599 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --filter-out-after "^scalar.ph:" --version 22; RUN: opt -mtriple=aarch64-none-linux-gnu -S -passes=loop-vectorize,instcombine -mattr=+sve -force-vector-width=16 -force-vector-interleave=1 -scalable-vectorization=on -prefer-predicate-over-epilogue=scalar-epilogue %s 2>&1 | FileCheck %s -check-prefix=SCALAR_TAIL_FOLDING3; RUN: opt -mtriple=aarch64-none-linux-gnu -S -passes=loop-vectorize,instcombine -mattr=+sve -force-vector-width=16 -force-vector-interleave=1 -scalable-vectorization=on -prefer-predicate-over-epilogue=predicate-dont-vectorize %s 2>&1 | FileCheck %s -check-prefix=PREDICATED_TAIL_FOLDING4 5target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"6 7; Tests are taken from llvm/test/Transforms/LoopVectorize/interleaved-accesses-masked-group.ll8;9; Expected to contain interleave2/deinterleave2 instructions10;11; void masked_strided1(const unsigned char* restrict p,12; unsigned char* restrict q,13; unsigned char guard) {14; for(ix=0; ix < 1024; ++ix) {15; if (ix > guard) {16; char left = p[2*ix];17; char right = p[2*ix + 1];18; char max = max(left, right);19; q[2*ix] = max;20; q[2*ix+1] = 0 - max;21; }22; }23;}24define dso_local void @masked_strided1(ptr noalias nocapture readonly %p, ptr noalias nocapture %q, i8 zeroext %guard) local_unnamed_addr #0 {25; SCALAR_TAIL_FOLDING-LABEL: define dso_local void @masked_strided126; SCALAR_TAIL_FOLDING-SAME: (ptr noalias readonly captures(none) [[P:%.*]], ptr noalias captures(none) [[Q:%.*]], i8 zeroext [[GUARD:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {27; SCALAR_TAIL_FOLDING-NEXT: entry:28; SCALAR_TAIL_FOLDING-NEXT: [[CONV:%.*]] = zext i8 [[GUARD]] to i3229; SCALAR_TAIL_FOLDING-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()30; SCALAR_TAIL_FOLDING-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ugt i32 [[TMP0]], 6431; SCALAR_TAIL_FOLDING-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]32; SCALAR_TAIL_FOLDING: vector.ph:33; SCALAR_TAIL_FOLDING-NEXT: [[TMP1:%.*]] = call i32 @llvm.vscale.i32()34; SCALAR_TAIL_FOLDING-NEXT: [[TMP2:%.*]] = shl nuw i32 [[TMP1]], 435; SCALAR_TAIL_FOLDING-NEXT: [[N_MOD_VF:%.*]] = urem i32 1024, [[TMP2]]36; SCALAR_TAIL_FOLDING-NEXT: [[N_VEC:%.*]] = sub nuw nsw i32 1024, [[N_MOD_VF]]37; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV]], i64 038; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer39; SCALAR_TAIL_FOLDING-NEXT: [[TMP3:%.*]] = call <vscale x 16 x i32> @llvm.stepvector.nxv16i32()40; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP2]], i64 041; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer42; SCALAR_TAIL_FOLDING-NEXT: br label [[VECTOR_BODY:%.*]]43; SCALAR_TAIL_FOLDING: vector.body:44; SCALAR_TAIL_FOLDING-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]45; SCALAR_TAIL_FOLDING-NEXT: [[VEC_IND:%.*]] = phi <vscale x 16 x i32> [ [[TMP3]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]46; SCALAR_TAIL_FOLDING-NEXT: [[TMP4:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]47; SCALAR_TAIL_FOLDING-NEXT: [[TMP5:%.*]] = shl i32 [[INDEX]], 148; SCALAR_TAIL_FOLDING-NEXT: [[TMP6:%.*]] = sext i32 [[TMP5]] to i6449; SCALAR_TAIL_FOLDING-NEXT: [[TMP7:%.*]] = getelementptr i8, ptr [[P]], i64 [[TMP6]]50; SCALAR_TAIL_FOLDING-NEXT: [[INTERLEAVED_MASK:%.*]] = call <vscale x 32 x i1> @llvm.vector.interleave2.nxv32i1(<vscale x 16 x i1> [[TMP4]], <vscale x 16 x i1> [[TMP4]])51; SCALAR_TAIL_FOLDING-NEXT: [[WIDE_MASKED_VEC:%.*]] = call <vscale x 32 x i8> @llvm.masked.load.nxv32i8.p0(ptr align 1 [[TMP7]], <vscale x 32 x i1> [[INTERLEAVED_MASK]], <vscale x 32 x i8> poison)52; SCALAR_TAIL_FOLDING-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.vector.deinterleave2.nxv32i8(<vscale x 32 x i8> [[WIDE_MASKED_VEC]])53; SCALAR_TAIL_FOLDING-NEXT: [[TMP8:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 054; SCALAR_TAIL_FOLDING-NEXT: [[TMP9:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 155; SCALAR_TAIL_FOLDING-NEXT: [[TMP10:%.*]] = call <vscale x 16 x i8> @llvm.smax.nxv16i8(<vscale x 16 x i8> [[TMP8]], <vscale x 16 x i8> [[TMP9]])56; SCALAR_TAIL_FOLDING-NEXT: [[TMP11:%.*]] = sext i32 [[TMP5]] to i6457; SCALAR_TAIL_FOLDING-NEXT: [[TMP12:%.*]] = getelementptr i8, ptr [[Q]], i64 [[TMP11]]58; SCALAR_TAIL_FOLDING-NEXT: [[TMP13:%.*]] = sub <vscale x 16 x i8> zeroinitializer, [[TMP10]]59; SCALAR_TAIL_FOLDING-NEXT: [[INTERLEAVED_VEC:%.*]] = call <vscale x 32 x i8> @llvm.vector.interleave2.nxv32i8(<vscale x 16 x i8> [[TMP10]], <vscale x 16 x i8> [[TMP13]])60; SCALAR_TAIL_FOLDING-NEXT: [[INTERLEAVED_MASK3:%.*]] = call <vscale x 32 x i1> @llvm.vector.interleave2.nxv32i1(<vscale x 16 x i1> [[TMP4]], <vscale x 16 x i1> [[TMP4]])61; SCALAR_TAIL_FOLDING-NEXT: call void @llvm.masked.store.nxv32i8.p0(<vscale x 32 x i8> [[INTERLEAVED_VEC]], ptr align 1 [[TMP12]], <vscale x 32 x i1> [[INTERLEAVED_MASK3]])62; SCALAR_TAIL_FOLDING-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP2]]63; SCALAR_TAIL_FOLDING-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT2]]64; SCALAR_TAIL_FOLDING-NEXT: [[TMP14:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]65; SCALAR_TAIL_FOLDING-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]66; SCALAR_TAIL_FOLDING: middle.block:67; SCALAR_TAIL_FOLDING-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N_MOD_VF]], 068; SCALAR_TAIL_FOLDING-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]69; SCALAR_TAIL_FOLDING: scalar.ph:70;71; PREDICATED_TAIL_FOLDING-LABEL: define dso_local void @masked_strided172; PREDICATED_TAIL_FOLDING-SAME: (ptr noalias readonly captures(none) [[P:%.*]], ptr noalias captures(none) [[Q:%.*]], i8 zeroext [[GUARD:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {73; PREDICATED_TAIL_FOLDING-NEXT: entry:74; PREDICATED_TAIL_FOLDING-NEXT: br label [[VECTOR_PH:%.*]]75; PREDICATED_TAIL_FOLDING: vector.ph:76; PREDICATED_TAIL_FOLDING-NEXT: [[CONV:%.*]] = zext i8 [[GUARD]] to i3277; PREDICATED_TAIL_FOLDING-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()78; PREDICATED_TAIL_FOLDING-NEXT: [[TMP1:%.*]] = shl nuw i32 [[TMP0]], 479; PREDICATED_TAIL_FOLDING-NEXT: [[TMP2:%.*]] = call i32 @llvm.vscale.i32()80; PREDICATED_TAIL_FOLDING-NEXT: [[TMP3:%.*]] = shl nuw i32 [[TMP2]], 481; PREDICATED_TAIL_FOLDING-NEXT: [[TMP4:%.*]] = sub i32 1024, [[TMP3]]82; PREDICATED_TAIL_FOLDING-NEXT: [[TMP5:%.*]] = icmp ult i32 [[TMP2]], 6483; PREDICATED_TAIL_FOLDING-NEXT: [[TMP6:%.*]] = select i1 [[TMP5]], i32 [[TMP4]], i32 084; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i32(i32 0, i32 1024)85; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV]], i64 086; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer87; PREDICATED_TAIL_FOLDING-NEXT: [[TMP7:%.*]] = call <vscale x 16 x i32> @llvm.stepvector.nxv16i32()88; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP1]], i64 089; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer90; PREDICATED_TAIL_FOLDING-NEXT: br label [[VECTOR_BODY:%.*]]91; PREDICATED_TAIL_FOLDING: vector.body:92; PREDICATED_TAIL_FOLDING-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]93; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 16 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[VECTOR_PH]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]94; PREDICATED_TAIL_FOLDING-NEXT: [[VEC_IND:%.*]] = phi <vscale x 16 x i32> [ [[TMP7]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]95; PREDICATED_TAIL_FOLDING-NEXT: [[TMP8:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]96; PREDICATED_TAIL_FOLDING-NEXT: [[TMP9:%.*]] = select <vscale x 16 x i1> [[ACTIVE_LANE_MASK]], <vscale x 16 x i1> [[TMP8]], <vscale x 16 x i1> zeroinitializer97; PREDICATED_TAIL_FOLDING-NEXT: [[TMP10:%.*]] = shl i32 [[INDEX]], 198; PREDICATED_TAIL_FOLDING-NEXT: [[TMP11:%.*]] = sext i32 [[TMP10]] to i6499; PREDICATED_TAIL_FOLDING-NEXT: [[TMP12:%.*]] = getelementptr i8, ptr [[P]], i64 [[TMP11]]100; PREDICATED_TAIL_FOLDING-NEXT: [[INTERLEAVED_MASK:%.*]] = call <vscale x 32 x i1> @llvm.vector.interleave2.nxv32i1(<vscale x 16 x i1> [[TMP9]], <vscale x 16 x i1> [[TMP9]])101; PREDICATED_TAIL_FOLDING-NEXT: [[WIDE_MASKED_VEC:%.*]] = call <vscale x 32 x i8> @llvm.masked.load.nxv32i8.p0(ptr align 1 [[TMP12]], <vscale x 32 x i1> [[INTERLEAVED_MASK]], <vscale x 32 x i8> poison)102; PREDICATED_TAIL_FOLDING-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.vector.deinterleave2.nxv32i8(<vscale x 32 x i8> [[WIDE_MASKED_VEC]])103; PREDICATED_TAIL_FOLDING-NEXT: [[TMP13:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 0104; PREDICATED_TAIL_FOLDING-NEXT: [[TMP14:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 1105; PREDICATED_TAIL_FOLDING-NEXT: [[TMP15:%.*]] = call <vscale x 16 x i8> @llvm.smax.nxv16i8(<vscale x 16 x i8> [[TMP13]], <vscale x 16 x i8> [[TMP14]])106; PREDICATED_TAIL_FOLDING-NEXT: [[TMP16:%.*]] = sext i32 [[TMP10]] to i64107; PREDICATED_TAIL_FOLDING-NEXT: [[TMP17:%.*]] = getelementptr i8, ptr [[Q]], i64 [[TMP16]]108; PREDICATED_TAIL_FOLDING-NEXT: [[TMP18:%.*]] = sub <vscale x 16 x i8> zeroinitializer, [[TMP15]]109; PREDICATED_TAIL_FOLDING-NEXT: [[INTERLEAVED_VEC:%.*]] = call <vscale x 32 x i8> @llvm.vector.interleave2.nxv32i8(<vscale x 16 x i8> [[TMP15]], <vscale x 16 x i8> [[TMP18]])110; PREDICATED_TAIL_FOLDING-NEXT: [[INTERLEAVED_MASK3:%.*]] = call <vscale x 32 x i1> @llvm.vector.interleave2.nxv32i1(<vscale x 16 x i1> [[TMP9]], <vscale x 16 x i1> [[TMP9]])111; PREDICATED_TAIL_FOLDING-NEXT: call void @llvm.masked.store.nxv32i8.p0(<vscale x 32 x i8> [[INTERLEAVED_VEC]], ptr align 1 [[TMP17]], <vscale x 32 x i1> [[INTERLEAVED_MASK3]])112; PREDICATED_TAIL_FOLDING-NEXT: [[INDEX_NEXT]] = add i32 [[INDEX]], [[TMP1]]113; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i32(i32 [[INDEX]], i32 [[TMP6]])114; PREDICATED_TAIL_FOLDING-NEXT: [[TMP19:%.*]] = extractelement <vscale x 16 x i1> [[ACTIVE_LANE_MASK_NEXT]], i64 0115; PREDICATED_TAIL_FOLDING-NEXT: [[VEC_IND_NEXT]] = add <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT2]]116; PREDICATED_TAIL_FOLDING-NEXT: br i1 [[TMP19]], label [[VECTOR_BODY]], label [[MIDDLE_BLOCK:%.*]], !llvm.loop [[LOOP0:![0-9]+]]117; PREDICATED_TAIL_FOLDING: middle.block:118; PREDICATED_TAIL_FOLDING-NEXT: br label [[FOR_END:%.*]]119; PREDICATED_TAIL_FOLDING: for.end:120; PREDICATED_TAIL_FOLDING-NEXT: ret void121;122entry:123 %conv = zext i8 %guard to i32124 br label %for.body125 126for.body:127 %ix.024 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]128 %cmp1 = icmp ugt i32 %ix.024, %conv129 br i1 %cmp1, label %if.then, label %for.inc130 131if.then:132 %mul = shl nuw nsw i32 %ix.024, 1133 %arrayidx = getelementptr inbounds i8, ptr %p, i32 %mul134 %0 = load i8, ptr %arrayidx, align 1135 %add = or disjoint i32 %mul, 1136 %arrayidx4 = getelementptr inbounds i8, ptr %p, i32 %add137 %1 = load i8, ptr %arrayidx4, align 1138 %cmp.i = icmp slt i8 %0, %1139 %spec.select.i = select i1 %cmp.i, i8 %1, i8 %0140 %arrayidx6 = getelementptr inbounds i8, ptr %q, i32 %mul141 store i8 %spec.select.i, ptr %arrayidx6, align 1142 %sub = sub i8 0, %spec.select.i143 %arrayidx11 = getelementptr inbounds i8, ptr %q, i32 %add144 store i8 %sub, ptr %arrayidx11, align 1145 br label %for.inc146 147for.inc:148 %inc = add nuw nsw i32 %ix.024, 1149 %exitcond = icmp eq i32 %inc, 1024150 br i1 %exitcond, label %for.end, label %for.body151 152for.end:153 ret void154}155 156; Expected not to have interleave2/deinterleave2 instructions157;158; void masked_strided2(const unsigned char* restrict p,159; unsigned char* restrict q,160; unsigned char guard1,161; unsigned char guard2) {162; for(ix=0; ix < 1024; ++ix) {163; if (ix > guard1) {164; q[2*ix] = 1;165; }166; q[2*ix+1] = 2;167; }168;}169define dso_local void @masked_strided2(ptr noalias nocapture readnone %p, ptr noalias nocapture %q, i8 zeroext %guard) local_unnamed_addr #0 {170; SCALAR_TAIL_FOLDING-LABEL: define dso_local void @masked_strided2171; SCALAR_TAIL_FOLDING-SAME: (ptr noalias readnone captures(none) [[P:%.*]], ptr noalias captures(none) [[Q:%.*]], i8 zeroext [[GUARD:%.*]]) local_unnamed_addr #[[ATTR0]] {172; SCALAR_TAIL_FOLDING-NEXT: entry:173; SCALAR_TAIL_FOLDING-NEXT: [[CONV:%.*]] = zext i8 [[GUARD]] to i32174; SCALAR_TAIL_FOLDING-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()175; SCALAR_TAIL_FOLDING-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ugt i32 [[TMP0]], 64176; SCALAR_TAIL_FOLDING-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]177; SCALAR_TAIL_FOLDING: vector.ph:178; SCALAR_TAIL_FOLDING-NEXT: [[TMP1:%.*]] = call i32 @llvm.vscale.i32()179; SCALAR_TAIL_FOLDING-NEXT: [[TMP2:%.*]] = shl nuw i32 [[TMP1]], 4180; SCALAR_TAIL_FOLDING-NEXT: [[N_MOD_VF:%.*]] = urem i32 1024, [[TMP2]]181; SCALAR_TAIL_FOLDING-NEXT: [[N_VEC:%.*]] = sub nuw nsw i32 1024, [[N_MOD_VF]]182; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV]], i64 0183; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer184; SCALAR_TAIL_FOLDING-NEXT: [[TMP3:%.*]] = call <vscale x 16 x i32> @llvm.stepvector.nxv16i32()185; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP2]], i64 0186; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer187; SCALAR_TAIL_FOLDING-NEXT: br label [[VECTOR_BODY:%.*]]188; SCALAR_TAIL_FOLDING: vector.body:189; SCALAR_TAIL_FOLDING-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]190; SCALAR_TAIL_FOLDING-NEXT: [[VEC_IND:%.*]] = phi <vscale x 16 x i32> [ [[TMP3]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]191; SCALAR_TAIL_FOLDING-NEXT: [[TMP4:%.*]] = shl nuw nsw <vscale x 16 x i32> [[VEC_IND]], splat (i32 1)192; SCALAR_TAIL_FOLDING-NEXT: [[TMP5:%.*]] = zext nneg <vscale x 16 x i32> [[TMP4]] to <vscale x 16 x i64>193; SCALAR_TAIL_FOLDING-NEXT: [[TMP6:%.*]] = getelementptr inbounds i8, ptr [[Q]], <vscale x 16 x i64> [[TMP5]]194; SCALAR_TAIL_FOLDING-NEXT: call void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> splat (i8 1), <vscale x 16 x ptr> align 1 [[TMP6]], <vscale x 16 x i1> splat (i1 true))195; SCALAR_TAIL_FOLDING-NEXT: [[TMP7:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]196; SCALAR_TAIL_FOLDING-NEXT: [[TMP8:%.*]] = or disjoint <vscale x 16 x i32> [[TMP4]], splat (i32 1)197; SCALAR_TAIL_FOLDING-NEXT: [[TMP9:%.*]] = zext nneg <vscale x 16 x i32> [[TMP8]] to <vscale x 16 x i64>198; SCALAR_TAIL_FOLDING-NEXT: [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[Q]], <vscale x 16 x i64> [[TMP9]]199; SCALAR_TAIL_FOLDING-NEXT: call void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> splat (i8 2), <vscale x 16 x ptr> align 1 [[TMP10]], <vscale x 16 x i1> [[TMP7]])200; SCALAR_TAIL_FOLDING-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP2]]201; SCALAR_TAIL_FOLDING-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT2]]202; SCALAR_TAIL_FOLDING-NEXT: [[TMP11:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]203; SCALAR_TAIL_FOLDING-NEXT: br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]204; SCALAR_TAIL_FOLDING: middle.block:205; SCALAR_TAIL_FOLDING-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N_MOD_VF]], 0206; SCALAR_TAIL_FOLDING-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]207; SCALAR_TAIL_FOLDING: scalar.ph:208;209; PREDICATED_TAIL_FOLDING-LABEL: define dso_local void @masked_strided2210; PREDICATED_TAIL_FOLDING-SAME: (ptr noalias readnone captures(none) [[P:%.*]], ptr noalias captures(none) [[Q:%.*]], i8 zeroext [[GUARD:%.*]]) local_unnamed_addr #[[ATTR0]] {211; PREDICATED_TAIL_FOLDING-NEXT: entry:212; PREDICATED_TAIL_FOLDING-NEXT: br label [[VECTOR_PH:%.*]]213; PREDICATED_TAIL_FOLDING: vector.ph:214; PREDICATED_TAIL_FOLDING-NEXT: [[CONV:%.*]] = zext i8 [[GUARD]] to i32215; PREDICATED_TAIL_FOLDING-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()216; PREDICATED_TAIL_FOLDING-NEXT: [[TMP1:%.*]] = shl nuw i32 [[TMP0]], 4217; PREDICATED_TAIL_FOLDING-NEXT: [[TMP2:%.*]] = call i32 @llvm.vscale.i32()218; PREDICATED_TAIL_FOLDING-NEXT: [[TMP3:%.*]] = shl nuw i32 [[TMP2]], 4219; PREDICATED_TAIL_FOLDING-NEXT: [[TMP4:%.*]] = sub i32 1024, [[TMP3]]220; PREDICATED_TAIL_FOLDING-NEXT: [[TMP5:%.*]] = icmp ult i32 [[TMP2]], 64221; PREDICATED_TAIL_FOLDING-NEXT: [[TMP6:%.*]] = select i1 [[TMP5]], i32 [[TMP4]], i32 0222; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i32(i32 0, i32 1024)223; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV]], i64 0224; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer225; PREDICATED_TAIL_FOLDING-NEXT: [[TMP7:%.*]] = call <vscale x 16 x i32> @llvm.stepvector.nxv16i32()226; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP1]], i64 0227; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer228; PREDICATED_TAIL_FOLDING-NEXT: br label [[VECTOR_BODY:%.*]]229; PREDICATED_TAIL_FOLDING: vector.body:230; PREDICATED_TAIL_FOLDING-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]231; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 16 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[VECTOR_PH]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]232; PREDICATED_TAIL_FOLDING-NEXT: [[VEC_IND:%.*]] = phi <vscale x 16 x i32> [ [[TMP7]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]233; PREDICATED_TAIL_FOLDING-NEXT: [[TMP8:%.*]] = shl nuw nsw <vscale x 16 x i32> [[VEC_IND]], splat (i32 1)234; PREDICATED_TAIL_FOLDING-NEXT: [[TMP9:%.*]] = zext nneg <vscale x 16 x i32> [[TMP8]] to <vscale x 16 x i64>235; PREDICATED_TAIL_FOLDING-NEXT: [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[Q]], <vscale x 16 x i64> [[TMP9]]236; PREDICATED_TAIL_FOLDING-NEXT: call void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> splat (i8 1), <vscale x 16 x ptr> align 1 [[TMP10]], <vscale x 16 x i1> [[ACTIVE_LANE_MASK]])237; PREDICATED_TAIL_FOLDING-NEXT: [[TMP11:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]238; PREDICATED_TAIL_FOLDING-NEXT: [[TMP12:%.*]] = select <vscale x 16 x i1> [[ACTIVE_LANE_MASK]], <vscale x 16 x i1> [[TMP11]], <vscale x 16 x i1> zeroinitializer239; PREDICATED_TAIL_FOLDING-NEXT: [[TMP13:%.*]] = or disjoint <vscale x 16 x i32> [[TMP8]], splat (i32 1)240; PREDICATED_TAIL_FOLDING-NEXT: [[TMP14:%.*]] = zext nneg <vscale x 16 x i32> [[TMP13]] to <vscale x 16 x i64>241; PREDICATED_TAIL_FOLDING-NEXT: [[TMP15:%.*]] = getelementptr inbounds i8, ptr [[Q]], <vscale x 16 x i64> [[TMP14]]242; PREDICATED_TAIL_FOLDING-NEXT: call void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> splat (i8 2), <vscale x 16 x ptr> align 1 [[TMP15]], <vscale x 16 x i1> [[TMP12]])243; PREDICATED_TAIL_FOLDING-NEXT: [[INDEX_NEXT]] = add i32 [[INDEX]], [[TMP1]]244; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i32(i32 [[INDEX]], i32 [[TMP6]])245; PREDICATED_TAIL_FOLDING-NEXT: [[TMP16:%.*]] = extractelement <vscale x 16 x i1> [[ACTIVE_LANE_MASK_NEXT]], i64 0246; PREDICATED_TAIL_FOLDING-NEXT: [[VEC_IND_NEXT]] = add <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT2]]247; PREDICATED_TAIL_FOLDING-NEXT: br i1 [[TMP16]], label [[VECTOR_BODY]], label [[MIDDLE_BLOCK:%.*]], !llvm.loop [[LOOP3:![0-9]+]]248; PREDICATED_TAIL_FOLDING: middle.block:249; PREDICATED_TAIL_FOLDING-NEXT: br label [[FOR_END:%.*]]250; PREDICATED_TAIL_FOLDING: for.end:251; PREDICATED_TAIL_FOLDING-NEXT: ret void252;253entry:254 %conv = zext i8 %guard to i32255 br label %for.body256 257for.body:258 %ix.012 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]259 %mul = shl nuw nsw i32 %ix.012, 1260 %arrayidx = getelementptr inbounds i8, ptr %q, i32 %mul261 store i8 1, ptr %arrayidx, align 1262 %cmp1 = icmp ugt i32 %ix.012, %conv263 br i1 %cmp1, label %if.then, label %for.inc264 265if.then:266 %add = or disjoint i32 %mul, 1267 %arrayidx3 = getelementptr inbounds i8, ptr %q, i32 %add268 store i8 2, ptr %arrayidx3, align 1269 br label %for.inc270 271for.inc:272 %inc = add nuw nsw i32 %ix.012, 1273 %exitcond = icmp eq i32 %inc, 1024274 br i1 %exitcond, label %for.end, label %for.body275 276for.end:277 ret void278}279 280; Expected not to have interleave2/deinterleave2 instructions281;282; void masked_strided3(const unsigned char* restrict p,283; unsigned char* restrict q,284; unsigned char guard1,285; unsigned char guard2) {286; for(ix=0; ix < 1024; ++ix) {287; if (ix > guard1) {288; q[2*ix] = 1;289; }290; if (ix > guard2) {291; q[2*ix+1] = 2;292; }293; }294;}295define dso_local void @masked_strided3(ptr noalias nocapture readnone %p, ptr noalias nocapture %q, i8 zeroext %guard1, i8 zeroext %guard2) local_unnamed_addr #0 {296; SCALAR_TAIL_FOLDING-LABEL: define dso_local void @masked_strided3297; SCALAR_TAIL_FOLDING-SAME: (ptr noalias readnone captures(none) [[P:%.*]], ptr noalias captures(none) [[Q:%.*]], i8 zeroext [[GUARD1:%.*]], i8 zeroext [[GUARD2:%.*]]) local_unnamed_addr #[[ATTR0]] {298; SCALAR_TAIL_FOLDING-NEXT: entry:299; SCALAR_TAIL_FOLDING-NEXT: [[CONV:%.*]] = zext i8 [[GUARD1]] to i32300; SCALAR_TAIL_FOLDING-NEXT: [[CONV3:%.*]] = zext i8 [[GUARD2]] to i32301; SCALAR_TAIL_FOLDING-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()302; SCALAR_TAIL_FOLDING-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ugt i32 [[TMP0]], 64303; SCALAR_TAIL_FOLDING-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]304; SCALAR_TAIL_FOLDING: vector.ph:305; SCALAR_TAIL_FOLDING-NEXT: [[TMP1:%.*]] = call i32 @llvm.vscale.i32()306; SCALAR_TAIL_FOLDING-NEXT: [[TMP2:%.*]] = shl nuw i32 [[TMP1]], 4307; SCALAR_TAIL_FOLDING-NEXT: [[N_MOD_VF:%.*]] = urem i32 1024, [[TMP2]]308; SCALAR_TAIL_FOLDING-NEXT: [[N_VEC:%.*]] = sub nuw nsw i32 1024, [[N_MOD_VF]]309; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV]], i64 0310; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer311; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV3]], i64 0312; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer313; SCALAR_TAIL_FOLDING-NEXT: [[TMP3:%.*]] = call <vscale x 16 x i32> @llvm.stepvector.nxv16i32()314; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP2]], i64 0315; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT3]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer316; SCALAR_TAIL_FOLDING-NEXT: br label [[VECTOR_BODY:%.*]]317; SCALAR_TAIL_FOLDING: vector.body:318; SCALAR_TAIL_FOLDING-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]319; SCALAR_TAIL_FOLDING-NEXT: [[VEC_IND:%.*]] = phi <vscale x 16 x i32> [ [[TMP3]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]320; SCALAR_TAIL_FOLDING-NEXT: [[TMP4:%.*]] = shl nuw nsw <vscale x 16 x i32> [[VEC_IND]], splat (i32 1)321; SCALAR_TAIL_FOLDING-NEXT: [[TMP5:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]322; SCALAR_TAIL_FOLDING-NEXT: [[TMP6:%.*]] = zext nneg <vscale x 16 x i32> [[TMP4]] to <vscale x 16 x i64>323; SCALAR_TAIL_FOLDING-NEXT: [[TMP7:%.*]] = getelementptr inbounds i8, ptr [[Q]], <vscale x 16 x i64> [[TMP6]]324; SCALAR_TAIL_FOLDING-NEXT: call void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> splat (i8 1), <vscale x 16 x ptr> align 1 [[TMP7]], <vscale x 16 x i1> [[TMP5]])325; SCALAR_TAIL_FOLDING-NEXT: [[TMP8:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT2]]326; SCALAR_TAIL_FOLDING-NEXT: [[TMP9:%.*]] = or disjoint <vscale x 16 x i32> [[TMP4]], splat (i32 1)327; SCALAR_TAIL_FOLDING-NEXT: [[TMP10:%.*]] = zext nneg <vscale x 16 x i32> [[TMP9]] to <vscale x 16 x i64>328; SCALAR_TAIL_FOLDING-NEXT: [[TMP11:%.*]] = getelementptr inbounds i8, ptr [[Q]], <vscale x 16 x i64> [[TMP10]]329; SCALAR_TAIL_FOLDING-NEXT: call void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> splat (i8 2), <vscale x 16 x ptr> align 1 [[TMP11]], <vscale x 16 x i1> [[TMP8]])330; SCALAR_TAIL_FOLDING-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP2]]331; SCALAR_TAIL_FOLDING-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT4]]332; SCALAR_TAIL_FOLDING-NEXT: [[TMP12:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]333; SCALAR_TAIL_FOLDING-NEXT: br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]334; SCALAR_TAIL_FOLDING: middle.block:335; SCALAR_TAIL_FOLDING-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N_MOD_VF]], 0336; SCALAR_TAIL_FOLDING-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]337; SCALAR_TAIL_FOLDING: scalar.ph:338;339; PREDICATED_TAIL_FOLDING-LABEL: define dso_local void @masked_strided3340; PREDICATED_TAIL_FOLDING-SAME: (ptr noalias readnone captures(none) [[P:%.*]], ptr noalias captures(none) [[Q:%.*]], i8 zeroext [[GUARD1:%.*]], i8 zeroext [[GUARD2:%.*]]) local_unnamed_addr #[[ATTR0]] {341; PREDICATED_TAIL_FOLDING-NEXT: entry:342; PREDICATED_TAIL_FOLDING-NEXT: br label [[VECTOR_PH:%.*]]343; PREDICATED_TAIL_FOLDING: vector.ph:344; PREDICATED_TAIL_FOLDING-NEXT: [[CONV3:%.*]] = zext i8 [[GUARD2]] to i32345; PREDICATED_TAIL_FOLDING-NEXT: [[CONV:%.*]] = zext i8 [[GUARD1]] to i32346; PREDICATED_TAIL_FOLDING-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()347; PREDICATED_TAIL_FOLDING-NEXT: [[TMP1:%.*]] = shl nuw i32 [[TMP0]], 4348; PREDICATED_TAIL_FOLDING-NEXT: [[TMP2:%.*]] = call i32 @llvm.vscale.i32()349; PREDICATED_TAIL_FOLDING-NEXT: [[TMP3:%.*]] = shl nuw i32 [[TMP2]], 4350; PREDICATED_TAIL_FOLDING-NEXT: [[TMP4:%.*]] = sub i32 1024, [[TMP3]]351; PREDICATED_TAIL_FOLDING-NEXT: [[TMP5:%.*]] = icmp ult i32 [[TMP2]], 64352; PREDICATED_TAIL_FOLDING-NEXT: [[TMP6:%.*]] = select i1 [[TMP5]], i32 [[TMP4]], i32 0353; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i32(i32 0, i32 1024)354; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV]], i64 0355; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer356; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV3]], i64 0357; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer358; PREDICATED_TAIL_FOLDING-NEXT: [[TMP7:%.*]] = call <vscale x 16 x i32> @llvm.stepvector.nxv16i32()359; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP1]], i64 0360; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT3]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer361; PREDICATED_TAIL_FOLDING-NEXT: br label [[VECTOR_BODY:%.*]]362; PREDICATED_TAIL_FOLDING: vector.body:363; PREDICATED_TAIL_FOLDING-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]364; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 16 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[VECTOR_PH]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]365; PREDICATED_TAIL_FOLDING-NEXT: [[VEC_IND:%.*]] = phi <vscale x 16 x i32> [ [[TMP7]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]366; PREDICATED_TAIL_FOLDING-NEXT: [[TMP8:%.*]] = shl nuw nsw <vscale x 16 x i32> [[VEC_IND]], splat (i32 1)367; PREDICATED_TAIL_FOLDING-NEXT: [[TMP9:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]368; PREDICATED_TAIL_FOLDING-NEXT: [[TMP10:%.*]] = select <vscale x 16 x i1> [[ACTIVE_LANE_MASK]], <vscale x 16 x i1> [[TMP9]], <vscale x 16 x i1> zeroinitializer369; PREDICATED_TAIL_FOLDING-NEXT: [[TMP11:%.*]] = zext nneg <vscale x 16 x i32> [[TMP8]] to <vscale x 16 x i64>370; PREDICATED_TAIL_FOLDING-NEXT: [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[Q]], <vscale x 16 x i64> [[TMP11]]371; PREDICATED_TAIL_FOLDING-NEXT: call void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> splat (i8 1), <vscale x 16 x ptr> align 1 [[TMP12]], <vscale x 16 x i1> [[TMP10]])372; PREDICATED_TAIL_FOLDING-NEXT: [[TMP13:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT2]]373; PREDICATED_TAIL_FOLDING-NEXT: [[TMP14:%.*]] = select <vscale x 16 x i1> [[ACTIVE_LANE_MASK]], <vscale x 16 x i1> [[TMP13]], <vscale x 16 x i1> zeroinitializer374; PREDICATED_TAIL_FOLDING-NEXT: [[TMP15:%.*]] = or disjoint <vscale x 16 x i32> [[TMP8]], splat (i32 1)375; PREDICATED_TAIL_FOLDING-NEXT: [[TMP16:%.*]] = zext nneg <vscale x 16 x i32> [[TMP15]] to <vscale x 16 x i64>376; PREDICATED_TAIL_FOLDING-NEXT: [[TMP17:%.*]] = getelementptr inbounds i8, ptr [[Q]], <vscale x 16 x i64> [[TMP16]]377; PREDICATED_TAIL_FOLDING-NEXT: call void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> splat (i8 2), <vscale x 16 x ptr> align 1 [[TMP17]], <vscale x 16 x i1> [[TMP14]])378; PREDICATED_TAIL_FOLDING-NEXT: [[INDEX_NEXT]] = add i32 [[INDEX]], [[TMP1]]379; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i32(i32 [[INDEX]], i32 [[TMP6]])380; PREDICATED_TAIL_FOLDING-NEXT: [[TMP18:%.*]] = extractelement <vscale x 16 x i1> [[ACTIVE_LANE_MASK_NEXT]], i64 0381; PREDICATED_TAIL_FOLDING-NEXT: [[VEC_IND_NEXT]] = add <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT4]]382; PREDICATED_TAIL_FOLDING-NEXT: br i1 [[TMP18]], label [[VECTOR_BODY]], label [[MIDDLE_BLOCK:%.*]], !llvm.loop [[LOOP4:![0-9]+]]383; PREDICATED_TAIL_FOLDING: middle.block:384; PREDICATED_TAIL_FOLDING-NEXT: br label [[FOR_END:%.*]]385; PREDICATED_TAIL_FOLDING: for.end:386; PREDICATED_TAIL_FOLDING-NEXT: ret void387;388entry:389 %conv = zext i8 %guard1 to i32390 %conv3 = zext i8 %guard2 to i32391 br label %for.body392 393for.body:394 %ix.018 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]395 %mul = shl nuw nsw i32 %ix.018, 1396 %cmp1 = icmp ugt i32 %ix.018, %conv397 br i1 %cmp1, label %if.then, label %if.end398 399if.then:400 %arrayidx = getelementptr inbounds i8, ptr %q, i32 %mul401 store i8 1, ptr %arrayidx, align 1402 br label %if.end403 404if.end:405 %cmp4 = icmp ugt i32 %ix.018, %conv3406 br i1 %cmp4, label %if.then6, label %for.inc407 408if.then6:409 %add = or disjoint i32 %mul, 1410 %arrayidx7 = getelementptr inbounds i8, ptr %q, i32 %add411 store i8 2, ptr %arrayidx7, align 1412 br label %for.inc413 414for.inc:415 %inc = add nuw nsw i32 %ix.018, 1416 %exitcond = icmp eq i32 %inc, 1024417 br i1 %exitcond, label %for.end, label %for.body418 419for.end:420 ret void421}422 423; Expected to contain interleave2/deinterleave2 instructions424;425; void masked_strided_factor4(const unsigned char* restrict p,426; unsigned char* restrict q,427; unsigned char guard) {428; for(ix=0; ix < 1024; ++ix) {429; if (ix > guard) {430; char left1 = p[4*ix];431; char right1 = p[4*ix + 1];432; char left2 = p[4*ix + 2];433; char right2 = p[4*ix + 3];434; char max1 = max(left1, right1);435; char max2 = max(left2, right2);436; q[4*ix] = max1;437; q[4*ix + 1] = 0 - max1;438; q[4*ix + 2] = max2;439; q[4*ix + 3] = 0 - max2;440; }441; }442;}443define dso_local void @masked_strided_factor4(ptr noalias nocapture readonly %p, ptr noalias nocapture %q, i8 zeroext %guard) local_unnamed_addr #0 {444; SCALAR_TAIL_FOLDING-LABEL: define dso_local void @masked_strided_factor4445; SCALAR_TAIL_FOLDING-SAME: (ptr noalias readonly captures(none) [[P:%.*]], ptr noalias captures(none) [[Q:%.*]], i8 zeroext [[GUARD:%.*]]) local_unnamed_addr #[[ATTR0]] {446; SCALAR_TAIL_FOLDING-NEXT: entry:447; SCALAR_TAIL_FOLDING-NEXT: [[CONV:%.*]] = zext i8 [[GUARD]] to i32448; SCALAR_TAIL_FOLDING-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()449; SCALAR_TAIL_FOLDING-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ugt i32 [[TMP0]], 64450; SCALAR_TAIL_FOLDING-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]451; SCALAR_TAIL_FOLDING: vector.ph:452; SCALAR_TAIL_FOLDING-NEXT: [[TMP1:%.*]] = call i32 @llvm.vscale.i32()453; SCALAR_TAIL_FOLDING-NEXT: [[TMP2:%.*]] = shl nuw i32 [[TMP1]], 4454; SCALAR_TAIL_FOLDING-NEXT: [[N_MOD_VF:%.*]] = urem i32 1024, [[TMP2]]455; SCALAR_TAIL_FOLDING-NEXT: [[N_VEC:%.*]] = sub nuw nsw i32 1024, [[N_MOD_VF]]456; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV]], i64 0457; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer458; SCALAR_TAIL_FOLDING-NEXT: [[TMP3:%.*]] = call <vscale x 16 x i32> @llvm.stepvector.nxv16i32()459; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP2]], i64 0460; SCALAR_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer461; SCALAR_TAIL_FOLDING-NEXT: br label [[VECTOR_BODY:%.*]]462; SCALAR_TAIL_FOLDING: vector.body:463; SCALAR_TAIL_FOLDING-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]464; SCALAR_TAIL_FOLDING-NEXT: [[VEC_IND:%.*]] = phi <vscale x 16 x i32> [ [[TMP3]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]465; SCALAR_TAIL_FOLDING-NEXT: [[TMP4:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]466; SCALAR_TAIL_FOLDING-NEXT: [[TMP5:%.*]] = shl i32 [[INDEX]], 2467; SCALAR_TAIL_FOLDING-NEXT: [[TMP6:%.*]] = sext i32 [[TMP5]] to i64468; SCALAR_TAIL_FOLDING-NEXT: [[TMP7:%.*]] = getelementptr i8, ptr [[P]], i64 [[TMP6]]469; SCALAR_TAIL_FOLDING-NEXT: [[INTERLEAVED_MASK:%.*]] = call <vscale x 64 x i1> @llvm.vector.interleave4.nxv64i1(<vscale x 16 x i1> [[TMP4]], <vscale x 16 x i1> [[TMP4]], <vscale x 16 x i1> [[TMP4]], <vscale x 16 x i1> [[TMP4]])470; SCALAR_TAIL_FOLDING-NEXT: [[WIDE_MASKED_VEC:%.*]] = call <vscale x 64 x i8> @llvm.masked.load.nxv64i8.p0(ptr align 1 [[TMP7]], <vscale x 64 x i1> [[INTERLEAVED_MASK]], <vscale x 64 x i8> poison)471; SCALAR_TAIL_FOLDING-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.vector.deinterleave4.nxv64i8(<vscale x 64 x i8> [[WIDE_MASKED_VEC]])472; SCALAR_TAIL_FOLDING-NEXT: [[TMP8:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 0473; SCALAR_TAIL_FOLDING-NEXT: [[TMP9:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 1474; SCALAR_TAIL_FOLDING-NEXT: [[TMP10:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 2475; SCALAR_TAIL_FOLDING-NEXT: [[TMP11:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 3476; SCALAR_TAIL_FOLDING-NEXT: [[TMP12:%.*]] = call <vscale x 16 x i8> @llvm.smax.nxv16i8(<vscale x 16 x i8> [[TMP8]], <vscale x 16 x i8> [[TMP9]])477; SCALAR_TAIL_FOLDING-NEXT: [[TMP13:%.*]] = sub <vscale x 16 x i8> zeroinitializer, [[TMP12]]478; SCALAR_TAIL_FOLDING-NEXT: [[TMP14:%.*]] = call <vscale x 16 x i8> @llvm.smax.nxv16i8(<vscale x 16 x i8> [[TMP10]], <vscale x 16 x i8> [[TMP11]])479; SCALAR_TAIL_FOLDING-NEXT: [[TMP15:%.*]] = sub <vscale x 16 x i8> zeroinitializer, [[TMP14]]480; SCALAR_TAIL_FOLDING-NEXT: [[TMP16:%.*]] = sext i32 [[TMP5]] to i64481; SCALAR_TAIL_FOLDING-NEXT: [[TMP17:%.*]] = getelementptr i8, ptr [[Q]], i64 [[TMP16]]482; SCALAR_TAIL_FOLDING-NEXT: [[INTERLEAVED_VEC:%.*]] = call <vscale x 64 x i8> @llvm.vector.interleave4.nxv64i8(<vscale x 16 x i8> [[TMP12]], <vscale x 16 x i8> [[TMP13]], <vscale x 16 x i8> [[TMP14]], <vscale x 16 x i8> [[TMP15]])483; SCALAR_TAIL_FOLDING-NEXT: [[INTERLEAVED_MASK3:%.*]] = call <vscale x 64 x i1> @llvm.vector.interleave4.nxv64i1(<vscale x 16 x i1> [[TMP4]], <vscale x 16 x i1> [[TMP4]], <vscale x 16 x i1> [[TMP4]], <vscale x 16 x i1> [[TMP4]])484; SCALAR_TAIL_FOLDING-NEXT: call void @llvm.masked.store.nxv64i8.p0(<vscale x 64 x i8> [[INTERLEAVED_VEC]], ptr align 1 [[TMP17]], <vscale x 64 x i1> [[INTERLEAVED_MASK3]])485; SCALAR_TAIL_FOLDING-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP2]]486; SCALAR_TAIL_FOLDING-NEXT: [[VEC_IND_NEXT]] = add nuw nsw <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT2]]487; SCALAR_TAIL_FOLDING-NEXT: [[TMP18:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]488; SCALAR_TAIL_FOLDING-NEXT: br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]489; SCALAR_TAIL_FOLDING: middle.block:490; SCALAR_TAIL_FOLDING-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N_MOD_VF]], 0491; SCALAR_TAIL_FOLDING-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]492; SCALAR_TAIL_FOLDING: scalar.ph:493;494; PREDICATED_TAIL_FOLDING-LABEL: define dso_local void @masked_strided_factor4495; PREDICATED_TAIL_FOLDING-SAME: (ptr noalias readonly captures(none) [[P:%.*]], ptr noalias captures(none) [[Q:%.*]], i8 zeroext [[GUARD:%.*]]) local_unnamed_addr #[[ATTR0]] {496; PREDICATED_TAIL_FOLDING-NEXT: entry:497; PREDICATED_TAIL_FOLDING-NEXT: br label [[VECTOR_PH:%.*]]498; PREDICATED_TAIL_FOLDING: vector.ph:499; PREDICATED_TAIL_FOLDING-NEXT: [[CONV:%.*]] = zext i8 [[GUARD]] to i32500; PREDICATED_TAIL_FOLDING-NEXT: [[TMP0:%.*]] = call i32 @llvm.vscale.i32()501; PREDICATED_TAIL_FOLDING-NEXT: [[TMP1:%.*]] = shl nuw i32 [[TMP0]], 4502; PREDICATED_TAIL_FOLDING-NEXT: [[TMP2:%.*]] = call i32 @llvm.vscale.i32()503; PREDICATED_TAIL_FOLDING-NEXT: [[TMP3:%.*]] = shl nuw i32 [[TMP2]], 4504; PREDICATED_TAIL_FOLDING-NEXT: [[TMP4:%.*]] = sub i32 1024, [[TMP3]]505; PREDICATED_TAIL_FOLDING-NEXT: [[TMP5:%.*]] = icmp ult i32 [[TMP2]], 64506; PREDICATED_TAIL_FOLDING-NEXT: [[TMP6:%.*]] = select i1 [[TMP5]], i32 [[TMP4]], i32 0507; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i32(i32 0, i32 1024)508; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV]], i64 0509; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer510; PREDICATED_TAIL_FOLDING-NEXT: [[TMP7:%.*]] = call <vscale x 16 x i32> @llvm.stepvector.nxv16i32()511; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP1]], i64 0512; PREDICATED_TAIL_FOLDING-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer513; PREDICATED_TAIL_FOLDING-NEXT: br label [[VECTOR_BODY:%.*]]514; PREDICATED_TAIL_FOLDING: vector.body:515; PREDICATED_TAIL_FOLDING-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]516; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 16 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[VECTOR_PH]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]517; PREDICATED_TAIL_FOLDING-NEXT: [[VEC_IND:%.*]] = phi <vscale x 16 x i32> [ [[TMP7]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]518; PREDICATED_TAIL_FOLDING-NEXT: [[TMP8:%.*]] = icmp ugt <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]519; PREDICATED_TAIL_FOLDING-NEXT: [[TMP9:%.*]] = select <vscale x 16 x i1> [[ACTIVE_LANE_MASK]], <vscale x 16 x i1> [[TMP8]], <vscale x 16 x i1> zeroinitializer520; PREDICATED_TAIL_FOLDING-NEXT: [[TMP10:%.*]] = shl i32 [[INDEX]], 2521; PREDICATED_TAIL_FOLDING-NEXT: [[TMP11:%.*]] = sext i32 [[TMP10]] to i64522; PREDICATED_TAIL_FOLDING-NEXT: [[TMP12:%.*]] = getelementptr i8, ptr [[P]], i64 [[TMP11]]523; PREDICATED_TAIL_FOLDING-NEXT: [[INTERLEAVED_MASK:%.*]] = call <vscale x 64 x i1> @llvm.vector.interleave4.nxv64i1(<vscale x 16 x i1> [[TMP9]], <vscale x 16 x i1> [[TMP9]], <vscale x 16 x i1> [[TMP9]], <vscale x 16 x i1> [[TMP9]])524; PREDICATED_TAIL_FOLDING-NEXT: [[WIDE_MASKED_VEC:%.*]] = call <vscale x 64 x i8> @llvm.masked.load.nxv64i8.p0(ptr align 1 [[TMP12]], <vscale x 64 x i1> [[INTERLEAVED_MASK]], <vscale x 64 x i8> poison)525; PREDICATED_TAIL_FOLDING-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.vector.deinterleave4.nxv64i8(<vscale x 64 x i8> [[WIDE_MASKED_VEC]])526; PREDICATED_TAIL_FOLDING-NEXT: [[TMP13:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 0527; PREDICATED_TAIL_FOLDING-NEXT: [[TMP14:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 1528; PREDICATED_TAIL_FOLDING-NEXT: [[TMP15:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 2529; PREDICATED_TAIL_FOLDING-NEXT: [[TMP16:%.*]] = extractvalue { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[STRIDED_VEC]], 3530; PREDICATED_TAIL_FOLDING-NEXT: [[TMP17:%.*]] = call <vscale x 16 x i8> @llvm.smax.nxv16i8(<vscale x 16 x i8> [[TMP13]], <vscale x 16 x i8> [[TMP14]])531; PREDICATED_TAIL_FOLDING-NEXT: [[TMP18:%.*]] = sub <vscale x 16 x i8> zeroinitializer, [[TMP17]]532; PREDICATED_TAIL_FOLDING-NEXT: [[TMP19:%.*]] = call <vscale x 16 x i8> @llvm.smax.nxv16i8(<vscale x 16 x i8> [[TMP15]], <vscale x 16 x i8> [[TMP16]])533; PREDICATED_TAIL_FOLDING-NEXT: [[TMP20:%.*]] = sub <vscale x 16 x i8> zeroinitializer, [[TMP19]]534; PREDICATED_TAIL_FOLDING-NEXT: [[TMP21:%.*]] = sext i32 [[TMP10]] to i64535; PREDICATED_TAIL_FOLDING-NEXT: [[TMP22:%.*]] = getelementptr i8, ptr [[Q]], i64 [[TMP21]]536; PREDICATED_TAIL_FOLDING-NEXT: [[INTERLEAVED_VEC:%.*]] = call <vscale x 64 x i8> @llvm.vector.interleave4.nxv64i8(<vscale x 16 x i8> [[TMP17]], <vscale x 16 x i8> [[TMP18]], <vscale x 16 x i8> [[TMP19]], <vscale x 16 x i8> [[TMP20]])537; PREDICATED_TAIL_FOLDING-NEXT: [[INTERLEAVED_MASK3:%.*]] = call <vscale x 64 x i1> @llvm.vector.interleave4.nxv64i1(<vscale x 16 x i1> [[TMP9]], <vscale x 16 x i1> [[TMP9]], <vscale x 16 x i1> [[TMP9]], <vscale x 16 x i1> [[TMP9]])538; PREDICATED_TAIL_FOLDING-NEXT: call void @llvm.masked.store.nxv64i8.p0(<vscale x 64 x i8> [[INTERLEAVED_VEC]], ptr align 1 [[TMP22]], <vscale x 64 x i1> [[INTERLEAVED_MASK3]])539; PREDICATED_TAIL_FOLDING-NEXT: [[INDEX_NEXT]] = add i32 [[INDEX]], [[TMP1]]540; PREDICATED_TAIL_FOLDING-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i32(i32 [[INDEX]], i32 [[TMP6]])541; PREDICATED_TAIL_FOLDING-NEXT: [[TMP23:%.*]] = extractelement <vscale x 16 x i1> [[ACTIVE_LANE_MASK_NEXT]], i64 0542; PREDICATED_TAIL_FOLDING-NEXT: [[VEC_IND_NEXT]] = add <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT2]]543; PREDICATED_TAIL_FOLDING-NEXT: br i1 [[TMP23]], label [[VECTOR_BODY]], label [[MIDDLE_BLOCK:%.*]], !llvm.loop [[LOOP5:![0-9]+]]544; PREDICATED_TAIL_FOLDING: middle.block:545; PREDICATED_TAIL_FOLDING-NEXT: br label [[FOR_END:%.*]]546; PREDICATED_TAIL_FOLDING: for.end:547; PREDICATED_TAIL_FOLDING-NEXT: ret void548;549entry:550 %conv = zext i8 %guard to i32551 br label %for.body552 553for.body:554 %ix.024 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]555 %cmp1 = icmp ugt i32 %ix.024, %conv556 br i1 %cmp1, label %if.then, label %for.inc557 558if.then:559 %idx0 = shl nuw nsw i32 %ix.024, 2560 %idx1 = add i32 %idx0, 1561 %idx2 = add i32 %idx0, 2562 %idx3 = add i32 %idx0, 3563 564 %array1idx0 = getelementptr inbounds i8, ptr %p, i32 %idx0565 %0 = load i8, ptr %array1idx0, align 1566 %array1idx1 = getelementptr inbounds i8, ptr %p, i32 %idx1567 %1 = load i8, ptr %array1idx1, align 1568 %array1idx2 = getelementptr inbounds i8, ptr %p, i32 %idx2569 %2 = load i8, ptr %array1idx2, align 1570 %array1idx3 = getelementptr inbounds i8, ptr %p, i32 %idx3571 %3 = load i8, ptr %array1idx3, align 1572 573 %cmp.i1 = icmp slt i8 %0, %1574 %spec.select.i1 = select i1 %cmp.i1, i8 %1, i8 %0575 %sub1 = sub i8 0, %spec.select.i1576 %cmp.i2 = icmp slt i8 %2, %3577 %spec.select.i2 = select i1 %cmp.i2, i8 %3, i8 %2578 %sub2 = sub i8 0, %spec.select.i2579 580 %array3idx0 = getelementptr inbounds i8, ptr %q, i32 %idx0581 store i8 %spec.select.i1, ptr %array3idx0, align 1582 %array3idx1 = getelementptr inbounds i8, ptr %q, i32 %idx1583 store i8 %sub1, ptr %array3idx1, align 1584 %array3idx2 = getelementptr inbounds i8, ptr %q, i32 %idx2585 store i8 %spec.select.i2, ptr %array3idx2, align 1586 %array3idx3 = getelementptr inbounds i8, ptr %q, i32 %idx3587 store i8 %sub2, ptr %array3idx3, align 1588 589 br label %for.inc590 591for.inc:592 %inc = add nuw nsw i32 %ix.024, 1593 %exitcond = icmp eq i32 %inc, 1024594 br i1 %exitcond, label %for.end, label %for.body595 596for.end:597 ret void598}599