1 ; RUN: opt < %s -S -passes=loop-unroll -mcpu=nehalem | FileCheck %s
2 ; RUN: opt < %s -S -passes=loop-unroll -unroll-runtime=0 | FileCheck -check-prefix=CHECK-NOUNRL %s
3 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
4 target triple = "x86_64-unknown-linux-gnu"
6 define void @foo(ptr noalias nocapture readnone %ip, double %alpha, ptr noalias nocapture %a, ptr noalias nocapture readonly %b) #0 {
10 vector.body: ; preds = %vector.body, %entry
11 %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]
12 %0 = getelementptr inbounds double, ptr %b, i64 %index
13 %wide.load = load <2 x double>, ptr %0, align 8
14 %.sum9 = or i64 %index, 2
15 %1 = getelementptr double, ptr %b, i64 %.sum9
16 %wide.load8 = load <2 x double>, ptr %1, align 8
17 %2 = fadd <2 x double> %wide.load, <double 1.000000e+00, double 1.000000e+00>
18 %3 = fadd <2 x double> %wide.load8, <double 1.000000e+00, double 1.000000e+00>
19 %4 = getelementptr inbounds double, ptr %a, i64 %index
20 store <2 x double> %2, ptr %4, align 8
21 %.sum10 = or i64 %index, 2
22 %5 = getelementptr double, ptr %a, i64 %.sum10
23 store <2 x double> %3, ptr %5, align 8
24 %index.next = add i64 %index, 4
25 %6 = icmp eq i64 %index.next, 1600
26 br i1 %6, label %for.end, label %vector.body
28 ; FIXME: We should probably unroll this loop by a factor of 2, but the cost
29 ; model needs to be fixed to account for instructions likely to be folded
30 ; as part of an addressing mode.
32 ; CHECK-NOUNRL-LABEL: @foo
34 for.end: ; preds = %vector.body
38 define void @bar(ptr noalias nocapture readnone %ip, double %alpha, ptr noalias nocapture %a, ptr noalias nocapture readonly %b) #0 {
42 vector.body: ; preds = %vector.body, %entry
43 %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]
44 %v0 = getelementptr inbounds double, ptr %b, i64 %index
45 %wide.load = load <2 x double>, ptr %v0, align 8
46 %v4 = fadd <2 x double> %wide.load, <double 1.000000e+00, double 1.000000e+00>
47 %v5 = fmul <2 x double> %v4, <double 8.000000e+00, double 8.000000e+00>
48 %v6 = getelementptr inbounds double, ptr %a, i64 %index
49 store <2 x double> %v5, ptr %v6, align 8
50 %index.next = add i64 %index, 2
51 %v10 = icmp eq i64 %index.next, 1600
52 br i1 %v10, label %for.end, label %vector.body
54 ; FIXME: We should probably unroll this loop by a factor of 2, but the cost
55 ; model needs to first to fixed to account for instructions likely to be folded
56 ; as part of an addressing mode.
64 ; CHECK-NOUNRL-LABEL: @bar
66 ; CHECK-NOUNRL-NEXT: fmul
67 ; CHECK-NOUNRL-NOT: fadd
69 for.end: ; preds = %vector.body
73 define zeroext i16 @test1(ptr nocapture readonly %arr, i32 %n) #0 {
75 %cmp25 = icmp eq i32 %n, 0
76 br i1 %cmp25, label %for.end, label %for.body
78 for.body: ; preds = %entry, %for.body
79 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
80 %reduction.026 = phi i16 [ %add14, %for.body ], [ 0, %entry ]
81 %arrayidx = getelementptr inbounds i16, ptr %arr, i64 %indvars.iv
82 %0 = load i16, ptr %arrayidx, align 2
84 %add = add i16 %mul, %reduction.026
85 %sext = mul i64 %indvars.iv, 12884901888
86 %idxprom3 = ashr exact i64 %sext, 32
87 %arrayidx4 = getelementptr inbounds i16, ptr %arr, i64 %idxprom3
88 %1 = load i16, ptr %arrayidx4, align 2
90 %add7 = add i16 %add, %mul2
91 %sext28 = mul i64 %indvars.iv, 21474836480
92 %idxprom10 = ashr exact i64 %sext28, 32
93 %arrayidx11 = getelementptr inbounds i16, ptr %arr, i64 %idxprom10
94 %2 = load i16, ptr %arrayidx11, align 2
96 %add14 = add i16 %add7, %mul3
97 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
98 %lftr.wideiv = trunc i64 %indvars.iv.next to i32
99 %exitcond = icmp eq i32 %lftr.wideiv, %n
100 br i1 %exitcond, label %for.end, label %for.body
102 for.end: ; preds = %for.body, %entry
103 %reduction.0.lcssa = phi i16 [ 0, %entry ], [ %add14, %for.body ]
104 ret i16 %reduction.0.lcssa
106 ; This loop is too large to be partially unrolled (size=16)
108 ; CHECK-LABEL: @test1
115 ; CHECK-NOUNRL-LABEL: @test1
120 ; CHECK-NOUNRL-NOT: br
123 attributes #0 = { nounwind uwtable }