1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -passes=loop-vectorize,dce -force-vector-interleave=1 -force-vector-width=4 -S | FileCheck %s 3 4target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128" 5target triple = "x86_64-unknown-linux-gnu" 6 7; PR15794 8; incorrect addition of llvm.mem.parallel_loop_access metadata is undefined 9; behaviour. Vectorizer ignores the memory dependency checks and goes ahead and 10; vectorizes this loop with uniform stores which has an output dependency. 11 12; void foo(int *a, int *b, int k, int m) { 13; for (int i = 0; i < m; i++) { 14; for (int j = 0; j < m; j++) { 15; a[i] = a[i + j + k] + 1; <<< 16; } 17; b[i] = b[i] + 3; 18; } 19; } 20 21; Function Attrs: nounwind uwtable 22define void @foo(ptr nocapture %a, ptr nocapture %b, i32 %k, i32 %m) #0 { 23; CHECK-LABEL: @foo( 24; CHECK-NEXT: entry: 25; CHECK-NEXT: [[CMP27:%.*]] = icmp sgt i32 [[M:%.*]], 0 26; CHECK-NEXT: br i1 [[CMP27]], label [[FOR_BODY3_LR_PH_US_PREHEADER:%.*]], label [[FOR_END15:%.*]] 27; CHECK: for.body3.lr.ph.us.preheader: 28; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[M]], -1 29; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[K:%.*]] to i64 30; CHECK-NEXT: [[TMP2:%.*]] = zext i32 [[M]] to i64 31; CHECK-NEXT: br label [[FOR_BODY3_LR_PH_US:%.*]] 32; CHECK: for.end.us: 33; CHECK-NEXT: [[ARRAYIDX9_US:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV33:%.*]] 34; CHECK-NEXT: [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX9_US]], align 4, !llvm.mem.parallel_loop_access [[META0:![0-9]+]] 35; CHECK-NEXT: [[ADD10_US:%.*]] = add nsw i32 [[TMP3]], 3 36; CHECK-NEXT: store i32 [[ADD10_US]], ptr [[ARRAYIDX9_US]], align 4, !llvm.mem.parallel_loop_access [[META0]] 37; CHECK-NEXT: [[INDVARS_IV_NEXT34:%.*]] = add i64 [[INDVARS_IV33]], 1 38; CHECK-NEXT: [[LFTR_WIDEIV35:%.*]] = trunc i64 [[INDVARS_IV_NEXT34]] to i32 39; CHECK-NEXT: [[EXITCOND36:%.*]] = icmp eq i32 [[LFTR_WIDEIV35]], [[M]] 40; CHECK-NEXT: br i1 [[EXITCOND36]], label [[FOR_END15_LOOPEXIT:%.*]], label [[FOR_BODY3_LR_PH_US]], !llvm.loop [[LOOP2:![0-9]+]] 41; CHECK: for.body3.us: 42; CHECK-NEXT: [[INDVARS_IV29:%.*]] = phi i64 [ [[BC_RESUME_VAL:%.*]], [[SCALAR_PH:%.*]] ], [ [[INDVARS_IV_NEXT30:%.*]], [[FOR_BODY3_US:%.*]] ] 43; CHECK-NEXT: [[TMP4:%.*]] = trunc i64 [[INDVARS_IV29]] to i32 44; CHECK-NEXT: [[ADD4_US:%.*]] = add i32 [[ADD_US:%.*]], [[TMP4]] 45; CHECK-NEXT: [[IDXPROM_US:%.*]] = sext i32 [[ADD4_US]] to i64 46; CHECK-NEXT: [[ARRAYIDX_US:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[IDXPROM_US]] 47; CHECK-NEXT: [[TMP5:%.*]] = load i32, ptr [[ARRAYIDX_US]], align 4, !llvm.mem.parallel_loop_access [[META0]] 48; CHECK-NEXT: [[ADD5_US:%.*]] = add nsw i32 [[TMP5]], 1 49; CHECK-NEXT: store i32 [[ADD5_US]], ptr [[ARRAYIDX7_US:%.*]], align 4, !llvm.mem.parallel_loop_access [[META0]] 50; CHECK-NEXT: [[INDVARS_IV_NEXT30]] = add i64 [[INDVARS_IV29]], 1 51; CHECK-NEXT: [[LFTR_WIDEIV31:%.*]] = trunc i64 [[INDVARS_IV_NEXT30]] to i32 52; CHECK-NEXT: [[EXITCOND32:%.*]] = icmp eq i32 [[LFTR_WIDEIV31]], [[M]] 53; CHECK-NEXT: br i1 [[EXITCOND32]], label [[FOR_END_US:%.*]], label [[FOR_BODY3_US]], !llvm.loop [[LOOP3:![0-9]+]] 54; CHECK: for.body3.lr.ph.us: 55; CHECK-NEXT: [[INDVARS_IV33]] = phi i64 [ [[INDVARS_IV_NEXT34]], [[FOR_END_US]] ], [ 0, [[FOR_BODY3_LR_PH_US_PREHEADER]] ] 56; CHECK-NEXT: [[TMP6:%.*]] = add i64 [[TMP1]], [[INDVARS_IV33]] 57; CHECK-NEXT: [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32 58; CHECK-NEXT: [[TMP8:%.*]] = trunc i64 [[INDVARS_IV33]] to i32 59; CHECK-NEXT: [[ADD_US]] = add i32 [[TMP8]], [[K]] 60; CHECK-NEXT: [[ARRAYIDX7_US]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV33]] 61; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 4 62; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH]], label [[VECTOR_SCEVCHECK:%.*]] 63; CHECK: vector.scevcheck: 64; CHECK-NEXT: [[TMP9:%.*]] = add i32 [[TMP7]], [[TMP0]] 65; CHECK-NEXT: [[TMP10:%.*]] = icmp slt i32 [[TMP9]], [[TMP7]] 66; CHECK-NEXT: br i1 [[TMP10]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 67; CHECK: vector.ph: 68; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 4 69; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]] 70; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 71; CHECK: vector.body: 72; CHECK-NEXT: [[OFFSET_IDX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 73; CHECK-NEXT: [[TMP11:%.*]] = trunc i64 [[OFFSET_IDX]] to i32 74; CHECK-NEXT: [[TMP12:%.*]] = add i32 [[TMP11]], 0 75; CHECK-NEXT: [[TMP13:%.*]] = add i32 [[ADD_US]], [[TMP12]] 76; CHECK-NEXT: [[TMP14:%.*]] = sext i32 [[TMP13]] to i64 77; CHECK-NEXT: [[TMP15:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP14]] 78; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[TMP15]], i32 0 79; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP16]], align 4 80; CHECK-NEXT: [[TMP17:%.*]] = add nsw <4 x i32> [[WIDE_LOAD]], splat (i32 1) 81; CHECK-NEXT: [[TMP18:%.*]] = extractelement <4 x i32> [[TMP17]], i32 3 82; CHECK-NEXT: store i32 [[TMP18]], ptr [[ARRAYIDX7_US]], align 4, !llvm.mem.parallel_loop_access [[META0]] 83; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[OFFSET_IDX]], 4 84; CHECK-NEXT: [[TMP19:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 85; CHECK-NEXT: br i1 [[TMP19]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]] 86; CHECK: middle.block: 87; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]] 88; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_END_US]], label [[SCALAR_PH]] 89; CHECK: scalar.ph: 90; CHECK-NEXT: [[BC_RESUME_VAL]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[VECTOR_SCEVCHECK]] ], [ 0, [[FOR_BODY3_LR_PH_US]] ] 91; CHECK-NEXT: br label [[FOR_BODY3_US]] 92; CHECK: for.end15.loopexit: 93; CHECK-NEXT: br label [[FOR_END15]] 94; CHECK: for.end15: 95; CHECK-NEXT: ret void 96; 97entry: 98 %cmp27 = icmp sgt i32 %m, 0 99 br i1 %cmp27, label %for.body3.lr.ph.us, label %for.end15 100 101for.end.us: ; preds = %for.body3.us 102 %arrayidx9.us = getelementptr inbounds i32, ptr %b, i64 %indvars.iv33 103 %0 = load i32, ptr %arrayidx9.us, align 4, !llvm.mem.parallel_loop_access !3 104 %add10.us = add nsw i32 %0, 3 105 store i32 %add10.us, ptr %arrayidx9.us, align 4, !llvm.mem.parallel_loop_access !3 106 %indvars.iv.next34 = add i64 %indvars.iv33, 1 107 %lftr.wideiv35 = trunc i64 %indvars.iv.next34 to i32 108 %exitcond36 = icmp eq i32 %lftr.wideiv35, %m 109 br i1 %exitcond36, label %for.end15, label %for.body3.lr.ph.us, !llvm.loop !5 110 111for.body3.us: ; preds = %for.body3.us, %for.body3.lr.ph.us 112 %indvars.iv29 = phi i64 [ 0, %for.body3.lr.ph.us ], [ %indvars.iv.next30, %for.body3.us ] 113 %1 = trunc i64 %indvars.iv29 to i32 114 %add4.us = add i32 %add.us, %1 115 %idxprom.us = sext i32 %add4.us to i64 116 %arrayidx.us = getelementptr inbounds i32, ptr %a, i64 %idxprom.us 117 %2 = load i32, ptr %arrayidx.us, align 4, !llvm.mem.parallel_loop_access !3 118 %add5.us = add nsw i32 %2, 1 119 store i32 %add5.us, ptr %arrayidx7.us, align 4, !llvm.mem.parallel_loop_access !3 120 %indvars.iv.next30 = add i64 %indvars.iv29, 1 121 %lftr.wideiv31 = trunc i64 %indvars.iv.next30 to i32 122 %exitcond32 = icmp eq i32 %lftr.wideiv31, %m 123 br i1 %exitcond32, label %for.end.us, label %for.body3.us, !llvm.loop !4 124 125for.body3.lr.ph.us: ; preds = %for.end.us, %entry 126 %indvars.iv33 = phi i64 [ %indvars.iv.next34, %for.end.us ], [ 0, %entry ] 127 %3 = trunc i64 %indvars.iv33 to i32 128 %add.us = add i32 %3, %k 129 %arrayidx7.us = getelementptr inbounds i32, ptr %a, i64 %indvars.iv33 130 br label %for.body3.us 131 132for.end15: ; preds = %for.end.us, %entry 133 ret void 134} 135 136; Same test as above, but without the invalid parallel_loop_access metadata. 137 138; Here we can see the vectorizer does the mem dep checks and decides it is 139; unsafe to vectorize. 140define void @no-par-mem-metadata(ptr nocapture %a, ptr nocapture %b, i32 %k, i32 %m) #0 { 141; CHECK-LABEL: @no-par-mem-metadata( 142; CHECK-NEXT: entry: 143; CHECK-NEXT: [[CMP27:%.*]] = icmp sgt i32 [[M:%.*]], 0 144; CHECK-NEXT: br i1 [[CMP27]], label [[FOR_BODY3_LR_PH_US_PREHEADER:%.*]], label [[FOR_END15:%.*]] 145; CHECK: for.body3.lr.ph.us.preheader: 146; CHECK-NEXT: br label [[FOR_BODY3_LR_PH_US:%.*]] 147; CHECK: for.end.us: 148; CHECK-NEXT: [[ARRAYIDX9_US:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV33:%.*]] 149; CHECK-NEXT: [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX9_US]], align 4 150; CHECK-NEXT: [[ADD10_US:%.*]] = add nsw i32 [[TMP0]], 3 151; CHECK-NEXT: store i32 [[ADD10_US]], ptr [[ARRAYIDX9_US]], align 4 152; CHECK-NEXT: [[INDVARS_IV_NEXT34:%.*]] = add i64 [[INDVARS_IV33]], 1 153; CHECK-NEXT: [[LFTR_WIDEIV35:%.*]] = trunc i64 [[INDVARS_IV_NEXT34]] to i32 154; CHECK-NEXT: [[EXITCOND36:%.*]] = icmp eq i32 [[LFTR_WIDEIV35]], [[M]] 155; CHECK-NEXT: br i1 [[EXITCOND36]], label [[FOR_END15_LOOPEXIT:%.*]], label [[FOR_BODY3_LR_PH_US]], !llvm.loop [[LOOP2]] 156; CHECK: for.body3.us: 157; CHECK-NEXT: [[INDVARS_IV29:%.*]] = phi i64 [ 0, [[FOR_BODY3_LR_PH_US]] ], [ [[INDVARS_IV_NEXT30:%.*]], [[FOR_BODY3_US:%.*]] ] 158; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[INDVARS_IV29]] to i32 159; CHECK-NEXT: [[ADD4_US:%.*]] = add i32 [[ADD_US:%.*]], [[TMP1]] 160; CHECK-NEXT: [[IDXPROM_US:%.*]] = sext i32 [[ADD4_US]] to i64 161; CHECK-NEXT: [[ARRAYIDX_US:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[IDXPROM_US]] 162; CHECK-NEXT: [[TMP2:%.*]] = load i32, ptr [[ARRAYIDX_US]], align 4 163; CHECK-NEXT: [[ADD5_US:%.*]] = add nsw i32 [[TMP2]], 1 164; CHECK-NEXT: store i32 [[ADD5_US]], ptr [[ARRAYIDX7_US:%.*]], align 4 165; CHECK-NEXT: [[INDVARS_IV_NEXT30]] = add i64 [[INDVARS_IV29]], 1 166; CHECK-NEXT: [[LFTR_WIDEIV31:%.*]] = trunc i64 [[INDVARS_IV_NEXT30]] to i32 167; CHECK-NEXT: [[EXITCOND32:%.*]] = icmp eq i32 [[LFTR_WIDEIV31]], [[M]] 168; CHECK-NEXT: br i1 [[EXITCOND32]], label [[FOR_END_US:%.*]], label [[FOR_BODY3_US]], !llvm.loop [[LOOP1:![0-9]+]] 169; CHECK: for.body3.lr.ph.us: 170; CHECK-NEXT: [[INDVARS_IV33]] = phi i64 [ [[INDVARS_IV_NEXT34]], [[FOR_END_US]] ], [ 0, [[FOR_BODY3_LR_PH_US_PREHEADER]] ] 171; CHECK-NEXT: [[TMP3:%.*]] = trunc i64 [[INDVARS_IV33]] to i32 172; CHECK-NEXT: [[ADD_US]] = add i32 [[TMP3]], [[K:%.*]] 173; CHECK-NEXT: [[ARRAYIDX7_US]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV33]] 174; CHECK-NEXT: br label [[FOR_BODY3_US]] 175; CHECK: for.end15.loopexit: 176; CHECK-NEXT: br label [[FOR_END15]] 177; CHECK: for.end15: 178; CHECK-NEXT: ret void 179; 180entry: 181 %cmp27 = icmp sgt i32 %m, 0 182 br i1 %cmp27, label %for.body3.lr.ph.us, label %for.end15 183 184for.end.us: ; preds = %for.body3.us 185 %arrayidx9.us = getelementptr inbounds i32, ptr %b, i64 %indvars.iv33 186 %0 = load i32, ptr %arrayidx9.us, align 4 187 %add10.us = add nsw i32 %0, 3 188 store i32 %add10.us, ptr %arrayidx9.us, align 4 189 %indvars.iv.next34 = add i64 %indvars.iv33, 1 190 %lftr.wideiv35 = trunc i64 %indvars.iv.next34 to i32 191 %exitcond36 = icmp eq i32 %lftr.wideiv35, %m 192 br i1 %exitcond36, label %for.end15, label %for.body3.lr.ph.us, !llvm.loop !5 193 194for.body3.us: ; preds = %for.body3.us, %for.body3.lr.ph.us 195 %indvars.iv29 = phi i64 [ 0, %for.body3.lr.ph.us ], [ %indvars.iv.next30, %for.body3.us ] 196 %1 = trunc i64 %indvars.iv29 to i32 197 %add4.us = add i32 %add.us, %1 198 %idxprom.us = sext i32 %add4.us to i64 199 %arrayidx.us = getelementptr inbounds i32, ptr %a, i64 %idxprom.us 200 %2 = load i32, ptr %arrayidx.us, align 4 201 %add5.us = add nsw i32 %2, 1 202 store i32 %add5.us, ptr %arrayidx7.us, align 4 203 %indvars.iv.next30 = add i64 %indvars.iv29, 1 204 %lftr.wideiv31 = trunc i64 %indvars.iv.next30 to i32 205 %exitcond32 = icmp eq i32 %lftr.wideiv31, %m 206 br i1 %exitcond32, label %for.end.us, label %for.body3.us, !llvm.loop !4 207 208for.body3.lr.ph.us: ; preds = %for.end.us, %entry 209 %indvars.iv33 = phi i64 [ %indvars.iv.next34, %for.end.us ], [ 0, %entry ] 210 %3 = trunc i64 %indvars.iv33 to i32 211 %add.us = add i32 %3, %k 212 %arrayidx7.us = getelementptr inbounds i32, ptr %a, i64 %indvars.iv33 213 br label %for.body3.us 214 215for.end15: ; preds = %for.end.us, %entry 216 ret void 217} 218 219attributes #0 = { nounwind uwtable "less-precise-fpmad"="false" "frame-pointer"="none" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "unsafe-fp-math"="false" "use-soft-float"="false" } 220 221!3 = !{!4, !5} 222!4 = !{!4} 223!5 = !{!5} 224 225