1 ; RUN: opt < %s -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -enable-if-conversion -dce -instcombine -S | FileCheck %s
3 target 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"
5 ; This is the loop in this example:
7 ;int function0(int *a, int *b, int start, int end) {
9 ; for (int i=start; i<end; ++i) {
12 ; if (a[i] > b[i]) <------ notice the IF inside the loop.
15 ; a[i] = k; <---- K is a phi node that becomes vector-select.
19 ;CHECK-LABEL: @function0(
20 ;CHECK: load <4 x i32>
21 ;CHECK: icmp sgt <4 x i32>
24 ;CHECK: select <4 x i1>
26 define i32 @function0(i32* nocapture %a, i32* nocapture %b, i32 %start, i32 %end) nounwind uwtable ssp {
28 %cmp16 = icmp slt i32 %start, %end
29 br i1 %cmp16, label %for.body.lr.ph, label %for.end
32 %0 = sext i32 %start to i64
36 %indvars.iv = phi i64 [ %0, %for.body.lr.ph ], [ %indvars.iv.next, %if.end ]
37 %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
38 %1 = load i32, i32* %arrayidx, align 4
39 %arrayidx4 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
40 %2 = load i32, i32* %arrayidx4, align 4
41 %cmp5 = icmp sgt i32 %1, %2
42 br i1 %cmp5, label %if.then, label %if.end
46 %add = add i32 %mul, 3
50 %k.0 = phi i32 [ %add, %if.then ], [ %1, %for.body ]
51 store i32 %k.0, i32* %arrayidx, align 4
52 %indvars.iv.next = add i64 %indvars.iv, 1
53 %3 = trunc i64 %indvars.iv.next to i32
54 %cmp = icmp slt i32 %3, %end
55 br i1 %cmp, label %for.body, label %for.end
63 ; int func(int *A, int n) {
65 ; for (int i = 0; i < n; ++i)
72 ;CHECK-LABEL: @reduction_func(
73 ;CHECK: load <4 x i32>
74 ;CHECK: icmp slt <4 x i32>
76 ;CHECK: select <4 x i1>
78 define i32 @reduction_func(i32* nocapture %A, i32 %n) nounwind uwtable readonly ssp {
80 %cmp10 = icmp sgt i32 %n, 0
81 br i1 %cmp10, label %for.body, label %for.end
83 for.body: ; preds = %entry, %for.inc
84 %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ]
85 %sum.011 = phi i32 [ %sum.1, %for.inc ], [ 0, %entry ]
86 %arrayidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
87 %0 = load i32, i32* %arrayidx, align 4
88 %cmp1 = icmp sgt i32 %0, 30
89 br i1 %cmp1, label %if.then, label %for.inc
91 if.then: ; preds = %for.body
92 %add = add i32 %sum.011, 2
93 %add4 = add i32 %add, %0
96 for.inc: ; preds = %for.body, %if.then
97 %sum.1 = phi i32 [ %add4, %if.then ], [ %sum.011, %for.body ]
98 %indvars.iv.next = add i64 %indvars.iv, 1
99 %lftr.wideiv = trunc i64 %indvars.iv.next to i32
100 %exitcond = icmp eq i32 %lftr.wideiv, %n
101 br i1 %exitcond, label %for.end, label %for.body
103 for.end: ; preds = %for.inc, %entry
104 %sum.0.lcssa = phi i32 [ 0, %entry ], [ %sum.1, %for.inc ]
108 @a = common global [1 x i32*] zeroinitializer, align 8
109 @c = common global i32* null, align 8
111 ; We use to if convert this loop. This is not safe because there is a trapping
112 ; constant expression.
115 ; CHECK-LABEL: trapping_constant_expression
116 ; CHECK-NOT: or <4 x i32>
118 define i32 @trapping_constant_expression() {
123 %inc3 = phi i32 [ 0, %entry ], [ %inc, %cond.end ]
124 %or2 = phi i32 [ 0, %entry ], [ %or, %cond.end ]
125 br i1 icmp eq (i32** getelementptr inbounds ([1 x i32*], [1 x i32*]* @a, i64 0, i64 0), i32** @c), label %cond.false, label %cond.end
131 %cond = phi i32 [ sdiv (i32 1, i32 zext (i1 icmp eq (i32** getelementptr inbounds ([1 x i32*], [1 x i32*]* @a, i64 0, i64 0), i32** @c) to i32)), %cond.false ], [ 0, %for.body ]
132 %or = or i32 %or2, %cond
133 %inc = add nsw i32 %inc3, 1
134 %cmp = icmp slt i32 %inc, 128
135 br i1 %cmp, label %for.body, label %for.end
141 ; Neither should we if-convert if there is an instruction operand that is a
142 ; trapping constant expression.
145 ; CHECK-LABEL: trapping_constant_expression2
146 ; CHECK-NOT: or <4 x i32>
148 define i32 @trapping_constant_expression2() {
153 %inc3 = phi i32 [ 0, %entry ], [ %inc, %cond.end ]
154 %or2 = phi i32 [ 0, %entry ], [ %or, %cond.end ]
155 br i1 icmp eq (i32** getelementptr inbounds ([1 x i32*], [1 x i32*]* @a, i64 0, i64 0), i32** @c), label %cond.false, label %cond.end
158 %cond.1 = or i32 %inc3, sdiv (i32 1, i32 zext (i1 icmp eq (i32** getelementptr inbounds ([1 x i32*], [1 x i32*]* @a, i64 0, i64 1), i32** @c) to i32))
162 %cond = phi i32 [ %cond.1, %cond.false ], [ %inc3, %for.body ]
163 %or = or i32 %or2, %cond
164 %inc = add nsw i32 %inc3, 1
165 %cmp = icmp slt i32 %inc, 128
166 br i1 %cmp, label %for.body, label %for.end
172 ; Handle PHI with single incoming value having a full mask.
175 ; CHECK-LABEL: PR34523
178 define void @PR34523() {
182 bb2: ; preds = %bb4, %bb1
183 %i = phi i16 [ undef, %bb1 ], [ %_tmp2, %bb4 ]
187 %_tmp1 = phi [1 x [1 x i32]]* [ undef, %bb2 ]
191 %_tmp2 = add i16 %i, 1
192 %_tmp3 = icmp slt i16 %_tmp2, 2
193 br i1 %_tmp3, label %bb2, label %bb5