1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -basic-aa -dse -S | FileCheck %s 3; RUN: opt < %s -aa-pipeline=basic-aa -passes=dse -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*) 11declare void @llvm.matrix.column.major.store(<6 x float>, float*, i64, i1, i32, i32) 12 13define void @test1(i32* %Q, i32* %P) { 14; CHECK-LABEL: @test1( 15; CHECK-NEXT: store i32 0, i32* [[P:%.*]], align 4 16; CHECK-NEXT: ret void 17; 18 %DEAD = load i32, i32* %Q 19 store i32 %DEAD, i32* %P 20 store i32 0, i32* %P 21 ret void 22} 23 24; PR8677 25@g = global i32 1 26 27define i32 @test3(i32* %g_addr) nounwind { 28; CHECK-LABEL: @test3( 29; CHECK-NEXT: [[G_VALUE:%.*]] = load i32, i32* [[G_ADDR:%.*]], align 4 30; CHECK-NEXT: store i32 -1, i32* @g, align 4 31; CHECK-NEXT: store i32 [[G_VALUE]], i32* [[G_ADDR]], align 4 32; CHECK-NEXT: [[TMP3:%.*]] = load i32, i32* @g, align 4 33; CHECK-NEXT: ret i32 [[TMP3]] 34; 35 %g_value = load i32, i32* %g_addr, align 4 36 store i32 -1, i32* @g, align 4 37 store i32 %g_value, i32* %g_addr, align 4 38 %tmp3 = load i32, i32* @g, align 4 39 ret i32 %tmp3 40} 41 42 43define void @test4(i32* %Q) { 44; CHECK-LABEL: @test4( 45; CHECK-NEXT: [[A:%.*]] = load i32, i32* [[Q:%.*]], align 4 46; CHECK-NEXT: store volatile i32 [[A]], i32* [[Q]], align 4 47; CHECK-NEXT: ret void 48; 49 %a = load i32, i32* %Q 50 store volatile i32 %a, i32* %Q 51 ret void 52} 53 54; PR8576 - Should delete store of 10 even though p/q are may aliases. 55define void @test2(i32 *%p, i32 *%q) { 56; CHECK-LABEL: @test2( 57; CHECK-NEXT: store i32 20, i32* [[Q:%.*]], align 4 58; CHECK-NEXT: store i32 30, i32* [[P:%.*]], align 4 59; CHECK-NEXT: ret void 60; 61 store i32 10, i32* %p, align 4 62 store i32 20, i32* %q, align 4 63 store i32 30, i32* %p, align 4 64 ret void 65} 66 67; Should delete store of 10 even though memset is a may-store to P (P and Q may 68; alias). 69define void @test6(i32 *%p, i8 *%q) { 70; CHECK-LABEL: @test6( 71; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[Q:%.*]], i8 42, i64 900, i1 false) 72; CHECK-NEXT: store i32 30, i32* [[P:%.*]], align 4 73; CHECK-NEXT: ret void 74; 75 store i32 10, i32* %p, align 4 ;; dead. 76 call void @llvm.memset.p0i8.i64(i8* %q, i8 42, i64 900, i1 false) 77 store i32 30, i32* %p, align 4 78 ret void 79} 80 81; Should delete store of 10 even though memset is a may-store to P (P and Q may 82; alias). 83define void @test6_atomic(i32* align 4 %p, i8* align 4 %q) { 84; CHECK-LABEL: @test6_atomic( 85; CHECK-NEXT: call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 4 [[Q:%.*]], i8 42, i64 900, i32 4) 86; CHECK-NEXT: store atomic i32 30, i32* [[P:%.*]] unordered, align 4 87; CHECK-NEXT: ret void 88; 89 store atomic i32 10, i32* %p unordered, align 4 ;; dead. 90 call void @llvm.memset.element.unordered.atomic.p0i8.i64(i8* align 4 %q, i8 42, i64 900, i32 4) 91 store atomic i32 30, i32* %p unordered, align 4 92 ret void 93} 94 95; Should delete store of 10 even though memcpy is a may-store to P (P and Q may 96; alias). 97define void @test7(i32 *%p, i8 *%q, i8* noalias %r) { 98; CHECK-LABEL: @test7( 99; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[Q:%.*]], i8* [[R:%.*]], i64 900, i1 false) 100; CHECK-NEXT: store i32 30, i32* [[P:%.*]], align 4 101; CHECK-NEXT: ret void 102; 103 store i32 10, i32* %p, align 4 ;; dead. 104 call void @llvm.memcpy.p0i8.p0i8.i64(i8* %q, i8* %r, i64 900, i1 false) 105 store i32 30, i32* %p, align 4 106 ret void 107} 108 109; Should delete store of 10 even though memcpy is a may-store to P (P and Q may 110; alias). 111define void @test7_atomic(i32* align 4 %p, i8* align 4 %q, i8* noalias align 4 %r) { 112; CHECK-LABEL: @test7_atomic( 113; CHECK-NEXT: call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 4 [[Q:%.*]], i8* align 4 [[R:%.*]], i64 900, i32 4) 114; CHECK-NEXT: store atomic i32 30, i32* [[P:%.*]] unordered, align 4 115; CHECK-NEXT: ret void 116; 117 store atomic i32 10, i32* %p unordered, align 4 ;; dead. 118 call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 4 %q, i8* align 4 %r, i64 900, i32 4) 119 store atomic i32 30, i32* %p unordered, align 4 120 ret void 121} 122 123; Do not delete stores that are only partially killed. 124define i32 @test8() { 125; CHECK-LABEL: @test8( 126; CHECK-NEXT: [[V:%.*]] = alloca i32, align 4 127; CHECK-NEXT: store i32 1234567, i32* [[V]], align 4 128; CHECK-NEXT: [[X:%.*]] = load i32, i32* [[V]], align 4 129; CHECK-NEXT: ret i32 [[X]] 130; 131 %V = alloca i32 132 store i32 1234567, i32* %V 133 %V2 = bitcast i32* %V to i8* 134 store i8 0, i8* %V2 135 %X = load i32, i32* %V 136 ret i32 %X 137 138} 139 140; Test for byval handling. 141%struct.x = type { i32, i32, i32, i32 } 142define void @test9(%struct.x* byval(%struct.x) %a) nounwind { 143; CHECK-LABEL: @test9( 144; CHECK-NEXT: ret void 145; 146 %tmp2 = getelementptr %struct.x, %struct.x* %a, i32 0, i32 0 147 store i32 1, i32* %tmp2, align 4 148 ret void 149} 150 151; Test for inalloca handling. 152define void @test9_2(%struct.x* inalloca(%struct.x) %a) nounwind { 153; CHECK-LABEL: @test9_2( 154; CHECK-NEXT: ret void 155; 156 %tmp2 = getelementptr %struct.x, %struct.x* %a, i32 0, i32 0 157 store i32 1, i32* %tmp2, align 4 158 ret void 159} 160 161; Test for preallocated handling. 162define void @test9_3(%struct.x* preallocated(%struct.x) %a) nounwind { 163; CHECK-LABEL: @test9_3( 164; CHECK-NEXT: ret void 165; 166 %tmp2 = getelementptr %struct.x, %struct.x* %a, i32 0, i32 0 167 store i32 1, i32* %tmp2, align 4 168 ret void 169} 170 171; va_arg has fuzzy dependence, the store shouldn't be zapped. 172define double @test10(i8* %X) { 173; CHECK-LABEL: @test10( 174; CHECK-NEXT: [[X_ADDR:%.*]] = alloca i8*, align 8 175; CHECK-NEXT: store i8* [[X:%.*]], i8** [[X_ADDR]], align 8 176; CHECK-NEXT: [[TMP_0:%.*]] = va_arg i8** [[X_ADDR]], double 177; CHECK-NEXT: ret double [[TMP_0]] 178; 179 %X_addr = alloca i8* 180 store i8* %X, i8** %X_addr 181 %tmp.0 = va_arg i8** %X_addr, double 182 ret double %tmp.0 183} 184 185; DSE should delete the dead trampoline. 186declare void @test11f() 187define void @test11() { 188; CHECK-LABEL: @test11( 189; CHECK-NEXT: ret void 190; 191 %storage = alloca [10 x i8], align 16 ; <[10 x i8]*> [#uses=1] 192 %cast = getelementptr [10 x i8], [10 x i8]* %storage, i32 0, i32 0 ; <i8*> [#uses=1] 193 call void @llvm.init.trampoline( i8* %cast, i8* bitcast (void ()* @test11f to i8*), i8* null ) ; <i8*> [#uses=1] 194 ret void 195} 196 197; Specialized store intrinsics should be removed if dead. 198define void @test_matrix_store(i64 %stride) { 199; CHECK-LABEL: @test_matrix_store( 200; CHECK-NEXT: ret void 201; 202 %a = alloca [6 x float] 203 %cast = bitcast [6 x float]* %a to float* 204 call void @llvm.matrix.column.major.store(<6 x float> zeroinitializer, float* %cast, i64 %stride, i1 false, i32 3, i32 2) 205 ret void 206} 207 208; %P doesn't escape, the DEAD instructions should be removed. 209declare void @may_unwind() 210define i32* @test_malloc_no_escape_before_return() { 211; CHECK-LABEL: @test_malloc_no_escape_before_return( 212; CHECK-NEXT: [[PTR:%.*]] = tail call i8* @malloc(i32 4) 213; CHECK-NEXT: [[P:%.*]] = bitcast i8* [[PTR]] to i32* 214; CHECK-NEXT: call void @may_unwind() 215; CHECK-NEXT: store i32 0, i32* [[P]], align 4 216; CHECK-NEXT: ret i32* [[P]] 217; 218 %ptr = tail call i8* @malloc(i32 4) 219 %P = bitcast i8* %ptr to i32* 220 %DEAD = load i32, i32* %P 221 %DEAD2 = add i32 %DEAD, 1 222 store i32 %DEAD2, i32* %P 223 call void @may_unwind() 224 store i32 0, i32* %P 225 ret i32* %P 226} 227 228define i32* @test_custom_malloc_no_escape_before_return() { 229; CHECK-LABEL: @test_custom_malloc_no_escape_before_return( 230; CHECK-NEXT: [[PTR:%.*]] = tail call i8* @custom_malloc(i32 4) 231; CHECK-NEXT: [[P:%.*]] = bitcast i8* [[PTR]] to i32* 232; CHECK-NEXT: call void @may_unwind() 233; CHECK-NEXT: store i32 0, i32* [[P]], align 4 234; CHECK-NEXT: ret i32* [[P]] 235; 236 %ptr = tail call i8* @custom_malloc(i32 4) 237 %P = bitcast i8* %ptr to i32* 238 %DEAD = load i32, i32* %P 239 %DEAD2 = add i32 %DEAD, 1 240 store i32 %DEAD2, i32* %P 241 call void @may_unwind() 242 store i32 0, i32* %P 243 ret i32* %P 244} 245 246define i32 addrspace(1)* @test13_addrspacecast() { 247; CHECK-LABEL: @test13_addrspacecast( 248; CHECK-NEXT: [[P:%.*]] = tail call i8* @malloc(i32 4) 249; CHECK-NEXT: [[P_BC:%.*]] = bitcast i8* [[P]] to i32* 250; CHECK-NEXT: [[P:%.*]] = addrspacecast i32* [[P_BC]] to i32 addrspace(1)* 251; CHECK-NEXT: call void @may_unwind() 252; CHECK-NEXT: store i32 0, i32 addrspace(1)* [[P]], align 4 253; CHECK-NEXT: ret i32 addrspace(1)* [[P]] 254; 255 %p = tail call i8* @malloc(i32 4) 256 %p.bc = bitcast i8* %p to i32* 257 %P = addrspacecast i32* %p.bc to i32 addrspace(1)* 258 %DEAD = load i32, i32 addrspace(1)* %P 259 %DEAD2 = add i32 %DEAD, 1 260 store i32 %DEAD2, i32 addrspace(1)* %P 261 call void @may_unwind() 262 store i32 0, i32 addrspace(1)* %P 263 ret i32 addrspace(1)* %P 264} 265 266 267declare noalias i8* @malloc(i32) willreturn 268declare noalias i8* @custom_malloc(i32) willreturn 269declare noalias i8* @calloc(i32, i32) willreturn 270 271define void @test14(i32* %Q) { 272; CHECK-LABEL: @test14( 273; CHECK-NEXT: ret void 274; 275 %P = alloca i32 276 %DEAD = load i32, i32* %Q 277 store i32 %DEAD, i32* %P 278 ret void 279 280} 281 282; The store here is not dead because the byval call reads it. 283declare void @test19f({i32}* byval({i32}) align 4 %P) 284 285define void @test19({i32}* nocapture byval({i32}) align 4 %arg5) nounwind ssp { 286; CHECK-LABEL: @test19( 287; CHECK-NEXT: bb: 288; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds { i32 }, { i32 }* [[ARG5:%.*]], i32 0, i32 0 289; CHECK-NEXT: store i32 912, i32* [[TMP7]], align 4 290; CHECK-NEXT: call void @test19f({ i32 }* byval({ i32 }) align 4 [[ARG5]]) 291; CHECK-NEXT: ret void 292; 293bb: 294 %tmp7 = getelementptr inbounds {i32}, {i32}* %arg5, i32 0, i32 0 295 store i32 912, i32* %tmp7 296 call void @test19f({i32}* byval({i32}) align 4 %arg5) 297 ret void 298 299} 300 301define void @malloc_no_escape() { 302; CHECK-LABEL: @malloc_no_escape( 303; CHECK-NEXT: ret void 304; 305 %m = call i8* @malloc(i32 24) 306 store i8 0, i8* %m 307 ret void 308} 309 310define void @custom_malloc_no_escape() { 311; CHECK-LABEL: @custom_malloc_no_escape( 312; CHECK-NEXT: [[M:%.*]] = call i8* @custom_malloc(i32 24) 313; CHECK-NEXT: ret void 314; 315 %m = call i8* @custom_malloc(i32 24) 316 store i8 0, i8* %m 317 ret void 318} 319 320define void @test21() { 321; CHECK-LABEL: @test21( 322; CHECK-NEXT: ret void 323; 324 %m = call i8* @calloc(i32 9, i32 7) 325 store i8 0, i8* %m 326 ret void 327} 328 329; Currently elimination of stores at the end of a function is limited to a 330; single underlying object, for compile-time. This case appears to not be 331; very important in practice. 332define void @test22(i1 %i, i32 %k, i32 %m) nounwind { 333; CHECK-LABEL: @test22( 334; CHECK-NEXT: [[K_ADDR:%.*]] = alloca i32, align 4 335; CHECK-NEXT: [[M_ADDR:%.*]] = alloca i32, align 4 336; CHECK-NEXT: [[K_ADDR_M_ADDR:%.*]] = select i1 [[I:%.*]], i32* [[K_ADDR]], i32* [[M_ADDR]] 337; CHECK-NEXT: store i32 0, i32* [[K_ADDR_M_ADDR]], align 4 338; CHECK-NEXT: ret void 339; 340 %k.addr = alloca i32 341 %m.addr = alloca i32 342 %k.addr.m.addr = select i1 %i, i32* %k.addr, i32* %m.addr 343 store i32 0, i32* %k.addr.m.addr, align 4 344 ret void 345} 346 347; PR13547 348declare noalias i8* @strdup(i8* nocapture) nounwind 349define noalias i8* @test23() nounwind uwtable ssp { 350; CHECK-LABEL: @test23( 351; CHECK-NEXT: [[X:%.*]] = alloca [2 x i8], align 1 352; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [2 x i8], [2 x i8]* [[X]], i64 0, i64 0 353; CHECK-NEXT: store i8 97, i8* [[ARRAYIDX]], align 1 354; CHECK-NEXT: [[ARRAYIDX1:%.*]] = getelementptr inbounds [2 x i8], [2 x i8]* [[X]], i64 0, i64 1 355; CHECK-NEXT: store i8 0, i8* [[ARRAYIDX1]], align 1 356; CHECK-NEXT: [[CALL:%.*]] = call i8* @strdup(i8* [[ARRAYIDX]]) #[[ATTR5:[0-9]+]] 357; CHECK-NEXT: ret i8* [[CALL]] 358; 359 %x = alloca [2 x i8], align 1 360 %arrayidx = getelementptr inbounds [2 x i8], [2 x i8]* %x, i64 0, i64 0 361 store i8 97, i8* %arrayidx, align 1 362 %arrayidx1 = getelementptr inbounds [2 x i8], [2 x i8]* %x, i64 0, i64 1 363 store i8 0, i8* %arrayidx1, align 1 364 %call = call i8* @strdup(i8* %arrayidx) nounwind 365 ret i8* %call 366} 367 368; Make sure same sized store to later element is deleted 369define void @test24([2 x i32]* %a, i32 %b, i32 %c) nounwind { 370; CHECK-LABEL: @test24( 371; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds [2 x i32], [2 x i32]* [[A:%.*]], i64 0, i64 0 372; CHECK-NEXT: store i32 [[B:%.*]], i32* [[TMP1]], align 4 373; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds [2 x i32], [2 x i32]* [[A]], i64 0, i64 1 374; CHECK-NEXT: store i32 [[C:%.*]], i32* [[TMP2]], align 4 375; CHECK-NEXT: ret void 376; 377 %1 = getelementptr inbounds [2 x i32], [2 x i32]* %a, i64 0, i64 0 378 store i32 0, i32* %1, align 4 379 %2 = getelementptr inbounds [2 x i32], [2 x i32]* %a, i64 0, i64 1 380 store i32 0, i32* %2, align 4 381 %3 = getelementptr inbounds [2 x i32], [2 x i32]* %a, i64 0, i64 0 382 store i32 %b, i32* %3, align 4 383 %4 = getelementptr inbounds [2 x i32], [2 x i32]* %a, i64 0, i64 1 384 store i32 %c, i32* %4, align 4 385 ret void 386} 387 388; Check another case like PR13547 where strdup is not like malloc. 389define i8* @test25(i8* %p) nounwind { 390; CHECK-LABEL: @test25( 391; CHECK-NEXT: [[P_4:%.*]] = getelementptr i8, i8* [[P:%.*]], i64 4 392; CHECK-NEXT: [[TMP:%.*]] = load i8, i8* [[P_4]], align 1 393; CHECK-NEXT: store i8 0, i8* [[P_4]], align 1 394; CHECK-NEXT: [[Q:%.*]] = call i8* @strdup(i8* [[P]]) #[[ATTR10:[0-9]+]] 395; CHECK-NEXT: store i8 [[TMP]], i8* [[P_4]], align 1 396; CHECK-NEXT: ret i8* [[Q]] 397; 398 %p.4 = getelementptr i8, i8* %p, i64 4 399 %tmp = load i8, i8* %p.4, align 1 400 store i8 0, i8* %p.4, align 1 401 %q = call i8* @strdup(i8* %p) nounwind optsize 402 store i8 %tmp, i8* %p.4, align 1 403 ret i8* %q 404} 405 406; Don't remove redundant store because of may-aliased store. 407define i32 @test28(i1 %c, i32* %p, i32* %p2, i32 %i) { 408; CHECK-LABEL: @test28( 409; CHECK-NEXT: entry: 410; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[P:%.*]], align 4 411; CHECK-NEXT: store i32 [[I:%.*]], i32* [[P2:%.*]], align 4 412; CHECK-NEXT: br i1 [[C:%.*]], label [[BB1:%.*]], label [[BB2:%.*]] 413; CHECK: bb1: 414; CHECK-NEXT: br label [[BB3:%.*]] 415; CHECK: bb2: 416; CHECK-NEXT: br label [[BB3]] 417; CHECK: bb3: 418; CHECK-NEXT: store i32 [[V]], i32* [[P]], align 4 419; CHECK-NEXT: ret i32 0 420; 421entry: 422 %v = load i32, i32* %p, align 4 423 424 ; Might overwrite value at %p 425 store i32 %i, i32* %p2, align 4 426 br i1 %c, label %bb1, label %bb2 427bb1: 428 br label %bb3 429bb2: 430 br label %bb3 431bb3: 432 store i32 %v, i32* %p, align 4 433 ret i32 0 434} 435 436; Don't remove redundant store because of may-aliased store. 437define i32 @test29(i1 %c, i32* %p, i32* %p2, i32 %i) { 438; CHECK-LABEL: @test29( 439; CHECK-NEXT: entry: 440; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[P:%.*]], align 4 441; CHECK-NEXT: br i1 [[C:%.*]], label [[BB1:%.*]], label [[BB2:%.*]] 442; CHECK: bb1: 443; CHECK-NEXT: br label [[BB3:%.*]] 444; CHECK: bb2: 445; CHECK-NEXT: store i32 [[I:%.*]], i32* [[P2:%.*]], align 4 446; CHECK-NEXT: br label [[BB3]] 447; CHECK: bb3: 448; CHECK-NEXT: store i32 [[V]], i32* [[P]], align 4 449; CHECK-NEXT: ret i32 0 450; 451entry: 452 %v = load i32, i32* %p, align 4 453 br i1 %c, label %bb1, label %bb2 454bb1: 455 br label %bb3 456bb2: 457 ; Might overwrite value at %p 458 store i32 %i, i32* %p2, align 4 459 br label %bb3 460bb3: 461 store i32 %v, i32* %p, align 4 462 ret i32 0 463} 464 465declare void @unknown_func() 466 467; Don't remove redundant store because of unknown call. 468define i32 @test30(i1 %c, i32* %p, i32 %i) { 469; CHECK-LABEL: @test30( 470; CHECK-NEXT: entry: 471; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[P:%.*]], align 4 472; CHECK-NEXT: br i1 [[C:%.*]], label [[BB1:%.*]], label [[BB2:%.*]] 473; CHECK: bb1: 474; CHECK-NEXT: br label [[BB3:%.*]] 475; CHECK: bb2: 476; CHECK-NEXT: call void @unknown_func() 477; CHECK-NEXT: br label [[BB3]] 478; CHECK: bb3: 479; CHECK-NEXT: store i32 [[V]], i32* [[P]], align 4 480; CHECK-NEXT: ret i32 0 481; 482entry: 483 %v = load i32, i32* %p, align 4 484 br i1 %c, label %bb1, label %bb2 485bb1: 486 br label %bb3 487bb2: 488 ; Might overwrite value at %p 489 call void @unknown_func() 490 br label %bb3 491bb3: 492 store i32 %v, i32* %p, align 4 493 ret i32 0 494} 495 496; Don't remove redundant store in a loop with a may-alias store. 497define i32 @test32(i1 %c, i32* %p, i32 %i) { 498; CHECK-LABEL: @test32( 499; CHECK-NEXT: entry: 500; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[P:%.*]], align 4 501; CHECK-NEXT: br label [[BB1:%.*]] 502; CHECK: bb1: 503; CHECK-NEXT: store i32 [[V]], i32* [[P]], align 4 504; CHECK-NEXT: call void @unknown_func() 505; CHECK-NEXT: br i1 undef, label [[BB1]], label [[BB2:%.*]] 506; CHECK: bb2: 507; CHECK-NEXT: ret i32 0 508; 509entry: 510 %v = load i32, i32* %p, align 4 511 br label %bb1 512bb1: 513 store i32 %v, i32* %p, align 4 514 ; Might read and overwrite value at %p 515 call void @unknown_func() 516 br i1 undef, label %bb1, label %bb2 517bb2: 518 ret i32 0 519} 520 521; We cannot remove any stores, because @unknown_func may unwind and the caller 522; may read %p while unwinding. 523define void @test34(i32* noalias %p) { 524; CHECK-LABEL: @test34( 525; CHECK-NEXT: store i32 1, i32* [[P:%.*]], align 4 526; CHECK-NEXT: call void @unknown_func() 527; CHECK-NEXT: store i32 0, i32* [[P]], align 4 528; CHECK-NEXT: ret void 529; 530 store i32 1, i32* %p 531 call void @unknown_func() 532 store i32 0, i32* %p 533 ret void 534} 535 536; Remove redundant store even with an unwinding function in the same block 537define void @test35(i32* noalias %p) { 538; CHECK-LABEL: @test35( 539; CHECK-NEXT: call void @unknown_func() 540; CHECK-NEXT: store i32 0, i32* [[P:%.*]], align 4 541; CHECK-NEXT: ret void 542; 543 call void @unknown_func() 544 store i32 1, i32* %p 545 store i32 0, i32* %p 546 ret void 547} 548 549; We cannot optimize away the first memmove since %P could overlap with %Q. 550define void @test36(i8* %P, i8* %Q) { 551; CHECK-LABEL: @test36( 552; CHECK-NEXT: tail call void @llvm.memmove.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 553; CHECK-NEXT: tail call void @llvm.memmove.p0i8.p0i8.i64(i8* [[P]], i8* [[Q]], i64 12, i1 false) 554; CHECK-NEXT: ret void 555; 556 557 tail call void @llvm.memmove.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 558 tail call void @llvm.memmove.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 559 ret void 560} 561 562define void @test36_atomic(i8* %P, i8* %Q) { 563; CHECK-LABEL: @test36_atomic( 564; 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) 565; 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) 566; CHECK-NEXT: ret void 567; 568 569 tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 570 tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 571 ret void 572} 573 574define void @test37(i8* %P, i8* %Q, i8* %R) { 575; CHECK-LABEL: @test37( 576; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 577; CHECK-NEXT: tail call void @llvm.memmove.p0i8.p0i8.i64(i8* [[P]], i8* [[R:%.*]], i64 12, i1 false) 578; CHECK-NEXT: ret void 579; 580 581 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 582 tail call void @llvm.memmove.p0i8.p0i8.i64(i8* %P, i8* %R, i64 12, i1 false) 583 ret void 584} 585 586define void @test37_atomic(i8* %P, i8* %Q, i8* %R) { 587; CHECK-LABEL: @test37_atomic( 588; 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) 589; 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) 590; CHECK-NEXT: ret void 591; 592 593 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 594 tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %R, i64 12, i32 1) 595 ret void 596} 597 598; See PR11763 - LLVM allows memcpy's source and destination to be equal (but not 599; inequal and overlapping). 600define void @test38(i8* %P, i8* %Q, i8* %R) { 601; CHECK-LABEL: @test38( 602; CHECK-NEXT: tail call void @llvm.memmove.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 603; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P]], i8* [[R:%.*]], i64 12, i1 false) 604; CHECK-NEXT: ret void 605; 606 607 tail call void @llvm.memmove.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 608 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %R, i64 12, i1 false) 609 ret void 610} 611 612; See PR11763 - LLVM allows memcpy's source and destination to be equal (but not 613; inequal and overlapping). 614define void @test38_atomic(i8* %P, i8* %Q, i8* %R) { 615; CHECK-LABEL: @test38_atomic( 616; 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) 617; 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) 618; CHECK-NEXT: ret void 619; 620 621 tail call void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 622 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %R, i64 12, i32 1) 623 ret void 624} 625 626define void @test39(i8* %P, i8* %Q, i8* %R) { 627; CHECK-LABEL: @test39( 628; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P:%.*]], i8* [[Q:%.*]], i64 12, i1 false) 629; CHECK-NEXT: tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[P]], i8* [[R:%.*]], i64 8, i1 false) 630; CHECK-NEXT: ret void 631; 632 633 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %Q, i64 12, i1 false) 634 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %P, i8* %R, i64 8, i1 false) 635 ret void 636} 637 638define void @test39_atomic(i8* %P, i8* %Q, i8* %R) { 639; CHECK-LABEL: @test39_atomic( 640; 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) 641; 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) 642; CHECK-NEXT: ret void 643; 644 645 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %Q, i64 12, i32 1) 646 tail call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* align 1 %P, i8* align 1 %R, i64 8, i32 1) 647 ret void 648} 649 650declare void @llvm.memmove.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i1) 651declare void @llvm.memmove.element.unordered.atomic.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i32) 652 653declare void @llvm.lifetime.start.p0i8(i64, i8* nocapture) nounwind 654declare void @llvm.lifetime.end.p0i8(i64, i8* nocapture) nounwind 655define void @test40(i32** noalias %Pp, i32* noalias %Q) { 656; CHECK-LABEL: @test40( 657; CHECK-NEXT: entry: 658; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 659; CHECK-NEXT: [[AC:%.*]] = bitcast i32* [[A]] to i8* 660; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull [[AC]]) 661; CHECK-NEXT: [[TMP0:%.*]] = bitcast i32** [[PP:%.*]] to i8** 662; CHECK-NEXT: [[PC:%.*]] = load i8*, i8** [[TMP0]], align 8 663; CHECK-NEXT: [[QC:%.*]] = bitcast i32* [[Q:%.*]] to i8* 664; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* nonnull align 4 [[AC]], i8* align 4 [[QC]], i64 4, i1 false) 665; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[PC]], i8* nonnull align 4 [[AC]], i64 4, i1 true) 666; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull [[AC]]) 667; CHECK-NEXT: ret void 668; 669entry: 670 %A = alloca i32, align 4 671 %Ac = bitcast i32* %A to i8* 672 call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %Ac) 673 %0 = bitcast i32** %Pp to i8** 674 %Pc = load i8*, i8** %0, align 8 675 %Qc = bitcast i32* %Q to i8* 676 call void @llvm.memcpy.p0i8.p0i8.i64(i8* nonnull align 4 %Ac, i8* align 4 %Qc, i64 4, i1 false) 677 call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %Pc, i8* nonnull align 4 %Ac, i64 4, i1 true) 678 call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %Ac) 679 ret void 680} 681 682declare void @free(i8* nocapture) 683 684; We cannot remove `store i32 1, i32* %p`, because @unknown_func may unwind 685; and the caller may read %p while unwinding. 686define void @test41(i32* noalias %P) { 687; CHECK-LABEL: @test41( 688; CHECK-NEXT: [[P2:%.*]] = bitcast i32* [[P:%.*]] to i8* 689; CHECK-NEXT: store i32 1, i32* [[P]], align 4 690; CHECK-NEXT: call void @unknown_func() 691; CHECK-NEXT: call void @free(i8* [[P2]]) 692; CHECK-NEXT: ret void 693; 694 %P2 = bitcast i32* %P to i8* 695 store i32 1, i32* %P 696 call void @unknown_func() 697 store i32 2, i32* %P 698 call void @free(i8* %P2) 699 ret void 700} 701 702define void @test42(i32* %P, i32* %Q) { 703; CHECK-LABEL: @test42( 704; CHECK-NEXT: store i32 1, i32* [[P:%.*]], align 4 705; CHECK-NEXT: [[P2:%.*]] = bitcast i32* [[P]] to i8* 706; CHECK-NEXT: store i32 2, i32* [[Q:%.*]], align 4 707; CHECK-NEXT: store i8 3, i8* [[P2]], align 1 708; CHECK-NEXT: ret void 709; 710 store i32 1, i32* %P 711 %P2 = bitcast i32* %P to i8* 712 store i32 2, i32* %Q 713 store i8 3, i8* %P2 714 ret void 715} 716 717define void @test42a(i32* %P, i32* %Q) { 718; CHECK-LABEL: @test42a( 719; CHECK-NEXT: store atomic i32 1, i32* [[P:%.*]] unordered, align 4 720; CHECK-NEXT: [[P2:%.*]] = bitcast i32* [[P]] to i8* 721; CHECK-NEXT: store atomic i32 2, i32* [[Q:%.*]] unordered, align 4 722; CHECK-NEXT: store atomic i8 3, i8* [[P2]] unordered, align 4 723; CHECK-NEXT: ret void 724; 725 store atomic i32 1, i32* %P unordered, align 4 726 %P2 = bitcast i32* %P to i8* 727 store atomic i32 2, i32* %Q unordered, align 4 728 store atomic i8 3, i8* %P2 unordered, align 4 729 ret void 730} 731 732define void @test43a(i32* %P, i32* noalias %Q) { 733; CHECK-LABEL: @test43a( 734; CHECK-NEXT: entry: 735; CHECK-NEXT: store atomic i32 50331649, i32* [[P:%.*]] unordered, align 4 736; CHECK-NEXT: store atomic i32 2, i32* [[Q:%.*]] unordered, align 4 737; CHECK-NEXT: ret void 738; 739entry: 740 store atomic i32 1, i32* %P unordered, align 4 741 %P2 = bitcast i32* %P to i8* 742 store atomic i32 2, i32* %Q unordered, align 4 743 store atomic i8 3, i8* %P2 unordered, align 4 744 ret void 745} 746 747; Some tests where volatile may block removing a store. 748 749; Here we can remove the first non-volatile store. We cannot remove the 750; volatile store. 751define void @test44_volatile(i32* %P) { 752; CHECK-LABEL: @test44_volatile( 753; CHECK-NEXT: store volatile i32 2, i32* [[P:%.*]], align 4 754; CHECK-NEXT: store i32 3, i32* [[P]], align 4 755; CHECK-NEXT: ret void 756; 757 store i32 1, i32* %P, align 4 758 store volatile i32 2, i32* %P, align 4 759 store i32 3, i32* %P, align 4 760 ret void 761} 762 763define void @test45_volatile(i32* %P) { 764; CHECK-LABEL: @test45_volatile( 765; CHECK-NEXT: store volatile i32 2, i32* [[P:%.*]], align 4 766; CHECK-NEXT: store volatile i32 3, i32* [[P]], align 4 767; CHECK-NEXT: ret void 768; 769 store i32 1, i32* %P, align 4 770 store volatile i32 2, i32* %P, align 4 771 store volatile i32 3, i32* %P, align 4 772 ret void 773} 774 775define void @test46_volatile(i32* %P) { 776; CHECK-LABEL: @test46_volatile( 777; CHECK-NEXT: store volatile i32 2, i32* [[P:%.*]], align 4 778; CHECK-NEXT: store volatile i32 3, i32* [[P]], align 4 779; CHECK-NEXT: ret void 780; 781 store volatile i32 2, i32* %P, align 4 782 store i32 1, i32* %P, align 4 783 store volatile i32 3, i32* %P, align 4 784 ret void 785} 786 787define void @test47_volatile(i32* %P) { 788; CHECK-LABEL: @test47_volatile( 789; CHECK-NEXT: store volatile i32 2, i32* [[P:%.*]], align 4 790; CHECK-NEXT: store volatile i32 3, i32* [[P]], align 4 791; CHECK-NEXT: ret void 792; 793 store volatile i32 2, i32* %P, align 4 794 store volatile i32 3, i32* %P, align 4 795 ret void 796} 797 798define i32 @test48(i32* %P, i32* noalias %Q, i32* %R) { 799; CHECK-LABEL: @test48( 800; CHECK-NEXT: store i32 2, i32* [[P:%.*]], align 4 801; CHECK-NEXT: store i32 3, i32* [[Q:%.*]], align 4 802; CHECK-NEXT: [[L:%.*]] = load i32, i32* [[R:%.*]], align 4 803; CHECK-NEXT: ret i32 [[L]] 804; 805 store i32 1, i32* %Q 806 store i32 2, i32* %P 807 store i32 3, i32* %Q 808 %l = load i32, i32* %R 809 ret i32 %l 810} 811