1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE2
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+ssse3 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSSE3
4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+sse4.1 | FileCheck %s --check-prefix=ALL --check-prefix=SSE --check-prefix=SSE41
5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+avx | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX1
6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=x86-64 -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX --check-prefix=AVX2
7
8;
9; Unary shuffle indices from registers
10;
11
12define <2 x double> @var_shuffle_v2f64_v2f64_xx_i64(<2 x double> %x, i64 %i0, i64 %i1) nounwind {
13; SSE-LABEL: var_shuffle_v2f64_v2f64_xx_i64:
14; SSE:       # BB#0:
15; SSE-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
16; SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
17; SSE-NEXT:    movhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
18; SSE-NEXT:    retq
19;
20; AVX-LABEL: var_shuffle_v2f64_v2f64_xx_i64:
21; AVX:       # BB#0:
22; AVX-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
23; AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
24; AVX-NEXT:    vmovhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
25; AVX-NEXT:    retq
26  %x0 = extractelement <2 x double> %x, i64 %i0
27  %x1 = extractelement <2 x double> %x, i64 %i1
28  %r0 = insertelement <2 x double> undef, double %x0, i32 0
29  %r1 = insertelement <2 x double>   %r0, double %x1, i32 1
30  ret <2 x double> %r1
31}
32
33define <2 x i64> @var_shuffle_v2i64_v2i64_xx_i64(<2 x i64> %x, i32 %i0, i32 %i1) nounwind {
34; SSE-LABEL: var_shuffle_v2i64_v2i64_xx_i64:
35; SSE:       # BB#0:
36; SSE-NEXT:    movslq %edi, %rax
37; SSE-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
38; SSE-NEXT:    movslq %esi, %rcx
39; SSE-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
40; SSE-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
41; SSE-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
42; SSE-NEXT:    retq
43;
44; AVX-LABEL: var_shuffle_v2i64_v2i64_xx_i64:
45; AVX:       # BB#0:
46; AVX-NEXT:    movslq %edi, %rax
47; AVX-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
48; AVX-NEXT:    movslq %esi, %rcx
49; AVX-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
50; AVX-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
51; AVX-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
52; AVX-NEXT:    retq
53  %x0 = extractelement <2 x i64> %x, i32 %i0
54  %x1 = extractelement <2 x i64> %x, i32 %i1
55  %r0 = insertelement <2 x i64> undef, i64 %x0, i32 0
56  %r1 = insertelement <2 x i64>   %r0, i64 %x1, i32 1
57  ret <2 x i64> %r1
58}
59
60define <4 x float> @var_shuffle_v4f32_v4f32_xxxx_i32(<4 x float> %x, i32 %i0, i32 %i1, i32 %i2, i32 %i3) nounwind {
61; SSE2-LABEL: var_shuffle_v4f32_v4f32_xxxx_i32:
62; SSE2:       # BB#0:
63; SSE2-NEXT:    movslq %edi, %rax
64; SSE2-NEXT:    movslq %esi, %rsi
65; SSE2-NEXT:    movslq %edx, %rdx
66; SSE2-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
67; SSE2-NEXT:    movslq %ecx, %rcx
68; SSE2-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
69; SSE2-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
70; SSE2-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
71; SSE2-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
72; SSE2-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
73; SSE2-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
74; SSE2-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
75; SSE2-NEXT:    retq
76;
77; SSSE3-LABEL: var_shuffle_v4f32_v4f32_xxxx_i32:
78; SSSE3:       # BB#0:
79; SSSE3-NEXT:    movslq %edi, %rax
80; SSSE3-NEXT:    movslq %esi, %rsi
81; SSSE3-NEXT:    movslq %edx, %rdx
82; SSSE3-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
83; SSSE3-NEXT:    movslq %ecx, %rcx
84; SSSE3-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
85; SSSE3-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
86; SSSE3-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
87; SSSE3-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
88; SSSE3-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
89; SSSE3-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
90; SSSE3-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
91; SSSE3-NEXT:    retq
92;
93; SSE41-LABEL: var_shuffle_v4f32_v4f32_xxxx_i32:
94; SSE41:       # BB#0:
95; SSE41-NEXT:    movslq %edi, %rax
96; SSE41-NEXT:    movslq %esi, %rsi
97; SSE41-NEXT:    movslq %edx, %rdx
98; SSE41-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
99; SSE41-NEXT:    movslq %ecx, %rcx
100; SSE41-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
101; SSE41-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
102; SSE41-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
103; SSE41-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
104; SSE41-NEXT:    retq
105;
106; AVX-LABEL: var_shuffle_v4f32_v4f32_xxxx_i32:
107; AVX:       # BB#0:
108; AVX-NEXT:    movslq %edi, %rax
109; AVX-NEXT:    movslq %esi, %rsi
110; AVX-NEXT:    movslq %edx, %rdx
111; AVX-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
112; AVX-NEXT:    movslq %ecx, %rcx
113; AVX-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
114; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
115; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
116; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
117; AVX-NEXT:    retq
118  %x0 = extractelement <4 x float> %x, i32 %i0
119  %x1 = extractelement <4 x float> %x, i32 %i1
120  %x2 = extractelement <4 x float> %x, i32 %i2
121  %x3 = extractelement <4 x float> %x, i32 %i3
122  %r0 = insertelement <4 x float> undef, float %x0, i32 0
123  %r1 = insertelement <4 x float>   %r0, float %x1, i32 1
124  %r2 = insertelement <4 x float>   %r1, float %x2, i32 2
125  %r3 = insertelement <4 x float>   %r2, float %x3, i32 3
126  ret <4 x float> %r3
127}
128
129define <4 x i32> @var_shuffle_v4i32_v4i32_xxxx_i32(<4 x i32> %x, i32 %i0, i32 %i1, i32 %i2, i32 %i3) nounwind {
130; SSE2-LABEL: var_shuffle_v4i32_v4i32_xxxx_i32:
131; SSE2:       # BB#0:
132; SSE2-NEXT:    movslq %edi, %rax
133; SSE2-NEXT:    movslq %esi, %rsi
134; SSE2-NEXT:    movslq %edx, %rdx
135; SSE2-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
136; SSE2-NEXT:    movslq %ecx, %rcx
137; SSE2-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
138; SSE2-NEXT:    movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
139; SSE2-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero
140; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
141; SSE2-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero
142; SSE2-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
143; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
144; SSE2-NEXT:    retq
145;
146; SSSE3-LABEL: var_shuffle_v4i32_v4i32_xxxx_i32:
147; SSSE3:       # BB#0:
148; SSSE3-NEXT:    movslq %edi, %rax
149; SSSE3-NEXT:    movslq %esi, %rsi
150; SSSE3-NEXT:    movslq %edx, %rdx
151; SSSE3-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
152; SSSE3-NEXT:    movslq %ecx, %rcx
153; SSSE3-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
154; SSSE3-NEXT:    movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
155; SSSE3-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero
156; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
157; SSSE3-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero
158; SSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
159; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
160; SSSE3-NEXT:    retq
161;
162; SSE41-LABEL: var_shuffle_v4i32_v4i32_xxxx_i32:
163; SSE41:       # BB#0:
164; SSE41-NEXT:    movslq %edi, %rax
165; SSE41-NEXT:    movslq %esi, %rsi
166; SSE41-NEXT:    movslq %edx, %rdx
167; SSE41-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
168; SSE41-NEXT:    movslq %ecx, %rcx
169; SSE41-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
170; SSE41-NEXT:    pinsrd $1, -24(%rsp,%rsi,4), %xmm0
171; SSE41-NEXT:    pinsrd $2, -24(%rsp,%rdx,4), %xmm0
172; SSE41-NEXT:    pinsrd $3, -24(%rsp,%rcx,4), %xmm0
173; SSE41-NEXT:    retq
174;
175; AVX-LABEL: var_shuffle_v4i32_v4i32_xxxx_i32:
176; AVX:       # BB#0:
177; AVX-NEXT:    movslq %edi, %rax
178; AVX-NEXT:    movslq %esi, %rsi
179; AVX-NEXT:    movslq %edx, %rdx
180; AVX-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
181; AVX-NEXT:    movslq %ecx, %rcx
182; AVX-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
183; AVX-NEXT:    vpinsrd $1, -24(%rsp,%rsi,4), %xmm0, %xmm0
184; AVX-NEXT:    vpinsrd $2, -24(%rsp,%rdx,4), %xmm0, %xmm0
185; AVX-NEXT:    vpinsrd $3, -24(%rsp,%rcx,4), %xmm0, %xmm0
186; AVX-NEXT:    retq
187  %x0 = extractelement <4 x i32> %x, i32 %i0
188  %x1 = extractelement <4 x i32> %x, i32 %i1
189  %x2 = extractelement <4 x i32> %x, i32 %i2
190  %x3 = extractelement <4 x i32> %x, i32 %i3
191  %r0 = insertelement <4 x i32> undef, i32 %x0, i32 0
192  %r1 = insertelement <4 x i32>   %r0, i32 %x1, i32 1
193  %r2 = insertelement <4 x i32>   %r1, i32 %x2, i32 2
194  %r3 = insertelement <4 x i32>   %r2, i32 %x3, i32 3
195  ret <4 x i32> %r3
196}
197
198define <8 x i16> @var_shuffle_v8i16_v8i16_xxxxxxxx_i16(<8 x i16> %x, i16 %i0, i16 %i1, i16 %i2, i16 %i3, i16 %i4, i16 %i5, i16 %i6, i16 %i7) nounwind {
199; SSE2-LABEL: var_shuffle_v8i16_v8i16_xxxxxxxx_i16:
200; SSE2:       # BB#0:
201; SSE2-NEXT:    movswq %di, %rax
202; SSE2-NEXT:    movswq %si, %rsi
203; SSE2-NEXT:    movswq %dx, %rdx
204; SSE2-NEXT:    movswq %cx, %r10
205; SSE2-NEXT:    movswq %r8w, %r11
206; SSE2-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
207; SSE2-NEXT:    movswq %r9w, %r8
208; SSE2-NEXT:    movswq {{[0-9]+}}(%rsp), %rcx
209; SSE2-NEXT:    movswq {{[0-9]+}}(%rsp), %rdi
210; SSE2-NEXT:    movzwl -24(%rsp,%rcx,2), %ecx
211; SSE2-NEXT:    movzwl -24(%rsp,%rdi,2), %edi
212; SSE2-NEXT:    movzwl -24(%rsp,%rax,2), %eax
213; SSE2-NEXT:    movzwl -24(%rsp,%rsi,2), %esi
214; SSE2-NEXT:    movd %ecx, %xmm0
215; SSE2-NEXT:    movzwl -24(%rsp,%rdx,2), %ecx
216; SSE2-NEXT:    movd %ecx, %xmm1
217; SSE2-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
218; SSE2-NEXT:    movzwl -24(%rsp,%r10,2), %ecx
219; SSE2-NEXT:    movd %eax, %xmm0
220; SSE2-NEXT:    movzwl -24(%rsp,%r11,2), %eax
221; SSE2-NEXT:    movd %eax, %xmm2
222; SSE2-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
223; SSE2-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
224; SSE2-NEXT:    movd %edi, %xmm1
225; SSE2-NEXT:    movd %ecx, %xmm2
226; SSE2-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
227; SSE2-NEXT:    movd %esi, %xmm1
228; SSE2-NEXT:    movzwl -24(%rsp,%r8,2), %eax
229; SSE2-NEXT:    movd %eax, %xmm3
230; SSE2-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
231; SSE2-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
232; SSE2-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
233; SSE2-NEXT:    retq
234;
235; SSSE3-LABEL: var_shuffle_v8i16_v8i16_xxxxxxxx_i16:
236; SSSE3:       # BB#0:
237; SSSE3-NEXT:    movswq %di, %rax
238; SSSE3-NEXT:    movswq %si, %rsi
239; SSSE3-NEXT:    movswq %dx, %rdx
240; SSSE3-NEXT:    movswq %cx, %r10
241; SSSE3-NEXT:    movswq %r8w, %r11
242; SSSE3-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
243; SSSE3-NEXT:    movswq %r9w, %r8
244; SSSE3-NEXT:    movswq {{[0-9]+}}(%rsp), %rcx
245; SSSE3-NEXT:    movswq {{[0-9]+}}(%rsp), %rdi
246; SSSE3-NEXT:    movzwl -24(%rsp,%rcx,2), %ecx
247; SSSE3-NEXT:    movzwl -24(%rsp,%rdi,2), %edi
248; SSSE3-NEXT:    movzwl -24(%rsp,%rax,2), %eax
249; SSSE3-NEXT:    movzwl -24(%rsp,%rsi,2), %esi
250; SSSE3-NEXT:    movd %ecx, %xmm0
251; SSSE3-NEXT:    movzwl -24(%rsp,%rdx,2), %ecx
252; SSSE3-NEXT:    movd %ecx, %xmm1
253; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
254; SSSE3-NEXT:    movzwl -24(%rsp,%r10,2), %ecx
255; SSSE3-NEXT:    movd %eax, %xmm0
256; SSSE3-NEXT:    movzwl -24(%rsp,%r11,2), %eax
257; SSSE3-NEXT:    movd %eax, %xmm2
258; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
259; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
260; SSSE3-NEXT:    movd %edi, %xmm1
261; SSSE3-NEXT:    movd %ecx, %xmm2
262; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
263; SSSE3-NEXT:    movd %esi, %xmm1
264; SSSE3-NEXT:    movzwl -24(%rsp,%r8,2), %eax
265; SSSE3-NEXT:    movd %eax, %xmm3
266; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
267; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
268; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
269; SSSE3-NEXT:    retq
270;
271; SSE41-LABEL: var_shuffle_v8i16_v8i16_xxxxxxxx_i16:
272; SSE41:       # BB#0:
273; SSE41-NEXT:    pushq %rbx
274; SSE41-NEXT:    movswq %di, %rax
275; SSE41-NEXT:    movswq %si, %rbx
276; SSE41-NEXT:    movswq %dx, %r11
277; SSE41-NEXT:    movswq %cx, %r10
278; SSE41-NEXT:    movswq %r8w, %rdi
279; SSE41-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
280; SSE41-NEXT:    movswq %r9w, %rcx
281; SSE41-NEXT:    movswq {{[0-9]+}}(%rsp), %rdx
282; SSE41-NEXT:    movswq {{[0-9]+}}(%rsp), %rsi
283; SSE41-NEXT:    movzwl -16(%rsp,%rdx,2), %edx
284; SSE41-NEXT:    movzwl -16(%rsp,%rsi,2), %esi
285; SSE41-NEXT:    movzwl -16(%rsp,%rax,2), %eax
286; SSE41-NEXT:    movd %eax, %xmm0
287; SSE41-NEXT:    pinsrw $1, -16(%rsp,%rbx,2), %xmm0
288; SSE41-NEXT:    pinsrw $2, -16(%rsp,%r11,2), %xmm0
289; SSE41-NEXT:    pinsrw $3, -16(%rsp,%r10,2), %xmm0
290; SSE41-NEXT:    pinsrw $4, -16(%rsp,%rdi,2), %xmm0
291; SSE41-NEXT:    pinsrw $5, -16(%rsp,%rcx,2), %xmm0
292; SSE41-NEXT:    pinsrw $6, %edx, %xmm0
293; SSE41-NEXT:    pinsrw $7, %esi, %xmm0
294; SSE41-NEXT:    popq %rbx
295; SSE41-NEXT:    retq
296;
297; AVX-LABEL: var_shuffle_v8i16_v8i16_xxxxxxxx_i16:
298; AVX:       # BB#0:
299; AVX-NEXT:    pushq %r14
300; AVX-NEXT:    pushq %rbx
301; AVX-NEXT:    movswq %di, %r10
302; AVX-NEXT:    movswq %si, %r11
303; AVX-NEXT:    movswq %dx, %r14
304; AVX-NEXT:    movswq %cx, %rcx
305; AVX-NEXT:    movswq %r8w, %rdi
306; AVX-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
307; AVX-NEXT:    movswq %r9w, %rax
308; AVX-NEXT:    movswq {{[0-9]+}}(%rsp), %rsi
309; AVX-NEXT:    movswq {{[0-9]+}}(%rsp), %rdx
310; AVX-NEXT:    movzwl -24(%rsp,%rsi,2), %esi
311; AVX-NEXT:    movzwl -24(%rsp,%rdx,2), %edx
312; AVX-NEXT:    movzwl -24(%rsp,%r10,2), %ebx
313; AVX-NEXT:    vmovd %ebx, %xmm0
314; AVX-NEXT:    vpinsrw $1, -24(%rsp,%r11,2), %xmm0, %xmm0
315; AVX-NEXT:    vpinsrw $2, -24(%rsp,%r14,2), %xmm0, %xmm0
316; AVX-NEXT:    vpinsrw $3, -24(%rsp,%rcx,2), %xmm0, %xmm0
317; AVX-NEXT:    vpinsrw $4, -24(%rsp,%rdi,2), %xmm0, %xmm0
318; AVX-NEXT:    vpinsrw $5, -24(%rsp,%rax,2), %xmm0, %xmm0
319; AVX-NEXT:    vpinsrw $6, %esi, %xmm0, %xmm0
320; AVX-NEXT:    vpinsrw $7, %edx, %xmm0, %xmm0
321; AVX-NEXT:    popq %rbx
322; AVX-NEXT:    popq %r14
323; AVX-NEXT:    retq
324  %x0 = extractelement <8 x i16> %x, i16 %i0
325  %x1 = extractelement <8 x i16> %x, i16 %i1
326  %x2 = extractelement <8 x i16> %x, i16 %i2
327  %x3 = extractelement <8 x i16> %x, i16 %i3
328  %x4 = extractelement <8 x i16> %x, i16 %i4
329  %x5 = extractelement <8 x i16> %x, i16 %i5
330  %x6 = extractelement <8 x i16> %x, i16 %i6
331  %x7 = extractelement <8 x i16> %x, i16 %i7
332  %r0 = insertelement <8 x i16> undef, i16 %x0, i32 0
333  %r1 = insertelement <8 x i16>   %r0, i16 %x1, i32 1
334  %r2 = insertelement <8 x i16>   %r1, i16 %x2, i32 2
335  %r3 = insertelement <8 x i16>   %r2, i16 %x3, i32 3
336  %r4 = insertelement <8 x i16>   %r3, i16 %x4, i32 4
337  %r5 = insertelement <8 x i16>   %r4, i16 %x5, i32 5
338  %r6 = insertelement <8 x i16>   %r5, i16 %x6, i32 6
339  %r7 = insertelement <8 x i16>   %r6, i16 %x7, i32 7
340  ret <8 x i16> %r7
341}
342
343define <16 x i8> @var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8(<16 x i8> %x, i8 %i0, i8 %i1, i8 %i2, i8 %i3, i8 %i4, i8 %i5, i8 %i6, i8 %i7, i8 %i8, i8 %i9, i8 %i10, i8 %i11, i8 %i12, i8 %i13, i8 %i14, i8 %i15) nounwind {
344; SSE2-LABEL: var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8:
345; SSE2:       # BB#0:
346; SSE2-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
347; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %r10
348; SSE2-NEXT:    leaq -{{[0-9]+}}(%rsp), %r11
349; SSE2-NEXT:    movzbl (%r10,%r11), %eax
350; SSE2-NEXT:    movd %eax, %xmm15
351; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
352; SSE2-NEXT:    movzbl (%rax,%r11), %eax
353; SSE2-NEXT:    movd %eax, %xmm8
354; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
355; SSE2-NEXT:    movzbl (%rax,%r11), %eax
356; SSE2-NEXT:    movd %eax, %xmm9
357; SSE2-NEXT:    movsbq %dl, %rax
358; SSE2-NEXT:    movzbl (%rax,%r11), %eax
359; SSE2-NEXT:    movd %eax, %xmm3
360; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
361; SSE2-NEXT:    movzbl (%rax,%r11), %eax
362; SSE2-NEXT:    movd %eax, %xmm10
363; SSE2-NEXT:    movsbq %dil, %rax
364; SSE2-NEXT:    movzbl (%rax,%r11), %eax
365; SSE2-NEXT:    movd %eax, %xmm0
366; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
367; SSE2-NEXT:    movzbl (%rax,%r11), %eax
368; SSE2-NEXT:    movd %eax, %xmm11
369; SSE2-NEXT:    movsbq %r8b, %rax
370; SSE2-NEXT:    movzbl (%rax,%r11), %eax
371; SSE2-NEXT:    movd %eax, %xmm7
372; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
373; SSE2-NEXT:    movzbl (%rax,%r11), %eax
374; SSE2-NEXT:    movd %eax, %xmm2
375; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
376; SSE2-NEXT:    movzbl (%rax,%r11), %eax
377; SSE2-NEXT:    movd %eax, %xmm12
378; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
379; SSE2-NEXT:    movzbl (%rax,%r11), %eax
380; SSE2-NEXT:    movd %eax, %xmm13
381; SSE2-NEXT:    movsbq %cl, %rax
382; SSE2-NEXT:    movzbl (%rax,%r11), %eax
383; SSE2-NEXT:    movd %eax, %xmm6
384; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
385; SSE2-NEXT:    movzbl (%rax,%r11), %eax
386; SSE2-NEXT:    movd %eax, %xmm14
387; SSE2-NEXT:    movsbq %sil, %rax
388; SSE2-NEXT:    movzbl (%rax,%r11), %eax
389; SSE2-NEXT:    movd %eax, %xmm5
390; SSE2-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
391; SSE2-NEXT:    movzbl (%rax,%r11), %eax
392; SSE2-NEXT:    movd %eax, %xmm4
393; SSE2-NEXT:    movsbq %r9b, %rax
394; SSE2-NEXT:    movzbl (%rax,%r11), %eax
395; SSE2-NEXT:    movd %eax, %xmm1
396; SSE2-NEXT:    punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm8[0],xmm15[1],xmm8[1],xmm15[2],xmm8[2],xmm15[3],xmm8[3],xmm15[4],xmm8[4],xmm15[5],xmm8[5],xmm15[6],xmm8[6],xmm15[7],xmm8[7]
397; SSE2-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3],xmm3[4],xmm9[4],xmm3[5],xmm9[5],xmm3[6],xmm9[6],xmm3[7],xmm9[7]
398; SSE2-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm15[0],xmm3[1],xmm15[1],xmm3[2],xmm15[2],xmm3[3],xmm15[3],xmm3[4],xmm15[4],xmm3[5],xmm15[5],xmm3[6],xmm15[6],xmm3[7],xmm15[7]
399; SSE2-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3],xmm0[4],xmm10[4],xmm0[5],xmm10[5],xmm0[6],xmm10[6],xmm0[7],xmm10[7]
400; SSE2-NEXT:    punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3],xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
401; SSE2-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3],xmm0[4],xmm7[4],xmm0[5],xmm7[5],xmm0[6],xmm7[6],xmm0[7],xmm7[7]
402; SSE2-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
403; SSE2-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm12[0],xmm2[1],xmm12[1],xmm2[2],xmm12[2],xmm2[3],xmm12[3],xmm2[4],xmm12[4],xmm2[5],xmm12[5],xmm2[6],xmm12[6],xmm2[7],xmm12[7]
404; SSE2-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm13[0],xmm6[1],xmm13[1],xmm6[2],xmm13[2],xmm6[3],xmm13[3],xmm6[4],xmm13[4],xmm6[5],xmm13[5],xmm6[6],xmm13[6],xmm6[7],xmm13[7]
405; SSE2-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3],xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
406; SSE2-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm14[0],xmm5[1],xmm14[1],xmm5[2],xmm14[2],xmm5[3],xmm14[3],xmm5[4],xmm14[4],xmm5[5],xmm14[5],xmm5[6],xmm14[6],xmm5[7],xmm14[7]
407; SSE2-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
408; SSE2-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3],xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
409; SSE2-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
410; SSE2-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
411; SSE2-NEXT:    retq
412;
413; SSSE3-LABEL: var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8:
414; SSSE3:       # BB#0:
415; SSSE3-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
416; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %r10
417; SSSE3-NEXT:    leaq -{{[0-9]+}}(%rsp), %r11
418; SSSE3-NEXT:    movzbl (%r10,%r11), %eax
419; SSSE3-NEXT:    movd %eax, %xmm15
420; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
421; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
422; SSSE3-NEXT:    movd %eax, %xmm8
423; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
424; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
425; SSSE3-NEXT:    movd %eax, %xmm9
426; SSSE3-NEXT:    movsbq %dl, %rax
427; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
428; SSSE3-NEXT:    movd %eax, %xmm3
429; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
430; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
431; SSSE3-NEXT:    movd %eax, %xmm10
432; SSSE3-NEXT:    movsbq %dil, %rax
433; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
434; SSSE3-NEXT:    movd %eax, %xmm0
435; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
436; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
437; SSSE3-NEXT:    movd %eax, %xmm11
438; SSSE3-NEXT:    movsbq %r8b, %rax
439; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
440; SSSE3-NEXT:    movd %eax, %xmm7
441; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
442; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
443; SSSE3-NEXT:    movd %eax, %xmm2
444; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
445; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
446; SSSE3-NEXT:    movd %eax, %xmm12
447; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
448; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
449; SSSE3-NEXT:    movd %eax, %xmm13
450; SSSE3-NEXT:    movsbq %cl, %rax
451; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
452; SSSE3-NEXT:    movd %eax, %xmm6
453; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
454; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
455; SSSE3-NEXT:    movd %eax, %xmm14
456; SSSE3-NEXT:    movsbq %sil, %rax
457; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
458; SSSE3-NEXT:    movd %eax, %xmm5
459; SSSE3-NEXT:    movsbq {{[0-9]+}}(%rsp), %rax
460; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
461; SSSE3-NEXT:    movd %eax, %xmm4
462; SSSE3-NEXT:    movsbq %r9b, %rax
463; SSSE3-NEXT:    movzbl (%rax,%r11), %eax
464; SSSE3-NEXT:    movd %eax, %xmm1
465; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm15 = xmm15[0],xmm8[0],xmm15[1],xmm8[1],xmm15[2],xmm8[2],xmm15[3],xmm8[3],xmm15[4],xmm8[4],xmm15[5],xmm8[5],xmm15[6],xmm8[6],xmm15[7],xmm8[7]
466; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3],xmm3[4],xmm9[4],xmm3[5],xmm9[5],xmm3[6],xmm9[6],xmm3[7],xmm9[7]
467; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm15[0],xmm3[1],xmm15[1],xmm3[2],xmm15[2],xmm3[3],xmm15[3],xmm3[4],xmm15[4],xmm3[5],xmm15[5],xmm3[6],xmm15[6],xmm3[7],xmm15[7]
468; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm10[0],xmm0[1],xmm10[1],xmm0[2],xmm10[2],xmm0[3],xmm10[3],xmm0[4],xmm10[4],xmm0[5],xmm10[5],xmm0[6],xmm10[6],xmm0[7],xmm10[7]
469; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3],xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
470; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3],xmm0[4],xmm7[4],xmm0[5],xmm7[5],xmm0[6],xmm7[6],xmm0[7],xmm7[7]
471; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
472; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm12[0],xmm2[1],xmm12[1],xmm2[2],xmm12[2],xmm2[3],xmm12[3],xmm2[4],xmm12[4],xmm2[5],xmm12[5],xmm2[6],xmm12[6],xmm2[7],xmm12[7]
473; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm13[0],xmm6[1],xmm13[1],xmm6[2],xmm13[2],xmm6[3],xmm13[3],xmm6[4],xmm13[4],xmm6[5],xmm13[5],xmm6[6],xmm13[6],xmm6[7],xmm13[7]
474; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3],xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
475; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm14[0],xmm5[1],xmm14[1],xmm5[2],xmm14[2],xmm5[3],xmm14[3],xmm5[4],xmm14[4],xmm5[5],xmm14[5],xmm5[6],xmm14[6],xmm5[7],xmm14[7]
476; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
477; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1],xmm5[2],xmm1[2],xmm5[3],xmm1[3],xmm5[4],xmm1[4],xmm5[5],xmm1[5],xmm5[6],xmm1[6],xmm5[7],xmm1[7]
478; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm6[0],xmm5[1],xmm6[1],xmm5[2],xmm6[2],xmm5[3],xmm6[3],xmm5[4],xmm6[4],xmm5[5],xmm6[5],xmm5[6],xmm6[6],xmm5[7],xmm6[7]
479; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm5[0],xmm0[1],xmm5[1],xmm0[2],xmm5[2],xmm0[3],xmm5[3],xmm0[4],xmm5[4],xmm0[5],xmm5[5],xmm0[6],xmm5[6],xmm0[7],xmm5[7]
480; SSSE3-NEXT:    retq
481;
482; SSE41-LABEL: var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8:
483; SSE41:       # BB#0:
484; SSE41-NEXT:    pushq %rbp
485; SSE41-NEXT:    pushq %r15
486; SSE41-NEXT:    pushq %r14
487; SSE41-NEXT:    pushq %r13
488; SSE41-NEXT:    pushq %r12
489; SSE41-NEXT:    pushq %rbx
490; SSE41-NEXT:    movsbq %dil, %r15
491; SSE41-NEXT:    movsbq %sil, %r14
492; SSE41-NEXT:    movsbq %dl, %r11
493; SSE41-NEXT:    movsbq %cl, %r10
494; SSE41-NEXT:    movsbq %r8b, %r8
495; SSE41-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
496; SSE41-NEXT:    movsbq %r9b, %r9
497; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %r12
498; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %r13
499; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %rbp
500; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %rbx
501; SSE41-NEXT:    leaq -{{[0-9]+}}(%rsp), %rax
502; SSE41-NEXT:    movzbl (%r15,%rax), %ecx
503; SSE41-NEXT:    movd %ecx, %xmm0
504; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %r15
505; SSE41-NEXT:    pinsrb $1, (%r14,%rax), %xmm0
506; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %r14
507; SSE41-NEXT:    pinsrb $2, (%r11,%rax), %xmm0
508; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %r11
509; SSE41-NEXT:    pinsrb $3, (%r10,%rax), %xmm0
510; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %r10
511; SSE41-NEXT:    pinsrb $4, (%r8,%rax), %xmm0
512; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %rcx
513; SSE41-NEXT:    pinsrb $5, (%r9,%rax), %xmm0
514; SSE41-NEXT:    movsbq {{[0-9]+}}(%rsp), %rdx
515; SSE41-NEXT:    movzbl (%r12,%rax), %esi
516; SSE41-NEXT:    movzbl (%r13,%rax), %edi
517; SSE41-NEXT:    movzbl (%rbp,%rax), %ebp
518; SSE41-NEXT:    movzbl (%rbx,%rax), %ebx
519; SSE41-NEXT:    movzbl (%r15,%rax), %r8d
520; SSE41-NEXT:    movzbl (%r14,%rax), %r9d
521; SSE41-NEXT:    movzbl (%r11,%rax), %r11d
522; SSE41-NEXT:    movzbl (%r10,%rax), %r10d
523; SSE41-NEXT:    movzbl (%rcx,%rax), %ecx
524; SSE41-NEXT:    movzbl (%rdx,%rax), %eax
525; SSE41-NEXT:    pinsrb $6, %esi, %xmm0
526; SSE41-NEXT:    pinsrb $7, %edi, %xmm0
527; SSE41-NEXT:    pinsrb $8, %ebp, %xmm0
528; SSE41-NEXT:    pinsrb $9, %ebx, %xmm0
529; SSE41-NEXT:    pinsrb $10, %r8d, %xmm0
530; SSE41-NEXT:    pinsrb $11, %r9d, %xmm0
531; SSE41-NEXT:    pinsrb $12, %r11d, %xmm0
532; SSE41-NEXT:    pinsrb $13, %r10d, %xmm0
533; SSE41-NEXT:    pinsrb $14, %ecx, %xmm0
534; SSE41-NEXT:    pinsrb $15, %eax, %xmm0
535; SSE41-NEXT:    popq %rbx
536; SSE41-NEXT:    popq %r12
537; SSE41-NEXT:    popq %r13
538; SSE41-NEXT:    popq %r14
539; SSE41-NEXT:    popq %r15
540; SSE41-NEXT:    popq %rbp
541; SSE41-NEXT:    retq
542;
543; AVX-LABEL: var_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8:
544; AVX:       # BB#0:
545; AVX-NEXT:    pushq %rbp
546; AVX-NEXT:    pushq %r15
547; AVX-NEXT:    pushq %r14
548; AVX-NEXT:    pushq %r13
549; AVX-NEXT:    pushq %r12
550; AVX-NEXT:    pushq %rbx
551; AVX-NEXT:    movsbq %dil, %r10
552; AVX-NEXT:    movsbq %sil, %r11
553; AVX-NEXT:    movsbq %dl, %r14
554; AVX-NEXT:    movsbq %cl, %r15
555; AVX-NEXT:    movsbq %r8b, %r8
556; AVX-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
557; AVX-NEXT:    movsbq %r9b, %r9
558; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %r12
559; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %r13
560; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %rbp
561; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %rcx
562; AVX-NEXT:    leaq -{{[0-9]+}}(%rsp), %rdi
563; AVX-NEXT:    movzbl (%r10,%rdi), %eax
564; AVX-NEXT:    vmovd %eax, %xmm0
565; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %r10
566; AVX-NEXT:    vpinsrb $1, (%r11,%rdi), %xmm0, %xmm0
567; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %r11
568; AVX-NEXT:    vpinsrb $2, (%r14,%rdi), %xmm0, %xmm0
569; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %r14
570; AVX-NEXT:    vpinsrb $3, (%r15,%rdi), %xmm0, %xmm0
571; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %r15
572; AVX-NEXT:    vpinsrb $4, (%r8,%rdi), %xmm0, %xmm0
573; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %r8
574; AVX-NEXT:    vpinsrb $5, (%r9,%rdi), %xmm0, %xmm0
575; AVX-NEXT:    movsbq {{[0-9]+}}(%rsp), %rsi
576; AVX-NEXT:    movzbl (%r12,%rdi), %edx
577; AVX-NEXT:    movzbl (%r13,%rdi), %ebx
578; AVX-NEXT:    movzbl (%rbp,%rdi), %ebp
579; AVX-NEXT:    movzbl (%rcx,%rdi), %ecx
580; AVX-NEXT:    movzbl (%r10,%rdi), %eax
581; AVX-NEXT:    movzbl (%r11,%rdi), %r9d
582; AVX-NEXT:    movzbl (%r14,%rdi), %r10d
583; AVX-NEXT:    movzbl (%r15,%rdi), %r11d
584; AVX-NEXT:    movzbl (%r8,%rdi), %r8d
585; AVX-NEXT:    movzbl (%rsi,%rdi), %esi
586; AVX-NEXT:    vpinsrb $6, %edx, %xmm0, %xmm0
587; AVX-NEXT:    vpinsrb $7, %ebx, %xmm0, %xmm0
588; AVX-NEXT:    vpinsrb $8, %ebp, %xmm0, %xmm0
589; AVX-NEXT:    vpinsrb $9, %ecx, %xmm0, %xmm0
590; AVX-NEXT:    vpinsrb $10, %eax, %xmm0, %xmm0
591; AVX-NEXT:    vpinsrb $11, %r9d, %xmm0, %xmm0
592; AVX-NEXT:    vpinsrb $12, %r10d, %xmm0, %xmm0
593; AVX-NEXT:    vpinsrb $13, %r11d, %xmm0, %xmm0
594; AVX-NEXT:    vpinsrb $14, %r8d, %xmm0, %xmm0
595; AVX-NEXT:    vpinsrb $15, %esi, %xmm0, %xmm0
596; AVX-NEXT:    popq %rbx
597; AVX-NEXT:    popq %r12
598; AVX-NEXT:    popq %r13
599; AVX-NEXT:    popq %r14
600; AVX-NEXT:    popq %r15
601; AVX-NEXT:    popq %rbp
602; AVX-NEXT:    retq
603  %x0  = extractelement <16 x i8> %x, i8 %i0
604  %x1  = extractelement <16 x i8> %x, i8 %i1
605  %x2  = extractelement <16 x i8> %x, i8 %i2
606  %x3  = extractelement <16 x i8> %x, i8 %i3
607  %x4  = extractelement <16 x i8> %x, i8 %i4
608  %x5  = extractelement <16 x i8> %x, i8 %i5
609  %x6  = extractelement <16 x i8> %x, i8 %i6
610  %x7  = extractelement <16 x i8> %x, i8 %i7
611  %x8  = extractelement <16 x i8> %x, i8 %i8
612  %x9  = extractelement <16 x i8> %x, i8 %i9
613  %x10 = extractelement <16 x i8> %x, i8 %i10
614  %x11 = extractelement <16 x i8> %x, i8 %i11
615  %x12 = extractelement <16 x i8> %x, i8 %i12
616  %x13 = extractelement <16 x i8> %x, i8 %i13
617  %x14 = extractelement <16 x i8> %x, i8 %i14
618  %x15 = extractelement <16 x i8> %x, i8 %i15
619  %r0  = insertelement <16 x i8> undef, i8 %x0 , i32 0
620  %r1  = insertelement <16 x i8>  %r0 , i8 %x1 , i32 1
621  %r2  = insertelement <16 x i8>  %r1 , i8 %x2 , i32 2
622  %r3  = insertelement <16 x i8>  %r2 , i8 %x3 , i32 3
623  %r4  = insertelement <16 x i8>  %r3 , i8 %x4 , i32 4
624  %r5  = insertelement <16 x i8>  %r4 , i8 %x5 , i32 5
625  %r6  = insertelement <16 x i8>  %r5 , i8 %x6 , i32 6
626  %r7  = insertelement <16 x i8>  %r6 , i8 %x7 , i32 7
627  %r8  = insertelement <16 x i8>  %r7 , i8 %x8 , i32 8
628  %r9  = insertelement <16 x i8>  %r8 , i8 %x9 , i32 9
629  %r10 = insertelement <16 x i8>  %r9 , i8 %x10, i32 10
630  %r11 = insertelement <16 x i8>  %r10, i8 %x11, i32 11
631  %r12 = insertelement <16 x i8>  %r11, i8 %x12, i32 12
632  %r13 = insertelement <16 x i8>  %r12, i8 %x13, i32 13
633  %r14 = insertelement <16 x i8>  %r13, i8 %x14, i32 14
634  %r15 = insertelement <16 x i8>  %r14, i8 %x15, i32 15
635  ret <16 x i8> %r15
636}
637
638;
639; Unary shuffle indices from memory
640;
641
642define <4 x i32> @mem_shuffle_v4i32_v4i32_xxxx_i32(<4 x i32> %x, i32* %i) nounwind {
643; SSE2-LABEL: mem_shuffle_v4i32_v4i32_xxxx_i32:
644; SSE2:       # BB#0:
645; SSE2-NEXT:    movslq (%rdi), %rax
646; SSE2-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
647; SSE2-NEXT:    movslq 4(%rdi), %rcx
648; SSE2-NEXT:    movslq 8(%rdi), %rdx
649; SSE2-NEXT:    movslq 12(%rdi), %rsi
650; SSE2-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
651; SSE2-NEXT:    movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
652; SSE2-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero
653; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
654; SSE2-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero
655; SSE2-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
656; SSE2-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
657; SSE2-NEXT:    retq
658;
659; SSSE3-LABEL: mem_shuffle_v4i32_v4i32_xxxx_i32:
660; SSSE3:       # BB#0:
661; SSSE3-NEXT:    movslq (%rdi), %rax
662; SSSE3-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
663; SSSE3-NEXT:    movslq 4(%rdi), %rcx
664; SSSE3-NEXT:    movslq 8(%rdi), %rdx
665; SSSE3-NEXT:    movslq 12(%rdi), %rsi
666; SSSE3-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
667; SSSE3-NEXT:    movd {{.*#+}} xmm1 = mem[0],zero,zero,zero
668; SSSE3-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero
669; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
670; SSSE3-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero
671; SSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
672; SSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
673; SSSE3-NEXT:    retq
674;
675; SSE41-LABEL: mem_shuffle_v4i32_v4i32_xxxx_i32:
676; SSE41:       # BB#0:
677; SSE41-NEXT:    movslq (%rdi), %rax
678; SSE41-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
679; SSE41-NEXT:    movslq 4(%rdi), %rcx
680; SSE41-NEXT:    movslq 8(%rdi), %rdx
681; SSE41-NEXT:    movslq 12(%rdi), %rsi
682; SSE41-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
683; SSE41-NEXT:    pinsrd $1, -24(%rsp,%rcx,4), %xmm0
684; SSE41-NEXT:    pinsrd $2, -24(%rsp,%rdx,4), %xmm0
685; SSE41-NEXT:    pinsrd $3, -24(%rsp,%rsi,4), %xmm0
686; SSE41-NEXT:    retq
687;
688; AVX-LABEL: mem_shuffle_v4i32_v4i32_xxxx_i32:
689; AVX:       # BB#0:
690; AVX-NEXT:    movslq (%rdi), %rax
691; AVX-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
692; AVX-NEXT:    movslq 4(%rdi), %rcx
693; AVX-NEXT:    movslq 8(%rdi), %rdx
694; AVX-NEXT:    movslq 12(%rdi), %rsi
695; AVX-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
696; AVX-NEXT:    vpinsrd $1, -24(%rsp,%rcx,4), %xmm0, %xmm0
697; AVX-NEXT:    vpinsrd $2, -24(%rsp,%rdx,4), %xmm0, %xmm0
698; AVX-NEXT:    vpinsrd $3, -24(%rsp,%rsi,4), %xmm0, %xmm0
699; AVX-NEXT:    retq
700  %p0  = getelementptr inbounds i32, i32* %i, i64 0
701  %p1  = getelementptr inbounds i32, i32* %i, i64 1
702  %p2  = getelementptr inbounds i32, i32* %i, i64 2
703  %p3  = getelementptr inbounds i32, i32* %i, i64 3
704  %i0  = load i32, i32* %p0, align 4
705  %i1  = load i32, i32* %p1, align 4
706  %i2  = load i32, i32* %p2, align 4
707  %i3  = load i32, i32* %p3, align 4
708  %x0 = extractelement <4 x i32> %x, i32 %i0
709  %x1 = extractelement <4 x i32> %x, i32 %i1
710  %x2 = extractelement <4 x i32> %x, i32 %i2
711  %x3 = extractelement <4 x i32> %x, i32 %i3
712  %r0 = insertelement <4 x i32> undef, i32 %x0, i32 0
713  %r1 = insertelement <4 x i32>   %r0, i32 %x1, i32 1
714  %r2 = insertelement <4 x i32>   %r1, i32 %x2, i32 2
715  %r3 = insertelement <4 x i32>   %r2, i32 %x3, i32 3
716  ret <4 x i32> %r3
717}
718
719define <16 x i8> @mem_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8(<16 x i8> %x, i8* %i) nounwind {
720; SSE2-LABEL: mem_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8:
721; SSE2:       # BB#0:
722; SSE2-NEXT:    movsbq (%rdi), %rcx
723; SSE2-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
724; SSE2-NEXT:    leaq -{{[0-9]+}}(%rsp), %rax
725; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
726; SSE2-NEXT:    movd %ecx, %xmm0
727; SSE2-NEXT:    movsbq 8(%rdi), %rcx
728; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
729; SSE2-NEXT:    movd %ecx, %xmm8
730; SSE2-NEXT:    movsbq 12(%rdi), %rcx
731; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
732; SSE2-NEXT:    movd %ecx, %xmm9
733; SSE2-NEXT:    movsbq 4(%rdi), %rcx
734; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
735; SSE2-NEXT:    movd %ecx, %xmm3
736; SSE2-NEXT:    movsbq 14(%rdi), %rcx
737; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
738; SSE2-NEXT:    movd %ecx, %xmm10
739; SSE2-NEXT:    movsbq 6(%rdi), %rcx
740; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
741; SSE2-NEXT:    movd %ecx, %xmm5
742; SSE2-NEXT:    movsbq 10(%rdi), %rcx
743; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
744; SSE2-NEXT:    movd %ecx, %xmm11
745; SSE2-NEXT:    movsbq 2(%rdi), %rcx
746; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
747; SSE2-NEXT:    movd %ecx, %xmm7
748; SSE2-NEXT:    movsbq 15(%rdi), %rcx
749; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
750; SSE2-NEXT:    movd %ecx, %xmm12
751; SSE2-NEXT:    movsbq 7(%rdi), %rcx
752; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
753; SSE2-NEXT:    movd %ecx, %xmm2
754; SSE2-NEXT:    movsbq 11(%rdi), %rcx
755; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
756; SSE2-NEXT:    movd %ecx, %xmm13
757; SSE2-NEXT:    movsbq 3(%rdi), %rcx
758; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
759; SSE2-NEXT:    movd %ecx, %xmm6
760; SSE2-NEXT:    movsbq 13(%rdi), %rcx
761; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
762; SSE2-NEXT:    movd %ecx, %xmm14
763; SSE2-NEXT:    movsbq 5(%rdi), %rcx
764; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
765; SSE2-NEXT:    movd %ecx, %xmm4
766; SSE2-NEXT:    movsbq 9(%rdi), %rcx
767; SSE2-NEXT:    movzbl (%rcx,%rax), %ecx
768; SSE2-NEXT:    movd %ecx, %xmm15
769; SSE2-NEXT:    movsbq 1(%rdi), %rcx
770; SSE2-NEXT:    movzbl (%rcx,%rax), %eax
771; SSE2-NEXT:    movd %eax, %xmm1
772; SSE2-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3],xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
773; SSE2-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3],xmm3[4],xmm9[4],xmm3[5],xmm9[5],xmm3[6],xmm9[6],xmm3[7],xmm9[7]
774; SSE2-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
775; SSE2-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm10[0],xmm5[1],xmm10[1],xmm5[2],xmm10[2],xmm5[3],xmm10[3],xmm5[4],xmm10[4],xmm5[5],xmm10[5],xmm5[6],xmm10[6],xmm5[7],xmm10[7]
776; SSE2-NEXT:    punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3],xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
777; SSE2-NEXT:    punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
778; SSE2-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3],xmm0[4],xmm7[4],xmm0[5],xmm7[5],xmm0[6],xmm7[6],xmm0[7],xmm7[7]
779; SSE2-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm12[0],xmm2[1],xmm12[1],xmm2[2],xmm12[2],xmm2[3],xmm12[3],xmm2[4],xmm12[4],xmm2[5],xmm12[5],xmm2[6],xmm12[6],xmm2[7],xmm12[7]
780; SSE2-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm13[0],xmm6[1],xmm13[1],xmm6[2],xmm13[2],xmm6[3],xmm13[3],xmm6[4],xmm13[4],xmm6[5],xmm13[5],xmm6[6],xmm13[6],xmm6[7],xmm13[7]
781; SSE2-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3],xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
782; SSE2-NEXT:    punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3],xmm4[4],xmm14[4],xmm4[5],xmm14[5],xmm4[6],xmm14[6],xmm4[7],xmm14[7]
783; SSE2-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3],xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
784; SSE2-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
785; SSE2-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
786; SSE2-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
787; SSE2-NEXT:    retq
788;
789; SSSE3-LABEL: mem_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8:
790; SSSE3:       # BB#0:
791; SSSE3-NEXT:    movsbq (%rdi), %rcx
792; SSSE3-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
793; SSSE3-NEXT:    leaq -{{[0-9]+}}(%rsp), %rax
794; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
795; SSSE3-NEXT:    movd %ecx, %xmm0
796; SSSE3-NEXT:    movsbq 8(%rdi), %rcx
797; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
798; SSSE3-NEXT:    movd %ecx, %xmm8
799; SSSE3-NEXT:    movsbq 12(%rdi), %rcx
800; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
801; SSSE3-NEXT:    movd %ecx, %xmm9
802; SSSE3-NEXT:    movsbq 4(%rdi), %rcx
803; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
804; SSSE3-NEXT:    movd %ecx, %xmm3
805; SSSE3-NEXT:    movsbq 14(%rdi), %rcx
806; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
807; SSSE3-NEXT:    movd %ecx, %xmm10
808; SSSE3-NEXT:    movsbq 6(%rdi), %rcx
809; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
810; SSSE3-NEXT:    movd %ecx, %xmm5
811; SSSE3-NEXT:    movsbq 10(%rdi), %rcx
812; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
813; SSSE3-NEXT:    movd %ecx, %xmm11
814; SSSE3-NEXT:    movsbq 2(%rdi), %rcx
815; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
816; SSSE3-NEXT:    movd %ecx, %xmm7
817; SSSE3-NEXT:    movsbq 15(%rdi), %rcx
818; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
819; SSSE3-NEXT:    movd %ecx, %xmm12
820; SSSE3-NEXT:    movsbq 7(%rdi), %rcx
821; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
822; SSSE3-NEXT:    movd %ecx, %xmm2
823; SSSE3-NEXT:    movsbq 11(%rdi), %rcx
824; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
825; SSSE3-NEXT:    movd %ecx, %xmm13
826; SSSE3-NEXT:    movsbq 3(%rdi), %rcx
827; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
828; SSSE3-NEXT:    movd %ecx, %xmm6
829; SSSE3-NEXT:    movsbq 13(%rdi), %rcx
830; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
831; SSSE3-NEXT:    movd %ecx, %xmm14
832; SSSE3-NEXT:    movsbq 5(%rdi), %rcx
833; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
834; SSSE3-NEXT:    movd %ecx, %xmm4
835; SSSE3-NEXT:    movsbq 9(%rdi), %rcx
836; SSSE3-NEXT:    movzbl (%rcx,%rax), %ecx
837; SSSE3-NEXT:    movd %ecx, %xmm15
838; SSSE3-NEXT:    movsbq 1(%rdi), %rcx
839; SSSE3-NEXT:    movzbl (%rcx,%rax), %eax
840; SSSE3-NEXT:    movd %eax, %xmm1
841; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3],xmm0[4],xmm8[4],xmm0[5],xmm8[5],xmm0[6],xmm8[6],xmm0[7],xmm8[7]
842; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm9[0],xmm3[1],xmm9[1],xmm3[2],xmm9[2],xmm3[3],xmm9[3],xmm3[4],xmm9[4],xmm3[5],xmm9[5],xmm3[6],xmm9[6],xmm3[7],xmm9[7]
843; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3],xmm0[4],xmm3[4],xmm0[5],xmm3[5],xmm0[6],xmm3[6],xmm0[7],xmm3[7]
844; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm10[0],xmm5[1],xmm10[1],xmm5[2],xmm10[2],xmm5[3],xmm10[3],xmm5[4],xmm10[4],xmm5[5],xmm10[5],xmm5[6],xmm10[6],xmm5[7],xmm10[7]
845; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm11[0],xmm7[1],xmm11[1],xmm7[2],xmm11[2],xmm7[3],xmm11[3],xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
846; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm7 = xmm7[0],xmm5[0],xmm7[1],xmm5[1],xmm7[2],xmm5[2],xmm7[3],xmm5[3],xmm7[4],xmm5[4],xmm7[5],xmm5[5],xmm7[6],xmm5[6],xmm7[7],xmm5[7]
847; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1],xmm0[2],xmm7[2],xmm0[3],xmm7[3],xmm0[4],xmm7[4],xmm0[5],xmm7[5],xmm0[6],xmm7[6],xmm0[7],xmm7[7]
848; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm12[0],xmm2[1],xmm12[1],xmm2[2],xmm12[2],xmm2[3],xmm12[3],xmm2[4],xmm12[4],xmm2[5],xmm12[5],xmm2[6],xmm12[6],xmm2[7],xmm12[7]
849; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm13[0],xmm6[1],xmm13[1],xmm6[2],xmm13[2],xmm6[3],xmm13[3],xmm6[4],xmm13[4],xmm6[5],xmm13[5],xmm6[6],xmm13[6],xmm6[7],xmm13[7]
850; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm2[0],xmm6[1],xmm2[1],xmm6[2],xmm2[2],xmm6[3],xmm2[3],xmm6[4],xmm2[4],xmm6[5],xmm2[5],xmm6[6],xmm2[6],xmm6[7],xmm2[7]
851; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm14[0],xmm4[1],xmm14[1],xmm4[2],xmm14[2],xmm4[3],xmm14[3],xmm4[4],xmm14[4],xmm4[5],xmm14[5],xmm4[6],xmm14[6],xmm4[7],xmm14[7]
852; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3],xmm1[4],xmm15[4],xmm1[5],xmm15[5],xmm1[6],xmm15[6],xmm1[7],xmm15[7]
853; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7]
854; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm6[0],xmm1[1],xmm6[1],xmm1[2],xmm6[2],xmm1[3],xmm6[3],xmm1[4],xmm6[4],xmm1[5],xmm6[5],xmm1[6],xmm6[6],xmm1[7],xmm6[7]
855; SSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
856; SSSE3-NEXT:    retq
857;
858; SSE41-LABEL: mem_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8:
859; SSE41:       # BB#0:
860; SSE41-NEXT:    pushq %rbp
861; SSE41-NEXT:    pushq %r15
862; SSE41-NEXT:    pushq %r14
863; SSE41-NEXT:    pushq %r13
864; SSE41-NEXT:    pushq %r12
865; SSE41-NEXT:    pushq %rbx
866; SSE41-NEXT:    movsbq (%rdi), %rax
867; SSE41-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
868; SSE41-NEXT:    movsbq 1(%rdi), %r15
869; SSE41-NEXT:    movsbq 2(%rdi), %r8
870; SSE41-NEXT:    movsbq 3(%rdi), %r9
871; SSE41-NEXT:    movsbq 4(%rdi), %r10
872; SSE41-NEXT:    movsbq 5(%rdi), %r11
873; SSE41-NEXT:    movsbq 6(%rdi), %r14
874; SSE41-NEXT:    movsbq 7(%rdi), %r12
875; SSE41-NEXT:    movsbq 8(%rdi), %r13
876; SSE41-NEXT:    movsbq 9(%rdi), %rdx
877; SSE41-NEXT:    movsbq 10(%rdi), %rcx
878; SSE41-NEXT:    movsbq 11(%rdi), %rsi
879; SSE41-NEXT:    movsbq 12(%rdi), %rbx
880; SSE41-NEXT:    leaq -{{[0-9]+}}(%rsp), %rbp
881; SSE41-NEXT:    movzbl (%rax,%rbp), %eax
882; SSE41-NEXT:    movd %eax, %xmm0
883; SSE41-NEXT:    movsbq 13(%rdi), %rax
884; SSE41-NEXT:    pinsrb $1, (%r15,%rbp), %xmm0
885; SSE41-NEXT:    movsbq 14(%rdi), %r15
886; SSE41-NEXT:    movsbq 15(%rdi), %rdi
887; SSE41-NEXT:    movzbl (%rdi,%rbp), %edi
888; SSE41-NEXT:    movzbl (%r15,%rbp), %r15d
889; SSE41-NEXT:    movzbl (%rax,%rbp), %eax
890; SSE41-NEXT:    movzbl (%rbx,%rbp), %ebx
891; SSE41-NEXT:    movzbl (%rsi,%rbp), %esi
892; SSE41-NEXT:    movzbl (%rcx,%rbp), %ecx
893; SSE41-NEXT:    movzbl (%rdx,%rbp), %edx
894; SSE41-NEXT:    movzbl (%r13,%rbp), %r13d
895; SSE41-NEXT:    movzbl (%r12,%rbp), %r12d
896; SSE41-NEXT:    movzbl (%r14,%rbp), %r14d
897; SSE41-NEXT:    movzbl (%r11,%rbp), %r11d
898; SSE41-NEXT:    movzbl (%r10,%rbp), %r10d
899; SSE41-NEXT:    movzbl (%r9,%rbp), %r9d
900; SSE41-NEXT:    movzbl (%r8,%rbp), %ebp
901; SSE41-NEXT:    pinsrb $2, %ebp, %xmm0
902; SSE41-NEXT:    pinsrb $3, %r9d, %xmm0
903; SSE41-NEXT:    pinsrb $4, %r10d, %xmm0
904; SSE41-NEXT:    pinsrb $5, %r11d, %xmm0
905; SSE41-NEXT:    pinsrb $6, %r14d, %xmm0
906; SSE41-NEXT:    pinsrb $7, %r12d, %xmm0
907; SSE41-NEXT:    pinsrb $8, %r13d, %xmm0
908; SSE41-NEXT:    pinsrb $9, %edx, %xmm0
909; SSE41-NEXT:    pinsrb $10, %ecx, %xmm0
910; SSE41-NEXT:    pinsrb $11, %esi, %xmm0
911; SSE41-NEXT:    pinsrb $12, %ebx, %xmm0
912; SSE41-NEXT:    pinsrb $13, %eax, %xmm0
913; SSE41-NEXT:    pinsrb $14, %r15d, %xmm0
914; SSE41-NEXT:    pinsrb $15, %edi, %xmm0
915; SSE41-NEXT:    popq %rbx
916; SSE41-NEXT:    popq %r12
917; SSE41-NEXT:    popq %r13
918; SSE41-NEXT:    popq %r14
919; SSE41-NEXT:    popq %r15
920; SSE41-NEXT:    popq %rbp
921; SSE41-NEXT:    retq
922;
923; AVX-LABEL: mem_shuffle_v16i8_v16i8_xxxxxxxxxxxxxxxx_i8:
924; AVX:       # BB#0:
925; AVX-NEXT:    pushq %rbp
926; AVX-NEXT:    pushq %r15
927; AVX-NEXT:    pushq %r14
928; AVX-NEXT:    pushq %r13
929; AVX-NEXT:    pushq %r12
930; AVX-NEXT:    pushq %rbx
931; AVX-NEXT:    movsbq (%rdi), %rsi
932; AVX-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
933; AVX-NEXT:    movsbq 1(%rdi), %r15
934; AVX-NEXT:    movsbq 2(%rdi), %r8
935; AVX-NEXT:    movsbq 3(%rdi), %r9
936; AVX-NEXT:    movsbq 4(%rdi), %r10
937; AVX-NEXT:    movsbq 5(%rdi), %r11
938; AVX-NEXT:    movsbq 6(%rdi), %r14
939; AVX-NEXT:    movsbq 7(%rdi), %r12
940; AVX-NEXT:    movsbq 8(%rdi), %r13
941; AVX-NEXT:    movsbq 9(%rdi), %rdx
942; AVX-NEXT:    movsbq 10(%rdi), %rax
943; AVX-NEXT:    movsbq 11(%rdi), %rcx
944; AVX-NEXT:    movsbq 12(%rdi), %rbx
945; AVX-NEXT:    leaq -{{[0-9]+}}(%rsp), %rbp
946; AVX-NEXT:    movzbl (%rsi,%rbp), %esi
947; AVX-NEXT:    vmovd %esi, %xmm0
948; AVX-NEXT:    movsbq 13(%rdi), %rsi
949; AVX-NEXT:    vpinsrb $1, (%r15,%rbp), %xmm0, %xmm0
950; AVX-NEXT:    movsbq 14(%rdi), %r15
951; AVX-NEXT:    movsbq 15(%rdi), %rdi
952; AVX-NEXT:    movzbl (%rdi,%rbp), %edi
953; AVX-NEXT:    movzbl (%r15,%rbp), %r15d
954; AVX-NEXT:    movzbl (%rsi,%rbp), %esi
955; AVX-NEXT:    movzbl (%rbx,%rbp), %ebx
956; AVX-NEXT:    movzbl (%rcx,%rbp), %ecx
957; AVX-NEXT:    movzbl (%rax,%rbp), %eax
958; AVX-NEXT:    movzbl (%rdx,%rbp), %edx
959; AVX-NEXT:    movzbl (%r13,%rbp), %r13d
960; AVX-NEXT:    movzbl (%r12,%rbp), %r12d
961; AVX-NEXT:    movzbl (%r14,%rbp), %r14d
962; AVX-NEXT:    movzbl (%r11,%rbp), %r11d
963; AVX-NEXT:    movzbl (%r10,%rbp), %r10d
964; AVX-NEXT:    movzbl (%r9,%rbp), %r9d
965; AVX-NEXT:    movzbl (%r8,%rbp), %ebp
966; AVX-NEXT:    vpinsrb $2, %ebp, %xmm0, %xmm0
967; AVX-NEXT:    vpinsrb $3, %r9d, %xmm0, %xmm0
968; AVX-NEXT:    vpinsrb $4, %r10d, %xmm0, %xmm0
969; AVX-NEXT:    vpinsrb $5, %r11d, %xmm0, %xmm0
970; AVX-NEXT:    vpinsrb $6, %r14d, %xmm0, %xmm0
971; AVX-NEXT:    vpinsrb $7, %r12d, %xmm0, %xmm0
972; AVX-NEXT:    vpinsrb $8, %r13d, %xmm0, %xmm0
973; AVX-NEXT:    vpinsrb $9, %edx, %xmm0, %xmm0
974; AVX-NEXT:    vpinsrb $10, %eax, %xmm0, %xmm0
975; AVX-NEXT:    vpinsrb $11, %ecx, %xmm0, %xmm0
976; AVX-NEXT:    vpinsrb $12, %ebx, %xmm0, %xmm0
977; AVX-NEXT:    vpinsrb $13, %esi, %xmm0, %xmm0
978; AVX-NEXT:    vpinsrb $14, %r15d, %xmm0, %xmm0
979; AVX-NEXT:    vpinsrb $15, %edi, %xmm0, %xmm0
980; AVX-NEXT:    popq %rbx
981; AVX-NEXT:    popq %r12
982; AVX-NEXT:    popq %r13
983; AVX-NEXT:    popq %r14
984; AVX-NEXT:    popq %r15
985; AVX-NEXT:    popq %rbp
986; AVX-NEXT:    retq
987  %p0  = getelementptr inbounds i8, i8* %i, i64 0
988  %p1  = getelementptr inbounds i8, i8* %i, i64 1
989  %p2  = getelementptr inbounds i8, i8* %i, i64 2
990  %p3  = getelementptr inbounds i8, i8* %i, i64 3
991  %p4  = getelementptr inbounds i8, i8* %i, i64 4
992  %p5  = getelementptr inbounds i8, i8* %i, i64 5
993  %p6  = getelementptr inbounds i8, i8* %i, i64 6
994  %p7  = getelementptr inbounds i8, i8* %i, i64 7
995  %p8  = getelementptr inbounds i8, i8* %i, i64 8
996  %p9  = getelementptr inbounds i8, i8* %i, i64 9
997  %p10 = getelementptr inbounds i8, i8* %i, i64 10
998  %p11 = getelementptr inbounds i8, i8* %i, i64 11
999  %p12 = getelementptr inbounds i8, i8* %i, i64 12
1000  %p13 = getelementptr inbounds i8, i8* %i, i64 13
1001  %p14 = getelementptr inbounds i8, i8* %i, i64 14
1002  %p15 = getelementptr inbounds i8, i8* %i, i64 15
1003  %i0  = load i8, i8* %p0 , align 4
1004  %i1  = load i8, i8* %p1 , align 4
1005  %i2  = load i8, i8* %p2 , align 4
1006  %i3  = load i8, i8* %p3 , align 4
1007  %i4  = load i8, i8* %p4 , align 4
1008  %i5  = load i8, i8* %p5 , align 4
1009  %i6  = load i8, i8* %p6 , align 4
1010  %i7  = load i8, i8* %p7 , align 4
1011  %i8  = load i8, i8* %p8 , align 4
1012  %i9  = load i8, i8* %p9 , align 4
1013  %i10 = load i8, i8* %p10, align 4
1014  %i11 = load i8, i8* %p11, align 4
1015  %i12 = load i8, i8* %p12, align 4
1016  %i13 = load i8, i8* %p13, align 4
1017  %i14 = load i8, i8* %p14, align 4
1018  %i15 = load i8, i8* %p15, align 4
1019  %x0  = extractelement <16 x i8> %x, i8 %i0
1020  %x1  = extractelement <16 x i8> %x, i8 %i1
1021  %x2  = extractelement <16 x i8> %x, i8 %i2
1022  %x3  = extractelement <16 x i8> %x, i8 %i3
1023  %x4  = extractelement <16 x i8> %x, i8 %i4
1024  %x5  = extractelement <16 x i8> %x, i8 %i5
1025  %x6  = extractelement <16 x i8> %x, i8 %i6
1026  %x7  = extractelement <16 x i8> %x, i8 %i7
1027  %x8  = extractelement <16 x i8> %x, i8 %i8
1028  %x9  = extractelement <16 x i8> %x, i8 %i9
1029  %x10 = extractelement <16 x i8> %x, i8 %i10
1030  %x11 = extractelement <16 x i8> %x, i8 %i11
1031  %x12 = extractelement <16 x i8> %x, i8 %i12
1032  %x13 = extractelement <16 x i8> %x, i8 %i13
1033  %x14 = extractelement <16 x i8> %x, i8 %i14
1034  %x15 = extractelement <16 x i8> %x, i8 %i15
1035  %r0  = insertelement <16 x i8> undef, i8 %x0 , i32 0
1036  %r1  = insertelement <16 x i8>  %r0 , i8 %x1 , i32 1
1037  %r2  = insertelement <16 x i8>  %r1 , i8 %x2 , i32 2
1038  %r3  = insertelement <16 x i8>  %r2 , i8 %x3 , i32 3
1039  %r4  = insertelement <16 x i8>  %r3 , i8 %x4 , i32 4
1040  %r5  = insertelement <16 x i8>  %r4 , i8 %x5 , i32 5
1041  %r6  = insertelement <16 x i8>  %r5 , i8 %x6 , i32 6
1042  %r7  = insertelement <16 x i8>  %r6 , i8 %x7 , i32 7
1043  %r8  = insertelement <16 x i8>  %r7 , i8 %x8 , i32 8
1044  %r9  = insertelement <16 x i8>  %r8 , i8 %x9 , i32 9
1045  %r10 = insertelement <16 x i8>  %r9 , i8 %x10, i32 10
1046  %r11 = insertelement <16 x i8>  %r10, i8 %x11, i32 11
1047  %r12 = insertelement <16 x i8>  %r11, i8 %x12, i32 12
1048  %r13 = insertelement <16 x i8>  %r12, i8 %x13, i32 13
1049  %r14 = insertelement <16 x i8>  %r13, i8 %x14, i32 14
1050  %r15 = insertelement <16 x i8>  %r14, i8 %x15, i32 15
1051  ret <16 x i8> %r15
1052}
1053
1054;
1055; Binary shuffle indices from registers
1056;
1057
1058define <4 x float> @var_shuffle_v4f32_v4f32_x0yx_i32(<4 x float> %x, <4 x float> %y, i32 %i0, i32 %i1, i32 %i2, i32 %i3) nounwind {
1059; SSE-LABEL: var_shuffle_v4f32_v4f32_x0yx_i32:
1060; SSE:       # BB#0:
1061; SSE-NEXT:    movslq %edi, %rax
1062; SSE-NEXT:    movaps %xmm1, -{{[0-9]+}}(%rsp)
1063; SSE-NEXT:    movslq %edx, %rdx
1064; SSE-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
1065; SSE-NEXT:    movslq %ecx, %rcx
1066; SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
1067; SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
1068; SSE-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
1069; SSE-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
1070; SSE-NEXT:    unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1071; SSE-NEXT:    retq
1072;
1073; AVX-LABEL: var_shuffle_v4f32_v4f32_x0yx_i32:
1074; AVX:       # BB#0:
1075; AVX-NEXT:    movslq %edi, %rax
1076; AVX-NEXT:    vmovaps %xmm1, -{{[0-9]+}}(%rsp)
1077; AVX-NEXT:    movslq %edx, %rdx
1078; AVX-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
1079; AVX-NEXT:    movslq %ecx, %rcx
1080; AVX-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
1081; AVX-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
1082; AVX-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
1083; AVX-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
1084; AVX-NEXT:    vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm1[0]
1085; AVX-NEXT:    retq
1086  %x0 = extractelement <4 x float> %x, i32 %i0
1087  %x1 = extractelement <4 x float> %x, i32 %i1
1088  %y2 = extractelement <4 x float> %y, i32 %i2
1089  %x3 = extractelement <4 x float> %x, i32 %i3
1090  %r0 = insertelement <4 x float> undef, float %x0, i32 0
1091  %r1 = insertelement <4 x float>   %r0, float 0.0, i32 1
1092  %r2 = insertelement <4 x float>   %r1, float %y2, i32 2
1093  %r3 = insertelement <4 x float>   %r2, float %x3, i32 3
1094  ret <4 x float> %r3
1095}
1096
1097define <8 x i16> @var_shuffle_v8i16_v8i16_xyxyxy00_i16(<8 x i16> %x, <8 x i16> %y, i16 %i0, i16 %i1, i16 %i2, i16 %i3, i16 %i4, i16 %i5, i16 %i6, i16 %i7) nounwind {
1098; SSE2-LABEL: var_shuffle_v8i16_v8i16_xyxyxy00_i16:
1099; SSE2:       # BB#0:
1100; SSE2-NEXT:    movswq %di, %r10
1101; SSE2-NEXT:    movswq %si, %rsi
1102; SSE2-NEXT:    movswq %dx, %r11
1103; SSE2-NEXT:    movswq %cx, %rcx
1104; SSE2-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
1105; SSE2-NEXT:    movswq %r8w, %rdi
1106; SSE2-NEXT:    movaps %xmm1, -{{[0-9]+}}(%rsp)
1107; SSE2-NEXT:    movswq %r9w, %rax
1108; SSE2-NEXT:    movzwl -24(%rsp,%rsi,2), %esi
1109; SSE2-NEXT:    xorl %edx, %edx
1110; SSE2-NEXT:    movd %edx, %xmm0
1111; SSE2-NEXT:    movzwl -24(%rsp,%rcx,2), %ecx
1112; SSE2-NEXT:    movd %ecx, %xmm1
1113; SSE2-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
1114; SSE2-NEXT:    movd %esi, %xmm2
1115; SSE2-NEXT:    movzwl -24(%rsp,%rax,2), %eax
1116; SSE2-NEXT:    movd %eax, %xmm3
1117; SSE2-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
1118; SSE2-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
1119; SSE2-NEXT:    movzwl -40(%rsp,%r10,2), %eax
1120; SSE2-NEXT:    movzwl -40(%rsp,%r11,2), %ecx
1121; SSE2-NEXT:    movd %ecx, %xmm1
1122; SSE2-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
1123; SSE2-NEXT:    movd %eax, %xmm0
1124; SSE2-NEXT:    movzwl -40(%rsp,%rdi,2), %eax
1125; SSE2-NEXT:    movd %eax, %xmm3
1126; SSE2-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
1127; SSE2-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1128; SSE2-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
1129; SSE2-NEXT:    retq
1130;
1131; SSSE3-LABEL: var_shuffle_v8i16_v8i16_xyxyxy00_i16:
1132; SSSE3:       # BB#0:
1133; SSSE3-NEXT:    movswq %di, %r10
1134; SSSE3-NEXT:    movswq %si, %rsi
1135; SSSE3-NEXT:    movswq %dx, %r11
1136; SSSE3-NEXT:    movswq %cx, %rcx
1137; SSSE3-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
1138; SSSE3-NEXT:    movswq %r8w, %rdi
1139; SSSE3-NEXT:    movaps %xmm1, -{{[0-9]+}}(%rsp)
1140; SSSE3-NEXT:    movswq %r9w, %rax
1141; SSSE3-NEXT:    movzwl -24(%rsp,%rsi,2), %esi
1142; SSSE3-NEXT:    xorl %edx, %edx
1143; SSSE3-NEXT:    movd %edx, %xmm0
1144; SSSE3-NEXT:    movzwl -24(%rsp,%rcx,2), %ecx
1145; SSSE3-NEXT:    movd %ecx, %xmm1
1146; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
1147; SSSE3-NEXT:    movd %esi, %xmm2
1148; SSSE3-NEXT:    movzwl -24(%rsp,%rax,2), %eax
1149; SSSE3-NEXT:    movd %eax, %xmm3
1150; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
1151; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
1152; SSSE3-NEXT:    movzwl -40(%rsp,%r10,2), %eax
1153; SSSE3-NEXT:    movzwl -40(%rsp,%r11,2), %ecx
1154; SSSE3-NEXT:    movd %ecx, %xmm1
1155; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
1156; SSSE3-NEXT:    movd %eax, %xmm0
1157; SSSE3-NEXT:    movzwl -40(%rsp,%rdi,2), %eax
1158; SSSE3-NEXT:    movd %eax, %xmm3
1159; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
1160; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
1161; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
1162; SSSE3-NEXT:    retq
1163;
1164; SSE41-LABEL: var_shuffle_v8i16_v8i16_xyxyxy00_i16:
1165; SSE41:       # BB#0:
1166; SSE41-NEXT:    movswq %di, %rax
1167; SSE41-NEXT:    movswq %si, %rsi
1168; SSE41-NEXT:    movswq %dx, %rdx
1169; SSE41-NEXT:    movswq %cx, %r10
1170; SSE41-NEXT:    movdqa %xmm0, -{{[0-9]+}}(%rsp)
1171; SSE41-NEXT:    movswq %r8w, %rdi
1172; SSE41-NEXT:    movaps %xmm1, -{{[0-9]+}}(%rsp)
1173; SSE41-NEXT:    movswq %r9w, %rcx
1174; SSE41-NEXT:    movzwl -40(%rsp,%rax,2), %eax
1175; SSE41-NEXT:    movd %eax, %xmm1
1176; SSE41-NEXT:    pinsrw $1, -24(%rsp,%rsi,2), %xmm1
1177; SSE41-NEXT:    pinsrw $2, -40(%rsp,%rdx,2), %xmm1
1178; SSE41-NEXT:    pinsrw $3, -24(%rsp,%r10,2), %xmm1
1179; SSE41-NEXT:    pinsrw $4, -40(%rsp,%rdi,2), %xmm1
1180; SSE41-NEXT:    pinsrw $5, -24(%rsp,%rcx,2), %xmm1
1181; SSE41-NEXT:    pxor %xmm0, %xmm0
1182; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm1[0,1,2,3,4,5],xmm0[6,7]
1183; SSE41-NEXT:    retq
1184;
1185; AVX1-LABEL: var_shuffle_v8i16_v8i16_xyxyxy00_i16:
1186; AVX1:       # BB#0:
1187; AVX1-NEXT:    movswq %di, %r10
1188; AVX1-NEXT:    movswq %si, %r11
1189; AVX1-NEXT:    movswq %dx, %rdx
1190; AVX1-NEXT:    movswq %cx, %rcx
1191; AVX1-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
1192; AVX1-NEXT:    movswq %r8w, %rdi
1193; AVX1-NEXT:    vmovdqa %xmm1, -{{[0-9]+}}(%rsp)
1194; AVX1-NEXT:    movswq %r9w, %rax
1195; AVX1-NEXT:    movzwl -40(%rsp,%r10,2), %esi
1196; AVX1-NEXT:    vmovd %esi, %xmm0
1197; AVX1-NEXT:    vpinsrw $1, -24(%rsp,%r11,2), %xmm0, %xmm0
1198; AVX1-NEXT:    vpinsrw $2, -40(%rsp,%rdx,2), %xmm0, %xmm0
1199; AVX1-NEXT:    vpinsrw $3, -24(%rsp,%rcx,2), %xmm0, %xmm0
1200; AVX1-NEXT:    vpinsrw $4, -40(%rsp,%rdi,2), %xmm0, %xmm0
1201; AVX1-NEXT:    vpinsrw $5, -24(%rsp,%rax,2), %xmm0, %xmm0
1202; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
1203; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5],xmm1[6,7]
1204; AVX1-NEXT:    retq
1205;
1206; AVX2-LABEL: var_shuffle_v8i16_v8i16_xyxyxy00_i16:
1207; AVX2:       # BB#0:
1208; AVX2-NEXT:    movswq %di, %r10
1209; AVX2-NEXT:    movswq %si, %r11
1210; AVX2-NEXT:    movswq %dx, %rdx
1211; AVX2-NEXT:    movswq %cx, %rcx
1212; AVX2-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
1213; AVX2-NEXT:    movswq %r8w, %rdi
1214; AVX2-NEXT:    vmovdqa %xmm1, -{{[0-9]+}}(%rsp)
1215; AVX2-NEXT:    movswq %r9w, %rax
1216; AVX2-NEXT:    movzwl -40(%rsp,%r10,2), %esi
1217; AVX2-NEXT:    vmovd %esi, %xmm0
1218; AVX2-NEXT:    vpinsrw $1, -24(%rsp,%r11,2), %xmm0, %xmm0
1219; AVX2-NEXT:    vpinsrw $2, -40(%rsp,%rdx,2), %xmm0, %xmm0
1220; AVX2-NEXT:    vpinsrw $3, -24(%rsp,%rcx,2), %xmm0, %xmm0
1221; AVX2-NEXT:    vpinsrw $4, -40(%rsp,%rdi,2), %xmm0, %xmm0
1222; AVX2-NEXT:    vpinsrw $5, -24(%rsp,%rax,2), %xmm0, %xmm0
1223; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
1224; AVX2-NEXT:    vpblendd {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[3]
1225; AVX2-NEXT:    retq
1226  %x0 = extractelement <8 x i16> %x, i16 %i0
1227  %y1 = extractelement <8 x i16> %y, i16 %i1
1228  %x2 = extractelement <8 x i16> %x, i16 %i2
1229  %y3 = extractelement <8 x i16> %y, i16 %i3
1230  %x4 = extractelement <8 x i16> %x, i16 %i4
1231  %y5 = extractelement <8 x i16> %y, i16 %i5
1232  %x6 = extractelement <8 x i16> %x, i16 %i6
1233  %x7 = extractelement <8 x i16> %x, i16 %i7
1234  %r0 = insertelement <8 x i16> undef, i16 %x0, i32 0
1235  %r1 = insertelement <8 x i16>   %r0, i16 %y1, i32 1
1236  %r2 = insertelement <8 x i16>   %r1, i16 %x2, i32 2
1237  %r3 = insertelement <8 x i16>   %r2, i16 %y3, i32 3
1238  %r4 = insertelement <8 x i16>   %r3, i16 %x4, i32 4
1239  %r5 = insertelement <8 x i16>   %r4, i16 %y5, i32 5
1240  %r6 = insertelement <8 x i16>   %r5, i16   0, i32 6
1241  %r7 = insertelement <8 x i16>   %r6, i16   0, i32 7
1242  ret <8 x i16> %r7
1243}
1244