1; RUN: opt -S -basicaa -licm < %s | FileCheck %s 2target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" 3target triple = "x86_64-unknown-linux-gnu" 4 5; This test represents the following function: 6; void test1(int * __restrict__ a, int * __restrict__ b, int &c, int n) { 7; for (int i = 0; i < n; ++i) 8; if (a[i] > 0) 9; a[i] = c*b[i]; 10; } 11; and we want to hoist the load of %c out of the loop. This can be done only 12; because the dereferenceable attribute is on %c. 13 14; CHECK-LABEL: @test1 15; CHECK: load i32* %c, align 4 16; CHECK: for.body: 17 18define void @test1(i32* noalias nocapture %a, i32* noalias nocapture readonly %b, i32* nocapture readonly nonnull dereferenceable(4) %c, i32 %n) #0 { 19entry: 20 %cmp11 = icmp sgt i32 %n, 0 21 br i1 %cmp11, label %for.body, label %for.end 22 23for.body: ; preds = %entry, %for.inc 24 %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ] 25 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 26 %0 = load i32* %arrayidx, align 4 27 %cmp1 = icmp sgt i32 %0, 0 28 br i1 %cmp1, label %if.then, label %for.inc 29 30if.then: ; preds = %for.body 31 %1 = load i32* %c, align 4 32 %arrayidx3 = getelementptr inbounds i32* %b, i64 %indvars.iv 33 %2 = load i32* %arrayidx3, align 4 34 %mul = mul nsw i32 %2, %1 35 store i32 %mul, i32* %arrayidx, align 4 36 br label %for.inc 37 38for.inc: ; preds = %for.body, %if.then 39 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 40 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 41 %exitcond = icmp eq i32 %lftr.wideiv, %n 42 br i1 %exitcond, label %for.end, label %for.body 43 44for.end: ; preds = %for.inc, %entry 45 ret void 46} 47 48; This is the same as @test1, but without the dereferenceable attribute on %c. 49; Without this attribute, we should not hoist the load of %c. 50 51; CHECK-LABEL: @test2 52; CHECK: if.then: 53; CHECK: load i32* %c, align 4 54 55define void @test2(i32* noalias nocapture %a, i32* noalias nocapture readonly %b, i32* nocapture readonly nonnull %c, i32 %n) #0 { 56entry: 57 %cmp11 = icmp sgt i32 %n, 0 58 br i1 %cmp11, label %for.body, label %for.end 59 60for.body: ; preds = %entry, %for.inc 61 %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ] 62 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 63 %0 = load i32* %arrayidx, align 4 64 %cmp1 = icmp sgt i32 %0, 0 65 br i1 %cmp1, label %if.then, label %for.inc 66 67if.then: ; preds = %for.body 68 %1 = load i32* %c, align 4 69 %arrayidx3 = getelementptr inbounds i32* %b, i64 %indvars.iv 70 %2 = load i32* %arrayidx3, align 4 71 %mul = mul nsw i32 %2, %1 72 store i32 %mul, i32* %arrayidx, align 4 73 br label %for.inc 74 75for.inc: ; preds = %for.body, %if.then 76 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 77 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 78 %exitcond = icmp eq i32 %lftr.wideiv, %n 79 br i1 %exitcond, label %for.end, label %for.body 80 81for.end: ; preds = %for.inc, %entry 82 ret void 83} 84 85; This test represents the following function: 86; void test3(int * restrict a, int * restrict b, int c[static 3], int n) { 87; for (int i = 0; i < n; ++i) 88; if (a[i] > 0) 89; a[i] = c[2]*b[i]; 90; } 91; and we want to hoist the load of c[2] out of the loop. This can be done only 92; because the dereferenceable attribute is on %c. 93 94; CHECK-LABEL: @test3 95; CHECK: load i32* %c2, align 4 96; CHECK: for.body: 97 98define void @test3(i32* noalias nocapture %a, i32* noalias nocapture readonly %b, i32* nocapture readonly dereferenceable(12) %c, i32 %n) #0 { 99entry: 100 %cmp11 = icmp sgt i32 %n, 0 101 br i1 %cmp11, label %for.body, label %for.end 102 103for.body: ; preds = %entry, %for.inc 104 %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ] 105 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 106 %0 = load i32* %arrayidx, align 4 107 %cmp1 = icmp sgt i32 %0, 0 108 br i1 %cmp1, label %if.then, label %for.inc 109 110if.then: ; preds = %for.body 111 %c2 = getelementptr inbounds i32* %c, i64 2 112 %1 = load i32* %c2, align 4 113 %arrayidx3 = getelementptr inbounds i32* %b, i64 %indvars.iv 114 %2 = load i32* %arrayidx3, align 4 115 %mul = mul nsw i32 %2, %1 116 store i32 %mul, i32* %arrayidx, align 4 117 br label %for.inc 118 119for.inc: ; preds = %for.body, %if.then 120 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 121 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 122 %exitcond = icmp eq i32 %lftr.wideiv, %n 123 br i1 %exitcond, label %for.end, label %for.body 124 125for.end: ; preds = %for.inc, %entry 126 ret void 127} 128 129; This is the same as @test3, but with a dereferenceable attribute on %c with a 130; size too small to cover c[2] (and so we should not hoist it). 131 132; CHECK-LABEL: @test4 133; CHECK: if.then: 134; CHECK: load i32* %c2, align 4 135 136define void @test4(i32* noalias nocapture %a, i32* noalias nocapture readonly %b, i32* nocapture readonly dereferenceable(11) %c, i32 %n) #0 { 137entry: 138 %cmp11 = icmp sgt i32 %n, 0 139 br i1 %cmp11, label %for.body, label %for.end 140 141for.body: ; preds = %entry, %for.inc 142 %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ] 143 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 144 %0 = load i32* %arrayidx, align 4 145 %cmp1 = icmp sgt i32 %0, 0 146 br i1 %cmp1, label %if.then, label %for.inc 147 148if.then: ; preds = %for.body 149 %c2 = getelementptr inbounds i32* %c, i64 2 150 %1 = load i32* %c2, align 4 151 %arrayidx3 = getelementptr inbounds i32* %b, i64 %indvars.iv 152 %2 = load i32* %arrayidx3, align 4 153 %mul = mul nsw i32 %2, %1 154 store i32 %mul, i32* %arrayidx, align 4 155 br label %for.inc 156 157for.inc: ; preds = %for.body, %if.then 158 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 159 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 160 %exitcond = icmp eq i32 %lftr.wideiv, %n 161 br i1 %exitcond, label %for.end, label %for.body 162 163for.end: ; preds = %for.inc, %entry 164 ret void 165} 166 167attributes #0 = { nounwind uwtable } 168 169