• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512cd | FileCheck %s --check-prefix=ALL --check-prefix=AVX512 --check-prefix=AVX512CD
3
4define <8 x i64> @testv8i64(<8 x i64> %in) nounwind {
5; ALL-LABEL: testv8i64:
6; ALL:       ## BB#0:
7; ALL-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
8; ALL-NEXT:    vpextrq $1, %xmm1, %rax
9; ALL-NEXT:    tzcntq %rax, %rax
10; ALL-NEXT:    vmovq %rax, %xmm2
11; ALL-NEXT:    vmovq %xmm1, %rax
12; ALL-NEXT:    tzcntq %rax, %rax
13; ALL-NEXT:    vmovq %rax, %xmm1
14; ALL-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
15; ALL-NEXT:    vextracti32x4 $2, %zmm0, %xmm2
16; ALL-NEXT:    vpextrq $1, %xmm2, %rax
17; ALL-NEXT:    tzcntq %rax, %rax
18; ALL-NEXT:    vmovq %rax, %xmm3
19; ALL-NEXT:    vmovq %xmm2, %rax
20; ALL-NEXT:    tzcntq %rax, %rax
21; ALL-NEXT:    vmovq %rax, %xmm2
22; ALL-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
23; ALL-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
24; ALL-NEXT:    vextracti32x4 $1, %zmm0, %xmm2
25; ALL-NEXT:    vpextrq $1, %xmm2, %rax
26; ALL-NEXT:    tzcntq %rax, %rax
27; ALL-NEXT:    vmovq %rax, %xmm3
28; ALL-NEXT:    vmovq %xmm2, %rax
29; ALL-NEXT:    tzcntq %rax, %rax
30; ALL-NEXT:    vmovq %rax, %xmm2
31; ALL-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
32; ALL-NEXT:    vpextrq $1, %xmm0, %rax
33; ALL-NEXT:    tzcntq %rax, %rax
34; ALL-NEXT:    vmovq %rax, %xmm3
35; ALL-NEXT:    vmovq %xmm0, %rax
36; ALL-NEXT:    tzcntq %rax, %rax
37; ALL-NEXT:    vmovq %rax, %xmm0
38; ALL-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
39; ALL-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
40; ALL-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
41; ALL-NEXT:    retq
42  %out = call <8 x i64> @llvm.cttz.v8i64(<8 x i64> %in, i1 0)
43  ret <8 x i64> %out
44}
45
46define <8 x i64> @testv8i64u(<8 x i64> %in) nounwind {
47; ALL-LABEL: testv8i64u:
48; ALL:       ## BB#0:
49; ALL-NEXT:    vpxord %zmm1, %zmm1, %zmm1
50; ALL-NEXT:    vpsubq %zmm0, %zmm1, %zmm1
51; ALL-NEXT:    vpandq %zmm1, %zmm0, %zmm0
52; ALL-NEXT:    vplzcntq %zmm0, %zmm0
53; ALL-NEXT:    vpbroadcastq {{.*}}(%rip), %zmm1
54; ALL-NEXT:    vpsubq %zmm0, %zmm1, %zmm0
55; ALL-NEXT:    retq
56  %out = call <8 x i64> @llvm.cttz.v8i64(<8 x i64> %in, i1 -1)
57  ret <8 x i64> %out
58}
59
60define <16 x i32> @testv16i32(<16 x i32> %in) nounwind {
61; ALL-LABEL: testv16i32:
62; ALL:       ## BB#0:
63; ALL-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
64; ALL-NEXT:    vpextrd $1, %xmm1, %eax
65; ALL-NEXT:    tzcntl %eax, %eax
66; ALL-NEXT:    vmovd %xmm1, %ecx
67; ALL-NEXT:    tzcntl %ecx, %ecx
68; ALL-NEXT:    vmovd %ecx, %xmm2
69; ALL-NEXT:    vpinsrd $1, %eax, %xmm2, %xmm2
70; ALL-NEXT:    vpextrd $2, %xmm1, %eax
71; ALL-NEXT:    tzcntl %eax, %eax
72; ALL-NEXT:    vpinsrd $2, %eax, %xmm2, %xmm2
73; ALL-NEXT:    vpextrd $3, %xmm1, %eax
74; ALL-NEXT:    tzcntl %eax, %eax
75; ALL-NEXT:    vpinsrd $3, %eax, %xmm2, %xmm1
76; ALL-NEXT:    vextracti32x4 $2, %zmm0, %xmm2
77; ALL-NEXT:    vpextrd $1, %xmm2, %eax
78; ALL-NEXT:    tzcntl %eax, %eax
79; ALL-NEXT:    vmovd %xmm2, %ecx
80; ALL-NEXT:    tzcntl %ecx, %ecx
81; ALL-NEXT:    vmovd %ecx, %xmm3
82; ALL-NEXT:    vpinsrd $1, %eax, %xmm3, %xmm3
83; ALL-NEXT:    vpextrd $2, %xmm2, %eax
84; ALL-NEXT:    tzcntl %eax, %eax
85; ALL-NEXT:    vpinsrd $2, %eax, %xmm3, %xmm3
86; ALL-NEXT:    vpextrd $3, %xmm2, %eax
87; ALL-NEXT:    tzcntl %eax, %eax
88; ALL-NEXT:    vpinsrd $3, %eax, %xmm3, %xmm2
89; ALL-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
90; ALL-NEXT:    vextracti32x4 $1, %zmm0, %xmm2
91; ALL-NEXT:    vpextrd $1, %xmm2, %eax
92; ALL-NEXT:    tzcntl %eax, %eax
93; ALL-NEXT:    vmovd %xmm2, %ecx
94; ALL-NEXT:    tzcntl %ecx, %ecx
95; ALL-NEXT:    vmovd %ecx, %xmm3
96; ALL-NEXT:    vpinsrd $1, %eax, %xmm3, %xmm3
97; ALL-NEXT:    vpextrd $2, %xmm2, %eax
98; ALL-NEXT:    tzcntl %eax, %eax
99; ALL-NEXT:    vpinsrd $2, %eax, %xmm3, %xmm3
100; ALL-NEXT:    vpextrd $3, %xmm2, %eax
101; ALL-NEXT:    tzcntl %eax, %eax
102; ALL-NEXT:    vpinsrd $3, %eax, %xmm3, %xmm2
103; ALL-NEXT:    vpextrd $1, %xmm0, %eax
104; ALL-NEXT:    tzcntl %eax, %eax
105; ALL-NEXT:    vmovd %xmm0, %ecx
106; ALL-NEXT:    tzcntl %ecx, %ecx
107; ALL-NEXT:    vmovd %ecx, %xmm3
108; ALL-NEXT:    vpinsrd $1, %eax, %xmm3, %xmm3
109; ALL-NEXT:    vpextrd $2, %xmm0, %eax
110; ALL-NEXT:    tzcntl %eax, %eax
111; ALL-NEXT:    vpinsrd $2, %eax, %xmm3, %xmm3
112; ALL-NEXT:    vpextrd $3, %xmm0, %eax
113; ALL-NEXT:    tzcntl %eax, %eax
114; ALL-NEXT:    vpinsrd $3, %eax, %xmm3, %xmm0
115; ALL-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
116; ALL-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
117; ALL-NEXT:    retq
118  %out = call <16 x i32> @llvm.cttz.v16i32(<16 x i32> %in, i1 0)
119  ret <16 x i32> %out
120}
121
122define <16 x i32> @testv16i32u(<16 x i32> %in) nounwind {
123; ALL-LABEL: testv16i32u:
124; ALL:       ## BB#0:
125; ALL-NEXT:    vpxord %zmm1, %zmm1, %zmm1
126; ALL-NEXT:    vpsubd %zmm0, %zmm1, %zmm1
127; ALL-NEXT:    vpandd %zmm1, %zmm0, %zmm0
128; ALL-NEXT:    vplzcntd %zmm0, %zmm0
129; ALL-NEXT:    vpbroadcastd {{.*}}(%rip), %zmm1
130; ALL-NEXT:    vpsubd %zmm0, %zmm1, %zmm0
131; ALL-NEXT:    retq
132  %out = call <16 x i32> @llvm.cttz.v16i32(<16 x i32> %in, i1 -1)
133  ret <16 x i32> %out
134}
135
136define <32 x i16> @testv32i16(<32 x i16> %in) nounwind {
137; ALL-LABEL: testv32i16:
138; ALL:       ## BB#0:
139; ALL-NEXT:    vpxor %ymm2, %ymm2, %ymm2
140; ALL-NEXT:    vpsubw %ymm0, %ymm2, %ymm3
141; ALL-NEXT:    vpand %ymm3, %ymm0, %ymm0
142; ALL-NEXT:    vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
143; ALL-NEXT:    vpsubw %ymm3, %ymm0, %ymm0
144; ALL-NEXT:    vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
145; ALL-NEXT:    vpand %ymm4, %ymm0, %ymm5
146; ALL-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
147; ALL-NEXT:    vpshufb %ymm5, %ymm6, %ymm5
148; ALL-NEXT:    vpsrlw $4, %ymm0, %ymm0
149; ALL-NEXT:    vpand %ymm4, %ymm0, %ymm0
150; ALL-NEXT:    vpshufb %ymm0, %ymm6, %ymm0
151; ALL-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
152; ALL-NEXT:    vpsllw $8, %ymm0, %ymm5
153; ALL-NEXT:    vpaddb %ymm0, %ymm5, %ymm0
154; ALL-NEXT:    vpsrlw $8, %ymm0, %ymm0
155; ALL-NEXT:    vpsubw %ymm1, %ymm2, %ymm2
156; ALL-NEXT:    vpand %ymm2, %ymm1, %ymm1
157; ALL-NEXT:    vpsubw %ymm3, %ymm1, %ymm1
158; ALL-NEXT:    vpand %ymm4, %ymm1, %ymm2
159; ALL-NEXT:    vpshufb %ymm2, %ymm6, %ymm2
160; ALL-NEXT:    vpsrlw $4, %ymm1, %ymm1
161; ALL-NEXT:    vpand %ymm4, %ymm1, %ymm1
162; ALL-NEXT:    vpshufb %ymm1, %ymm6, %ymm1
163; ALL-NEXT:    vpaddb %ymm2, %ymm1, %ymm1
164; ALL-NEXT:    vpsllw $8, %ymm1, %ymm2
165; ALL-NEXT:    vpaddb %ymm1, %ymm2, %ymm1
166; ALL-NEXT:    vpsrlw $8, %ymm1, %ymm1
167; ALL-NEXT:    retq
168  %out = call <32 x i16> @llvm.cttz.v32i16(<32 x i16> %in, i1 0)
169  ret <32 x i16> %out
170}
171
172define <32 x i16> @testv32i16u(<32 x i16> %in) nounwind {
173; ALL-LABEL: testv32i16u:
174; ALL:       ## BB#0:
175; ALL-NEXT:    vpxor %ymm2, %ymm2, %ymm2
176; ALL-NEXT:    vpsubw %ymm0, %ymm2, %ymm3
177; ALL-NEXT:    vpand %ymm3, %ymm0, %ymm0
178; ALL-NEXT:    vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
179; ALL-NEXT:    vpsubw %ymm3, %ymm0, %ymm0
180; ALL-NEXT:    vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
181; ALL-NEXT:    vpand %ymm4, %ymm0, %ymm5
182; ALL-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
183; ALL-NEXT:    vpshufb %ymm5, %ymm6, %ymm5
184; ALL-NEXT:    vpsrlw $4, %ymm0, %ymm0
185; ALL-NEXT:    vpand %ymm4, %ymm0, %ymm0
186; ALL-NEXT:    vpshufb %ymm0, %ymm6, %ymm0
187; ALL-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
188; ALL-NEXT:    vpsllw $8, %ymm0, %ymm5
189; ALL-NEXT:    vpaddb %ymm0, %ymm5, %ymm0
190; ALL-NEXT:    vpsrlw $8, %ymm0, %ymm0
191; ALL-NEXT:    vpsubw %ymm1, %ymm2, %ymm2
192; ALL-NEXT:    vpand %ymm2, %ymm1, %ymm1
193; ALL-NEXT:    vpsubw %ymm3, %ymm1, %ymm1
194; ALL-NEXT:    vpand %ymm4, %ymm1, %ymm2
195; ALL-NEXT:    vpshufb %ymm2, %ymm6, %ymm2
196; ALL-NEXT:    vpsrlw $4, %ymm1, %ymm1
197; ALL-NEXT:    vpand %ymm4, %ymm1, %ymm1
198; ALL-NEXT:    vpshufb %ymm1, %ymm6, %ymm1
199; ALL-NEXT:    vpaddb %ymm2, %ymm1, %ymm1
200; ALL-NEXT:    vpsllw $8, %ymm1, %ymm2
201; ALL-NEXT:    vpaddb %ymm1, %ymm2, %ymm1
202; ALL-NEXT:    vpsrlw $8, %ymm1, %ymm1
203; ALL-NEXT:    retq
204  %out = call <32 x i16> @llvm.cttz.v32i16(<32 x i16> %in, i1 -1)
205  ret <32 x i16> %out
206}
207
208define <64 x i8> @testv64i8(<64 x i8> %in) nounwind {
209; ALL-LABEL: testv64i8:
210; ALL:       ## BB#0:
211; ALL-NEXT:    vpxor %ymm2, %ymm2, %ymm2
212; ALL-NEXT:    vpsubb %ymm0, %ymm2, %ymm3
213; ALL-NEXT:    vpand %ymm3, %ymm0, %ymm0
214; ALL-NEXT:    vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
215; ALL-NEXT:    vpsubb %ymm3, %ymm0, %ymm0
216; ALL-NEXT:    vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
217; ALL-NEXT:    vpand %ymm4, %ymm0, %ymm5
218; ALL-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
219; ALL-NEXT:    vpshufb %ymm5, %ymm6, %ymm5
220; ALL-NEXT:    vpsrlw $4, %ymm0, %ymm0
221; ALL-NEXT:    vpand %ymm4, %ymm0, %ymm0
222; ALL-NEXT:    vpshufb %ymm0, %ymm6, %ymm0
223; ALL-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
224; ALL-NEXT:    vpsubb %ymm1, %ymm2, %ymm2
225; ALL-NEXT:    vpand %ymm2, %ymm1, %ymm1
226; ALL-NEXT:    vpsubb %ymm3, %ymm1, %ymm1
227; ALL-NEXT:    vpand %ymm4, %ymm1, %ymm2
228; ALL-NEXT:    vpshufb %ymm2, %ymm6, %ymm2
229; ALL-NEXT:    vpsrlw $4, %ymm1, %ymm1
230; ALL-NEXT:    vpand %ymm4, %ymm1, %ymm1
231; ALL-NEXT:    vpshufb %ymm1, %ymm6, %ymm1
232; ALL-NEXT:    vpaddb %ymm2, %ymm1, %ymm1
233; ALL-NEXT:    retq
234  %out = call <64 x i8> @llvm.cttz.v64i8(<64 x i8> %in, i1 0)
235  ret <64 x i8> %out
236}
237
238define <64 x i8> @testv64i8u(<64 x i8> %in) nounwind {
239; ALL-LABEL: testv64i8u:
240; ALL:       ## BB#0:
241; ALL-NEXT:    vpxor %ymm2, %ymm2, %ymm2
242; ALL-NEXT:    vpsubb %ymm0, %ymm2, %ymm3
243; ALL-NEXT:    vpand %ymm3, %ymm0, %ymm0
244; ALL-NEXT:    vmovdqa {{.*#+}} ymm3 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
245; ALL-NEXT:    vpsubb %ymm3, %ymm0, %ymm0
246; ALL-NEXT:    vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
247; ALL-NEXT:    vpand %ymm4, %ymm0, %ymm5
248; ALL-NEXT:    vmovdqa {{.*#+}} ymm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
249; ALL-NEXT:    vpshufb %ymm5, %ymm6, %ymm5
250; ALL-NEXT:    vpsrlw $4, %ymm0, %ymm0
251; ALL-NEXT:    vpand %ymm4, %ymm0, %ymm0
252; ALL-NEXT:    vpshufb %ymm0, %ymm6, %ymm0
253; ALL-NEXT:    vpaddb %ymm5, %ymm0, %ymm0
254; ALL-NEXT:    vpsubb %ymm1, %ymm2, %ymm2
255; ALL-NEXT:    vpand %ymm2, %ymm1, %ymm1
256; ALL-NEXT:    vpsubb %ymm3, %ymm1, %ymm1
257; ALL-NEXT:    vpand %ymm4, %ymm1, %ymm2
258; ALL-NEXT:    vpshufb %ymm2, %ymm6, %ymm2
259; ALL-NEXT:    vpsrlw $4, %ymm1, %ymm1
260; ALL-NEXT:    vpand %ymm4, %ymm1, %ymm1
261; ALL-NEXT:    vpshufb %ymm1, %ymm6, %ymm1
262; ALL-NEXT:    vpaddb %ymm2, %ymm1, %ymm1
263; ALL-NEXT:    retq
264  %out = call <64 x i8> @llvm.cttz.v64i8(<64 x i8> %in, i1 -1)
265  ret <64 x i8> %out
266}
267
268declare <8 x i64> @llvm.cttz.v8i64(<8 x i64>, i1)
269declare <16 x i32> @llvm.cttz.v16i32(<16 x i32>, i1)
270declare <32 x i16> @llvm.cttz.v32i16(<32 x i16>, i1)
271declare <64 x i8> @llvm.cttz.v64i8(<64 x i8>, i1)
272