1 //===- LegalizerTest.cpp --------------------------------------------------===// 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/CodeGen/GlobalISel/Legalizer.h" 10 #include "GISelMITest.h" 11 #include "llvm/CodeGen/GlobalISel/LostDebugLocObserver.h" 12 13 #define DEBUG_TYPE "legalizer-test" 14 15 using namespace LegalizeActions; 16 using namespace LegalizeMutations; 17 using namespace LegalityPredicates; 18 19 namespace { 20 21 ::testing::AssertionResult isNullMIPtr(const MachineInstr *MI) { 22 if (MI == nullptr) 23 return ::testing::AssertionSuccess(); 24 std::string MIBuffer; 25 raw_string_ostream MISStream(MIBuffer); 26 MI->print(MISStream, /*IsStandalone=*/true, /*SkipOpers=*/false, 27 /*SkipDebugLoc=*/false, /*AddNewLine=*/false); 28 return ::testing::AssertionFailure() 29 << "unable to legalize instruction: " << MISStream.str(); 30 } 31 32 DefineLegalizerInfo(ALegalizer, { 33 auto p0 = LLT::pointer(0, 64); 34 auto s8 = LLT::scalar(8); 35 auto v2s8 = LLT::fixed_vector(2, 8); 36 auto v2s16 = LLT::fixed_vector(2, 16); 37 getActionDefinitionsBuilder(G_LOAD) 38 .legalForTypesWithMemDesc({{s16, p0, s8, 8}}) 39 .scalarize(0) 40 .clampScalar(0, s16, s16); 41 getActionDefinitionsBuilder(G_PTR_ADD).legalFor({{p0, s64}}); 42 getActionDefinitionsBuilder(G_CONSTANT).legalFor({s32, s64}); 43 getActionDefinitionsBuilder(G_BUILD_VECTOR) 44 .legalFor({{v2s16, s16}}) 45 .clampScalar(1, s16, s16); 46 getActionDefinitionsBuilder(G_BUILD_VECTOR_TRUNC).legalFor({{v2s8, s16}}); 47 getActionDefinitionsBuilder(G_ANYEXT).legalFor({{s32, s16}}); 48 getActionDefinitionsBuilder(G_ZEXT).legalFor({{s32, s16}}); 49 getActionDefinitionsBuilder(G_SEXT).legalFor({{s32, s16}}); 50 getActionDefinitionsBuilder(G_AND).legalFor({s32}); 51 getActionDefinitionsBuilder(G_SEXT_INREG).lower(); 52 getActionDefinitionsBuilder(G_ASHR).legalFor({{s32, s32}}); 53 getActionDefinitionsBuilder(G_SHL).legalFor({{s32, s32}}); 54 }) 55 56 TEST_F(AArch64GISelMITest, BasicLegalizerTest) { 57 StringRef MIRString = R"( 58 %vptr:_(p0) = COPY $x4 59 %v:_(<2 x s8>) = G_LOAD %vptr:_(p0) :: (load (<2 x s8>), align 1) 60 $h4 = COPY %v:_(<2 x s8>) 61 )"; 62 setUp(MIRString.rtrim(' ')); 63 if (!TM) 64 GTEST_SKIP(); 65 66 ALegalizerInfo LI(MF->getSubtarget()); 67 LostDebugLocObserver LocObserver(DEBUG_TYPE); 68 69 Legalizer::MFResult Result = Legalizer::legalizeMachineFunction( 70 *MF, LI, {&LocObserver}, LocObserver, B); 71 72 EXPECT_TRUE(isNullMIPtr(Result.FailedOn)); 73 EXPECT_TRUE(Result.Changed); 74 75 StringRef CheckString = R"( 76 CHECK: %vptr:_(p0) = COPY $x4 77 CHECK-NEXT: [[LOAD_0:%[0-9]+]]:_(s16) = G_LOAD %vptr:_(p0) :: (load (s8)) 78 CHECK-NEXT: [[OFFSET_1:%[0-9]+]]:_(s64) = G_CONSTANT i64 1 79 CHECK-NEXT: [[VPTR_1:%[0-9]+]]:_(p0) = G_PTR_ADD %vptr:_, [[OFFSET_1]]:_(s64) 80 CHECK-NEXT: [[LOAD_1:%[0-9]+]]:_(s16) = G_LOAD [[VPTR_1]]:_(p0) :: (load (s8) from unknown-address + 1) 81 CHECK-NEXT: %v:_(<2 x s8>) = G_BUILD_VECTOR_TRUNC [[LOAD_0]]:_(s16), [[LOAD_1]]:_(s16) 82 CHECK-NEXT: $h4 = COPY %v:_(<2 x s8>) 83 )"; 84 85 EXPECT_TRUE(CheckMachineFunction(*MF, CheckString)) << *MF; 86 } 87 88 // Making sure the legalization finishes successfully w/o failure to combine 89 // away all the legalization artifacts regardless of the order of their 90 // creation. 91 TEST_F(AArch64GISelMITest, UnorderedArtifactCombiningTest) { 92 StringRef MIRString = R"( 93 %vptr:_(p0) = COPY $x4 94 %v:_(<2 x s8>) = G_LOAD %vptr:_(p0) :: (load (<2 x s8>), align 1) 95 %v0:_(s8), %v1:_(s8) = G_UNMERGE_VALUES %v:_(<2 x s8>) 96 %v0_ext:_(s16) = G_ANYEXT %v0:_(s8) 97 $h4 = COPY %v0_ext:_(s16) 98 )"; 99 setUp(MIRString.rtrim(' ')); 100 if (!TM) 101 GTEST_SKIP(); 102 103 ALegalizerInfo LI(MF->getSubtarget()); 104 LostDebugLocObserver LocObserver(DEBUG_TYPE); 105 106 // The events here unfold as follows: 107 // 1. First, the function is scanned pre-forming the worklist of artifacts: 108 // 109 // UNMERGE (1): pushed into the worklist first, will be processed last. 110 // | 111 // ANYEXT (2) 112 // 113 // 2. Second, the load is scalarized, and then its destination is widened, 114 // forming the following chain of legalization artifacts: 115 // 116 // TRUNC (4): created last, will be processed first. 117 // | 118 // BUILD_VECTOR (3) 119 // | 120 // UNMERGE (1): pushed into the worklist first, will be processed last. 121 // | 122 // ANYEXT (2) 123 // 124 // 3. Third, the artifacts are attempted to be combined in pairs, looking 125 // through the def-use chain from the roots towards the leafs, visiting the 126 // roots in order they happen to be in the worklist: 127 // (4) - (trunc): can not be combined; 128 // (3) - (build_vector (trunc)): can not be combined; 129 // (2) - (anyext (unmerge)): can not be combined; 130 // (1) - (unmerge (build_vector)): combined and eliminated; 131 // 132 // leaving the function in the following state: 133 // 134 // TRUNC (1): moved to non-artifact instructions worklist first. 135 // | 136 // ANYEXT (2): also moved to non-artifact instructions worklist. 137 // 138 // Every other instruction is successfully legalized in full. 139 // If combining (unmerge (build_vector)) does not re-insert every artifact 140 // that had its def-use chain modified (shortened) into the artifact 141 // worklist (here it's just ANYEXT), the process moves on onto the next 142 // outer loop iteration of the top-level legalization algorithm here, w/o 143 // performing all the artifact combines possible. Let's consider this 144 // scenario first: 145 // 4.A. Neither TRUNC, nor ANYEXT can be legalized in isolation, both of them 146 // get moved to the retry worklist, but no additional artifacts were 147 // created in the process, thus algorithm concludes no progress could be 148 // made, and fails. 149 // 4.B. If, however, combining (unmerge (build_vector)) had re-inserted 150 // ANYEXT into the worklist (as ANYEXT's source changes, not by value, 151 // but by implementation), (anyext (trunc)) combine happens next, which 152 // fully eliminates all the artifacts and legalization succeeds. 153 // 154 // We're looking into making sure that (4.B) happens here, not (4.A). Note 155 // that in that case the first scan through the artifacts worklist, while not 156 // being done in any guaranteed order, only needs to find the innermost 157 // pair(s) of artifacts that could be immediately combined out. After that 158 // the process follows def-use chains, making them shorter at each step, thus 159 // combining everything that can be combined in O(n) time. 160 Legalizer::MFResult Result = Legalizer::legalizeMachineFunction( 161 *MF, LI, {&LocObserver}, LocObserver, B); 162 163 EXPECT_TRUE(isNullMIPtr(Result.FailedOn)); 164 EXPECT_TRUE(Result.Changed); 165 166 StringRef CheckString = R"( 167 CHECK: %vptr:_(p0) = COPY $x4 168 CHECK-NEXT: [[LOAD_0:%[0-9]+]]:_(s16) = G_LOAD %vptr:_(p0) :: (load (s8)) 169 CHECK: $h4 = COPY [[LOAD_0]]:_(s16) 170 )"; 171 172 EXPECT_TRUE(CheckMachineFunction(*MF, CheckString)) << *MF; 173 } 174 175 TEST_F(AArch64GISelMITest, UnorderedArtifactCombiningManyCopiesTest) { 176 StringRef MIRString = R"( 177 %vptr:_(p0) = COPY $x4 178 %v:_(<2 x s8>) = G_LOAD %vptr:_(p0) :: (load (<2 x s8>), align 1) 179 %vc0:_(<2 x s8>) = COPY %v:_(<2 x s8>) 180 %vc1:_(<2 x s8>) = COPY %v:_(<2 x s8>) 181 %vc00:_(s8), %vc01:_(s8) = G_UNMERGE_VALUES %vc0:_(<2 x s8>) 182 %vc10:_(s8), %vc11:_(s8) = G_UNMERGE_VALUES %vc1:_(<2 x s8>) 183 %v0t:_(s8) = COPY %vc00:_(s8) 184 %v0:_(s8) = COPY %v0t:_(s8) 185 %v1t:_(s8) = COPY %vc11:_(s8) 186 %v1:_(s8) = COPY %v1t:_(s8) 187 %v0_zext:_(s32) = G_ZEXT %v0:_(s8) 188 %v1_sext:_(s32) = G_SEXT %v1:_(s8) 189 $w4 = COPY %v0_zext:_(s32) 190 $w5 = COPY %v1_sext:_(s32) 191 )"; 192 setUp(MIRString.rtrim(' ')); 193 if (!TM) 194 GTEST_SKIP(); 195 196 ALegalizerInfo LI(MF->getSubtarget()); 197 LostDebugLocObserver LocObserver(DEBUG_TYPE); 198 199 Legalizer::MFResult Result = Legalizer::legalizeMachineFunction( 200 *MF, LI, {&LocObserver}, LocObserver, B); 201 202 EXPECT_TRUE(isNullMIPtr(Result.FailedOn)); 203 EXPECT_TRUE(Result.Changed); 204 205 StringRef CheckString = R"( 206 CHECK: %vptr:_(p0) = COPY $x4 207 CHECK-NEXT: [[LOAD_0:%[0-9]+]]:_(s16) = G_LOAD %vptr:_(p0) :: (load (s8)) 208 CHECK-NEXT: [[OFFSET_1:%[0-9]+]]:_(s64) = G_CONSTANT i64 1 209 CHECK-NEXT: [[VPTR_1:%[0-9]+]]:_(p0) = G_PTR_ADD %vptr:_, [[OFFSET_1]]:_(s64) 210 CHECK-NEXT: [[LOAD_1:%[0-9]+]]:_(s16) = G_LOAD [[VPTR_1]]:_(p0) :: (load (s8) from unknown-address + 1) 211 CHECK-NEXT: [[FF_MASK:%[0-9]+]]:_(s32) = G_CONSTANT i32 255 212 CHECK-NEXT: [[V0_EXT:%[0-9]+]]:_(s32) = G_ANYEXT [[LOAD_0]]:_(s16) 213 CHECK-NEXT: %v0_zext:_(s32) = G_AND [[V0_EXT]]:_, [[FF_MASK]]:_ 214 CHECK-NEXT: [[V1_EXT:%[0-9]+]]:_(s32) = G_ANYEXT [[LOAD_1]]:_(s16) 215 CHECK-NEXT: [[SHAMNT:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 216 CHECK-NEXT: [[V1_SHL:%[0-9]+]]:_(s32) = G_SHL [[V1_EXT]]:_, [[SHAMNT]]:_(s32) 217 CHECK-NEXT: %v1_sext:_(s32) = G_ASHR [[V1_SHL]]:_, [[SHAMNT]]:_(s32) 218 CHECK-NEXT: $w4 = COPY %v0_zext:_(s32) 219 CHECK-NEXT: $w5 = COPY %v1_sext:_(s32) 220 )"; 221 222 EXPECT_TRUE(CheckMachineFunction(*MF, CheckString)) << *MF; 223 } 224 225 } // namespace 226