ::Test { |
| 46 |
protected: |
| 47 |
SmallVector<LRStartEndInfo> |
| 48 |
setupOverlapProblem(const SmallVectorImpl<LRPosInfoIndexes> &Segments, |
| 49 |
ilist<IndexListEntry> &IndexList) { |
| 50 |
SmallVector<LRStartEndInfo> PositionsToReturn; |
| 51 |
PositionsToReturn.reserve(Segments.size()); |
| 52 |
for (auto CurrentPosIndexInfo : Segments) { |
| 53 |
LRStartEndInfo CurrentPosInfo = {}; |
| 54 |
CurrentPosInfo.Pos = CurrentPosIndexInfo.PhysReg; |
| 55 |
PositionsToReturn.push_back(CurrentPosInfo); |
| 56 |
} |
| 57 |
size_t CurrentSegmentIndex = 0; |
| 58 |
size_t CurrentIndex = 0; |
| 59 |
while (CurrentSegmentIndex < Segments.size()) { |
| 60 |
auto *CurrentLEMem = static_cast<IndexListEntry *>( |
| 61 |
Allocator.Allocate(sizeof(IndexListEntry), alignof(IndexListEntry))); |
| 62 |
auto *CurrentListEntry = |
| 63 |
new (CurrentLEMem) IndexListEntry(nullptr, CurrentIndex); |
| 64 |
IndexList.push_back(CurrentListEntry); |
| 65 |
for (size_t CurrentPosInfoIndex = 0; |
| 66 |
CurrentPosInfoIndex < Segments.size(); ++CurrentPosInfoIndex) { |
| 67 |
if ((CurrentIndex / SlotIndex::InstrDist) == |
| 68 |
Segments[CurrentPosInfoIndex].StartIndex) { |
| 69 |
PositionsToReturn[CurrentPosInfoIndex].Begin = |
| 70 |
SlotIndex(CurrentListEntry, 0); |
| 71 |
} else if ((CurrentIndex / SlotIndex::InstrDist) == |
| 72 |
Segments[CurrentPosInfoIndex].EndIndex) { |
| 73 |
PositionsToReturn[CurrentPosInfoIndex].End = |
| 74 |
SlotIndex(CurrentListEntry, 0); |
| 75 |
++CurrentSegmentIndex; |
| 76 |
} |
| 77 |
} |
| 78 |
CurrentIndex += SlotIndex::InstrDist; |
| 79 |
} |
| 80 |
return PositionsToReturn; |
| 81 |
} |
| 82 |
|
| 83 |
NoInferenceModelRunner setupModelRunner() { |
| 84 |
const std::vector<TensorSpec> Inputs{ |
| 85 |
TensorSpec::createSpec<int64_t>("instructions", InstructionsShape), |
| 86 |
TensorSpec::createSpec<int64_t>("instructions_mapping", |
| 87 |
InstructionsMappingShape), |
| 88 |
TensorSpec::createSpec<float>("mbb_frequencies", MBBFrequencyShape), |
| 89 |
TensorSpec::createSpec<int64_t>("mbb_mapping", InstructionsShape)}; |
| 90 |
LLVMContext Ctx; |
| 91 |
return NoInferenceModelRunner(Ctx, Inputs); |
| 92 |
} |
| 93 |
|
| 94 |
std::vector<int64_t> |
| 95 |
getExpectedMappingMatrix(SmallVectorImpl<LRPosInfoIndexes> &OverlapSetup) { |
| 96 |
std::vector<int64_t> ExpectedMappingMatrix( |
| 97 |
NumberOfInterferences * ModelMaxSupportedInstructionCount, 0); |
| 98 |
for (auto NewSegment : OverlapSetup) { |
| 99 |
for (size_t CurrentIndex = NewSegment.StartIndex; |
| 100 |
CurrentIndex <= NewSegment.EndIndex; ++CurrentIndex) { |
| 101 |
ExpectedMappingMatrix[NewSegment.PhysReg * |
| 102 |
ModelMaxSupportedInstructionCount + |
| 103 |
CurrentIndex] = 1; |
| 104 |
} |
| 105 |
} |
| 106 |
return ExpectedMappingMatrix; |
| 107 |
} |
| 108 |
|
| 109 |
void runOverlapTest(SmallVectorImpl<LRPosInfoIndexes> &OverlapSetup) { |
| 110 |
ilist<IndexListEntry> IndexList; |
| 111 |
auto OverlapProblem = setupOverlapProblem(OverlapSetup, IndexList); |
| 112 |
NoInferenceModelRunner ModelRunner = setupModelRunner(); |
| 113 |
size_t MaxIndex = 0; |
| 114 |
for (size_t CurrentOverlap = 0; CurrentOverlap < OverlapSetup.size(); |
| 115 |
++CurrentOverlap) { |
| 116 |
if (OverlapSetup[CurrentOverlap].EndIndex > |
| 117 |
OverlapSetup[MaxIndex].EndIndex) { |
| 118 |
MaxIndex = CurrentOverlap; |
| 119 |
} |
| 120 |
} |
| 121 |
SlotIndex LastIndex = OverlapProblem[MaxIndex].End; |
| 122 |
extractInstructionFeatures( |
| 123 |
OverlapProblem, &ModelRunner, |
| 124 |
[](SlotIndex InputSlot) -> int { return 0; }, |
| 125 |
[](SlotIndex InputSlot) -> float { return 0.0f; }, |
| 126 |
[](SlotIndex InputSlot) -> MachineBasicBlock * { return nullptr; }, 0, |
| 127 |
1, 2, 3, LastIndex); |
| 128 |
std::vector<int64_t> MappingMatrix( |
| 129 |
ModelRunner.getTensor<int64_t>(1), |
| 130 |
ModelRunner.getTensor<int64_t>(1) + |
| 131 |
NumberOfInterferences * ModelMaxSupportedInstructionCount); |
| 132 |
ASSERT_THAT(MappingMatrix, |
| 133 |
ContainerEq(getExpectedMappingMatrix(OverlapSetup))); |
| 134 |
IndexList.clearAndLeakNodesUnsafely(); |
| 135 |
} |
| 136 |
|
| 137 |
BumpPtrAllocator Allocator; |
| 138 |
}; |
| 139 |
|
| 140 |
// meta tests to ensure that test setup works correctly |
| 141 |
|
| 142 |
TEST_F(RegallocDevelopmentFeaturesTest, |
| 143 |
MetaOverlapInstructionDistancesAreCorrect) { |
| 144 |
SmallVector<LRPosInfoIndexes, 2> OverlapSetup; |
| 145 |
OverlapSetup.push_back({0, 5, 0}); |
| 146 |
OverlapSetup.push_back({5, 10, 0}); |
| 147 |
ilist<IndexListEntry> IndexList; |
| 148 |
auto OverlapProblem = setupOverlapProblem(OverlapSetup, IndexList); |
| 149 |
ASSERT_EQ(OverlapProblem[0].End.distance(OverlapProblem[1].End), |
| 150 |
5 * SlotIndex::InstrDist); |
| 151 |
ASSERT_EQ(OverlapProblem[0].End.distance(OverlapProblem[1].Begin), 0); |
| 152 |
} |
| 153 |
|
| 154 |
TEST_F(RegallocDevelopmentFeaturesTest, MetaSlotIndicesAreValid) { |
| 155 |
SmallVector<LRPosInfoIndexes, 1> OverlapSetup; |
| 156 |
OverlapSetup.push_back({0, 10, 0}); |
| 157 |
ilist<IndexListEntry> IndexList; |
| 158 |
auto OverlapProblem = setupOverlapProblem(OverlapSetup, IndexList); |
| 159 |
ASSERT_TRUE(OverlapProblem[0].Begin.isValid()); |
| 160 |
ASSERT_TRUE(OverlapProblem[0].End.isValid()); |
| 161 |
} |
| 162 |
|
| 163 |
// Testing of feature extraction for per-instruction features |
| 164 |
|
| 165 |
TEST_F(RegallocDevelopmentFeaturesTest, InstructionOpcodesAreCorrect) { |
| 166 |
SmallVector<LRPosInfoIndexes, 1> OverlapSetup; |
| 167 |
OverlapSetup.push_back({0, ModelMaxSupportedInstructionCount - 1, 0}); |
| 168 |
ilist<IndexListEntry> IndexList; |
| 169 |
auto OverlapProblem = setupOverlapProblem(OverlapSetup, IndexList); |
| 170 |
NoInferenceModelRunner ModelRunner = setupModelRunner(); |
| 171 |
SlotIndex LastIndex = OverlapProblem[0].End; |
| 172 |
SlotIndex FirstIndex = OverlapProblem[0].Begin; |
| 173 |
extractInstructionFeatures( |
| 174 |
OverlapProblem, &ModelRunner, |
| 175 |
[FirstIndex](SlotIndex InputSlot) -> int { |
| 176 |
return FirstIndex.distance(InputSlot) / SlotIndex::InstrDist; |
| 177 |
}, |
| 178 |
[](SlotIndex InputSlot) -> float { return 0.0f; }, |
| 179 |
[](SlotIndex InputSlot) -> MachineBasicBlock * { return nullptr; }, 0, 1, |
| 180 |
2, 3, LastIndex); |
| 181 |
for (size_t CurrentInstructionIndex = 0; |
| 182 |
CurrentInstructionIndex < ModelMaxSupportedInstructionCount; |
| 183 |
++CurrentInstructionIndex) { |
| 184 |
ASSERT_EQ( |
| 185 |
(size_t)ModelRunner.getTensor<int64_t>(0)[CurrentInstructionIndex], |
| 186 |
CurrentInstructionIndex); |
| 187 |
} |
| 188 |
} |
| 189 |
|
| 190 |
TEST_F(RegallocDevelopmentFeaturesTest, FullOverlap) { |
| 191 |
SmallVector<LRPosInfoIndexes, 2> OverlapSetup; |
| 192 |
OverlapSetup.push_back({0, ModelMaxSupportedInstructionCount - 1, 0}); |
| 193 |
OverlapSetup.push_back({0, ModelMaxSupportedInstructionCount - 1, 1}); |
| 194 |
runOverlapTest(OverlapSetup); |
| 195 |
} |
| 196 |
|
| 197 |
TEST_F(RegallocDevelopmentFeaturesTest, PartialOverlap) { |
| 198 |
SmallVector<LRPosInfoIndexes, 2> OverlapSetup; |
| 199 |
OverlapSetup.push_back({0, 20, 0}); |
| 200 |
OverlapSetup.push_back({15, 30, 1}); |
| 201 |
runOverlapTest(OverlapSetup); |
| 202 |
} |
| 203 |
|
| 204 |
TEST_F(RegallocDevelopmentFeaturesTest, PartialOverlapOpposite) { |
| 205 |
SmallVector<LRPosInfoIndexes, 2> OverlapSetup; |
| 206 |
OverlapSetup.push_back({15, 30, 1}); |
| 207 |
OverlapSetup.push_back({0, 20, 0}); |
| 208 |
runOverlapTest(OverlapSetup); |
| 209 |
} |
| 210 |
|
| 211 |
TEST_F(RegallocDevelopmentFeaturesTest, InternalOverlap) { |
| 212 |
SmallVector<LRPosInfoIndexes, 2> OverlapSetup; |
| 213 |
OverlapSetup.push_back({0, 30, 0}); |
| 214 |
OverlapSetup.push_back({10, 20, 1}); |
| 215 |
runOverlapTest(OverlapSetup); |
| 216 |
} |
| 217 |
|
| 218 |
TEST_F(RegallocDevelopmentFeaturesTest, TripleInternalOverlap) { |
| 219 |
SmallVector<LRPosInfoIndexes, 3> OverlapSetup; |
| 220 |
OverlapSetup.push_back({0, 30, 0}); |
| 221 |
OverlapSetup.push_back({10, 25, 1}); |
| 222 |
OverlapSetup.push_back({15, 20, 2}); |
| 223 |
runOverlapTest(OverlapSetup); |
| 224 |
} |
| 225 |
|
| 226 |
TEST_F(RegallocDevelopmentFeaturesTest, InternalMultiOverlap) { |
| 227 |
SmallVector<LRPosInfoIndexes, 3> OverlapSetup; |
| 228 |
OverlapSetup.push_back({0, 45, 0}); |
| 229 |
OverlapSetup.push_back({30, 40, 1}); |
| 230 |
OverlapSetup.push_back({35, 60, 2}); |
| 231 |
runOverlapTest(OverlapSetup); |
| 232 |
} |
| 233 |
|
| 234 |
TEST_F(RegallocDevelopmentFeaturesTest, SingleMBBTest) { |
| 235 |
NoInferenceModelRunner ModelRunner = setupModelRunner(); |
| 236 |
SlotIndex CurrentIndex; |
| 237 |
// set index to 1 so we can ensure that the mapping actually get set |
| 238 |
std::map<MachineBasicBlock *, size_t> VisitedMBBs = {{nullptr, 1}}; |
| 239 |
extractMBBFrequency( |
| 240 |
CurrentIndex, 0, VisitedMBBs, |
| 241 |
[](SlotIndex InputSlot) -> float { return 1.0f; }, nullptr, &ModelRunner, |
| 242 |
2, 3); |
| 243 |
ASSERT_FLOAT_EQ(ModelRunner.getTensor<float>(2)[1], 1.0f); |
| 244 |
ASSERT_EQ(ModelRunner.getTensor<int64_t>(3)[0], 1); |
| 245 |
} |
| 246 |
|
| 247 |
TEST_F(RegallocDevelopmentFeaturesTest, MBBFullTruncated) { |
| 248 |
SmallVector<LRPosInfoIndexes, 1> OverlapSetup; |
| 249 |
OverlapSetup.push_back({0, ModelMaxSupportedInstructionCount - 1, 0}); |
| 250 |
ilist<IndexListEntry> IndexList; |
| 251 |
auto OverlapProblem = setupOverlapProblem(OverlapSetup, IndexList); |
| 252 |
NoInferenceModelRunner ModelRunner = setupModelRunner(); |
| 253 |
SlotIndex LastIndex = OverlapProblem[0].End; |
| 254 |
SlotIndex FirstIndex = OverlapProblem[0].Begin; |
| 255 |
|
| 256 |
LLVMContext Ctx; |
| 257 |
Module Mod("Module", Ctx); |
| 258 |
auto MF = createMachineFunction(Ctx, Mod); |
| 259 |
std::array<MachineBasicBlock *, ModelMaxSupportedInstructionCount> |
| 260 |
MBBsForTest; |
| 261 |
for (size_t I = 0; I < ModelMaxSupportedInstructionCount; ++I) { |
| 262 |
MBBsForTest[I] = MF->CreateMachineBasicBlock(); |
| 263 |
} |
| 264 |
|
| 265 |
extractInstructionFeatures( |
| 266 |
OverlapProblem, &ModelRunner, |
| 267 |
[](SlotIndex InputSlot) -> int { return 0; }, |
| 268 |
[FirstIndex](SlotIndex InputSlot) -> float { |
| 269 |
return static_cast<float>(FirstIndex.distance(InputSlot) / |
| 270 |
SlotIndex::InstrDist); |
| 271 |
}, |
| 272 |
[FirstIndex, MBBsForTest](SlotIndex InputSlot) -> MachineBasicBlock * { |
| 273 |
return MBBsForTest[FirstIndex.distance(InputSlot) / |
| 274 |
SlotIndex::InstrDist]; |
| 275 |
}, |
| 276 |
0, 1, 2, 3, LastIndex); |
| 277 |
for (size_t MBBIndex = 0; MBBIndex < ModelMaxSupportedMBBCount; ++MBBIndex) { |
| 278 |
ASSERT_FLOAT_EQ(ModelRunner.getTensor<float>(2)[MBBIndex], |
| 279 |
static_cast<float>(MBBIndex)); |
| 280 |
ASSERT_EQ(ModelRunner.getTensor<int64_t>(3)[MBBIndex], |
| 281 |
static_cast<int64_t>(MBBIndex)); |
| 282 |
} |
| 283 |
// the rest of the mapping values should be zero (truncated to 100 MBBs) |
| 284 |
for (size_t MBBIndex = ModelMaxSupportedMBBCount; |
| 285 |
MBBIndex < ModelMaxSupportedInstructionCount; ++MBBIndex) { |
| 286 |
ASSERT_EQ(ModelRunner.getTensor<int64_t>(3)[MBBIndex], |
| 287 |
static_cast<int64_t>(0)); |
| 288 |
} |
| 289 |
} |
| 290 |
|
| 291 |
} // end namespace |