372 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=slp-vectorizer -slp-threshold=-999 -S -mtriple=x86_64-unknown-linux-gnu -mcpu=skylake < %s | FileCheck %s3 4declare i64 @may_inf_loop_ro() nounwind readonly5 6; Base case without allocas or stacksave7define void @basecase(ptr %a, ptr %b, ptr %c) {8; CHECK-LABEL: @basecase(9; CHECK-NEXT: [[TMP1:%.*]] = load <2 x ptr>, ptr [[A:%.*]], align 810; CHECK-NEXT: store ptr null, ptr [[A]], align 811; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i8, <2 x ptr> [[TMP1]], <2 x i32> splat (i32 1)12; CHECK-NEXT: store <2 x ptr> [[TMP2]], ptr [[B:%.*]], align 813; CHECK-NEXT: ret void14;15 16 %v1 = load ptr, ptr %a17 store ptr zeroinitializer, ptr %a18 %a2 = getelementptr ptr, ptr %a, i32 119 %v2 = load ptr, ptr %a220 21 %add1 = getelementptr i8, ptr %v1, i32 122 %add2 = getelementptr i8, ptr %v2, i32 123 24 store ptr %add1, ptr %b25 %b2 = getelementptr ptr, ptr %b, i32 126 store ptr %add2, ptr %b227 ret void28}29 30; Using two allocas and a buildvector31define void @allocas(ptr %a, ptr %b, ptr %c) {32; CHECK-LABEL: @allocas(33; CHECK-NEXT: [[V1:%.*]] = alloca i8, align 134; CHECK-NEXT: [[V2:%.*]] = alloca i8, align 135; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x ptr> poison, ptr [[V1]], i32 036; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x ptr> [[TMP1]], ptr [[V2]], i32 137; CHECK-NEXT: [[TMP3:%.*]] = getelementptr i8, <2 x ptr> [[TMP2]], <2 x i32> splat (i32 1)38; CHECK-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[V1]], i32 139; CHECK-NEXT: store ptr [[TMP4]], ptr [[A:%.*]], align 840; CHECK-NEXT: store <2 x ptr> [[TMP3]], ptr [[B:%.*]], align 841; CHECK-NEXT: ret void42;43 44 %v1 = alloca i845 %add1 = getelementptr i8, ptr %v1, i32 146 store ptr %add1, ptr %a47 %v2 = alloca i848 49 %add2 = getelementptr i8, ptr %v2, i32 150 51 store ptr %add1, ptr %b52 %b2 = getelementptr ptr, ptr %b, i32 153 store ptr %add2, ptr %b254 ret void55}56 57; Allocas can not be speculated above a potentially non-returning call58define void @allocas_speculation(ptr %a, ptr %b, ptr %c) {59; CHECK-LABEL: @allocas_speculation(60; CHECK-NEXT: [[V1:%.*]] = alloca i8, align 161; CHECK-NEXT: [[ADD1:%.*]] = getelementptr i8, ptr [[V1]], i32 162; CHECK-NEXT: store ptr [[ADD1]], ptr [[A:%.*]], align 863; CHECK-NEXT: [[TMP1:%.*]] = call i64 @may_inf_loop_ro()64; CHECK-NEXT: [[V2:%.*]] = alloca i8, align 165; CHECK-NEXT: [[ADD2:%.*]] = getelementptr i8, ptr [[V2]], i32 166; CHECK-NEXT: store ptr [[ADD1]], ptr [[B:%.*]], align 867; CHECK-NEXT: [[B2:%.*]] = getelementptr ptr, ptr [[B]], i32 168; CHECK-NEXT: store ptr [[ADD2]], ptr [[B2]], align 869; CHECK-NEXT: ret void70;71 72 %v1 = alloca i873 %add1 = getelementptr i8, ptr %v1, i32 174 store ptr %add1, ptr %a75 call i64 @may_inf_loop_ro()76 %v2 = alloca i877 78 %add2 = getelementptr i8, ptr %v2, i32 179 80 store ptr %add1, ptr %b81 %b2 = getelementptr ptr, ptr %b, i32 182 store ptr %add2, ptr %b283 ret void84}85 86; We must be careful not to lift the inalloca alloc above the stacksave here.87; We used to miscompile this example before adding explicit dependency handling88; for stacksave.89define void @stacksave(ptr %a, ptr %b, ptr %c) {90; CHECK-LABEL: @stacksave(91; CHECK-NEXT: [[V1:%.*]] = alloca i8, align 192; CHECK-NEXT: [[ADD1:%.*]] = getelementptr i8, ptr [[V1]], i32 193; CHECK-NEXT: store ptr [[ADD1]], ptr [[A:%.*]], align 894; CHECK-NEXT: [[STACK:%.*]] = call ptr @llvm.stacksave.p0()95; CHECK-NEXT: [[V2:%.*]] = alloca inalloca i8, align 196; CHECK-NEXT: call void @use(ptr inalloca(i8) [[V2]]) #[[ATTR4:[0-9]+]]97; CHECK-NEXT: call void @llvm.stackrestore.p0(ptr [[STACK]])98; CHECK-NEXT: [[ADD2:%.*]] = getelementptr i8, ptr [[V2]], i32 199; CHECK-NEXT: store ptr [[ADD1]], ptr [[B:%.*]], align 8100; CHECK-NEXT: [[B2:%.*]] = getelementptr ptr, ptr [[B]], i32 1101; CHECK-NEXT: store ptr [[ADD2]], ptr [[B2]], align 8102; CHECK-NEXT: ret void103;104 105 %v1 = alloca i8106 %add1 = getelementptr i8, ptr %v1, i32 1107 store ptr %add1, ptr %a108 109 %stack = call ptr @llvm.stacksave()110 %v2 = alloca inalloca i8111 call void @use(ptr inalloca(i8) %v2) readnone112 call void @llvm.stackrestore(ptr %stack)113 114 %add2 = getelementptr i8, ptr %v2, i32 1115 116 store ptr %add1, ptr %b117 %b2 = getelementptr ptr, ptr %b, i32 1118 store ptr %add2, ptr %b2119 ret void120}121 122define void @stacksave2(ptr %a, ptr %b, ptr %c) {123; CHECK-LABEL: @stacksave2(124; CHECK-NEXT: [[V1:%.*]] = alloca i8, align 1125; CHECK-NEXT: [[STACK:%.*]] = call ptr @llvm.stacksave.p0()126; CHECK-NEXT: [[V2:%.*]] = alloca inalloca i8, align 1127; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x ptr> poison, ptr [[V1]], i32 0128; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x ptr> [[TMP1]], ptr [[V2]], i32 1129; CHECK-NEXT: [[TMP3:%.*]] = getelementptr i8, <2 x ptr> [[TMP2]], <2 x i32> splat (i32 1)130; CHECK-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[V1]], i32 1131; CHECK-NEXT: store ptr [[TMP4]], ptr [[A:%.*]], align 8132; CHECK-NEXT: call void @use(ptr inalloca(i8) [[V2]]) #[[ATTR5:[0-9]+]]133; CHECK-NEXT: call void @llvm.stackrestore.p0(ptr [[STACK]])134; CHECK-NEXT: store <2 x ptr> [[TMP3]], ptr [[B:%.*]], align 8135; CHECK-NEXT: ret void136;137 138 %v1 = alloca i8139 %add1 = getelementptr i8, ptr %v1, i32 1140 141 %stack = call ptr @llvm.stacksave()142 store ptr %add1, ptr %a143 %v2 = alloca inalloca i8144 call void @use(ptr inalloca(i8) %v2) readonly145 call void @llvm.stackrestore(ptr %stack)146 147 %add2 = getelementptr i8, ptr %v2, i32 1148 149 store ptr %add1, ptr %b150 %b2 = getelementptr ptr, ptr %b, i32 1151 store ptr %add2, ptr %b2152 ret void153}154 155define void @stacksave3(ptr %a, ptr %b, ptr %c) {156; CHECK-LABEL: @stacksave3(157; CHECK-NEXT: [[STACK:%.*]] = call ptr @llvm.stacksave.p0()158; CHECK-NEXT: [[V1:%.*]] = alloca i8, align 1159; CHECK-NEXT: [[V2:%.*]] = alloca inalloca i8, align 1160; CHECK-NEXT: call void @use(ptr inalloca(i8) [[V2]]) #[[ATTR4]]161; CHECK-NEXT: call void @llvm.stackrestore.p0(ptr [[STACK]])162; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x ptr> poison, ptr [[V1]], i32 0163; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x ptr> [[TMP1]], ptr [[V2]], i32 1164; CHECK-NEXT: [[TMP3:%.*]] = getelementptr i8, <2 x ptr> [[TMP2]], <2 x i32> splat (i32 1)165; CHECK-NEXT: store <2 x ptr> [[TMP3]], ptr [[B:%.*]], align 8166; CHECK-NEXT: ret void167;168 169 %stack = call ptr @llvm.stacksave()170 %v1 = alloca i8171 172 %v2 = alloca inalloca i8173 call void @use(ptr inalloca(i8) %v2) readnone174 call void @llvm.stackrestore(ptr %stack)175 176 %add1 = getelementptr i8, ptr %v1, i32 1177 %add2 = getelementptr i8, ptr %v2, i32 1178 179 store ptr %add1, ptr %b180 %b2 = getelementptr ptr, ptr %b, i32 1181 store ptr %add2, ptr %b2182 ret void183}184 185; Here we have an alloca which needs to stay under the stacksave, but is not186; directly part of the vectorization tree. Instead, the stacksave is187; encountered during dependency scanning via the memory chain.188define void @stacksave4(ptr %a, ptr %b, ptr %c) {189; CHECK-LABEL: @stacksave4(190; CHECK-NEXT: [[TMP1:%.*]] = load <2 x ptr>, ptr [[A:%.*]], align 8191; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i8, <2 x ptr> [[TMP1]], <2 x i32> splat (i32 1)192; CHECK-NEXT: [[STACK:%.*]] = call ptr @llvm.stacksave.p0()193; CHECK-NEXT: [[X:%.*]] = alloca inalloca i8, align 1194; CHECK-NEXT: call void @use(ptr inalloca(i8) [[X]]) #[[ATTR4]]195; CHECK-NEXT: call void @llvm.stackrestore.p0(ptr [[STACK]])196; CHECK-NEXT: store <2 x ptr> [[TMP2]], ptr [[B:%.*]], align 8197; CHECK-NEXT: ret void198;199 200 %v1 = load ptr, ptr %a201 %a2 = getelementptr ptr, ptr %a, i32 1202 %v2 = load ptr, ptr %a2203 204 %add1 = getelementptr i8, ptr %v1, i32 1205 %add2 = getelementptr i8, ptr %v2, i32 1206 207 %stack = call ptr @llvm.stacksave()208 %x = alloca inalloca i8209 call void @use(ptr inalloca(i8) %x) readnone210 call void @llvm.stackrestore(ptr %stack)211 212 store ptr %add1, ptr %b213 %b2 = getelementptr ptr, ptr %b, i32 1214 store ptr %add2, ptr %b2215 ret void216}217 218define void @stacksave5(ptr %a, ptr %b, ptr %c) {219; CHECK-LABEL: @stacksave5(220; CHECK-NEXT: [[TMP1:%.*]] = load <2 x ptr>, ptr [[A:%.*]], align 8221; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i8, <2 x ptr> [[TMP1]], <2 x i32> splat (i32 1)222; CHECK-NEXT: [[STACK:%.*]] = call ptr @llvm.stacksave.p0()223; CHECK-NEXT: [[X:%.*]] = alloca inalloca i8, align 1224; CHECK-NEXT: call void @use(ptr inalloca(i8) [[X]]) #[[ATTR4]]225; CHECK-NEXT: call void @llvm.stackrestore.p0(ptr [[STACK]])226; CHECK-NEXT: store <2 x ptr> [[TMP2]], ptr [[B:%.*]], align 8227; CHECK-NEXT: ret void228;229 230 %v1 = load ptr, ptr %a231 %a2 = getelementptr ptr, ptr %a, i32 1232 %v2 = load ptr, ptr %a2233 234 %add1 = getelementptr i8, ptr %v1, i32 1235 %add2 = getelementptr i8, ptr %v2, i32 1236 237 %stack = call ptr @llvm.stacksave()238 %x = alloca inalloca i8239 call void @use(ptr inalloca(i8) %x) readnone240 call void @llvm.stackrestore(ptr %stack)241 242 store ptr %add1, ptr %b243 %b2 = getelementptr ptr, ptr %b, i32 1244 store ptr %add2, ptr %b2245 ret void246}247 248; Reordering the second alloca above the stackrestore while249; leaving the write to it below would introduce a write-after-free250; bug.251define void @stackrestore1(ptr %a, ptr %b, ptr %c) {252; CHECK-LABEL: @stackrestore1(253; CHECK-NEXT: [[STACK:%.*]] = call ptr @llvm.stacksave.p0()254; CHECK-NEXT: [[V1:%.*]] = alloca i8, align 1255; CHECK-NEXT: store i8 0, ptr [[V1]], align 1256; CHECK-NEXT: call void @llvm.stackrestore.p0(ptr [[STACK]])257; CHECK-NEXT: [[V2:%.*]] = alloca i8, align 1258; CHECK-NEXT: store i8 0, ptr [[V2]], align 1259; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x ptr> poison, ptr [[V1]], i32 0260; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x ptr> [[TMP1]], ptr [[V2]], i32 1261; CHECK-NEXT: [[TMP3:%.*]] = getelementptr i8, <2 x ptr> [[TMP2]], <2 x i32> splat (i32 1)262; CHECK-NEXT: store <2 x ptr> [[TMP3]], ptr [[B:%.*]], align 8263; CHECK-NEXT: ret void264;265 266 %stack = call ptr @llvm.stacksave()267 %v1 = alloca i8268 store i8 0, ptr %v1269 call void @llvm.stackrestore(ptr %stack)270 %v2 = alloca i8271 store i8 0, ptr %v2272 273 %add1 = getelementptr i8, ptr %v1, i32 1274 %add2 = getelementptr i8, ptr %v2, i32 1275 276 store ptr %add1, ptr %b277 %b2 = getelementptr ptr, ptr %b, i32 1278 store ptr %add2, ptr %b2279 ret void280}281 282declare void @use(ptr inalloca(i8))283declare ptr @llvm.stacksave()284declare void @llvm.stackrestore(ptr)285 286; The next set are reduced from previous regressions.287 288declare ptr @wibble(ptr)289declare void @quux(ptr inalloca(i32))290 291define void @ham() #1 {292; CHECK-LABEL: @ham(293; CHECK-NEXT: [[VAR2:%.*]] = alloca i8, align 1294; CHECK-NEXT: [[VAR3:%.*]] = alloca i8, align 1295; CHECK-NEXT: [[VAR4:%.*]] = alloca i8, align 1296; CHECK-NEXT: [[VAR5:%.*]] = alloca i8, align 1297; CHECK-NEXT: [[VAR12:%.*]] = alloca [12 x ptr], align 8298; CHECK-NEXT: [[VAR15:%.*]] = call ptr @wibble(ptr [[VAR2]])299; CHECK-NEXT: [[VAR16:%.*]] = call ptr @wibble(ptr [[VAR3]])300; CHECK-NEXT: [[VAR17:%.*]] = call ptr @wibble(ptr [[VAR4]])301; CHECK-NEXT: [[VAR23:%.*]] = call ptr @llvm.stacksave.p0()302; CHECK-NEXT: [[VAR24:%.*]] = alloca inalloca i32, align 4303; CHECK-NEXT: call void @quux(ptr inalloca(i32) [[VAR24]])304; CHECK-NEXT: call void @llvm.stackrestore.p0(ptr [[VAR23]])305; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x ptr> poison, ptr [[VAR4]], i32 0306; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <4 x ptr> [[TMP1]], <4 x ptr> poison, <4 x i32> zeroinitializer307; CHECK-NEXT: store <4 x ptr> [[TMP2]], ptr [[VAR12]], align 8308; CHECK-NEXT: [[VAR36:%.*]] = getelementptr inbounds [12 x ptr], ptr [[VAR12]], i32 0, i32 4309; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x ptr> [[TMP1]], ptr [[VAR5]], i32 1310; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <4 x ptr> [[TMP3]], <4 x ptr> poison, <4 x i32> <i32 0, i32 1, i32 1, i32 1>311; CHECK-NEXT: store <4 x ptr> [[TMP4]], ptr [[VAR36]], align 8312; CHECK-NEXT: ret void313;314 %var2 = alloca i8315 %var3 = alloca i8316 %var4 = alloca i8317 %var5 = alloca i8318 %var12 = alloca [12 x ptr]319 %var15 = call ptr @wibble(ptr %var2)320 %var16 = call ptr @wibble(ptr %var3)321 %var17 = call ptr @wibble(ptr %var4)322 %var23 = call ptr @llvm.stacksave()323 %var24 = alloca inalloca i32324 call void @quux(ptr inalloca(i32) %var24)325 call void @llvm.stackrestore(ptr %var23)326 store ptr %var4, ptr %var12327 %var33 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 1328 store ptr %var4, ptr %var33329 %var34 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 2330 store ptr %var4, ptr %var34331 %var35 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 3332 store ptr %var4, ptr %var35333 %var36 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 4334 store ptr %var4, ptr %var36335 %var37 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 5336 store ptr %var5, ptr %var37337 %var38 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 6338 store ptr %var5, ptr %var38339 %var39 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 7340 store ptr %var5, ptr %var39341 ret void342}343 344define void @spam() #1 {345; CHECK-LABEL: @spam(346; CHECK-NEXT: [[VAR4:%.*]] = alloca i8, align 1347; CHECK-NEXT: [[VAR5:%.*]] = alloca i8, align 1348; CHECK-NEXT: [[VAR12:%.*]] = alloca [12 x ptr], align 8349; CHECK-NEXT: [[VAR36:%.*]] = getelementptr inbounds [12 x ptr], ptr [[VAR12]], i32 0, i32 4350; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x ptr> poison, ptr [[VAR4]], i32 0351; CHECK-NEXT: [[TMP2:%.*]] = insertelement <4 x ptr> [[TMP1]], ptr [[VAR5]], i32 1352; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x ptr> [[TMP2]], <4 x ptr> poison, <4 x i32> <i32 0, i32 1, i32 1, i32 1>353; CHECK-NEXT: store <4 x ptr> [[TMP3]], ptr [[VAR36]], align 8354; CHECK-NEXT: ret void355;356 %var4 = alloca i8357 %var5 = alloca i8358 %var12 = alloca [12 x ptr]359 %var36 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 4360 store ptr %var4, ptr %var36361 %var37 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 5362 store ptr %var5, ptr %var37363 %var38 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 6364 store ptr %var5, ptr %var38365 %var39 = getelementptr inbounds [12 x ptr], ptr %var12, i32 0, i32 7366 store ptr %var5, ptr %var39367 ret void368}369 370attributes #0 = { nofree nosync nounwind willreturn }371attributes #1 = { "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+sse3,+x87" }372