327 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=loop-vectorize,dce,instcombine -scalable-vectorization=on -force-target-instruction-cost=1 -force-target-supports-scalable-vectors < %s -S | FileCheck %s3 4; Test that we can add on the induction variable5; for (long long i = 0; i < n; i++) {6; a[i] = b[i] + i;7; }8; with an unroll factor (interleave count) of 2.9 10define void @add_ind64_unrolled(ptr noalias nocapture %a, ptr noalias nocapture readonly %b, i64 %n) {11; CHECK-LABEL: @add_ind64_unrolled(12; CHECK-NEXT: entry:13; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()14; CHECK-NEXT: [[TMP1:%.*]] = shl nuw i64 [[TMP0]], 215; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]]16; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]17; CHECK: vector.ph:18; CHECK-NEXT: [[TMP4:%.*]] = call i64 @llvm.vscale.i64()19; CHECK-NEXT: [[TMP8:%.*]] = shl nuw i64 [[TMP4]], 120; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 2 x i64> poison, i64 [[TMP8]], i64 021; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 2 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer22; CHECK-NEXT: [[TMP5:%.*]] = shl i64 [[TMP4]], 223; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP5]]24; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]25; CHECK-NEXT: [[TMP6:%.*]] = call <vscale x 2 x i64> @llvm.stepvector.nxv2i64()26; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]27; CHECK: vector.body:28; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]29; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 2 x i64> [ [[TMP6]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]30; CHECK-NEXT: [[STEP_ADD:%.*]] = add <vscale x 2 x i64> [[VEC_IND]], [[DOTSPLAT]]31; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds i64, ptr [[B:%.*]], i64 [[INDEX]]32; CHECK-NEXT: [[TMP10:%.*]] = call i64 @llvm.vscale.i64()33; CHECK-NEXT: [[DOTIDX:%.*]] = shl i64 [[TMP10]], 434; CHECK-NEXT: [[TMP11:%.*]] = getelementptr inbounds i8, ptr [[TMP9]], i64 [[DOTIDX]]35; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 2 x i64>, ptr [[TMP9]], align 836; CHECK-NEXT: [[WIDE_LOAD2:%.*]] = load <vscale x 2 x i64>, ptr [[TMP11]], align 837; CHECK-NEXT: [[TMP12:%.*]] = add nsw <vscale x 2 x i64> [[WIDE_LOAD]], [[VEC_IND]]38; CHECK-NEXT: [[TMP13:%.*]] = add nsw <vscale x 2 x i64> [[WIDE_LOAD2]], [[STEP_ADD]]39; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds i64, ptr [[A:%.*]], i64 [[INDEX]]40; CHECK-NEXT: [[TMP15:%.*]] = call i64 @llvm.vscale.i64()41; CHECK-NEXT: [[DOTIDX3:%.*]] = shl i64 [[TMP15]], 442; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i8, ptr [[TMP14]], i64 [[DOTIDX3]]43; CHECK-NEXT: store <vscale x 2 x i64> [[TMP12]], ptr [[TMP14]], align 844; CHECK-NEXT: store <vscale x 2 x i64> [[TMP13]], ptr [[TMP16]], align 845; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP5]]46; CHECK-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 2 x i64> [[STEP_ADD]], [[DOTSPLAT]]47; CHECK-NEXT: [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]48; CHECK-NEXT: br i1 [[TMP17]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]49; CHECK: middle.block:50; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N_MOD_VF]], 051; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]52; CHECK: scalar.ph:53; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]54; CHECK-NEXT: br label [[FOR_BODY:%.*]]55; CHECK: for.body:56; CHECK-NEXT: [[I_08:%.*]] = phi i64 [ [[INC:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]57; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i64, ptr [[B]], i64 [[I_08]]58; CHECK-NEXT: [[TMP18:%.*]] = load i64, ptr [[ARRAYIDX]], align 859; CHECK-NEXT: [[ADD:%.*]] = add nsw i64 [[TMP18]], [[I_08]]60; CHECK-NEXT: [[ARRAYIDX1:%.*]] = getelementptr inbounds i64, ptr [[A]], i64 [[I_08]]61; CHECK-NEXT: store i64 [[ADD]], ptr [[ARRAYIDX1]], align 862; CHECK-NEXT: [[INC]] = add nuw nsw i64 [[I_08]], 163; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INC]], [[N]]64; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]65; CHECK: exit:66; CHECK-NEXT: ret void67;68entry:69 br label %for.body70 71for.body: ; preds = %entry, %for.body72 %i.08 = phi i64 [ %inc, %for.body ], [ 0, %entry ]73 %arrayidx = getelementptr inbounds i64, ptr %b, i64 %i.0874 %0 = load i64, ptr %arrayidx, align 875 %add = add nsw i64 %0, %i.0876 %arrayidx1 = getelementptr inbounds i64, ptr %a, i64 %i.0877 store i64 %add, ptr %arrayidx1, align 878 %inc = add nuw nsw i64 %i.08, 179 %exitcond.not = icmp eq i64 %inc, %n80 br i1 %exitcond.not, label %exit, label %for.body, !llvm.loop !081 82exit: ; preds = %for.body83 ret void84}85 86 87; Same as above, except we test with a vectorisation factor of (1, scalable)88 89define void @add_ind64_unrolled_nxv1i64(ptr noalias nocapture %a, ptr noalias nocapture readonly %b, i64 %n) {90; CHECK-LABEL: @add_ind64_unrolled_nxv1i64(91; CHECK-NEXT: entry:92; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()93; CHECK-NEXT: [[TMP1:%.*]] = shl nuw i64 [[TMP0]], 194; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]]95; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]96; CHECK: vector.ph:97; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()98; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 1 x i64> poison, i64 [[TMP2]], i64 099; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 1 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 1 x i64> poison, <vscale x 1 x i32> zeroinitializer100; CHECK-NEXT: [[TMP3:%.*]] = shl i64 [[TMP2]], 1101; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]]102; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]103; CHECK-NEXT: [[TMP6:%.*]] = call <vscale x 1 x i64> @llvm.stepvector.nxv1i64()104; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]105; CHECK: vector.body:106; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]107; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 1 x i64> [ [[TMP6]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]108; CHECK-NEXT: [[STEP_ADD:%.*]] = add <vscale x 1 x i64> [[VEC_IND]], [[DOTSPLAT]]109; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds i64, ptr [[B:%.*]], i64 [[INDEX]]110; CHECK-NEXT: [[TMP9:%.*]] = call i64 @llvm.vscale.i64()111; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds i64, ptr [[TMP8]], i64 [[TMP9]]112; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 1 x i64>, ptr [[TMP8]], align 8113; CHECK-NEXT: [[WIDE_LOAD2:%.*]] = load <vscale x 1 x i64>, ptr [[TMP10]], align 8114; CHECK-NEXT: [[TMP11:%.*]] = add nsw <vscale x 1 x i64> [[WIDE_LOAD]], [[VEC_IND]]115; CHECK-NEXT: [[TMP12:%.*]] = add nsw <vscale x 1 x i64> [[WIDE_LOAD2]], [[STEP_ADD]]116; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds i64, ptr [[A:%.*]], i64 [[INDEX]]117; CHECK-NEXT: [[TMP14:%.*]] = call i64 @llvm.vscale.i64()118; CHECK-NEXT: [[TMP15:%.*]] = getelementptr inbounds i64, ptr [[TMP13]], i64 [[TMP14]]119; CHECK-NEXT: store <vscale x 1 x i64> [[TMP11]], ptr [[TMP13]], align 8120; CHECK-NEXT: store <vscale x 1 x i64> [[TMP12]], ptr [[TMP15]], align 8121; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]122; CHECK-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 1 x i64> [[STEP_ADD]], [[DOTSPLAT]]123; CHECK-NEXT: [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]124; CHECK-NEXT: br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]125; CHECK: middle.block:126; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N_MOD_VF]], 0127; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]128; CHECK: scalar.ph:129; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]130; CHECK-NEXT: br label [[FOR_BODY:%.*]]131; CHECK: for.body:132; CHECK-NEXT: [[I_08:%.*]] = phi i64 [ [[INC:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]133; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i64, ptr [[B]], i64 [[I_08]]134; CHECK-NEXT: [[TMP17:%.*]] = load i64, ptr [[ARRAYIDX]], align 8135; CHECK-NEXT: [[ADD:%.*]] = add nsw i64 [[TMP17]], [[I_08]]136; CHECK-NEXT: [[ARRAYIDX1:%.*]] = getelementptr inbounds i64, ptr [[A]], i64 [[I_08]]137; CHECK-NEXT: store i64 [[ADD]], ptr [[ARRAYIDX1]], align 8138; CHECK-NEXT: [[INC]] = add nuw nsw i64 [[I_08]], 1139; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INC]], [[N]]140; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]141; CHECK: exit:142; CHECK-NEXT: ret void143;144entry:145 br label %for.body146 147for.body: ; preds = %entry, %for.body148 %i.08 = phi i64 [ %inc, %for.body ], [ 0, %entry ]149 %arrayidx = getelementptr inbounds i64, ptr %b, i64 %i.08150 %0 = load i64, ptr %arrayidx, align 8151 %add = add nsw i64 %0, %i.08152 %arrayidx1 = getelementptr inbounds i64, ptr %a, i64 %i.08153 store i64 %add, ptr %arrayidx1, align 8154 %inc = add nuw nsw i64 %i.08, 1155 %exitcond.not = icmp eq i64 %inc, %n156 br i1 %exitcond.not, label %exit, label %for.body, !llvm.loop !9157 158exit: ; preds = %for.body159 ret void160}161 162 163; Test that we can vectorize a separate induction variable (not used for the branch)164; int r = 0;165; for (long long i = 0; i < n; i++) {166; a[i] = r;167; r += 2;168; }169; with an unroll factor (interleave count) of 1.170 171 172define void @add_unique_ind32(ptr noalias nocapture %a, i64 %n) {173; CHECK-LABEL: @add_unique_ind32(174; CHECK-NEXT: entry:175; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()176; CHECK-NEXT: [[TMP1:%.*]] = shl nuw i64 [[TMP0]], 2177; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]]178; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]179; CHECK: vector.ph:180; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()181; CHECK-NEXT: [[TMP3:%.*]] = shl nuw i64 [[TMP2]], 2182; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]]183; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]184; CHECK-NEXT: [[DOTCAST:%.*]] = trunc i64 [[N_VEC]] to i32185; CHECK-NEXT: [[IND_END:%.*]] = shl i32 [[DOTCAST]], 1186; CHECK-NEXT: [[TMP6:%.*]] = call <vscale x 4 x i32> @llvm.stepvector.nxv4i32()187; CHECK-NEXT: [[TMP7:%.*]] = shl nuw nsw <vscale x 4 x i32> [[TMP6]], splat (i32 1)188; CHECK-NEXT: [[TMP8:%.*]] = trunc i64 [[TMP3]] to i32189; CHECK-NEXT: [[TMP9:%.*]] = shl nuw nsw i32 [[TMP8]], 1190; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[TMP9]], i64 0191; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[DOTSPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer192; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]193; CHECK: vector.body:194; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]195; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 4 x i32> [ [[TMP7]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]196; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[INDEX]]197; CHECK-NEXT: store <vscale x 4 x i32> [[VEC_IND]], ptr [[TMP10]], align 4198; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]199; CHECK-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 4 x i32> [[VEC_IND]], [[DOTSPLAT]]200; CHECK-NEXT: [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]201; CHECK-NEXT: br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP7:![0-9]+]]202; CHECK: middle.block:203; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N_MOD_VF]], 0204; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]205; CHECK: scalar.ph:206; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]207; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]208; CHECK-NEXT: br label [[FOR_BODY:%.*]]209; CHECK: for.body:210; CHECK-NEXT: [[I_08:%.*]] = phi i64 [ [[INC:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]211; CHECK-NEXT: [[R_07:%.*]] = phi i32 [ [[ADD:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]212; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I_08]]213; CHECK-NEXT: store i32 [[R_07]], ptr [[ARRAYIDX]], align 4214; CHECK-NEXT: [[ADD]] = add nuw nsw i32 [[R_07]], 2215; CHECK-NEXT: [[INC]] = add nuw nsw i64 [[I_08]], 1216; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INC]], [[N]]217; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]218; CHECK: exit:219; CHECK-NEXT: ret void220;221entry:222 br label %for.body223 224for.body: ; preds = %entry, %for.body225 %i.08 = phi i64 [ %inc, %for.body ], [ 0, %entry ]226 %r.07 = phi i32 [ %add, %for.body ], [ 0, %entry ]227 %arrayidx = getelementptr inbounds i32, ptr %a, i64 %i.08228 store i32 %r.07, ptr %arrayidx, align 4229 %add = add nuw nsw i32 %r.07, 2230 %inc = add nuw nsw i64 %i.08, 1231 %exitcond.not = icmp eq i64 %inc, %n232 br i1 %exitcond.not, label %exit, label %for.body, !llvm.loop !6233 234exit: ; preds = %for.body235 ret void236}237 238 239; Test that we can vectorize a separate FP induction variable (not used for the branch)240; float r = 0;241; for (long long i = 0; i < n; i++) {242; a[i] = r;243; r += 2;244; }245; with an unroll factor (interleave count) of 1.246 247define void @add_unique_indf32(ptr noalias nocapture %a, i64 %n) {248; CHECK-LABEL: @add_unique_indf32(249; CHECK-NEXT: entry:250; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()251; CHECK-NEXT: [[TMP1:%.*]] = shl nuw i64 [[TMP0]], 2252; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]]253; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]254; CHECK: vector.ph:255; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()256; CHECK-NEXT: [[TMP3:%.*]] = shl nuw i64 [[TMP2]], 2257; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]]258; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]259; CHECK-NEXT: [[DOTCAST:%.*]] = sitofp i64 [[N_VEC]] to float260; CHECK-NEXT: [[TMP4:%.*]] = fmul float [[DOTCAST]], 2.000000e+00261; CHECK-NEXT: [[IND_END:%.*]] = fadd float [[TMP4]], 0.000000e+00262; CHECK-NEXT: [[TMP7:%.*]] = call <vscale x 4 x i32> @llvm.stepvector.nxv4i32()263; CHECK-NEXT: [[TMP8:%.*]] = uitofp <vscale x 4 x i32> [[TMP7]] to <vscale x 4 x float>264; CHECK-NEXT: [[TMP9:%.*]] = fmul <vscale x 4 x float> [[TMP8]], splat (float 2.000000e+00)265; CHECK-NEXT: [[INDUCTION:%.*]] = fadd <vscale x 4 x float> [[TMP9]], zeroinitializer266; CHECK-NEXT: [[TMP12:%.*]] = uitofp i64 [[TMP3]] to float267; CHECK-NEXT: [[TMP13:%.*]] = fmul float [[TMP12]], 2.000000e+00268; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x float> poison, float [[TMP13]], i64 0269; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x float> [[DOTSPLATINSERT]], <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer270; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]271; CHECK: vector.body:272; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]273; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 4 x float> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]274; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds float, ptr [[A:%.*]], i64 [[INDEX]]275; CHECK-NEXT: store <vscale x 4 x float> [[VEC_IND]], ptr [[TMP14]], align 4276; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]277; CHECK-NEXT: [[VEC_IND_NEXT]] = fadd <vscale x 4 x float> [[VEC_IND]], [[DOTSPLAT]]278; CHECK-NEXT: [[TMP15:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]279; CHECK-NEXT: br i1 [[TMP15]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]280; CHECK: middle.block:281; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N_MOD_VF]], 0282; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]283; CHECK: scalar.ph:284; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]285; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi float [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0.000000e+00, [[ENTRY]] ]286; CHECK-NEXT: br label [[FOR_BODY:%.*]]287; CHECK: for.body:288; CHECK-NEXT: [[I_08:%.*]] = phi i64 [ [[INC:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]289; CHECK-NEXT: [[R_07:%.*]] = phi float [ [[ADD:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]290; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[I_08]]291; CHECK-NEXT: store float [[R_07]], ptr [[ARRAYIDX]], align 4292; CHECK-NEXT: [[ADD]] = fadd float [[R_07]], 2.000000e+00293; CHECK-NEXT: [[INC]] = add nuw nsw i64 [[I_08]], 1294; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INC]], [[N]]295; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]296; CHECK: exit:297; CHECK-NEXT: ret void298;299entry:300 br label %for.body301 302for.body: ; preds = %entry, %for.body303 %i.08 = phi i64 [ %inc, %for.body ], [ 0, %entry ]304 %r.07 = phi float [ %add, %for.body ], [ 0.000000e+00, %entry ]305 %arrayidx = getelementptr inbounds float, ptr %a, i64 %i.08306 store float %r.07, ptr %arrayidx, align 4307 %add = fadd float %r.07, 2.000000e+00308 %inc = add nuw nsw i64 %i.08, 1309 %exitcond.not = icmp eq i64 %inc, %n310 br i1 %exitcond.not, label %exit, label %for.body, !llvm.loop !6311 312exit: ; preds = %for.body313 ret void314}315 316!0 = distinct !{!0, !1, !2, !3, !4, !5}317!1 = !{!"llvm.loop.mustprogress"}318!2 = !{!"llvm.loop.vectorize.width", i32 2}319!3 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}320!4 = !{!"llvm.loop.interleave.count", i32 2}321!5 = !{!"llvm.loop.vectorize.enable", i1 true}322!6 = distinct !{!6, !1, !7, !3, !8, !5}323!7 = !{!"llvm.loop.vectorize.width", i32 4}324!8 = !{!"llvm.loop.interleave.count", i32 1}325!9 = distinct !{!9, !1, !10, !3, !4, !5}326!10 = !{!"llvm.loop.vectorize.width", i32 1}327