1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -basic-aa -dse -enable-dse-memoryssa=false -S | FileCheck %s 3; RUN: opt < %s -aa-pipeline=basic-aa -passes=dse -enable-dse-memoryssa=false -S | FileCheck %s 4target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128" 5 6declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i1) nounwind 7declare void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* nocapture, i8, i64, i32) nounwind 8declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i1) nounwind 9declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i32) nounwind 10declare void @llvm.init.trampoline(i8*, i8*, i8*) 11 12define void @test1(i32* %Q, i32* %P) { 13; CHECK-LABEL: @test1( 14; CHECK-NEXT: store i32 0, i32* [[P:%.*]], align 4 15; CHECK-NEXT: ret void 16; 17 %DEAD = load i32, i32* %Q 18 store i32 %DEAD, i32* %P 19 store i32 0, i32* %P 20 ret void 21} 22 23; PR8576 - Should delete store of 10 even though p/q are may aliases. 24define void @test2(i32 *%p, i32 *%q) { 25; CHECK-LABEL: @test2( 26; CHECK-NEXT: store i32 20, i32* [[Q:%.*]], align 4 27; CHECK-NEXT: store i32 30, i32* [[P:%.*]], align 4 28; CHECK-NEXT: ret void 29; 30 store i32 10, i32* %p, align 4 31 store i32 20, i32* %q, align 4 32 store i32 30, i32* %p, align 4 33 ret void 34} 35 36 37; PR8677 38@g = global i32 1 39 40define i32 @test3(i32* %g_addr) nounwind { 41; CHECK-LABEL: @test3( 42; CHECK-NEXT: [[G_VALUE:%.*]] = load i32, i32* [[G_ADDR:%.*]], align 4 43; CHECK-NEXT: store i32 -1, i32* @g, align 4 44; CHECK-NEXT: store i32 [[G_VALUE]], i32* [[G_ADDR]], align 4 45; CHECK-NEXT: [[TMP3:%.*]] = load i32, i32* @g, align 4 46; CHECK-NEXT: ret i32 [[TMP3]] 47; 48 %g_value = load i32, i32* %g_addr, align 4 49 store i32 -1, i32* @g, align 4 50 store i32 %g_value, i32* %g_addr, align 4 51 %tmp3 = load i32, i32* @g, align 4 52 ret i32 %tmp3 53} 54 55 56define void @test4(i32* %Q) { 57; CHECK-LABEL: @test4( 58; CHECK-NEXT: [[A:%.*]] = load i32, i32* [[Q:%.*]], align 4 59; CHECK-NEXT: store volatile i32 [[A]], i32* [[Q]], align 4 60; CHECK-NEXT: ret void 61; 62 %a = load i32, i32* %Q 63 store volatile i32 %a, i32* %Q 64 ret void 65} 66 67define void @test5(i32* %Q) { 68; CHECK-LABEL: @test5( 69; CHECK-NEXT: [[A:%.*]] = load volatile i32, i32* [[Q:%.*]], align 4 70; CHECK-NEXT: ret void 71; 72 %a = load volatile i32, i32* %Q 73 store i32 %a, i32* %Q 74 ret void 75} 76 77; Should delete store of 10 even though memset is a may-store to P (P and Q may 78; alias). 79define void @test6(i32 *%p, i8 *%q) { 80; CHECK-LABEL: @test6( 81; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[Q:%.*]], i8 42, i64 900, i1 false) 82; CHECK-NEXT: store i32 30, i32* [[P:%.*]], align 4 83; CHECK-NEXT: ret void 84; 85 store i32 10, i32* %p, align 4 ;; dead. 86 call void @llvm.memset.p0i8.i64(i8* %q, i8 42, i64 900, i1 false) 87 store i32 30, i32* %p, align 4 88 ret void 89} 90 91; Should delete store of 10 even though memset is a may-store to P (P and Q may 92; alias). 93define void @test6_atomic(i32* align 4 %p, i8* align 4 %q) { 94; CHECK-LABEL: @test6_atomic( 95; CHECK-NEXT: call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 4 [[Q:%.*]], i8 42, i64 900, i32 4) 96; CHECK-NEXT: store atomic i32 30, i32* [[P:%.*]] unordered, align 4 97; CHECK-NEXT: ret void 98; 99 store atomic i32 10, i32* %p unordered, align 4 ;; dead. 100 call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 4 %q, i8 42, i64 900, i32 4) 101 store atomic i32 30, i32* %p unordered, align 4 102 ret void 103} 104 105; Should delete store of 10 even though memcpy is a may-store to P (P and Q may 106; alias). 107define void @test7(i32 *%p, i8 *%q, i8* noalias %r) { 108; CHECK-LABEL: @test7( 109; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[Q:%.*]], i8* [[R:%.*]], i64 900, i1 false) 110; CHECK-NEXT: store i32 30, i32* [[P:%.*]], align 4 111; CHECK-NEXT: ret void 112; 113 store i32 10, i32* %p, align 4 ;; dead. 114 call void @llvm.memcpy.p0i8.p0i8.i64(i8* %q, i8* %r, i64 900, i1 false) 115 store i32 30, i32* %p, align 4 116 ret void 117} 118 119; Should delete store of 10 even though memcpy is a may-store to P (P and Q may 120; alias). 121define void @test7_atomic(i32* align 4 %p, i8* align 4 %q, i8* noalias align 4 %r) { 122; CHECK-LABEL: @test7_atomic( 123; CHECK-NEXT: call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 4 [[Q:%.*]], i8* align 4 [[R:%.*]], i64 900, i32 4) 124; CHECK-NEXT: store atomic i32 30, i32* [[P:%.*]] unordered, align 4 125; CHECK-NEXT: ret void 126; 127 store atomic i32 10, i32* %p unordered, align 4 ;; dead. 128 call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 4 %q, i8* align 4 %r, i64 900, i32 4) 129 store atomic i32 30, i32* %p unordered, align 4 130 ret void 131} 132 133; Do not delete stores that are only partially killed. 134define i32 @test8() { 135; CHECK-LABEL: @test8( 136; CHECK-NEXT: [[V:%.*]] = alloca i32, align 4 137; CHECK-NEXT: store i32 1234567, i32* [[V]], align 4 138; CHECK-NEXT: [[X:%.*]] = load i32, i32* [[V]], align 4 139; CHECK-NEXT: ret i32 [[X]] 140; 141 %V = alloca i32 142 store i32 1234567, i32* %V 143 %V2 = bitcast i32* %V to i8* 144 store i8 0, i8* %V2 145 %X = load i32, i32* %V 146 ret i32 %X 147 148} 149 150 151; Test for byval handling. 152%struct.x = type { i32, i32, i32, i32 } 153define void @test9(%struct.x* byval(%struct.x) %a) nounwind { 154; CHECK-LABEL: @test9( 155; CHECK-NEXT: ret void 156; 157 %tmp2 = getelementptr %struct.x, %struct.x* %a, i32 0, i32 0 158 store i32 1, i32* %tmp2, align 4 159 ret void 160} 161 162; Test for inalloca handling. 163define void @test9_2(%struct.x* inalloca %a) nounwind { 164; CHECK-LABEL: @test9_2( 165; CHECK-NEXT: ret void 166; 167 %tmp2 = getelementptr %struct.x, %struct.x* %a, i32 0, i32 0 168 store i32 1, i32* %tmp2, align 4 169 ret void 170} 171 172; Test for preallocated handling. 173define void @test9_3(%struct.x* preallocated(%struct.x) %a) nounwind { 174; CHECK-LABEL: @test9_3( 175; CHECK-NEXT: ret void 176; 177 %tmp2 = getelementptr %struct.x, %struct.x* %a, i32 0, i32 0 178 store i32 1, i32* %tmp2, align 4 179 ret void 180} 181 182; va_arg has fuzzy dependence, the store shouldn't be zapped. 183define double @test10(i8* %X) { 184; CHECK-LABEL: @test10( 185; CHECK-NEXT: [[X_ADDR:%.*]] = alloca i8*, align 8 186; CHECK-NEXT: store i8* [[X:%.*]], i8** [[X_ADDR]], align 8 187; CHECK-NEXT: [[TMP_0:%.*]] = va_arg i8** [[X_ADDR]], double 188; CHECK-NEXT: ret double [[TMP_0]] 189; 190 %X_addr = alloca i8* 191 store i8* %X, i8** %X_addr 192 %tmp.0 = va_arg i8** %X_addr, double 193 ret double %tmp.0 194} 195 196 197; DSE should delete the dead trampoline. 198declare void @test11f() 199define void @test11() { 200; CHECK-LABEL: @test11( 201; CHECK-NEXT: ret void 202; 203 %storage = alloca [10 x i8], align 16 ; <[10 x i8]*> [#uses=1] 204 %cast = getelementptr [10 x i8], [10 x i8]* %storage, i32 0, i32 0 ; <i8*> [#uses=1] 205 call void @llvm.init.trampoline( i8* %cast, i8* bitcast (void ()* @test11f to i8*), i8* null ) ; <i8*> [#uses=1] 206 ret void 207} 208 209 210; PR2599 - load -> store to same address. 211define void @test12({ i32, i32 }* %x) nounwind { 212; CHECK-LABEL: @test12( 213; CHECK-NEXT: [[TMP7:%.*]] = getelementptr { i32, i32 }, { i32, i32 }* [[X:%.*]], i32 0, i32 1 214; CHECK-NEXT: [[TMP8:%.*]] = load i32, i32* [[TMP7]], align 4 215; CHECK-NEXT: [[TMP17:%.*]] = sub i32 0, [[TMP8]] 216; CHECK-NEXT: store i32 [[TMP17]], i32* [[TMP7]], align 4 217; CHECK-NEXT: ret void 218; 219 %tmp4 = getelementptr { i32, i32 }, { i32, i32 }* %x, i32 0, i32 0 220 %tmp5 = load i32, i32* %tmp4, align 4 221 %tmp7 = getelementptr { i32, i32 }, { i32, i32 }* %x, i32 0, i32 1 222 %tmp8 = load i32, i32* %tmp7, align 4 223 %tmp17 = sub i32 0, %tmp8 224 store i32 %tmp5, i32* %tmp4, align 4 225 store i32 %tmp17, i32* %tmp7, align 4 226 ret void 227} 228 229 230; %P doesn't escape, the DEAD instructions should be removed. 231declare void @test13f() 232define i32* @test13() { 233; CHECK-LABEL: @test13( 234; CHECK-NEXT: [[PTR:%.*]] = tail call i8* @malloc(i32 4) 235; CHECK-NEXT: [[P:%.*]] = bitcast i8* [[PTR]] to i32* 236; CHECK-NEXT: call void @test13f() 237; CHECK-NEXT: store i32 0, i32* [[P]], align 4 238; CHECK-NEXT: ret i32* [[P]] 239; 240 %ptr = tail call i8* @malloc(i32 4) 241 %P = bitcast i8* %ptr to i32* 242 %DEAD = load i32, i32* %P 243 %DEAD2 = add i32 %DEAD, 1 244 store i32 %DEAD2, i32* %P 245 call void @test13f( ) 246 store i32 0, i32* %P 247 ret i32* %P 248} 249 250define i32 addrspace(1)* @test13_addrspacecast() { 251; CHECK-LABEL: @test13_addrspacecast( 252; CHECK-NEXT: [[P:%.*]] = tail call i8* @malloc(i32 4) 253; CHECK-NEXT: [[P_BC:%.*]] = bitcast i8* [[P]] to i32* 254; CHECK-NEXT: [[P:%.*]] = addrspacecast i32* [[P_BC]] to i32 addrspace(1)* 255; CHECK-NEXT: call void @test13f() 256; CHECK-NEXT: store i32 0, i32 addrspace(1)* [[P]], align 4 257; CHECK-NEXT: ret i32 addrspace(1)* [[P]] 258; 259 %p = tail call i8* @malloc(i32 4) 260 %p.bc = bitcast i8* %p to i32* 261 %P = addrspacecast i32* %p.bc to i32 addrspace(1)* 262 %DEAD = load i32, i32 addrspace(1)* %P 263 %DEAD2 = add i32 %DEAD, 1 264 store i32 %DEAD2, i32 addrspace(1)* %P 265 call void @test13f( ) 266 store i32 0, i32 addrspace(1)* %P 267 ret i32 addrspace(1)* %P 268} 269 270declare noalias i8* @malloc(i32) 271declare noalias i8* @calloc(i32, i32) 272declare noalias i8* @aligned_alloc(i32, i32) 273declare void @free(i8*) 274 275 276define void @test14(i32* %Q) { 277; CHECK-LABEL: @test14( 278; CHECK-NEXT: ret void 279; 280 %P = alloca i32 281 %DEAD = load i32, i32* %Q 282 store i32 %DEAD, i32* %P 283 ret void 284 285} 286 287; Dead store on an aligned_alloc: should know that %M doesn't alias with %A. 288define i32 @test14a(i8* %M, i8 %value) { 289; CHECK-LABEL: @test14a( 290; CHECK-NEXT: [[A:%.*]] = tail call i8* @aligned_alloc(i32 32, i32 1024) 291; CHECK-NEXT: tail call void @free(i8* [[A]]) 292; CHECK-NEXT: ret i32 0 293; 294 %A = tail call i8* @aligned_alloc(i32 32, i32 1024) 295 store i8 %value, i8* %A 296 tail call void @free(i8* %A) 297 ret i32 0 298} 299 300; PR8701 301 302;; Fully dead overwrite of memcpy. 303define void @test15(i8* %P, i8* %Q) nounwind ssp { 304; CHECK-LABEL: @test15( 305; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 306; CHECK-NEXT: ret void 307; 308 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 309 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 310 ret void 311} 312 313;; Fully dead overwrite of memcpy. 314define void @test15_atomic(i8* %P, i8* %Q) nounwind ssp { 315; CHECK-LABEL: @test15_atomic( 316; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 317; CHECK-NEXT: ret void 318; 319 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 320 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 321 ret void 322} 323 324;; Fully dead overwrite of memcpy. 325define void @test15_atomic_weaker(i8* %P, i8* %Q) nounwind ssp { 326; CHECK-LABEL: @test15_atomic_weaker( 327; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 328; CHECK-NEXT: ret void 329; 330 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i1 false) 331 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 332 ret void 333} 334 335;; Fully dead overwrite of memcpy. 336define void @test15_atomic_weaker_2(i8* %P, i8* %Q) nounwind ssp { 337; CHECK-LABEL: @test15_atomic_weaker_2( 338; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i1 false) 339; CHECK-NEXT: ret void 340; 341 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 342 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i1 false) 343 ret void 344} 345 346;; Full overwrite of smaller memcpy. 347define void @test16(i8* %P, i8* %Q) nounwind ssp { 348; CHECK-LABEL: @test16( 349; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 350; CHECK-NEXT: ret void 351; 352 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 8, i1 false) 353 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 354 ret void 355} 356 357;; Full overwrite of smaller memcpy. 358define void @test16_atomic(i8* %P, i8* %Q) nounwind ssp { 359; CHECK-LABEL: @test16_atomic( 360; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 361; CHECK-NEXT: ret void 362; 363 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 8, i32 1) 364 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 365 ret void 366} 367 368;; Full overwrite of smaller memory where overwrite has stronger atomicity 369define void @test16_atomic_weaker(i8* %P, i8* %Q) nounwind ssp { 370; CHECK-LABEL: @test16_atomic_weaker( 371; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 372; CHECK-NEXT: ret void 373; 374 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 8, i1 false) 375 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 376 ret void 377} 378 379;; Full overwrite of smaller memory where overwrite has weaker atomicity. 380define void @test16_atomic_weaker_2(i8* %P, i8* %Q) nounwind ssp { 381; CHECK-LABEL: @test16_atomic_weaker_2( 382; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i1 false) 383; CHECK-NEXT: ret void 384; 385 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 8, i32 1) 386 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i1 false) 387 ret void 388} 389 390;; Overwrite of memset by memcpy. 391define void @test17(i8* %P, i8* noalias %Q) nounwind ssp { 392; CHECK-LABEL: @test17( 393; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 394; CHECK-NEXT: ret void 395; 396 tail call void @llvm.memset.p0i8.i64(i8* %P, i8 42, i64 8, i1 false) 397 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 398 ret void 399} 400 401;; Overwrite of memset by memcpy. 402define void @test17_atomic(i8* %P, i8* noalias %Q) nounwind ssp { 403; CHECK-LABEL: @test17_atomic( 404; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 405; CHECK-NEXT: ret void 406; 407 tail call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 1 %P, i8 42, i64 8, i32 1) 408 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 409 ret void 410} 411 412;; Overwrite of memset by memcpy. Overwrite is stronger atomicity. We can 413;; remove the memset. 414define void @test17_atomic_weaker(i8* %P, i8* noalias %Q) nounwind ssp { 415; CHECK-LABEL: @test17_atomic_weaker( 416; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 417; CHECK-NEXT: ret void 418; 419 tail call void @llvm.memset.p0i8.i64(i8* align 1 %P, i8 42, i64 8, i1 false) 420 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 421 ret void 422} 423 424;; Overwrite of memset by memcpy. Overwrite is weaker atomicity. We can remove 425;; the memset. 426define void @test17_atomic_weaker_2(i8* %P, i8* noalias %Q) nounwind ssp { 427; CHECK-LABEL: @test17_atomic_weaker_2( 428; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i1 false) 429; CHECK-NEXT: ret void 430; 431 tail call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 1 %P, i8 42, i64 8, i32 1) 432 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i1 false) 433 ret void 434} 435 436; Should not delete the volatile memset. 437define void @test17v(i8* %P, i8* %Q) nounwind ssp { 438; CHECK-LABEL: @test17v( 439; CHECK-NEXT: tail call void @llvm.memset.p0i8.i64(i8* [[P:%.*]], i8 42, i64 8, i1 true) 440; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P]], i8* [[Q:%.*]], i64 12, i1 false) 441; CHECK-NEXT: ret void 442; 443 tail call void @llvm.memset.p0i8.i64(i8* %P, i8 42, i64 8, i1 true) 444 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 445 ret void 446} 447 448; PR8728 449; Do not delete instruction where possible situation is: 450; A = B 451; A = A 452; 453; NB! See PR11763 - currently LLVM allows memcpy's source and destination to be 454; equal (but not inequal and overlapping). 455define void @test18(i8* %P, i8* %Q, i8* %R) nounwind ssp { 456; CHECK-LABEL: @test18( 457; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 458; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P]], i8* [[R:%.*]], i64 12, i1 false) 459; CHECK-NEXT: ret void 460; 461 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 462 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %R, i64 12, i1 false) 463 ret void 464} 465 466define void @test18_atomic(i8* %P, i8* %Q, i8* %R) nounwind ssp { 467; CHECK-LABEL: @test18_atomic( 468; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 469; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[R:%.*]], i64 12, i32 1) 470; CHECK-NEXT: ret void 471; 472 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 473 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %R, i64 12, i32 1) 474 ret void 475} 476 477 478; The store here is not dead because the byval call reads it. 479declare void @test19f({i32}* byval({i32}) align 4 %P) 480 481define void @test19({i32} * nocapture byval({i32}) align 4 %arg5) nounwind ssp { 482; CHECK-LABEL: @test19( 483; CHECK-NEXT: bb: 484; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds { i32 }, { i32 }* [[ARG5:%.*]], i32 0, i32 0 485; CHECK-NEXT: store i32 912, i32* [[TMP7]], align 4 486; CHECK-NEXT: call void @test19f({ i32 }* byval({ i32 }) align 4 [[ARG5]]) 487; CHECK-NEXT: ret void 488; 489bb: 490 %tmp7 = getelementptr inbounds {i32}, {i32}* %arg5, i32 0, i32 0 491 store i32 912, i32* %tmp7 492 call void @test19f({i32}* byval({i32}) align 4 %arg5) 493 ret void 494 495} 496 497define void @test20() { 498; CHECK-LABEL: @test20( 499; CHECK-NEXT: ret void 500; 501 %m = call i8* @malloc(i32 24) 502 store i8 0, i8* %m 503 ret void 504} 505 506define void @test21() { 507; CHECK-LABEL: @test21( 508; CHECK-NEXT: ret void 509; 510 %m = call i8* @calloc(i32 9, i32 7) 511 store i8 0, i8* %m 512 ret void 513} 514 515define void @test22(i1 %i, i32 %k, i32 %m) nounwind { 516; CHECK-LABEL: @test22( 517; CHECK-NEXT: ret void 518; 519 %k.addr = alloca i32 520 %m.addr = alloca i32 521 %k.addr.m.addr = select i1 %i, i32* %k.addr, i32* %m.addr 522 store i32 0, i32* %k.addr.m.addr, align 4 523 ret void 524} 525 526; PR13547 527declare noalias i8* @strdup(i8* nocapture) nounwind 528define noalias i8* @test23() nounwind uwtable ssp { 529; CHECK-LABEL: @test23( 530; CHECK-NEXT: [[X:%.*]] = alloca [2 x i8], align 1 531; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [2 x i8], [2 x i8]* [[X]], i64 0, i64 0 532; CHECK-NEXT: store i8 97, i8* [[ARRAYIDX]], align 1 533; CHECK-NEXT: [[ARRAYIDX1:%.*]] = getelementptr inbounds [2 x i8], [2 x i8]* [[X]], i64 0, i64 1 534; CHECK-NEXT: store i8 0, i8* [[ARRAYIDX1]], align 1 535; CHECK-NEXT: [[CALL:%.*]] = call i8* @strdup(i8* [[ARRAYIDX]]) [[ATTR5:#.*]] 536; CHECK-NEXT: ret i8* [[CALL]] 537; 538 %x = alloca [2 x i8], align 1 539 %arrayidx = getelementptr inbounds [2 x i8], [2 x i8]* %x, i64 0, i64 0 540 store i8 97, i8* %arrayidx, align 1 541 %arrayidx1 = getelementptr inbounds [2 x i8], [2 x i8]* %x, i64 0, i64 1 542 store i8 0, i8* %arrayidx1, align 1 543 %call = call i8* @strdup(i8* %arrayidx) nounwind 544 ret i8* %call 545} 546 547; Make sure same sized store to later element is deleted 548define void @test24([2 x i32]* %a, i32 %b, i32 %c) nounwind { 549; CHECK-LABEL: @test24( 550; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds [2 x i32], [2 x i32]* [[A:%.*]], i64 0, i64 0 551; CHECK-NEXT: store i32 [[B:%.*]], i32* [[TMP1]], align 4 552; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds [2 x i32], [2 x i32]* [[A]], i64 0, i64 1 553; CHECK-NEXT: store i32 [[C:%.*]], i32* [[TMP2]], align 4 554; CHECK-NEXT: ret void 555; 556 %1 = getelementptr inbounds [2 x i32], [2 x i32]* %a, i64 0, i64 0 557 store i32 0, i32* %1, align 4 558 %2 = getelementptr inbounds [2 x i32], [2 x i32]* %a, i64 0, i64 1 559 store i32 0, i32* %2, align 4 560 %3 = getelementptr inbounds [2 x i32], [2 x i32]* %a, i64 0, i64 0 561 store i32 %b, i32* %3, align 4 562 %4 = getelementptr inbounds [2 x i32], [2 x i32]* %a, i64 0, i64 1 563 store i32 %c, i32* %4, align 4 564 ret void 565} 566 567; Check another case like PR13547 where strdup is not like malloc. 568define i8* @test25(i8* %p) nounwind { 569; CHECK-LABEL: @test25( 570; CHECK-NEXT: [[P_4:%.*]] = getelementptr i8, i8* [[P:%.*]], i64 4 571; CHECK-NEXT: [[TMP:%.*]] = load i8, i8* [[P_4]], align 1 572; CHECK-NEXT: store i8 0, i8* [[P_4]], align 1 573; CHECK-NEXT: [[Q:%.*]] = call i8* @strdup(i8* [[P]]) [[ATTR8:#.*]] 574; CHECK-NEXT: store i8 [[TMP]], i8* [[P_4]], align 1 575; CHECK-NEXT: ret i8* [[Q]] 576; 577 %p.4 = getelementptr i8, i8* %p, i64 4 578 %tmp = load i8, i8* %p.4, align 1 579 store i8 0, i8* %p.4, align 1 580 %q = call i8* @strdup(i8* %p) nounwind optsize 581 store i8 %tmp, i8* %p.4, align 1 582 ret i8* %q 583} 584 585; Remove redundant store if loaded value is in another block. 586define i32 @test26(i1 %c, i32* %p) { 587; CHECK-LABEL: @test26( 588; CHECK-NEXT: entry: 589; CHECK-NEXT: br i1 [[C:%.*]], label [[BB1:%.*]], label [[BB2:%.*]] 590; CHECK: bb1: 591; CHECK-NEXT: br label [[BB3:%.*]] 592; CHECK: bb2: 593; CHECK-NEXT: br label [[BB3]] 594; CHECK: bb3: 595; CHECK-NEXT: ret i32 0 596; 597entry: 598 %v = load i32, i32* %p, align 4 599 br i1 %c, label %bb1, label %bb2 600bb1: 601 br label %bb3 602bb2: 603 store i32 %v, i32* %p, align 4 604 br label %bb3 605bb3: 606 ret i32 0 607} 608 609; Remove redundant store if loaded value is in another block. 610define i32 @test27(i1 %c, i32* %p) { 611; CHECK-LABEL: @test27( 612; CHECK-NEXT: entry: 613; CHECK-NEXT: br i1 [[C:%.*]], label [[BB1:%.*]], label [[BB2:%.*]] 614; CHECK: bb1: 615; CHECK-NEXT: br label [[BB3:%.*]] 616; CHECK: bb2: 617; CHECK-NEXT: br label [[BB3]] 618; CHECK: bb3: 619; CHECK-NEXT: ret i32 0 620; 621entry: 622 %v = load i32, i32* %p, align 4 623 br i1 %c, label %bb1, label %bb2 624bb1: 625 br label %bb3 626bb2: 627 br label %bb3 628bb3: 629 store i32 %v, i32* %p, align 4 630 ret i32 0 631} 632 633; Don't remove redundant store because of may-aliased store. 634define i32 @test28(i1 %c, i32* %p, i32* %p2, i32 %i) { 635; CHECK-LABEL: @test28( 636; CHECK-NEXT: entry: 637; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[P:%.*]], align 4 638; CHECK-NEXT: store i32 [[I:%.*]], i32* [[P2:%.*]], align 4 639; CHECK-NEXT: br i1 [[C:%.*]], label [[BB1:%.*]], label [[BB2:%.*]] 640; CHECK: bb1: 641; CHECK-NEXT: br label [[BB3:%.*]] 642; CHECK: bb2: 643; CHECK-NEXT: br label [[BB3]] 644; CHECK: bb3: 645; CHECK-NEXT: store i32 [[V]], i32* [[P]], align 4 646; CHECK-NEXT: ret i32 0 647; 648entry: 649 %v = load i32, i32* %p, align 4 650 651 ; Might overwrite value at %p 652 store i32 %i, i32* %p2, align 4 653 br i1 %c, label %bb1, label %bb2 654bb1: 655 br label %bb3 656bb2: 657 br label %bb3 658bb3: 659 store i32 %v, i32* %p, align 4 660 ret i32 0 661} 662 663; Don't remove redundant store because of may-aliased store. 664define i32 @test29(i1 %c, i32* %p, i32* %p2, i32 %i) { 665; CHECK-LABEL: @test29( 666; CHECK-NEXT: entry: 667; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[P:%.*]], align 4 668; CHECK-NEXT: br i1 [[C:%.*]], label [[BB1:%.*]], label [[BB2:%.*]] 669; CHECK: bb1: 670; CHECK-NEXT: br label [[BB3:%.*]] 671; CHECK: bb2: 672; CHECK-NEXT: store i32 [[I:%.*]], i32* [[P2:%.*]], align 4 673; CHECK-NEXT: br label [[BB3]] 674; CHECK: bb3: 675; CHECK-NEXT: store i32 [[V]], i32* [[P]], align 4 676; CHECK-NEXT: ret i32 0 677; 678entry: 679 %v = load i32, i32* %p, align 4 680 br i1 %c, label %bb1, label %bb2 681bb1: 682 br label %bb3 683bb2: 684 ; Might overwrite value at %p 685 store i32 %i, i32* %p2, align 4 686 br label %bb3 687bb3: 688 store i32 %v, i32* %p, align 4 689 ret i32 0 690} 691 692declare void @unknown_func() 693 694; Don't remove redundant store because of unknown call. 695define i32 @test30(i1 %c, i32* %p, i32 %i) { 696; CHECK-LABEL: @test30( 697; CHECK-NEXT: entry: 698; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[P:%.*]], align 4 699; CHECK-NEXT: br i1 [[C:%.*]], label [[BB1:%.*]], label [[BB2:%.*]] 700; CHECK: bb1: 701; CHECK-NEXT: br label [[BB3:%.*]] 702; CHECK: bb2: 703; CHECK-NEXT: call void @unknown_func() 704; CHECK-NEXT: br label [[BB3]] 705; CHECK: bb3: 706; CHECK-NEXT: store i32 [[V]], i32* [[P]], align 4 707; CHECK-NEXT: ret i32 0 708; 709entry: 710 %v = load i32, i32* %p, align 4 711 br i1 %c, label %bb1, label %bb2 712bb1: 713 br label %bb3 714bb2: 715 ; Might overwrite value at %p 716 call void @unknown_func() 717 br label %bb3 718bb3: 719 store i32 %v, i32* %p, align 4 720 ret i32 0 721} 722 723; Remove redundant store if loaded value is in another block inside a loop. 724define i32 @test31(i1 %c, i32* %p, i32 %i) { 725; CHECK-LABEL: @test31( 726; CHECK-NEXT: entry: 727; CHECK-NEXT: br label [[BB1:%.*]] 728; CHECK: bb1: 729; CHECK-NEXT: br i1 undef, label [[BB1]], label [[BB2:%.*]] 730; CHECK: bb2: 731; CHECK-NEXT: ret i32 0 732; 733entry: 734 %v = load i32, i32* %p, align 4 735 br label %bb1 736bb1: 737 store i32 %v, i32* %p, align 4 738 br i1 undef, label %bb1, label %bb2 739bb2: 740 ret i32 0 741} 742 743; Don't remove redundant store in a loop with a may-alias store. 744define i32 @test32(i1 %c, i32* %p, i32 %i) { 745; CHECK-LABEL: @test32( 746; CHECK-NEXT: entry: 747; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[P:%.*]], align 4 748; CHECK-NEXT: br label [[BB1:%.*]] 749; CHECK: bb1: 750; CHECK-NEXT: store i32 [[V]], i32* [[P]], align 4 751; CHECK-NEXT: call void @unknown_func() 752; CHECK-NEXT: br i1 undef, label [[BB1]], label [[BB2:%.*]] 753; CHECK: bb2: 754; CHECK-NEXT: ret i32 0 755; 756entry: 757 %v = load i32, i32* %p, align 4 758 br label %bb1 759bb1: 760 store i32 %v, i32* %p, align 4 761 ; Might read and overwrite value at %p 762 call void @unknown_func() 763 br i1 undef, label %bb1, label %bb2 764bb2: 765 ret i32 0 766} 767 768; Remove redundant store, which is in the lame loop as the load. 769define i32 @test33(i1 %c, i32* %p, i32 %i) { 770; CHECK-LABEL: @test33( 771; CHECK-NEXT: entry: 772; CHECK-NEXT: br label [[BB1:%.*]] 773; CHECK: bb1: 774; CHECK-NEXT: br label [[BB2:%.*]] 775; CHECK: bb2: 776; CHECK-NEXT: call void @unknown_func() 777; CHECK-NEXT: br i1 undef, label [[BB1]], label [[BB3:%.*]] 778; CHECK: bb3: 779; CHECK-NEXT: ret i32 0 780; 781entry: 782 br label %bb1 783bb1: 784 %v = load i32, i32* %p, align 4 785 br label %bb2 786bb2: 787 store i32 %v, i32* %p, align 4 788 ; Might read and overwrite value at %p, but doesn't matter. 789 call void @unknown_func() 790 br i1 undef, label %bb1, label %bb3 791bb3: 792 ret i32 0 793} 794 795; Don't remove redundant store: unknown_func could unwind 796define void @test34(i32* noalias %p) { 797; CHECK-LABEL: @test34( 798; CHECK-NEXT: store i32 1, i32* [[P:%.*]], align 4 799; CHECK-NEXT: call void @unknown_func() 800; CHECK-NEXT: store i32 0, i32* [[P]], align 4 801; CHECK-NEXT: ret void 802; 803 store i32 1, i32* %p 804 call void @unknown_func() 805 store i32 0, i32* %p 806 ret void 807} 808 809; Remove redundant store even with an unwinding function in the same block 810define void @test35(i32* noalias %p) { 811; CHECK-LABEL: @test35( 812; CHECK-NEXT: call void @unknown_func() 813; CHECK-NEXT: store i32 0, i32* [[P:%.*]], align 4 814; CHECK-NEXT: ret void 815; 816 call void @unknown_func() 817 store i32 1, i32* %p 818 store i32 0, i32* %p 819 ret void 820} 821 822; We cannot optimize away the first memmove since %P could overlap with %Q. 823define void @test36(i8* %P, i8* %Q) { 824; CHECK-LABEL: @test36( 825; CHECK-NEXT: tail call void @llvm.memmove.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 826; CHECK-NEXT: tail call void @llvm.memmove.p0i8.p0i8.i64(i8* [[P]], i8* [[Q]], i64 12, i1 false) 827; CHECK-NEXT: ret void 828; 829 830 tail call void @llvm.memmove.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 831 tail call void @llvm.memmove.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 832 ret void 833} 834 835define void @test36_atomic(i8* %P, i8* %Q) { 836; CHECK-LABEL: @test36_atomic( 837; CHECK-NEXT: tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 838; CHECK-NEXT: tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[Q]], i64 12, i32 1) 839; CHECK-NEXT: ret void 840; 841 842 tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 843 tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 844 ret void 845} 846 847define void @test37(i8* %P, i8* %Q, i8* %R) { 848; CHECK-LABEL: @test37( 849; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 850; CHECK-NEXT: tail call void @llvm.memmove.p0i8.p0i8.i64(i8* [[P]], i8* [[R:%.*]], i64 12, i1 false) 851; CHECK-NEXT: ret void 852; 853 854 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 855 tail call void @llvm.memmove.p0i8.p0i8.i64(i8* %P, i8* %R, i64 12, i1 false) 856 ret void 857} 858 859define void @test37_atomic(i8* %P, i8* %Q, i8* %R) { 860; CHECK-LABEL: @test37_atomic( 861; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 862; CHECK-NEXT: tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[R:%.*]], i64 12, i32 1) 863; CHECK-NEXT: ret void 864; 865 866 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 867 tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %R, i64 12, i32 1) 868 ret void 869} 870 871; Same caveat about memcpy as in @test18 applies here. 872define void @test38(i8* %P, i8* %Q, i8* %R) { 873; CHECK-LABEL: @test38( 874; CHECK-NEXT: tail call void @llvm.memmove.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 875; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P]], i8* [[R:%.*]], i64 12, i1 false) 876; CHECK-NEXT: ret void 877; 878 879 tail call void @llvm.memmove.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 880 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %R, i64 12, i1 false) 881 ret void 882} 883 884define void @test38_atomic(i8* %P, i8* %Q, i8* %R) { 885; CHECK-LABEL: @test38_atomic( 886; CHECK-NEXT: tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 887; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[R:%.*]], i64 12, i32 1) 888; CHECK-NEXT: ret void 889; 890 891 tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 892 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %R, i64 12, i32 1) 893 ret void 894} 895 896define void @test39(i8* %P, i8* %Q, i8* %R) { 897; CHECK-LABEL: @test39( 898; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 899; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P]], i8* [[R:%.*]], i64 8, i1 false) 900; CHECK-NEXT: ret void 901; 902 903 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 904 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %R, i64 8, i1 false) 905 ret void 906} 907 908define void @test39_atomic(i8* %P, i8* %Q, i8* %R) { 909; CHECK-LABEL: @test39_atomic( 910; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P:%.*]], i8* align 1 [[Q:%.*]], i64 12, i32 1) 911; CHECK-NEXT: tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 [[P]], i8* align 1 [[R:%.*]], i64 8, i32 1) 912; CHECK-NEXT: ret void 913; 914 915 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 916 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %R, i64 8, i32 1) 917 ret void 918} 919 920define i32 @test40() { 921; CHECK-LABEL: @test40( 922; CHECK-NEXT: entry: 923; CHECK-NEXT: [[M:%.*]] = call i8* @calloc(i32 9, i32 20) 924; CHECK-NEXT: br label [[LOOP:%.*]] 925; CHECK: loop: 926; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[LOOP]] ] 927; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 928; CHECK-NEXT: [[P_NEXT:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[INDVARS_IV_NEXT]] 929; CHECK-NEXT: store i8 1, i8* [[P_NEXT]], align 1 930; CHECK-NEXT: [[P:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[INDVARS_IV]] 931; CHECK-NEXT: store i8 0, i8* [[P]], align 1 932; CHECK-NEXT: [[CONTINUE:%.*]] = icmp ugt i64 [[INDVARS_IV]], 15 933; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[RETURN:%.*]] 934; CHECK: return: 935; CHECK-NEXT: ret i32 0 936; 937entry: 938 %m = call i8* @calloc(i32 9, i32 20) 939 br label %loop 940loop: 941 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %loop ] 942 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 943 %p.next = getelementptr inbounds i8, i8* %m, i64 %indvars.iv.next 944 store i8 1, i8* %p.next 945 %p = getelementptr inbounds i8, i8* %m, i64 %indvars.iv 946 store i8 0, i8* %p 947 %continue = icmp ugt i64 %indvars.iv, 15 948 br i1 %continue, label %loop, label %return 949return: 950 ret i32 0 951} 952 953define i32 @test41() { 954; CHECK-LABEL: @test41( 955; CHECK-NEXT: entry: 956; CHECK-NEXT: [[M:%.*]] = call i8* @calloc(i32 9, i32 20) 957; CHECK-NEXT: br label [[LOOP:%.*]] 958; CHECK: loop: 959; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[CONT:%.*]] ] 960; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 961; CHECK-NEXT: [[P_NEXT:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[INDVARS_IV_NEXT]] 962; CHECK-NEXT: store i8 1, i8* [[P_NEXT]], align 1 963; CHECK-NEXT: br label [[CONT]] 964; CHECK: cont: 965; CHECK-NEXT: [[P:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[INDVARS_IV]] 966; CHECK-NEXT: store i8 0, i8* [[P]], align 1 967; CHECK-NEXT: [[CONTINUE:%.*]] = icmp ugt i64 [[INDVARS_IV]], 15 968; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[RETURN:%.*]] 969; CHECK: return: 970; CHECK-NEXT: ret i32 0 971; 972entry: 973 %m = call i8* @calloc(i32 9, i32 20) 974 br label %loop 975loop: 976 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %cont ] 977 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 978 %p.next = getelementptr inbounds i8, i8* %m, i64 %indvars.iv.next 979 store i8 1, i8* %p.next 980 br label %cont 981 982cont: 983 %p = getelementptr inbounds i8, i8* %m, i64 %indvars.iv 984 store i8 0, i8* %p 985 %continue = icmp ugt i64 %indvars.iv, 15 986 br i1 %continue, label %loop, label %return 987 988return: 989 ret i32 0 990} 991 992; The store is redundant here, but currently we fail to eliminate it. 993; We are walking from the store up to the calloc and translate phis as 994; needed. In this case we fail to translate %p while going over the 995; backedge. Because of that we conservatively assume that zero initialized 996; memory is clobbered. 997define i32 @test42() { 998; CHECK-LABEL: @test42( 999; CHECK-NEXT: entry: 1000; CHECK-NEXT: [[M:%.*]] = call i8* @calloc(i32 9, i32 20) 1001; CHECK-NEXT: br label [[LOOP:%.*]] 1002; CHECK: loop: 1003; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[CONT:%.*]] ] 1004; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 1005; CHECK-NEXT: br label [[CONT]] 1006; CHECK: cont: 1007; CHECK-NEXT: [[P:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[INDVARS_IV]] 1008; CHECK-NEXT: store i8 0, i8* [[P]], align 1 1009; CHECK-NEXT: [[CONTINUE:%.*]] = icmp ugt i64 [[INDVARS_IV]], 15 1010; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[RETURN:%.*]] 1011; CHECK: return: 1012; CHECK-NEXT: ret i32 0 1013; 1014entry: 1015 %m = call i8* @calloc(i32 9, i32 20) 1016 br label %loop 1017loop: 1018 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %cont ] 1019 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 1020 br label %cont 1021 1022cont: 1023 %p = getelementptr inbounds i8, i8* %m, i64 %indvars.iv 1024 store i8 0, i8* %p 1025 %continue = icmp ugt i64 %indvars.iv, 15 1026 br i1 %continue, label %loop, label %return 1027 1028return: 1029 ret i32 0 1030} 1031 1032define i32 @test43() { 1033; CHECK-LABEL: @test43( 1034; CHECK-NEXT: entry: 1035; CHECK-NEXT: [[M:%.*]] = call i8* @calloc(i32 9, i32 20) 1036; CHECK-NEXT: br label [[LOOP:%.*]] 1037; CHECK: loop: 1038; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[CONT_2:%.*]] ] 1039; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 1040; CHECK-NEXT: [[P_NEXT:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[INDVARS_IV_NEXT]] 1041; CHECK-NEXT: store i8 1, i8* [[P_NEXT]], align 1 1042; CHECK-NEXT: br label [[CONT:%.*]] 1043; CHECK: cont: 1044; CHECK-NEXT: [[P:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[INDVARS_IV]] 1045; CHECK-NEXT: store i8 0, i8* [[P]], align 1 1046; CHECK-NEXT: br label [[CONT_2]] 1047; CHECK: cont.2: 1048; CHECK-NEXT: [[CONTINUE:%.*]] = icmp ugt i64 [[INDVARS_IV]], 15 1049; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[RETURN:%.*]] 1050; CHECK: return: 1051; CHECK-NEXT: ret i32 0 1052; 1053entry: 1054 %m = call i8* @calloc(i32 9, i32 20) 1055 br label %loop 1056loop: 1057 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %cont.2 ] 1058 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 1059 %p.next = getelementptr inbounds i8, i8* %m, i64 %indvars.iv.next 1060 store i8 1, i8* %p.next 1061 br label %cont 1062 1063cont: 1064 %p = getelementptr inbounds i8, i8* %m, i64 %indvars.iv 1065 store i8 0, i8* %p 1066 br label %cont.2 1067 1068cont.2: 1069 %continue = icmp ugt i64 %indvars.iv, 15 1070 br i1 %continue, label %loop, label %return 1071 1072return: 1073 ret i32 0 1074} 1075 1076define i32 @test44() { 1077; CHECK-LABEL: @test44( 1078; CHECK-NEXT: entry: 1079; CHECK-NEXT: [[M:%.*]] = call i8* @calloc(i32 9, i32 20) 1080; CHECK-NEXT: br label [[LOOP:%.*]] 1081; CHECK: loop: 1082; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[CONT_2:%.*]] ] 1083; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 1084; CHECK-NEXT: [[P_NEXT:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[INDVARS_IV_NEXT]] 1085; CHECK-NEXT: [[P:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[INDVARS_IV]] 1086; CHECK-NEXT: store i8 0, i8* [[P]], align 1 1087; CHECK-NEXT: br label [[CONT:%.*]] 1088; CHECK: cont: 1089; CHECK-NEXT: store i8 1, i8* [[P_NEXT]], align 1 1090; CHECK-NEXT: br label [[CONT_2]] 1091; CHECK: cont.2: 1092; CHECK-NEXT: [[CONTINUE:%.*]] = icmp ugt i64 [[INDVARS_IV]], 15 1093; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[RETURN:%.*]] 1094; CHECK: return: 1095; CHECK-NEXT: ret i32 0 1096; 1097entry: 1098 %m = call i8* @calloc(i32 9, i32 20) 1099 br label %loop 1100loop: 1101 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %cont.2 ] 1102 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 1103 %p.next = getelementptr inbounds i8, i8* %m, i64 %indvars.iv.next 1104 %p = getelementptr inbounds i8, i8* %m, i64 %indvars.iv 1105 store i8 0, i8* %p 1106 br label %cont 1107 1108cont: 1109 store i8 1, i8* %p.next 1110 br label %cont.2 1111 1112cont.2: 1113 %continue = icmp ugt i64 %indvars.iv, 15 1114 br i1 %continue, label %loop, label %return 1115 1116return: 1117 ret i32 0 1118} 1119 1120; This is an example which can potentially benefit from PHI translation. 1121; Current implementation doesn't handle this case though. This is because 1122; we don't visit the same block with different addresses while looking for 1123; clobbering instructions. 1124define i32 @test45(i1 %c) { 1125; CHECK-LABEL: @test45( 1126; CHECK-NEXT: entry: 1127; CHECK-NEXT: [[M:%.*]] = call i8* @calloc(i32 9, i32 20) 1128; CHECK-NEXT: br i1 [[C:%.*]], label [[TRUE:%.*]], label [[FALSE:%.*]] 1129; CHECK: true: 1130; CHECK-NEXT: [[P_1:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 1 1131; CHECK-NEXT: store i8 1, i8* [[P_1]], align 1 1132; CHECK-NEXT: br label [[CONT:%.*]] 1133; CHECK: false: 1134; CHECK-NEXT: [[P_2:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 2 1135; CHECK-NEXT: store i8 1, i8* [[P_2]], align 1 1136; CHECK-NEXT: br label [[CONT]] 1137; CHECK: cont: 1138; CHECK-NEXT: [[OFFSET:%.*]] = phi i64 [ 2, [[TRUE]] ], [ 1, [[FALSE]] ] 1139; CHECK-NEXT: [[P:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 [[OFFSET]] 1140; CHECK-NEXT: store i8 0, i8* [[P]], align 1 1141; CHECK-NEXT: br label [[RETURN:%.*]] 1142; CHECK: return: 1143; CHECK-NEXT: ret i32 0 1144; 1145entry: 1146 %m = call i8* @calloc(i32 9, i32 20) 1147 br i1 %c, label %true, label %false 1148 1149true: 1150 %p.1 = getelementptr inbounds i8, i8* %m, i64 1 1151 store i8 1, i8* %p.1 1152 br label %cont 1153 1154false: 1155 %p.2 = getelementptr inbounds i8, i8* %m, i64 2 1156 store i8 1, i8* %p.2 1157 br label %cont 1158 1159cont: 1160 %offset = phi i64 [ 2, %true ], [ 1, %false ] 1161 %p = getelementptr inbounds i8, i8* %m, i64 %offset 1162 store i8 0, i8* %p 1163 br label %return 1164 1165return: 1166 ret i32 0 1167} 1168 1169; This is test45 modified in a way to demonstrate PHI translation 1170; improving the accuracy of the analysis (on a slightly convoluted 1171; case though). 1172define i32 @test46(i1 %c) { 1173; CHECK-LABEL: @test46( 1174; CHECK-NEXT: entry: 1175; CHECK-NEXT: [[M:%.*]] = call i8* @calloc(i32 9, i32 20) 1176; CHECK-NEXT: [[P_1:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 1 1177; CHECK-NEXT: [[P_2:%.*]] = getelementptr inbounds i8, i8* [[M]], i64 2 1178; CHECK-NEXT: br i1 [[C:%.*]], label [[TRUE:%.*]], label [[FALSE:%.*]] 1179; CHECK: true: 1180; CHECK-NEXT: store i8 1, i8* [[P_1]], align 1 1181; CHECK-NEXT: br label [[CONT:%.*]] 1182; CHECK: false: 1183; CHECK-NEXT: store i8 1, i8* [[P_1]], align 1 1184; CHECK-NEXT: br label [[CONT]] 1185; CHECK: cont: 1186; CHECK-NEXT: br label [[RETURN:%.*]] 1187; CHECK: return: 1188; CHECK-NEXT: ret i32 0 1189; 1190entry: 1191 %m = call i8* @calloc(i32 9, i32 20) 1192 %p.1 = getelementptr inbounds i8, i8* %m, i64 1 1193 %p.2 = getelementptr inbounds i8, i8* %m, i64 2 1194 br i1 %c, label %true, label %false 1195 1196true: 1197 store i8 1, i8* %p.1 1198 br label %cont 1199 1200false: 1201 store i8 1, i8* %p.1 1202 br label %cont 1203 1204cont: 1205 %offset = phi i64 [ 2, %true ], [ 2, %false ] 1206 %p = getelementptr inbounds i8, i8* %m, i64 %offset 1207 store i8 0, i8* %p 1208 br label %return 1209 1210return: 1211 ret i32 0 1212} 1213 1214declare void @llvm.memmove.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i1) 1215declare void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i32) 1216