| blob | 80491942fc24d1c513f79771e8ca8a101f2946a8 |
1 ; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py
2 ; RUN: opt < %s -mtriple=x86_64-unknown-linux-gnu -passes="print<cost-model>" 2>&1 -disable-output -cost-kind=latency -mattr=+sse2 | FileCheck %s -check-prefixes=SSE
3 ; RUN: opt < %s -mtriple=x86_64-unknown-linux-gnu -passes="print<cost-model>" 2>&1 -disable-output -cost-kind=latency -mattr=+sse2 | FileCheck %s -check-prefixes=SSE2
4 ; RUN: opt < %s -mtriple=x86_64-unknown-linux-gnu -passes="print<cost-model>" 2>&1 -disable-output -cost-kind=latency -mattr=+sse3 | FileCheck %s -check-prefixes=SSE3
5 ; RUN: opt < %s -mtriple=x86_64-unknown-linux-gnu -passes="print<cost-model>" 2>&1 -disable-output -cost-kind=latency -mattr=+sse3 | FileCheck %s -check-prefixes=AVX
6 ; RUN: opt < %s -mtriple=x86_64-unknown-linux-gnu -passes="print<cost-model>" 2>&1 -disable-output -cost-kind=latency -mattr=+sse3 | FileCheck %s -check-prefixes=AVX2
7 ; RUN: opt < %s -mtriple=x86_64-unknown-linux-gnu -passes="print<cost-model>" 2>&1 -disable-output -cost-kind=latency -mattr=+sse3 | FileCheck %s -check-prefixes=AVX512
9 ; This test checks that the cost of a splat-load shuffle is correctly detected.
10 ; If there is a combined load+broadcast instruction, like `movddup` it should
11 ; return 0.
12 ;
13 ; TODO: AVX `vbroadcast*` seems to support more types than the
14 ; 2xdouble type of `movddup`:
15 ; - `vbroadcastss` supports 4xfloat, 8xfloat
16 ; - `vbroadcastsd` supports 4xdouble
18 ; NOTE: The code in this test is a hack. Since TTI cannot currently detect a
19 ; proper broadcast pattern from a scalar load (like the one that follows),
20 ; we use a vector load as the shuffle's operand to trigger the pattern.
21 ;
22 ; %load = load double, double *%ptr
23 ; %insert = insertelement <2 x double> poison, double %load, i32 0
24 ; %bcast = shufflevector <2 x double> %insert, <2 x double> poison, <2 x i32> zeroinitializer
26 define void @shuffle_load() {
27 ; SSE-LABEL: 'shuffle_load'
28 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi8 = load <2 x i8>, ptr undef, align 2
29 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi8 = shufflevector <2 x i8> %ld_2xi8, <2 x i8> undef, <2 x i32> zeroinitializer
30 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi8 = load <4 x i8>, ptr undef, align 4
31 ; SSE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_4xi8 = shufflevector <4 x i8> %ld_4xi8, <4 x i8> undef, <4 x i32> zeroinitializer
32 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi8 = load <8 x i8>, ptr undef, align 8
33 ; SSE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi8 = shufflevector <8 x i8> %ld_8xi8, <8 x i8> undef, <8 x i32> zeroinitializer
34 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi8 = load <16 x i8>, ptr undef, align 16
35 ; SSE-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_16xi8 = shufflevector <16 x i8> %ld_16xi8, <16 x i8> undef, <16 x i32> zeroinitializer
36 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi8 = load <32 x i8>, ptr undef, align 32
37 ; SSE-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_32xi8 = shufflevector <32 x i8> %ld_32xi8, <32 x i8> undef, <32 x i32> zeroinitializer
38 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_64xi8 = load <64 x i8>, ptr undef, align 64
39 ; SSE-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_64xi8 = shufflevector <64 x i8> %ld_64xi8, <64 x i8> undef, <64 x i32> zeroinitializer
40 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi16 = load <2 x i16>, ptr undef, align 4
41 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi16 = shufflevector <2 x i16> %ld_2xi16, <2 x i16> undef, <2 x i32> zeroinitializer
42 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi16 = load <4 x i16>, ptr undef, align 8
43 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi16 = shufflevector <4 x i16> %ld_4xi16, <4 x i16> undef, <4 x i32> zeroinitializer
44 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi16 = load <8 x i16>, ptr undef, align 16
45 ; SSE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi16 = shufflevector <8 x i16> %ld_8xi16, <8 x i16> undef, <8 x i32> zeroinitializer
46 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi16 = load <16 x i16>, ptr undef, align 32
47 ; SSE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xi16 = shufflevector <16 x i16> %ld_16xi16, <16 x i16> undef, <16 x i32> zeroinitializer
48 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi16 = load <32 x i16>, ptr undef, align 64
49 ; SSE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xi16 = shufflevector <32 x i16> %ld_32xi16, <32 x i16> undef, <32 x i32> zeroinitializer
50 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi32 = load <2 x i32>, ptr undef, align 8
51 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi32 = shufflevector <2 x i32> %ld_2xi32, <2 x i32> undef, <2 x i32> zeroinitializer
52 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi32 = load <4 x i32>, ptr undef, align 16
53 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi32 = shufflevector <4 x i32> %ld_4xi32, <4 x i32> undef, <4 x i32> zeroinitializer
54 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi32 = load <8 x i32>, ptr undef, align 32
55 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi32 = shufflevector <8 x i32> %ld_8xi32, <8 x i32> undef, <8 x i32> zeroinitializer
56 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi32 = load <16 x i32>, ptr undef, align 64
57 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xi32 = shufflevector <16 x i32> %ld_16xi32, <16 x i32> undef, <16 x i32> zeroinitializer
58 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi64 = load <2 x i64>, ptr undef, align 16
59 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi64 = shufflevector <2 x i64> %ld_2xi64, <2 x i64> undef, <2 x i32> zeroinitializer
60 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi64 = load <4 x i64>, ptr undef, align 32
61 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi64 = shufflevector <4 x i64> %ld_4xi64, <4 x i64> undef, <4 x i32> zeroinitializer
62 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi64 = load <8 x i64>, ptr undef, align 64
63 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi64 = shufflevector <8 x i64> %ld_8xi64, <8 x i64> undef, <8 x i32> zeroinitializer
64 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf16 = load <2 x half>, ptr undef, align 4
65 ; SSE-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf16 = shufflevector <2 x half> %ld_2xf16, <2 x half> undef, <2 x i32> zeroinitializer
66 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf16 = load <4 x half>, ptr undef, align 8
67 ; SSE-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf16 = shufflevector <4 x half> %ld_4xf16, <4 x half> undef, <4 x i32> zeroinitializer
68 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf16 = load <8 x half>, ptr undef, align 16
69 ; SSE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xf16 = shufflevector <8 x half> %ld_8xf16, <8 x half> undef, <8 x i32> zeroinitializer
70 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf16 = load <16 x half>, ptr undef, align 32
71 ; SSE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xf16 = shufflevector <16 x half> %ld_16xf16, <16 x half> undef, <16 x i32> zeroinitializer
72 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xf16 = load <32 x half>, ptr undef, align 64
73 ; SSE-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xf16 = shufflevector <32 x half> %ld_32xf16, <32 x half> undef, <32 x i32> zeroinitializer
74 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf32 = load <2 x float>, ptr undef, align 8
75 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xf32 = shufflevector <2 x float> %ld_2xf32, <2 x float> undef, <2 x i32> zeroinitializer
76 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf32 = load <4 x float>, ptr undef, align 16
77 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xf32 = shufflevector <4 x float> %ld_4xf32, <4 x float> undef, <4 x i32> zeroinitializer
78 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf32 = load <8 x float>, ptr undef, align 32
79 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xf32 = shufflevector <8 x float> %ld_8xf32, <8 x float> undef, <8 x i32> zeroinitializer
80 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf32 = load <16 x float>, ptr undef, align 64
81 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xf32 = shufflevector <16 x float> %ld_16xf32, <16 x float> undef, <16 x i32> zeroinitializer
82 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf64 = load <2 x double>, ptr undef, align 16
83 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xf64 = shufflevector <2 x double> %ld_2xf64, <2 x double> undef, <2 x i32> zeroinitializer
84 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf64 = load <4 x double>, ptr undef, align 32
85 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xf64 = shufflevector <4 x double> %ld_4xf64, <4 x double> undef, <4 x i32> zeroinitializer
86 ; SSE-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf64 = load <8 x double>, ptr undef, align 64
87 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xf64 = shufflevector <8 x double> %ld_8xf64, <8 x double> undef, <8 x i32> zeroinitializer
88 ; SSE-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
89 ;
90 ; SSE2-LABEL: 'shuffle_load'
91 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi8 = load <2 x i8>, ptr undef, align 2
92 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi8 = shufflevector <2 x i8> %ld_2xi8, <2 x i8> undef, <2 x i32> zeroinitializer
93 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi8 = load <4 x i8>, ptr undef, align 4
94 ; SSE2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_4xi8 = shufflevector <4 x i8> %ld_4xi8, <4 x i8> undef, <4 x i32> zeroinitializer
95 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi8 = load <8 x i8>, ptr undef, align 8
96 ; SSE2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi8 = shufflevector <8 x i8> %ld_8xi8, <8 x i8> undef, <8 x i32> zeroinitializer
97 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi8 = load <16 x i8>, ptr undef, align 16
98 ; SSE2-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_16xi8 = shufflevector <16 x i8> %ld_16xi8, <16 x i8> undef, <16 x i32> zeroinitializer
99 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi8 = load <32 x i8>, ptr undef, align 32
100 ; SSE2-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_32xi8 = shufflevector <32 x i8> %ld_32xi8, <32 x i8> undef, <32 x i32> zeroinitializer
101 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_64xi8 = load <64 x i8>, ptr undef, align 64
102 ; SSE2-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_64xi8 = shufflevector <64 x i8> %ld_64xi8, <64 x i8> undef, <64 x i32> zeroinitializer
103 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi16 = load <2 x i16>, ptr undef, align 4
104 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi16 = shufflevector <2 x i16> %ld_2xi16, <2 x i16> undef, <2 x i32> zeroinitializer
105 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi16 = load <4 x i16>, ptr undef, align 8
106 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi16 = shufflevector <4 x i16> %ld_4xi16, <4 x i16> undef, <4 x i32> zeroinitializer
107 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi16 = load <8 x i16>, ptr undef, align 16
108 ; SSE2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi16 = shufflevector <8 x i16> %ld_8xi16, <8 x i16> undef, <8 x i32> zeroinitializer
109 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi16 = load <16 x i16>, ptr undef, align 32
110 ; SSE2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xi16 = shufflevector <16 x i16> %ld_16xi16, <16 x i16> undef, <16 x i32> zeroinitializer
111 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi16 = load <32 x i16>, ptr undef, align 64
112 ; SSE2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xi16 = shufflevector <32 x i16> %ld_32xi16, <32 x i16> undef, <32 x i32> zeroinitializer
113 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi32 = load <2 x i32>, ptr undef, align 8
114 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi32 = shufflevector <2 x i32> %ld_2xi32, <2 x i32> undef, <2 x i32> zeroinitializer
115 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi32 = load <4 x i32>, ptr undef, align 16
116 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi32 = shufflevector <4 x i32> %ld_4xi32, <4 x i32> undef, <4 x i32> zeroinitializer
117 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi32 = load <8 x i32>, ptr undef, align 32
118 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi32 = shufflevector <8 x i32> %ld_8xi32, <8 x i32> undef, <8 x i32> zeroinitializer
119 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi32 = load <16 x i32>, ptr undef, align 64
120 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xi32 = shufflevector <16 x i32> %ld_16xi32, <16 x i32> undef, <16 x i32> zeroinitializer
121 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi64 = load <2 x i64>, ptr undef, align 16
122 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi64 = shufflevector <2 x i64> %ld_2xi64, <2 x i64> undef, <2 x i32> zeroinitializer
123 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi64 = load <4 x i64>, ptr undef, align 32
124 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi64 = shufflevector <4 x i64> %ld_4xi64, <4 x i64> undef, <4 x i32> zeroinitializer
125 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi64 = load <8 x i64>, ptr undef, align 64
126 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi64 = shufflevector <8 x i64> %ld_8xi64, <8 x i64> undef, <8 x i32> zeroinitializer
127 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf16 = load <2 x half>, ptr undef, align 4
128 ; SSE2-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf16 = shufflevector <2 x half> %ld_2xf16, <2 x half> undef, <2 x i32> zeroinitializer
129 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf16 = load <4 x half>, ptr undef, align 8
130 ; SSE2-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf16 = shufflevector <4 x half> %ld_4xf16, <4 x half> undef, <4 x i32> zeroinitializer
131 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf16 = load <8 x half>, ptr undef, align 16
132 ; SSE2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xf16 = shufflevector <8 x half> %ld_8xf16, <8 x half> undef, <8 x i32> zeroinitializer
133 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf16 = load <16 x half>, ptr undef, align 32
134 ; SSE2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xf16 = shufflevector <16 x half> %ld_16xf16, <16 x half> undef, <16 x i32> zeroinitializer
135 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xf16 = load <32 x half>, ptr undef, align 64
136 ; SSE2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xf16 = shufflevector <32 x half> %ld_32xf16, <32 x half> undef, <32 x i32> zeroinitializer
137 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf32 = load <2 x float>, ptr undef, align 8
138 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xf32 = shufflevector <2 x float> %ld_2xf32, <2 x float> undef, <2 x i32> zeroinitializer
139 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf32 = load <4 x float>, ptr undef, align 16
140 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xf32 = shufflevector <4 x float> %ld_4xf32, <4 x float> undef, <4 x i32> zeroinitializer
141 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf32 = load <8 x float>, ptr undef, align 32
142 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xf32 = shufflevector <8 x float> %ld_8xf32, <8 x float> undef, <8 x i32> zeroinitializer
143 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf32 = load <16 x float>, ptr undef, align 64
144 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xf32 = shufflevector <16 x float> %ld_16xf32, <16 x float> undef, <16 x i32> zeroinitializer
145 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf64 = load <2 x double>, ptr undef, align 16
146 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xf64 = shufflevector <2 x double> %ld_2xf64, <2 x double> undef, <2 x i32> zeroinitializer
147 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf64 = load <4 x double>, ptr undef, align 32
148 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xf64 = shufflevector <4 x double> %ld_4xf64, <4 x double> undef, <4 x i32> zeroinitializer
149 ; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf64 = load <8 x double>, ptr undef, align 64
150 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xf64 = shufflevector <8 x double> %ld_8xf64, <8 x double> undef, <8 x i32> zeroinitializer
151 ; SSE2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
152 ;
153 ; SSE3-LABEL: 'shuffle_load'
154 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi8 = load <2 x i8>, ptr undef, align 2
155 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi8 = shufflevector <2 x i8> %ld_2xi8, <2 x i8> undef, <2 x i32> zeroinitializer
156 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi8 = load <4 x i8>, ptr undef, align 4
157 ; SSE3-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_4xi8 = shufflevector <4 x i8> %ld_4xi8, <4 x i8> undef, <4 x i32> zeroinitializer
158 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi8 = load <8 x i8>, ptr undef, align 8
159 ; SSE3-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi8 = shufflevector <8 x i8> %ld_8xi8, <8 x i8> undef, <8 x i32> zeroinitializer
160 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi8 = load <16 x i8>, ptr undef, align 16
161 ; SSE3-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_16xi8 = shufflevector <16 x i8> %ld_16xi8, <16 x i8> undef, <16 x i32> zeroinitializer
162 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi8 = load <32 x i8>, ptr undef, align 32
163 ; SSE3-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_32xi8 = shufflevector <32 x i8> %ld_32xi8, <32 x i8> undef, <32 x i32> zeroinitializer
164 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_64xi8 = load <64 x i8>, ptr undef, align 64
165 ; SSE3-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_64xi8 = shufflevector <64 x i8> %ld_64xi8, <64 x i8> undef, <64 x i32> zeroinitializer
166 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi16 = load <2 x i16>, ptr undef, align 4
167 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi16 = shufflevector <2 x i16> %ld_2xi16, <2 x i16> undef, <2 x i32> zeroinitializer
168 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi16 = load <4 x i16>, ptr undef, align 8
169 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi16 = shufflevector <4 x i16> %ld_4xi16, <4 x i16> undef, <4 x i32> zeroinitializer
170 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi16 = load <8 x i16>, ptr undef, align 16
171 ; SSE3-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi16 = shufflevector <8 x i16> %ld_8xi16, <8 x i16> undef, <8 x i32> zeroinitializer
172 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi16 = load <16 x i16>, ptr undef, align 32
173 ; SSE3-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xi16 = shufflevector <16 x i16> %ld_16xi16, <16 x i16> undef, <16 x i32> zeroinitializer
174 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi16 = load <32 x i16>, ptr undef, align 64
175 ; SSE3-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xi16 = shufflevector <32 x i16> %ld_32xi16, <32 x i16> undef, <32 x i32> zeroinitializer
176 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi32 = load <2 x i32>, ptr undef, align 8
177 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi32 = shufflevector <2 x i32> %ld_2xi32, <2 x i32> undef, <2 x i32> zeroinitializer
178 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi32 = load <4 x i32>, ptr undef, align 16
179 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi32 = shufflevector <4 x i32> %ld_4xi32, <4 x i32> undef, <4 x i32> zeroinitializer
180 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi32 = load <8 x i32>, ptr undef, align 32
181 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi32 = shufflevector <8 x i32> %ld_8xi32, <8 x i32> undef, <8 x i32> zeroinitializer
182 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi32 = load <16 x i32>, ptr undef, align 64
183 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xi32 = shufflevector <16 x i32> %ld_16xi32, <16 x i32> undef, <16 x i32> zeroinitializer
184 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi64 = load <2 x i64>, ptr undef, align 16
185 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi64 = shufflevector <2 x i64> %ld_2xi64, <2 x i64> undef, <2 x i32> zeroinitializer
186 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi64 = load <4 x i64>, ptr undef, align 32
187 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi64 = shufflevector <4 x i64> %ld_4xi64, <4 x i64> undef, <4 x i32> zeroinitializer
188 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi64 = load <8 x i64>, ptr undef, align 64
189 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi64 = shufflevector <8 x i64> %ld_8xi64, <8 x i64> undef, <8 x i32> zeroinitializer
190 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf16 = load <2 x half>, ptr undef, align 4
191 ; SSE3-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf16 = shufflevector <2 x half> %ld_2xf16, <2 x half> undef, <2 x i32> zeroinitializer
192 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf16 = load <4 x half>, ptr undef, align 8
193 ; SSE3-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf16 = shufflevector <4 x half> %ld_4xf16, <4 x half> undef, <4 x i32> zeroinitializer
194 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf16 = load <8 x half>, ptr undef, align 16
195 ; SSE3-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xf16 = shufflevector <8 x half> %ld_8xf16, <8 x half> undef, <8 x i32> zeroinitializer
196 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf16 = load <16 x half>, ptr undef, align 32
197 ; SSE3-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xf16 = shufflevector <16 x half> %ld_16xf16, <16 x half> undef, <16 x i32> zeroinitializer
198 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xf16 = load <32 x half>, ptr undef, align 64
199 ; SSE3-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xf16 = shufflevector <32 x half> %ld_32xf16, <32 x half> undef, <32 x i32> zeroinitializer
200 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf32 = load <2 x float>, ptr undef, align 8
201 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xf32 = shufflevector <2 x float> %ld_2xf32, <2 x float> undef, <2 x i32> zeroinitializer
202 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf32 = load <4 x float>, ptr undef, align 16
203 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xf32 = shufflevector <4 x float> %ld_4xf32, <4 x float> undef, <4 x i32> zeroinitializer
204 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf32 = load <8 x float>, ptr undef, align 32
205 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xf32 = shufflevector <8 x float> %ld_8xf32, <8 x float> undef, <8 x i32> zeroinitializer
206 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf32 = load <16 x float>, ptr undef, align 64
207 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xf32 = shufflevector <16 x float> %ld_16xf32, <16 x float> undef, <16 x i32> zeroinitializer
208 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf64 = load <2 x double>, ptr undef, align 16
209 ; SSE3-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf64 = shufflevector <2 x double> %ld_2xf64, <2 x double> undef, <2 x i32> zeroinitializer
210 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf64 = load <4 x double>, ptr undef, align 32
211 ; SSE3-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf64 = shufflevector <4 x double> %ld_4xf64, <4 x double> undef, <4 x i32> zeroinitializer
212 ; SSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf64 = load <8 x double>, ptr undef, align 64
213 ; SSE3-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_8xf64 = shufflevector <8 x double> %ld_8xf64, <8 x double> undef, <8 x i32> zeroinitializer
214 ; SSE3-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
215 ;
216 ; AVX-LABEL: 'shuffle_load'
217 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi8 = load <2 x i8>, ptr undef, align 2
218 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi8 = shufflevector <2 x i8> %ld_2xi8, <2 x i8> undef, <2 x i32> zeroinitializer
219 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi8 = load <4 x i8>, ptr undef, align 4
220 ; AVX-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_4xi8 = shufflevector <4 x i8> %ld_4xi8, <4 x i8> undef, <4 x i32> zeroinitializer
221 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi8 = load <8 x i8>, ptr undef, align 8
222 ; AVX-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi8 = shufflevector <8 x i8> %ld_8xi8, <8 x i8> undef, <8 x i32> zeroinitializer
223 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi8 = load <16 x i8>, ptr undef, align 16
224 ; AVX-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_16xi8 = shufflevector <16 x i8> %ld_16xi8, <16 x i8> undef, <16 x i32> zeroinitializer
225 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi8 = load <32 x i8>, ptr undef, align 32
226 ; AVX-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_32xi8 = shufflevector <32 x i8> %ld_32xi8, <32 x i8> undef, <32 x i32> zeroinitializer
227 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_64xi8 = load <64 x i8>, ptr undef, align 64
228 ; AVX-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_64xi8 = shufflevector <64 x i8> %ld_64xi8, <64 x i8> undef, <64 x i32> zeroinitializer
229 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi16 = load <2 x i16>, ptr undef, align 4
230 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi16 = shufflevector <2 x i16> %ld_2xi16, <2 x i16> undef, <2 x i32> zeroinitializer
231 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi16 = load <4 x i16>, ptr undef, align 8
232 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi16 = shufflevector <4 x i16> %ld_4xi16, <4 x i16> undef, <4 x i32> zeroinitializer
233 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi16 = load <8 x i16>, ptr undef, align 16
234 ; AVX-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi16 = shufflevector <8 x i16> %ld_8xi16, <8 x i16> undef, <8 x i32> zeroinitializer
235 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi16 = load <16 x i16>, ptr undef, align 32
236 ; AVX-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xi16 = shufflevector <16 x i16> %ld_16xi16, <16 x i16> undef, <16 x i32> zeroinitializer
237 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi16 = load <32 x i16>, ptr undef, align 64
238 ; AVX-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xi16 = shufflevector <32 x i16> %ld_32xi16, <32 x i16> undef, <32 x i32> zeroinitializer
239 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi32 = load <2 x i32>, ptr undef, align 8
240 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi32 = shufflevector <2 x i32> %ld_2xi32, <2 x i32> undef, <2 x i32> zeroinitializer
241 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi32 = load <4 x i32>, ptr undef, align 16
242 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi32 = shufflevector <4 x i32> %ld_4xi32, <4 x i32> undef, <4 x i32> zeroinitializer
243 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi32 = load <8 x i32>, ptr undef, align 32
244 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi32 = shufflevector <8 x i32> %ld_8xi32, <8 x i32> undef, <8 x i32> zeroinitializer
245 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi32 = load <16 x i32>, ptr undef, align 64
246 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xi32 = shufflevector <16 x i32> %ld_16xi32, <16 x i32> undef, <16 x i32> zeroinitializer
247 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi64 = load <2 x i64>, ptr undef, align 16
248 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi64 = shufflevector <2 x i64> %ld_2xi64, <2 x i64> undef, <2 x i32> zeroinitializer
249 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi64 = load <4 x i64>, ptr undef, align 32
250 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi64 = shufflevector <4 x i64> %ld_4xi64, <4 x i64> undef, <4 x i32> zeroinitializer
251 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi64 = load <8 x i64>, ptr undef, align 64
252 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi64 = shufflevector <8 x i64> %ld_8xi64, <8 x i64> undef, <8 x i32> zeroinitializer
253 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf16 = load <2 x half>, ptr undef, align 4
254 ; AVX-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf16 = shufflevector <2 x half> %ld_2xf16, <2 x half> undef, <2 x i32> zeroinitializer
255 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf16 = load <4 x half>, ptr undef, align 8
256 ; AVX-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf16 = shufflevector <4 x half> %ld_4xf16, <4 x half> undef, <4 x i32> zeroinitializer
257 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf16 = load <8 x half>, ptr undef, align 16
258 ; AVX-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xf16 = shufflevector <8 x half> %ld_8xf16, <8 x half> undef, <8 x i32> zeroinitializer
259 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf16 = load <16 x half>, ptr undef, align 32
260 ; AVX-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xf16 = shufflevector <16 x half> %ld_16xf16, <16 x half> undef, <16 x i32> zeroinitializer
261 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xf16 = load <32 x half>, ptr undef, align 64
262 ; AVX-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xf16 = shufflevector <32 x half> %ld_32xf16, <32 x half> undef, <32 x i32> zeroinitializer
263 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf32 = load <2 x float>, ptr undef, align 8
264 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xf32 = shufflevector <2 x float> %ld_2xf32, <2 x float> undef, <2 x i32> zeroinitializer
265 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf32 = load <4 x float>, ptr undef, align 16
266 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xf32 = shufflevector <4 x float> %ld_4xf32, <4 x float> undef, <4 x i32> zeroinitializer
267 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf32 = load <8 x float>, ptr undef, align 32
268 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xf32 = shufflevector <8 x float> %ld_8xf32, <8 x float> undef, <8 x i32> zeroinitializer
269 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf32 = load <16 x float>, ptr undef, align 64
270 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xf32 = shufflevector <16 x float> %ld_16xf32, <16 x float> undef, <16 x i32> zeroinitializer
271 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf64 = load <2 x double>, ptr undef, align 16
272 ; AVX-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf64 = shufflevector <2 x double> %ld_2xf64, <2 x double> undef, <2 x i32> zeroinitializer
273 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf64 = load <4 x double>, ptr undef, align 32
274 ; AVX-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf64 = shufflevector <4 x double> %ld_4xf64, <4 x double> undef, <4 x i32> zeroinitializer
275 ; AVX-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf64 = load <8 x double>, ptr undef, align 64
276 ; AVX-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_8xf64 = shufflevector <8 x double> %ld_8xf64, <8 x double> undef, <8 x i32> zeroinitializer
277 ; AVX-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
278 ;
279 ; AVX2-LABEL: 'shuffle_load'
280 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi8 = load <2 x i8>, ptr undef, align 2
281 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi8 = shufflevector <2 x i8> %ld_2xi8, <2 x i8> undef, <2 x i32> zeroinitializer
282 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi8 = load <4 x i8>, ptr undef, align 4
283 ; AVX2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_4xi8 = shufflevector <4 x i8> %ld_4xi8, <4 x i8> undef, <4 x i32> zeroinitializer
284 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi8 = load <8 x i8>, ptr undef, align 8
285 ; AVX2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi8 = shufflevector <8 x i8> %ld_8xi8, <8 x i8> undef, <8 x i32> zeroinitializer
286 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi8 = load <16 x i8>, ptr undef, align 16
287 ; AVX2-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_16xi8 = shufflevector <16 x i8> %ld_16xi8, <16 x i8> undef, <16 x i32> zeroinitializer
288 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi8 = load <32 x i8>, ptr undef, align 32
289 ; AVX2-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_32xi8 = shufflevector <32 x i8> %ld_32xi8, <32 x i8> undef, <32 x i32> zeroinitializer
290 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_64xi8 = load <64 x i8>, ptr undef, align 64
291 ; AVX2-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_64xi8 = shufflevector <64 x i8> %ld_64xi8, <64 x i8> undef, <64 x i32> zeroinitializer
292 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi16 = load <2 x i16>, ptr undef, align 4
293 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi16 = shufflevector <2 x i16> %ld_2xi16, <2 x i16> undef, <2 x i32> zeroinitializer
294 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi16 = load <4 x i16>, ptr undef, align 8
295 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi16 = shufflevector <4 x i16> %ld_4xi16, <4 x i16> undef, <4 x i32> zeroinitializer
296 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi16 = load <8 x i16>, ptr undef, align 16
297 ; AVX2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi16 = shufflevector <8 x i16> %ld_8xi16, <8 x i16> undef, <8 x i32> zeroinitializer
298 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi16 = load <16 x i16>, ptr undef, align 32
299 ; AVX2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xi16 = shufflevector <16 x i16> %ld_16xi16, <16 x i16> undef, <16 x i32> zeroinitializer
300 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi16 = load <32 x i16>, ptr undef, align 64
301 ; AVX2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xi16 = shufflevector <32 x i16> %ld_32xi16, <32 x i16> undef, <32 x i32> zeroinitializer
302 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi32 = load <2 x i32>, ptr undef, align 8
303 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi32 = shufflevector <2 x i32> %ld_2xi32, <2 x i32> undef, <2 x i32> zeroinitializer
304 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi32 = load <4 x i32>, ptr undef, align 16
305 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi32 = shufflevector <4 x i32> %ld_4xi32, <4 x i32> undef, <4 x i32> zeroinitializer
306 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi32 = load <8 x i32>, ptr undef, align 32
307 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi32 = shufflevector <8 x i32> %ld_8xi32, <8 x i32> undef, <8 x i32> zeroinitializer
308 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi32 = load <16 x i32>, ptr undef, align 64
309 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xi32 = shufflevector <16 x i32> %ld_16xi32, <16 x i32> undef, <16 x i32> zeroinitializer
310 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi64 = load <2 x i64>, ptr undef, align 16
311 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi64 = shufflevector <2 x i64> %ld_2xi64, <2 x i64> undef, <2 x i32> zeroinitializer
312 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi64 = load <4 x i64>, ptr undef, align 32
313 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi64 = shufflevector <4 x i64> %ld_4xi64, <4 x i64> undef, <4 x i32> zeroinitializer
314 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi64 = load <8 x i64>, ptr undef, align 64
315 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi64 = shufflevector <8 x i64> %ld_8xi64, <8 x i64> undef, <8 x i32> zeroinitializer
316 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf16 = load <2 x half>, ptr undef, align 4
317 ; AVX2-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf16 = shufflevector <2 x half> %ld_2xf16, <2 x half> undef, <2 x i32> zeroinitializer
318 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf16 = load <4 x half>, ptr undef, align 8
319 ; AVX2-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf16 = shufflevector <4 x half> %ld_4xf16, <4 x half> undef, <4 x i32> zeroinitializer
320 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf16 = load <8 x half>, ptr undef, align 16
321 ; AVX2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xf16 = shufflevector <8 x half> %ld_8xf16, <8 x half> undef, <8 x i32> zeroinitializer
322 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf16 = load <16 x half>, ptr undef, align 32
323 ; AVX2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xf16 = shufflevector <16 x half> %ld_16xf16, <16 x half> undef, <16 x i32> zeroinitializer
324 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xf16 = load <32 x half>, ptr undef, align 64
325 ; AVX2-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xf16 = shufflevector <32 x half> %ld_32xf16, <32 x half> undef, <32 x i32> zeroinitializer
326 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf32 = load <2 x float>, ptr undef, align 8
327 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xf32 = shufflevector <2 x float> %ld_2xf32, <2 x float> undef, <2 x i32> zeroinitializer
328 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf32 = load <4 x float>, ptr undef, align 16
329 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xf32 = shufflevector <4 x float> %ld_4xf32, <4 x float> undef, <4 x i32> zeroinitializer
330 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf32 = load <8 x float>, ptr undef, align 32
331 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xf32 = shufflevector <8 x float> %ld_8xf32, <8 x float> undef, <8 x i32> zeroinitializer
332 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf32 = load <16 x float>, ptr undef, align 64
333 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xf32 = shufflevector <16 x float> %ld_16xf32, <16 x float> undef, <16 x i32> zeroinitializer
334 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf64 = load <2 x double>, ptr undef, align 16
335 ; AVX2-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf64 = shufflevector <2 x double> %ld_2xf64, <2 x double> undef, <2 x i32> zeroinitializer
336 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf64 = load <4 x double>, ptr undef, align 32
337 ; AVX2-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf64 = shufflevector <4 x double> %ld_4xf64, <4 x double> undef, <4 x i32> zeroinitializer
338 ; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf64 = load <8 x double>, ptr undef, align 64
339 ; AVX2-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_8xf64 = shufflevector <8 x double> %ld_8xf64, <8 x double> undef, <8 x i32> zeroinitializer
340 ; AVX2-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
341 ;
342 ; AVX512-LABEL: 'shuffle_load'
343 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi8 = load <2 x i8>, ptr undef, align 2
344 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi8 = shufflevector <2 x i8> %ld_2xi8, <2 x i8> undef, <2 x i32> zeroinitializer
345 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi8 = load <4 x i8>, ptr undef, align 4
346 ; AVX512-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_4xi8 = shufflevector <4 x i8> %ld_4xi8, <4 x i8> undef, <4 x i32> zeroinitializer
347 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi8 = load <8 x i8>, ptr undef, align 8
348 ; AVX512-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi8 = shufflevector <8 x i8> %ld_8xi8, <8 x i8> undef, <8 x i32> zeroinitializer
349 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi8 = load <16 x i8>, ptr undef, align 16
350 ; AVX512-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_16xi8 = shufflevector <16 x i8> %ld_16xi8, <16 x i8> undef, <16 x i32> zeroinitializer
351 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi8 = load <32 x i8>, ptr undef, align 32
352 ; AVX512-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_32xi8 = shufflevector <32 x i8> %ld_32xi8, <32 x i8> undef, <32 x i32> zeroinitializer
353 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_64xi8 = load <64 x i8>, ptr undef, align 64
354 ; AVX512-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %sf_64xi8 = shufflevector <64 x i8> %ld_64xi8, <64 x i8> undef, <64 x i32> zeroinitializer
355 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi16 = load <2 x i16>, ptr undef, align 4
356 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi16 = shufflevector <2 x i16> %ld_2xi16, <2 x i16> undef, <2 x i32> zeroinitializer
357 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi16 = load <4 x i16>, ptr undef, align 8
358 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi16 = shufflevector <4 x i16> %ld_4xi16, <4 x i16> undef, <4 x i32> zeroinitializer
359 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi16 = load <8 x i16>, ptr undef, align 16
360 ; AVX512-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xi16 = shufflevector <8 x i16> %ld_8xi16, <8 x i16> undef, <8 x i32> zeroinitializer
361 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi16 = load <16 x i16>, ptr undef, align 32
362 ; AVX512-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xi16 = shufflevector <16 x i16> %ld_16xi16, <16 x i16> undef, <16 x i32> zeroinitializer
363 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xi16 = load <32 x i16>, ptr undef, align 64
364 ; AVX512-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xi16 = shufflevector <32 x i16> %ld_32xi16, <32 x i16> undef, <32 x i32> zeroinitializer
365 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi32 = load <2 x i32>, ptr undef, align 8
366 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi32 = shufflevector <2 x i32> %ld_2xi32, <2 x i32> undef, <2 x i32> zeroinitializer
367 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi32 = load <4 x i32>, ptr undef, align 16
368 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi32 = shufflevector <4 x i32> %ld_4xi32, <4 x i32> undef, <4 x i32> zeroinitializer
369 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi32 = load <8 x i32>, ptr undef, align 32
370 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi32 = shufflevector <8 x i32> %ld_8xi32, <8 x i32> undef, <8 x i32> zeroinitializer
371 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xi32 = load <16 x i32>, ptr undef, align 64
372 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xi32 = shufflevector <16 x i32> %ld_16xi32, <16 x i32> undef, <16 x i32> zeroinitializer
373 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xi64 = load <2 x i64>, ptr undef, align 16
374 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xi64 = shufflevector <2 x i64> %ld_2xi64, <2 x i64> undef, <2 x i32> zeroinitializer
375 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xi64 = load <4 x i64>, ptr undef, align 32
376 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xi64 = shufflevector <4 x i64> %ld_4xi64, <4 x i64> undef, <4 x i32> zeroinitializer
377 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xi64 = load <8 x i64>, ptr undef, align 64
378 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xi64 = shufflevector <8 x i64> %ld_8xi64, <8 x i64> undef, <8 x i32> zeroinitializer
379 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf16 = load <2 x half>, ptr undef, align 4
380 ; AVX512-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf16 = shufflevector <2 x half> %ld_2xf16, <2 x half> undef, <2 x i32> zeroinitializer
381 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf16 = load <4 x half>, ptr undef, align 8
382 ; AVX512-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf16 = shufflevector <4 x half> %ld_4xf16, <4 x half> undef, <4 x i32> zeroinitializer
383 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf16 = load <8 x half>, ptr undef, align 16
384 ; AVX512-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_8xf16 = shufflevector <8 x half> %ld_8xf16, <8 x half> undef, <8 x i32> zeroinitializer
385 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf16 = load <16 x half>, ptr undef, align 32
386 ; AVX512-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_16xf16 = shufflevector <16 x half> %ld_16xf16, <16 x half> undef, <16 x i32> zeroinitializer
387 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_32xf16 = load <32 x half>, ptr undef, align 64
388 ; AVX512-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %sf_32xf16 = shufflevector <32 x half> %ld_32xf16, <32 x half> undef, <32 x i32> zeroinitializer
389 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf32 = load <2 x float>, ptr undef, align 8
390 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_2xf32 = shufflevector <2 x float> %ld_2xf32, <2 x float> undef, <2 x i32> zeroinitializer
391 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf32 = load <4 x float>, ptr undef, align 16
392 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_4xf32 = shufflevector <4 x float> %ld_4xf32, <4 x float> undef, <4 x i32> zeroinitializer
393 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf32 = load <8 x float>, ptr undef, align 32
394 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_8xf32 = shufflevector <8 x float> %ld_8xf32, <8 x float> undef, <8 x i32> zeroinitializer
395 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_16xf32 = load <16 x float>, ptr undef, align 64
396 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %sf_16xf32 = shufflevector <16 x float> %ld_16xf32, <16 x float> undef, <16 x i32> zeroinitializer
397 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_2xf64 = load <2 x double>, ptr undef, align 16
398 ; AVX512-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_2xf64 = shufflevector <2 x double> %ld_2xf64, <2 x double> undef, <2 x i32> zeroinitializer
399 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_4xf64 = load <4 x double>, ptr undef, align 32
400 ; AVX512-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_4xf64 = shufflevector <4 x double> %ld_4xf64, <4 x double> undef, <4 x i32> zeroinitializer
401 ; AVX512-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %ld_8xf64 = load <8 x double>, ptr undef, align 64
402 ; AVX512-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %sf_8xf64 = shufflevector <8 x double> %ld_8xf64, <8 x double> undef, <8 x i32> zeroinitializer
403 ; AVX512-NEXT: Cost Model: Found an estimated cost of 1 for instruction: ret void
404 ;
405 %ld_2xi8 = load <2 x i8>, ptr undef
406 %sf_2xi8 = shufflevector <2 x i8> %ld_2xi8, <2 x i8> undef, <2 x i32> zeroinitializer
407 %ld_4xi8 = load <4 x i8>, ptr undef
408 %sf_4xi8 = shufflevector <4 x i8> %ld_4xi8, <4 x i8> undef, <4 x i32> zeroinitializer
409 %ld_8xi8 = load <8 x i8>, ptr undef
410 %sf_8xi8 = shufflevector <8 x i8> %ld_8xi8, <8 x i8> undef, <8 x i32> zeroinitializer
411 %ld_16xi8 = load <16 x i8>, ptr undef
412 %sf_16xi8 = shufflevector <16 x i8> %ld_16xi8, <16 x i8> undef, <16 x i32> zeroinitializer
413 %ld_32xi8 = load <32 x i8>, ptr undef
414 %sf_32xi8 = shufflevector <32 x i8> %ld_32xi8, <32 x i8> undef, <32 x i32> zeroinitializer
415 %ld_64xi8 = load <64 x i8>, ptr undef
416 %sf_64xi8 = shufflevector <64 x i8> %ld_64xi8, <64 x i8> undef, <64 x i32> zeroinitializer
418 %ld_2xi16 = load <2 x i16>, ptr undef
419 %sf_2xi16 = shufflevector <2 x i16> %ld_2xi16, <2 x i16> undef, <2 x i32> zeroinitializer
420 %ld_4xi16 = load <4 x i16>, ptr undef
421 %sf_4xi16 = shufflevector <4 x i16> %ld_4xi16, <4 x i16> undef, <4 x i32> zeroinitializer
422 %ld_8xi16 = load <8 x i16>, ptr undef
423 %sf_8xi16 = shufflevector <8 x i16> %ld_8xi16, <8 x i16> undef, <8 x i32> zeroinitializer
424 %ld_16xi16 = load <16 x i16>, ptr undef
425 %sf_16xi16 = shufflevector <16 x i16> %ld_16xi16, <16 x i16> undef, <16 x i32> zeroinitializer
426 %ld_32xi16 = load <32 x i16>, ptr undef
427 %sf_32xi16 = shufflevector <32 x i16> %ld_32xi16, <32 x i16> undef, <32 x i32> zeroinitializer
429 %ld_2xi32 = load <2 x i32>, ptr undef
430 %sf_2xi32 = shufflevector <2 x i32> %ld_2xi32, <2 x i32> undef, <2 x i32> zeroinitializer
431 %ld_4xi32 = load <4 x i32>, ptr undef
432 %sf_4xi32 = shufflevector <4 x i32> %ld_4xi32, <4 x i32> undef, <4 x i32> zeroinitializer
433 %ld_8xi32 = load <8 x i32>, ptr undef
434 %sf_8xi32 = shufflevector <8 x i32> %ld_8xi32, <8 x i32> undef, <8 x i32> zeroinitializer
435 %ld_16xi32 = load <16 x i32>, ptr undef
436 %sf_16xi32 = shufflevector <16 x i32> %ld_16xi32, <16 x i32> undef, <16 x i32> zeroinitializer
438 %ld_2xi64 = load <2 x i64>, ptr undef
439 %sf_2xi64 = shufflevector <2 x i64> %ld_2xi64, <2 x i64> undef, <2 x i32> zeroinitializer
440 %ld_4xi64 = load <4 x i64>, ptr undef
441 %sf_4xi64 = shufflevector <4 x i64> %ld_4xi64, <4 x i64> undef, <4 x i32> zeroinitializer
442 %ld_8xi64 = load <8 x i64>, ptr undef
443 %sf_8xi64 = shufflevector <8 x i64> %ld_8xi64, <8 x i64> undef, <8 x i32> zeroinitializer
445 %ld_2xf16 = load <2 x half>, ptr undef
446 %sf_2xf16 = shufflevector <2 x half> %ld_2xf16, <2 x half> undef, <2 x i32> zeroinitializer
447 %ld_4xf16 = load <4 x half>, ptr undef
448 %sf_4xf16 = shufflevector <4 x half> %ld_4xf16, <4 x half> undef, <4 x i32> zeroinitializer
449 %ld_8xf16 = load <8 x half>, ptr undef
450 %sf_8xf16 = shufflevector <8 x half> %ld_8xf16, <8 x half> undef, <8 x i32> zeroinitializer
451 %ld_16xf16 = load <16 x half>, ptr undef
452 %sf_16xf16 = shufflevector <16 x half> %ld_16xf16, <16 x half> undef, <16 x i32> zeroinitializer
453 %ld_32xf16 = load <32 x half>, ptr undef
454 %sf_32xf16 = shufflevector <32 x half> %ld_32xf16, <32 x half> undef, <32 x i32> zeroinitializer
456 %ld_2xf32 = load <2 x float>, ptr undef
457 %sf_2xf32 = shufflevector <2 x float> %ld_2xf32, <2 x float> undef, <2 x i32> zeroinitializer
458 %ld_4xf32 = load <4 x float>, ptr undef
459 %sf_4xf32 = shufflevector <4 x float> %ld_4xf32, <4 x float> undef, <4 x i32> zeroinitializer
460 %ld_8xf32 = load <8 x float>, ptr undef
461 %sf_8xf32 = shufflevector <8 x float> %ld_8xf32, <8 x float> undef, <8 x i32> zeroinitializer
462 %ld_16xf32 = load <16 x float>, ptr undef
463 %sf_16xf32 = shufflevector <16 x float> %ld_16xf32, <16 x float> undef, <16 x i32> zeroinitializer
465 %ld_2xf64 = load <2 x double>, ptr undef
466 %sf_2xf64 = shufflevector <2 x double> %ld_2xf64, <2 x double> undef, <2 x i32> zeroinitializer
467 %ld_4xf64 = load <4 x double>, ptr undef
468 %sf_4xf64 = shufflevector <4 x double> %ld_4xf64, <4 x double> undef, <4 x i32> zeroinitializer
469 %ld_8xf64 = load <8 x double>, ptr undef
470 %sf_8xf64 = shufflevector <8 x double> %ld_8xf64, <8 x double> undef, <8 x i32> zeroinitializer
472 ret void
473 }