1; RUN: opt < %s -sroa -S | FileCheck %s 2; RUN: opt -debugify -sroa -S < %s | FileCheck %s -check-prefix DEBUGLOC 3 4target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-n8:16:32:64" 5 6declare void @llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i1) 7 8define void @test1({ i8, i8 }* %a, { i8, i8 }* %b) { 9; CHECK-LABEL: @test1( 10; CHECK: %[[gep_a0:.*]] = getelementptr inbounds { i8, i8 }, { i8, i8 }* %a, i64 0, i32 0 11; CHECK: %[[a0:.*]] = load i8, i8* %[[gep_a0]], align 16 12; CHECK: %[[gep_a1:.*]] = getelementptr inbounds { i8, i8 }, { i8, i8 }* %a, i64 0, i32 1 13; CHECK: %[[a1:.*]] = load i8, i8* %[[gep_a1]], align 1 14; CHECK: %[[gep_b0:.*]] = getelementptr inbounds { i8, i8 }, { i8, i8 }* %b, i64 0, i32 0 15; CHECK: store i8 %[[a0]], i8* %[[gep_b0]], align 16 16; CHECK: %[[gep_b1:.*]] = getelementptr inbounds { i8, i8 }, { i8, i8 }* %b, i64 0, i32 1 17; CHECK: store i8 %[[a1]], i8* %[[gep_b1]], align 1 18; CHECK: ret void 19 20entry: 21 %alloca = alloca { i8, i8 }, align 16 22 %gep_a = getelementptr { i8, i8 }, { i8, i8 }* %a, i32 0, i32 0 23 %gep_alloca = getelementptr { i8, i8 }, { i8, i8 }* %alloca, i32 0, i32 0 24 %gep_b = getelementptr { i8, i8 }, { i8, i8 }* %b, i32 0, i32 0 25 26 store i8 420, i8* %gep_alloca, align 16 27 28 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %gep_alloca, i8* align 16 %gep_a, i32 2, i1 false) 29 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 16 %gep_b, i8* align 16 %gep_alloca, i32 2, i1 false) 30 ret void 31} 32 33define void @test2() { 34; CHECK-LABEL: @test2( 35; CHECK: alloca i16 36; CHECK: load i8, i8* %{{.*}} 37; CHECK: store i8 42, i8* %{{.*}} 38; CHECK: ret void 39 40; Check that when sroa rewrites the alloca partition 41; it preserves the original DebugLocation. 42; DEBUGLOC-LABEL: @test2( 43; DEBUGLOC: {{.*}} = alloca {{.*}} !dbg ![[DbgLoc:[0-9]+]] 44; DEBUGLOC-LABEL: } 45; 46; DEBUGLOC: ![[DbgLoc]] = !DILocation(line: 9, 47 48entry: 49 %a = alloca { i8, i8, i8, i8 }, align 2 ; "line 9" to -debugify 50 %gep1 = getelementptr { i8, i8, i8, i8 }, { i8, i8, i8, i8 }* %a, i32 0, i32 1 51 %cast1 = bitcast i8* %gep1 to i16* 52 store volatile i16 0, i16* %cast1 53 %gep2 = getelementptr { i8, i8, i8, i8 }, { i8, i8, i8, i8 }* %a, i32 0, i32 2 54 %result = load i8, i8* %gep2 55 store i8 42, i8* %gep2 56 ret void 57} 58 59define void @PR13920(<2 x i64>* %a, i16* %b) { 60; Test that alignments on memcpy intrinsics get propagated to loads and stores. 61; CHECK-LABEL: @PR13920( 62; CHECK: load <2 x i64>, <2 x i64>* %a, align 2 63; CHECK: store <2 x i64> {{.*}}, <2 x i64>* {{.*}}, align 2 64; CHECK: ret void 65 66entry: 67 %aa = alloca <2 x i64>, align 16 68 %aptr = bitcast <2 x i64>* %a to i8* 69 %aaptr = bitcast <2 x i64>* %aa to i8* 70 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 2 %aaptr, i8* align 2 %aptr, i32 16, i1 false) 71 %bptr = bitcast i16* %b to i8* 72 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 2 %bptr, i8* align 2 %aaptr, i32 16, i1 false) 73 ret void 74} 75 76define void @test3(i8* %x) { 77; Test that when we promote an alloca to a type with lower ABI alignment, we 78; provide the needed explicit alignment that code using the alloca may be 79; expecting. However, also check that any offset within an alloca can in turn 80; reduce the alignment. 81; CHECK-LABEL: @test3( 82; CHECK: alloca [22 x i8], align 8 83; CHECK: alloca [18 x i8], align 2 84; CHECK: ret void 85 86entry: 87 %a = alloca { i8*, i8*, i8* } 88 %b = alloca { i8*, i8*, i8* } 89 %a_raw = bitcast { i8*, i8*, i8* }* %a to i8* 90 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 8 %a_raw, i8* align 8 %x, i32 22, i1 false) 91 %b_raw = bitcast { i8*, i8*, i8* }* %b to i8* 92 %b_gep = getelementptr i8, i8* %b_raw, i32 6 93 call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 2 %b_gep, i8* align 2 %x, i32 18, i1 false) 94 ret void 95} 96 97define void @test5() { 98; Test that we preserve underaligned loads and stores when splitting. The use 99; of volatile in this test case is just to force the loads and stores to not be 100; split or promoted out of existence. 101; 102; CHECK-LABEL: @test5( 103; CHECK: alloca [9 x i8] 104; CHECK: alloca [9 x i8] 105; CHECK: store volatile double 0.0{{.*}}, double* %{{.*}}, align 1 106; CHECK: load volatile i16, i16* %{{.*}}, align 1 107; CHECK: load double, double* %{{.*}}, align 1 108; CHECK: store volatile double %{{.*}}, double* %{{.*}}, align 1 109; CHECK: load volatile i16, i16* %{{.*}}, align 1 110; CHECK: ret void 111 112entry: 113 %a = alloca [18 x i8] 114 %raw1 = getelementptr inbounds [18 x i8], [18 x i8]* %a, i32 0, i32 0 115 %ptr1 = bitcast i8* %raw1 to double* 116 store volatile double 0.0, double* %ptr1, align 1 117 %weird_gep1 = getelementptr inbounds [18 x i8], [18 x i8]* %a, i32 0, i32 7 118 %weird_cast1 = bitcast i8* %weird_gep1 to i16* 119 %weird_load1 = load volatile i16, i16* %weird_cast1, align 1 120 121 %raw2 = getelementptr inbounds [18 x i8], [18 x i8]* %a, i32 0, i32 9 122 %ptr2 = bitcast i8* %raw2 to double* 123 %d1 = load double, double* %ptr1, align 1 124 store volatile double %d1, double* %ptr2, align 1 125 %weird_gep2 = getelementptr inbounds [18 x i8], [18 x i8]* %a, i32 0, i32 16 126 %weird_cast2 = bitcast i8* %weird_gep2 to i16* 127 %weird_load2 = load volatile i16, i16* %weird_cast2, align 1 128 129 ret void 130} 131 132define void @test6() { 133; We should set the alignment on all load and store operations; make sure 134; we choose an appropriate alignment. 135; CHECK-LABEL: @test6( 136; CHECK: alloca double, align 8{{$}} 137; CHECK: alloca double, align 8{{$}} 138; CHECK: store{{.*}}, align 8 139; CHECK: load{{.*}}, align 8 140; CHECK: store{{.*}}, align 8 141; CHECK-NOT: align 142; CHECK: ret void 143 144entry: 145 %a = alloca [16 x i8] 146 %raw1 = getelementptr inbounds [16 x i8], [16 x i8]* %a, i32 0, i32 0 147 %ptr1 = bitcast i8* %raw1 to double* 148 store volatile double 0.0, double* %ptr1, align 1 149 150 %raw2 = getelementptr inbounds [16 x i8], [16 x i8]* %a, i32 0, i32 8 151 %ptr2 = bitcast i8* %raw2 to double* 152 %val = load double, double* %ptr1, align 1 153 store volatile double %val, double* %ptr2, align 1 154 155 ret void 156} 157 158define void @test7(i8* %out) { 159; Test that we properly compute the destination alignment when rewriting 160; memcpys as direct loads or stores. 161; CHECK-LABEL: @test7( 162; CHECK-NOT: alloca 163 164entry: 165 %a = alloca [16 x i8] 166 %raw1 = getelementptr inbounds [16 x i8], [16 x i8]* %a, i32 0, i32 0 167 %ptr1 = bitcast i8* %raw1 to double* 168 %raw2 = getelementptr inbounds [16 x i8], [16 x i8]* %a, i32 0, i32 8 169 %ptr2 = bitcast i8* %raw2 to double* 170 171 call void @llvm.memcpy.p0i8.p0i8.i32(i8* %raw1, i8* %out, i32 16, i1 false) 172; CHECK: %[[val2:.*]] = load double, double* %{{.*}}, align 1 173; CHECK: %[[val1:.*]] = load double, double* %{{.*}}, align 1 174 175 %val1 = load double, double* %ptr2, align 1 176 %val2 = load double, double* %ptr1, align 1 177 178 store double %val1, double* %ptr1, align 1 179 store double %val2, double* %ptr2, align 1 180 181 call void @llvm.memcpy.p0i8.p0i8.i32(i8* %out, i8* %raw1, i32 16, i1 false) 182; CHECK: store double %[[val1]], double* %{{.*}}, align 1 183; CHECK: store double %[[val2]], double* %{{.*}}, align 1 184 185 ret void 186; CHECK: ret void 187} 188 189define void @test8() { 190; CHECK-LABEL: @test8( 191; CHECK: load i32, {{.*}}, align 1 192; CHECK: load i32, {{.*}}, align 1 193; CHECK: load i32, {{.*}}, align 1 194; CHECK: load i32, {{.*}}, align 1 195; CHECK: load i32, {{.*}}, align 1 196 197 %ptr = alloca [5 x i32], align 1 198 %ptr.8 = bitcast [5 x i32]* %ptr to i8* 199 call void @populate(i8* %ptr.8) 200 %val = load [5 x i32], [5 x i32]* %ptr, align 1 201 ret void 202} 203 204define void @test9() { 205; CHECK-LABEL: @test9( 206; CHECK: load i32, {{.*}}, align 8 207; CHECK: load i32, {{.*}}, align 4 208; CHECK: load i32, {{.*}}, align 8 209; CHECK: load i32, {{.*}}, align 4 210; CHECK: load i32, {{.*}}, align 8 211 212 %ptr = alloca [5 x i32], align 8 213 %ptr.8 = bitcast [5 x i32]* %ptr to i8* 214 call void @populate(i8* %ptr.8) 215 %val = load [5 x i32], [5 x i32]* %ptr, align 8 216 ret void 217} 218 219define void @test10() { 220; CHECK-LABEL: @test10( 221; CHECK: load i32, {{.*}}, align 2 222; CHECK: load i8, {{.*}}, align 2 223; CHECK: load i8, {{.*}}, align 1 224; CHECK: load i8, {{.*}}, align 2 225; CHECK: load i16, {{.*}}, align 2 226 227 %ptr = alloca {i32, i8, i8, {i8, i16}}, align 2 228 %ptr.8 = bitcast {i32, i8, i8, {i8, i16}}* %ptr to i8* 229 call void @populate(i8* %ptr.8) 230 %val = load {i32, i8, i8, {i8, i16}}, {i32, i8, i8, {i8, i16}}* %ptr, align 2 231 ret void 232} 233 234%struct = type { i32, i32 } 235define dso_local i32 @pr45010(%struct* %A) { 236; CHECK-LABEL: @pr45010 237; CHECK: load atomic volatile i32, {{.*}}, align 4 238 239 %B = alloca %struct, align 4 240 %A.i = getelementptr inbounds %struct, %struct* %A, i32 0, i32 0 241 %B.i = getelementptr inbounds %struct, %struct* %B, i32 0, i32 0 242 %1 = load i32, i32* %A.i, align 4 243 store atomic volatile i32 %1, i32* %B.i release, align 4 244 %2 = bitcast %struct* %B to i32* 245 %x = load atomic volatile i32, i32* %2 acquire, align 4 246 ret i32 %x 247} 248 249declare void @populate(i8*) 250