xref: /llvm-project/llvm/test/Analysis/CostModel/AMDGPU/shufflevector.ll (revision 07d284d4ebffd58d4b2934769b4e11fedd0b106e) 
1; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py
2; RUN: opt < %s -passes="print<cost-model>" 2>&1 -disable-output -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx1010 -S | FileCheck -check-prefixes=ALL,GFX9-10 %s
3; RUN: opt < %s -passes="print<cost-model>" 2>&1 -disable-output -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 -S | FileCheck -check-prefixes=ALL,GFX9-10 %s
4; RUN: opt < %s -passes="print<cost-model>" 2>&1 -disable-output -mtriple=amdgcn-unknown-amdhsa -mcpu=fiji -S | FileCheck -check-prefixes=ALL,VI %s
5; RUN: opt < %s -passes="print<cost-model>" 2>&1 -disable-output -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx1010 -cost-kind=code-size -S | FileCheck -check-prefixes=ALL-SIZE,GFX9-10-SIZE %s
6; RUN: opt < %s -passes="print<cost-model>" 2>&1 -disable-output -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 -cost-kind=code-size -S | FileCheck -check-prefixes=ALL-SIZE,GFX9-10-SIZE %s
7; RUN: opt < %s -passes="print<cost-model>" 2>&1 -disable-output -mtriple=amdgcn-unknown-amdhsa -mcpu=fiji -cost-kind=code-size -S | FileCheck -check-prefixes=ALL-SIZE,VI-SIZE %s
8; END.
9
10define amdgpu_kernel void @shufflevector_i16(<2 x i16> %vec1, <2 x i16> %vec2) {
11; GFX9-10-LABEL: 'shufflevector_i16'
12; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf00 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> zeroinitializer
13; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 1>
14; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf10 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 0>
15; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf11 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 1>
16; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf02 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 2>
17; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %shuf20 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 0>
18; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf22 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 2>
19; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf03 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 3>
20; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf30 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 0>
21; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf33 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 3>
22; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf12 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 2>
23; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf21 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 1>
24; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf13 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 3>
25; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf31 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 1>
26; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 3>
27; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf32 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 2>
28; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf000 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> zeroinitializer
29; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf001 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
30; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf010 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
31; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf011 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
32; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf100 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
33; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf101 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
34; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf110 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
35; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf111 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
36; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf002 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
37; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf020 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
38; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf022 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
39; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf200 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
40; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf202 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
41; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf220 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
42; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf222 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
43; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf112 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
44; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf121 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
45; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf122 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
46; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf211 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
47; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf212 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
48; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf221 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
49; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf00_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> zeroinitializer
50; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 1>
51; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf10_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 0>
52; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf11_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 1>
53; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf02_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 2>
54; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %shuf20_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 0>
55; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf22_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 2>
56; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf03_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 3>
57; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf30_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 0>
58; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf33_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 3>
59; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf12_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 2>
60; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf21_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 1>
61; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf13_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 3>
62; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf31_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 1>
63; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 3>
64; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf32_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 2>
65; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf000_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> zeroinitializer
66; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf001_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
67; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf010_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
68; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf011_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
69; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf100_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
70; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf101_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
71; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf110_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
72; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf111_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
73; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf002_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
74; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf020_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
75; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf022_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
76; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf200_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
77; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf202_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
78; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf220_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
79; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf222_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
80; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf112_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
81; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf121_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
82; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf122_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
83; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf211_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
84; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf212_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
85; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf221_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
86; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: ret void
87;
88; VI-LABEL: 'shufflevector_i16'
89; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf00 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> zeroinitializer
90; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 1>
91; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf10 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 0>
92; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf11 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 1>
93; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf02 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 2>
94; VI-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %shuf20 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 0>
95; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf22 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 2>
96; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf03 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 3>
97; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf30 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 0>
98; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf33 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 3>
99; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf12 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 2>
100; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf21 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 1>
101; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf13 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 3>
102; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf31 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 1>
103; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 3>
104; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf32 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 2>
105; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf000 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> zeroinitializer
106; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf001 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
107; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf010 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
108; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf011 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
109; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf100 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
110; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf101 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
111; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf110 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
112; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf111 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
113; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf002 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
114; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf020 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
115; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf022 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
116; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf200 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
117; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf202 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
118; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf220 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
119; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf222 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
120; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf112 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
121; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf121 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
122; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf122 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
123; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf211 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
124; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf212 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
125; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf221 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
126; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf00_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> zeroinitializer
127; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 1>
128; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf10_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 0>
129; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf11_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 1>
130; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf02_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 2>
131; VI-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %shuf20_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 0>
132; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf22_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 2>
133; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf03_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 3>
134; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf30_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 0>
135; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf33_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 3>
136; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf12_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 2>
137; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf21_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 1>
138; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf13_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 3>
139; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf31_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 1>
140; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 3>
141; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf32_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 2>
142; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf000_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> zeroinitializer
143; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf001_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
144; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf010_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
145; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf011_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
146; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf100_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
147; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf101_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
148; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf110_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
149; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf111_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
150; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf002_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
151; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf020_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
152; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf022_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
153; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf200_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
154; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf202_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
155; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf220_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
156; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf222_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
157; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf112_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
158; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf121_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
159; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf122_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
160; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf211_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
161; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf212_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
162; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf221_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
163; VI-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: ret void
164;
165; GFX9-10-SIZE-LABEL: 'shufflevector_i16'
166; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf00 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> zeroinitializer
167; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 1>
168; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf10 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 0>
169; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf11 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 1>
170; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf02 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 2>
171; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %shuf20 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 0>
172; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf22 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 2>
173; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf03 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 3>
174; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf30 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 0>
175; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf33 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 3>
176; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf12 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 2>
177; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf21 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 1>
178; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf13 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 3>
179; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf31 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 1>
180; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 3>
181; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf32 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 2>
182; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf000 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> zeroinitializer
183; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf001 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
184; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf010 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
185; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf011 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
186; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf100 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
187; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf101 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
188; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf110 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
189; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf111 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
190; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf002 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
191; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf020 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
192; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf022 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
193; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf200 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
194; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf202 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
195; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf220 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
196; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf222 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
197; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf112 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
198; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf121 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
199; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf122 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
200; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf211 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
201; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf212 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
202; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf221 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
203; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf00_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> zeroinitializer
204; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 1>
205; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf10_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 0>
206; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf11_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 1>
207; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf02_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 2>
208; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %shuf20_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 0>
209; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf22_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 2>
210; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf03_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 3>
211; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf30_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 0>
212; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf33_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 3>
213; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf12_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 2>
214; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf21_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 1>
215; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf13_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 3>
216; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf31_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 1>
217; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 3>
218; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf32_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 2>
219; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf000_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> zeroinitializer
220; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf001_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
221; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf010_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
222; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf011_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
223; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf100_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
224; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf101_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
225; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf110_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
226; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf111_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
227; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf002_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
228; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf020_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
229; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf022_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
230; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf200_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
231; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf202_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
232; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf220_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
233; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf222_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
234; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf112_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
235; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf121_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
236; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf122_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
237; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf211_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
238; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf212_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
239; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf221_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
240; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
241;
242; VI-SIZE-LABEL: 'shufflevector_i16'
243; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf00 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> zeroinitializer
244; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 1>
245; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf10 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 0>
246; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf11 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 1>
247; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf02 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 2>
248; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %shuf20 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 0>
249; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf22 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 2>
250; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf03 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 3>
251; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf30 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 0>
252; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf33 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 3>
253; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf12 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 2>
254; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf21 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 1>
255; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf13 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 3>
256; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf31 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 1>
257; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 3>
258; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf32 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 2>
259; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf000 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> zeroinitializer
260; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf001 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
261; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf010 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
262; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf011 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
263; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf100 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
264; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf101 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
265; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf110 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
266; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf111 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
267; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf002 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
268; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf020 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
269; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf022 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
270; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf200 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
271; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf202 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
272; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf220 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
273; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf222 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
274; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf112 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
275; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf121 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
276; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf122 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
277; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf211 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
278; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf212 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
279; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf221 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
280; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf00_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> zeroinitializer
281; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 1>
282; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf10_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 0>
283; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf11_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 1>
284; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf02_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 2>
285; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %shuf20_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 0>
286; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf22_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 2>
287; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf03_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 3>
288; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf30_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 0>
289; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf33_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 3>
290; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf12_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 2>
291; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf21_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 1>
292; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf13_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 3>
293; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf31_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 1>
294; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 3>
295; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf32_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 2>
296; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf000_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> zeroinitializer
297; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf001_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
298; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf010_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
299; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf011_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
300; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf100_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
301; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf101_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
302; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf110_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
303; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf111_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
304; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf002_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
305; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf020_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
306; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf022_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
307; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf200_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
308; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf202_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
309; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf220_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
310; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf222_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
311; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf112_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
312; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf121_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
313; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf122_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
314; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf211_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
315; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf212_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
316; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf221_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
317; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
318;
319  %shuf00 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> zeroinitializer
320  %shuf01 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 1>
321  %shuf10 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 0>
322  %shuf11 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 1>
323  %shuf02 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 2>
324  %shuf20 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 0>
325  %shuf22 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 2>
326  %shuf03 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 0, i32 3>
327  %shuf30 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 0>
328  %shuf33 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 3>
329  %shuf12 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 2>
330  %shuf21 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 1>
331  %shuf13 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 1, i32 3>
332  %shuf31 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 1>
333  %shuf23 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 2, i32 3>
334  %shuf32 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <2 x i32> <i32 3, i32 2>
335  %shuf000 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 0>
336  %shuf001 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
337  %shuf010 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
338  %shuf011 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
339  %shuf100 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
340  %shuf101 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
341  %shuf110 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
342  %shuf111 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
343  %shuf002 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
344  %shuf020 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
345  %shuf022 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
346  %shuf200 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
347  %shuf202 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
348  %shuf220 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
349  %shuf222 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
350  %shuf112 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
351  %shuf121 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
352  %shuf122 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
353  %shuf211 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
354  %shuf212 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
355  %shuf221 = shufflevector <2 x i16> %vec1, <2 x i16> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
356  %shuf00_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> zeroinitializer
357  %shuf01_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 1>
358  %shuf10_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 0>
359  %shuf11_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 1>
360  %shuf02_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 2>
361  %shuf20_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 0>
362  %shuf22_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 2>
363  %shuf03_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 0, i32 3>
364  %shuf30_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 0>
365  %shuf33_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 3>
366  %shuf12_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 2>
367  %shuf21_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 1>
368  %shuf13_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 1, i32 3>
369  %shuf31_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 1>
370  %shuf23_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 2, i32 3>
371  %shuf32_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <2 x i32> <i32 3, i32 2>
372  %shuf000_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 0>
373  %shuf001_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
374  %shuf010_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
375  %shuf011_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
376  %shuf100_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
377  %shuf101_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
378  %shuf110_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
379  %shuf111_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
380  %shuf002_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
381  %shuf020_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
382  %shuf022_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
383  %shuf200_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
384  %shuf202_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
385  %shuf220_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
386  %shuf222_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
387  %shuf112_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
388  %shuf121_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
389  %shuf122_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
390  %shuf211_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
391  %shuf212_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
392  %shuf221_2 = shufflevector <2 x i16> %vec1, <2 x i16> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
393  ret void
394}
395
396; Should not assert
397define amdgpu_kernel void @shufflevector_i8(<2 x i8> %vec1, <2 x i8> %vec2) {
398; ALL-LABEL: 'shufflevector_i8'
399; ALL-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf00 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> zeroinitializer
400; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 0, i32 1>
401; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf10 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 0>
402; ALL-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf11 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 1>
403; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf02 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 0, i32 2>
404; ALL-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf20 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 0>
405; ALL-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf22 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 2>
406; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf03 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 0, i32 3>
407; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf30 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 0>
408; ALL-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf33 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 3>
409; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf12 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 2>
410; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf21 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 1>
411; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf13 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 3>
412; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf31 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 1>
413; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 3>
414; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf32 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 2>
415; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf000 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> zeroinitializer
416; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf001 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
417; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf010 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
418; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf011 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
419; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf100 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
420; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf101 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
421; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf110 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
422; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf111 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
423; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf002 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
424; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf020 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
425; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf022 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
426; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf200 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
427; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf202 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
428; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf220 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
429; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf222 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
430; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf112 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
431; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf121 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
432; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf122 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
433; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf211 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
434; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf212 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
435; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf221 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
436; ALL-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf00_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> zeroinitializer
437; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 0, i32 1>
438; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf10_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 0>
439; ALL-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf11_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 1>
440; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf02_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 0, i32 2>
441; ALL-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf20_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 0>
442; ALL-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf22_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 2>
443; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf03_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 0, i32 3>
444; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf30_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 0>
445; ALL-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf33_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 3>
446; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf12_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 2>
447; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf21_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 1>
448; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf13_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 3>
449; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf31_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 1>
450; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 3>
451; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf32_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 2>
452; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf000_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> zeroinitializer
453; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf001_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
454; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf010_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
455; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf011_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
456; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf100_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
457; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf101_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
458; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf110_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
459; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf111_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
460; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf002_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
461; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf020_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
462; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf022_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
463; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf200_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
464; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf202_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
465; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf220_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
466; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf222_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
467; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf112_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
468; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf121_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
469; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf122_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
470; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf211_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
471; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf212_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
472; ALL-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf221_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
473; ALL-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: ret void
474;
475; ALL-SIZE-LABEL: 'shufflevector_i8'
476; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf00 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> zeroinitializer
477; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 0, i32 1>
478; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf10 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 0>
479; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf11 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 1>
480; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf02 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 0, i32 2>
481; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf20 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 0>
482; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf22 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 2>
483; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf03 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 0, i32 3>
484; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf30 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 0>
485; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf33 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 3>
486; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf12 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 2>
487; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf21 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 1>
488; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf13 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 3>
489; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf31 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 1>
490; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 3>
491; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf32 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 2>
492; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf000 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> zeroinitializer
493; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf001 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
494; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf010 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
495; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf011 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
496; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf100 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
497; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf101 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
498; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf110 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
499; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf111 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
500; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf002 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
501; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf020 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
502; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf022 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
503; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf200 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
504; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf202 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
505; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf220 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
506; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf222 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
507; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf112 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
508; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf121 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
509; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf122 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
510; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf211 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
511; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf212 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
512; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf221 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
513; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf00_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> zeroinitializer
514; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 0, i32 1>
515; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf10_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 0>
516; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf11_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 1>
517; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf02_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 0, i32 2>
518; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %shuf20_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 0>
519; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf22_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 2>
520; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf03_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 0, i32 3>
521; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf30_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 0>
522; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %shuf33_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 3>
523; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf12_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 2>
524; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf21_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 1>
525; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf13_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 3>
526; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf31_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 1>
527; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 3>
528; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %shuf32_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 2>
529; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf000_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> zeroinitializer
530; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf001_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
531; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf010_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
532; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf011_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
533; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf100_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
534; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf101_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
535; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf110_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
536; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf111_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
537; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf002_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
538; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf020_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
539; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf022_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
540; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf200_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
541; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf202_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
542; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf220_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
543; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf222_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
544; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf112_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
545; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf121_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
546; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf122_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
547; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf211_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
548; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf212_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
549; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %shuf221_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
550; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
551;
552  %shuf00 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> zeroinitializer
553  %shuf01 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 0, i32 1>
554  %shuf10 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 0>
555  %shuf11 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 1>
556  %shuf02 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 0, i32 2>
557  %shuf20 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 0>
558  %shuf22 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 2>
559  %shuf03 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 0, i32 3>
560  %shuf30 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 0>
561  %shuf33 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 3>
562  %shuf12 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 2>
563  %shuf21 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 1>
564  %shuf13 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 1, i32 3>
565  %shuf31 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 1>
566  %shuf23 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 2, i32 3>
567  %shuf32 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <2 x i32> <i32 3, i32 2>
568  %shuf000 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 0, i32 0>
569  %shuf001 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
570  %shuf010 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
571  %shuf011 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
572  %shuf100 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
573  %shuf101 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
574  %shuf110 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
575  %shuf111 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
576  %shuf002 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
577  %shuf020 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
578  %shuf022 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
579  %shuf200 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
580  %shuf202 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
581  %shuf220 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
582  %shuf222 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
583  %shuf112 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
584  %shuf121 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
585  %shuf122 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
586  %shuf211 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
587  %shuf212 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
588  %shuf221 = shufflevector <2 x i8> %vec1, <2 x i8> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
589  %shuf00_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> zeroinitializer
590  %shuf01_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 0, i32 1>
591  %shuf10_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 0>
592  %shuf11_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 1>
593  %shuf02_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 0, i32 2>
594  %shuf20_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 0>
595  %shuf22_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 2>
596  %shuf03_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 0, i32 3>
597  %shuf30_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 0>
598  %shuf33_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 3>
599  %shuf12_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 2>
600  %shuf21_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 1>
601  %shuf13_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 1, i32 3>
602  %shuf31_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 1>
603  %shuf23_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 2, i32 3>
604  %shuf32_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <2 x i32> <i32 3, i32 2>
605  %shuf000_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 0, i32 0>
606  %shuf001_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
607  %shuf010_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
608  %shuf011_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
609  %shuf100_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
610  %shuf101_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
611  %shuf110_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
612  %shuf111_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
613  %shuf002_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
614  %shuf020_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
615  %shuf022_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
616  %shuf200_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
617  %shuf202_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
618  %shuf220_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
619  %shuf222_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
620  %shuf112_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
621  %shuf121_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
622  %shuf122_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
623  %shuf211_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
624  %shuf212_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
625  %shuf221_2 = shufflevector <2 x i8> %vec1, <2 x i8> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
626  ret void
627}
628
629define amdgpu_kernel void @shufflevector_i32(<2 x i32> %vec1, <2 x i32> %vec2) {
630; ALL-LABEL: 'shufflevector_i32'
631; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf00 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> zeroinitializer
632; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 0, i32 1>
633; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf10 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 0>
634; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf11 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 1>
635; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf02 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 0, i32 2>
636; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf20 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 0>
637; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf22 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 2>
638; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf03 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 0, i32 3>
639; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf30 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 0>
640; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf33 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 3>
641; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf12 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 2>
642; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf21 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 1>
643; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf13 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 3>
644; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf31 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 1>
645; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 3>
646; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf32 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 2>
647; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf000 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> zeroinitializer
648; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf001 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
649; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf010 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
650; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf011 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
651; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf100 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
652; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf101 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
653; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf110 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
654; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf111 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
655; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf002 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
656; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf020 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
657; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf022 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
658; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf200 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
659; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf202 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
660; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf220 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
661; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf222 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
662; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf112 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
663; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf121 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
664; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf122 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
665; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf211 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
666; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf212 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
667; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf221 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
668; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf00_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> zeroinitializer
669; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 0, i32 1>
670; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf10_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 0>
671; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf11_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 1>
672; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf02_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 0, i32 2>
673; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf20_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 0>
674; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf22_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 2>
675; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf03_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 0, i32 3>
676; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf30_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 0>
677; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf33_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 3>
678; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf12_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 2>
679; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf21_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 1>
680; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf13_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 3>
681; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf31_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 1>
682; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 3>
683; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf32_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 2>
684; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf000_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> zeroinitializer
685; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf001_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
686; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf010_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
687; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf011_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
688; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf100_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
689; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf101_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
690; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf110_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
691; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf111_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
692; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf002_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
693; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf020_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
694; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf022_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
695; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf200_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
696; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf202_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
697; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf220_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
698; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf222_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
699; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf112_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
700; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf121_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
701; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf122_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
702; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf211_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
703; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf212_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
704; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf221_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
705; ALL-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: ret void
706;
707; ALL-SIZE-LABEL: 'shufflevector_i32'
708; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf00 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> zeroinitializer
709; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 0, i32 1>
710; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf10 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 0>
711; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf11 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 1>
712; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf02 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 0, i32 2>
713; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf20 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 0>
714; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf22 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 2>
715; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf03 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 0, i32 3>
716; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf30 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 0>
717; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf33 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 3>
718; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf12 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 2>
719; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf21 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 1>
720; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf13 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 3>
721; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf31 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 1>
722; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 3>
723; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf32 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 2>
724; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf000 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> zeroinitializer
725; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf001 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
726; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf010 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
727; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf011 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
728; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf100 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
729; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf101 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
730; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf110 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
731; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf111 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
732; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf002 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
733; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf020 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
734; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf022 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
735; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf200 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
736; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf202 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
737; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf220 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
738; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf222 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
739; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf112 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
740; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf121 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
741; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf122 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
742; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf211 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
743; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf212 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
744; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf221 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
745; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf00_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> zeroinitializer
746; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf01_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 0, i32 1>
747; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf10_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 0>
748; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf11_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 1>
749; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf02_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 0, i32 2>
750; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf20_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 0>
751; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf22_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 2>
752; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf03_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 0, i32 3>
753; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf30_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 0>
754; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf33_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 3>
755; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf12_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 2>
756; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf21_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 1>
757; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf13_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 3>
758; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf31_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 1>
759; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf23_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 3>
760; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf32_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 2>
761; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf000_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> zeroinitializer
762; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf001_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
763; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf010_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
764; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf011_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
765; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf100_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
766; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf101_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
767; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf110_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
768; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf111_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
769; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf002_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
770; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf020_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
771; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf022_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
772; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf200_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
773; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf202_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
774; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf220_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
775; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf222_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
776; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf112_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
777; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf121_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
778; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf122_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
779; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf211_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
780; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf212_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
781; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %shuf221_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
782; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
783;
784  %shuf00 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> zeroinitializer
785  %shuf01 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 0, i32 1>
786  %shuf10 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 0>
787  %shuf11 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 1>
788  %shuf02 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 0, i32 2>
789  %shuf20 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 0>
790  %shuf22 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 2>
791  %shuf03 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 0, i32 3>
792  %shuf30 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 0>
793  %shuf33 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 3>
794  %shuf12 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 2>
795  %shuf21 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 1>
796  %shuf13 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 1, i32 3>
797  %shuf31 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 1>
798  %shuf23 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 2, i32 3>
799  %shuf32 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <2 x i32> <i32 3, i32 2>
800  %shuf000 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 0, i32 0>
801  %shuf001 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 0, i32 1>
802  %shuf010 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 1, i32 0>
803  %shuf011 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 1, i32 1>
804  %shuf100 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 0, i32 0>
805  %shuf101 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 0, i32 1>
806  %shuf110 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 1, i32 0>
807  %shuf111 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 1, i32 1>
808  %shuf002 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 0, i32 2>
809  %shuf020 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 2, i32 0>
810  %shuf022 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 0, i32 2, i32 2>
811  %shuf200 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 0, i32 0>
812  %shuf202 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 0, i32 2>
813  %shuf220 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 2, i32 0>
814  %shuf222 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 2, i32 2>
815  %shuf112 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 1, i32 2>
816  %shuf121 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 2, i32 1>
817  %shuf122 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 1, i32 2, i32 2>
818  %shuf211 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 1, i32 1>
819  %shuf212 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 1, i32 2>
820  %shuf221 = shufflevector <2 x i32> %vec1, <2 x i32> %vec1, <3 x i32> <i32 2, i32 2, i32 1>
821  %shuf00_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> zeroinitializer
822  %shuf01_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 0, i32 1>
823  %shuf10_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 0>
824  %shuf11_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 1>
825  %shuf02_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 0, i32 2>
826  %shuf20_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 0>
827  %shuf22_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 2>
828  %shuf03_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 0, i32 3>
829  %shuf30_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 0>
830  %shuf33_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 3>
831  %shuf12_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 2>
832  %shuf21_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 1>
833  %shuf13_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 1, i32 3>
834  %shuf31_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 1>
835  %shuf23_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 2, i32 3>
836  %shuf32_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <2 x i32> <i32 3, i32 2>
837  %shuf000_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 0, i32 0>
838  %shuf001_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 0, i32 1>
839  %shuf010_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 1, i32 0>
840  %shuf011_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 1, i32 1>
841  %shuf100_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 0, i32 0>
842  %shuf101_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 0, i32 1>
843  %shuf110_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 1, i32 0>
844  %shuf111_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 1, i32 1>
845  %shuf002_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 0, i32 2>
846  %shuf020_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 2, i32 0>
847  %shuf022_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 0, i32 2, i32 2>
848  %shuf200_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 0, i32 0>
849  %shuf202_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 0, i32 2>
850  %shuf220_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 2, i32 0>
851  %shuf222_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 2, i32 2>
852  %shuf112_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 1, i32 2>
853  %shuf121_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 2, i32 1>
854  %shuf122_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 1, i32 2, i32 2>
855  %shuf211_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 1, i32 1>
856  %shuf212_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 1, i32 2>
857  %shuf221_2 = shufflevector <2 x i32> %vec1, <2 x i32> %vec2, <3 x i32> <i32 2, i32 2, i32 1>
858  ret void
859}
860
861; Other shuffle cases
862define void @shuffle(<2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i8> %i8v4, <4 x i8> %i8v4_2, <6 x i8> %i8v6, <6 x i8> %i8v6_2, <8 x i8> %i8v8, <8 x i8> %i8v8_2, <16 x i8> %i8v16, <16 x i8> %i8v16_2, <2 x i16> %i16v2, <2 x i16> %i16v2_2, <4 x i16> %i16v4, <4 x i16> %i16v4_2, <8 x i16> %i16v8, <8 x i16> %i16v8_2, <2 x i32> %i32v2, <2 x i32> %i32v2_2, <4 x i32> %i32v4, <4 x i32> %i32v4_2, <2 x float> %floatv2, <2 x float> %floatv2_2, <4 x float> %floatv4, <4 x float> %floatv4_2,<2 x i64> %i64v2, <2 x i64> %i64v2_2,<2 x double> %doublev2, <2 x double> %doublev2_2) {
863; GFX9-10-LABEL: 'shuffle'
864; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v2i8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <2 x i32> <i32 1, i32 0>
865; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v2i8_2_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <2 x i32> <i32 1, i32 0>
866; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v2i8_4 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
867; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v2i8_4_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
868; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v4i8_4 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
869; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v4i8_4_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
870; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v2i8_8 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
871; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v2i8_8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
872; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v4i8_8 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
873; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v4i8_8_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
874; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v6i8_8 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
875; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v6i8_8_2 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
876; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v8i8_8 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
877; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v8i8_8_2 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
878; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %v16i8_16 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
879; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %v16i8_16_2 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16_2, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
880; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i16_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2, <2 x i32> <i32 1, i32 0>
881; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i16_2_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2_2, <2 x i32> <i32 1, i32 0>
882; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i16_4 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
883; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i16_4_2 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
884; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i16_8 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
885; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i16_8_2 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
886; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i32_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2, <2 x i32> <i32 1, i32 0>
887; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i32_2_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2_2, <2 x i32> <i32 1, i32 0>
888; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_4 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
889; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_4_2 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
890; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f32_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2, <2 x i32> <i32 1, i32 0>
891; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f32_2_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2_2, <2 x i32> <i32 1, i32 0>
892; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_4 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
893; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_4_2 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
894; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i64_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2, <2 x i32> <i32 1, i32 0>
895; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i64_2_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2_2, <2 x i32> <i32 1, i32 0>
896; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f64_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2, <2 x i32> <i32 1, i32 0>
897; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f64_2_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2_2, <2 x i32> <i32 1, i32 0>
898; GFX9-10-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: ret void
899;
900; VI-LABEL: 'shuffle'
901; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v2i8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <2 x i32> <i32 1, i32 0>
902; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v2i8_2_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <2 x i32> <i32 1, i32 0>
903; VI-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v2i8_4 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
904; VI-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v2i8_4_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
905; VI-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v4i8_4 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
906; VI-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v4i8_4_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
907; VI-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v2i8_8 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
908; VI-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v2i8_8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
909; VI-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v4i8_8 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
910; VI-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v4i8_8_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
911; VI-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v6i8_8 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
912; VI-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v6i8_8_2 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
913; VI-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v8i8_8 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
914; VI-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v8i8_8_2 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
915; VI-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %v16i8_16 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
916; VI-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %v16i8_16_2 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16_2, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
917; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2i16_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2, <2 x i32> <i32 1, i32 0>
918; VI-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2i16_2_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2_2, <2 x i32> <i32 1, i32 0>
919; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i16_4 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
920; VI-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i16_4_2 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
921; VI-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i16_8 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
922; VI-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i16_8_2 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
923; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i32_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2, <2 x i32> <i32 1, i32 0>
924; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i32_2_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2_2, <2 x i32> <i32 1, i32 0>
925; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_4 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
926; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_4_2 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
927; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f32_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2, <2 x i32> <i32 1, i32 0>
928; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f32_2_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2_2, <2 x i32> <i32 1, i32 0>
929; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_4 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
930; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_4_2 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
931; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i64_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2, <2 x i32> <i32 1, i32 0>
932; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i64_2_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2_2, <2 x i32> <i32 1, i32 0>
933; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f64_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2, <2 x i32> <i32 1, i32 0>
934; VI-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f64_2_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2_2, <2 x i32> <i32 1, i32 0>
935; VI-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: ret void
936;
937; GFX9-10-SIZE-LABEL: 'shuffle'
938; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v2i8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <2 x i32> <i32 1, i32 0>
939; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v2i8_2_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <2 x i32> <i32 1, i32 0>
940; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v2i8_4 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
941; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v2i8_4_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
942; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v4i8_4 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
943; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v4i8_4_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
944; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v2i8_8 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
945; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v2i8_8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
946; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v4i8_8 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
947; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v4i8_8_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
948; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v6i8_8 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
949; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v6i8_8_2 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
950; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v8i8_8 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
951; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v8i8_8_2 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
952; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %v16i8_16 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
953; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %v16i8_16_2 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16_2, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
954; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i16_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2, <2 x i32> <i32 1, i32 0>
955; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i16_2_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2_2, <2 x i32> <i32 1, i32 0>
956; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i16_4 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
957; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i16_4_2 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
958; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i16_8 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
959; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i16_8_2 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
960; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i32_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2, <2 x i32> <i32 1, i32 0>
961; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i32_2_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2_2, <2 x i32> <i32 1, i32 0>
962; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_4 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
963; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_4_2 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
964; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f32_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2, <2 x i32> <i32 1, i32 0>
965; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f32_2_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2_2, <2 x i32> <i32 1, i32 0>
966; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_4 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
967; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_4_2 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
968; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i64_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2, <2 x i32> <i32 1, i32 0>
969; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i64_2_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2_2, <2 x i32> <i32 1, i32 0>
970; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f64_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2, <2 x i32> <i32 1, i32 0>
971; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f64_2_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2_2, <2 x i32> <i32 1, i32 0>
972; GFX9-10-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
973;
974; VI-SIZE-LABEL: 'shuffle'
975; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v2i8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <2 x i32> <i32 1, i32 0>
976; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v2i8_2_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <2 x i32> <i32 1, i32 0>
977; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v2i8_4 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
978; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v2i8_4_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
979; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v4i8_4 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
980; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v4i8_4_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
981; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v2i8_8 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
982; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v2i8_8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
983; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v4i8_8 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
984; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v4i8_8_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
985; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v6i8_8 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
986; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v6i8_8_2 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
987; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v8i8_8 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
988; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v8i8_8_2 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
989; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %v16i8_16 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
990; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 32 for instruction: %v16i8_16_2 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16_2, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
991; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2i16_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2, <2 x i32> <i32 1, i32 0>
992; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %v2i16_2_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2_2, <2 x i32> <i32 1, i32 0>
993; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i16_4 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
994; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i16_4_2 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
995; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i16_8 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
996; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i16_8_2 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
997; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i32_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2, <2 x i32> <i32 1, i32 0>
998; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i32_2_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2_2, <2 x i32> <i32 1, i32 0>
999; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_4 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1000; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_4_2 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1001; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f32_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2, <2 x i32> <i32 1, i32 0>
1002; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f32_2_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2_2, <2 x i32> <i32 1, i32 0>
1003; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_4 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1004; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_4_2 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1005; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i64_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2, <2 x i32> <i32 1, i32 0>
1006; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2i64_2_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2_2, <2 x i32> <i32 1, i32 0>
1007; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f64_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2, <2 x i32> <i32 1, i32 0>
1008; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v2f64_2_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2_2, <2 x i32> <i32 1, i32 0>
1009; VI-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
1010;
1011  %v2i8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <2 x i32> <i32 1, i32 0>
1012  %v2i8_2_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <2 x i32> <i32 1, i32 0>
1013  %v2i8_4 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1014  %v2i8_4_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1015  %v4i8_4 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1016  %v4i8_4_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1017  %v2i8_8 = shufflevector <2 x i8>  %i8v2, <2 x i8>  %i8v2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1018  %v2i8_8_2 = shufflevector <2 x i8>  %i8v2, <2 x i8>  %i8v2_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1019  %v4i8_8 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1020  %v4i8_8_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1021  %v6i8_8 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1022  %v6i8_8_2 = shufflevector <6 x i8> %i8v6, <6 x i8> %i8v6_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1023  %v8i8_8 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1024  %v8i8_8_2 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1025  %v16i8_16 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1026  %v16i8_16_2 = shufflevector <16 x i8> %i8v16, <16 x i8> %i8v16_2, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1027  %v2i16_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2, <2 x i32> <i32 1, i32 0>
1028  %v2i16_2_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2_2, <2 x i32> <i32 1, i32 0>
1029  %v4i16_4 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1030  %v4i16_4_2 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1031  %v8i16_8 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1032  %v8i16_8_2 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8_2, <8 x i32> <i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1>
1033  %v2i32_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2, <2 x i32> <i32 1, i32 0>
1034  %v2i32_2_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2_2, <2 x i32> <i32 1, i32 0>
1035  %v4i32_4 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1036  %v4i32_4_2 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1037  %v2f32_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2, <2 x i32> <i32 1, i32 0>
1038  %v2f32_2_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2_2, <2 x i32> <i32 1, i32 0>
1039  %v4f32_4 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1040  %v4f32_4_2 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4_2, <4 x i32> <i32 1, i32 3, i32 2, i32 0>
1041  %v2i64_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2, <2 x i32> <i32 1, i32 0>
1042  %v2i64_2_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2_2, <2 x i32> <i32 1, i32 0>
1043  %v2f64_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2, <2 x i32> <i32 1, i32 0>
1044  %v2f64_2_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2_2, <2 x i32> <i32 1, i32 0>
1045  ret void
1046}
1047
1048define void @concat(<2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i8> %i8v4, <4 x i8> %i8v4_2, <8 x i8> %i8v8, <8 x i8> %i8v8_2, <2 x half> %halfv2, <2 x half> %halfv2_2, <4 x half> %halfv4, <4 x half> %halfv4_2, <8 x half> %halfv8, <8 x half> %halfv8_2, <2 x i16> %i16v2, <2 x i16> %i16v2_2, <4 x i16> %i16v4, <4 x i16> %i16v4_2, <8 x i16> %i16v8, <8 x i16> %i16v8_2, <2 x i32> %i32v2, <2 x i32> %i32v2_2, <4 x i32> %i32v4, <4 x i32> %i32v4_2, <2 x float> %floatv2, <2 x float> %floatv2_2, <4 x float> %floatv4, <4 x float> %floatv4_2,<2 x i64> %i64v2, <2 x i64> %i64v2_2,<2 x double> %doublev2, <2 x double> %doublev2_2) {
1049; ALL-LABEL: 'concat'
1050; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i8 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1051; ALL-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i8 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1052; ALL-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v16i8 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1053; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i16 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1054; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8i16 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1055; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v16i16 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1056; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1057; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8i32 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1058; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i64 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1059; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f16 = shufflevector <2 x half> %halfv2, <2 x half> %halfv2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1060; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8f16 = shufflevector <4 x half> %halfv4, <4 x half> %halfv4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1061; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v16f16 = shufflevector <8 x half> %halfv8, <8 x half> %halfv8, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1062; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1063; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8f32 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1064; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f64 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1065; ALL-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1066; ALL-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i8_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1067; ALL-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v16i8_2 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8_2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1068; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i16_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1069; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8i16_2 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1070; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v16i16_2 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8_2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1071; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1072; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8i32_2 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1073; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i64_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1074; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f16_2 = shufflevector <2 x half> %halfv2, <2 x half> %halfv2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1075; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8f16_2 = shufflevector <4 x half> %halfv4, <4 x half> %halfv4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1076; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v16f16_2 = shufflevector <8 x half> %halfv8, <8 x half> %halfv8_2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1077; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1078; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8f32_2 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1079; ALL-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f64_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1080; ALL-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: ret void
1081;
1082; ALL-SIZE-LABEL: 'concat'
1083; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i8 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1084; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i8 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1085; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v16i8 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1086; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i16 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1087; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8i16 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1088; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v16i16 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1089; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1090; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8i32 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1091; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i64 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1092; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f16 = shufflevector <2 x half> %halfv2, <2 x half> %halfv2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1093; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8f16 = shufflevector <4 x half> %halfv4, <4 x half> %halfv4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1094; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v16f16 = shufflevector <8 x half> %halfv8, <8 x half> %halfv8, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1095; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1096; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8f32 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1097; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f64 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1098; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 4 for instruction: %v4i8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1099; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %v8i8_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1100; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 16 for instruction: %v16i8_2 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8_2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1101; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i16_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1102; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8i16_2 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1103; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v16i16_2 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8_2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1104; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i32_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1105; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8i32_2 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1106; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4i64_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1107; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f16_2 = shufflevector <2 x half> %halfv2, <2 x half> %halfv2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1108; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8f16_2 = shufflevector <4 x half> %halfv4, <4 x half> %halfv4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1109; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v16f16_2 = shufflevector <8 x half> %halfv8, <8 x half> %halfv8_2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1110; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f32_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1111; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v8f32_2 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1112; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: %v4f64_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1113; ALL-SIZE-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
1114;
1115  %v4i8 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1116  %v8i8 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1117  %v16i8 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1118  %v4i16 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1119  %v8i16 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1120  %v16i16 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1121  %v4i32 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1122  %v8i32 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1123  %v4i64 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1124  %v4f16 = shufflevector <2 x half> %halfv2, <2 x half> %halfv2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1125  %v8f16 = shufflevector <4 x half> %halfv4, <4 x half> %halfv4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1126  %v16f16 = shufflevector <8 x half> %halfv8, <8 x half> %halfv8, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1127  %v4f32 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1128  %v8f32 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1129  %v4f64 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1130  %v4i8_2 = shufflevector <2 x i8> %i8v2, <2 x i8> %i8v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1131  %v8i8_2 = shufflevector <4 x i8> %i8v4, <4 x i8> %i8v4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1132  %v16i8_2 = shufflevector <8 x i8> %i8v8, <8 x i8> %i8v8_2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1133  %v4i16_2 = shufflevector <2 x i16> %i16v2, <2 x i16> %i16v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1134  %v8i16_2 = shufflevector <4 x i16> %i16v4, <4 x i16> %i16v4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1135  %v16i16_2 = shufflevector <8 x i16> %i16v8, <8 x i16> %i16v8_2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1136  %v4i32_2 = shufflevector <2 x i32> %i32v2, <2 x i32> %i32v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1137  %v8i32_2 = shufflevector <4 x i32> %i32v4, <4 x i32> %i32v4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1138  %v4i64_2 = shufflevector <2 x i64> %i64v2, <2 x i64> %i64v2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1139  %v4f16_2 = shufflevector <2 x half> %halfv2, <2 x half> %halfv2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1140  %v8f16_2 = shufflevector <4 x half> %halfv4, <4 x half> %halfv4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1141  %v16f16_2 = shufflevector <8 x half> %halfv8, <8 x half> %halfv8_2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1142  %v4f32_2 = shufflevector <2 x float> %floatv2, <2 x float> %floatv2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1143  %v8f32_2 = shufflevector <4 x float> %floatv4, <4 x float> %floatv4_2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1144  %v4f64_2 = shufflevector <2 x double> %doublev2, <2 x double> %doublev2_2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1145  ret void
1146}
1147