1; RUN: opt -loop-accesses -analyze -enable-new-pm=0 %s 2>&1 | FileCheck %s 2; RUN: opt -disable-output -passes='require<scalar-evolution>,require<aa>,loop(print-access-info)' %s 2>&1 | FileCheck %s 3 4target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128" 5 6; CHECK-LABEL: function 'forked_ptrs_different_base_same_offset': 7; CHECK-NEXT: for.body: 8; CHECK-NEXT: Report: cannot identify array bounds 9; CHECK-NEXT: Dependences: 10; CHECK-NEXT: Run-time memory checks: 11; CHECK-NEXT: Grouped accesses: 12; CHECK-EMPTY: 13; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop. 14; CHECK-NEXT: SCEV assumptions: 15; CHECK-EMPTY: 16; CHECK-NEXT: Expressions re-written: 17 18;;;; Derived from the following C code 19;; void forked_ptrs_different_base_same_offset(float *A, float *B, float *C, int *D) { 20;; for (int i=0; i<100; i++) { 21;; if (D[i] != 0) { 22;; C[i] = A[i]; 23;; } else { 24;; C[i] = B[i]; 25;; } 26;; } 27;; } 28 29define dso_local void @forked_ptrs_different_base_same_offset(float* nocapture readonly %Base1, float* nocapture readonly %Base2, float* nocapture %Dest, i32* nocapture readonly %Preds) { 30entry: 31 br label %for.body 32 33for.cond.cleanup: 34 ret void 35 36for.body: 37 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 38 %arrayidx = getelementptr inbounds i32, i32* %Preds, i64 %indvars.iv 39 %0 = load i32, i32* %arrayidx, align 4 40 %cmp1.not = icmp eq i32 %0, 0 41 %spec.select = select i1 %cmp1.not, float* %Base2, float* %Base1 42 %.sink.in = getelementptr inbounds float, float* %spec.select, i64 %indvars.iv 43 %.sink = load float, float* %.sink.in, align 4 44 %1 = getelementptr inbounds float, float* %Dest, i64 %indvars.iv 45 store float %.sink, float* %1, align 4 46 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 47 %exitcond.not = icmp eq i64 %indvars.iv.next, 100 48 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body 49} 50 51; CHECK-LABEL: function 'forked_ptrs_same_base_different_offset': 52; CHECK-NEXT: for.body: 53; CHECK-NEXT: Report: cannot identify array bounds 54; CHECK-NEXT: Dependences: 55; CHECK-NEXT: Run-time memory checks: 56; CHECK-NEXT: Grouped accesses: 57; CHECK-EMPTY: 58; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop. 59; CHECK-NEXT: SCEV assumptions: 60; CHECK-EMPTY: 61; CHECK-NEXT: Expressions re-written: 62 63;;;; Derived from the following C code 64;; void forked_ptrs_same_base_different_offset(float *A, float *B, int *C) { 65;; int offset; 66;; for (int i = 0; i < 100; i++) { 67;; if (C[i] != 0) 68;; offset = i; 69;; else 70;; offset = i+1; 71;; B[i] = A[offset]; 72;; } 73;; } 74 75define dso_local void @forked_ptrs_same_base_different_offset(float* nocapture readonly %Base, float* nocapture %Dest, i32* nocapture readonly %Preds) { 76entry: 77 br label %for.body 78 79for.cond.cleanup: ; preds = %for.body 80 ret void 81 82for.body: ; preds = %entry, %for.body 83 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 84 %i.014 = phi i32 [ 0, %entry ], [ %add, %for.body ] 85 %arrayidx = getelementptr inbounds i32, i32* %Preds, i64 %indvars.iv 86 %0 = load i32, i32* %arrayidx, align 4 87 %cmp1.not = icmp eq i32 %0, 0 88 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 89 %add = add nuw nsw i32 %i.014, 1 90 %1 = trunc i64 %indvars.iv to i32 91 %offset.0 = select i1 %cmp1.not, i32 %add, i32 %1 92 %idxprom213 = zext i32 %offset.0 to i64 93 %arrayidx3 = getelementptr inbounds float, float* %Base, i64 %idxprom213 94 %2 = load float, float* %arrayidx3, align 4 95 %arrayidx5 = getelementptr inbounds float, float* %Dest, i64 %indvars.iv 96 store float %2, float* %arrayidx5, align 4 97 %exitcond.not = icmp eq i64 %indvars.iv.next, 100 98 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body 99} 100 101;;;; Cases that can be handled by a forked pointer but are not currently allowed. 102 103; CHECK-LABEL: function 'forked_ptrs_uniform_and_strided_forks': 104; CHECK-NEXT: for.body: 105; CHECK-NEXT: Report: cannot identify array bounds 106; CHECK-NEXT: Dependences: 107; CHECK-NEXT: Run-time memory checks: 108; CHECK-NEXT: Grouped accesses: 109; CHECK-EMPTY: 110; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop. 111; CHECK-NEXT: SCEV assumptions: 112; CHECK-EMPTY: 113; CHECK-NEXT: Expressions re-written: 114 115;;;; Derived from forked_ptrs_same_base_different_offset with a manually 116;;;; added uniform offset and a mul to provide a stride 117 118define dso_local void @forked_ptrs_uniform_and_strided_forks(float* nocapture readonly %Base, float* nocapture %Dest, i32* nocapture readonly %Preds) { 119entry: 120 br label %for.body 121 122for.cond.cleanup: ; preds = %for.body 123 ret void 124 125for.body: ; preds = %entry, %for.body 126 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 127 %i.014 = phi i32 [ 0, %entry ], [ %add, %for.body ] 128 %arrayidx = getelementptr inbounds i32, i32* %Preds, i64 %indvars.iv 129 %0 = load i32, i32* %arrayidx, align 4 130 %cmp1.not = icmp eq i32 %0, 0 131 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 132 %add = add nuw nsw i32 %i.014, 1 133 %1 = trunc i64 %indvars.iv to i32 134 %mul = mul i32 %1, 3 135 %offset.0 = select i1 %cmp1.not, i32 4, i32 %mul 136 %idxprom213 = sext i32 %offset.0 to i64 137 %arrayidx3 = getelementptr inbounds float, float* %Base, i64 %idxprom213 138 %2 = load float, float* %arrayidx3, align 4 139 %arrayidx5 = getelementptr inbounds float, float* %Dest, i64 %indvars.iv 140 store float %2, float* %arrayidx5, align 4 141 %exitcond.not = icmp eq i64 %indvars.iv.next, 100 142 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body 143} 144 145; CHECK-LABEL: function 'forked_ptrs_gather_and_contiguous_forks': 146; CHECK-NEXT: for.body: 147; CHECK-NEXT: Report: cannot identify array bounds 148; CHECK-NEXT: Dependences: 149; CHECK-NEXT: Run-time memory checks: 150; CHECK-NEXT: Grouped accesses: 151; CHECK-EMPTY: 152; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop. 153; CHECK-NEXT: SCEV assumptions: 154; CHECK-EMPTY: 155; CHECK-NEXT: Expressions re-written: 156 157;;;; Derived from forked_ptrs_same_base_different_offset with a gather 158;;;; added using Preds as an index array in addition to the per-iteration 159;;;; condition. 160 161define dso_local void @forked_ptrs_gather_and_contiguous_forks(float* nocapture readonly %Base1, float* nocapture readonly %Base2, float* nocapture %Dest, i32* nocapture readonly %Preds) { 162entry: 163 br label %for.body 164 165for.cond.cleanup: ; preds = %for.body 166 ret void 167 168for.body: ; preds = %entry, %for.body 169 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 170 %arrayidx = getelementptr inbounds i32, i32* %Preds, i64 %indvars.iv 171 %0 = load i32, i32* %arrayidx, align 4 172 %cmp1.not = icmp eq i32 %0, 0 173 %arrayidx9 = getelementptr inbounds float, float* %Base2, i64 %indvars.iv 174 %idxprom4 = sext i32 %0 to i64 175 %arrayidx5 = getelementptr inbounds float, float* %Base1, i64 %idxprom4 176 %.sink.in = select i1 %cmp1.not, float* %arrayidx9, float* %arrayidx5 177 %.sink = load float, float* %.sink.in, align 4 178 %1 = getelementptr inbounds float, float* %Dest, i64 %indvars.iv 179 store float %.sink, float* %1, align 4 180 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 181 %exitcond.not = icmp eq i64 %indvars.iv.next, 100 182 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body 183} 184