453 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -S -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=1 < %s 2>&1 | FileCheck %s3 4target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"5 6; Check that the vectorizer identifies the %p.09 phi,7; as an induction variable, despite the potential overflow8; due to the truncation from 32bit to 8bit.9; SCEV will detect the pattern "sext(trunc(%p.09)) + %step"10; and generate the required runtime checks under which11; we can assume no overflow. We check here that we generate12; exactly two runtime checks:13; 1) an overflow check:14; {0,+,(trunc i32 %step to i8)}<%for.body> Added Flags: <nssw>15; 2) an equality check verifying that the step of the induction16; is equal to sext(trunc(step)):17; Equal predicate: %step == (sext i8 (trunc i32 %step to i8) to i32)18;19; See also pr30654.20;21; int a[N];22; void doit1(int n, int step) {23; int i;24; char p = 0;25; for (i = 0; i < n; i++) {26; a[i] = p;27; p = p + step;28; }29; }30;31 32@a = common local_unnamed_addr global [250 x i32] zeroinitializer, align 1633 34; Function Attrs: norecurse nounwind uwtable35define void @doit1(i32 %n, i32 %step) local_unnamed_addr {36; CHECK-LABEL: @doit1(37; CHECK-NEXT: entry:38; CHECK-NEXT: [[CMP7:%.*]] = icmp sgt i32 [[N:%.*]], 039; CHECK-NEXT: br i1 [[CMP7]], label [[FOR_BODY_PREHEADER:%.*]], label [[FOR_END:%.*]]40; CHECK: for.body.preheader:41; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i6442; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 443; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]]44; CHECK: vector.scevcheck:45; CHECK-NEXT: [[TMP0:%.*]] = add nsw i64 [[WIDE_TRIP_COUNT]], -146; CHECK-NEXT: [[TMP1:%.*]] = trunc i32 [[STEP:%.*]] to i847; CHECK-NEXT: [[TMP2:%.*]] = sub i8 0, [[TMP1]]48; CHECK-NEXT: [[TMP3:%.*]] = icmp slt i8 [[TMP1]], 049; CHECK-NEXT: [[TMP4:%.*]] = select i1 [[TMP3]], i8 [[TMP2]], i8 [[TMP1]]50; CHECK-NEXT: [[TMP5:%.*]] = trunc i64 [[TMP0]] to i851; CHECK-NEXT: [[MUL:%.*]] = call { i8, i1 } @llvm.umul.with.overflow.i8(i8 [[TMP4]], i8 [[TMP5]])52; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i8, i1 } [[MUL]], 053; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i8, i1 } [[MUL]], 154; CHECK-NEXT: [[TMP6:%.*]] = sub i8 0, [[MUL_RESULT]]55; CHECK-NEXT: [[TMP7:%.*]] = icmp slt i8 [[MUL_RESULT]], 056; CHECK-NEXT: [[TMP8:%.*]] = icmp sgt i8 [[TMP6]], 057; CHECK-NEXT: [[TMP9:%.*]] = select i1 [[TMP3]], i1 [[TMP8]], i1 [[TMP7]]58; CHECK-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[MUL_OVERFLOW]]59; CHECK-NEXT: [[TMP11:%.*]] = icmp ugt i64 [[TMP0]], 25560; CHECK-NEXT: [[TMP12:%.*]] = icmp ne i8 [[TMP1]], 061; CHECK-NEXT: [[TMP13:%.*]] = and i1 [[TMP11]], [[TMP12]]62; CHECK-NEXT: [[TMP14:%.*]] = or i1 [[TMP10]], [[TMP13]]63; CHECK-NEXT: [[TMP15:%.*]] = sext i8 [[TMP1]] to i3264; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[STEP]], [[TMP15]]65; CHECK-NEXT: [[TMP16:%.*]] = or i1 [[TMP14]], [[IDENT_CHECK]]66; CHECK-NEXT: br i1 [[TMP16]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]67; CHECK: vector.ph:68; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 469; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]70; CHECK-NEXT: [[DOTCAST:%.*]] = trunc i64 [[N_VEC]] to i3271; CHECK-NEXT: [[IND_END:%.*]] = mul i32 [[DOTCAST]], [[STEP]]72; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[STEP]], i64 073; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <4 x i32> [[DOTSPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer74; CHECK-NEXT: [[TMP19:%.*]] = mul nsw <4 x i32> <i32 0, i32 1, i32 2, i32 3>, [[DOTSPLAT]]75; CHECK-NEXT: [[INDUCTION:%.*]] = add nsw <4 x i32> zeroinitializer, [[TMP19]]76; CHECK-NEXT: [[TMP18:%.*]] = mul nsw i32 [[STEP]], 477; CHECK-NEXT: [[DOTSPLATINSERT2:%.*]] = insertelement <4 x i32> poison, i32 [[TMP18]], i64 078; CHECK-NEXT: [[DOTSPLAT3:%.*]] = shufflevector <4 x i32> [[DOTSPLATINSERT2]], <4 x i32> poison, <4 x i32> zeroinitializer79; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]80; CHECK: vector.body:81; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]82; CHECK-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]83; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 [[INDEX]]84; CHECK-NEXT: store <4 x i32> [[VEC_IND]], ptr [[TMP20]], align 485; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 486; CHECK-NEXT: [[VEC_IND_NEXT]] = add nsw <4 x i32> [[VEC_IND]], [[DOTSPLAT3]]87; CHECK-NEXT: [[TMP22:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]88; CHECK-NEXT: br i1 [[TMP22]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]89; CHECK: middle.block:90; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]91; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]92; CHECK: scalar.ph:93; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_PREHEADER]] ], [ 0, [[VECTOR_SCEVCHECK]] ]94; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_PREHEADER]] ], [ 0, [[VECTOR_SCEVCHECK]] ]95; CHECK-NEXT: br label [[FOR_BODY:%.*]]96; CHECK: for.body:97; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]98; CHECK-NEXT: [[P_09:%.*]] = phi i32 [ [[ADD:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]99; CHECK-NEXT: [[SEXT:%.*]] = shl i32 [[P_09]], 24100; CHECK-NEXT: [[CONV:%.*]] = ashr exact i32 [[SEXT]], 24101; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 [[INDVARS_IV]]102; CHECK-NEXT: store i32 [[CONV]], ptr [[ARRAYIDX]], align 4103; CHECK-NEXT: [[ADD]] = add nsw i32 [[CONV]], [[STEP]]104; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1105; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]106; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]107; CHECK: for.end.loopexit:108; CHECK-NEXT: br label [[FOR_END]]109; CHECK: for.end:110; CHECK-NEXT: ret void111;112entry:113 %cmp7 = icmp sgt i32 %n, 0114 br i1 %cmp7, label %for.body.preheader, label %for.end115 116for.body.preheader:117 %wide.trip.count = zext i32 %n to i64118 br label %for.body119 120for.body:121 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]122 %p.09 = phi i32 [ %add, %for.body ], [ 0, %for.body.preheader ]123 %sext = shl i32 %p.09, 24124 %conv = ashr exact i32 %sext, 24125 %arrayidx = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 %indvars.iv126 store i32 %conv, ptr %arrayidx, align 4127 %add = add nsw i32 %conv, %step128 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1129 %exitcond = icmp eq i64 %indvars.iv.next, %wide.trip.count130 br i1 %exitcond, label %for.end.loopexit, label %for.body131 132for.end.loopexit:133 br label %for.end134 135for.end:136 ret void137}138 139; Same as above, but for checking the SCEV "zext(trunc(%p.09)) + %step".140; Here we expect the following two predicates to be added for runtime checking:141; 1) {0,+,(trunc i32 %step to i8)}<%for.body> Added Flags: <nusw>142; 2) Equal predicate: %step == (sext i8 (trunc i32 %step to i8) to i32)143;144; int a[N];145; void doit2(int n, int step) {146; int i;147; unsigned char p = 0;148; for (i = 0; i < n; i++) {149; a[i] = p;150; p = p + step;151; }152; }153;154 155 156; Function Attrs: norecurse nounwind uwtable157define void @doit2(i32 %n, i32 %step) local_unnamed_addr {158; CHECK-LABEL: @doit2(159; CHECK-NEXT: entry:160; CHECK-NEXT: [[CMP7:%.*]] = icmp sgt i32 [[N:%.*]], 0161; CHECK-NEXT: br i1 [[CMP7]], label [[FOR_BODY_PREHEADER:%.*]], label [[FOR_END:%.*]]162; CHECK: for.body.preheader:163; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64164; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4165; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]]166; CHECK: vector.scevcheck:167; CHECK-NEXT: [[TMP0:%.*]] = add nsw i64 [[WIDE_TRIP_COUNT]], -1168; CHECK-NEXT: [[TMP1:%.*]] = trunc i32 [[STEP:%.*]] to i8169; CHECK-NEXT: [[TMP2:%.*]] = sub i8 0, [[TMP1]]170; CHECK-NEXT: [[TMP3:%.*]] = icmp slt i8 [[TMP1]], 0171; CHECK-NEXT: [[TMP4:%.*]] = select i1 [[TMP3]], i8 [[TMP2]], i8 [[TMP1]]172; CHECK-NEXT: [[TMP5:%.*]] = trunc i64 [[TMP0]] to i8173; CHECK-NEXT: [[MUL:%.*]] = call { i8, i1 } @llvm.umul.with.overflow.i8(i8 [[TMP4]], i8 [[TMP5]])174; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i8, i1 } [[MUL]], 0175; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i8, i1 } [[MUL]], 1176; CHECK-NEXT: [[TMP6:%.*]] = sub i8 0, [[MUL_RESULT]]177; CHECK-NEXT: [[TMP7:%.*]] = icmp ugt i8 [[TMP6]], 0178; CHECK-NEXT: [[TMP8:%.*]] = select i1 [[TMP3]], i1 [[TMP7]], i1 false179; CHECK-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[MUL_OVERFLOW]]180; CHECK-NEXT: [[TMP10:%.*]] = icmp ugt i64 [[TMP0]], 255181; CHECK-NEXT: [[TMP11:%.*]] = icmp ne i8 [[TMP1]], 0182; CHECK-NEXT: [[TMP12:%.*]] = and i1 [[TMP10]], [[TMP11]]183; CHECK-NEXT: [[TMP13:%.*]] = or i1 [[TMP9]], [[TMP12]]184; CHECK-NEXT: [[TMP14:%.*]] = sext i8 [[TMP1]] to i32185; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[STEP]], [[TMP14]]186; CHECK-NEXT: [[TMP15:%.*]] = or i1 [[TMP13]], [[IDENT_CHECK]]187; CHECK-NEXT: br i1 [[TMP15]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]188; CHECK: vector.ph:189; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4190; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]191; CHECK-NEXT: [[DOTCAST:%.*]] = trunc i64 [[N_VEC]] to i32192; CHECK-NEXT: [[IND_END:%.*]] = mul i32 [[DOTCAST]], [[STEP]]193; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[STEP]], i64 0194; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <4 x i32> [[DOTSPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer195; CHECK-NEXT: [[TMP18:%.*]] = mul nsw <4 x i32> <i32 0, i32 1, i32 2, i32 3>, [[DOTSPLAT]]196; CHECK-NEXT: [[INDUCTION:%.*]] = add nsw <4 x i32> zeroinitializer, [[TMP18]]197; CHECK-NEXT: [[TMP17:%.*]] = mul nsw i32 [[STEP]], 4198; CHECK-NEXT: [[DOTSPLATINSERT2:%.*]] = insertelement <4 x i32> poison, i32 [[TMP17]], i64 0199; CHECK-NEXT: [[DOTSPLAT3:%.*]] = shufflevector <4 x i32> [[DOTSPLATINSERT2]], <4 x i32> poison, <4 x i32> zeroinitializer200; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]201; CHECK: vector.body:202; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]203; CHECK-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]204; CHECK-NEXT: [[TMP19:%.*]] = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 [[INDEX]]205; CHECK-NEXT: store <4 x i32> [[VEC_IND]], ptr [[TMP19]], align 4206; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4207; CHECK-NEXT: [[VEC_IND_NEXT]] = add nsw <4 x i32> [[VEC_IND]], [[DOTSPLAT3]]208; CHECK-NEXT: [[TMP21:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]209; CHECK-NEXT: br i1 [[TMP21]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]210; CHECK: middle.block:211; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]212; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]213; CHECK: scalar.ph:214; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_PREHEADER]] ], [ 0, [[VECTOR_SCEVCHECK]] ]215; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_PREHEADER]] ], [ 0, [[VECTOR_SCEVCHECK]] ]216; CHECK-NEXT: br label [[FOR_BODY:%.*]]217; CHECK: for.body:218; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]219; CHECK-NEXT: [[P_09:%.*]] = phi i32 [ [[ADD:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]220; CHECK-NEXT: [[CONV:%.*]] = and i32 [[P_09]], 255221; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 [[INDVARS_IV]]222; CHECK-NEXT: store i32 [[CONV]], ptr [[ARRAYIDX]], align 4223; CHECK-NEXT: [[ADD]] = add nsw i32 [[CONV]], [[STEP]]224; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1225; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]226; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]227; CHECK: for.end.loopexit:228; CHECK-NEXT: br label [[FOR_END]]229; CHECK: for.end:230; CHECK-NEXT: ret void231;232entry:233 %cmp7 = icmp sgt i32 %n, 0234 br i1 %cmp7, label %for.body.preheader, label %for.end235 236for.body.preheader:237 %wide.trip.count = zext i32 %n to i64238 br label %for.body239 240for.body:241 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]242 %p.09 = phi i32 [ %add, %for.body ], [ 0, %for.body.preheader ]243 %conv = and i32 %p.09, 255244 %arrayidx = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 %indvars.iv245 store i32 %conv, ptr %arrayidx, align 4246 %add = add nsw i32 %conv, %step247 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1248 %exitcond = icmp eq i64 %indvars.iv.next, %wide.trip.count249 br i1 %exitcond, label %for.end.loopexit, label %for.body250 251for.end.loopexit:252 br label %for.end253 254for.end:255 ret void256}257 258; Here we check that the same phi scev analysis would fail259; to create the runtime checks because the step is not invariant.260; As a result vectorization will fail.261;262; int a[N];263; void doit3(int n, int step) {264; int i;265; char p = 0;266; for (i = 0; i < n; i++) {267; a[i] = p;268; p = p + step;269; step += 2;270; }271; }272;273 274 275; Function Attrs: norecurse nounwind uwtable276define void @doit3(i32 %n, i32 %step) local_unnamed_addr {277; CHECK-LABEL: @doit3(278; CHECK-NEXT: entry:279; CHECK-NEXT: [[CMP9:%.*]] = icmp sgt i32 [[N:%.*]], 0280; CHECK-NEXT: br i1 [[CMP9]], label [[FOR_BODY_PREHEADER:%.*]], label [[FOR_END:%.*]]281; CHECK: for.body.preheader:282; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64283; CHECK-NEXT: br label [[FOR_BODY:%.*]]284; CHECK: for.body:285; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_BODY_PREHEADER]] ]286; CHECK-NEXT: [[P_012:%.*]] = phi i32 [ [[ADD:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_BODY_PREHEADER]] ]287; CHECK-NEXT: [[STEP_ADDR_010:%.*]] = phi i32 [ [[ADD3:%.*]], [[FOR_BODY]] ], [ [[STEP:%.*]], [[FOR_BODY_PREHEADER]] ]288; CHECK-NEXT: [[SEXT:%.*]] = shl i32 [[P_012]], 24289; CHECK-NEXT: [[CONV:%.*]] = ashr exact i32 [[SEXT]], 24290; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 [[INDVARS_IV]]291; CHECK-NEXT: store i32 [[CONV]], ptr [[ARRAYIDX]], align 4292; CHECK-NEXT: [[ADD]] = add nsw i32 [[CONV]], [[STEP_ADDR_010]]293; CHECK-NEXT: [[ADD3]] = add nsw i32 [[STEP_ADDR_010]], 2294; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1295; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]296; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY]]297; CHECK: for.end.loopexit:298; CHECK-NEXT: br label [[FOR_END]]299; CHECK: for.end:300; CHECK-NEXT: ret void301;302entry:303 %cmp9 = icmp sgt i32 %n, 0304 br i1 %cmp9, label %for.body.preheader, label %for.end305 306for.body.preheader:307 %wide.trip.count = zext i32 %n to i64308 br label %for.body309 310for.body:311 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]312 %p.012 = phi i32 [ %add, %for.body ], [ 0, %for.body.preheader ]313 %step.addr.010 = phi i32 [ %add3, %for.body ], [ %step, %for.body.preheader ]314 %sext = shl i32 %p.012, 24315 %conv = ashr exact i32 %sext, 24316 %arrayidx = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 %indvars.iv317 store i32 %conv, ptr %arrayidx, align 4318 %add = add nsw i32 %conv, %step.addr.010319 %add3 = add nsw i32 %step.addr.010, 2320 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1321 %exitcond = icmp eq i64 %indvars.iv.next, %wide.trip.count322 br i1 %exitcond, label %for.end.loopexit, label %for.body323 324for.end.loopexit:325 br label %for.end326 327for.end:328 ret void329}330 331 332; Lastly, we also check the case where we can tell at compile time that333; the step of the induction is equal to sext(trunc(step)), in which case334; we don't have to check this equality at runtime (we only need the335; runtime overflow check). Therefore only the following overflow predicate336; will be added for runtime checking:337; {0,+,%cstep}<%for.body> Added Flags: <nssw>338;339; a[N];340; void doit4(int n, char cstep) {341; int i;342; char p = 0;343; int istep = cstep;344; for (i = 0; i < n; i++) {345; a[i] = p;346; p = p + istep;347; }348; }349 350 351; Function Attrs: norecurse nounwind uwtable352define void @doit4(i32 %n, i8 signext %cstep) local_unnamed_addr {353; CHECK-LABEL: @doit4(354; CHECK-NEXT: entry:355; CHECK-NEXT: [[CONV:%.*]] = sext i8 [[CSTEP:%.*]] to i32356; CHECK-NEXT: [[CMP10:%.*]] = icmp sgt i32 [[N:%.*]], 0357; CHECK-NEXT: br i1 [[CMP10]], label [[FOR_BODY_PREHEADER:%.*]], label [[FOR_END:%.*]]358; CHECK: for.body.preheader:359; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64360; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4361; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]]362; CHECK: vector.scevcheck:363; CHECK-NEXT: [[TMP0:%.*]] = add nsw i64 [[WIDE_TRIP_COUNT]], -1364; CHECK-NEXT: [[TMP1:%.*]] = sub i8 0, [[CSTEP]]365; CHECK-NEXT: [[TMP2:%.*]] = icmp slt i8 [[CSTEP]], 0366; CHECK-NEXT: [[TMP3:%.*]] = select i1 [[TMP2]], i8 [[TMP1]], i8 [[CSTEP]]367; CHECK-NEXT: [[TMP4:%.*]] = trunc i64 [[TMP0]] to i8368; CHECK-NEXT: [[MUL:%.*]] = call { i8, i1 } @llvm.umul.with.overflow.i8(i8 [[TMP3]], i8 [[TMP4]])369; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i8, i1 } [[MUL]], 0370; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i8, i1 } [[MUL]], 1371; CHECK-NEXT: [[TMP5:%.*]] = sub i8 0, [[MUL_RESULT]]372; CHECK-NEXT: [[TMP6:%.*]] = icmp slt i8 [[MUL_RESULT]], 0373; CHECK-NEXT: [[TMP7:%.*]] = icmp sgt i8 [[TMP5]], 0374; CHECK-NEXT: [[TMP8:%.*]] = select i1 [[TMP2]], i1 [[TMP7]], i1 [[TMP6]]375; CHECK-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[MUL_OVERFLOW]]376; CHECK-NEXT: [[TMP10:%.*]] = icmp ugt i64 [[TMP0]], 255377; CHECK-NEXT: [[TMP11:%.*]] = icmp ne i8 [[CSTEP]], 0378; CHECK-NEXT: [[TMP12:%.*]] = and i1 [[TMP10]], [[TMP11]]379; CHECK-NEXT: [[TMP13:%.*]] = or i1 [[TMP9]], [[TMP12]]380; CHECK-NEXT: br i1 [[TMP13]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]381; CHECK: vector.ph:382; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4383; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]384; CHECK-NEXT: [[DOTCAST:%.*]] = trunc i64 [[N_VEC]] to i32385; CHECK-NEXT: [[IND_END:%.*]] = mul i32 [[DOTCAST]], [[CONV]]386; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[CONV]], i64 0387; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <4 x i32> [[DOTSPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer388; CHECK-NEXT: [[TMP16:%.*]] = mul nsw <4 x i32> <i32 0, i32 1, i32 2, i32 3>, [[DOTSPLAT]]389; CHECK-NEXT: [[INDUCTION:%.*]] = add nsw <4 x i32> zeroinitializer, [[TMP16]]390; CHECK-NEXT: [[TMP15:%.*]] = mul nsw i32 [[CONV]], 4391; CHECK-NEXT: [[DOTSPLATINSERT2:%.*]] = insertelement <4 x i32> poison, i32 [[TMP15]], i64 0392; CHECK-NEXT: [[DOTSPLAT3:%.*]] = shufflevector <4 x i32> [[DOTSPLATINSERT2]], <4 x i32> poison, <4 x i32> zeroinitializer393; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]394; CHECK: vector.body:395; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]396; CHECK-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]397; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 [[INDEX]]398; CHECK-NEXT: store <4 x i32> [[VEC_IND]], ptr [[TMP17]], align 4399; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4400; CHECK-NEXT: [[VEC_IND_NEXT]] = add nsw <4 x i32> [[VEC_IND]], [[DOTSPLAT3]]401; CHECK-NEXT: [[TMP19:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]402; CHECK-NEXT: br i1 [[TMP19]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]403; CHECK: middle.block:404; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]405; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]406; CHECK: scalar.ph:407; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_PREHEADER]] ], [ 0, [[VECTOR_SCEVCHECK]] ]408; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_PREHEADER]] ], [ 0, [[VECTOR_SCEVCHECK]] ]409; CHECK-NEXT: br label [[FOR_BODY:%.*]]410; CHECK: for.body:411; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]412; CHECK-NEXT: [[P_011:%.*]] = phi i32 [ [[ADD:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]413; CHECK-NEXT: [[SEXT:%.*]] = shl i32 [[P_011]], 24414; CHECK-NEXT: [[CONV2:%.*]] = ashr exact i32 [[SEXT]], 24415; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 [[INDVARS_IV]]416; CHECK-NEXT: store i32 [[CONV2]], ptr [[ARRAYIDX]], align 4417; CHECK-NEXT: [[ADD]] = add nsw i32 [[CONV2]], [[CONV]]418; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1419; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]420; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP7:![0-9]+]]421; CHECK: for.end.loopexit:422; CHECK-NEXT: br label [[FOR_END]]423; CHECK: for.end:424; CHECK-NEXT: ret void425;426entry:427 %conv = sext i8 %cstep to i32428 %cmp10 = icmp sgt i32 %n, 0429 br i1 %cmp10, label %for.body.preheader, label %for.end430 431for.body.preheader:432 %wide.trip.count = zext i32 %n to i64433 br label %for.body434 435for.body:436 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]437 %p.011 = phi i32 [ %add, %for.body ], [ 0, %for.body.preheader ]438 %sext = shl i32 %p.011, 24439 %conv2 = ashr exact i32 %sext, 24440 %arrayidx = getelementptr inbounds [250 x i32], ptr @a, i64 0, i64 %indvars.iv441 store i32 %conv2, ptr %arrayidx, align 4442 %add = add nsw i32 %conv2, %conv443 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1444 %exitcond = icmp eq i64 %indvars.iv.next, %wide.trip.count445 br i1 %exitcond, label %for.end.loopexit, label %for.body446 447for.end.loopexit:448 br label %for.end449 450for.end:451 ret void452}453