1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx                       | FileCheck %s --check-prefix=AVX1
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2                      | FileCheck %s --check-prefix=AVX2
4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512bw,avx512vl,avx512f | FileCheck %s --check-prefix=AVX512
5
6; PR37427 - https://bugs.llvm.org/show_bug.cgi?id=37427
7
8define <8 x i32> @eq_zero(ptr %p, <8 x i32> %x, <8 x i32> %y) {
9; AVX1-LABEL: eq_zero:
10; AVX1:       # %bb.0:
11; AVX1-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
12; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
13; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm2, %xmm2
14; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm3
15; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
16; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm2
17; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
18; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
19; AVX1-NEXT:    retq
20;
21; AVX2-LABEL: eq_zero:
22; AVX2:       # %bb.0:
23; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
24; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
25; AVX2-NEXT:    vpcmpeqd %ymm3, %ymm2, %ymm2
26; AVX2-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
27; AVX2-NEXT:    retq
28;
29; AVX512-LABEL: eq_zero:
30; AVX512:       # %bb.0:
31; AVX512-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
32; AVX512-NEXT:    vptestnmb %xmm2, %xmm2, %k1
33; AVX512-NEXT:    vpblendmd %ymm0, %ymm1, %ymm0 {%k1}
34; AVX512-NEXT:    retq
35  %load = load <8 x i8>, ptr %p
36  %cmp = icmp eq <8 x i8> %load, zeroinitializer
37  %sel = select <8 x i1> %cmp, <8 x i32> %x, <8 x i32> %y
38  ret <8 x i32> %sel
39}
40
41define <4 x i64> @ne_zero(ptr %p, <4 x i64> %x, <4 x i64> %y) {
42; AVX1-LABEL: ne_zero:
43; AVX1:       # %bb.0:
44; AVX1-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
45; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
46; AVX1-NEXT:    vpcmpeqw %xmm3, %xmm2, %xmm2
47; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm3, %xmm3
48; AVX1-NEXT:    vpxor %xmm3, %xmm2, %xmm2
49; AVX1-NEXT:    vpmovsxwq %xmm2, %xmm3
50; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
51; AVX1-NEXT:    vpmovsxwq %xmm2, %xmm2
52; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
53; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
54; AVX1-NEXT:    retq
55;
56; AVX2-LABEL: ne_zero:
57; AVX2:       # %bb.0:
58; AVX2-NEXT:    vpmovzxwq {{.*#+}} ymm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
59; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
60; AVX2-NEXT:    vpcmpgtq %ymm3, %ymm2, %ymm2
61; AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
62; AVX2-NEXT:    retq
63;
64; AVX512-LABEL: ne_zero:
65; AVX512:       # %bb.0:
66; AVX512-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
67; AVX512-NEXT:    vptestmw %xmm2, %xmm2, %k1
68; AVX512-NEXT:    vpblendmq %ymm0, %ymm1, %ymm0 {%k1}
69; AVX512-NEXT:    retq
70  %load = load <4 x i16>, ptr %p
71  %cmp = icmp ne <4 x i16> %load, zeroinitializer
72  %sel = select <4 x i1> %cmp, <4 x i64> %x, <4 x i64> %y
73  ret <4 x i64> %sel
74}
75
76define <16 x i16> @sgt_zero(ptr %p, <16 x i16> %x, <16 x i16> %y) {
77; AVX1-LABEL: sgt_zero:
78; AVX1:       # %bb.0:
79; AVX1-NEXT:    vmovdqa (%rdi), %xmm2
80; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
81; AVX1-NEXT:    vpcmpgtb %xmm3, %xmm2, %xmm2
82; AVX1-NEXT:    vpmovsxbw %xmm2, %xmm3
83; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
84; AVX1-NEXT:    vpmovsxbw %xmm2, %xmm2
85; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
86; AVX1-NEXT:    vandnps %ymm1, %ymm2, %ymm1
87; AVX1-NEXT:    vandps %ymm2, %ymm0, %ymm0
88; AVX1-NEXT:    vorps %ymm1, %ymm0, %ymm0
89; AVX1-NEXT:    retq
90;
91; AVX2-LABEL: sgt_zero:
92; AVX2:       # %bb.0:
93; AVX2-NEXT:    vpmovsxbw (%rdi), %ymm2
94; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
95; AVX2-NEXT:    vpcmpgtw %ymm3, %ymm2, %ymm2
96; AVX2-NEXT:    vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
97; AVX2-NEXT:    retq
98;
99; AVX512-LABEL: sgt_zero:
100; AVX512:       # %bb.0:
101; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
102; AVX512-NEXT:    vpcmpltb (%rdi), %xmm2, %k1
103; AVX512-NEXT:    vpblendmw %ymm0, %ymm1, %ymm0 {%k1}
104; AVX512-NEXT:    retq
105  %load = load <16 x i8>, ptr %p
106  %cmp = icmp sgt <16 x i8> %load, zeroinitializer
107  %sel = select <16 x i1> %cmp, <16 x i16> %x, <16 x i16> %y
108  ret <16 x i16> %sel
109}
110
111define <8 x i32> @slt_zero(ptr %p, <8 x i32> %x, <8 x i32> %y) {
112; AVX1-LABEL: slt_zero:
113; AVX1:       # %bb.0:
114; AVX1-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
115; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm3
116; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
117; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm2
118; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
119; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
120; AVX1-NEXT:    retq
121;
122; AVX2-LABEL: slt_zero:
123; AVX2:       # %bb.0:
124; AVX2-NEXT:    vpmovsxbd (%rdi), %ymm2
125; AVX2-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
126; AVX2-NEXT:    retq
127;
128; AVX512-LABEL: slt_zero:
129; AVX512:       # %bb.0:
130; AVX512-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
131; AVX512-NEXT:    vpmovb2m %xmm2, %k1
132; AVX512-NEXT:    vpblendmd %ymm0, %ymm1, %ymm0 {%k1}
133; AVX512-NEXT:    retq
134  %load = load <8 x i8>, ptr %p
135  %cmp = icmp slt <8 x i8> %load, zeroinitializer
136  %sel = select <8 x i1> %cmp, <8 x i32> %x, <8 x i32> %y
137  ret <8 x i32> %sel
138}
139
140define <4 x double> @eq_zero_fp_select(ptr %p, <4 x double> %x, <4 x double> %y) {
141; AVX1-LABEL: eq_zero_fp_select:
142; AVX1:       # %bb.0:
143; AVX1-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
144; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
145; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm2, %xmm2
146; AVX1-NEXT:    vpmovsxbq %xmm2, %xmm3
147; AVX1-NEXT:    vpsrld $16, %xmm2, %xmm2
148; AVX1-NEXT:    vpmovsxbq %xmm2, %xmm2
149; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
150; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
151; AVX1-NEXT:    retq
152;
153; AVX2-LABEL: eq_zero_fp_select:
154; AVX2:       # %bb.0:
155; AVX2-NEXT:    vpmovzxbq {{.*#+}} ymm2 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero,mem[2],zero,zero,zero,zero,zero,zero,zero,mem[3],zero,zero,zero,zero,zero,zero,zero
156; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
157; AVX2-NEXT:    vpcmpeqq %ymm3, %ymm2, %ymm2
158; AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
159; AVX2-NEXT:    retq
160;
161; AVX512-LABEL: eq_zero_fp_select:
162; AVX512:       # %bb.0:
163; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
164; AVX512-NEXT:    vptestnmb %xmm2, %xmm2, %k1
165; AVX512-NEXT:    vblendmpd %ymm0, %ymm1, %ymm0 {%k1}
166; AVX512-NEXT:    retq
167  %load = load <4 x i8>, ptr %p
168  %cmp = icmp eq <4 x i8> %load, zeroinitializer
169  %sel = select <4 x i1> %cmp, <4 x double> %x, <4 x double> %y
170  ret <4 x double> %sel
171}
172
173define <8 x float> @ne_zero_fp_select(ptr %p, <8 x float> %x, <8 x float> %y) {
174; AVX1-LABEL: ne_zero_fp_select:
175; AVX1:       # %bb.0:
176; AVX1-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
177; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
178; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm2, %xmm2
179; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm3, %xmm3
180; AVX1-NEXT:    vpxor %xmm3, %xmm2, %xmm2
181; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm3
182; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
183; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm2
184; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
185; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
186; AVX1-NEXT:    retq
187;
188; AVX2-LABEL: ne_zero_fp_select:
189; AVX2:       # %bb.0:
190; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
191; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
192; AVX2-NEXT:    vpcmpgtd %ymm3, %ymm2, %ymm2
193; AVX2-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
194; AVX2-NEXT:    retq
195;
196; AVX512-LABEL: ne_zero_fp_select:
197; AVX512:       # %bb.0:
198; AVX512-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
199; AVX512-NEXT:    vptestmb %xmm2, %xmm2, %k1
200; AVX512-NEXT:    vblendmps %ymm0, %ymm1, %ymm0 {%k1}
201; AVX512-NEXT:    retq
202  %load = load <8 x i8>, ptr %p
203  %cmp = icmp ne <8 x i8> %load, zeroinitializer
204  %sel = select <8 x i1> %cmp, <8 x float> %x, <8 x float> %y
205  ret <8 x float> %sel
206}
207
208define <4 x double> @sgt_zero_fp_select(ptr %p, <4 x double> %x, <4 x double> %y) {
209; AVX1-LABEL: sgt_zero_fp_select:
210; AVX1:       # %bb.0:
211; AVX1-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
212; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
213; AVX1-NEXT:    vpcmpgtb %xmm3, %xmm2, %xmm2
214; AVX1-NEXT:    vpmovsxbq %xmm2, %xmm3
215; AVX1-NEXT:    vpsrld $16, %xmm2, %xmm2
216; AVX1-NEXT:    vpmovsxbq %xmm2, %xmm2
217; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
218; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
219; AVX1-NEXT:    retq
220;
221; AVX2-LABEL: sgt_zero_fp_select:
222; AVX2:       # %bb.0:
223; AVX2-NEXT:    vpmovsxbq (%rdi), %ymm2
224; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
225; AVX2-NEXT:    vpcmpgtq %ymm3, %ymm2, %ymm2
226; AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
227; AVX2-NEXT:    retq
228;
229; AVX512-LABEL: sgt_zero_fp_select:
230; AVX512:       # %bb.0:
231; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
232; AVX512-NEXT:    vpxor %xmm3, %xmm3, %xmm3
233; AVX512-NEXT:    vpcmpgtb %xmm3, %xmm2, %k1
234; AVX512-NEXT:    vblendmpd %ymm0, %ymm1, %ymm0 {%k1}
235; AVX512-NEXT:    retq
236  %load = load <4 x i8>, ptr %p
237  %cmp = icmp sgt <4 x i8> %load, zeroinitializer
238  %sel = select <4 x i1> %cmp, <4 x double> %x, <4 x double> %y
239  ret <4 x double> %sel
240}
241
242define <8 x float> @slt_zero_fp_select(ptr %p, <8 x float> %x, <8 x float> %y) {
243; AVX1-LABEL: slt_zero_fp_select:
244; AVX1:       # %bb.0:
245; AVX1-NEXT:    vpmovsxwd 8(%rdi), %xmm2
246; AVX1-NEXT:    vpmovsxwd (%rdi), %xmm3
247; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
248; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
249; AVX1-NEXT:    retq
250;
251; AVX2-LABEL: slt_zero_fp_select:
252; AVX2:       # %bb.0:
253; AVX2-NEXT:    vpmovsxwd (%rdi), %ymm2
254; AVX2-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
255; AVX2-NEXT:    retq
256;
257; AVX512-LABEL: slt_zero_fp_select:
258; AVX512:       # %bb.0:
259; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
260; AVX512-NEXT:    vpcmpgtw (%rdi), %xmm2, %k1
261; AVX512-NEXT:    vblendmps %ymm0, %ymm1, %ymm0 {%k1}
262; AVX512-NEXT:    retq
263  %load = load <8 x i16>, ptr %p
264  %cmp = icmp slt <8 x i16> %load, zeroinitializer
265  %sel = select <8 x i1> %cmp, <8 x float> %x, <8 x float> %y
266  ret <8 x float> %sel
267}
268
269