1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt < %s -basicaa -slp-vectorizer -S -mtriple=aarch64-unknown-unknown -mcpu=cortex-a53 | FileCheck %s
4 target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
6 ; This test is reduced from the matrix multiplication benchmark in the test-suite:
7 ; https://github.com/llvm/llvm-test-suite/tree/master/SingleSource/Benchmarks/Misc/matmul_f64_4x4.c
8 ; The operations here are expected to be vectorized to <2 x double>.
9 ; Otherwise, performance will suffer on Cortex-A53.
11 define void @wrap_mul4(double* nocapture %Out, [2 x double]* nocapture readonly %A, [4 x double]* nocapture readonly %B) {
12 ; CHECK-LABEL: @wrap_mul4(
13 ; CHECK-NEXT: [[ARRAYIDX1_I:%.*]] = getelementptr inbounds [2 x double], [2 x double]* [[A:%.*]], i64 0, i64 0
14 ; CHECK-NEXT: [[TEMP:%.*]] = load double, double* [[ARRAYIDX1_I]], align 8
15 ; CHECK-NEXT: [[ARRAYIDX3_I:%.*]] = getelementptr inbounds [4 x double], [4 x double]* [[B:%.*]], i64 0, i64 0
16 ; CHECK-NEXT: [[ARRAYIDX5_I:%.*]] = getelementptr inbounds [2 x double], [2 x double]* [[A]], i64 0, i64 1
17 ; CHECK-NEXT: [[TEMP2:%.*]] = load double, double* [[ARRAYIDX5_I]], align 8
18 ; CHECK-NEXT: [[ARRAYIDX7_I:%.*]] = getelementptr inbounds [4 x double], [4 x double]* [[B]], i64 1, i64 0
19 ; CHECK-NEXT: [[ARRAYIDX13_I:%.*]] = getelementptr inbounds [4 x double], [4 x double]* [[B]], i64 0, i64 1
20 ; CHECK-NEXT: [[TMP1:%.*]] = bitcast double* [[ARRAYIDX3_I]] to <2 x double>*
21 ; CHECK-NEXT: [[TMP2:%.*]] = load <2 x double>, <2 x double>* [[TMP1]], align 8
22 ; CHECK-NEXT: [[TMP3:%.*]] = insertelement <2 x double> undef, double [[TEMP]], i32 0
23 ; CHECK-NEXT: [[TMP4:%.*]] = insertelement <2 x double> [[TMP3]], double [[TEMP]], i32 1
24 ; CHECK-NEXT: [[TMP5:%.*]] = fmul <2 x double> [[TMP4]], [[TMP2]]
25 ; CHECK-NEXT: [[ARRAYIDX18_I:%.*]] = getelementptr inbounds [4 x double], [4 x double]* [[B]], i64 1, i64 1
26 ; CHECK-NEXT: [[TMP6:%.*]] = bitcast double* [[ARRAYIDX7_I]] to <2 x double>*
27 ; CHECK-NEXT: [[TMP7:%.*]] = load <2 x double>, <2 x double>* [[TMP6]], align 8
28 ; CHECK-NEXT: [[TMP8:%.*]] = insertelement <2 x double> undef, double [[TEMP2]], i32 0
29 ; CHECK-NEXT: [[TMP9:%.*]] = insertelement <2 x double> [[TMP8]], double [[TEMP2]], i32 1
30 ; CHECK-NEXT: [[TMP10:%.*]] = fmul <2 x double> [[TMP9]], [[TMP7]]
31 ; CHECK-NEXT: [[TMP11:%.*]] = fadd <2 x double> [[TMP5]], [[TMP10]]
32 ; CHECK-NEXT: [[ARRAYIDX25_I:%.*]] = getelementptr inbounds [4 x double], [4 x double]* [[B]], i64 0, i64 2
33 ; CHECK-NEXT: [[ARRAYIDX30_I:%.*]] = getelementptr inbounds [4 x double], [4 x double]* [[B]], i64 1, i64 2
34 ; CHECK-NEXT: [[ARRAYIDX37_I:%.*]] = getelementptr inbounds [4 x double], [4 x double]* [[B]], i64 0, i64 3
35 ; CHECK-NEXT: [[TMP12:%.*]] = bitcast double* [[ARRAYIDX25_I]] to <2 x double>*
36 ; CHECK-NEXT: [[TMP13:%.*]] = load <2 x double>, <2 x double>* [[TMP12]], align 8
37 ; CHECK-NEXT: [[TMP14:%.*]] = fmul <2 x double> [[TMP4]], [[TMP13]]
38 ; CHECK-NEXT: [[ARRAYIDX42_I:%.*]] = getelementptr inbounds [4 x double], [4 x double]* [[B]], i64 1, i64 3
39 ; CHECK-NEXT: [[TMP15:%.*]] = bitcast double* [[ARRAYIDX30_I]] to <2 x double>*
40 ; CHECK-NEXT: [[TMP16:%.*]] = load <2 x double>, <2 x double>* [[TMP15]], align 8
41 ; CHECK-NEXT: [[TMP17:%.*]] = fmul <2 x double> [[TMP9]], [[TMP16]]
42 ; CHECK-NEXT: [[TMP18:%.*]] = fadd <2 x double> [[TMP14]], [[TMP17]]
43 ; CHECK-NEXT: [[ARRAYIDX47_I:%.*]] = getelementptr inbounds [2 x double], [2 x double]* [[A]], i64 1, i64 0
44 ; CHECK-NEXT: [[TEMP10:%.*]] = load double, double* [[ARRAYIDX47_I]], align 8
45 ; CHECK-NEXT: [[ARRAYIDX52_I:%.*]] = getelementptr inbounds [2 x double], [2 x double]* [[A]], i64 1, i64 1
46 ; CHECK-NEXT: [[TEMP11:%.*]] = load double, double* [[ARRAYIDX52_I]], align 8
47 ; CHECK-NEXT: [[TMP19:%.*]] = insertelement <2 x double> undef, double [[TEMP10]], i32 0
48 ; CHECK-NEXT: [[TMP20:%.*]] = insertelement <2 x double> [[TMP19]], double [[TEMP10]], i32 1
49 ; CHECK-NEXT: [[TMP21:%.*]] = fmul <2 x double> [[TMP2]], [[TMP20]]
50 ; CHECK-NEXT: [[TMP22:%.*]] = insertelement <2 x double> undef, double [[TEMP11]], i32 0
51 ; CHECK-NEXT: [[TMP23:%.*]] = insertelement <2 x double> [[TMP22]], double [[TEMP11]], i32 1
52 ; CHECK-NEXT: [[TMP24:%.*]] = fmul <2 x double> [[TMP7]], [[TMP23]]
53 ; CHECK-NEXT: [[TMP25:%.*]] = fadd <2 x double> [[TMP21]], [[TMP24]]
54 ; CHECK-NEXT: [[TMP26:%.*]] = fmul <2 x double> [[TMP13]], [[TMP20]]
55 ; CHECK-NEXT: [[TMP27:%.*]] = fmul <2 x double> [[TMP16]], [[TMP23]]
56 ; CHECK-NEXT: [[TMP28:%.*]] = fadd <2 x double> [[TMP26]], [[TMP27]]
57 ; CHECK-NEXT: [[RES_I_SROA_4_0_OUT2_I_SROA_IDX2:%.*]] = getelementptr inbounds double, double* [[OUT:%.*]], i64 1
58 ; CHECK-NEXT: [[TMP29:%.*]] = bitcast double* [[OUT]] to <2 x double>*
59 ; CHECK-NEXT: store <2 x double> [[TMP11]], <2 x double>* [[TMP29]], align 8
60 ; CHECK-NEXT: [[RES_I_SROA_5_0_OUT2_I_SROA_IDX4:%.*]] = getelementptr inbounds double, double* [[OUT]], i64 2
61 ; CHECK-NEXT: [[RES_I_SROA_6_0_OUT2_I_SROA_IDX6:%.*]] = getelementptr inbounds double, double* [[OUT]], i64 3
62 ; CHECK-NEXT: [[TMP30:%.*]] = bitcast double* [[RES_I_SROA_5_0_OUT2_I_SROA_IDX4]] to <2 x double>*
63 ; CHECK-NEXT: store <2 x double> [[TMP18]], <2 x double>* [[TMP30]], align 8
64 ; CHECK-NEXT: [[RES_I_SROA_7_0_OUT2_I_SROA_IDX8:%.*]] = getelementptr inbounds double, double* [[OUT]], i64 4
65 ; CHECK-NEXT: [[RES_I_SROA_8_0_OUT2_I_SROA_IDX10:%.*]] = getelementptr inbounds double, double* [[OUT]], i64 5
66 ; CHECK-NEXT: [[TMP31:%.*]] = bitcast double* [[RES_I_SROA_7_0_OUT2_I_SROA_IDX8]] to <2 x double>*
67 ; CHECK-NEXT: store <2 x double> [[TMP25]], <2 x double>* [[TMP31]], align 8
68 ; CHECK-NEXT: [[RES_I_SROA_9_0_OUT2_I_SROA_IDX12:%.*]] = getelementptr inbounds double, double* [[OUT]], i64 6
69 ; CHECK-NEXT: [[RES_I_SROA_10_0_OUT2_I_SROA_IDX14:%.*]] = getelementptr inbounds double, double* [[OUT]], i64 7
70 ; CHECK-NEXT: [[TMP32:%.*]] = bitcast double* [[RES_I_SROA_9_0_OUT2_I_SROA_IDX12]] to <2 x double>*
71 ; CHECK-NEXT: store <2 x double> [[TMP28]], <2 x double>* [[TMP32]], align 8
72 ; CHECK-NEXT: ret void
74 %arrayidx1.i = getelementptr inbounds [2 x double], [2 x double]* %A, i64 0, i64 0
75 %temp = load double, double* %arrayidx1.i, align 8
76 %arrayidx3.i = getelementptr inbounds [4 x double], [4 x double]* %B, i64 0, i64 0
77 %temp1 = load double, double* %arrayidx3.i, align 8
78 %mul.i = fmul double %temp, %temp1
79 %arrayidx5.i = getelementptr inbounds [2 x double], [2 x double]* %A, i64 0, i64 1
80 %temp2 = load double, double* %arrayidx5.i, align 8
81 %arrayidx7.i = getelementptr inbounds [4 x double], [4 x double]* %B, i64 1, i64 0
82 %temp3 = load double, double* %arrayidx7.i, align 8
83 %mul8.i = fmul double %temp2, %temp3
84 %add.i = fadd double %mul.i, %mul8.i
85 %arrayidx13.i = getelementptr inbounds [4 x double], [4 x double]* %B, i64 0, i64 1
86 %temp4 = load double, double* %arrayidx13.i, align 8
87 %mul14.i = fmul double %temp, %temp4
88 %arrayidx18.i = getelementptr inbounds [4 x double], [4 x double]* %B, i64 1, i64 1
89 %temp5 = load double, double* %arrayidx18.i, align 8
90 %mul19.i = fmul double %temp2, %temp5
91 %add20.i = fadd double %mul14.i, %mul19.i
92 %arrayidx25.i = getelementptr inbounds [4 x double], [4 x double]* %B, i64 0, i64 2
93 %temp6 = load double, double* %arrayidx25.i, align 8
94 %mul26.i = fmul double %temp, %temp6
95 %arrayidx30.i = getelementptr inbounds [4 x double], [4 x double]* %B, i64 1, i64 2
96 %temp7 = load double, double* %arrayidx30.i, align 8
97 %mul31.i = fmul double %temp2, %temp7
98 %add32.i = fadd double %mul26.i, %mul31.i
99 %arrayidx37.i = getelementptr inbounds [4 x double], [4 x double]* %B, i64 0, i64 3
100 %temp8 = load double, double* %arrayidx37.i, align 8
101 %mul38.i = fmul double %temp, %temp8
102 %arrayidx42.i = getelementptr inbounds [4 x double], [4 x double]* %B, i64 1, i64 3
103 %temp9 = load double, double* %arrayidx42.i, align 8
104 %mul43.i = fmul double %temp2, %temp9
105 %add44.i = fadd double %mul38.i, %mul43.i
106 %arrayidx47.i = getelementptr inbounds [2 x double], [2 x double]* %A, i64 1, i64 0
107 %temp10 = load double, double* %arrayidx47.i, align 8
108 %mul50.i = fmul double %temp1, %temp10
109 %arrayidx52.i = getelementptr inbounds [2 x double], [2 x double]* %A, i64 1, i64 1
110 %temp11 = load double, double* %arrayidx52.i, align 8
111 %mul55.i = fmul double %temp3, %temp11
112 %add56.i = fadd double %mul50.i, %mul55.i
113 %mul62.i = fmul double %temp4, %temp10
114 %mul67.i = fmul double %temp5, %temp11
115 %add68.i = fadd double %mul62.i, %mul67.i
116 %mul74.i = fmul double %temp6, %temp10
117 %mul79.i = fmul double %temp7, %temp11
118 %add80.i = fadd double %mul74.i, %mul79.i
119 %mul86.i = fmul double %temp8, %temp10
120 %mul91.i = fmul double %temp9, %temp11
121 %add92.i = fadd double %mul86.i, %mul91.i
122 store double %add.i, double* %Out, align 8
123 %Res.i.sroa.4.0.Out2.i.sroa_idx2 = getelementptr inbounds double, double* %Out, i64 1
124 store double %add20.i, double* %Res.i.sroa.4.0.Out2.i.sroa_idx2, align 8
125 %Res.i.sroa.5.0.Out2.i.sroa_idx4 = getelementptr inbounds double, double* %Out, i64 2
126 store double %add32.i, double* %Res.i.sroa.5.0.Out2.i.sroa_idx4, align 8
127 %Res.i.sroa.6.0.Out2.i.sroa_idx6 = getelementptr inbounds double, double* %Out, i64 3
128 store double %add44.i, double* %Res.i.sroa.6.0.Out2.i.sroa_idx6, align 8
129 %Res.i.sroa.7.0.Out2.i.sroa_idx8 = getelementptr inbounds double, double* %Out, i64 4
130 store double %add56.i, double* %Res.i.sroa.7.0.Out2.i.sroa_idx8, align 8
131 %Res.i.sroa.8.0.Out2.i.sroa_idx10 = getelementptr inbounds double, double* %Out, i64 5
132 store double %add68.i, double* %Res.i.sroa.8.0.Out2.i.sroa_idx10, align 8
133 %Res.i.sroa.9.0.Out2.i.sroa_idx12 = getelementptr inbounds double, double* %Out, i64 6
134 store double %add80.i, double* %Res.i.sroa.9.0.Out2.i.sroa_idx12, align 8
135 %Res.i.sroa.10.0.Out2.i.sroa_idx14 = getelementptr inbounds double, double* %Out, i64 7
136 store double %add92.i, double* %Res.i.sroa.10.0.Out2.i.sroa_idx14, align 8