| blob | f424483c53e2ced7a6370b23582f010e47c844c3 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2,+gfni | FileCheck %s --check-prefixes=GFNISSE
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx,+gfni | FileCheck %s --check-prefixes=GFNIAVX,GFNIAVX1OR2,GFNIAVX1
4 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+gfni | FileCheck %s --check-prefixes=GFNIAVX,GFNIAVX1OR2,GFNIAVX2
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+gfni | FileCheck %s --check-prefixes=GFNIAVX,GFNIAVX512,GFNIAVX512VL
6 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw,+avx512vl,+gfni | FileCheck %s --check-prefixes=GFNIAVX,GFNIAVX512,GFNIAVX512BW
8 define <16 x i8> @testv16i8(<16 x i8> %in) nounwind {
9 ; GFNISSE-LABEL: testv16i8:
10 ; GFNISSE: # %bb.0:
11 ; GFNISSE-NEXT: pcmpeqd %xmm1, %xmm1
12 ; GFNISSE-NEXT: paddb %xmm0, %xmm1
13 ; GFNISSE-NEXT: pandn %xmm1, %xmm0
14 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
15 ; GFNISSE-NEXT: pand %xmm0, %xmm2
16 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
17 ; GFNISSE-NEXT: movdqa %xmm1, %xmm3
18 ; GFNISSE-NEXT: pshufb %xmm2, %xmm3
19 ; GFNISSE-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
20 ; GFNISSE-NEXT: pshufb %xmm0, %xmm1
21 ; GFNISSE-NEXT: paddb %xmm3, %xmm1
22 ; GFNISSE-NEXT: movdqa %xmm1, %xmm0
23 ; GFNISSE-NEXT: retq
24 ;
25 ; GFNIAVX1OR2-LABEL: testv16i8:
26 ; GFNIAVX1OR2: # %bb.0:
27 ; GFNIAVX1OR2-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
28 ; GFNIAVX1OR2-NEXT: vpaddb %xmm1, %xmm0, %xmm1
29 ; GFNIAVX1OR2-NEXT: vpandn %xmm1, %xmm0, %xmm0
30 ; GFNIAVX1OR2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
31 ; GFNIAVX1OR2-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
32 ; GFNIAVX1OR2-NEXT: vpshufb %xmm1, %xmm2, %xmm1
33 ; GFNIAVX1OR2-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
34 ; GFNIAVX1OR2-NEXT: vpshufb %xmm0, %xmm2, %xmm0
35 ; GFNIAVX1OR2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
36 ; GFNIAVX1OR2-NEXT: retq
37 ;
38 ; GFNIAVX512-LABEL: testv16i8:
39 ; GFNIAVX512: # %bb.0:
40 ; GFNIAVX512-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
41 ; GFNIAVX512-NEXT: vpaddb %xmm1, %xmm0, %xmm1
42 ; GFNIAVX512-NEXT: vpandn %xmm1, %xmm0, %xmm0
43 ; GFNIAVX512-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm1
44 ; GFNIAVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
45 ; GFNIAVX512-NEXT: vpshufb %xmm1, %xmm2, %xmm1
46 ; GFNIAVX512-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to2}, %xmm0, %xmm0
47 ; GFNIAVX512-NEXT: vpshufb %xmm0, %xmm2, %xmm0
48 ; GFNIAVX512-NEXT: vpaddb %xmm1, %xmm0, %xmm0
49 ; GFNIAVX512-NEXT: retq
50 %out = call <16 x i8> @llvm.cttz.v16i8(<16 x i8> %in, i1 0)
51 ret <16 x i8> %out
52 }
54 define <16 x i8> @testv16i8u(<16 x i8> %in) nounwind {
55 ; GFNISSE-LABEL: testv16i8u:
56 ; GFNISSE: # %bb.0:
57 ; GFNISSE-NEXT: pcmpeqd %xmm1, %xmm1
58 ; GFNISSE-NEXT: paddb %xmm0, %xmm1
59 ; GFNISSE-NEXT: pandn %xmm1, %xmm0
60 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
61 ; GFNISSE-NEXT: pand %xmm0, %xmm2
62 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm1 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
63 ; GFNISSE-NEXT: movdqa %xmm1, %xmm3
64 ; GFNISSE-NEXT: pshufb %xmm2, %xmm3
65 ; GFNISSE-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
66 ; GFNISSE-NEXT: pshufb %xmm0, %xmm1
67 ; GFNISSE-NEXT: paddb %xmm3, %xmm1
68 ; GFNISSE-NEXT: movdqa %xmm1, %xmm0
69 ; GFNISSE-NEXT: retq
70 ;
71 ; GFNIAVX1OR2-LABEL: testv16i8u:
72 ; GFNIAVX1OR2: # %bb.0:
73 ; GFNIAVX1OR2-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
74 ; GFNIAVX1OR2-NEXT: vpaddb %xmm1, %xmm0, %xmm1
75 ; GFNIAVX1OR2-NEXT: vpandn %xmm1, %xmm0, %xmm0
76 ; GFNIAVX1OR2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
77 ; GFNIAVX1OR2-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
78 ; GFNIAVX1OR2-NEXT: vpshufb %xmm1, %xmm2, %xmm1
79 ; GFNIAVX1OR2-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
80 ; GFNIAVX1OR2-NEXT: vpshufb %xmm0, %xmm2, %xmm0
81 ; GFNIAVX1OR2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
82 ; GFNIAVX1OR2-NEXT: retq
83 ;
84 ; GFNIAVX512-LABEL: testv16i8u:
85 ; GFNIAVX512: # %bb.0:
86 ; GFNIAVX512-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
87 ; GFNIAVX512-NEXT: vpaddb %xmm1, %xmm0, %xmm1
88 ; GFNIAVX512-NEXT: vpandn %xmm1, %xmm0, %xmm0
89 ; GFNIAVX512-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm0, %xmm1
90 ; GFNIAVX512-NEXT: vmovdqa {{.*#+}} xmm2 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
91 ; GFNIAVX512-NEXT: vpshufb %xmm1, %xmm2, %xmm1
92 ; GFNIAVX512-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to2}, %xmm0, %xmm0
93 ; GFNIAVX512-NEXT: vpshufb %xmm0, %xmm2, %xmm0
94 ; GFNIAVX512-NEXT: vpaddb %xmm1, %xmm0, %xmm0
95 ; GFNIAVX512-NEXT: retq
96 %out = call <16 x i8> @llvm.cttz.v16i8(<16 x i8> %in, i1 -1)
97 ret <16 x i8> %out
98 }
100 define <32 x i8> @testv32i8(<32 x i8> %in) nounwind {
101 ; GFNISSE-LABEL: testv32i8:
102 ; GFNISSE: # %bb.0:
103 ; GFNISSE-NEXT: pcmpeqd %xmm4, %xmm4
104 ; GFNISSE-NEXT: movdqa %xmm0, %xmm2
105 ; GFNISSE-NEXT: paddb %xmm4, %xmm2
106 ; GFNISSE-NEXT: pandn %xmm2, %xmm0
107 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm5 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
108 ; GFNISSE-NEXT: movdqa %xmm0, %xmm3
109 ; GFNISSE-NEXT: pand %xmm5, %xmm3
110 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm2 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
111 ; GFNISSE-NEXT: movdqa %xmm2, %xmm6
112 ; GFNISSE-NEXT: pshufb %xmm3, %xmm6
113 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm7 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
114 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm7, %xmm0
115 ; GFNISSE-NEXT: movdqa %xmm2, %xmm3
116 ; GFNISSE-NEXT: pshufb %xmm0, %xmm3
117 ; GFNISSE-NEXT: paddb %xmm6, %xmm3
118 ; GFNISSE-NEXT: paddb %xmm1, %xmm4
119 ; GFNISSE-NEXT: pandn %xmm4, %xmm1
120 ; GFNISSE-NEXT: pand %xmm1, %xmm5
121 ; GFNISSE-NEXT: movdqa %xmm2, %xmm0
122 ; GFNISSE-NEXT: pshufb %xmm5, %xmm0
123 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm7, %xmm1
124 ; GFNISSE-NEXT: pshufb %xmm1, %xmm2
125 ; GFNISSE-NEXT: paddb %xmm0, %xmm2
126 ; GFNISSE-NEXT: movdqa %xmm3, %xmm0
127 ; GFNISSE-NEXT: movdqa %xmm2, %xmm1
128 ; GFNISSE-NEXT: retq
129 ;
130 ; GFNIAVX1-LABEL: testv32i8:
131 ; GFNIAVX1: # %bb.0:
132 ; GFNIAVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
133 ; GFNIAVX1-NEXT: vpcmpeqd %xmm2, %xmm2, %xmm2
134 ; GFNIAVX1-NEXT: vpaddb %xmm2, %xmm1, %xmm3
135 ; GFNIAVX1-NEXT: vpandn %xmm3, %xmm1, %xmm1
136 ; GFNIAVX1-NEXT: vbroadcastss {{.*#+}} xmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
137 ; GFNIAVX1-NEXT: vpand %xmm3, %xmm1, %xmm4
138 ; GFNIAVX1-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
139 ; GFNIAVX1-NEXT: vpshufb %xmm4, %xmm5, %xmm4
140 ; GFNIAVX1-NEXT: vmovddup {{.*#+}} xmm6 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
141 ; GFNIAVX1-NEXT: # xmm6 = mem[0,0]
142 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm6, %xmm1, %xmm1
143 ; GFNIAVX1-NEXT: vpshufb %xmm1, %xmm5, %xmm1
144 ; GFNIAVX1-NEXT: vpaddb %xmm4, %xmm1, %xmm1
145 ; GFNIAVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm2
146 ; GFNIAVX1-NEXT: vpandn %xmm2, %xmm0, %xmm0
147 ; GFNIAVX1-NEXT: vpand %xmm3, %xmm0, %xmm2
148 ; GFNIAVX1-NEXT: vpshufb %xmm2, %xmm5, %xmm2
149 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm6, %xmm0, %xmm0
150 ; GFNIAVX1-NEXT: vpshufb %xmm0, %xmm5, %xmm0
151 ; GFNIAVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
152 ; GFNIAVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
153 ; GFNIAVX1-NEXT: retq
154 ;
155 ; GFNIAVX2-LABEL: testv32i8:
156 ; GFNIAVX2: # %bb.0:
157 ; GFNIAVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
158 ; GFNIAVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm1
159 ; GFNIAVX2-NEXT: vpandn %ymm1, %ymm0, %ymm0
160 ; GFNIAVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
161 ; GFNIAVX2-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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]
162 ; GFNIAVX2-NEXT: # ymm2 = mem[0,1,0,1]
163 ; GFNIAVX2-NEXT: vpshufb %ymm1, %ymm2, %ymm1
164 ; GFNIAVX2-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
165 ; GFNIAVX2-NEXT: vpshufb %ymm0, %ymm2, %ymm0
166 ; GFNIAVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
167 ; GFNIAVX2-NEXT: retq
168 ;
169 ; GFNIAVX512-LABEL: testv32i8:
170 ; GFNIAVX512: # %bb.0:
171 ; GFNIAVX512-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
172 ; GFNIAVX512-NEXT: vpaddb %ymm1, %ymm0, %ymm1
173 ; GFNIAVX512-NEXT: vpandn %ymm1, %ymm0, %ymm0
174 ; GFNIAVX512-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %ymm0, %ymm1
175 ; GFNIAVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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]
176 ; GFNIAVX512-NEXT: # ymm2 = mem[0,1,0,1]
177 ; GFNIAVX512-NEXT: vpshufb %ymm1, %ymm2, %ymm1
178 ; GFNIAVX512-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %ymm0, %ymm0
179 ; GFNIAVX512-NEXT: vpshufb %ymm0, %ymm2, %ymm0
180 ; GFNIAVX512-NEXT: vpaddb %ymm1, %ymm0, %ymm0
181 ; GFNIAVX512-NEXT: retq
182 %out = call <32 x i8> @llvm.cttz.v32i8(<32 x i8> %in, i1 0)
183 ret <32 x i8> %out
184 }
186 define <32 x i8> @testv32i8u(<32 x i8> %in) nounwind {
187 ; GFNISSE-LABEL: testv32i8u:
188 ; GFNISSE: # %bb.0:
189 ; GFNISSE-NEXT: pcmpeqd %xmm4, %xmm4
190 ; GFNISSE-NEXT: movdqa %xmm0, %xmm2
191 ; GFNISSE-NEXT: paddb %xmm4, %xmm2
192 ; GFNISSE-NEXT: pandn %xmm2, %xmm0
193 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm5 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
194 ; GFNISSE-NEXT: movdqa %xmm0, %xmm3
195 ; GFNISSE-NEXT: pand %xmm5, %xmm3
196 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm2 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
197 ; GFNISSE-NEXT: movdqa %xmm2, %xmm6
198 ; GFNISSE-NEXT: pshufb %xmm3, %xmm6
199 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm7 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
200 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm7, %xmm0
201 ; GFNISSE-NEXT: movdqa %xmm2, %xmm3
202 ; GFNISSE-NEXT: pshufb %xmm0, %xmm3
203 ; GFNISSE-NEXT: paddb %xmm6, %xmm3
204 ; GFNISSE-NEXT: paddb %xmm1, %xmm4
205 ; GFNISSE-NEXT: pandn %xmm4, %xmm1
206 ; GFNISSE-NEXT: pand %xmm1, %xmm5
207 ; GFNISSE-NEXT: movdqa %xmm2, %xmm0
208 ; GFNISSE-NEXT: pshufb %xmm5, %xmm0
209 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm7, %xmm1
210 ; GFNISSE-NEXT: pshufb %xmm1, %xmm2
211 ; GFNISSE-NEXT: paddb %xmm0, %xmm2
212 ; GFNISSE-NEXT: movdqa %xmm3, %xmm0
213 ; GFNISSE-NEXT: movdqa %xmm2, %xmm1
214 ; GFNISSE-NEXT: retq
215 ;
216 ; GFNIAVX1-LABEL: testv32i8u:
217 ; GFNIAVX1: # %bb.0:
218 ; GFNIAVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
219 ; GFNIAVX1-NEXT: vpcmpeqd %xmm2, %xmm2, %xmm2
220 ; GFNIAVX1-NEXT: vpaddb %xmm2, %xmm1, %xmm3
221 ; GFNIAVX1-NEXT: vpandn %xmm3, %xmm1, %xmm1
222 ; GFNIAVX1-NEXT: vbroadcastss {{.*#+}} xmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
223 ; GFNIAVX1-NEXT: vpand %xmm3, %xmm1, %xmm4
224 ; GFNIAVX1-NEXT: vmovdqa {{.*#+}} xmm5 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
225 ; GFNIAVX1-NEXT: vpshufb %xmm4, %xmm5, %xmm4
226 ; GFNIAVX1-NEXT: vmovddup {{.*#+}} xmm6 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
227 ; GFNIAVX1-NEXT: # xmm6 = mem[0,0]
228 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm6, %xmm1, %xmm1
229 ; GFNIAVX1-NEXT: vpshufb %xmm1, %xmm5, %xmm1
230 ; GFNIAVX1-NEXT: vpaddb %xmm4, %xmm1, %xmm1
231 ; GFNIAVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm2
232 ; GFNIAVX1-NEXT: vpandn %xmm2, %xmm0, %xmm0
233 ; GFNIAVX1-NEXT: vpand %xmm3, %xmm0, %xmm2
234 ; GFNIAVX1-NEXT: vpshufb %xmm2, %xmm5, %xmm2
235 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm6, %xmm0, %xmm0
236 ; GFNIAVX1-NEXT: vpshufb %xmm0, %xmm5, %xmm0
237 ; GFNIAVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
238 ; GFNIAVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
239 ; GFNIAVX1-NEXT: retq
240 ;
241 ; GFNIAVX2-LABEL: testv32i8u:
242 ; GFNIAVX2: # %bb.0:
243 ; GFNIAVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
244 ; GFNIAVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm1
245 ; GFNIAVX2-NEXT: vpandn %ymm1, %ymm0, %ymm0
246 ; GFNIAVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1
247 ; GFNIAVX2-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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]
248 ; GFNIAVX2-NEXT: # ymm2 = mem[0,1,0,1]
249 ; GFNIAVX2-NEXT: vpshufb %ymm1, %ymm2, %ymm1
250 ; GFNIAVX2-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
251 ; GFNIAVX2-NEXT: vpshufb %ymm0, %ymm2, %ymm0
252 ; GFNIAVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
253 ; GFNIAVX2-NEXT: retq
254 ;
255 ; GFNIAVX512-LABEL: testv32i8u:
256 ; GFNIAVX512: # %bb.0:
257 ; GFNIAVX512-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
258 ; GFNIAVX512-NEXT: vpaddb %ymm1, %ymm0, %ymm1
259 ; GFNIAVX512-NEXT: vpandn %ymm1, %ymm0, %ymm0
260 ; GFNIAVX512-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %ymm0, %ymm1
261 ; GFNIAVX512-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [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]
262 ; GFNIAVX512-NEXT: # ymm2 = mem[0,1,0,1]
263 ; GFNIAVX512-NEXT: vpshufb %ymm1, %ymm2, %ymm1
264 ; GFNIAVX512-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %ymm0, %ymm0
265 ; GFNIAVX512-NEXT: vpshufb %ymm0, %ymm2, %ymm0
266 ; GFNIAVX512-NEXT: vpaddb %ymm1, %ymm0, %ymm0
267 ; GFNIAVX512-NEXT: retq
268 %out = call <32 x i8> @llvm.cttz.v32i8(<32 x i8> %in, i1 -1)
269 ret <32 x i8> %out
270 }
272 define <64 x i8> @testv64i8(<64 x i8> %in) nounwind {
273 ; GFNISSE-LABEL: testv64i8:
274 ; GFNISSE: # %bb.0:
275 ; GFNISSE-NEXT: movdqa %xmm1, %xmm5
276 ; GFNISSE-NEXT: movdqa %xmm0, %xmm1
277 ; GFNISSE-NEXT: pcmpeqd %xmm6, %xmm6
278 ; GFNISSE-NEXT: paddb %xmm6, %xmm0
279 ; GFNISSE-NEXT: pandn %xmm0, %xmm1
280 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm7 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
281 ; GFNISSE-NEXT: movdqa %xmm1, %xmm0
282 ; GFNISSE-NEXT: pand %xmm7, %xmm0
283 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
284 ; GFNISSE-NEXT: movdqa %xmm4, %xmm9
285 ; GFNISSE-NEXT: pshufb %xmm0, %xmm9
286 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm8 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
287 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm8, %xmm1
288 ; GFNISSE-NEXT: movdqa %xmm4, %xmm0
289 ; GFNISSE-NEXT: pshufb %xmm1, %xmm0
290 ; GFNISSE-NEXT: paddb %xmm9, %xmm0
291 ; GFNISSE-NEXT: movdqa %xmm5, %xmm1
292 ; GFNISSE-NEXT: paddb %xmm6, %xmm1
293 ; GFNISSE-NEXT: pandn %xmm1, %xmm5
294 ; GFNISSE-NEXT: movdqa %xmm5, %xmm1
295 ; GFNISSE-NEXT: pand %xmm7, %xmm1
296 ; GFNISSE-NEXT: movdqa %xmm4, %xmm9
297 ; GFNISSE-NEXT: pshufb %xmm1, %xmm9
298 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm8, %xmm5
299 ; GFNISSE-NEXT: movdqa %xmm4, %xmm1
300 ; GFNISSE-NEXT: pshufb %xmm5, %xmm1
301 ; GFNISSE-NEXT: paddb %xmm9, %xmm1
302 ; GFNISSE-NEXT: movdqa %xmm2, %xmm5
303 ; GFNISSE-NEXT: paddb %xmm6, %xmm5
304 ; GFNISSE-NEXT: pandn %xmm5, %xmm2
305 ; GFNISSE-NEXT: movdqa %xmm2, %xmm5
306 ; GFNISSE-NEXT: pand %xmm7, %xmm5
307 ; GFNISSE-NEXT: movdqa %xmm4, %xmm9
308 ; GFNISSE-NEXT: pshufb %xmm5, %xmm9
309 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm8, %xmm2
310 ; GFNISSE-NEXT: movdqa %xmm4, %xmm5
311 ; GFNISSE-NEXT: pshufb %xmm2, %xmm5
312 ; GFNISSE-NEXT: paddb %xmm9, %xmm5
313 ; GFNISSE-NEXT: paddb %xmm3, %xmm6
314 ; GFNISSE-NEXT: pandn %xmm6, %xmm3
315 ; GFNISSE-NEXT: pand %xmm3, %xmm7
316 ; GFNISSE-NEXT: movdqa %xmm4, %xmm2
317 ; GFNISSE-NEXT: pshufb %xmm7, %xmm2
318 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm8, %xmm3
319 ; GFNISSE-NEXT: pshufb %xmm3, %xmm4
320 ; GFNISSE-NEXT: paddb %xmm2, %xmm4
321 ; GFNISSE-NEXT: movdqa %xmm5, %xmm2
322 ; GFNISSE-NEXT: movdqa %xmm4, %xmm3
323 ; GFNISSE-NEXT: retq
324 ;
325 ; GFNIAVX1-LABEL: testv64i8:
326 ; GFNIAVX1: # %bb.0:
327 ; GFNIAVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
328 ; GFNIAVX1-NEXT: vpcmpeqd %xmm3, %xmm3, %xmm3
329 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm2, %xmm4
330 ; GFNIAVX1-NEXT: vpandn %xmm4, %xmm2, %xmm2
331 ; GFNIAVX1-NEXT: vbroadcastss {{.*#+}} xmm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
332 ; GFNIAVX1-NEXT: vpand %xmm4, %xmm2, %xmm5
333 ; GFNIAVX1-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
334 ; GFNIAVX1-NEXT: vpshufb %xmm5, %xmm6, %xmm5
335 ; GFNIAVX1-NEXT: vmovddup {{.*#+}} xmm7 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
336 ; GFNIAVX1-NEXT: # xmm7 = mem[0,0]
337 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm7, %xmm2, %xmm2
338 ; GFNIAVX1-NEXT: vpshufb %xmm2, %xmm6, %xmm2
339 ; GFNIAVX1-NEXT: vpaddb %xmm5, %xmm2, %xmm2
340 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm0, %xmm5
341 ; GFNIAVX1-NEXT: vpandn %xmm5, %xmm0, %xmm0
342 ; GFNIAVX1-NEXT: vpand %xmm4, %xmm0, %xmm5
343 ; GFNIAVX1-NEXT: vpshufb %xmm5, %xmm6, %xmm5
344 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm7, %xmm0, %xmm0
345 ; GFNIAVX1-NEXT: vpshufb %xmm0, %xmm6, %xmm0
346 ; GFNIAVX1-NEXT: vpaddb %xmm5, %xmm0, %xmm0
347 ; GFNIAVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
348 ; GFNIAVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
349 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm2, %xmm5
350 ; GFNIAVX1-NEXT: vpandn %xmm5, %xmm2, %xmm2
351 ; GFNIAVX1-NEXT: vpand %xmm4, %xmm2, %xmm5
352 ; GFNIAVX1-NEXT: vpshufb %xmm5, %xmm6, %xmm5
353 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm7, %xmm2, %xmm2
354 ; GFNIAVX1-NEXT: vpshufb %xmm2, %xmm6, %xmm2
355 ; GFNIAVX1-NEXT: vpaddb %xmm5, %xmm2, %xmm2
356 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm3
357 ; GFNIAVX1-NEXT: vpandn %xmm3, %xmm1, %xmm1
358 ; GFNIAVX1-NEXT: vpand %xmm4, %xmm1, %xmm3
359 ; GFNIAVX1-NEXT: vpshufb %xmm3, %xmm6, %xmm3
360 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm7, %xmm1, %xmm1
361 ; GFNIAVX1-NEXT: vpshufb %xmm1, %xmm6, %xmm1
362 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
363 ; GFNIAVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1
364 ; GFNIAVX1-NEXT: retq
365 ;
366 ; GFNIAVX2-LABEL: testv64i8:
367 ; GFNIAVX2: # %bb.0:
368 ; GFNIAVX2-NEXT: vpcmpeqd %ymm2, %ymm2, %ymm2
369 ; GFNIAVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm3
370 ; GFNIAVX2-NEXT: vpandn %ymm3, %ymm0, %ymm0
371 ; GFNIAVX2-NEXT: vpbroadcastb {{.*#+}} ymm3 = [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]
372 ; GFNIAVX2-NEXT: vpand %ymm3, %ymm0, %ymm4
373 ; GFNIAVX2-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [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]
374 ; GFNIAVX2-NEXT: # ymm5 = mem[0,1,0,1]
375 ; GFNIAVX2-NEXT: vpshufb %ymm4, %ymm5, %ymm4
376 ; GFNIAVX2-NEXT: vpbroadcastq {{.*#+}} ymm6 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
377 ; GFNIAVX2-NEXT: vgf2p8affineqb $0, %ymm6, %ymm0, %ymm0
378 ; GFNIAVX2-NEXT: vpshufb %ymm0, %ymm5, %ymm0
379 ; GFNIAVX2-NEXT: vpaddb %ymm4, %ymm0, %ymm0
380 ; GFNIAVX2-NEXT: vpaddb %ymm2, %ymm1, %ymm2
381 ; GFNIAVX2-NEXT: vpandn %ymm2, %ymm1, %ymm1
382 ; GFNIAVX2-NEXT: vpand %ymm3, %ymm1, %ymm2
383 ; GFNIAVX2-NEXT: vpshufb %ymm2, %ymm5, %ymm2
384 ; GFNIAVX2-NEXT: vgf2p8affineqb $0, %ymm6, %ymm1, %ymm1
385 ; GFNIAVX2-NEXT: vpshufb %ymm1, %ymm5, %ymm1
386 ; GFNIAVX2-NEXT: vpaddb %ymm2, %ymm1, %ymm1
387 ; GFNIAVX2-NEXT: retq
388 ;
389 ; GFNIAVX512VL-LABEL: testv64i8:
390 ; GFNIAVX512VL: # %bb.0:
391 ; GFNIAVX512VL-NEXT: vextracti64x4 $1, %zmm0, %ymm1
392 ; GFNIAVX512VL-NEXT: vpcmpeqd %ymm2, %ymm2, %ymm2
393 ; GFNIAVX512VL-NEXT: vpaddb %ymm2, %ymm1, %ymm3
394 ; GFNIAVX512VL-NEXT: vpandn %ymm3, %ymm1, %ymm1
395 ; GFNIAVX512VL-NEXT: vpbroadcastd {{.*#+}} ymm3 = [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]
396 ; GFNIAVX512VL-NEXT: vpand %ymm3, %ymm1, %ymm4
397 ; GFNIAVX512VL-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [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]
398 ; GFNIAVX512VL-NEXT: # ymm5 = mem[0,1,0,1]
399 ; GFNIAVX512VL-NEXT: vpshufb %ymm4, %ymm5, %ymm4
400 ; GFNIAVX512VL-NEXT: vpbroadcastq {{.*#+}} ymm6 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
401 ; GFNIAVX512VL-NEXT: vgf2p8affineqb $0, %ymm6, %ymm1, %ymm1
402 ; GFNIAVX512VL-NEXT: vpshufb %ymm1, %ymm5, %ymm1
403 ; GFNIAVX512VL-NEXT: vpaddb %ymm4, %ymm1, %ymm1
404 ; GFNIAVX512VL-NEXT: vpaddb %ymm2, %ymm0, %ymm2
405 ; GFNIAVX512VL-NEXT: vpandn %ymm2, %ymm0, %ymm0
406 ; GFNIAVX512VL-NEXT: vpand %ymm3, %ymm0, %ymm2
407 ; GFNIAVX512VL-NEXT: vpshufb %ymm2, %ymm5, %ymm2
408 ; GFNIAVX512VL-NEXT: vgf2p8affineqb $0, %ymm6, %ymm0, %ymm0
409 ; GFNIAVX512VL-NEXT: vpshufb %ymm0, %ymm5, %ymm0
410 ; GFNIAVX512VL-NEXT: vpaddb %ymm2, %ymm0, %ymm0
411 ; GFNIAVX512VL-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
412 ; GFNIAVX512VL-NEXT: retq
413 ;
414 ; GFNIAVX512BW-LABEL: testv64i8:
415 ; GFNIAVX512BW: # %bb.0:
416 ; GFNIAVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1
417 ; GFNIAVX512BW-NEXT: vpaddb %zmm1, %zmm0, %zmm1
418 ; GFNIAVX512BW-NEXT: vpandnq %zmm1, %zmm0, %zmm0
419 ; GFNIAVX512BW-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %zmm1
420 ; GFNIAVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
421 ; GFNIAVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
422 ; GFNIAVX512BW-NEXT: vpshufb %zmm1, %zmm2, %zmm1
423 ; GFNIAVX512BW-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %zmm0, %zmm0
424 ; GFNIAVX512BW-NEXT: vpshufb %zmm0, %zmm2, %zmm0
425 ; GFNIAVX512BW-NEXT: vpaddb %zmm1, %zmm0, %zmm0
426 ; GFNIAVX512BW-NEXT: retq
427 %out = call <64 x i8> @llvm.cttz.v64i8(<64 x i8> %in, i1 0)
428 ret <64 x i8> %out
429 }
431 define <64 x i8> @testv64i8u(<64 x i8> %in) nounwind {
432 ; GFNISSE-LABEL: testv64i8u:
433 ; GFNISSE: # %bb.0:
434 ; GFNISSE-NEXT: movdqa %xmm1, %xmm5
435 ; GFNISSE-NEXT: movdqa %xmm0, %xmm1
436 ; GFNISSE-NEXT: pcmpeqd %xmm6, %xmm6
437 ; GFNISSE-NEXT: paddb %xmm6, %xmm0
438 ; GFNISSE-NEXT: pandn %xmm0, %xmm1
439 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm7 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
440 ; GFNISSE-NEXT: movdqa %xmm1, %xmm0
441 ; GFNISSE-NEXT: pand %xmm7, %xmm0
442 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
443 ; GFNISSE-NEXT: movdqa %xmm4, %xmm9
444 ; GFNISSE-NEXT: pshufb %xmm0, %xmm9
445 ; GFNISSE-NEXT: movdqa {{.*#+}} xmm8 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
446 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm8, %xmm1
447 ; GFNISSE-NEXT: movdqa %xmm4, %xmm0
448 ; GFNISSE-NEXT: pshufb %xmm1, %xmm0
449 ; GFNISSE-NEXT: paddb %xmm9, %xmm0
450 ; GFNISSE-NEXT: movdqa %xmm5, %xmm1
451 ; GFNISSE-NEXT: paddb %xmm6, %xmm1
452 ; GFNISSE-NEXT: pandn %xmm1, %xmm5
453 ; GFNISSE-NEXT: movdqa %xmm5, %xmm1
454 ; GFNISSE-NEXT: pand %xmm7, %xmm1
455 ; GFNISSE-NEXT: movdqa %xmm4, %xmm9
456 ; GFNISSE-NEXT: pshufb %xmm1, %xmm9
457 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm8, %xmm5
458 ; GFNISSE-NEXT: movdqa %xmm4, %xmm1
459 ; GFNISSE-NEXT: pshufb %xmm5, %xmm1
460 ; GFNISSE-NEXT: paddb %xmm9, %xmm1
461 ; GFNISSE-NEXT: movdqa %xmm2, %xmm5
462 ; GFNISSE-NEXT: paddb %xmm6, %xmm5
463 ; GFNISSE-NEXT: pandn %xmm5, %xmm2
464 ; GFNISSE-NEXT: movdqa %xmm2, %xmm5
465 ; GFNISSE-NEXT: pand %xmm7, %xmm5
466 ; GFNISSE-NEXT: movdqa %xmm4, %xmm9
467 ; GFNISSE-NEXT: pshufb %xmm5, %xmm9
468 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm8, %xmm2
469 ; GFNISSE-NEXT: movdqa %xmm4, %xmm5
470 ; GFNISSE-NEXT: pshufb %xmm2, %xmm5
471 ; GFNISSE-NEXT: paddb %xmm9, %xmm5
472 ; GFNISSE-NEXT: paddb %xmm3, %xmm6
473 ; GFNISSE-NEXT: pandn %xmm6, %xmm3
474 ; GFNISSE-NEXT: pand %xmm3, %xmm7
475 ; GFNISSE-NEXT: movdqa %xmm4, %xmm2
476 ; GFNISSE-NEXT: pshufb %xmm7, %xmm2
477 ; GFNISSE-NEXT: gf2p8affineqb $0, %xmm8, %xmm3
478 ; GFNISSE-NEXT: pshufb %xmm3, %xmm4
479 ; GFNISSE-NEXT: paddb %xmm2, %xmm4
480 ; GFNISSE-NEXT: movdqa %xmm5, %xmm2
481 ; GFNISSE-NEXT: movdqa %xmm4, %xmm3
482 ; GFNISSE-NEXT: retq
483 ;
484 ; GFNIAVX1-LABEL: testv64i8u:
485 ; GFNIAVX1: # %bb.0:
486 ; GFNIAVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
487 ; GFNIAVX1-NEXT: vpcmpeqd %xmm3, %xmm3, %xmm3
488 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm2, %xmm4
489 ; GFNIAVX1-NEXT: vpandn %xmm4, %xmm2, %xmm2
490 ; GFNIAVX1-NEXT: vbroadcastss {{.*#+}} xmm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
491 ; GFNIAVX1-NEXT: vpand %xmm4, %xmm2, %xmm5
492 ; GFNIAVX1-NEXT: vmovdqa {{.*#+}} xmm6 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
493 ; GFNIAVX1-NEXT: vpshufb %xmm5, %xmm6, %xmm5
494 ; GFNIAVX1-NEXT: vmovddup {{.*#+}} xmm7 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
495 ; GFNIAVX1-NEXT: # xmm7 = mem[0,0]
496 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm7, %xmm2, %xmm2
497 ; GFNIAVX1-NEXT: vpshufb %xmm2, %xmm6, %xmm2
498 ; GFNIAVX1-NEXT: vpaddb %xmm5, %xmm2, %xmm2
499 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm0, %xmm5
500 ; GFNIAVX1-NEXT: vpandn %xmm5, %xmm0, %xmm0
501 ; GFNIAVX1-NEXT: vpand %xmm4, %xmm0, %xmm5
502 ; GFNIAVX1-NEXT: vpshufb %xmm5, %xmm6, %xmm5
503 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm7, %xmm0, %xmm0
504 ; GFNIAVX1-NEXT: vpshufb %xmm0, %xmm6, %xmm0
505 ; GFNIAVX1-NEXT: vpaddb %xmm5, %xmm0, %xmm0
506 ; GFNIAVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
507 ; GFNIAVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
508 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm2, %xmm5
509 ; GFNIAVX1-NEXT: vpandn %xmm5, %xmm2, %xmm2
510 ; GFNIAVX1-NEXT: vpand %xmm4, %xmm2, %xmm5
511 ; GFNIAVX1-NEXT: vpshufb %xmm5, %xmm6, %xmm5
512 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm7, %xmm2, %xmm2
513 ; GFNIAVX1-NEXT: vpshufb %xmm2, %xmm6, %xmm2
514 ; GFNIAVX1-NEXT: vpaddb %xmm5, %xmm2, %xmm2
515 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm3
516 ; GFNIAVX1-NEXT: vpandn %xmm3, %xmm1, %xmm1
517 ; GFNIAVX1-NEXT: vpand %xmm4, %xmm1, %xmm3
518 ; GFNIAVX1-NEXT: vpshufb %xmm3, %xmm6, %xmm3
519 ; GFNIAVX1-NEXT: vgf2p8affineqb $0, %xmm7, %xmm1, %xmm1
520 ; GFNIAVX1-NEXT: vpshufb %xmm1, %xmm6, %xmm1
521 ; GFNIAVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
522 ; GFNIAVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1
523 ; GFNIAVX1-NEXT: retq
524 ;
525 ; GFNIAVX2-LABEL: testv64i8u:
526 ; GFNIAVX2: # %bb.0:
527 ; GFNIAVX2-NEXT: vpcmpeqd %ymm2, %ymm2, %ymm2
528 ; GFNIAVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm3
529 ; GFNIAVX2-NEXT: vpandn %ymm3, %ymm0, %ymm0
530 ; GFNIAVX2-NEXT: vpbroadcastb {{.*#+}} ymm3 = [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]
531 ; GFNIAVX2-NEXT: vpand %ymm3, %ymm0, %ymm4
532 ; GFNIAVX2-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [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]
533 ; GFNIAVX2-NEXT: # ymm5 = mem[0,1,0,1]
534 ; GFNIAVX2-NEXT: vpshufb %ymm4, %ymm5, %ymm4
535 ; GFNIAVX2-NEXT: vpbroadcastq {{.*#+}} ymm6 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
536 ; GFNIAVX2-NEXT: vgf2p8affineqb $0, %ymm6, %ymm0, %ymm0
537 ; GFNIAVX2-NEXT: vpshufb %ymm0, %ymm5, %ymm0
538 ; GFNIAVX2-NEXT: vpaddb %ymm4, %ymm0, %ymm0
539 ; GFNIAVX2-NEXT: vpaddb %ymm2, %ymm1, %ymm2
540 ; GFNIAVX2-NEXT: vpandn %ymm2, %ymm1, %ymm1
541 ; GFNIAVX2-NEXT: vpand %ymm3, %ymm1, %ymm2
542 ; GFNIAVX2-NEXT: vpshufb %ymm2, %ymm5, %ymm2
543 ; GFNIAVX2-NEXT: vgf2p8affineqb $0, %ymm6, %ymm1, %ymm1
544 ; GFNIAVX2-NEXT: vpshufb %ymm1, %ymm5, %ymm1
545 ; GFNIAVX2-NEXT: vpaddb %ymm2, %ymm1, %ymm1
546 ; GFNIAVX2-NEXT: retq
547 ;
548 ; GFNIAVX512VL-LABEL: testv64i8u:
549 ; GFNIAVX512VL: # %bb.0:
550 ; GFNIAVX512VL-NEXT: vextracti64x4 $1, %zmm0, %ymm1
551 ; GFNIAVX512VL-NEXT: vpcmpeqd %ymm2, %ymm2, %ymm2
552 ; GFNIAVX512VL-NEXT: vpaddb %ymm2, %ymm1, %ymm3
553 ; GFNIAVX512VL-NEXT: vpandn %ymm3, %ymm1, %ymm1
554 ; GFNIAVX512VL-NEXT: vpbroadcastd {{.*#+}} ymm3 = [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]
555 ; GFNIAVX512VL-NEXT: vpand %ymm3, %ymm1, %ymm4
556 ; GFNIAVX512VL-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [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]
557 ; GFNIAVX512VL-NEXT: # ymm5 = mem[0,1,0,1]
558 ; GFNIAVX512VL-NEXT: vpshufb %ymm4, %ymm5, %ymm4
559 ; GFNIAVX512VL-NEXT: vpbroadcastq {{.*#+}} ymm6 = [0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16,0,0,0,0,128,64,32,16]
560 ; GFNIAVX512VL-NEXT: vgf2p8affineqb $0, %ymm6, %ymm1, %ymm1
561 ; GFNIAVX512VL-NEXT: vpshufb %ymm1, %ymm5, %ymm1
562 ; GFNIAVX512VL-NEXT: vpaddb %ymm4, %ymm1, %ymm1
563 ; GFNIAVX512VL-NEXT: vpaddb %ymm2, %ymm0, %ymm2
564 ; GFNIAVX512VL-NEXT: vpandn %ymm2, %ymm0, %ymm0
565 ; GFNIAVX512VL-NEXT: vpand %ymm3, %ymm0, %ymm2
566 ; GFNIAVX512VL-NEXT: vpshufb %ymm2, %ymm5, %ymm2
567 ; GFNIAVX512VL-NEXT: vgf2p8affineqb $0, %ymm6, %ymm0, %ymm0
568 ; GFNIAVX512VL-NEXT: vpshufb %ymm0, %ymm5, %ymm0
569 ; GFNIAVX512VL-NEXT: vpaddb %ymm2, %ymm0, %ymm0
570 ; GFNIAVX512VL-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
571 ; GFNIAVX512VL-NEXT: retq
572 ;
573 ; GFNIAVX512BW-LABEL: testv64i8u:
574 ; GFNIAVX512BW: # %bb.0:
575 ; GFNIAVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1
576 ; GFNIAVX512BW-NEXT: vpaddb %zmm1, %zmm0, %zmm1
577 ; GFNIAVX512BW-NEXT: vpandnq %zmm1, %zmm0, %zmm0
578 ; GFNIAVX512BW-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %zmm1
579 ; GFNIAVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
580 ; GFNIAVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
581 ; GFNIAVX512BW-NEXT: vpshufb %zmm1, %zmm2, %zmm1
582 ; GFNIAVX512BW-NEXT: vgf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %zmm0, %zmm0
583 ; GFNIAVX512BW-NEXT: vpshufb %zmm0, %zmm2, %zmm0
584 ; GFNIAVX512BW-NEXT: vpaddb %zmm1, %zmm0, %zmm0
585 ; GFNIAVX512BW-NEXT: retq
586 %out = call <64 x i8> @llvm.cttz.v64i8(<64 x i8> %in, i1 -1)
587 ret <64 x i8> %out
588 }
589 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
590 ; GFNIAVX: {{.*}}