xref: /llvm-project/llvm/test/CodeGen/X86/haddsub-4.ll (revision 69a322fed19b977d15be9500d8653496b73673e9) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3           | FileCheck %s --check-prefix=SSE
3; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3,fast-hops | FileCheck %s --check-prefix=SSE
4; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx             | FileCheck %s --check-prefixes=AVX,AVX1
5; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx,fast-hops   | FileCheck %s --check-prefixes=AVX,AVX1
6; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2            | FileCheck %s --check-prefixes=AVX,AVX2
7; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2,fast-hops  | FileCheck %s --check-prefixes=AVX,AVX2
8
9define <8 x i16> @hadd_reverse_v8i16(<8 x i16> %a0, <8 x i16> %a1) nounwind {
10; SSE-LABEL: hadd_reverse_v8i16:
11; SSE:       # %bb.0:
12; SSE-NEXT:    phaddw %xmm1, %xmm0
13; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
14; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
15; SSE-NEXT:    retq
16;
17; AVX-LABEL: hadd_reverse_v8i16:
18; AVX:       # %bb.0:
19; AVX-NEXT:    vphaddw %xmm1, %xmm0, %xmm0
20; AVX-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
21; AVX-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
22; AVX-NEXT:    retq
23  %lhs = shufflevector <8 x i16> %a0, <8 x i16> %a1, <8 x i32> <i32 7, i32 5, i32 3, i32 1, i32 15, i32 13, i32 11, i32 9>
24  %rhs = shufflevector <8 x i16> %a0, <8 x i16> %a1, <8 x i32> <i32 6, i32 4, i32 2, i32 0, i32 14, i32 12, i32 10, i32 8>
25  %add = add <8 x i16> %lhs, %rhs
26  ret <8 x i16> %add
27}
28
29define <8 x i16> @hadd_reverse2_v8i16(<8 x i16> %a0, <8 x i16> %a1) nounwind {
30; SSE-LABEL: hadd_reverse2_v8i16:
31; SSE:       # %bb.0:
32; SSE-NEXT:    movdqa {{.*#+}} xmm2 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1]
33; SSE-NEXT:    pshufb %xmm2, %xmm0
34; SSE-NEXT:    pshufb %xmm2, %xmm1
35; SSE-NEXT:    phaddw %xmm1, %xmm0
36; SSE-NEXT:    retq
37;
38; AVX-LABEL: hadd_reverse2_v8i16:
39; AVX:       # %bb.0:
40; AVX-NEXT:    vmovdqa {{.*#+}} xmm2 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1]
41; AVX-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
42; AVX-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
43; AVX-NEXT:    vphaddw %xmm1, %xmm0, %xmm0
44; AVX-NEXT:    retq
45  %shuf0 = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
46  %shuf1 = shufflevector <8 x i16> %a1, <8 x i16> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
47  %lhs = shufflevector <8 x i16> %shuf0, <8 x i16> %shuf1, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>
48  %rhs = shufflevector <8 x i16> %shuf0, <8 x i16> %shuf1, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
49  %add = add <8 x i16> %lhs, %rhs
50  ret <8 x i16> %add
51}
52
53define <8 x float> @hadd_reverse_v8f32(<8 x float> %a0, <8 x float> %a1) {
54; SSE-LABEL: hadd_reverse_v8f32:
55; SSE:       # %bb.0:
56; SSE-NEXT:    movaps %xmm0, %xmm4
57; SSE-NEXT:    haddps %xmm3, %xmm1
58; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,0,3,2]
59; SSE-NEXT:    haddps %xmm2, %xmm4
60; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,0,3,2]
61; SSE-NEXT:    movaps %xmm1, %xmm0
62; SSE-NEXT:    movaps %xmm4, %xmm1
63; SSE-NEXT:    retq
64;
65; AVX1-LABEL: hadd_reverse_v8f32:
66; AVX1:       # %bb.0:
67; AVX1-NEXT:    vhaddps %ymm1, %ymm0, %ymm0
68; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]
69; AVX1-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]
70; AVX1-NEXT:    retq
71;
72; AVX2-LABEL: hadd_reverse_v8f32:
73; AVX2:       # %bb.0:
74; AVX2-NEXT:    vhaddps %ymm1, %ymm0, %ymm0
75; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]
76; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,3,0,1]
77; AVX2-NEXT:    retq
78  %lhs = shufflevector <8 x float> %a0, <8 x float> %a1, <8 x i32> <i32 7, i32 5, i32 15, i32 13, i32 3, i32 1, i32 11, i32 9>
79  %rhs = shufflevector <8 x float> %a0, <8 x float> %a1, <8 x i32> <i32 6, i32 4, i32 14, i32 12, i32 2, i32 0, i32 10, i32 8>
80  %add = fadd <8 x float> %lhs, %rhs
81  ret <8 x float> %add
82}
83
84define <8 x float> @hadd_reverse2_v8f32(<8 x float> %a0, <8 x float> %a1) {
85; SSE-LABEL: hadd_reverse2_v8f32:
86; SSE:       # %bb.0:
87; SSE-NEXT:    movaps %xmm0, %xmm4
88; SSE-NEXT:    haddps %xmm3, %xmm1
89; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,0,3,2]
90; SSE-NEXT:    haddps %xmm2, %xmm4
91; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,0,3,2]
92; SSE-NEXT:    movaps %xmm1, %xmm0
93; SSE-NEXT:    movaps %xmm4, %xmm1
94; SSE-NEXT:    retq
95;
96; AVX1-LABEL: hadd_reverse2_v8f32:
97; AVX1:       # %bb.0:
98; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]
99; AVX1-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[3,2,1,0,7,6,5,4]
100; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,0,1]
101; AVX1-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[3,2,1,0,7,6,5,4]
102; AVX1-NEXT:    vhaddps %ymm1, %ymm0, %ymm0
103; AVX1-NEXT:    retq
104;
105; AVX2-LABEL: hadd_reverse2_v8f32:
106; AVX2:       # %bb.0:
107; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[3,2,1,0,7,6,5,4]
108; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,3,0,1]
109; AVX2-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[3,2,1,0,7,6,5,4]
110; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm1[2,3,0,1]
111; AVX2-NEXT:    vhaddps %ymm1, %ymm0, %ymm0
112; AVX2-NEXT:    retq
113  %shuf0 = shufflevector <8 x float> %a0, <8 x float> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
114  %shuf1 = shufflevector <8 x float> %a1, <8 x float> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
115  %lhs = shufflevector <8 x float> %shuf0, <8 x float> %shuf1, <8 x i32> <i32 0, i32 2, i32 8, i32 10, i32 4, i32 6, i32 12, i32 14>
116  %rhs = shufflevector <8 x float> %shuf0, <8 x float> %shuf1, <8 x i32> <i32 1, i32 3, i32 9, i32 11, i32 5, i32 7, i32 13, i32 15>
117  %add = fadd <8 x float> %lhs, %rhs
118  ret <8 x float> %add
119}
120
121define <8 x float> @hadd_reverse3_v8f32(<8 x float> %a0, <8 x float> %a1) {
122; SSE-LABEL: hadd_reverse3_v8f32:
123; SSE:       # %bb.0:
124; SSE-NEXT:    haddps %xmm1, %xmm3
125; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,0,3,2]
126; SSE-NEXT:    haddps %xmm0, %xmm2
127; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,0,3,2]
128; SSE-NEXT:    movaps %xmm3, %xmm0
129; SSE-NEXT:    movaps %xmm2, %xmm1
130; SSE-NEXT:    retq
131;
132; AVX1-LABEL: hadd_reverse3_v8f32:
133; AVX1:       # %bb.0:
134; AVX1-NEXT:    vhaddps %ymm0, %ymm1, %ymm0
135; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]
136; AVX1-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]
137; AVX1-NEXT:    retq
138;
139; AVX2-LABEL: hadd_reverse3_v8f32:
140; AVX2:       # %bb.0:
141; AVX2-NEXT:    vhaddps %ymm0, %ymm1, %ymm0
142; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]
143; AVX2-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[2,3,0,1]
144; AVX2-NEXT:    retq
145  %shuf0 = shufflevector <8 x float> %a0, <8 x float> %a1, <8 x i32> <i32 0, i32 2, i32 8, i32 10, i32 4, i32 6, i32 12, i32 14>
146  %shuf1 = shufflevector <8 x float> %a0, <8 x float> %a1, <8 x i32> <i32 1, i32 3, i32 9, i32 11, i32 5, i32 7, i32 13, i32 15>
147  %add = fadd <8 x float> %shuf0, %shuf1
148  %shuf2 = shufflevector <8 x float> %add, <8 x float> poison, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
149  ret <8 x float> %shuf2
150}
151
152define <16 x i16> @hadd_reverse_v16i16(<16 x i16> %a0, <16 x i16> %a1) nounwind {
153; SSE-LABEL: hadd_reverse_v16i16:
154; SSE:       # %bb.0:
155; SSE-NEXT:    phaddw %xmm3, %xmm1
156; SSE-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[3,2,1,0,4,5,6,7]
157; SSE-NEXT:    pshufhw {{.*#+}} xmm3 = xmm1[0,1,2,3,7,6,5,4]
158; SSE-NEXT:    phaddw %xmm2, %xmm0
159; SSE-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
160; SSE-NEXT:    pshufhw {{.*#+}} xmm1 = xmm0[0,1,2,3,7,6,5,4]
161; SSE-NEXT:    movdqa %xmm3, %xmm0
162; SSE-NEXT:    retq
163;
164; AVX1-LABEL: hadd_reverse_v16i16:
165; AVX1:       # %bb.0:
166; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
167; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
168; AVX1-NEXT:    vphaddw %xmm2, %xmm3, %xmm2
169; AVX1-NEXT:    vpshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7]
170; AVX1-NEXT:    vpshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4]
171; AVX1-NEXT:    vphaddw %xmm1, %xmm0, %xmm0
172; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
173; AVX1-NEXT:    vpshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
174; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
175; AVX1-NEXT:    retq
176;
177; AVX2-LABEL: hadd_reverse_v16i16:
178; AVX2:       # %bb.0:
179; AVX2-NEXT:    vphaddw %ymm1, %ymm0, %ymm0
180; AVX2-NEXT:    vpshuflw {{.*#+}} ymm0 = ymm0[3,2,1,0,4,5,6,7,11,10,9,8,12,13,14,15]
181; AVX2-NEXT:    vpshufhw {{.*#+}} ymm0 = ymm0[0,1,2,3,7,6,5,4,8,9,10,11,15,14,13,12]
182; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1]
183; AVX2-NEXT:    retq
184  %lhs = shufflevector <16 x i16> %a0, <16 x i16> %a1, <16 x i32> <i32 15, i32 13, i32 11, i32 9, i32 31, i32 29, i32 27, i32 25, i32 7, i32 5, i32 3, i32 1, i32 23, i32 21, i32 19, i32 17>
185  %rhs = shufflevector <16 x i16> %a0, <16 x i16> %a1, <16 x i32> <i32 14, i32 12, i32 10, i32 8, i32 30, i32 28, i32 26, i32 24, i32 6, i32 4, i32 2, i32 0, i32 22, i32 20, i32 18, i32 16>
186  %add = add <16 x i16> %lhs, %rhs
187  ret <16 x i16> %add
188}
189
190define <16 x i16> @hadd_reverse2_v16i16(<16 x i16> %a0, <16 x i16> %a1) nounwind {
191; SSE-LABEL: hadd_reverse2_v16i16:
192; SSE:       # %bb.0:
193; SSE-NEXT:    movdqa %xmm0, %xmm4
194; SSE-NEXT:    movdqa {{.*#+}} xmm0 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1]
195; SSE-NEXT:    pshufb %xmm0, %xmm4
196; SSE-NEXT:    pshufb %xmm0, %xmm1
197; SSE-NEXT:    pshufb %xmm0, %xmm2
198; SSE-NEXT:    phaddw %xmm2, %xmm4
199; SSE-NEXT:    pshufb %xmm0, %xmm3
200; SSE-NEXT:    phaddw %xmm3, %xmm1
201; SSE-NEXT:    movdqa %xmm1, %xmm0
202; SSE-NEXT:    movdqa %xmm4, %xmm1
203; SSE-NEXT:    retq
204;
205; AVX1-LABEL: hadd_reverse2_v16i16:
206; AVX1:       # %bb.0:
207; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
208; AVX1-NEXT:    vmovdqa {{.*#+}} xmm3 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1]
209; AVX1-NEXT:    vpshufb %xmm3, %xmm2, %xmm2
210; AVX1-NEXT:    vpshufb %xmm3, %xmm0, %xmm0
211; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm4
212; AVX1-NEXT:    vpshufb %xmm3, %xmm4, %xmm4
213; AVX1-NEXT:    vphaddw %xmm4, %xmm2, %xmm2
214; AVX1-NEXT:    vpshufb %xmm3, %xmm1, %xmm1
215; AVX1-NEXT:    vphaddw %xmm1, %xmm0, %xmm0
216; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
217; AVX1-NEXT:    retq
218;
219; AVX2-LABEL: hadd_reverse2_v16i16:
220; AVX2:       # %bb.0:
221; AVX2-NEXT:    vmovdqa {{.*#+}} ymm2 = [14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1,30,31,28,29,26,27,24,25,22,23,20,21,18,19,16,17]
222; AVX2-NEXT:    vpshufb %ymm2, %ymm0, %ymm0
223; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[2,3,0,1]
224; AVX2-NEXT:    vpshufb %ymm2, %ymm1, %ymm1
225; AVX2-NEXT:    vpermq {{.*#+}} ymm1 = ymm1[2,3,0,1]
226; AVX2-NEXT:    vphaddw %ymm1, %ymm0, %ymm0
227; AVX2-NEXT:    retq
228  %shuf0 = shufflevector <16 x i16> %a0, <16 x i16> undef, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
229  %shuf1 = shufflevector <16 x i16> %a1, <16 x i16> undef, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
230  %lhs = shufflevector <16 x i16> %shuf0, <16 x i16> %shuf1, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 16, i32 18, i32 20, i32 22, i32 8, i32 10, i32 12, i32 14, i32 24, i32 26, i32 28, i32 30>
231  %rhs = shufflevector <16 x i16> %shuf0, <16 x i16> %shuf1, <16 x i32> <i32 1, i32 3, i32 5, i32 7, i32 17, i32 19, i32 21, i32 23, i32 9, i32 11, i32 13, i32 15, i32 25, i32 27, i32 29, i32 31>
232  %add = add <16 x i16> %lhs, %rhs
233  ret <16 x i16> %add
234}
235
236define <8 x double> @hadd_reverse_v8f64(<8 x double> %a0, <8 x double> %a1) nounwind {
237; SSE-LABEL: hadd_reverse_v8f64:
238; SSE:       # %bb.0:
239; SSE-NEXT:    movapd %xmm1, %xmm8
240; SSE-NEXT:    movapd %xmm0, %xmm9
241; SSE-NEXT:    haddpd %xmm7, %xmm3
242; SSE-NEXT:    haddpd %xmm6, %xmm2
243; SSE-NEXT:    haddpd %xmm5, %xmm8
244; SSE-NEXT:    haddpd %xmm4, %xmm9
245; SSE-NEXT:    movapd %xmm3, %xmm0
246; SSE-NEXT:    movapd %xmm2, %xmm1
247; SSE-NEXT:    movapd %xmm8, %xmm2
248; SSE-NEXT:    movapd %xmm9, %xmm3
249; SSE-NEXT:    retq
250;
251; AVX1-LABEL: hadd_reverse_v8f64:
252; AVX1:       # %bb.0:
253; AVX1-NEXT:    vhaddpd %ymm3, %ymm1, %ymm1
254; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm1[2,3,0,1]
255; AVX1-NEXT:    vhaddpd %ymm2, %ymm0, %ymm0
256; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm0[2,3,0,1]
257; AVX1-NEXT:    vmovapd %ymm3, %ymm0
258; AVX1-NEXT:    retq
259;
260; AVX2-LABEL: hadd_reverse_v8f64:
261; AVX2:       # %bb.0:
262; AVX2-NEXT:    vhaddpd %ymm3, %ymm1, %ymm1
263; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm1[2,3,0,1]
264; AVX2-NEXT:    vhaddpd %ymm2, %ymm0, %ymm0
265; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm0[2,3,0,1]
266; AVX2-NEXT:    vmovapd %ymm3, %ymm0
267; AVX2-NEXT:    retq
268  %lhs = shufflevector <8 x double> %a0, <8 x double> %a1, <8 x i32> <i32 7, i32 15, i32 5, i32 13, i32 3, i32 11, i32 1, i32 9>
269  %rhs = shufflevector <8 x double> %a0, <8 x double> %a1, <8 x i32> <i32 6, i32 14, i32 4, i32 12, i32 2, i32 10, i32 0, i32 8>
270  %fadd = fadd <8 x double> %lhs, %rhs
271  ret <8 x double> %fadd
272}
273
274define <8 x double> @hadd_reverse2_v8f64(<8 x double> %a0, <8 x double> %a1) nounwind {
275; SSE-LABEL: hadd_reverse2_v8f64:
276; SSE:       # %bb.0:
277; SSE-NEXT:    movapd %xmm1, %xmm8
278; SSE-NEXT:    movapd %xmm0, %xmm9
279; SSE-NEXT:    haddpd %xmm7, %xmm3
280; SSE-NEXT:    haddpd %xmm6, %xmm2
281; SSE-NEXT:    haddpd %xmm5, %xmm8
282; SSE-NEXT:    haddpd %xmm4, %xmm9
283; SSE-NEXT:    movapd %xmm3, %xmm0
284; SSE-NEXT:    movapd %xmm2, %xmm1
285; SSE-NEXT:    movapd %xmm8, %xmm2
286; SSE-NEXT:    movapd %xmm9, %xmm3
287; SSE-NEXT:    retq
288;
289; AVX1-LABEL: hadd_reverse2_v8f64:
290; AVX1:       # %bb.0:
291; AVX1-NEXT:    vhaddpd %ymm3, %ymm1, %ymm1
292; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm3 = ymm1[2,3,0,1]
293; AVX1-NEXT:    vhaddpd %ymm2, %ymm0, %ymm0
294; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm0[2,3,0,1]
295; AVX1-NEXT:    vmovapd %ymm3, %ymm0
296; AVX1-NEXT:    retq
297;
298; AVX2-LABEL: hadd_reverse2_v8f64:
299; AVX2:       # %bb.0:
300; AVX2-NEXT:    vhaddpd %ymm3, %ymm1, %ymm1
301; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm1[2,3,0,1]
302; AVX2-NEXT:    vhaddpd %ymm2, %ymm0, %ymm0
303; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm0[2,3,0,1]
304; AVX2-NEXT:    vmovapd %ymm3, %ymm0
305; AVX2-NEXT:    retq
306  %shuf0 = shufflevector <8 x double> %a0, <8 x double> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
307  %shuf1 = shufflevector <8 x double> %a1, <8 x double> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
308  %lhs = shufflevector <8 x double> %shuf0, <8 x double> %shuf1, <8 x i32> <i32 0, i32 8, i32 2, i32 10, i32 4, i32 12, i32 6, i32 14>
309  %rhs = shufflevector <8 x double> %shuf0, <8 x double> %shuf1, <8 x i32> <i32 1, i32 9, i32 3, i32 11, i32 5, i32 13, i32 7, i32 15>
310  %fadd = fadd <8 x double> %lhs, %rhs
311  ret <8 x double> %fadd
312}
313
314define <16 x float> @hadd_reverse_v16f32(<16 x float> %a0, <16 x float> %a1) nounwind {
315; SSE-LABEL: hadd_reverse_v16f32:
316; SSE:       # %bb.0:
317; SSE-NEXT:    movaps %xmm0, %xmm8
318; SSE-NEXT:    haddps %xmm3, %xmm2
319; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[3,2,1,0]
320; SSE-NEXT:    haddps %xmm7, %xmm6
321; SSE-NEXT:    shufps {{.*#+}} xmm6 = xmm6[3,2,1,0]
322; SSE-NEXT:    haddps %xmm1, %xmm8
323; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[3,2,1,0]
324; SSE-NEXT:    haddps %xmm5, %xmm4
325; SSE-NEXT:    shufps {{.*#+}} xmm4 = xmm4[3,2,1,0]
326; SSE-NEXT:    movaps %xmm2, %xmm0
327; SSE-NEXT:    movaps %xmm6, %xmm1
328; SSE-NEXT:    movaps %xmm8, %xmm2
329; SSE-NEXT:    movaps %xmm4, %xmm3
330; SSE-NEXT:    retq
331;
332; AVX1-LABEL: hadd_reverse_v16f32:
333; AVX1:       # %bb.0:
334; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm4 = ymm0[2,3],ymm2[2,3]
335; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
336; AVX1-NEXT:    vhaddps %ymm0, %ymm4, %ymm2
337; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm1[2,3],ymm3[2,3]
338; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
339; AVX1-NEXT:    vhaddps %ymm1, %ymm0, %ymm0
340; AVX1-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]
341; AVX1-NEXT:    vshufps {{.*#+}} ymm1 = ymm2[1,0,3,2,5,4,7,6]
342; AVX1-NEXT:    retq
343;
344; AVX2-LABEL: hadd_reverse_v16f32:
345; AVX2:       # %bb.0:
346; AVX2-NEXT:    vhaddps %ymm3, %ymm1, %ymm1
347; AVX2-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[1,0,3,2,5,4,7,6]
348; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm1[2,0,3,1]
349; AVX2-NEXT:    vhaddps %ymm2, %ymm0, %ymm0
350; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]
351; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm0[2,0,3,1]
352; AVX2-NEXT:    vmovaps %ymm3, %ymm0
353; AVX2-NEXT:    retq
354  %lhs = shufflevector <16 x float> %a0, <16 x float> %a1, <16 x i32> <i32 15, i32 13, i32 11, i32 9, i32 31, i32 29, i32 27, i32 25, i32 7, i32 5, i32 3, i32 1, i32 23, i32 21, i32 19, i32 17>
355  %rhs = shufflevector <16 x float> %a0, <16 x float> %a1, <16 x i32> <i32 14, i32 12, i32 10, i32 8, i32 30, i32 28, i32 26, i32 24, i32 6, i32 4, i32 2, i32 0, i32 22, i32 20, i32 18, i32 16>
356  %fadd = fadd <16 x float> %lhs, %rhs
357  ret <16 x float> %fadd
358}
359
360define <16 x float> @hadd_reverse2_v16f32(<16 x float> %a0, <16 x float> %a1) nounwind {
361; SSE-LABEL: hadd_reverse2_v16f32:
362; SSE:       # %bb.0:
363; SSE-NEXT:    movaps %xmm1, %xmm8
364; SSE-NEXT:    movaps %xmm0, %xmm9
365; SSE-NEXT:    haddps %xmm7, %xmm3
366; SSE-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,0,3,2]
367; SSE-NEXT:    haddps %xmm6, %xmm2
368; SSE-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,0,3,2]
369; SSE-NEXT:    haddps %xmm5, %xmm8
370; SSE-NEXT:    shufps {{.*#+}} xmm8 = xmm8[1,0,3,2]
371; SSE-NEXT:    haddps %xmm4, %xmm9
372; SSE-NEXT:    shufps {{.*#+}} xmm9 = xmm9[1,0,3,2]
373; SSE-NEXT:    movaps %xmm3, %xmm0
374; SSE-NEXT:    movaps %xmm2, %xmm1
375; SSE-NEXT:    movaps %xmm8, %xmm2
376; SSE-NEXT:    movaps %xmm9, %xmm3
377; SSE-NEXT:    retq
378;
379; AVX1-LABEL: hadd_reverse2_v16f32:
380; AVX1:       # %bb.0:
381; AVX1-NEXT:    vhaddps %ymm3, %ymm1, %ymm1
382; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm1 = ymm1[2,3,0,1]
383; AVX1-NEXT:    vshufps {{.*#+}} ymm3 = ymm1[1,0,3,2,5,4,7,6]
384; AVX1-NEXT:    vhaddps %ymm2, %ymm0, %ymm0
385; AVX1-NEXT:    vperm2f128 {{.*#+}} ymm0 = ymm0[2,3,0,1]
386; AVX1-NEXT:    vshufps {{.*#+}} ymm1 = ymm0[1,0,3,2,5,4,7,6]
387; AVX1-NEXT:    vmovaps %ymm3, %ymm0
388; AVX1-NEXT:    retq
389;
390; AVX2-LABEL: hadd_reverse2_v16f32:
391; AVX2:       # %bb.0:
392; AVX2-NEXT:    vhaddps %ymm3, %ymm1, %ymm1
393; AVX2-NEXT:    vshufps {{.*#+}} ymm1 = ymm1[1,0,3,2,5,4,7,6]
394; AVX2-NEXT:    vpermpd {{.*#+}} ymm3 = ymm1[2,3,0,1]
395; AVX2-NEXT:    vhaddps %ymm2, %ymm0, %ymm0
396; AVX2-NEXT:    vshufps {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6]
397; AVX2-NEXT:    vpermpd {{.*#+}} ymm1 = ymm0[2,3,0,1]
398; AVX2-NEXT:    vmovaps %ymm3, %ymm0
399; AVX2-NEXT:    retq
400  %shuf0 = shufflevector <16 x float> %a0, <16 x float> undef, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
401  %shuf1 = shufflevector <16 x float> %a1, <16 x float> undef, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
402  %lhs = shufflevector <16 x float> %shuf0, <16 x float> %shuf1, <16 x i32> <i32 0, i32 2, i32 16, i32 18, i32 4, i32 6, i32 20, i32 22, i32 8, i32 10, i32 24, i32 26, i32 12, i32 14, i32 28, i32 30>
403  %rhs = shufflevector <16 x float> %shuf0, <16 x float> %shuf1, <16 x i32> <i32 1, i32 3, i32 17, i32 19, i32 5, i32 7, i32 21, i32 23, i32 9, i32 11, i32 25, i32 27, i32 13, i32 15, i32 29, i32 31>
404  %fadd = fadd <16 x float> %lhs, %rhs
405  ret <16 x float> %fadd
406}
407