1; RUN: opt < %s -basicaa -licm -S | FileCheck %s 2 3declare i32 @strlen(i8*) readonly 4 5declare void @foo() 6 7; Sink readonly function. 8define i32 @test1(i8* %P) { 9 br label %Loop 10 11Loop: ; preds = %Loop, %0 12 %A = call i32 @strlen( i8* %P ) readonly 13 br i1 false, label %Loop, label %Out 14 15Out: ; preds = %Loop 16 ret i32 %A 17; CHECK-LABEL: @test1( 18; CHECK: Out: 19; CHECK-NEXT: call i32 @strlen 20; CHECK-NEXT: ret i32 %A 21} 22 23declare double @sin(double) readnone 24 25; Sink readnone function out of loop with unknown memory behavior. 26define double @test2(double %X) { 27 br label %Loop 28 29Loop: ; preds = %Loop, %0 30 call void @foo( ) 31 %A = call double @sin( double %X ) readnone 32 br i1 true, label %Loop, label %Out 33 34Out: ; preds = %Loop 35 ret double %A 36; CHECK-LABEL: @test2( 37; CHECK: Out: 38; CHECK-NEXT: call double @sin 39; CHECK-NEXT: ret double %A 40} 41 42; This testcase checks to make sure the sinker does not cause problems with 43; critical edges. 44define void @test3() { 45Entry: 46 br i1 false, label %Loop, label %Exit 47Loop: 48 %X = add i32 0, 1 49 br i1 false, label %Loop, label %Exit 50Exit: 51 %Y = phi i32 [ 0, %Entry ], [ %X, %Loop ] 52 ret void 53 54; CHECK-LABEL: @test3( 55; CHECK: Exit.loopexit: 56; CHECK-NEXT: %X.le = add i32 0, 1 57; CHECK-NEXT: br label %Exit 58 59} 60 61; If the result of an instruction is only used outside of the loop, sink 62; the instruction to the exit blocks instead of executing it on every 63; iteration of the loop. 64; 65define i32 @test4(i32 %N) { 66Entry: 67 br label %Loop 68Loop: ; preds = %Loop, %Entry 69 %N_addr.0.pn = phi i32 [ %dec, %Loop ], [ %N, %Entry ] 70 %tmp.6 = mul i32 %N, %N_addr.0.pn ; <i32> [#uses=1] 71 %tmp.7 = sub i32 %tmp.6, %N ; <i32> [#uses=1] 72 %dec = add i32 %N_addr.0.pn, -1 ; <i32> [#uses=1] 73 %tmp.1 = icmp ne i32 %N_addr.0.pn, 1 ; <i1> [#uses=1] 74 br i1 %tmp.1, label %Loop, label %Out 75Out: ; preds = %Loop 76 ret i32 %tmp.7 77; CHECK-LABEL: @test4( 78; CHECK: Out: 79; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %N_addr.0.pn 80; CHECK-NEXT: mul i32 %N, %[[LCSSAPHI]] 81; CHECK-NEXT: sub i32 %tmp.6.le, %N 82; CHECK-NEXT: ret i32 83} 84 85; To reduce register pressure, if a load is hoistable out of the loop, and the 86; result of the load is only used outside of the loop, sink the load instead of 87; hoisting it! 88; 89@X = global i32 5 ; <i32*> [#uses=1] 90 91define i32 @test5(i32 %N) { 92Entry: 93 br label %Loop 94Loop: ; preds = %Loop, %Entry 95 %N_addr.0.pn = phi i32 [ %dec, %Loop ], [ %N, %Entry ] 96 %tmp.6 = load i32, i32* @X ; <i32> [#uses=1] 97 %dec = add i32 %N_addr.0.pn, -1 ; <i32> [#uses=1] 98 %tmp.1 = icmp ne i32 %N_addr.0.pn, 1 ; <i1> [#uses=1] 99 br i1 %tmp.1, label %Loop, label %Out 100Out: ; preds = %Loop 101 ret i32 %tmp.6 102; CHECK-LABEL: @test5( 103; CHECK: Out: 104; CHECK-NEXT: %tmp.6.le = load i32, i32* @X 105; CHECK-NEXT: ret i32 %tmp.6.le 106} 107 108 109 110; The loop sinker was running from the bottom of the loop to the top, causing 111; it to miss opportunities to sink instructions that depended on sinking other 112; instructions from the loop. Instead they got hoisted, which is better than 113; leaving them in the loop, but increases register pressure pointlessly. 114 115 %Ty = type { i32, i32 } 116@X2 = external global %Ty 117 118define i32 @test6() { 119 br label %Loop 120Loop: 121 %dead = getelementptr %Ty, %Ty* @X2, i64 0, i32 0 122 %sunk2 = load i32, i32* %dead 123 br i1 false, label %Loop, label %Out 124Out: ; preds = %Loop 125 ret i32 %sunk2 126; CHECK-LABEL: @test6( 127; CHECK: Out: 128; CHECK-NEXT: %dead.le = getelementptr %Ty, %Ty* @X2, i64 0, i32 0 129; CHECK-NEXT: %sunk2.le = load i32, i32* %dead.le 130; CHECK-NEXT: ret i32 %sunk2.le 131} 132 133 134 135; This testcase ensures that we can sink instructions from loops with 136; multiple exits. 137; 138define i32 @test7(i32 %N, i1 %C) { 139Entry: 140 br label %Loop 141Loop: ; preds = %ContLoop, %Entry 142 %N_addr.0.pn = phi i32 [ %dec, %ContLoop ], [ %N, %Entry ] 143 %tmp.6 = mul i32 %N, %N_addr.0.pn 144 %tmp.7 = sub i32 %tmp.6, %N ; <i32> [#uses=2] 145 %dec = add i32 %N_addr.0.pn, -1 ; <i32> [#uses=1] 146 br i1 %C, label %ContLoop, label %Out1 147ContLoop: 148 %tmp.1 = icmp ne i32 %N_addr.0.pn, 1 149 br i1 %tmp.1, label %Loop, label %Out2 150Out1: ; preds = %Loop 151 ret i32 %tmp.7 152Out2: ; preds = %ContLoop 153 ret i32 %tmp.7 154; CHECK-LABEL: @test7( 155; CHECK: Out1: 156; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %N_addr.0.pn 157; CHECK-NEXT: mul i32 %N, %[[LCSSAPHI]] 158; CHECK-NEXT: sub i32 %tmp.6.le, %N 159; CHECK-NEXT: ret 160; CHECK: Out2: 161; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %N_addr.0.pn 162; CHECK-NEXT: mul i32 %N, %[[LCSSAPHI]] 163; CHECK-NEXT: sub i32 %tmp.6.le4, %N 164; CHECK-NEXT: ret 165} 166 167 168; This testcase checks to make sure we can sink values which are only live on 169; some exits out of the loop, and that we can do so without breaking dominator 170; info. 171define i32 @test8(i1 %C1, i1 %C2, i32* %P, i32* %Q) { 172Entry: 173 br label %Loop 174Loop: ; preds = %Cont, %Entry 175 br i1 %C1, label %Cont, label %exit1 176Cont: ; preds = %Loop 177 %X = load i32, i32* %P ; <i32> [#uses=2] 178 store i32 %X, i32* %Q 179 %V = add i32 %X, 1 ; <i32> [#uses=1] 180 br i1 %C2, label %Loop, label %exit2 181exit1: ; preds = %Loop 182 ret i32 0 183exit2: ; preds = %Cont 184 ret i32 %V 185; CHECK-LABEL: @test8( 186; CHECK: exit1: 187; CHECK-NEXT: ret i32 0 188; CHECK: exit2: 189; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %X 190; CHECK-NEXT: %V.le = add i32 %[[LCSSAPHI]], 1 191; CHECK-NEXT: ret i32 %V.le 192} 193 194 195define void @test9() { 196loopentry.2.i: 197 br i1 false, label %no_exit.1.i.preheader, label %loopentry.3.i.preheader 198no_exit.1.i.preheader: ; preds = %loopentry.2.i 199 br label %no_exit.1.i 200no_exit.1.i: ; preds = %endif.8.i, %no_exit.1.i.preheader 201 br i1 false, label %return.i, label %endif.8.i 202endif.8.i: ; preds = %no_exit.1.i 203 %inc.1.i = add i32 0, 1 ; <i32> [#uses=1] 204 br i1 false, label %no_exit.1.i, label %loopentry.3.i.preheader.loopexit 205loopentry.3.i.preheader.loopexit: ; preds = %endif.8.i 206 br label %loopentry.3.i.preheader 207loopentry.3.i.preheader: ; preds = %loopentry.3.i.preheader.loopexit, %loopentry.2.i 208 %arg_num.0.i.ph13000 = phi i32 [ 0, %loopentry.2.i ], [ %inc.1.i, %loopentry.3.i.preheader.loopexit ] ; <i32> [#uses=0] 209 ret void 210return.i: ; preds = %no_exit.1.i 211 ret void 212 213; CHECK-LABEL: @test9( 214; CHECK: loopentry.3.i.preheader.loopexit: 215; CHECK-NEXT: %inc.1.i.le = add i32 0, 1 216; CHECK-NEXT: br label %loopentry.3.i.preheader 217} 218 219 220; Potentially trapping instructions may be sunk as long as they are guaranteed 221; to be executed. 222define i32 @test10(i32 %N) { 223Entry: 224 br label %Loop 225Loop: ; preds = %Loop, %Entry 226 %N_addr.0.pn = phi i32 [ %dec, %Loop ], [ %N, %Entry ] ; <i32> [#uses=3] 227 %tmp.6 = sdiv i32 %N, %N_addr.0.pn ; <i32> [#uses=1] 228 %dec = add i32 %N_addr.0.pn, -1 ; <i32> [#uses=1] 229 %tmp.1 = icmp ne i32 %N_addr.0.pn, 0 ; <i1> [#uses=1] 230 br i1 %tmp.1, label %Loop, label %Out 231Out: ; preds = %Loop 232 ret i32 %tmp.6 233 234; CHECK-LABEL: @test10( 235; CHECK: Out: 236; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i32 [ %N_addr.0.pn 237; CHECK-NEXT: %tmp.6.le = sdiv i32 %N, %[[LCSSAPHI]] 238; CHECK-NEXT: ret i32 %tmp.6.le 239} 240 241; Should delete, not sink, dead instructions. 242define void @test11() { 243 br label %Loop 244Loop: 245 %dead = getelementptr %Ty, %Ty* @X2, i64 0, i32 0 246 br i1 false, label %Loop, label %Out 247Out: 248 ret void 249; CHECK-LABEL: @test11( 250; CHECK: Out: 251; CHECK-NEXT: ret void 252} 253 254@c = common global [1 x i32] zeroinitializer, align 4 255 256; Test a *many* way nested loop with multiple exit blocks both of which exit 257; multiple loop nests. This exercises LCSSA corner cases. 258define i32 @PR18753(i1* %a, i1* %b, i1* %c, i1* %d) { 259entry: 260 br label %l1.header 261 262l1.header: 263 %iv = phi i64 [ %iv.next, %l1.latch ], [ 0, %entry ] 264 %arrayidx.i = getelementptr inbounds [1 x i32], [1 x i32]* @c, i64 0, i64 %iv 265 br label %l2.header 266 267l2.header: 268 %x0 = load i1, i1* %c, align 4 269 br i1 %x0, label %l1.latch, label %l3.preheader 270 271l3.preheader: 272 br label %l3.header 273 274l3.header: 275 %x1 = load i1, i1* %d, align 4 276 br i1 %x1, label %l2.latch, label %l4.preheader 277 278l4.preheader: 279 br label %l4.header 280 281l4.header: 282 %x2 = load i1, i1* %a 283 br i1 %x2, label %l3.latch, label %l4.body 284 285l4.body: 286 call void @f(i32* %arrayidx.i) 287 %x3 = load i1, i1* %b 288 %l = trunc i64 %iv to i32 289 br i1 %x3, label %l4.latch, label %exit 290 291l4.latch: 292 call void @g() 293 %x4 = load i1, i1* %b, align 4 294 br i1 %x4, label %l4.header, label %exit 295 296l3.latch: 297 br label %l3.header 298 299l2.latch: 300 br label %l2.header 301 302l1.latch: 303 %iv.next = add nsw i64 %iv, 1 304 br label %l1.header 305 306exit: 307 %lcssa = phi i32 [ %l, %l4.latch ], [ %l, %l4.body ] 308; CHECK-LABEL: @PR18753( 309; CHECK: exit: 310; CHECK-NEXT: %[[LCSSAPHI:.*]] = phi i64 [ %iv, %l4.latch ], [ %iv, %l4.body ] 311; CHECK-NEXT: %l.le = trunc i64 %[[LCSSAPHI]] to i32 312; CHECK-NEXT: ret i32 %l.le 313 314 ret i32 %lcssa 315} 316 317; Can't sink stores out of exit blocks containing indirectbr instructions 318; because loop simplify does not create dedicated exits for such blocks. Test 319; that by sinking the store from lab21 to lab22, but not further. 320define void @test12() { 321; CHECK-LABEL: @test12 322 br label %lab4 323 324lab4: 325 br label %lab20 326 327lab5: 328 br label %lab20 329 330lab6: 331 br label %lab4 332 333lab7: 334 br i1 undef, label %lab8, label %lab13 335 336lab8: 337 br i1 undef, label %lab13, label %lab10 338 339lab10: 340 br label %lab7 341 342lab13: 343 ret void 344 345lab20: 346 br label %lab21 347 348lab21: 349; CHECK: lab21: 350; CHECK-NOT: store 351; CHECK: br i1 false, label %lab21, label %lab22 352 store i32 36127957, i32* undef, align 4 353 br i1 undef, label %lab21, label %lab22 354 355lab22: 356; CHECK: lab22: 357; CHECK: store 358; CHECK-NEXT: indirectbr i8* undef 359 indirectbr i8* undef, [label %lab5, label %lab6, label %lab7] 360} 361 362; Test that we don't crash when trying to sink stores and there's no preheader 363; available (which is used for creating loads that may be used by the SSA 364; updater) 365define void @test13() { 366; CHECK-LABEL: @test13 367 br label %lab59 368 369lab19: 370 br i1 undef, label %lab20, label %lab38 371 372lab20: 373 br label %lab60 374 375lab21: 376 br i1 undef, label %lab22, label %lab38 377 378lab22: 379 br label %lab38 380 381lab38: 382 ret void 383 384lab59: 385 indirectbr i8* undef, [label %lab60, label %lab38] 386 387lab60: 388; CHECK: lab60: 389; CHECK: store 390; CHECK-NEXT: indirectbr 391 store i32 2145244101, i32* undef, align 4 392 indirectbr i8* undef, [label %lab21, label %lab19] 393} 394 395declare void @f(i32*) 396 397declare void @g() 398