1; RUN: opt < %s -basic-aa -tbaa -licm -S | FileCheck %s 2; RUN: opt -aa-pipeline=tbaa,basic-aa -passes='require<aa>,require<targetir>,require<scalar-evolution>,require<opt-remark-emit>,loop(licm)' -S %s | FileCheck %s 3target 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" 4 5@X = global i32 7 ; <i32*> [#uses=4] 6 7define void @test1(i32 %i) { 8Entry: 9 br label %Loop 10; CHECK-LABEL: @test1( 11; CHECK: Entry: 12; CHECK-NEXT: load i32, i32* @X 13; CHECK-NEXT: br label %Loop 14 15 16Loop: ; preds = %Loop, %0 17 %j = phi i32 [ 0, %Entry ], [ %Next, %Loop ] ; <i32> [#uses=1] 18 %x = load i32, i32* @X ; <i32> [#uses=1] 19 %x2 = add i32 %x, 1 ; <i32> [#uses=1] 20 store i32 %x2, i32* @X 21 %Next = add i32 %j, 1 ; <i32> [#uses=2] 22 %cond = icmp eq i32 %Next, 0 ; <i1> [#uses=1] 23 br i1 %cond, label %Out, label %Loop 24 25Out: 26 ret void 27; CHECK: Out: 28; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %x2 29; CHECK-NEXT: store i32 %[[LCSSAPHI]], i32* @X 30; CHECK-NEXT: ret void 31 32} 33 34define void @test2(i32 %i) { 35Entry: 36 br label %Loop 37; CHECK-LABEL: @test2( 38; CHECK: Entry: 39; CHECK-NEXT: %.promoted = load i32, i32* getelementptr inbounds (i32, i32* @X, i64 1) 40; CHECK-NEXT: br label %Loop 41 42Loop: ; preds = %Loop, %0 43 %X1 = getelementptr i32, i32* @X, i64 1 ; <i32*> [#uses=1] 44 %A = load i32, i32* %X1 ; <i32> [#uses=1] 45 %V = add i32 %A, 1 ; <i32> [#uses=1] 46 %X2 = getelementptr i32, i32* @X, i64 1 ; <i32*> [#uses=1] 47 store i32 %V, i32* %X2 48 br i1 false, label %Loop, label %Exit 49 50Exit: ; preds = %Loop 51 ret void 52; CHECK: Exit: 53; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %V 54; CHECK-NEXT: store i32 %[[LCSSAPHI]], i32* getelementptr inbounds (i32, i32* @X, i64 1) 55; CHECK-NEXT: ret void 56} 57 58 59 60define void @test3(i32 %i) { 61; CHECK-LABEL: @test3( 62 br label %Loop 63Loop: 64 ; Should not promote this to a register 65 %x = load volatile i32, i32* @X 66 %x2 = add i32 %x, 1 67 store i32 %x2, i32* @X 68 br i1 true, label %Out, label %Loop 69 70; CHECK: Loop: 71; CHECK-NEXT: load volatile 72 73Out: ; preds = %Loop 74 ret void 75} 76 77define void @test3b(i32 %i) { 78; CHECK-LABEL: @test3b( 79; CHECK-LABEL: Loop: 80; CHECK: store volatile 81; CHECK-LABEL: Out: 82 br label %Loop 83Loop: 84 ; Should not promote this to a register 85 %x = load i32, i32* @X 86 %x2 = add i32 %x, 1 87 store volatile i32 %x2, i32* @X 88 br i1 true, label %Out, label %Loop 89 90Out: ; preds = %Loop 91 ret void 92} 93 94; PR8041 95define void @test4(i8* %x, i8 %n) { 96; CHECK-LABEL: @test4( 97 %handle1 = alloca i8* 98 %handle2 = alloca i8* 99 store i8* %x, i8** %handle1 100 br label %loop 101 102loop: 103 %tmp = getelementptr i8, i8* %x, i64 8 104 store i8* %tmp, i8** %handle2 105 br label %subloop 106 107subloop: 108 %count = phi i8 [ 0, %loop ], [ %nextcount, %subloop ] 109 %offsetx2 = load i8*, i8** %handle2 110 store i8 %n, i8* %offsetx2 111 %newoffsetx2 = getelementptr i8, i8* %offsetx2, i64 -1 112 store i8* %newoffsetx2, i8** %handle2 113 %nextcount = add i8 %count, 1 114 %innerexitcond = icmp sge i8 %nextcount, 8 115 br i1 %innerexitcond, label %innerexit, label %subloop 116 117; Should have promoted 'handle2' accesses. 118; CHECK: subloop: 119; CHECK-NEXT: phi i8* [ 120; CHECK-NEXT: %count = phi i8 [ 121; CHECK-NEXT: store i8 %n 122; CHECK-NOT: store 123; CHECK: br i1 124 125innerexit: 126 %offsetx1 = load i8*, i8** %handle1 127 %val = load i8, i8* %offsetx1 128 %cond = icmp eq i8 %val, %n 129 br i1 %cond, label %exit, label %loop 130 131; Should not have promoted offsetx1 loads. 132; CHECK: innerexit: 133; CHECK: %val = load i8, i8* %offsetx1 134; CHECK: %cond = icmp eq i8 %val, %n 135; CHECK: br i1 %cond, label %exit, label %loop 136 137exit: 138 ret void 139} 140 141define void @test5(i32 %i, i32** noalias %P2) { 142Entry: 143 br label %Loop 144; CHECK-LABEL: @test5( 145; CHECK: Entry: 146; CHECK-NEXT: load i32, i32* @X 147; CHECK-NEXT: br label %Loop 148 149 150Loop: ; preds = %Loop, %0 151 %j = phi i32 [ 0, %Entry ], [ %Next, %Loop ] ; <i32> [#uses=1] 152 %x = load i32, i32* @X ; <i32> [#uses=1] 153 %x2 = add i32 %x, 1 ; <i32> [#uses=1] 154 store i32 %x2, i32* @X 155 156 store atomic i32* @X, i32** %P2 monotonic, align 8 157 158 %Next = add i32 %j, 1 ; <i32> [#uses=2] 159 %cond = icmp eq i32 %Next, 0 ; <i1> [#uses=1] 160 br i1 %cond, label %Out, label %Loop 161 162Out: 163 ret void 164; CHECK: Out: 165; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %x2 166; CHECK-NEXT: store i32 %[[LCSSAPHI]], i32* @X 167; CHECK-NEXT: ret void 168 169} 170 171 172; PR14753 - Preserve TBAA tags when promoting values in a loop. 173define void @test6(i32 %n, float* nocapture %a, i32* %gi) { 174entry: 175 store i32 0, i32* %gi, align 4, !tbaa !0 176 %cmp1 = icmp slt i32 0, %n 177 br i1 %cmp1, label %for.body.lr.ph, label %for.end 178 179for.body.lr.ph: ; preds = %entry 180 br label %for.body 181 182for.body: ; preds = %for.body.lr.ph, %for.body 183 %storemerge2 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ] 184 %idxprom = sext i32 %storemerge2 to i64 185 %arrayidx = getelementptr inbounds float, float* %a, i64 %idxprom 186 store float 0.000000e+00, float* %arrayidx, align 4, !tbaa !3 187 %0 = load i32, i32* %gi, align 4, !tbaa !0 188 %inc = add nsw i32 %0, 1 189 store i32 %inc, i32* %gi, align 4, !tbaa !0 190 %cmp = icmp slt i32 %inc, %n 191 br i1 %cmp, label %for.body, label %for.cond.for.end_crit_edge 192 193for.cond.for.end_crit_edge: ; preds = %for.body 194 br label %for.end 195 196for.end: ; preds = %for.cond.for.end_crit_edge, %entry 197 ret void 198 199; CHECK: for.body.lr.ph: 200; CHECK-NEXT: %gi.promoted = load i32, i32* %gi, align 4, !tbaa !0 201; CHECK: for.cond.for.end_crit_edge: 202; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %inc 203; CHECK-NEXT: store i32 %[[LCSSAPHI]], i32* %gi, align 4, !tbaa !0 204} 205 206declare i32 @opaque(i32) argmemonly 207declare void @capture(i32*) 208 209; We can promote even if opaque may throw. 210define i32 @test7() { 211; CHECK-LABEL: @test7( 212; CHECK: entry: 213; CHECK-NEXT: %local = alloca 214; CHECK-NEXT: call void @capture(i32* %local) 215; CHECK-NEXT: load i32, i32* %local 216; CHECK-NEXT: br label %loop 217; CHECK: exit: 218; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %x2, %loop ] 219; CHECK-NEXT: store i32 %[[LCSSAPHI]], i32* %local 220; CHECK-NEXT: %ret = load i32, i32* %local 221; CHECK-NEXT: ret i32 %ret 222entry: 223 %local = alloca i32 224 call void @capture(i32* %local) 225 br label %loop 226 227loop: 228 %j = phi i32 [ 0, %entry ], [ %next, %loop ] 229 %x = load i32, i32* %local 230 %x2 = call i32 @opaque(i32 %x) ; Note this does not capture %local 231 store i32 %x2, i32* %local 232 %next = add i32 %j, 1 233 %cond = icmp eq i32 %next, 0 234 br i1 %cond, label %exit, label %loop 235 236exit: 237 %ret = load i32, i32* %local 238 ret i32 %ret 239} 240 241; Make sure we don't promote if the store is really control-flow dependent. 242define i32 @test7bad() { 243; CHECK-LABEL: @test7bad( 244; CHECK: entry: 245; CHECK-NEXT: %local = alloca 246; CHECK-NEXT: call void @capture(i32* %local) 247; CHECK-NEXT: br label %loop 248; CHECK: if: 249; CHECK-NEXT: store i32 %x2, i32* %local 250; CHECK-NEXT: br label %else 251; CHECK: exit: 252; CHECK-NEXT: %ret = load i32, i32* %local 253; CHECK-NEXT: ret i32 %ret 254entry: 255 %local = alloca i32 256 call void @capture(i32* %local) 257 br label %loop 258loop: 259 %j = phi i32 [ 0, %entry ], [ %next, %else ] 260 %x = load i32, i32* %local 261 %x2 = call i32 @opaque(i32 %x) ; Note this does not capture %local 262 %cmp = icmp eq i32 %x2, 0 263 br i1 %cmp, label %if, label %else 264 265if: 266 store i32 %x2, i32* %local 267 br label %else 268 269else: 270 %next = add i32 %j, 1 271 %cond = icmp eq i32 %next, 0 272 br i1 %cond, label %exit, label %loop 273 274exit: 275 %ret = load i32, i32* %local 276 ret i32 %ret 277} 278 279; Even if neither the load nor the store or guaranteed to execute because 280; opaque() may throw, we can still promote - the load not being guaranteed 281; doesn't block us, because %local is always dereferenceable. 282define i32 @test8() { 283; CHECK-LABEL: @test8( 284; CHECK: entry: 285; CHECK-NEXT: %local = alloca 286; CHECK-NEXT: call void @capture(i32* %local) 287; CHECK-NEXT: load i32, i32* %local 288; CHECK-NEXT: br label %loop 289; CHECK: exit: 290; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %x2, %loop ] 291; CHECK-NEXT: store i32 %[[LCSSAPHI]], i32* %local 292; CHECK-NEXT: %ret = load i32, i32* %local 293; CHECK-NEXT: ret i32 %ret 294entry: 295 %local = alloca i32 296 call void @capture(i32* %local) 297 br label %loop 298 299loop: 300 %j = phi i32 [ 0, %entry ], [ %next, %loop ] 301 %throwaway = call i32 @opaque(i32 %j) 302 %x = load i32, i32* %local 303 %x2 = call i32 @opaque(i32 %x) 304 store i32 %x2, i32* %local 305 %next = add i32 %j, 1 306 %cond = icmp eq i32 %next, 0 307 br i1 %cond, label %exit, label %loop 308 309exit: 310 %ret = load i32, i32* %local 311 ret i32 %ret 312} 313 314 315; If the store is "guaranteed modulo exceptions", and the load depends on 316; control flow, we can only promote if the pointer is otherwise known to be 317; dereferenceable 318define i32 @test9() { 319; CHECK-LABEL: @test9( 320; CHECK: entry: 321; CHECK-NEXT: %local = alloca 322; CHECK-NEXT: call void @capture(i32* %local) 323; CHECK-NEXT: load i32, i32* %local 324; CHECK-NEXT: br label %loop 325; CHECK: exit: 326; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %x2, %else ] 327; CHECK-NEXT: store i32 %[[LCSSAPHI]], i32* %local 328; CHECK-NEXT: %ret = load i32, i32* %local 329; CHECK-NEXT: ret i32 %ret 330entry: 331 %local = alloca i32 332 call void @capture(i32* %local) 333 br label %loop 334 335loop: 336 %j = phi i32 [ 0, %entry ], [ %next, %else ] 337 %j2 = call i32 @opaque(i32 %j) 338 %cmp = icmp eq i32 %j2, 0 339 br i1 %cmp, label %if, label %else 340 341if: 342 %x = load i32, i32* %local 343 br label %else 344 345else: 346 %x2 = phi i32 [ 0, %loop ], [ %x, %if] 347 store i32 %x2, i32* %local 348 %next = add i32 %j, 1 349 %cond = icmp eq i32 %next, 0 350 br i1 %cond, label %exit, label %loop 351 352exit: 353 %ret = load i32, i32* %local 354 ret i32 %ret 355} 356 357define i32 @test9bad(i32 %i) { 358; CHECK-LABEL: @test9bad( 359; CHECK: entry: 360; CHECK-NEXT: %local = alloca 361; CHECK-NEXT: call void @capture(i32* %local) 362; CHECK-NEXT: %notderef = getelementptr 363; CHECK-NEXT: br label %loop 364; CHECK: if: 365; CHECK-NEXT: load i32, i32* %notderef 366; CHECK-NEXT: br label %else 367; CHECK: exit: 368; CHECK-NEXT: %ret = load i32, i32* %notderef 369; CHECK-NEXT: ret i32 %ret 370entry: 371 %local = alloca i32 372 call void @capture(i32* %local) 373 %notderef = getelementptr i32, i32* %local, i32 %i 374 br label %loop 375 376loop: 377 %j = phi i32 [ 0, %entry ], [ %next, %else ] 378 %j2 = call i32 @opaque(i32 %j) 379 %cmp = icmp eq i32 %j2, 0 380 br i1 %cmp, label %if, label %else 381 382if: 383 %x = load i32, i32* %notderef 384 br label %else 385 386else: 387 %x2 = phi i32 [ 0, %loop ], [ %x, %if] 388 store i32 %x2, i32* %notderef 389 %next = add i32 %j, 1 390 %cond = icmp eq i32 %next, 0 391 br i1 %cond, label %exit, label %loop 392 393exit: 394 %ret = load i32, i32* %notderef 395 ret i32 %ret 396} 397 398define void @test10(i32 %i) { 399Entry: 400 br label %Loop 401; CHECK-LABEL: @test10( 402; CHECK: Entry: 403; CHECK-NEXT: load atomic i32, i32* @X unordered, align 4 404; CHECK-NEXT: br label %Loop 405 406 407Loop: ; preds = %Loop, %0 408 %j = phi i32 [ 0, %Entry ], [ %Next, %Loop ] ; <i32> [#uses=1] 409 %x = load atomic i32, i32* @X unordered, align 4 410 %x2 = add i32 %x, 1 411 store atomic i32 %x2, i32* @X unordered, align 4 412 %Next = add i32 %j, 1 413 %cond = icmp eq i32 %Next, 0 414 br i1 %cond, label %Out, label %Loop 415 416Out: 417 ret void 418; CHECK: Out: 419; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %x2 420; CHECK-NEXT: store atomic i32 %[[LCSSAPHI]], i32* @X unordered, align 4 421; CHECK-NEXT: ret void 422 423} 424 425; Early exit is known not to be taken on first iteration and thus doesn't 426; effect whether load is known to execute. 427define void @test11(i32 %i) { 428Entry: 429 br label %Loop 430; CHECK-LABEL: @test11( 431; CHECK: Entry: 432; CHECK-NEXT: load i32, i32* @X 433; CHECK-NEXT: br label %Loop 434 435 436Loop: ; preds = %Loop, %0 437 %j = phi i32 [ 0, %Entry ], [ %Next, %body ] ; <i32> [#uses=1] 438 %early.test = icmp ult i32 %j, 32 439 br i1 %early.test, label %body, label %Early 440body: 441 %x = load i32, i32* @X ; <i32> [#uses=1] 442 %x2 = add i32 %x, 1 ; <i32> [#uses=1] 443 store i32 %x2, i32* @X 444 %Next = add i32 %j, 1 ; <i32> [#uses=2] 445 %cond = icmp eq i32 %Next, 0 ; <i1> [#uses=1] 446 br i1 %cond, label %Out, label %Loop 447 448Early: 449; CHECK: Early: 450; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %x2 451; CHECK-NEXT: store i32 %[[LCSSAPHI]], i32* @X 452; CHECK-NEXT: ret void 453 ret void 454Out: 455 ret void 456; CHECK: Out: 457; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %x2 458; CHECK-NEXT: store i32 %[[LCSSAPHI]], i32* @X 459; CHECK-NEXT: ret void 460 461} 462 463!0 = !{!4, !4, i64 0} 464!1 = !{!"omnipotent char", !2} 465!2 = !{!"Simple C/C++ TBAA"} 466!3 = !{!5, !5, i64 0} 467!4 = !{!"int", !1} 468!5 = !{!"float", !1} 469