275 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -S -passes=loop-vectorize -prefer-predicate-over-epilogue=predicate-else-scalar-epilogue < %s | FileCheck --check-prefix=FORCED-TF %s3; RUN: opt -S -passes=loop-vectorize < %s | FileCheck %s4 5; This tests should produce the same result as with default options, and when tail folding6; is preferred, because the vectorizer can't fold the tail by masking (due to an7; outside user of %incdec.ptr in %end) and should fallback to a scalar epilogue.8;9; The first test (@basic_loop) simply relies on the command-line switches.10; The second test (@metadata) specificies its tail-folding preference via metadata.11; Both tests should always generate a scalar epilogue.12 13target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"14 15define void @basic_loop(ptr nocapture readonly %ptr, i32 %size, ptr %pos) {16; FORCED-TF-LABEL: @basic_loop(17; FORCED-TF-NEXT: header:18; FORCED-TF-NEXT: [[PTR0:%.*]] = load ptr, ptr [[POS:%.*]], align 419; FORCED-TF-NEXT: br label [[BODY:%.*]]20; FORCED-TF: body:21; FORCED-TF-NEXT: [[DEC66:%.*]] = phi i32 [ [[DEC:%.*]], [[BODY]] ], [ [[SIZE:%.*]], [[HEADER:%.*]] ]22; FORCED-TF-NEXT: [[BUFF:%.*]] = phi ptr [ [[INCDEC_PTR:%.*]], [[BODY]] ], [ [[PTR:%.*]], [[HEADER]] ]23; FORCED-TF-NEXT: [[INCDEC_PTR]] = getelementptr inbounds i8, ptr [[BUFF]], i32 124; FORCED-TF-NEXT: [[DEC]] = add nsw i32 [[DEC66]], -125; FORCED-TF-NEXT: [[TMP0:%.*]] = load i8, ptr [[INCDEC_PTR]], align 126; FORCED-TF-NEXT: store i8 [[TMP0]], ptr [[BUFF]], align 127; FORCED-TF-NEXT: [[TOBOOL11:%.*]] = icmp eq i32 [[DEC]], 028; FORCED-TF-NEXT: br i1 [[TOBOOL11]], label [[END:%.*]], label [[BODY]]29; FORCED-TF: end:30; FORCED-TF-NEXT: [[INCDEC_PTR_LCSSA:%.*]] = phi ptr [ [[INCDEC_PTR]], [[BODY]] ]31; FORCED-TF-NEXT: store ptr [[INCDEC_PTR_LCSSA]], ptr [[POS]], align 432; FORCED-TF-NEXT: ret void33;34; CHECK-LABEL: @basic_loop(35; CHECK-NEXT: header:36; CHECK-NEXT: [[PTR0:%.*]] = load ptr, ptr [[POS:%.*]], align 437; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[SIZE:%.*]], 438; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]39; CHECK: vector.ph:40; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[SIZE]], 441; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[SIZE]], [[N_MOD_VF]]42; CHECK-NEXT: [[TMP0:%.*]] = sub i32 [[SIZE]], [[N_VEC]]43; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i8, ptr [[PTR:%.*]], i32 [[N_VEC]]44; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]45; CHECK: vector.body:46; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]47; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[PTR]], i32 [[INDEX]]48; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP]], i32 149; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i8>, ptr [[TMP2]], align 150; CHECK-NEXT: store <4 x i8> [[WIDE_LOAD]], ptr [[NEXT_GEP]], align 151; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 452; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]53; CHECK-NEXT: br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]54; CHECK: middle.block:55; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[SIZE]], [[N_VEC]]56; CHECK-NEXT: br i1 [[CMP_N]], label [[END:%.*]], label [[SCALAR_PH]]57; CHECK: scalar.ph:58; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[TMP0]], [[MIDDLE_BLOCK]] ], [ [[SIZE]], [[HEADER:%.*]] ]59; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi ptr [ [[TMP1]], [[MIDDLE_BLOCK]] ], [ [[PTR]], [[HEADER]] ]60; CHECK-NEXT: br label [[BODY:%.*]]61; CHECK: body:62; CHECK-NEXT: [[DEC66:%.*]] = phi i32 [ [[DEC:%.*]], [[BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]63; CHECK-NEXT: [[BUFF:%.*]] = phi ptr [ [[INCDEC_PTR:%.*]], [[BODY]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]64; CHECK-NEXT: [[INCDEC_PTR]] = getelementptr inbounds i8, ptr [[BUFF]], i32 165; CHECK-NEXT: [[DEC]] = add nsw i32 [[DEC66]], -166; CHECK-NEXT: [[TMP4:%.*]] = load i8, ptr [[INCDEC_PTR]], align 167; CHECK-NEXT: store i8 [[TMP4]], ptr [[BUFF]], align 168; CHECK-NEXT: [[TOBOOL11:%.*]] = icmp eq i32 [[DEC]], 069; CHECK-NEXT: br i1 [[TOBOOL11]], label [[END]], label [[BODY]], !llvm.loop [[LOOP3:![0-9]+]]70; CHECK: end:71; CHECK-NEXT: [[INCDEC_PTR_LCSSA:%.*]] = phi ptr [ [[INCDEC_PTR]], [[BODY]] ], [ [[TMP1]], [[MIDDLE_BLOCK]] ]72; CHECK-NEXT: store ptr [[INCDEC_PTR_LCSSA]], ptr [[POS]], align 473; CHECK-NEXT: ret void74;75header:76 %ptr0 = load ptr, ptr %pos, align 477 br label %body78 79body:80 %dec66 = phi i32 [ %dec, %body ], [ %size, %header ]81 %buff = phi ptr [ %incdec.ptr, %body ], [ %ptr, %header ]82 %incdec.ptr = getelementptr inbounds i8, ptr %buff, i32 183 %dec = add nsw i32 %dec66, -184 %0 = load i8, ptr %incdec.ptr, align 185 store i8 %0, ptr %buff, align 186 %tobool11 = icmp eq i32 %dec, 087 br i1 %tobool11, label %end, label %body88 89end:90 store ptr %incdec.ptr, ptr %pos, align 491 ret void92}93 94define void @metadata(ptr nocapture readonly %ptr, i32 %size, ptr %pos) {95; FORCED-TF-LABEL: @metadata(96; FORCED-TF-NEXT: header:97; FORCED-TF-NEXT: [[PTR0:%.*]] = load ptr, ptr [[POS:%.*]], align 498; FORCED-TF-NEXT: br label [[VECTOR_PH:%.*]]99; FORCED-TF: vector.ph:100; FORCED-TF-NEXT: [[N_RND_UP:%.*]] = add i32 [[SIZE:%.*]], 3101; FORCED-TF-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N_RND_UP]], 4102; FORCED-TF-NEXT: [[N_VEC:%.*]] = sub i32 [[N_RND_UP]], [[N_MOD_VF]]103; FORCED-TF-NEXT: [[TRIP_COUNT_MINUS_1:%.*]] = sub i32 [[SIZE]], 1104; FORCED-TF-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TRIP_COUNT_MINUS_1]], i64 0105; FORCED-TF-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer106; FORCED-TF-NEXT: br label [[VECTOR_BODY:%.*]]107; FORCED-TF: vector.body:108; FORCED-TF-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_STORE_CONTINUE11:%.*]] ]109; FORCED-TF-NEXT: [[TMP0:%.*]] = add i32 [[INDEX]], 0110; FORCED-TF-NEXT: [[TMP1:%.*]] = add i32 [[INDEX]], 1111; FORCED-TF-NEXT: [[TMP2:%.*]] = add i32 [[INDEX]], 2112; FORCED-TF-NEXT: [[TMP3:%.*]] = add i32 [[INDEX]], 3113; FORCED-TF-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[PTR:%.*]], i32 [[TMP0]]114; FORCED-TF-NEXT: [[NEXT_GEP1:%.*]] = getelementptr i8, ptr [[PTR]], i32 [[TMP1]]115; FORCED-TF-NEXT: [[NEXT_GEP2:%.*]] = getelementptr i8, ptr [[PTR]], i32 [[TMP2]]116; FORCED-TF-NEXT: [[NEXT_GEP3:%.*]] = getelementptr i8, ptr [[PTR]], i32 [[TMP3]]117; FORCED-TF-NEXT: [[TMP4:%.*]] = insertelement <4 x ptr> poison, ptr [[NEXT_GEP]], i32 0118; FORCED-TF-NEXT: [[TMP5:%.*]] = insertelement <4 x ptr> [[TMP4]], ptr [[NEXT_GEP1]], i32 1119; FORCED-TF-NEXT: [[TMP6:%.*]] = insertelement <4 x ptr> [[TMP5]], ptr [[NEXT_GEP2]], i32 2120; FORCED-TF-NEXT: [[TMP7:%.*]] = insertelement <4 x ptr> [[TMP6]], ptr [[NEXT_GEP3]], i32 3121; FORCED-TF-NEXT: [[BROADCAST_SPLATINSERT4:%.*]] = insertelement <4 x i32> poison, i32 [[INDEX]], i64 0122; FORCED-TF-NEXT: [[BROADCAST_SPLAT5:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT4]], <4 x i32> poison, <4 x i32> zeroinitializer123; FORCED-TF-NEXT: [[VEC_IV:%.*]] = add <4 x i32> [[BROADCAST_SPLAT5]], <i32 0, i32 1, i32 2, i32 3>124; FORCED-TF-NEXT: [[TMP8:%.*]] = icmp ule <4 x i32> [[VEC_IV]], [[BROADCAST_SPLAT]]125; FORCED-TF-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP]], i32 1126; FORCED-TF-NEXT: [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP1]], i32 1127; FORCED-TF-NEXT: [[TMP11:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP2]], i32 1128; FORCED-TF-NEXT: [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP3]], i32 1129; FORCED-TF-NEXT: [[TMP13:%.*]] = insertelement <4 x ptr> poison, ptr [[TMP9]], i32 0130; FORCED-TF-NEXT: [[TMP14:%.*]] = insertelement <4 x ptr> [[TMP13]], ptr [[TMP10]], i32 1131; FORCED-TF-NEXT: [[TMP15:%.*]] = insertelement <4 x ptr> [[TMP14]], ptr [[TMP11]], i32 2132; FORCED-TF-NEXT: [[TMP16:%.*]] = insertelement <4 x ptr> [[TMP15]], ptr [[TMP12]], i32 3133; FORCED-TF-NEXT: [[TMP17:%.*]] = extractelement <4 x i1> [[TMP8]], i32 0134; FORCED-TF-NEXT: br i1 [[TMP17]], label [[PRED_STORE_IF:%.*]], label [[PRED_STORE_CONTINUE:%.*]]135; FORCED-TF: pred.store.if:136; FORCED-TF-NEXT: [[TMP18:%.*]] = load i8, ptr [[TMP9]], align 1137; FORCED-TF-NEXT: store i8 [[TMP18]], ptr [[NEXT_GEP]], align 1138; FORCED-TF-NEXT: br label [[PRED_STORE_CONTINUE]]139; FORCED-TF: pred.store.continue:140; FORCED-TF-NEXT: [[TMP19:%.*]] = extractelement <4 x i1> [[TMP8]], i32 1141; FORCED-TF-NEXT: br i1 [[TMP19]], label [[PRED_STORE_IF6:%.*]], label [[PRED_STORE_CONTINUE7:%.*]]142; FORCED-TF: pred.store.if6:143; FORCED-TF-NEXT: [[TMP20:%.*]] = load i8, ptr [[TMP10]], align 1144; FORCED-TF-NEXT: store i8 [[TMP20]], ptr [[NEXT_GEP1]], align 1145; FORCED-TF-NEXT: br label [[PRED_STORE_CONTINUE7]]146; FORCED-TF: pred.store.continue7:147; FORCED-TF-NEXT: [[TMP21:%.*]] = extractelement <4 x i1> [[TMP8]], i32 2148; FORCED-TF-NEXT: br i1 [[TMP21]], label [[PRED_STORE_IF8:%.*]], label [[PRED_STORE_CONTINUE9:%.*]]149; FORCED-TF: pred.store.if8:150; FORCED-TF-NEXT: [[TMP22:%.*]] = load i8, ptr [[TMP11]], align 1151; FORCED-TF-NEXT: store i8 [[TMP22]], ptr [[NEXT_GEP2]], align 1152; FORCED-TF-NEXT: br label [[PRED_STORE_CONTINUE9]]153; FORCED-TF: pred.store.continue9:154; FORCED-TF-NEXT: [[TMP23:%.*]] = extractelement <4 x i1> [[TMP8]], i32 3155; FORCED-TF-NEXT: br i1 [[TMP23]], label [[PRED_STORE_IF10:%.*]], label [[PRED_STORE_CONTINUE11]]156; FORCED-TF: pred.store.if10:157; FORCED-TF-NEXT: [[TMP24:%.*]] = load i8, ptr [[TMP12]], align 1158; FORCED-TF-NEXT: store i8 [[TMP24]], ptr [[NEXT_GEP3]], align 1159; FORCED-TF-NEXT: br label [[PRED_STORE_CONTINUE11]]160; FORCED-TF: pred.store.continue11:161; FORCED-TF-NEXT: [[INDEX_NEXT]] = add i32 [[INDEX]], 4162; FORCED-TF-NEXT: [[TMP25:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]163; FORCED-TF-NEXT: br i1 [[TMP25]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]164; FORCED-TF: middle.block:165; FORCED-TF-NEXT: [[TMP26:%.*]] = xor <4 x i1> [[TMP8]], splat (i1 true)166; FORCED-TF-NEXT: [[TMP27:%.*]] = call i64 @llvm.experimental.cttz.elts.i64.v4i1(<4 x i1> [[TMP26]], i1 false)167; FORCED-TF-NEXT: [[TMP28:%.*]] = sub i64 [[TMP27]], 1168; FORCED-TF-NEXT: [[TMP29:%.*]] = extractelement <4 x ptr> [[TMP16]], i64 [[TMP28]]169; FORCED-TF-NEXT: br label [[END:%.*]]170; FORCED-TF: end:171; FORCED-TF-NEXT: store ptr [[TMP29]], ptr [[POS]], align 4172; FORCED-TF-NEXT: ret void173;174; CHECK-LABEL: @metadata(175; CHECK-NEXT: header:176; CHECK-NEXT: [[PTR0:%.*]] = load ptr, ptr [[POS:%.*]], align 4177; CHECK-NEXT: br label [[VECTOR_PH:%.*]]178; CHECK: vector.ph:179; CHECK-NEXT: [[N_RND_UP:%.*]] = add i32 [[SIZE:%.*]], 3180; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[N_RND_UP]], 4181; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[N_RND_UP]], [[N_MOD_VF]]182; CHECK-NEXT: [[TRIP_COUNT_MINUS_1:%.*]] = sub i32 [[SIZE]], 1183; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TRIP_COUNT_MINUS_1]], i64 0184; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer185; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]186; CHECK: vector.body:187; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_STORE_CONTINUE11:%.*]] ]188; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[INDEX]], 0189; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[INDEX]], 1190; CHECK-NEXT: [[TMP2:%.*]] = add i32 [[INDEX]], 2191; CHECK-NEXT: [[TMP3:%.*]] = add i32 [[INDEX]], 3192; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[PTR:%.*]], i32 [[TMP0]]193; CHECK-NEXT: [[NEXT_GEP1:%.*]] = getelementptr i8, ptr [[PTR]], i32 [[TMP1]]194; CHECK-NEXT: [[NEXT_GEP2:%.*]] = getelementptr i8, ptr [[PTR]], i32 [[TMP2]]195; CHECK-NEXT: [[NEXT_GEP3:%.*]] = getelementptr i8, ptr [[PTR]], i32 [[TMP3]]196; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x ptr> poison, ptr [[NEXT_GEP]], i32 0197; CHECK-NEXT: [[TMP5:%.*]] = insertelement <4 x ptr> [[TMP4]], ptr [[NEXT_GEP1]], i32 1198; CHECK-NEXT: [[TMP6:%.*]] = insertelement <4 x ptr> [[TMP5]], ptr [[NEXT_GEP2]], i32 2199; CHECK-NEXT: [[TMP7:%.*]] = insertelement <4 x ptr> [[TMP6]], ptr [[NEXT_GEP3]], i32 3200; CHECK-NEXT: [[BROADCAST_SPLATINSERT4:%.*]] = insertelement <4 x i32> poison, i32 [[INDEX]], i64 0201; CHECK-NEXT: [[BROADCAST_SPLAT5:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT4]], <4 x i32> poison, <4 x i32> zeroinitializer202; CHECK-NEXT: [[VEC_IV:%.*]] = add <4 x i32> [[BROADCAST_SPLAT5]], <i32 0, i32 1, i32 2, i32 3>203; CHECK-NEXT: [[TMP8:%.*]] = icmp ule <4 x i32> [[VEC_IV]], [[BROADCAST_SPLAT]]204; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP]], i32 1205; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP1]], i32 1206; CHECK-NEXT: [[TMP11:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP2]], i32 1207; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[NEXT_GEP3]], i32 1208; CHECK-NEXT: [[TMP13:%.*]] = insertelement <4 x ptr> poison, ptr [[TMP9]], i32 0209; CHECK-NEXT: [[TMP14:%.*]] = insertelement <4 x ptr> [[TMP13]], ptr [[TMP10]], i32 1210; CHECK-NEXT: [[TMP15:%.*]] = insertelement <4 x ptr> [[TMP14]], ptr [[TMP11]], i32 2211; CHECK-NEXT: [[TMP16:%.*]] = insertelement <4 x ptr> [[TMP15]], ptr [[TMP12]], i32 3212; CHECK-NEXT: [[TMP17:%.*]] = extractelement <4 x i1> [[TMP8]], i32 0213; CHECK-NEXT: br i1 [[TMP17]], label [[PRED_STORE_IF:%.*]], label [[PRED_STORE_CONTINUE:%.*]]214; CHECK: pred.store.if:215; CHECK-NEXT: [[TMP18:%.*]] = load i8, ptr [[TMP9]], align 1216; CHECK-NEXT: store i8 [[TMP18]], ptr [[NEXT_GEP]], align 1217; CHECK-NEXT: br label [[PRED_STORE_CONTINUE]]218; CHECK: pred.store.continue:219; CHECK-NEXT: [[TMP19:%.*]] = extractelement <4 x i1> [[TMP8]], i32 1220; CHECK-NEXT: br i1 [[TMP19]], label [[PRED_STORE_IF6:%.*]], label [[PRED_STORE_CONTINUE7:%.*]]221; CHECK: pred.store.if6:222; CHECK-NEXT: [[TMP20:%.*]] = load i8, ptr [[TMP10]], align 1223; CHECK-NEXT: store i8 [[TMP20]], ptr [[NEXT_GEP1]], align 1224; CHECK-NEXT: br label [[PRED_STORE_CONTINUE7]]225; CHECK: pred.store.continue7:226; CHECK-NEXT: [[TMP21:%.*]] = extractelement <4 x i1> [[TMP8]], i32 2227; CHECK-NEXT: br i1 [[TMP21]], label [[PRED_STORE_IF8:%.*]], label [[PRED_STORE_CONTINUE9:%.*]]228; CHECK: pred.store.if8:229; CHECK-NEXT: [[TMP22:%.*]] = load i8, ptr [[TMP11]], align 1230; CHECK-NEXT: store i8 [[TMP22]], ptr [[NEXT_GEP2]], align 1231; CHECK-NEXT: br label [[PRED_STORE_CONTINUE9]]232; CHECK: pred.store.continue9:233; CHECK-NEXT: [[TMP23:%.*]] = extractelement <4 x i1> [[TMP8]], i32 3234; CHECK-NEXT: br i1 [[TMP23]], label [[PRED_STORE_IF10:%.*]], label [[PRED_STORE_CONTINUE11]]235; CHECK: pred.store.if10:236; CHECK-NEXT: [[TMP24:%.*]] = load i8, ptr [[TMP12]], align 1237; CHECK-NEXT: store i8 [[TMP24]], ptr [[NEXT_GEP3]], align 1238; CHECK-NEXT: br label [[PRED_STORE_CONTINUE11]]239; CHECK: pred.store.continue11:240; CHECK-NEXT: [[INDEX_NEXT]] = add i32 [[INDEX]], 4241; CHECK-NEXT: [[TMP25:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]242; CHECK-NEXT: br i1 [[TMP25]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]243; CHECK: middle.block:244; CHECK-NEXT: [[TMP26:%.*]] = xor <4 x i1> [[TMP8]], splat (i1 true)245; CHECK-NEXT: [[TMP27:%.*]] = call i64 @llvm.experimental.cttz.elts.i64.v4i1(<4 x i1> [[TMP26]], i1 false)246; CHECK-NEXT: [[TMP28:%.*]] = sub i64 [[TMP27]], 1247; CHECK-NEXT: [[TMP29:%.*]] = extractelement <4 x ptr> [[TMP16]], i64 [[TMP28]]248; CHECK-NEXT: br label [[END:%.*]]249; CHECK: end:250; CHECK-NEXT: store ptr [[TMP29]], ptr [[POS]], align 4251; CHECK-NEXT: ret void252;253header:254 %ptr0 = load ptr, ptr %pos, align 4255 br label %body256 257body:258 %dec66 = phi i32 [ %dec, %body ], [ %size, %header ]259 %buff = phi ptr [ %incdec.ptr, %body ], [ %ptr, %header ]260 %incdec.ptr = getelementptr inbounds i8, ptr %buff, i32 1261 %dec = add nsw i32 %dec66, -1262 %0 = load i8, ptr %incdec.ptr, align 1263 store i8 %0, ptr %buff, align 1264 %tobool11 = icmp eq i32 %dec, 0265 br i1 %tobool11, label %end, label %body, !llvm.loop !1266 267end:268 store ptr %incdec.ptr, ptr %pos, align 4269 ret void270}271 272!1 = distinct !{!1, !2, !3}273!2 = !{!"llvm.loop.vectorize.predicate.enable", i1 true}274!3 = !{!"llvm.loop.vectorize.enable", i1 true}275