xref: /llvm-project/llvm/test/CodeGen/X86/vsel-cmp-load.ll (revision 6cf8bde056f3ff58bde54dbe3b320d81ad5d66d6) 
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:    vpunpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
84; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
85; AVX1-NEXT:    vandnps %ymm1, %ymm2, %ymm1
86; AVX1-NEXT:    vandps %ymm2, %ymm0, %ymm0
87; AVX1-NEXT:    vorps %ymm1, %ymm0, %ymm0
88; AVX1-NEXT:    retq
89;
90; AVX2-LABEL: sgt_zero:
91; AVX2:       # %bb.0:
92; AVX2-NEXT:    vpmovsxbw (%rdi), %ymm2
93; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
94; AVX2-NEXT:    vpcmpgtw %ymm3, %ymm2, %ymm2
95; AVX2-NEXT:    vpblendvb %ymm2, %ymm0, %ymm1, %ymm0
96; AVX2-NEXT:    retq
97;
98; AVX512-LABEL: sgt_zero:
99; AVX512:       # %bb.0:
100; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
101; AVX512-NEXT:    vpcmpltb (%rdi), %xmm2, %k1
102; AVX512-NEXT:    vpblendmw %ymm0, %ymm1, %ymm0 {%k1}
103; AVX512-NEXT:    retq
104  %load = load <16 x i8>, ptr %p
105  %cmp = icmp sgt <16 x i8> %load, zeroinitializer
106  %sel = select <16 x i1> %cmp, <16 x i16> %x, <16 x i16> %y
107  ret <16 x i16> %sel
108}
109
110define <8 x i32> @slt_zero(ptr %p, <8 x i32> %x, <8 x i32> %y) {
111; AVX1-LABEL: slt_zero:
112; AVX1:       # %bb.0:
113; AVX1-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
114; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm3
115; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
116; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm2
117; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
118; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
119; AVX1-NEXT:    retq
120;
121; AVX2-LABEL: slt_zero:
122; AVX2:       # %bb.0:
123; AVX2-NEXT:    vpmovsxbd (%rdi), %ymm2
124; AVX2-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
125; AVX2-NEXT:    retq
126;
127; AVX512-LABEL: slt_zero:
128; AVX512:       # %bb.0:
129; AVX512-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
130; AVX512-NEXT:    vpmovb2m %xmm2, %k1
131; AVX512-NEXT:    vpblendmd %ymm0, %ymm1, %ymm0 {%k1}
132; AVX512-NEXT:    retq
133  %load = load <8 x i8>, ptr %p
134  %cmp = icmp slt <8 x i8> %load, zeroinitializer
135  %sel = select <8 x i1> %cmp, <8 x i32> %x, <8 x i32> %y
136  ret <8 x i32> %sel
137}
138
139define <4 x double> @eq_zero_fp_select(ptr %p, <4 x double> %x, <4 x double> %y) {
140; AVX1-LABEL: eq_zero_fp_select:
141; AVX1:       # %bb.0:
142; AVX1-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
143; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
144; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm2, %xmm2
145; AVX1-NEXT:    vpmovsxbq %xmm2, %xmm3
146; AVX1-NEXT:    vpsrld $16, %xmm2, %xmm2
147; AVX1-NEXT:    vpmovsxbq %xmm2, %xmm2
148; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
149; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
150; AVX1-NEXT:    retq
151;
152; AVX2-LABEL: eq_zero_fp_select:
153; AVX2:       # %bb.0:
154; 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
155; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
156; AVX2-NEXT:    vpcmpeqq %ymm3, %ymm2, %ymm2
157; AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
158; AVX2-NEXT:    retq
159;
160; AVX512-LABEL: eq_zero_fp_select:
161; AVX512:       # %bb.0:
162; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
163; AVX512-NEXT:    vptestnmb %xmm2, %xmm2, %k1
164; AVX512-NEXT:    vblendmpd %ymm0, %ymm1, %ymm0 {%k1}
165; AVX512-NEXT:    retq
166  %load = load <4 x i8>, ptr %p
167  %cmp = icmp eq <4 x i8> %load, zeroinitializer
168  %sel = select <4 x i1> %cmp, <4 x double> %x, <4 x double> %y
169  ret <4 x double> %sel
170}
171
172define <8 x float> @ne_zero_fp_select(ptr %p, <8 x float> %x, <8 x float> %y) {
173; AVX1-LABEL: ne_zero_fp_select:
174; AVX1:       # %bb.0:
175; AVX1-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
176; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
177; AVX1-NEXT:    vpcmpeqb %xmm3, %xmm2, %xmm2
178; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm3, %xmm3
179; AVX1-NEXT:    vpxor %xmm3, %xmm2, %xmm2
180; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm3
181; AVX1-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[1,1,1,1]
182; AVX1-NEXT:    vpmovsxbd %xmm2, %xmm2
183; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
184; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
185; AVX1-NEXT:    retq
186;
187; AVX2-LABEL: ne_zero_fp_select:
188; AVX2:       # %bb.0:
189; 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
190; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
191; AVX2-NEXT:    vpcmpgtd %ymm3, %ymm2, %ymm2
192; AVX2-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
193; AVX2-NEXT:    retq
194;
195; AVX512-LABEL: ne_zero_fp_select:
196; AVX512:       # %bb.0:
197; AVX512-NEXT:    vmovq {{.*#+}} xmm2 = mem[0],zero
198; AVX512-NEXT:    vptestmb %xmm2, %xmm2, %k1
199; AVX512-NEXT:    vblendmps %ymm0, %ymm1, %ymm0 {%k1}
200; AVX512-NEXT:    retq
201  %load = load <8 x i8>, ptr %p
202  %cmp = icmp ne <8 x i8> %load, zeroinitializer
203  %sel = select <8 x i1> %cmp, <8 x float> %x, <8 x float> %y
204  ret <8 x float> %sel
205}
206
207define <4 x double> @sgt_zero_fp_select(ptr %p, <4 x double> %x, <4 x double> %y) {
208; AVX1-LABEL: sgt_zero_fp_select:
209; AVX1:       # %bb.0:
210; AVX1-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
211; AVX1-NEXT:    vpxor %xmm3, %xmm3, %xmm3
212; AVX1-NEXT:    vpcmpgtb %xmm3, %xmm2, %xmm2
213; AVX1-NEXT:    vpmovsxbq %xmm2, %xmm3
214; AVX1-NEXT:    vpsrld $16, %xmm2, %xmm2
215; AVX1-NEXT:    vpmovsxbq %xmm2, %xmm2
216; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
217; AVX1-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
218; AVX1-NEXT:    retq
219;
220; AVX2-LABEL: sgt_zero_fp_select:
221; AVX2:       # %bb.0:
222; AVX2-NEXT:    vpmovsxbq (%rdi), %ymm2
223; AVX2-NEXT:    vpxor %xmm3, %xmm3, %xmm3
224; AVX2-NEXT:    vpcmpgtq %ymm3, %ymm2, %ymm2
225; AVX2-NEXT:    vblendvpd %ymm2, %ymm0, %ymm1, %ymm0
226; AVX2-NEXT:    retq
227;
228; AVX512-LABEL: sgt_zero_fp_select:
229; AVX512:       # %bb.0:
230; AVX512-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
231; AVX512-NEXT:    vpxor %xmm3, %xmm3, %xmm3
232; AVX512-NEXT:    vpcmpgtb %xmm3, %xmm2, %k1
233; AVX512-NEXT:    vblendmpd %ymm0, %ymm1, %ymm0 {%k1}
234; AVX512-NEXT:    retq
235  %load = load <4 x i8>, ptr %p
236  %cmp = icmp sgt <4 x i8> %load, zeroinitializer
237  %sel = select <4 x i1> %cmp, <4 x double> %x, <4 x double> %y
238  ret <4 x double> %sel
239}
240
241define <8 x float> @slt_zero_fp_select(ptr %p, <8 x float> %x, <8 x float> %y) {
242; AVX1-LABEL: slt_zero_fp_select:
243; AVX1:       # %bb.0:
244; AVX1-NEXT:    vpmovsxwd 8(%rdi), %xmm2
245; AVX1-NEXT:    vpmovsxwd (%rdi), %xmm3
246; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm3, %ymm2
247; AVX1-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
248; AVX1-NEXT:    retq
249;
250; AVX2-LABEL: slt_zero_fp_select:
251; AVX2:       # %bb.0:
252; AVX2-NEXT:    vpmovsxwd (%rdi), %ymm2
253; AVX2-NEXT:    vblendvps %ymm2, %ymm0, %ymm1, %ymm0
254; AVX2-NEXT:    retq
255;
256; AVX512-LABEL: slt_zero_fp_select:
257; AVX512:       # %bb.0:
258; AVX512-NEXT:    vpxor %xmm2, %xmm2, %xmm2
259; AVX512-NEXT:    vpcmpgtw (%rdi), %xmm2, %k1
260; AVX512-NEXT:    vblendmps %ymm0, %ymm1, %ymm0 {%k1}
261; AVX512-NEXT:    retq
262  %load = load <8 x i16>, ptr %p
263  %cmp = icmp slt <8 x i16> %load, zeroinitializer
264  %sel = select <8 x i1> %cmp, <8 x float> %x, <8 x float> %y
265  ret <8 x float> %sel
266}
267
268