xref: /llvm-project/llvm/test/CodeGen/X86/horizontal-reduce-smax.ll (revision 7b3bbd83c0c24087072ec5b22a76799ab31f87d5) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+sse2     | FileCheck %s --check-prefix=X86-SSE2
3; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+sse4.2   | FileCheck %s --check-prefix=X86-SSE42
4; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx      | FileCheck %s --check-prefixes=X86-AVX,X86-AVX1
5; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx2     | FileCheck %s --check-prefixes=X86-AVX,X86-AVX2
6; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse2   | FileCheck %s --check-prefix=X64-SSE2
7; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse4.2 | FileCheck %s --check-prefix=X64-SSE42
8; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx    | FileCheck %s --check-prefixes=X64-AVX,X64-AVX1OR2,X64-AVX1
9; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx2   | FileCheck %s --check-prefixes=X64-AVX,X64-AVX1OR2,X64-AVX2
10; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512bw,+avx512dq,+avx512vl | FileCheck %s --check-prefixes=X64-AVX,X64-AVX512
11
12;
13; 128-bit Vectors
14;
15
16define i64 @test_reduce_v2i64(<2 x i64> %a0) {
17; X86-SSE2-LABEL: test_reduce_v2i64:
18; X86-SSE2:       ## %bb.0:
19; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
20; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,0,2147483648,0]
21; X86-SSE2-NEXT:    movdqa %xmm0, %xmm3
22; X86-SSE2-NEXT:    pxor %xmm2, %xmm3
23; X86-SSE2-NEXT:    pxor %xmm1, %xmm2
24; X86-SSE2-NEXT:    movdqa %xmm3, %xmm4
25; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm4
26; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
27; X86-SSE2-NEXT:    pcmpeqd %xmm3, %xmm2
28; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
29; X86-SSE2-NEXT:    pand %xmm5, %xmm2
30; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
31; X86-SSE2-NEXT:    por %xmm2, %xmm3
32; X86-SSE2-NEXT:    pand %xmm3, %xmm0
33; X86-SSE2-NEXT:    pandn %xmm1, %xmm3
34; X86-SSE2-NEXT:    por %xmm0, %xmm3
35; X86-SSE2-NEXT:    movd %xmm3, %eax
36; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,1,1]
37; X86-SSE2-NEXT:    movd %xmm0, %edx
38; X86-SSE2-NEXT:    retl
39;
40; X86-SSE42-LABEL: test_reduce_v2i64:
41; X86-SSE42:       ## %bb.0:
42; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
43; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
44; X86-SSE42-NEXT:    pcmpgtq %xmm2, %xmm0
45; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
46; X86-SSE42-NEXT:    movd %xmm2, %eax
47; X86-SSE42-NEXT:    pextrd $1, %xmm2, %edx
48; X86-SSE42-NEXT:    retl
49;
50; X86-AVX-LABEL: test_reduce_v2i64:
51; X86-AVX:       ## %bb.0:
52; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
53; X86-AVX-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
54; X86-AVX-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
55; X86-AVX-NEXT:    vmovd %xmm0, %eax
56; X86-AVX-NEXT:    vpextrd $1, %xmm0, %edx
57; X86-AVX-NEXT:    retl
58;
59; X64-SSE2-LABEL: test_reduce_v2i64:
60; X64-SSE2:       ## %bb.0:
61; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
62; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
63; X64-SSE2-NEXT:    movdqa %xmm0, %xmm3
64; X64-SSE2-NEXT:    pxor %xmm2, %xmm3
65; X64-SSE2-NEXT:    pxor %xmm1, %xmm2
66; X64-SSE2-NEXT:    movdqa %xmm3, %xmm4
67; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm4
68; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
69; X64-SSE2-NEXT:    pcmpeqd %xmm3, %xmm2
70; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
71; X64-SSE2-NEXT:    pand %xmm5, %xmm2
72; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
73; X64-SSE2-NEXT:    por %xmm2, %xmm3
74; X64-SSE2-NEXT:    pand %xmm3, %xmm0
75; X64-SSE2-NEXT:    pandn %xmm1, %xmm3
76; X64-SSE2-NEXT:    por %xmm0, %xmm3
77; X64-SSE2-NEXT:    movq %xmm3, %rax
78; X64-SSE2-NEXT:    retq
79;
80; X64-SSE42-LABEL: test_reduce_v2i64:
81; X64-SSE42:       ## %bb.0:
82; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
83; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
84; X64-SSE42-NEXT:    pcmpgtq %xmm2, %xmm0
85; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
86; X64-SSE42-NEXT:    movq %xmm2, %rax
87; X64-SSE42-NEXT:    retq
88;
89; X64-AVX1OR2-LABEL: test_reduce_v2i64:
90; X64-AVX1OR2:       ## %bb.0:
91; X64-AVX1OR2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
92; X64-AVX1OR2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
93; X64-AVX1OR2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
94; X64-AVX1OR2-NEXT:    vmovq %xmm0, %rax
95; X64-AVX1OR2-NEXT:    retq
96;
97; X64-AVX512-LABEL: test_reduce_v2i64:
98; X64-AVX512:       ## %bb.0:
99; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
100; X64-AVX512-NEXT:    vpmaxsq %xmm1, %xmm0, %xmm0
101; X64-AVX512-NEXT:    vmovq %xmm0, %rax
102; X64-AVX512-NEXT:    retq
103  %1 = shufflevector <2 x i64> %a0, <2 x i64> undef, <2 x i32> <i32 1, i32 undef>
104  %2 = icmp sgt <2 x i64> %a0, %1
105  %3 = select <2 x i1> %2, <2 x i64> %a0, <2 x i64> %1
106  %4 = extractelement <2 x i64> %3, i32 0
107  ret i64 %4
108}
109
110define i32 @test_reduce_v4i32(<4 x i32> %a0) {
111; X86-SSE2-LABEL: test_reduce_v4i32:
112; X86-SSE2:       ## %bb.0:
113; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
114; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
115; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
116; X86-SSE2-NEXT:    pand %xmm2, %xmm0
117; X86-SSE2-NEXT:    pandn %xmm1, %xmm2
118; X86-SSE2-NEXT:    por %xmm0, %xmm2
119; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
120; X86-SSE2-NEXT:    movdqa %xmm2, %xmm1
121; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
122; X86-SSE2-NEXT:    pand %xmm1, %xmm2
123; X86-SSE2-NEXT:    pandn %xmm0, %xmm1
124; X86-SSE2-NEXT:    por %xmm2, %xmm1
125; X86-SSE2-NEXT:    movd %xmm1, %eax
126; X86-SSE2-NEXT:    retl
127;
128; X86-SSE42-LABEL: test_reduce_v4i32:
129; X86-SSE42:       ## %bb.0:
130; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
131; X86-SSE42-NEXT:    pmaxsd %xmm0, %xmm1
132; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
133; X86-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
134; X86-SSE42-NEXT:    movd %xmm0, %eax
135; X86-SSE42-NEXT:    retl
136;
137; X86-AVX-LABEL: test_reduce_v4i32:
138; X86-AVX:       ## %bb.0:
139; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
140; X86-AVX-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
141; X86-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
142; X86-AVX-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
143; X86-AVX-NEXT:    vmovd %xmm0, %eax
144; X86-AVX-NEXT:    retl
145;
146; X64-SSE2-LABEL: test_reduce_v4i32:
147; X64-SSE2:       ## %bb.0:
148; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
149; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
150; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
151; X64-SSE2-NEXT:    pand %xmm2, %xmm0
152; X64-SSE2-NEXT:    pandn %xmm1, %xmm2
153; X64-SSE2-NEXT:    por %xmm0, %xmm2
154; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
155; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
156; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
157; X64-SSE2-NEXT:    pand %xmm1, %xmm2
158; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
159; X64-SSE2-NEXT:    por %xmm2, %xmm1
160; X64-SSE2-NEXT:    movd %xmm1, %eax
161; X64-SSE2-NEXT:    retq
162;
163; X64-SSE42-LABEL: test_reduce_v4i32:
164; X64-SSE42:       ## %bb.0:
165; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
166; X64-SSE42-NEXT:    pmaxsd %xmm0, %xmm1
167; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
168; X64-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
169; X64-SSE42-NEXT:    movd %xmm0, %eax
170; X64-SSE42-NEXT:    retq
171;
172; X64-AVX-LABEL: test_reduce_v4i32:
173; X64-AVX:       ## %bb.0:
174; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
175; X64-AVX-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
176; X64-AVX-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
177; X64-AVX-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
178; X64-AVX-NEXT:    vmovd %xmm0, %eax
179; X64-AVX-NEXT:    retq
180  %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
181  %2 = icmp sgt <4 x i32> %a0, %1
182  %3 = select <4 x i1> %2, <4 x i32> %a0, <4 x i32> %1
183  %4 = shufflevector <4 x i32> %3, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
184  %5 = icmp sgt <4 x i32> %3, %4
185  %6 = select <4 x i1> %5, <4 x i32> %3, <4 x i32> %4
186  %7 = extractelement <4 x i32> %6, i32 0
187  ret i32 %7
188}
189
190define i16 @test_reduce_v8i16(<8 x i16> %a0) {
191; X86-SSE2-LABEL: test_reduce_v8i16:
192; X86-SSE2:       ## %bb.0:
193; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
194; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
195; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
196; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
197; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
198; X86-SSE2-NEXT:    psrld $16, %xmm1
199; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
200; X86-SSE2-NEXT:    movd %xmm1, %eax
201; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
202; X86-SSE2-NEXT:    retl
203;
204; X86-SSE42-LABEL: test_reduce_v8i16:
205; X86-SSE42:       ## %bb.0:
206; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
207; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
208; X86-SSE42-NEXT:    movd %xmm0, %eax
209; X86-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
210; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
211; X86-SSE42-NEXT:    retl
212;
213; X86-AVX-LABEL: test_reduce_v8i16:
214; X86-AVX:       ## %bb.0:
215; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
216; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
217; X86-AVX-NEXT:    vmovd %xmm0, %eax
218; X86-AVX-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
219; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax
220; X86-AVX-NEXT:    retl
221;
222; X64-SSE2-LABEL: test_reduce_v8i16:
223; X64-SSE2:       ## %bb.0:
224; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
225; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
226; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
227; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
228; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
229; X64-SSE2-NEXT:    psrld $16, %xmm1
230; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
231; X64-SSE2-NEXT:    movd %xmm1, %eax
232; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
233; X64-SSE2-NEXT:    retq
234;
235; X64-SSE42-LABEL: test_reduce_v8i16:
236; X64-SSE42:       ## %bb.0:
237; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
238; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
239; X64-SSE42-NEXT:    movd %xmm0, %eax
240; X64-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
241; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
242; X64-SSE42-NEXT:    retq
243;
244; X64-AVX1OR2-LABEL: test_reduce_v8i16:
245; X64-AVX1OR2:       ## %bb.0:
246; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
247; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm0
248; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax
249; X64-AVX1OR2-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
250; X64-AVX1OR2-NEXT:    ## kill: def $ax killed $ax killed $eax
251; X64-AVX1OR2-NEXT:    retq
252;
253; X64-AVX512-LABEL: test_reduce_v8i16:
254; X64-AVX512:       ## %bb.0:
255; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
256; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
257; X64-AVX512-NEXT:    vmovd %xmm0, %eax
258; X64-AVX512-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
259; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax
260; X64-AVX512-NEXT:    retq
261  %1  = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
262  %2  = icmp sgt <8 x i16> %a0, %1
263  %3  = select <8 x i1> %2, <8 x i16> %a0, <8 x i16> %1
264  %4  = shufflevector <8 x i16> %3, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
265  %5  = icmp sgt <8 x i16> %3, %4
266  %6  = select <8 x i1> %5, <8 x i16> %3, <8 x i16> %4
267  %7  = shufflevector <8 x i16> %6, <8 x i16> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
268  %8  = icmp sgt <8 x i16> %6, %7
269  %9  = select <8 x i1> %8, <8 x i16> %6, <8 x i16> %7
270  %10 = extractelement <8 x i16> %9, i32 0
271  ret i16 %10
272}
273
274define i8 @test_reduce_v16i8(<16 x i8> %a0) {
275; X86-SSE2-LABEL: test_reduce_v16i8:
276; X86-SSE2:       ## %bb.0:
277; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
278; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
279; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
280; X86-SSE2-NEXT:    pand %xmm2, %xmm0
281; X86-SSE2-NEXT:    pandn %xmm1, %xmm2
282; X86-SSE2-NEXT:    por %xmm0, %xmm2
283; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]
284; X86-SSE2-NEXT:    movdqa %xmm2, %xmm0
285; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm0
286; X86-SSE2-NEXT:    pand %xmm0, %xmm2
287; X86-SSE2-NEXT:    pandn %xmm1, %xmm0
288; X86-SSE2-NEXT:    por %xmm2, %xmm0
289; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
290; X86-SSE2-NEXT:    psrld $16, %xmm2
291; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
292; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1
293; X86-SSE2-NEXT:    pand %xmm1, %xmm0
294; X86-SSE2-NEXT:    pandn %xmm2, %xmm1
295; X86-SSE2-NEXT:    por %xmm0, %xmm1
296; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
297; X86-SSE2-NEXT:    psrlw $8, %xmm0
298; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2
299; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
300; X86-SSE2-NEXT:    pand %xmm2, %xmm1
301; X86-SSE2-NEXT:    pandn %xmm0, %xmm2
302; X86-SSE2-NEXT:    por %xmm1, %xmm2
303; X86-SSE2-NEXT:    movd %xmm2, %eax
304; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
305; X86-SSE2-NEXT:    retl
306;
307; X86-SSE42-LABEL: test_reduce_v16i8:
308; X86-SSE42:       ## %bb.0:
309; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
310; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
311; X86-SSE42-NEXT:    psrlw $8, %xmm1
312; X86-SSE42-NEXT:    pminub %xmm0, %xmm1
313; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0
314; X86-SSE42-NEXT:    movd %xmm0, %eax
315; X86-SSE42-NEXT:    xorb $127, %al
316; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
317; X86-SSE42-NEXT:    retl
318;
319; X86-AVX-LABEL: test_reduce_v16i8:
320; X86-AVX:       ## %bb.0:
321; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
322; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm1
323; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm0
324; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
325; X86-AVX-NEXT:    vmovd %xmm0, %eax
326; X86-AVX-NEXT:    xorb $127, %al
327; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax
328; X86-AVX-NEXT:    retl
329;
330; X64-SSE2-LABEL: test_reduce_v16i8:
331; X64-SSE2:       ## %bb.0:
332; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
333; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
334; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
335; X64-SSE2-NEXT:    pand %xmm2, %xmm0
336; X64-SSE2-NEXT:    pandn %xmm1, %xmm2
337; X64-SSE2-NEXT:    por %xmm0, %xmm2
338; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
339; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
340; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
341; X64-SSE2-NEXT:    pand %xmm1, %xmm2
342; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
343; X64-SSE2-NEXT:    por %xmm2, %xmm1
344; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0
345; X64-SSE2-NEXT:    psrld $16, %xmm0
346; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2
347; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
348; X64-SSE2-NEXT:    pand %xmm2, %xmm1
349; X64-SSE2-NEXT:    pandn %xmm0, %xmm2
350; X64-SSE2-NEXT:    por %xmm1, %xmm2
351; X64-SSE2-NEXT:    movdqa %xmm2, %xmm0
352; X64-SSE2-NEXT:    psrlw $8, %xmm0
353; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
354; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
355; X64-SSE2-NEXT:    pand %xmm1, %xmm2
356; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
357; X64-SSE2-NEXT:    por %xmm2, %xmm1
358; X64-SSE2-NEXT:    movd %xmm1, %eax
359; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
360; X64-SSE2-NEXT:    retq
361;
362; X64-SSE42-LABEL: test_reduce_v16i8:
363; X64-SSE42:       ## %bb.0:
364; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
365; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
366; X64-SSE42-NEXT:    psrlw $8, %xmm1
367; X64-SSE42-NEXT:    pminub %xmm0, %xmm1
368; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0
369; X64-SSE42-NEXT:    movd %xmm0, %eax
370; X64-SSE42-NEXT:    xorb $127, %al
371; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
372; X64-SSE42-NEXT:    retq
373;
374; X64-AVX1OR2-LABEL: test_reduce_v16i8:
375; X64-AVX1OR2:       ## %bb.0:
376; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
377; X64-AVX1OR2-NEXT:    vpsrlw $8, %xmm0, %xmm1
378; X64-AVX1OR2-NEXT:    vpminub %xmm1, %xmm0, %xmm0
379; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm0
380; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax
381; X64-AVX1OR2-NEXT:    xorb $127, %al
382; X64-AVX1OR2-NEXT:    ## kill: def $al killed $al killed $eax
383; X64-AVX1OR2-NEXT:    retq
384;
385; X64-AVX512-LABEL: test_reduce_v16i8:
386; X64-AVX512:       ## %bb.0:
387; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
388; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm1
389; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm0
390; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
391; X64-AVX512-NEXT:    vmovd %xmm0, %eax
392; X64-AVX512-NEXT:    xorb $127, %al
393; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax
394; X64-AVX512-NEXT:    retq
395  %1  = shufflevector <16 x i8> %a0, <16 x i8> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
396  %2  = icmp sgt <16 x i8> %a0, %1
397  %3  = select <16 x i1> %2, <16 x i8> %a0, <16 x i8> %1
398  %4  = shufflevector <16 x i8> %3, <16 x i8> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
399  %5  = icmp sgt <16 x i8> %3, %4
400  %6  = select <16 x i1> %5, <16 x i8> %3, <16 x i8> %4
401  %7  = shufflevector <16 x i8> %6, <16 x i8> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
402  %8  = icmp sgt <16 x i8> %6, %7
403  %9  = select <16 x i1> %8, <16 x i8> %6, <16 x i8> %7
404  %10 = shufflevector <16 x i8> %9, <16 x i8> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
405  %11 = icmp sgt <16 x i8> %9, %10
406  %12 = select <16 x i1> %11, <16 x i8> %9, <16 x i8> %10
407  %13 = extractelement <16 x i8> %12, i32 0
408  ret i8 %13
409}
410
411;
412; 256-bit Vectors
413;
414
415define i64 @test_reduce_v4i64(<4 x i64> %a0) {
416; X86-SSE2-LABEL: test_reduce_v4i64:
417; X86-SSE2:       ## %bb.0:
418; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,0,2147483648,0]
419; X86-SSE2-NEXT:    movdqa %xmm1, %xmm3
420; X86-SSE2-NEXT:    pxor %xmm2, %xmm3
421; X86-SSE2-NEXT:    movdqa %xmm0, %xmm4
422; X86-SSE2-NEXT:    pxor %xmm2, %xmm4
423; X86-SSE2-NEXT:    movdqa %xmm4, %xmm5
424; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm5
425; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
426; X86-SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
427; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
428; X86-SSE2-NEXT:    pand %xmm6, %xmm4
429; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm5[1,1,3,3]
430; X86-SSE2-NEXT:    por %xmm4, %xmm3
431; X86-SSE2-NEXT:    pand %xmm3, %xmm0
432; X86-SSE2-NEXT:    pandn %xmm1, %xmm3
433; X86-SSE2-NEXT:    por %xmm0, %xmm3
434; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
435; X86-SSE2-NEXT:    movdqa %xmm3, %xmm1
436; X86-SSE2-NEXT:    pxor %xmm2, %xmm1
437; X86-SSE2-NEXT:    pxor %xmm0, %xmm2
438; X86-SSE2-NEXT:    movdqa %xmm1, %xmm4
439; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm4
440; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[0,0,2,2]
441; X86-SSE2-NEXT:    pcmpeqd %xmm1, %xmm2
442; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
443; X86-SSE2-NEXT:    pand %xmm5, %xmm1
444; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
445; X86-SSE2-NEXT:    por %xmm1, %xmm2
446; X86-SSE2-NEXT:    pand %xmm2, %xmm3
447; X86-SSE2-NEXT:    pandn %xmm0, %xmm2
448; X86-SSE2-NEXT:    por %xmm3, %xmm2
449; X86-SSE2-NEXT:    movd %xmm2, %eax
450; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
451; X86-SSE2-NEXT:    movd %xmm0, %edx
452; X86-SSE2-NEXT:    retl
453;
454; X86-SSE42-LABEL: test_reduce_v4i64:
455; X86-SSE42:       ## %bb.0:
456; X86-SSE42-NEXT:    movdqa %xmm0, %xmm2
457; X86-SSE42-NEXT:    pcmpgtq %xmm1, %xmm0
458; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm1
459; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]
460; X86-SSE42-NEXT:    movdqa %xmm1, %xmm0
461; X86-SSE42-NEXT:    pcmpgtq %xmm2, %xmm0
462; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
463; X86-SSE42-NEXT:    movd %xmm2, %eax
464; X86-SSE42-NEXT:    pextrd $1, %xmm2, %edx
465; X86-SSE42-NEXT:    retl
466;
467; X86-AVX1-LABEL: test_reduce_v4i64:
468; X86-AVX1:       ## %bb.0:
469; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
470; X86-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
471; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
472; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
473; X86-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
474; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
475; X86-AVX1-NEXT:    vmovd %xmm0, %eax
476; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx
477; X86-AVX1-NEXT:    vzeroupper
478; X86-AVX1-NEXT:    retl
479;
480; X86-AVX2-LABEL: test_reduce_v4i64:
481; X86-AVX2:       ## %bb.0:
482; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
483; X86-AVX2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
484; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
485; X86-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
486; X86-AVX2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
487; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
488; X86-AVX2-NEXT:    vmovd %xmm0, %eax
489; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx
490; X86-AVX2-NEXT:    vzeroupper
491; X86-AVX2-NEXT:    retl
492;
493; X64-SSE2-LABEL: test_reduce_v4i64:
494; X64-SSE2:       ## %bb.0:
495; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [2147483648,2147483648]
496; X64-SSE2-NEXT:    movdqa %xmm1, %xmm3
497; X64-SSE2-NEXT:    pxor %xmm2, %xmm3
498; X64-SSE2-NEXT:    movdqa %xmm0, %xmm4
499; X64-SSE2-NEXT:    pxor %xmm2, %xmm4
500; X64-SSE2-NEXT:    movdqa %xmm4, %xmm5
501; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm5
502; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[0,0,2,2]
503; X64-SSE2-NEXT:    pcmpeqd %xmm3, %xmm4
504; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm4[1,1,3,3]
505; X64-SSE2-NEXT:    pand %xmm6, %xmm3
506; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm5[1,1,3,3]
507; X64-SSE2-NEXT:    por %xmm3, %xmm4
508; X64-SSE2-NEXT:    pand %xmm4, %xmm0
509; X64-SSE2-NEXT:    pandn %xmm1, %xmm4
510; X64-SSE2-NEXT:    por %xmm0, %xmm4
511; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
512; X64-SSE2-NEXT:    movdqa %xmm4, %xmm1
513; X64-SSE2-NEXT:    pxor %xmm2, %xmm1
514; X64-SSE2-NEXT:    pxor %xmm0, %xmm2
515; X64-SSE2-NEXT:    movdqa %xmm1, %xmm3
516; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm3
517; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
518; X64-SSE2-NEXT:    pcmpeqd %xmm1, %xmm2
519; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]
520; X64-SSE2-NEXT:    pand %xmm5, %xmm1
521; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
522; X64-SSE2-NEXT:    por %xmm1, %xmm2
523; X64-SSE2-NEXT:    pand %xmm2, %xmm4
524; X64-SSE2-NEXT:    pandn %xmm0, %xmm2
525; X64-SSE2-NEXT:    por %xmm4, %xmm2
526; X64-SSE2-NEXT:    movq %xmm2, %rax
527; X64-SSE2-NEXT:    retq
528;
529; X64-SSE42-LABEL: test_reduce_v4i64:
530; X64-SSE42:       ## %bb.0:
531; X64-SSE42-NEXT:    movdqa %xmm0, %xmm2
532; X64-SSE42-NEXT:    pcmpgtq %xmm1, %xmm0
533; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm1
534; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]
535; X64-SSE42-NEXT:    movdqa %xmm1, %xmm0
536; X64-SSE42-NEXT:    pcmpgtq %xmm2, %xmm0
537; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm2
538; X64-SSE42-NEXT:    movq %xmm2, %rax
539; X64-SSE42-NEXT:    retq
540;
541; X64-AVX1-LABEL: test_reduce_v4i64:
542; X64-AVX1:       ## %bb.0:
543; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
544; X64-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
545; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
546; X64-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
547; X64-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
548; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
549; X64-AVX1-NEXT:    vmovq %xmm0, %rax
550; X64-AVX1-NEXT:    vzeroupper
551; X64-AVX1-NEXT:    retq
552;
553; X64-AVX2-LABEL: test_reduce_v4i64:
554; X64-AVX2:       ## %bb.0:
555; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
556; X64-AVX2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
557; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
558; X64-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
559; X64-AVX2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
560; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
561; X64-AVX2-NEXT:    vmovq %xmm0, %rax
562; X64-AVX2-NEXT:    vzeroupper
563; X64-AVX2-NEXT:    retq
564;
565; X64-AVX512-LABEL: test_reduce_v4i64:
566; X64-AVX512:       ## %bb.0:
567; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
568; X64-AVX512-NEXT:    vpmaxsq %xmm1, %xmm0, %xmm0
569; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
570; X64-AVX512-NEXT:    vpmaxsq %xmm1, %xmm0, %xmm0
571; X64-AVX512-NEXT:    vmovq %xmm0, %rax
572; X64-AVX512-NEXT:    vzeroupper
573; X64-AVX512-NEXT:    retq
574  %1 = shufflevector <4 x i64> %a0, <4 x i64> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
575  %2 = icmp sgt <4 x i64> %a0, %1
576  %3 = select <4 x i1> %2, <4 x i64> %a0, <4 x i64> %1
577  %4 = shufflevector <4 x i64> %3, <4 x i64> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
578  %5 = icmp sgt <4 x i64> %3, %4
579  %6 = select <4 x i1> %5, <4 x i64> %3, <4 x i64> %4
580  %7 = extractelement <4 x i64> %6, i32 0
581  ret i64 %7
582}
583
584define i32 @test_reduce_v8i32(<8 x i32> %a0) {
585; X86-SSE2-LABEL: test_reduce_v8i32:
586; X86-SSE2:       ## %bb.0:
587; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
588; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
589; X86-SSE2-NEXT:    pand %xmm2, %xmm0
590; X86-SSE2-NEXT:    pandn %xmm1, %xmm2
591; X86-SSE2-NEXT:    por %xmm0, %xmm2
592; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
593; X86-SSE2-NEXT:    movdqa %xmm2, %xmm1
594; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
595; X86-SSE2-NEXT:    pand %xmm1, %xmm2
596; X86-SSE2-NEXT:    pandn %xmm0, %xmm1
597; X86-SSE2-NEXT:    por %xmm2, %xmm1
598; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
599; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2
600; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
601; X86-SSE2-NEXT:    pand %xmm2, %xmm1
602; X86-SSE2-NEXT:    pandn %xmm0, %xmm2
603; X86-SSE2-NEXT:    por %xmm1, %xmm2
604; X86-SSE2-NEXT:    movd %xmm2, %eax
605; X86-SSE2-NEXT:    retl
606;
607; X86-SSE42-LABEL: test_reduce_v8i32:
608; X86-SSE42:       ## %bb.0:
609; X86-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
610; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
611; X86-SSE42-NEXT:    pmaxsd %xmm0, %xmm1
612; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
613; X86-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
614; X86-SSE42-NEXT:    movd %xmm0, %eax
615; X86-SSE42-NEXT:    retl
616;
617; X86-AVX1-LABEL: test_reduce_v8i32:
618; X86-AVX1:       ## %bb.0:
619; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
620; X86-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
621; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
622; X86-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
623; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
624; X86-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
625; X86-AVX1-NEXT:    vmovd %xmm0, %eax
626; X86-AVX1-NEXT:    vzeroupper
627; X86-AVX1-NEXT:    retl
628;
629; X86-AVX2-LABEL: test_reduce_v8i32:
630; X86-AVX2:       ## %bb.0:
631; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
632; X86-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
633; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
634; X86-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
635; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
636; X86-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
637; X86-AVX2-NEXT:    vmovd %xmm0, %eax
638; X86-AVX2-NEXT:    vzeroupper
639; X86-AVX2-NEXT:    retl
640;
641; X64-SSE2-LABEL: test_reduce_v8i32:
642; X64-SSE2:       ## %bb.0:
643; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
644; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
645; X64-SSE2-NEXT:    pand %xmm2, %xmm0
646; X64-SSE2-NEXT:    pandn %xmm1, %xmm2
647; X64-SSE2-NEXT:    por %xmm0, %xmm2
648; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
649; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
650; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
651; X64-SSE2-NEXT:    pand %xmm1, %xmm2
652; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
653; X64-SSE2-NEXT:    por %xmm2, %xmm1
654; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
655; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2
656; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm2
657; X64-SSE2-NEXT:    pand %xmm2, %xmm1
658; X64-SSE2-NEXT:    pandn %xmm0, %xmm2
659; X64-SSE2-NEXT:    por %xmm1, %xmm2
660; X64-SSE2-NEXT:    movd %xmm2, %eax
661; X64-SSE2-NEXT:    retq
662;
663; X64-SSE42-LABEL: test_reduce_v8i32:
664; X64-SSE42:       ## %bb.0:
665; X64-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
666; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
667; X64-SSE42-NEXT:    pmaxsd %xmm0, %xmm1
668; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
669; X64-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
670; X64-SSE42-NEXT:    movd %xmm0, %eax
671; X64-SSE42-NEXT:    retq
672;
673; X64-AVX1-LABEL: test_reduce_v8i32:
674; X64-AVX1:       ## %bb.0:
675; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
676; X64-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
677; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
678; X64-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
679; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
680; X64-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
681; X64-AVX1-NEXT:    vmovd %xmm0, %eax
682; X64-AVX1-NEXT:    vzeroupper
683; X64-AVX1-NEXT:    retq
684;
685; X64-AVX2-LABEL: test_reduce_v8i32:
686; X64-AVX2:       ## %bb.0:
687; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
688; X64-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
689; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
690; X64-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
691; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
692; X64-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
693; X64-AVX2-NEXT:    vmovd %xmm0, %eax
694; X64-AVX2-NEXT:    vzeroupper
695; X64-AVX2-NEXT:    retq
696;
697; X64-AVX512-LABEL: test_reduce_v8i32:
698; X64-AVX512:       ## %bb.0:
699; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
700; X64-AVX512-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
701; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
702; X64-AVX512-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
703; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
704; X64-AVX512-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
705; X64-AVX512-NEXT:    vmovd %xmm0, %eax
706; X64-AVX512-NEXT:    vzeroupper
707; X64-AVX512-NEXT:    retq
708  %1  = shufflevector <8 x i32> %a0, <8 x i32> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
709  %2  = icmp sgt <8 x i32> %a0, %1
710  %3  = select <8 x i1> %2, <8 x i32> %a0, <8 x i32> %1
711  %4  = shufflevector <8 x i32> %3, <8 x i32> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
712  %5  = icmp sgt <8 x i32> %3, %4
713  %6  = select <8 x i1> %5, <8 x i32> %3, <8 x i32> %4
714  %7  = shufflevector <8 x i32> %6, <8 x i32> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
715  %8  = icmp sgt <8 x i32> %6, %7
716  %9  = select <8 x i1> %8, <8 x i32> %6, <8 x i32> %7
717  %10 = extractelement <8 x i32> %9, i32 0
718  ret i32 %10
719}
720
721define i16 @test_reduce_v16i16(<16 x i16> %a0) {
722; X86-SSE2-LABEL: test_reduce_v16i16:
723; X86-SSE2:       ## %bb.0:
724; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
725; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
726; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
727; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
728; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
729; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
730; X86-SSE2-NEXT:    psrld $16, %xmm1
731; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
732; X86-SSE2-NEXT:    movd %xmm1, %eax
733; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
734; X86-SSE2-NEXT:    retl
735;
736; X86-SSE42-LABEL: test_reduce_v16i16:
737; X86-SSE42:       ## %bb.0:
738; X86-SSE42-NEXT:    pmaxsw %xmm1, %xmm0
739; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
740; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
741; X86-SSE42-NEXT:    movd %xmm0, %eax
742; X86-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
743; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
744; X86-SSE42-NEXT:    retl
745;
746; X86-AVX1-LABEL: test_reduce_v16i16:
747; X86-AVX1:       ## %bb.0:
748; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
749; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
750; X86-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
751; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
752; X86-AVX1-NEXT:    vmovd %xmm0, %eax
753; X86-AVX1-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
754; X86-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax
755; X86-AVX1-NEXT:    vzeroupper
756; X86-AVX1-NEXT:    retl
757;
758; X86-AVX2-LABEL: test_reduce_v16i16:
759; X86-AVX2:       ## %bb.0:
760; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
761; X86-AVX2-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
762; X86-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
763; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
764; X86-AVX2-NEXT:    vmovd %xmm0, %eax
765; X86-AVX2-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
766; X86-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax
767; X86-AVX2-NEXT:    vzeroupper
768; X86-AVX2-NEXT:    retl
769;
770; X64-SSE2-LABEL: test_reduce_v16i16:
771; X64-SSE2:       ## %bb.0:
772; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
773; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
774; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
775; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
776; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
777; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
778; X64-SSE2-NEXT:    psrld $16, %xmm1
779; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
780; X64-SSE2-NEXT:    movd %xmm1, %eax
781; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
782; X64-SSE2-NEXT:    retq
783;
784; X64-SSE42-LABEL: test_reduce_v16i16:
785; X64-SSE42:       ## %bb.0:
786; X64-SSE42-NEXT:    pmaxsw %xmm1, %xmm0
787; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
788; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
789; X64-SSE42-NEXT:    movd %xmm0, %eax
790; X64-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
791; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
792; X64-SSE42-NEXT:    retq
793;
794; X64-AVX1-LABEL: test_reduce_v16i16:
795; X64-AVX1:       ## %bb.0:
796; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
797; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
798; X64-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
799; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
800; X64-AVX1-NEXT:    vmovd %xmm0, %eax
801; X64-AVX1-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
802; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax
803; X64-AVX1-NEXT:    vzeroupper
804; X64-AVX1-NEXT:    retq
805;
806; X64-AVX2-LABEL: test_reduce_v16i16:
807; X64-AVX2:       ## %bb.0:
808; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
809; X64-AVX2-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
810; X64-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
811; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
812; X64-AVX2-NEXT:    vmovd %xmm0, %eax
813; X64-AVX2-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
814; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax
815; X64-AVX2-NEXT:    vzeroupper
816; X64-AVX2-NEXT:    retq
817;
818; X64-AVX512-LABEL: test_reduce_v16i16:
819; X64-AVX512:       ## %bb.0:
820; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
821; X64-AVX512-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
822; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
823; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
824; X64-AVX512-NEXT:    vmovd %xmm0, %eax
825; X64-AVX512-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
826; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax
827; X64-AVX512-NEXT:    vzeroupper
828; X64-AVX512-NEXT:    retq
829  %1  = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
830  %2  = icmp sgt <16 x i16> %a0, %1
831  %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1
832  %4  = shufflevector <16 x i16> %3, <16 x i16> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
833  %5  = icmp sgt <16 x i16> %3, %4
834  %6  = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %4
835  %7  = shufflevector <16 x i16> %6, <16 x i16> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
836  %8  = icmp sgt <16 x i16> %6, %7
837  %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %7
838  %10 = shufflevector <16 x i16> %9, <16 x i16> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
839  %11 = icmp sgt <16 x i16> %9, %10
840  %12 = select <16 x i1> %11, <16 x i16> %9, <16 x i16> %10
841  %13 = extractelement <16 x i16> %12, i32 0
842  ret i16 %13
843}
844
845define i8 @test_reduce_v32i8(<32 x i8> %a0) {
846; X86-SSE2-LABEL: test_reduce_v32i8:
847; X86-SSE2:       ## %bb.0:
848; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
849; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
850; X86-SSE2-NEXT:    pand %xmm2, %xmm0
851; X86-SSE2-NEXT:    pandn %xmm1, %xmm2
852; X86-SSE2-NEXT:    por %xmm0, %xmm2
853; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
854; X86-SSE2-NEXT:    movdqa %xmm2, %xmm1
855; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
856; X86-SSE2-NEXT:    pand %xmm1, %xmm2
857; X86-SSE2-NEXT:    pandn %xmm0, %xmm1
858; X86-SSE2-NEXT:    por %xmm2, %xmm1
859; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]
860; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
861; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm0
862; X86-SSE2-NEXT:    pand %xmm0, %xmm1
863; X86-SSE2-NEXT:    pandn %xmm2, %xmm0
864; X86-SSE2-NEXT:    por %xmm1, %xmm0
865; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
866; X86-SSE2-NEXT:    psrld $16, %xmm2
867; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
868; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1
869; X86-SSE2-NEXT:    pand %xmm1, %xmm0
870; X86-SSE2-NEXT:    pandn %xmm2, %xmm1
871; X86-SSE2-NEXT:    por %xmm0, %xmm1
872; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
873; X86-SSE2-NEXT:    psrlw $8, %xmm0
874; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2
875; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
876; X86-SSE2-NEXT:    pand %xmm2, %xmm1
877; X86-SSE2-NEXT:    pandn %xmm0, %xmm2
878; X86-SSE2-NEXT:    por %xmm1, %xmm2
879; X86-SSE2-NEXT:    movd %xmm2, %eax
880; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
881; X86-SSE2-NEXT:    retl
882;
883; X86-SSE42-LABEL: test_reduce_v32i8:
884; X86-SSE42:       ## %bb.0:
885; X86-SSE42-NEXT:    pmaxsb %xmm1, %xmm0
886; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
887; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
888; X86-SSE42-NEXT:    psrlw $8, %xmm1
889; X86-SSE42-NEXT:    pminub %xmm0, %xmm1
890; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0
891; X86-SSE42-NEXT:    movd %xmm0, %eax
892; X86-SSE42-NEXT:    xorb $127, %al
893; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
894; X86-SSE42-NEXT:    retl
895;
896; X86-AVX1-LABEL: test_reduce_v32i8:
897; X86-AVX1:       ## %bb.0:
898; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
899; X86-AVX1-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
900; X86-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
901; X86-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1
902; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm0
903; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
904; X86-AVX1-NEXT:    vmovd %xmm0, %eax
905; X86-AVX1-NEXT:    xorb $127, %al
906; X86-AVX1-NEXT:    ## kill: def $al killed $al killed $eax
907; X86-AVX1-NEXT:    vzeroupper
908; X86-AVX1-NEXT:    retl
909;
910; X86-AVX2-LABEL: test_reduce_v32i8:
911; X86-AVX2:       ## %bb.0:
912; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
913; X86-AVX2-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
914; X86-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
915; X86-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1
916; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0
917; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
918; X86-AVX2-NEXT:    vmovd %xmm0, %eax
919; X86-AVX2-NEXT:    xorb $127, %al
920; X86-AVX2-NEXT:    ## kill: def $al killed $al killed $eax
921; X86-AVX2-NEXT:    vzeroupper
922; X86-AVX2-NEXT:    retl
923;
924; X64-SSE2-LABEL: test_reduce_v32i8:
925; X64-SSE2:       ## %bb.0:
926; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
927; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
928; X64-SSE2-NEXT:    pand %xmm2, %xmm0
929; X64-SSE2-NEXT:    pandn %xmm1, %xmm2
930; X64-SSE2-NEXT:    por %xmm0, %xmm2
931; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
932; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
933; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
934; X64-SSE2-NEXT:    pand %xmm1, %xmm2
935; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
936; X64-SSE2-NEXT:    por %xmm2, %xmm1
937; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
938; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2
939; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
940; X64-SSE2-NEXT:    pand %xmm2, %xmm1
941; X64-SSE2-NEXT:    pandn %xmm0, %xmm2
942; X64-SSE2-NEXT:    por %xmm1, %xmm2
943; X64-SSE2-NEXT:    movdqa %xmm2, %xmm0
944; X64-SSE2-NEXT:    psrld $16, %xmm0
945; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
946; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
947; X64-SSE2-NEXT:    pand %xmm1, %xmm2
948; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
949; X64-SSE2-NEXT:    por %xmm2, %xmm1
950; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0
951; X64-SSE2-NEXT:    psrlw $8, %xmm0
952; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2
953; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
954; X64-SSE2-NEXT:    pand %xmm2, %xmm1
955; X64-SSE2-NEXT:    pandn %xmm0, %xmm2
956; X64-SSE2-NEXT:    por %xmm1, %xmm2
957; X64-SSE2-NEXT:    movd %xmm2, %eax
958; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
959; X64-SSE2-NEXT:    retq
960;
961; X64-SSE42-LABEL: test_reduce_v32i8:
962; X64-SSE42:       ## %bb.0:
963; X64-SSE42-NEXT:    pmaxsb %xmm1, %xmm0
964; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
965; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
966; X64-SSE42-NEXT:    psrlw $8, %xmm1
967; X64-SSE42-NEXT:    pminub %xmm0, %xmm1
968; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0
969; X64-SSE42-NEXT:    movd %xmm0, %eax
970; X64-SSE42-NEXT:    xorb $127, %al
971; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
972; X64-SSE42-NEXT:    retq
973;
974; X64-AVX1-LABEL: test_reduce_v32i8:
975; X64-AVX1:       ## %bb.0:
976; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
977; X64-AVX1-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
978; X64-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
979; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1
980; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm0
981; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
982; X64-AVX1-NEXT:    vmovd %xmm0, %eax
983; X64-AVX1-NEXT:    xorb $127, %al
984; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax
985; X64-AVX1-NEXT:    vzeroupper
986; X64-AVX1-NEXT:    retq
987;
988; X64-AVX2-LABEL: test_reduce_v32i8:
989; X64-AVX2:       ## %bb.0:
990; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
991; X64-AVX2-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
992; X64-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
993; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1
994; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0
995; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
996; X64-AVX2-NEXT:    vmovd %xmm0, %eax
997; X64-AVX2-NEXT:    xorb $127, %al
998; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax
999; X64-AVX2-NEXT:    vzeroupper
1000; X64-AVX2-NEXT:    retq
1001;
1002; X64-AVX512-LABEL: test_reduce_v32i8:
1003; X64-AVX512:       ## %bb.0:
1004; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
1005; X64-AVX512-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
1006; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
1007; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm1
1008; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm0
1009; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
1010; X64-AVX512-NEXT:    vmovd %xmm0, %eax
1011; X64-AVX512-NEXT:    xorb $127, %al
1012; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax
1013; X64-AVX512-NEXT:    vzeroupper
1014; X64-AVX512-NEXT:    retq
1015  %1  = shufflevector <32 x i8> %a0, <32 x i8> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1016  %2  = icmp sgt <32 x i8> %a0, %1
1017  %3  = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %1
1018  %4  = shufflevector <32 x i8> %3, <32 x i8> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1019  %5  = icmp sgt <32 x i8> %3, %4
1020  %6  = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %4
1021  %7  = shufflevector <32 x i8> %6, <32 x i8> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1022  %8  = icmp sgt <32 x i8> %6, %7
1023  %9  = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %7
1024  %10 = shufflevector <32 x i8> %9, <32 x i8> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1025  %11 = icmp sgt <32 x i8> %9, %10
1026  %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %10
1027  %13 = shufflevector <32 x i8> %12, <32 x i8> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1028  %14 = icmp sgt <32 x i8> %12, %13
1029  %15 = select <32 x i1> %14, <32 x i8> %12, <32 x i8> %13
1030  %16 = extractelement <32 x i8> %15, i32 0
1031  ret i8 %16
1032}
1033
1034;
1035; 512-bit Vectors
1036;
1037
1038define i64 @test_reduce_v8i64(<8 x i64> %a0) {
1039; X86-SSE2-LABEL: test_reduce_v8i64:
1040; X86-SSE2:       ## %bb.0:
1041; X86-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,0,2147483648,0]
1042; X86-SSE2-NEXT:    movdqa %xmm2, %xmm5
1043; X86-SSE2-NEXT:    pxor %xmm4, %xmm5
1044; X86-SSE2-NEXT:    movdqa %xmm0, %xmm6
1045; X86-SSE2-NEXT:    pxor %xmm4, %xmm6
1046; X86-SSE2-NEXT:    movdqa %xmm6, %xmm7
1047; X86-SSE2-NEXT:    pcmpgtd %xmm5, %xmm7
1048; X86-SSE2-NEXT:    pcmpeqd %xmm5, %xmm6
1049; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[0,0,2,2]
1050; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
1051; X86-SSE2-NEXT:    pand %xmm5, %xmm6
1052; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
1053; X86-SSE2-NEXT:    por %xmm6, %xmm5
1054; X86-SSE2-NEXT:    pand %xmm5, %xmm0
1055; X86-SSE2-NEXT:    pandn %xmm2, %xmm5
1056; X86-SSE2-NEXT:    por %xmm0, %xmm5
1057; X86-SSE2-NEXT:    movdqa %xmm3, %xmm0
1058; X86-SSE2-NEXT:    pxor %xmm4, %xmm0
1059; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2
1060; X86-SSE2-NEXT:    pxor %xmm4, %xmm2
1061; X86-SSE2-NEXT:    movdqa %xmm2, %xmm6
1062; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm6
1063; X86-SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
1064; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm6[0,0,2,2]
1065; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm2[1,1,3,3]
1066; X86-SSE2-NEXT:    pand %xmm0, %xmm7
1067; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]
1068; X86-SSE2-NEXT:    por %xmm7, %xmm2
1069; X86-SSE2-NEXT:    pand %xmm2, %xmm1
1070; X86-SSE2-NEXT:    pandn %xmm3, %xmm2
1071; X86-SSE2-NEXT:    por %xmm1, %xmm2
1072; X86-SSE2-NEXT:    movdqa %xmm2, %xmm0
1073; X86-SSE2-NEXT:    pxor %xmm4, %xmm0
1074; X86-SSE2-NEXT:    movdqa %xmm5, %xmm1
1075; X86-SSE2-NEXT:    pxor %xmm4, %xmm1
1076; X86-SSE2-NEXT:    movdqa %xmm1, %xmm3
1077; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm3
1078; X86-SSE2-NEXT:    pcmpeqd %xmm0, %xmm1
1079; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[0,0,2,2]
1080; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]
1081; X86-SSE2-NEXT:    pand %xmm0, %xmm1
1082; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm3[1,1,3,3]
1083; X86-SSE2-NEXT:    por %xmm1, %xmm0
1084; X86-SSE2-NEXT:    pand %xmm0, %xmm5
1085; X86-SSE2-NEXT:    pandn %xmm2, %xmm0
1086; X86-SSE2-NEXT:    por %xmm5, %xmm0
1087; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1088; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
1089; X86-SSE2-NEXT:    pxor %xmm4, %xmm2
1090; X86-SSE2-NEXT:    pxor %xmm1, %xmm4
1091; X86-SSE2-NEXT:    movdqa %xmm2, %xmm3
1092; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm3
1093; X86-SSE2-NEXT:    pcmpeqd %xmm2, %xmm4
1094; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[0,0,2,2]
1095; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm4[1,1,3,3]
1096; X86-SSE2-NEXT:    pand %xmm2, %xmm4
1097; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm3[1,1,3,3]
1098; X86-SSE2-NEXT:    por %xmm4, %xmm2
1099; X86-SSE2-NEXT:    pand %xmm2, %xmm0
1100; X86-SSE2-NEXT:    pandn %xmm1, %xmm2
1101; X86-SSE2-NEXT:    por %xmm0, %xmm2
1102; X86-SSE2-NEXT:    movd %xmm2, %eax
1103; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
1104; X86-SSE2-NEXT:    movd %xmm0, %edx
1105; X86-SSE2-NEXT:    retl
1106;
1107; X86-SSE42-LABEL: test_reduce_v8i64:
1108; X86-SSE42:       ## %bb.0:
1109; X86-SSE42-NEXT:    movdqa %xmm0, %xmm4
1110; X86-SSE42-NEXT:    movdqa %xmm1, %xmm0
1111; X86-SSE42-NEXT:    pcmpgtq %xmm3, %xmm0
1112; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm3
1113; X86-SSE42-NEXT:    movdqa %xmm4, %xmm0
1114; X86-SSE42-NEXT:    pcmpgtq %xmm2, %xmm0
1115; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm4, %xmm2
1116; X86-SSE42-NEXT:    movapd %xmm2, %xmm0
1117; X86-SSE42-NEXT:    pcmpgtq %xmm3, %xmm0
1118; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm3
1119; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]
1120; X86-SSE42-NEXT:    movdqa %xmm3, %xmm0
1121; X86-SSE42-NEXT:    pcmpgtq %xmm1, %xmm0
1122; X86-SSE42-NEXT:    blendvpd %xmm0, %xmm3, %xmm1
1123; X86-SSE42-NEXT:    movd %xmm1, %eax
1124; X86-SSE42-NEXT:    pextrd $1, %xmm1, %edx
1125; X86-SSE42-NEXT:    retl
1126;
1127; X86-AVX1-LABEL: test_reduce_v8i64:
1128; X86-AVX1:       ## %bb.0:
1129; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
1130; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
1131; X86-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm4
1132; X86-AVX1-NEXT:    vblendvpd %xmm4, %xmm3, %xmm2, %xmm2
1133; X86-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm3
1134; X86-AVX1-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm0
1135; X86-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm0, %xmm1
1136; X86-AVX1-NEXT:    vblendvpd %xmm1, %xmm0, %xmm2, %xmm0
1137; X86-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
1138; X86-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
1139; X86-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
1140; X86-AVX1-NEXT:    vmovd %xmm0, %eax
1141; X86-AVX1-NEXT:    vpextrd $1, %xmm0, %edx
1142; X86-AVX1-NEXT:    vzeroupper
1143; X86-AVX1-NEXT:    retl
1144;
1145; X86-AVX2-LABEL: test_reduce_v8i64:
1146; X86-AVX2:       ## %bb.0:
1147; X86-AVX2-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm2
1148; X86-AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
1149; X86-AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm1
1150; X86-AVX2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
1151; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
1152; X86-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
1153; X86-AVX2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
1154; X86-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
1155; X86-AVX2-NEXT:    vmovd %xmm0, %eax
1156; X86-AVX2-NEXT:    vpextrd $1, %xmm0, %edx
1157; X86-AVX2-NEXT:    vzeroupper
1158; X86-AVX2-NEXT:    retl
1159;
1160; X64-SSE2-LABEL: test_reduce_v8i64:
1161; X64-SSE2:       ## %bb.0:
1162; X64-SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [2147483648,2147483648]
1163; X64-SSE2-NEXT:    movdqa %xmm2, %xmm5
1164; X64-SSE2-NEXT:    pxor %xmm4, %xmm5
1165; X64-SSE2-NEXT:    movdqa %xmm0, %xmm6
1166; X64-SSE2-NEXT:    pxor %xmm4, %xmm6
1167; X64-SSE2-NEXT:    movdqa %xmm6, %xmm7
1168; X64-SSE2-NEXT:    pcmpgtd %xmm5, %xmm7
1169; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[0,0,2,2]
1170; X64-SSE2-NEXT:    pcmpeqd %xmm5, %xmm6
1171; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm6[1,1,3,3]
1172; X64-SSE2-NEXT:    pand %xmm8, %xmm6
1173; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm7[1,1,3,3]
1174; X64-SSE2-NEXT:    por %xmm6, %xmm5
1175; X64-SSE2-NEXT:    pand %xmm5, %xmm0
1176; X64-SSE2-NEXT:    pandn %xmm2, %xmm5
1177; X64-SSE2-NEXT:    por %xmm0, %xmm5
1178; X64-SSE2-NEXT:    movdqa %xmm3, %xmm0
1179; X64-SSE2-NEXT:    pxor %xmm4, %xmm0
1180; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2
1181; X64-SSE2-NEXT:    pxor %xmm4, %xmm2
1182; X64-SSE2-NEXT:    movdqa %xmm2, %xmm6
1183; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm6
1184; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[0,0,2,2]
1185; X64-SSE2-NEXT:    pcmpeqd %xmm0, %xmm2
1186; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,3,3]
1187; X64-SSE2-NEXT:    pand %xmm7, %xmm0
1188; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]
1189; X64-SSE2-NEXT:    por %xmm0, %xmm2
1190; X64-SSE2-NEXT:    pand %xmm2, %xmm1
1191; X64-SSE2-NEXT:    pandn %xmm3, %xmm2
1192; X64-SSE2-NEXT:    por %xmm1, %xmm2
1193; X64-SSE2-NEXT:    movdqa %xmm2, %xmm0
1194; X64-SSE2-NEXT:    pxor %xmm4, %xmm0
1195; X64-SSE2-NEXT:    movdqa %xmm5, %xmm1
1196; X64-SSE2-NEXT:    pxor %xmm4, %xmm1
1197; X64-SSE2-NEXT:    movdqa %xmm1, %xmm3
1198; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm3
1199; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm3[0,0,2,2]
1200; X64-SSE2-NEXT:    pcmpeqd %xmm0, %xmm1
1201; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,3,3]
1202; X64-SSE2-NEXT:    pand %xmm6, %xmm0
1203; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[1,1,3,3]
1204; X64-SSE2-NEXT:    por %xmm0, %xmm1
1205; X64-SSE2-NEXT:    pand %xmm1, %xmm5
1206; X64-SSE2-NEXT:    pandn %xmm2, %xmm1
1207; X64-SSE2-NEXT:    por %xmm5, %xmm1
1208; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
1209; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2
1210; X64-SSE2-NEXT:    pxor %xmm4, %xmm2
1211; X64-SSE2-NEXT:    pxor %xmm0, %xmm4
1212; X64-SSE2-NEXT:    movdqa %xmm2, %xmm3
1213; X64-SSE2-NEXT:    pcmpgtd %xmm4, %xmm3
1214; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[0,0,2,2]
1215; X64-SSE2-NEXT:    pcmpeqd %xmm2, %xmm4
1216; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
1217; X64-SSE2-NEXT:    pand %xmm5, %xmm2
1218; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
1219; X64-SSE2-NEXT:    por %xmm2, %xmm3
1220; X64-SSE2-NEXT:    pand %xmm3, %xmm1
1221; X64-SSE2-NEXT:    pandn %xmm0, %xmm3
1222; X64-SSE2-NEXT:    por %xmm1, %xmm3
1223; X64-SSE2-NEXT:    movq %xmm3, %rax
1224; X64-SSE2-NEXT:    retq
1225;
1226; X64-SSE42-LABEL: test_reduce_v8i64:
1227; X64-SSE42:       ## %bb.0:
1228; X64-SSE42-NEXT:    movdqa %xmm0, %xmm4
1229; X64-SSE42-NEXT:    movdqa %xmm1, %xmm0
1230; X64-SSE42-NEXT:    pcmpgtq %xmm3, %xmm0
1231; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm1, %xmm3
1232; X64-SSE42-NEXT:    movdqa %xmm4, %xmm0
1233; X64-SSE42-NEXT:    pcmpgtq %xmm2, %xmm0
1234; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm4, %xmm2
1235; X64-SSE42-NEXT:    movapd %xmm2, %xmm0
1236; X64-SSE42-NEXT:    pcmpgtq %xmm3, %xmm0
1237; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm2, %xmm3
1238; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm3[2,3,2,3]
1239; X64-SSE42-NEXT:    movdqa %xmm3, %xmm0
1240; X64-SSE42-NEXT:    pcmpgtq %xmm1, %xmm0
1241; X64-SSE42-NEXT:    blendvpd %xmm0, %xmm3, %xmm1
1242; X64-SSE42-NEXT:    movq %xmm1, %rax
1243; X64-SSE42-NEXT:    retq
1244;
1245; X64-AVX1-LABEL: test_reduce_v8i64:
1246; X64-AVX1:       ## %bb.0:
1247; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
1248; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
1249; X64-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm3, %xmm4
1250; X64-AVX1-NEXT:    vblendvpd %xmm4, %xmm3, %xmm2, %xmm2
1251; X64-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm3
1252; X64-AVX1-NEXT:    vblendvpd %xmm3, %xmm0, %xmm1, %xmm0
1253; X64-AVX1-NEXT:    vpcmpgtq %xmm2, %xmm0, %xmm1
1254; X64-AVX1-NEXT:    vblendvpd %xmm1, %xmm0, %xmm2, %xmm0
1255; X64-AVX1-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
1256; X64-AVX1-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
1257; X64-AVX1-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
1258; X64-AVX1-NEXT:    vmovq %xmm0, %rax
1259; X64-AVX1-NEXT:    vzeroupper
1260; X64-AVX1-NEXT:    retq
1261;
1262; X64-AVX2-LABEL: test_reduce_v8i64:
1263; X64-AVX2:       ## %bb.0:
1264; X64-AVX2-NEXT:    vpcmpgtq %ymm1, %ymm0, %ymm2
1265; X64-AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
1266; X64-AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm1
1267; X64-AVX2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
1268; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
1269; X64-AVX2-NEXT:    vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
1270; X64-AVX2-NEXT:    vpcmpgtq %xmm1, %xmm0, %xmm2
1271; X64-AVX2-NEXT:    vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
1272; X64-AVX2-NEXT:    vmovq %xmm0, %rax
1273; X64-AVX2-NEXT:    vzeroupper
1274; X64-AVX2-NEXT:    retq
1275;
1276; X64-AVX512-LABEL: test_reduce_v8i64:
1277; X64-AVX512:       ## %bb.0:
1278; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
1279; X64-AVX512-NEXT:    vpmaxsq %zmm1, %zmm0, %zmm0
1280; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
1281; X64-AVX512-NEXT:    vpmaxsq %xmm1, %xmm0, %xmm0
1282; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1283; X64-AVX512-NEXT:    vpmaxsq %xmm1, %xmm0, %xmm0
1284; X64-AVX512-NEXT:    vmovq %xmm0, %rax
1285; X64-AVX512-NEXT:    vzeroupper
1286; X64-AVX512-NEXT:    retq
1287  %1  = shufflevector <8 x i64> %a0, <8 x i64> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
1288  %2  = icmp sgt <8 x i64> %a0, %1
1289  %3  = select <8 x i1> %2, <8 x i64> %a0, <8 x i64> %1
1290  %4  = shufflevector <8 x i64> %3, <8 x i64> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1291  %5  = icmp sgt <8 x i64> %3, %4
1292  %6  = select <8 x i1> %5, <8 x i64> %3, <8 x i64> %4
1293  %7  = shufflevector <8 x i64> %6, <8 x i64> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1294  %8  = icmp sgt <8 x i64> %6, %7
1295  %9  = select <8 x i1> %8, <8 x i64> %6, <8 x i64> %7
1296  %10 = extractelement <8 x i64> %9, i32 0
1297  ret i64 %10
1298}
1299
1300define i32 @test_reduce_v16i32(<16 x i32> %a0) {
1301; X86-SSE2-LABEL: test_reduce_v16i32:
1302; X86-SSE2:       ## %bb.0:
1303; X86-SSE2-NEXT:    movdqa %xmm1, %xmm4
1304; X86-SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
1305; X86-SSE2-NEXT:    pand %xmm4, %xmm1
1306; X86-SSE2-NEXT:    pandn %xmm3, %xmm4
1307; X86-SSE2-NEXT:    por %xmm1, %xmm4
1308; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
1309; X86-SSE2-NEXT:    pcmpgtd %xmm2, %xmm1
1310; X86-SSE2-NEXT:    pand %xmm1, %xmm0
1311; X86-SSE2-NEXT:    pandn %xmm2, %xmm1
1312; X86-SSE2-NEXT:    por %xmm0, %xmm1
1313; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
1314; X86-SSE2-NEXT:    pcmpgtd %xmm4, %xmm0
1315; X86-SSE2-NEXT:    pand %xmm0, %xmm1
1316; X86-SSE2-NEXT:    pandn %xmm4, %xmm0
1317; X86-SSE2-NEXT:    por %xmm1, %xmm0
1318; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1319; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
1320; X86-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
1321; X86-SSE2-NEXT:    pand %xmm2, %xmm0
1322; X86-SSE2-NEXT:    pandn %xmm1, %xmm2
1323; X86-SSE2-NEXT:    por %xmm0, %xmm2
1324; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
1325; X86-SSE2-NEXT:    movdqa %xmm2, %xmm1
1326; X86-SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
1327; X86-SSE2-NEXT:    pand %xmm1, %xmm2
1328; X86-SSE2-NEXT:    pandn %xmm0, %xmm1
1329; X86-SSE2-NEXT:    por %xmm2, %xmm1
1330; X86-SSE2-NEXT:    movd %xmm1, %eax
1331; X86-SSE2-NEXT:    retl
1332;
1333; X86-SSE42-LABEL: test_reduce_v16i32:
1334; X86-SSE42:       ## %bb.0:
1335; X86-SSE42-NEXT:    pmaxsd %xmm3, %xmm1
1336; X86-SSE42-NEXT:    pmaxsd %xmm2, %xmm0
1337; X86-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
1338; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1339; X86-SSE42-NEXT:    pmaxsd %xmm0, %xmm1
1340; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
1341; X86-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
1342; X86-SSE42-NEXT:    movd %xmm0, %eax
1343; X86-SSE42-NEXT:    retl
1344;
1345; X86-AVX1-LABEL: test_reduce_v16i32:
1346; X86-AVX1:       ## %bb.0:
1347; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
1348; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
1349; X86-AVX1-NEXT:    vpmaxsd %xmm2, %xmm3, %xmm2
1350; X86-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1351; X86-AVX1-NEXT:    vpmaxsd %xmm2, %xmm0, %xmm0
1352; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1353; X86-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1354; X86-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
1355; X86-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1356; X86-AVX1-NEXT:    vmovd %xmm0, %eax
1357; X86-AVX1-NEXT:    vzeroupper
1358; X86-AVX1-NEXT:    retl
1359;
1360; X86-AVX2-LABEL: test_reduce_v16i32:
1361; X86-AVX2:       ## %bb.0:
1362; X86-AVX2-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0
1363; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
1364; X86-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1365; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1366; X86-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1367; X86-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
1368; X86-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1369; X86-AVX2-NEXT:    vmovd %xmm0, %eax
1370; X86-AVX2-NEXT:    vzeroupper
1371; X86-AVX2-NEXT:    retl
1372;
1373; X64-SSE2-LABEL: test_reduce_v16i32:
1374; X64-SSE2:       ## %bb.0:
1375; X64-SSE2-NEXT:    movdqa %xmm1, %xmm4
1376; X64-SSE2-NEXT:    pcmpgtd %xmm3, %xmm4
1377; X64-SSE2-NEXT:    pand %xmm4, %xmm1
1378; X64-SSE2-NEXT:    pandn %xmm3, %xmm4
1379; X64-SSE2-NEXT:    por %xmm1, %xmm4
1380; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
1381; X64-SSE2-NEXT:    pcmpgtd %xmm2, %xmm1
1382; X64-SSE2-NEXT:    pand %xmm1, %xmm0
1383; X64-SSE2-NEXT:    pandn %xmm2, %xmm1
1384; X64-SSE2-NEXT:    por %xmm0, %xmm1
1385; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0
1386; X64-SSE2-NEXT:    pcmpgtd %xmm4, %xmm0
1387; X64-SSE2-NEXT:    pand %xmm0, %xmm1
1388; X64-SSE2-NEXT:    pandn %xmm4, %xmm0
1389; X64-SSE2-NEXT:    por %xmm1, %xmm0
1390; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1391; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
1392; X64-SSE2-NEXT:    pcmpgtd %xmm1, %xmm2
1393; X64-SSE2-NEXT:    pand %xmm2, %xmm0
1394; X64-SSE2-NEXT:    pandn %xmm1, %xmm2
1395; X64-SSE2-NEXT:    por %xmm0, %xmm2
1396; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
1397; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
1398; X64-SSE2-NEXT:    pcmpgtd %xmm0, %xmm1
1399; X64-SSE2-NEXT:    pand %xmm1, %xmm2
1400; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
1401; X64-SSE2-NEXT:    por %xmm2, %xmm1
1402; X64-SSE2-NEXT:    movd %xmm1, %eax
1403; X64-SSE2-NEXT:    retq
1404;
1405; X64-SSE42-LABEL: test_reduce_v16i32:
1406; X64-SSE42:       ## %bb.0:
1407; X64-SSE42-NEXT:    pmaxsd %xmm3, %xmm1
1408; X64-SSE42-NEXT:    pmaxsd %xmm2, %xmm0
1409; X64-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
1410; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1411; X64-SSE42-NEXT:    pmaxsd %xmm0, %xmm1
1412; X64-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
1413; X64-SSE42-NEXT:    pmaxsd %xmm1, %xmm0
1414; X64-SSE42-NEXT:    movd %xmm0, %eax
1415; X64-SSE42-NEXT:    retq
1416;
1417; X64-AVX1-LABEL: test_reduce_v16i32:
1418; X64-AVX1:       ## %bb.0:
1419; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
1420; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
1421; X64-AVX1-NEXT:    vpmaxsd %xmm2, %xmm3, %xmm2
1422; X64-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1423; X64-AVX1-NEXT:    vpmaxsd %xmm2, %xmm0, %xmm0
1424; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1425; X64-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1426; X64-AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
1427; X64-AVX1-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1428; X64-AVX1-NEXT:    vmovd %xmm0, %eax
1429; X64-AVX1-NEXT:    vzeroupper
1430; X64-AVX1-NEXT:    retq
1431;
1432; X64-AVX2-LABEL: test_reduce_v16i32:
1433; X64-AVX2:       ## %bb.0:
1434; X64-AVX2-NEXT:    vpmaxsd %ymm1, %ymm0, %ymm0
1435; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
1436; X64-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1437; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1438; X64-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1439; X64-AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
1440; X64-AVX2-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1441; X64-AVX2-NEXT:    vmovd %xmm0, %eax
1442; X64-AVX2-NEXT:    vzeroupper
1443; X64-AVX2-NEXT:    retq
1444;
1445; X64-AVX512-LABEL: test_reduce_v16i32:
1446; X64-AVX512:       ## %bb.0:
1447; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
1448; X64-AVX512-NEXT:    vpmaxsd %zmm1, %zmm0, %zmm0
1449; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
1450; X64-AVX512-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1451; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1452; X64-AVX512-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1453; X64-AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
1454; X64-AVX512-NEXT:    vpmaxsd %xmm1, %xmm0, %xmm0
1455; X64-AVX512-NEXT:    vmovd %xmm0, %eax
1456; X64-AVX512-NEXT:    vzeroupper
1457; X64-AVX512-NEXT:    retq
1458  %1  = shufflevector <16 x i32> %a0, <16 x i32> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1459  %2  = icmp sgt <16 x i32> %a0, %1
1460  %3  = select <16 x i1> %2, <16 x i32> %a0, <16 x i32> %1
1461  %4  = shufflevector <16 x i32> %3, <16 x i32> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1462  %5  = icmp sgt <16 x i32> %3, %4
1463  %6  = select <16 x i1> %5, <16 x i32> %3, <16 x i32> %4
1464  %7  = shufflevector <16 x i32> %6, <16 x i32> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1465  %8  = icmp sgt <16 x i32> %6, %7
1466  %9  = select <16 x i1> %8, <16 x i32> %6, <16 x i32> %7
1467  %10 = shufflevector <16 x i32> %9, <16 x i32> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1468  %11 = icmp sgt <16 x i32> %9, %10
1469  %12 = select <16 x i1> %11, <16 x i32> %9, <16 x i32> %10
1470  %13 = extractelement <16 x i32> %12, i32 0
1471  ret i32 %13
1472}
1473
1474define i16 @test_reduce_v32i16(<32 x i16> %a0) {
1475; X86-SSE2-LABEL: test_reduce_v32i16:
1476; X86-SSE2:       ## %bb.0:
1477; X86-SSE2-NEXT:    pmaxsw %xmm3, %xmm1
1478; X86-SSE2-NEXT:    pmaxsw %xmm2, %xmm0
1479; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
1480; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1481; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1482; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
1483; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
1484; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
1485; X86-SSE2-NEXT:    psrld $16, %xmm1
1486; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1487; X86-SSE2-NEXT:    movd %xmm1, %eax
1488; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
1489; X86-SSE2-NEXT:    retl
1490;
1491; X86-SSE42-LABEL: test_reduce_v32i16:
1492; X86-SSE42:       ## %bb.0:
1493; X86-SSE42-NEXT:    pmaxsw %xmm3, %xmm1
1494; X86-SSE42-NEXT:    pmaxsw %xmm2, %xmm0
1495; X86-SSE42-NEXT:    pmaxsw %xmm1, %xmm0
1496; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
1497; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
1498; X86-SSE42-NEXT:    movd %xmm0, %eax
1499; X86-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1500; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
1501; X86-SSE42-NEXT:    retl
1502;
1503; X86-AVX1-LABEL: test_reduce_v32i16:
1504; X86-AVX1:       ## %bb.0:
1505; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
1506; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
1507; X86-AVX1-NEXT:    vpmaxsw %xmm2, %xmm3, %xmm2
1508; X86-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
1509; X86-AVX1-NEXT:    vpmaxsw %xmm2, %xmm0, %xmm0
1510; X86-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
1511; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
1512; X86-AVX1-NEXT:    vmovd %xmm0, %eax
1513; X86-AVX1-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1514; X86-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax
1515; X86-AVX1-NEXT:    vzeroupper
1516; X86-AVX1-NEXT:    retl
1517;
1518; X86-AVX2-LABEL: test_reduce_v32i16:
1519; X86-AVX2:       ## %bb.0:
1520; X86-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
1521; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
1522; X86-AVX2-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
1523; X86-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
1524; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
1525; X86-AVX2-NEXT:    vmovd %xmm0, %eax
1526; X86-AVX2-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1527; X86-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax
1528; X86-AVX2-NEXT:    vzeroupper
1529; X86-AVX2-NEXT:    retl
1530;
1531; X64-SSE2-LABEL: test_reduce_v32i16:
1532; X64-SSE2:       ## %bb.0:
1533; X64-SSE2-NEXT:    pmaxsw %xmm3, %xmm1
1534; X64-SSE2-NEXT:    pmaxsw %xmm2, %xmm0
1535; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
1536; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1537; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1538; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
1539; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
1540; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
1541; X64-SSE2-NEXT:    psrld $16, %xmm1
1542; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1543; X64-SSE2-NEXT:    movd %xmm1, %eax
1544; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
1545; X64-SSE2-NEXT:    retq
1546;
1547; X64-SSE42-LABEL: test_reduce_v32i16:
1548; X64-SSE42:       ## %bb.0:
1549; X64-SSE42-NEXT:    pmaxsw %xmm3, %xmm1
1550; X64-SSE42-NEXT:    pmaxsw %xmm2, %xmm0
1551; X64-SSE42-NEXT:    pmaxsw %xmm1, %xmm0
1552; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1553; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
1554; X64-SSE42-NEXT:    movd %xmm0, %eax
1555; X64-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1556; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
1557; X64-SSE42-NEXT:    retq
1558;
1559; X64-AVX1-LABEL: test_reduce_v32i16:
1560; X64-AVX1:       ## %bb.0:
1561; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
1562; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
1563; X64-AVX1-NEXT:    vpmaxsw %xmm2, %xmm3, %xmm2
1564; X64-AVX1-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
1565; X64-AVX1-NEXT:    vpmaxsw %xmm2, %xmm0, %xmm0
1566; X64-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1567; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
1568; X64-AVX1-NEXT:    vmovd %xmm0, %eax
1569; X64-AVX1-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1570; X64-AVX1-NEXT:    ## kill: def $ax killed $ax killed $eax
1571; X64-AVX1-NEXT:    vzeroupper
1572; X64-AVX1-NEXT:    retq
1573;
1574; X64-AVX2-LABEL: test_reduce_v32i16:
1575; X64-AVX2:       ## %bb.0:
1576; X64-AVX2-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
1577; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
1578; X64-AVX2-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
1579; X64-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1580; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
1581; X64-AVX2-NEXT:    vmovd %xmm0, %eax
1582; X64-AVX2-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1583; X64-AVX2-NEXT:    ## kill: def $ax killed $ax killed $eax
1584; X64-AVX2-NEXT:    vzeroupper
1585; X64-AVX2-NEXT:    retq
1586;
1587; X64-AVX512-LABEL: test_reduce_v32i16:
1588; X64-AVX512:       ## %bb.0:
1589; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
1590; X64-AVX512-NEXT:    vpmaxsw %ymm1, %ymm0, %ymm0
1591; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
1592; X64-AVX512-NEXT:    vpmaxsw %xmm1, %xmm0, %xmm0
1593; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
1594; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
1595; X64-AVX512-NEXT:    vmovd %xmm0, %eax
1596; X64-AVX512-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1597; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax
1598; X64-AVX512-NEXT:    vzeroupper
1599; X64-AVX512-NEXT:    retq
1600  %1  = shufflevector <32 x i16> %a0, <32 x i16> undef, <32 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1601  %2  = icmp sgt <32 x i16> %a0, %1
1602  %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %1
1603  %4  = shufflevector <32 x i16> %3, <32 x i16> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1604  %5  = icmp sgt <32 x i16> %3, %4
1605  %6  = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %4
1606  %7  = shufflevector <32 x i16> %6, <32 x i16> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1607  %8  = icmp sgt <32 x i16> %6, %7
1608  %9  = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %7
1609  %10 = shufflevector <32 x i16> %9, <32 x i16> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1610  %11 = icmp sgt <32 x i16> %9, %10
1611  %12 = select <32 x i1> %11, <32 x i16> %9, <32 x i16> %10
1612  %13 = shufflevector <32 x i16> %12, <32 x i16> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1613  %14 = icmp sgt <32 x i16> %12, %13
1614  %15 = select <32 x i1> %14, <32 x i16> %12, <32 x i16> %13
1615  %16 = extractelement <32 x i16> %15, i32 0
1616  ret i16 %16
1617}
1618
1619define i8 @test_reduce_v64i8(<64 x i8> %a0) {
1620; X86-SSE2-LABEL: test_reduce_v64i8:
1621; X86-SSE2:       ## %bb.0:
1622; X86-SSE2-NEXT:    movdqa %xmm1, %xmm4
1623; X86-SSE2-NEXT:    pcmpgtb %xmm3, %xmm4
1624; X86-SSE2-NEXT:    pand %xmm4, %xmm1
1625; X86-SSE2-NEXT:    pandn %xmm3, %xmm4
1626; X86-SSE2-NEXT:    por %xmm1, %xmm4
1627; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
1628; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1
1629; X86-SSE2-NEXT:    pand %xmm1, %xmm0
1630; X86-SSE2-NEXT:    pandn %xmm2, %xmm1
1631; X86-SSE2-NEXT:    por %xmm0, %xmm1
1632; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
1633; X86-SSE2-NEXT:    pcmpgtb %xmm4, %xmm0
1634; X86-SSE2-NEXT:    pand %xmm0, %xmm1
1635; X86-SSE2-NEXT:    pandn %xmm4, %xmm0
1636; X86-SSE2-NEXT:    por %xmm1, %xmm0
1637; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1638; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
1639; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
1640; X86-SSE2-NEXT:    pand %xmm2, %xmm0
1641; X86-SSE2-NEXT:    pandn %xmm1, %xmm2
1642; X86-SSE2-NEXT:    por %xmm0, %xmm2
1643; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]
1644; X86-SSE2-NEXT:    movdqa %xmm2, %xmm0
1645; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm0
1646; X86-SSE2-NEXT:    pand %xmm0, %xmm2
1647; X86-SSE2-NEXT:    pandn %xmm1, %xmm0
1648; X86-SSE2-NEXT:    por %xmm2, %xmm0
1649; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
1650; X86-SSE2-NEXT:    psrld $16, %xmm2
1651; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
1652; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1
1653; X86-SSE2-NEXT:    pand %xmm1, %xmm0
1654; X86-SSE2-NEXT:    pandn %xmm2, %xmm1
1655; X86-SSE2-NEXT:    por %xmm0, %xmm1
1656; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
1657; X86-SSE2-NEXT:    psrlw $8, %xmm0
1658; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2
1659; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
1660; X86-SSE2-NEXT:    pand %xmm2, %xmm1
1661; X86-SSE2-NEXT:    pandn %xmm0, %xmm2
1662; X86-SSE2-NEXT:    por %xmm1, %xmm2
1663; X86-SSE2-NEXT:    movd %xmm2, %eax
1664; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
1665; X86-SSE2-NEXT:    retl
1666;
1667; X86-SSE42-LABEL: test_reduce_v64i8:
1668; X86-SSE42:       ## %bb.0:
1669; X86-SSE42-NEXT:    pmaxsb %xmm3, %xmm1
1670; X86-SSE42-NEXT:    pmaxsb %xmm2, %xmm0
1671; X86-SSE42-NEXT:    pmaxsb %xmm1, %xmm0
1672; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
1673; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
1674; X86-SSE42-NEXT:    psrlw $8, %xmm1
1675; X86-SSE42-NEXT:    pminub %xmm0, %xmm1
1676; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0
1677; X86-SSE42-NEXT:    movd %xmm0, %eax
1678; X86-SSE42-NEXT:    xorb $127, %al
1679; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
1680; X86-SSE42-NEXT:    retl
1681;
1682; X86-AVX1-LABEL: test_reduce_v64i8:
1683; X86-AVX1:       ## %bb.0:
1684; X86-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
1685; X86-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
1686; X86-AVX1-NEXT:    vpmaxsb %xmm2, %xmm3, %xmm2
1687; X86-AVX1-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
1688; X86-AVX1-NEXT:    vpmaxsb %xmm2, %xmm0, %xmm0
1689; X86-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
1690; X86-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1
1691; X86-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm0
1692; X86-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
1693; X86-AVX1-NEXT:    vmovd %xmm0, %eax
1694; X86-AVX1-NEXT:    xorb $127, %al
1695; X86-AVX1-NEXT:    ## kill: def $al killed $al killed $eax
1696; X86-AVX1-NEXT:    vzeroupper
1697; X86-AVX1-NEXT:    retl
1698;
1699; X86-AVX2-LABEL: test_reduce_v64i8:
1700; X86-AVX2:       ## %bb.0:
1701; X86-AVX2-NEXT:    vpmaxsb %ymm1, %ymm0, %ymm0
1702; X86-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
1703; X86-AVX2-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
1704; X86-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
1705; X86-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1
1706; X86-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0
1707; X86-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
1708; X86-AVX2-NEXT:    vmovd %xmm0, %eax
1709; X86-AVX2-NEXT:    xorb $127, %al
1710; X86-AVX2-NEXT:    ## kill: def $al killed $al killed $eax
1711; X86-AVX2-NEXT:    vzeroupper
1712; X86-AVX2-NEXT:    retl
1713;
1714; X64-SSE2-LABEL: test_reduce_v64i8:
1715; X64-SSE2:       ## %bb.0:
1716; X64-SSE2-NEXT:    movdqa %xmm1, %xmm4
1717; X64-SSE2-NEXT:    pcmpgtb %xmm3, %xmm4
1718; X64-SSE2-NEXT:    pand %xmm4, %xmm1
1719; X64-SSE2-NEXT:    pandn %xmm3, %xmm4
1720; X64-SSE2-NEXT:    por %xmm1, %xmm4
1721; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
1722; X64-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1
1723; X64-SSE2-NEXT:    pand %xmm1, %xmm0
1724; X64-SSE2-NEXT:    pandn %xmm2, %xmm1
1725; X64-SSE2-NEXT:    por %xmm0, %xmm1
1726; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0
1727; X64-SSE2-NEXT:    pcmpgtb %xmm4, %xmm0
1728; X64-SSE2-NEXT:    pand %xmm0, %xmm1
1729; X64-SSE2-NEXT:    pandn %xmm4, %xmm0
1730; X64-SSE2-NEXT:    por %xmm1, %xmm0
1731; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1732; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
1733; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
1734; X64-SSE2-NEXT:    pand %xmm2, %xmm0
1735; X64-SSE2-NEXT:    pandn %xmm1, %xmm2
1736; X64-SSE2-NEXT:    por %xmm0, %xmm2
1737; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
1738; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
1739; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
1740; X64-SSE2-NEXT:    pand %xmm1, %xmm2
1741; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
1742; X64-SSE2-NEXT:    por %xmm2, %xmm1
1743; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0
1744; X64-SSE2-NEXT:    psrld $16, %xmm0
1745; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2
1746; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
1747; X64-SSE2-NEXT:    pand %xmm2, %xmm1
1748; X64-SSE2-NEXT:    pandn %xmm0, %xmm2
1749; X64-SSE2-NEXT:    por %xmm1, %xmm2
1750; X64-SSE2-NEXT:    movdqa %xmm2, %xmm0
1751; X64-SSE2-NEXT:    psrlw $8, %xmm0
1752; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
1753; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
1754; X64-SSE2-NEXT:    pand %xmm1, %xmm2
1755; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
1756; X64-SSE2-NEXT:    por %xmm2, %xmm1
1757; X64-SSE2-NEXT:    movd %xmm1, %eax
1758; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
1759; X64-SSE2-NEXT:    retq
1760;
1761; X64-SSE42-LABEL: test_reduce_v64i8:
1762; X64-SSE42:       ## %bb.0:
1763; X64-SSE42-NEXT:    pmaxsb %xmm3, %xmm1
1764; X64-SSE42-NEXT:    pmaxsb %xmm2, %xmm0
1765; X64-SSE42-NEXT:    pmaxsb %xmm1, %xmm0
1766; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1767; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
1768; X64-SSE42-NEXT:    psrlw $8, %xmm1
1769; X64-SSE42-NEXT:    pminub %xmm0, %xmm1
1770; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0
1771; X64-SSE42-NEXT:    movd %xmm0, %eax
1772; X64-SSE42-NEXT:    xorb $127, %al
1773; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
1774; X64-SSE42-NEXT:    retq
1775;
1776; X64-AVX1-LABEL: test_reduce_v64i8:
1777; X64-AVX1:       ## %bb.0:
1778; X64-AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
1779; X64-AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
1780; X64-AVX1-NEXT:    vpmaxsb %xmm2, %xmm3, %xmm2
1781; X64-AVX1-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
1782; X64-AVX1-NEXT:    vpmaxsb %xmm2, %xmm0, %xmm0
1783; X64-AVX1-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1784; X64-AVX1-NEXT:    vpsrlw $8, %xmm0, %xmm1
1785; X64-AVX1-NEXT:    vpminub %xmm1, %xmm0, %xmm0
1786; X64-AVX1-NEXT:    vphminposuw %xmm0, %xmm0
1787; X64-AVX1-NEXT:    vmovd %xmm0, %eax
1788; X64-AVX1-NEXT:    xorb $127, %al
1789; X64-AVX1-NEXT:    ## kill: def $al killed $al killed $eax
1790; X64-AVX1-NEXT:    vzeroupper
1791; X64-AVX1-NEXT:    retq
1792;
1793; X64-AVX2-LABEL: test_reduce_v64i8:
1794; X64-AVX2:       ## %bb.0:
1795; X64-AVX2-NEXT:    vpmaxsb %ymm1, %ymm0, %ymm0
1796; X64-AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
1797; X64-AVX2-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
1798; X64-AVX2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1799; X64-AVX2-NEXT:    vpsrlw $8, %xmm0, %xmm1
1800; X64-AVX2-NEXT:    vpminub %xmm1, %xmm0, %xmm0
1801; X64-AVX2-NEXT:    vphminposuw %xmm0, %xmm0
1802; X64-AVX2-NEXT:    vmovd %xmm0, %eax
1803; X64-AVX2-NEXT:    xorb $127, %al
1804; X64-AVX2-NEXT:    ## kill: def $al killed $al killed $eax
1805; X64-AVX2-NEXT:    vzeroupper
1806; X64-AVX2-NEXT:    retq
1807;
1808; X64-AVX512-LABEL: test_reduce_v64i8:
1809; X64-AVX512:       ## %bb.0:
1810; X64-AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
1811; X64-AVX512-NEXT:    vpmaxsb %ymm1, %ymm0, %ymm0
1812; X64-AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
1813; X64-AVX512-NEXT:    vpmaxsb %xmm1, %xmm0, %xmm0
1814; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
1815; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm1
1816; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm0
1817; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
1818; X64-AVX512-NEXT:    vmovd %xmm0, %eax
1819; X64-AVX512-NEXT:    xorb $127, %al
1820; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax
1821; X64-AVX512-NEXT:    vzeroupper
1822; X64-AVX512-NEXT:    retq
1823  %1  = shufflevector <64 x i8> %a0, <64 x i8> undef, <64 x i32> <i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1824  %2  = icmp sgt <64 x i8> %a0, %1
1825  %3  = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %1
1826  %4  = shufflevector <64 x i8> %3, <64 x i8> undef, <64 x i32> <i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1827  %5  = icmp sgt <64 x i8> %3, %4
1828  %6  = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %4
1829  %7  = shufflevector <64 x i8> %6, <64 x i8> undef, <64 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1830  %8  = icmp sgt <64 x i8> %6, %7
1831  %9  = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %7
1832  %10 = shufflevector <64 x i8> %9, <64 x i8> undef, <64 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1833  %11 = icmp sgt <64 x i8> %9, %10
1834  %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %10
1835  %13 = shufflevector <64 x i8> %12, <64 x i8> undef, <64 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1836  %14 = icmp sgt <64 x i8> %12, %13
1837  %15 = select <64 x i1> %14, <64 x i8> %12, <64 x i8> %13
1838  %16 = shufflevector <64 x i8> %15, <64 x i8> undef, <64 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1839  %17 = icmp sgt <64 x i8> %15, %16
1840  %18 = select <64 x i1> %17, <64 x i8> %15, <64 x i8> %16
1841  %19 = extractelement <64 x i8> %18, i32 0
1842  ret i8 %19
1843}
1844
1845;
1846; Partial Vector Reductions
1847;
1848
1849define i16 @test_reduce_v16i16_v8i16(<16 x i16> %a0) {
1850; X86-SSE2-LABEL: test_reduce_v16i16_v8i16:
1851; X86-SSE2:       ## %bb.0:
1852; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1853; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1854; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
1855; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
1856; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
1857; X86-SSE2-NEXT:    psrld $16, %xmm1
1858; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1859; X86-SSE2-NEXT:    movd %xmm1, %eax
1860; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
1861; X86-SSE2-NEXT:    retl
1862;
1863; X86-SSE42-LABEL: test_reduce_v16i16_v8i16:
1864; X86-SSE42:       ## %bb.0:
1865; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
1866; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
1867; X86-SSE42-NEXT:    movd %xmm0, %eax
1868; X86-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1869; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
1870; X86-SSE42-NEXT:    retl
1871;
1872; X86-AVX-LABEL: test_reduce_v16i16_v8i16:
1873; X86-AVX:       ## %bb.0:
1874; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
1875; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
1876; X86-AVX-NEXT:    vmovd %xmm0, %eax
1877; X86-AVX-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1878; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax
1879; X86-AVX-NEXT:    vzeroupper
1880; X86-AVX-NEXT:    retl
1881;
1882; X64-SSE2-LABEL: test_reduce_v16i16_v8i16:
1883; X64-SSE2:       ## %bb.0:
1884; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1885; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1886; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
1887; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
1888; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
1889; X64-SSE2-NEXT:    psrld $16, %xmm1
1890; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1891; X64-SSE2-NEXT:    movd %xmm1, %eax
1892; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
1893; X64-SSE2-NEXT:    retq
1894;
1895; X64-SSE42-LABEL: test_reduce_v16i16_v8i16:
1896; X64-SSE42:       ## %bb.0:
1897; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1898; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
1899; X64-SSE42-NEXT:    movd %xmm0, %eax
1900; X64-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1901; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
1902; X64-SSE42-NEXT:    retq
1903;
1904; X64-AVX1OR2-LABEL: test_reduce_v16i16_v8i16:
1905; X64-AVX1OR2:       ## %bb.0:
1906; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1907; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm0
1908; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax
1909; X64-AVX1OR2-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1910; X64-AVX1OR2-NEXT:    ## kill: def $ax killed $ax killed $eax
1911; X64-AVX1OR2-NEXT:    vzeroupper
1912; X64-AVX1OR2-NEXT:    retq
1913;
1914; X64-AVX512-LABEL: test_reduce_v16i16_v8i16:
1915; X64-AVX512:       ## %bb.0:
1916; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
1917; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
1918; X64-AVX512-NEXT:    vmovd %xmm0, %eax
1919; X64-AVX512-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1920; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax
1921; X64-AVX512-NEXT:    vzeroupper
1922; X64-AVX512-NEXT:    retq
1923  %1  = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1924  %2  = icmp sgt <16 x i16> %a0, %1
1925  %3  = select <16 x i1> %2, <16 x i16> %a0, <16 x i16> %1
1926  %4  = shufflevector <16 x i16> %3, <16 x i16> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1927  %5  = icmp sgt <16 x i16> %3, %4
1928  %6  = select <16 x i1> %5, <16 x i16> %3, <16 x i16> %4
1929  %7  = shufflevector <16 x i16> %6, <16 x i16> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
1930  %8  = icmp sgt <16 x i16> %6, %7
1931  %9  = select <16 x i1> %8, <16 x i16> %6, <16 x i16> %7
1932  %10 = extractelement <16 x i16> %9, i32 0
1933  ret i16 %10
1934}
1935
1936define i16 @test_reduce_v32i16_v8i16(<32 x i16> %a0) {
1937; X86-SSE2-LABEL: test_reduce_v32i16_v8i16:
1938; X86-SSE2:       ## %bb.0:
1939; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1940; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1941; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
1942; X86-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
1943; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
1944; X86-SSE2-NEXT:    psrld $16, %xmm1
1945; X86-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1946; X86-SSE2-NEXT:    movd %xmm1, %eax
1947; X86-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
1948; X86-SSE2-NEXT:    retl
1949;
1950; X86-SSE42-LABEL: test_reduce_v32i16_v8i16:
1951; X86-SSE42:       ## %bb.0:
1952; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
1953; X86-SSE42-NEXT:    phminposuw %xmm0, %xmm0
1954; X86-SSE42-NEXT:    movd %xmm0, %eax
1955; X86-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1956; X86-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
1957; X86-SSE42-NEXT:    retl
1958;
1959; X86-AVX-LABEL: test_reduce_v32i16_v8i16:
1960; X86-AVX:       ## %bb.0:
1961; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
1962; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
1963; X86-AVX-NEXT:    vmovd %xmm0, %eax
1964; X86-AVX-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1965; X86-AVX-NEXT:    ## kill: def $ax killed $ax killed $eax
1966; X86-AVX-NEXT:    vzeroupper
1967; X86-AVX-NEXT:    retl
1968;
1969; X64-SSE2-LABEL: test_reduce_v32i16_v8i16:
1970; X64-SSE2:       ## %bb.0:
1971; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
1972; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1973; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
1974; X64-SSE2-NEXT:    pmaxsw %xmm1, %xmm0
1975; X64-SSE2-NEXT:    movdqa %xmm0, %xmm1
1976; X64-SSE2-NEXT:    psrld $16, %xmm1
1977; X64-SSE2-NEXT:    pmaxsw %xmm0, %xmm1
1978; X64-SSE2-NEXT:    movd %xmm1, %eax
1979; X64-SSE2-NEXT:    ## kill: def $ax killed $ax killed $eax
1980; X64-SSE2-NEXT:    retq
1981;
1982; X64-SSE42-LABEL: test_reduce_v32i16_v8i16:
1983; X64-SSE42:       ## %bb.0:
1984; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1985; X64-SSE42-NEXT:    phminposuw %xmm0, %xmm0
1986; X64-SSE42-NEXT:    movd %xmm0, %eax
1987; X64-SSE42-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1988; X64-SSE42-NEXT:    ## kill: def $ax killed $ax killed $eax
1989; X64-SSE42-NEXT:    retq
1990;
1991; X64-AVX1OR2-LABEL: test_reduce_v32i16_v8i16:
1992; X64-AVX1OR2:       ## %bb.0:
1993; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
1994; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm0
1995; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax
1996; X64-AVX1OR2-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
1997; X64-AVX1OR2-NEXT:    ## kill: def $ax killed $ax killed $eax
1998; X64-AVX1OR2-NEXT:    vzeroupper
1999; X64-AVX1OR2-NEXT:    retq
2000;
2001; X64-AVX512-LABEL: test_reduce_v32i16_v8i16:
2002; X64-AVX512:       ## %bb.0:
2003; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
2004; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
2005; X64-AVX512-NEXT:    vmovd %xmm0, %eax
2006; X64-AVX512-NEXT:    xorl $32767, %eax ## imm = 0x7FFF
2007; X64-AVX512-NEXT:    ## kill: def $ax killed $ax killed $eax
2008; X64-AVX512-NEXT:    vzeroupper
2009; X64-AVX512-NEXT:    retq
2010  %1  = shufflevector <32 x i16> %a0, <32 x i16> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2011  %2  = icmp sgt <32 x i16> %a0, %1
2012  %3  = select <32 x i1> %2, <32 x i16> %a0, <32 x i16> %1
2013  %4  = shufflevector <32 x i16> %3, <32 x i16> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2014  %5  = icmp sgt <32 x i16> %3, %4
2015  %6  = select <32 x i1> %5, <32 x i16> %3, <32 x i16> %4
2016  %7  = shufflevector <32 x i16> %6, <32 x i16> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2017  %8  = icmp sgt <32 x i16> %6, %7
2018  %9  = select <32 x i1> %8, <32 x i16> %6, <32 x i16> %7
2019  %10 = extractelement <32 x i16> %9, i32 0
2020  ret i16 %10
2021}
2022
2023define i8 @test_reduce_v32i8_v16i8(<32 x i8> %a0) {
2024; X86-SSE2-LABEL: test_reduce_v32i8_v16i8:
2025; X86-SSE2:       ## %bb.0:
2026; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
2027; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
2028; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
2029; X86-SSE2-NEXT:    pand %xmm2, %xmm0
2030; X86-SSE2-NEXT:    pandn %xmm1, %xmm2
2031; X86-SSE2-NEXT:    por %xmm0, %xmm2
2032; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]
2033; X86-SSE2-NEXT:    movdqa %xmm2, %xmm0
2034; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm0
2035; X86-SSE2-NEXT:    pand %xmm0, %xmm2
2036; X86-SSE2-NEXT:    pandn %xmm1, %xmm0
2037; X86-SSE2-NEXT:    por %xmm2, %xmm0
2038; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
2039; X86-SSE2-NEXT:    psrld $16, %xmm2
2040; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
2041; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1
2042; X86-SSE2-NEXT:    pand %xmm1, %xmm0
2043; X86-SSE2-NEXT:    pandn %xmm2, %xmm1
2044; X86-SSE2-NEXT:    por %xmm0, %xmm1
2045; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
2046; X86-SSE2-NEXT:    psrlw $8, %xmm0
2047; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2
2048; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
2049; X86-SSE2-NEXT:    pand %xmm2, %xmm1
2050; X86-SSE2-NEXT:    pandn %xmm0, %xmm2
2051; X86-SSE2-NEXT:    por %xmm1, %xmm2
2052; X86-SSE2-NEXT:    movd %xmm2, %eax
2053; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
2054; X86-SSE2-NEXT:    retl
2055;
2056; X86-SSE42-LABEL: test_reduce_v32i8_v16i8:
2057; X86-SSE42:       ## %bb.0:
2058; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
2059; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
2060; X86-SSE42-NEXT:    psrlw $8, %xmm1
2061; X86-SSE42-NEXT:    pminub %xmm0, %xmm1
2062; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0
2063; X86-SSE42-NEXT:    movd %xmm0, %eax
2064; X86-SSE42-NEXT:    xorb $127, %al
2065; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
2066; X86-SSE42-NEXT:    retl
2067;
2068; X86-AVX-LABEL: test_reduce_v32i8_v16i8:
2069; X86-AVX:       ## %bb.0:
2070; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
2071; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm1
2072; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm0
2073; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
2074; X86-AVX-NEXT:    vmovd %xmm0, %eax
2075; X86-AVX-NEXT:    xorb $127, %al
2076; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax
2077; X86-AVX-NEXT:    vzeroupper
2078; X86-AVX-NEXT:    retl
2079;
2080; X64-SSE2-LABEL: test_reduce_v32i8_v16i8:
2081; X64-SSE2:       ## %bb.0:
2082; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
2083; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
2084; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
2085; X64-SSE2-NEXT:    pand %xmm2, %xmm0
2086; X64-SSE2-NEXT:    pandn %xmm1, %xmm2
2087; X64-SSE2-NEXT:    por %xmm0, %xmm2
2088; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
2089; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
2090; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
2091; X64-SSE2-NEXT:    pand %xmm1, %xmm2
2092; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
2093; X64-SSE2-NEXT:    por %xmm2, %xmm1
2094; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0
2095; X64-SSE2-NEXT:    psrld $16, %xmm0
2096; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2
2097; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
2098; X64-SSE2-NEXT:    pand %xmm2, %xmm1
2099; X64-SSE2-NEXT:    pandn %xmm0, %xmm2
2100; X64-SSE2-NEXT:    por %xmm1, %xmm2
2101; X64-SSE2-NEXT:    movdqa %xmm2, %xmm0
2102; X64-SSE2-NEXT:    psrlw $8, %xmm0
2103; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
2104; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
2105; X64-SSE2-NEXT:    pand %xmm1, %xmm2
2106; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
2107; X64-SSE2-NEXT:    por %xmm2, %xmm1
2108; X64-SSE2-NEXT:    movd %xmm1, %eax
2109; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
2110; X64-SSE2-NEXT:    retq
2111;
2112; X64-SSE42-LABEL: test_reduce_v32i8_v16i8:
2113; X64-SSE42:       ## %bb.0:
2114; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
2115; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
2116; X64-SSE42-NEXT:    psrlw $8, %xmm1
2117; X64-SSE42-NEXT:    pminub %xmm0, %xmm1
2118; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0
2119; X64-SSE42-NEXT:    movd %xmm0, %eax
2120; X64-SSE42-NEXT:    xorb $127, %al
2121; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
2122; X64-SSE42-NEXT:    retq
2123;
2124; X64-AVX1OR2-LABEL: test_reduce_v32i8_v16i8:
2125; X64-AVX1OR2:       ## %bb.0:
2126; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
2127; X64-AVX1OR2-NEXT:    vpsrlw $8, %xmm0, %xmm1
2128; X64-AVX1OR2-NEXT:    vpminub %xmm1, %xmm0, %xmm0
2129; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm0
2130; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax
2131; X64-AVX1OR2-NEXT:    xorb $127, %al
2132; X64-AVX1OR2-NEXT:    ## kill: def $al killed $al killed $eax
2133; X64-AVX1OR2-NEXT:    vzeroupper
2134; X64-AVX1OR2-NEXT:    retq
2135;
2136; X64-AVX512-LABEL: test_reduce_v32i8_v16i8:
2137; X64-AVX512:       ## %bb.0:
2138; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
2139; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm1
2140; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm0
2141; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
2142; X64-AVX512-NEXT:    vmovd %xmm0, %eax
2143; X64-AVX512-NEXT:    xorb $127, %al
2144; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax
2145; X64-AVX512-NEXT:    vzeroupper
2146; X64-AVX512-NEXT:    retq
2147  %1  = shufflevector <32 x i8> %a0, <32 x i8> undef, <32 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2148  %2  = icmp sgt <32 x i8> %a0, %1
2149  %3  = select <32 x i1> %2, <32 x i8> %a0, <32 x i8> %1
2150  %4  = shufflevector <32 x i8> %3, <32 x i8> undef, <32 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2151  %5  = icmp sgt <32 x i8> %3, %4
2152  %6  = select <32 x i1> %5, <32 x i8> %3, <32 x i8> %4
2153  %7  = shufflevector <32 x i8> %6, <32 x i8> undef, <32 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2154  %8  = icmp sgt <32 x i8> %6, %7
2155  %9  = select <32 x i1> %8, <32 x i8> %6, <32 x i8> %7
2156  %10 = shufflevector <32 x i8> %9, <32 x i8> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2157  %11 = icmp sgt <32 x i8> %9, %10
2158  %12 = select <32 x i1> %11, <32 x i8> %9, <32 x i8> %10
2159  %13 = extractelement <32 x i8> %12, i32 0
2160  ret i8 %13
2161}
2162
2163define i8 @test_reduce_v64i8_v16i8(<64 x i8> %a0) {
2164; X86-SSE2-LABEL: test_reduce_v64i8_v16i8:
2165; X86-SSE2:       ## %bb.0:
2166; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
2167; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
2168; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
2169; X86-SSE2-NEXT:    pand %xmm2, %xmm0
2170; X86-SSE2-NEXT:    pandn %xmm1, %xmm2
2171; X86-SSE2-NEXT:    por %xmm0, %xmm2
2172; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,1,1,1]
2173; X86-SSE2-NEXT:    movdqa %xmm2, %xmm0
2174; X86-SSE2-NEXT:    pcmpgtb %xmm1, %xmm0
2175; X86-SSE2-NEXT:    pand %xmm0, %xmm2
2176; X86-SSE2-NEXT:    pandn %xmm1, %xmm0
2177; X86-SSE2-NEXT:    por %xmm2, %xmm0
2178; X86-SSE2-NEXT:    movdqa %xmm0, %xmm2
2179; X86-SSE2-NEXT:    psrld $16, %xmm2
2180; X86-SSE2-NEXT:    movdqa %xmm0, %xmm1
2181; X86-SSE2-NEXT:    pcmpgtb %xmm2, %xmm1
2182; X86-SSE2-NEXT:    pand %xmm1, %xmm0
2183; X86-SSE2-NEXT:    pandn %xmm2, %xmm1
2184; X86-SSE2-NEXT:    por %xmm0, %xmm1
2185; X86-SSE2-NEXT:    movdqa %xmm1, %xmm0
2186; X86-SSE2-NEXT:    psrlw $8, %xmm0
2187; X86-SSE2-NEXT:    movdqa %xmm1, %xmm2
2188; X86-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
2189; X86-SSE2-NEXT:    pand %xmm2, %xmm1
2190; X86-SSE2-NEXT:    pandn %xmm0, %xmm2
2191; X86-SSE2-NEXT:    por %xmm1, %xmm2
2192; X86-SSE2-NEXT:    movd %xmm2, %eax
2193; X86-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
2194; X86-SSE2-NEXT:    retl
2195;
2196; X86-SSE42-LABEL: test_reduce_v64i8_v16i8:
2197; X86-SSE42:       ## %bb.0:
2198; X86-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
2199; X86-SSE42-NEXT:    movdqa %xmm0, %xmm1
2200; X86-SSE42-NEXT:    psrlw $8, %xmm1
2201; X86-SSE42-NEXT:    pminub %xmm0, %xmm1
2202; X86-SSE42-NEXT:    phminposuw %xmm1, %xmm0
2203; X86-SSE42-NEXT:    movd %xmm0, %eax
2204; X86-SSE42-NEXT:    xorb $127, %al
2205; X86-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
2206; X86-SSE42-NEXT:    retl
2207;
2208; X86-AVX-LABEL: test_reduce_v64i8_v16i8:
2209; X86-AVX:       ## %bb.0:
2210; X86-AVX-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
2211; X86-AVX-NEXT:    vpsrlw $8, %xmm0, %xmm1
2212; X86-AVX-NEXT:    vpminub %xmm1, %xmm0, %xmm0
2213; X86-AVX-NEXT:    vphminposuw %xmm0, %xmm0
2214; X86-AVX-NEXT:    vmovd %xmm0, %eax
2215; X86-AVX-NEXT:    xorb $127, %al
2216; X86-AVX-NEXT:    ## kill: def $al killed $al killed $eax
2217; X86-AVX-NEXT:    vzeroupper
2218; X86-AVX-NEXT:    retl
2219;
2220; X64-SSE2-LABEL: test_reduce_v64i8_v16i8:
2221; X64-SSE2:       ## %bb.0:
2222; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
2223; X64-SSE2-NEXT:    movdqa %xmm0, %xmm2
2224; X64-SSE2-NEXT:    pcmpgtb %xmm1, %xmm2
2225; X64-SSE2-NEXT:    pand %xmm2, %xmm0
2226; X64-SSE2-NEXT:    pandn %xmm1, %xmm2
2227; X64-SSE2-NEXT:    por %xmm0, %xmm2
2228; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
2229; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
2230; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
2231; X64-SSE2-NEXT:    pand %xmm1, %xmm2
2232; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
2233; X64-SSE2-NEXT:    por %xmm2, %xmm1
2234; X64-SSE2-NEXT:    movdqa %xmm1, %xmm0
2235; X64-SSE2-NEXT:    psrld $16, %xmm0
2236; X64-SSE2-NEXT:    movdqa %xmm1, %xmm2
2237; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm2
2238; X64-SSE2-NEXT:    pand %xmm2, %xmm1
2239; X64-SSE2-NEXT:    pandn %xmm0, %xmm2
2240; X64-SSE2-NEXT:    por %xmm1, %xmm2
2241; X64-SSE2-NEXT:    movdqa %xmm2, %xmm0
2242; X64-SSE2-NEXT:    psrlw $8, %xmm0
2243; X64-SSE2-NEXT:    movdqa %xmm2, %xmm1
2244; X64-SSE2-NEXT:    pcmpgtb %xmm0, %xmm1
2245; X64-SSE2-NEXT:    pand %xmm1, %xmm2
2246; X64-SSE2-NEXT:    pandn %xmm0, %xmm1
2247; X64-SSE2-NEXT:    por %xmm2, %xmm1
2248; X64-SSE2-NEXT:    movd %xmm1, %eax
2249; X64-SSE2-NEXT:    ## kill: def $al killed $al killed $eax
2250; X64-SSE2-NEXT:    retq
2251;
2252; X64-SSE42-LABEL: test_reduce_v64i8_v16i8:
2253; X64-SSE42:       ## %bb.0:
2254; X64-SSE42-NEXT:    pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
2255; X64-SSE42-NEXT:    movdqa %xmm0, %xmm1
2256; X64-SSE42-NEXT:    psrlw $8, %xmm1
2257; X64-SSE42-NEXT:    pminub %xmm0, %xmm1
2258; X64-SSE42-NEXT:    phminposuw %xmm1, %xmm0
2259; X64-SSE42-NEXT:    movd %xmm0, %eax
2260; X64-SSE42-NEXT:    xorb $127, %al
2261; X64-SSE42-NEXT:    ## kill: def $al killed $al killed $eax
2262; X64-SSE42-NEXT:    retq
2263;
2264; X64-AVX1OR2-LABEL: test_reduce_v64i8_v16i8:
2265; X64-AVX1OR2:       ## %bb.0:
2266; X64-AVX1OR2-NEXT:    vpxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
2267; X64-AVX1OR2-NEXT:    vpsrlw $8, %xmm0, %xmm1
2268; X64-AVX1OR2-NEXT:    vpminub %xmm1, %xmm0, %xmm0
2269; X64-AVX1OR2-NEXT:    vphminposuw %xmm0, %xmm0
2270; X64-AVX1OR2-NEXT:    vmovd %xmm0, %eax
2271; X64-AVX1OR2-NEXT:    xorb $127, %al
2272; X64-AVX1OR2-NEXT:    ## kill: def $al killed $al killed $eax
2273; X64-AVX1OR2-NEXT:    vzeroupper
2274; X64-AVX1OR2-NEXT:    retq
2275;
2276; X64-AVX512-LABEL: test_reduce_v64i8_v16i8:
2277; X64-AVX512:       ## %bb.0:
2278; X64-AVX512-NEXT:    vpxord {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm0
2279; X64-AVX512-NEXT:    vpsrlw $8, %xmm0, %xmm1
2280; X64-AVX512-NEXT:    vpminub %xmm1, %xmm0, %xmm0
2281; X64-AVX512-NEXT:    vphminposuw %xmm0, %xmm0
2282; X64-AVX512-NEXT:    vmovd %xmm0, %eax
2283; X64-AVX512-NEXT:    xorb $127, %al
2284; X64-AVX512-NEXT:    ## kill: def $al killed $al killed $eax
2285; X64-AVX512-NEXT:    vzeroupper
2286; X64-AVX512-NEXT:    retq
2287  %1  = shufflevector <64 x i8> %a0, <64 x i8> undef, <64 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2288  %2  = icmp sgt <64 x i8> %a0, %1
2289  %3  = select <64 x i1> %2, <64 x i8> %a0, <64 x i8> %1
2290  %4  = shufflevector <64 x i8> %3, <64 x i8> undef, <64 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2291  %5  = icmp sgt <64 x i8> %3, %4
2292  %6  = select <64 x i1> %5, <64 x i8> %3, <64 x i8> %4
2293  %7  = shufflevector <64 x i8> %6, <64 x i8> undef, <64 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2294  %8  = icmp sgt <64 x i8> %6, %7
2295  %9  = select <64 x i1> %8, <64 x i8> %6, <64 x i8> %7
2296  %10 = shufflevector <64 x i8> %9, <64 x i8> undef, <64 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
2297  %11 = icmp sgt <64 x i8> %9, %10
2298  %12 = select <64 x i1> %11, <64 x i8> %9, <64 x i8> %10
2299  %13 = extractelement <64 x i8> %12, i32 0
2300  ret i8 %13
2301}
2302