1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s
3
4declare <2 x i64> @llvm.cttz.v2i64(<2 x i64>, i1)
5declare <2 x i64> @llvm.ctlz.v2i64(<2 x i64>, i1)
6declare <2 x i64> @llvm.ctpop.v2i64(<2 x i64>)
7
8define <2 x i64> @footz(<2 x i64> %a) nounwind {
9; CHECK-LABEL: footz:
10; CHECK:       # BB#0:
11; CHECK-NEXT:    pxor %xmm1, %xmm1
12; CHECK-NEXT:    pxor %xmm2, %xmm2
13; CHECK-NEXT:    psubq %xmm0, %xmm2
14; CHECK-NEXT:    pand %xmm0, %xmm2
15; CHECK-NEXT:    pcmpeqd %xmm3, %xmm3
16; CHECK-NEXT:    paddq %xmm2, %xmm3
17; CHECK-NEXT:    movdqa %xmm3, %xmm0
18; CHECK-NEXT:    psrlq $1, %xmm0
19; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
20; CHECK-NEXT:    psubq %xmm0, %xmm3
21; CHECK-NEXT:    movdqa {{.*#+}} xmm0 = [3689348814741910323,3689348814741910323]
22; CHECK-NEXT:    movdqa %xmm3, %xmm2
23; CHECK-NEXT:    pand %xmm0, %xmm2
24; CHECK-NEXT:    psrlq $2, %xmm3
25; CHECK-NEXT:    pand %xmm0, %xmm3
26; CHECK-NEXT:    paddq %xmm2, %xmm3
27; CHECK-NEXT:    movdqa %xmm3, %xmm0
28; CHECK-NEXT:    psrlq $4, %xmm0
29; CHECK-NEXT:    paddq %xmm3, %xmm0
30; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
31; CHECK-NEXT:    psadbw %xmm1, %xmm0
32; CHECK-NEXT:    retq
33  %c = call <2 x i64> @llvm.cttz.v2i64(<2 x i64> %a, i1 true)
34  ret <2 x i64> %c
35
36}
37define <2 x i64> @foolz(<2 x i64> %a) nounwind {
38; CHECK-LABEL: foolz:
39; CHECK:       # BB#0:
40; CHECK-NEXT:    movdqa %xmm0, %xmm1
41; CHECK-NEXT:    psrlq $1, %xmm1
42; CHECK-NEXT:    por %xmm0, %xmm1
43; CHECK-NEXT:    movdqa %xmm1, %xmm0
44; CHECK-NEXT:    psrlq $2, %xmm0
45; CHECK-NEXT:    por %xmm1, %xmm0
46; CHECK-NEXT:    movdqa %xmm0, %xmm1
47; CHECK-NEXT:    psrlq $4, %xmm1
48; CHECK-NEXT:    por %xmm0, %xmm1
49; CHECK-NEXT:    movdqa %xmm1, %xmm0
50; CHECK-NEXT:    psrlq $8, %xmm0
51; CHECK-NEXT:    por %xmm1, %xmm0
52; CHECK-NEXT:    movdqa %xmm0, %xmm1
53; CHECK-NEXT:    psrlq $16, %xmm1
54; CHECK-NEXT:    por %xmm0, %xmm1
55; CHECK-NEXT:    movdqa %xmm1, %xmm0
56; CHECK-NEXT:    psrlq $32, %xmm0
57; CHECK-NEXT:    por %xmm1, %xmm0
58; CHECK-NEXT:    pcmpeqd %xmm1, %xmm1
59; CHECK-NEXT:    pxor %xmm0, %xmm1
60; CHECK-NEXT:    movdqa %xmm1, %xmm0
61; CHECK-NEXT:    psrlq $1, %xmm0
62; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
63; CHECK-NEXT:    psubq %xmm0, %xmm1
64; CHECK-NEXT:    movdqa {{.*#+}} xmm0 = [3689348814741910323,3689348814741910323]
65; CHECK-NEXT:    movdqa %xmm1, %xmm2
66; CHECK-NEXT:    pand %xmm0, %xmm2
67; CHECK-NEXT:    psrlq $2, %xmm1
68; CHECK-NEXT:    pand %xmm0, %xmm1
69; CHECK-NEXT:    paddq %xmm2, %xmm1
70; CHECK-NEXT:    movdqa %xmm1, %xmm2
71; CHECK-NEXT:    psrlq $4, %xmm2
72; CHECK-NEXT:    paddq %xmm1, %xmm2
73; CHECK-NEXT:    pand {{.*}}(%rip), %xmm2
74; CHECK-NEXT:    pxor %xmm0, %xmm0
75; CHECK-NEXT:    psadbw %xmm2, %xmm0
76; CHECK-NEXT:    retq
77  %c = call <2 x i64> @llvm.ctlz.v2i64(<2 x i64> %a, i1 true)
78  ret <2 x i64> %c
79
80}
81
82define <2 x i64> @foopop(<2 x i64> %a) nounwind {
83; CHECK-LABEL: foopop:
84; CHECK:       # BB#0:
85; CHECK-NEXT:    movdqa %xmm0, %xmm1
86; CHECK-NEXT:    psrlq $1, %xmm1
87; CHECK-NEXT:    pand {{.*}}(%rip), %xmm1
88; CHECK-NEXT:    psubq %xmm1, %xmm0
89; CHECK-NEXT:    movdqa {{.*#+}} xmm1 = [3689348814741910323,3689348814741910323]
90; CHECK-NEXT:    movdqa %xmm0, %xmm2
91; CHECK-NEXT:    pand %xmm1, %xmm2
92; CHECK-NEXT:    psrlq $2, %xmm0
93; CHECK-NEXT:    pand %xmm1, %xmm0
94; CHECK-NEXT:    paddq %xmm2, %xmm0
95; CHECK-NEXT:    movdqa %xmm0, %xmm1
96; CHECK-NEXT:    psrlq $4, %xmm1
97; CHECK-NEXT:    paddq %xmm0, %xmm1
98; CHECK-NEXT:    pand {{.*}}(%rip), %xmm1
99; CHECK-NEXT:    pxor %xmm0, %xmm0
100; CHECK-NEXT:    psadbw %xmm0, %xmm1
101; CHECK-NEXT:    movdqa %xmm1, %xmm0
102; CHECK-NEXT:    retq
103  %c = call <2 x i64> @llvm.ctpop.v2i64(<2 x i64> %a)
104  ret <2 x i64> %c
105}
106
107declare <2 x i32> @llvm.cttz.v2i32(<2 x i32>, i1)
108declare <2 x i32> @llvm.ctlz.v2i32(<2 x i32>, i1)
109declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32>)
110
111define <2 x i32> @promtz(<2 x i32> %a) nounwind {
112; CHECK-LABEL: promtz:
113; CHECK:       # BB#0:
114; CHECK-NEXT:    por {{.*}}(%rip), %xmm0
115; CHECK-NEXT:    pxor %xmm1, %xmm1
116; CHECK-NEXT:    pxor %xmm2, %xmm2
117; CHECK-NEXT:    psubq %xmm0, %xmm2
118; CHECK-NEXT:    pand %xmm0, %xmm2
119; CHECK-NEXT:    pcmpeqd %xmm3, %xmm3
120; CHECK-NEXT:    paddq %xmm2, %xmm3
121; CHECK-NEXT:    movdqa %xmm3, %xmm0
122; CHECK-NEXT:    psrlq $1, %xmm0
123; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
124; CHECK-NEXT:    psubq %xmm0, %xmm3
125; CHECK-NEXT:    movdqa {{.*#+}} xmm0 = [3689348814741910323,3689348814741910323]
126; CHECK-NEXT:    movdqa %xmm3, %xmm2
127; CHECK-NEXT:    pand %xmm0, %xmm2
128; CHECK-NEXT:    psrlq $2, %xmm3
129; CHECK-NEXT:    pand %xmm0, %xmm3
130; CHECK-NEXT:    paddq %xmm2, %xmm3
131; CHECK-NEXT:    movdqa %xmm3, %xmm0
132; CHECK-NEXT:    psrlq $4, %xmm0
133; CHECK-NEXT:    paddq %xmm3, %xmm0
134; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
135; CHECK-NEXT:    psadbw %xmm1, %xmm0
136; CHECK-NEXT:    retq
137  %c = call <2 x i32> @llvm.cttz.v2i32(<2 x i32> %a, i1 false)
138  ret <2 x i32> %c
139
140}
141define <2 x i32> @promlz(<2 x i32> %a) nounwind {
142; CHECK-LABEL: promlz:
143; CHECK:       # BB#0:
144; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
145; CHECK-NEXT:    pxor %xmm1, %xmm1
146; CHECK-NEXT:    movdqa %xmm0, %xmm2
147; CHECK-NEXT:    psrlq $1, %xmm2
148; CHECK-NEXT:    por %xmm0, %xmm2
149; CHECK-NEXT:    movdqa %xmm2, %xmm0
150; CHECK-NEXT:    psrlq $2, %xmm0
151; CHECK-NEXT:    por %xmm2, %xmm0
152; CHECK-NEXT:    movdqa %xmm0, %xmm2
153; CHECK-NEXT:    psrlq $4, %xmm2
154; CHECK-NEXT:    por %xmm0, %xmm2
155; CHECK-NEXT:    movdqa %xmm2, %xmm0
156; CHECK-NEXT:    psrlq $8, %xmm0
157; CHECK-NEXT:    por %xmm2, %xmm0
158; CHECK-NEXT:    movdqa %xmm0, %xmm2
159; CHECK-NEXT:    psrlq $16, %xmm2
160; CHECK-NEXT:    por %xmm0, %xmm2
161; CHECK-NEXT:    movdqa %xmm2, %xmm0
162; CHECK-NEXT:    psrlq $32, %xmm0
163; CHECK-NEXT:    por %xmm2, %xmm0
164; CHECK-NEXT:    pcmpeqd %xmm2, %xmm2
165; CHECK-NEXT:    pxor %xmm0, %xmm2
166; CHECK-NEXT:    movdqa %xmm2, %xmm0
167; CHECK-NEXT:    psrlq $1, %xmm0
168; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
169; CHECK-NEXT:    psubq %xmm0, %xmm2
170; CHECK-NEXT:    movdqa {{.*#+}} xmm0 = [3689348814741910323,3689348814741910323]
171; CHECK-NEXT:    movdqa %xmm2, %xmm3
172; CHECK-NEXT:    pand %xmm0, %xmm3
173; CHECK-NEXT:    psrlq $2, %xmm2
174; CHECK-NEXT:    pand %xmm0, %xmm2
175; CHECK-NEXT:    paddq %xmm3, %xmm2
176; CHECK-NEXT:    movdqa %xmm2, %xmm0
177; CHECK-NEXT:    psrlq $4, %xmm0
178; CHECK-NEXT:    paddq %xmm2, %xmm0
179; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
180; CHECK-NEXT:    psadbw %xmm1, %xmm0
181; CHECK-NEXT:    psubq {{.*}}(%rip), %xmm0
182; CHECK-NEXT:    retq
183  %c = call <2 x i32> @llvm.ctlz.v2i32(<2 x i32> %a, i1 false)
184  ret <2 x i32> %c
185
186}
187
188define <2 x i32> @prompop(<2 x i32> %a) nounwind {
189; CHECK-LABEL: prompop:
190; CHECK:       # BB#0:
191; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
192; CHECK-NEXT:    pxor %xmm2, %xmm2
193; CHECK-NEXT:    movdqa %xmm0, %xmm1
194; CHECK-NEXT:    psrlq $1, %xmm1
195; CHECK-NEXT:    pand {{.*}}(%rip), %xmm1
196; CHECK-NEXT:    psubq %xmm1, %xmm0
197; CHECK-NEXT:    movdqa {{.*#+}} xmm1 = [3689348814741910323,3689348814741910323]
198; CHECK-NEXT:    movdqa %xmm0, %xmm3
199; CHECK-NEXT:    pand %xmm1, %xmm3
200; CHECK-NEXT:    psrlq $2, %xmm0
201; CHECK-NEXT:    pand %xmm1, %xmm0
202; CHECK-NEXT:    paddq %xmm3, %xmm0
203; CHECK-NEXT:    movdqa %xmm0, %xmm1
204; CHECK-NEXT:    psrlq $4, %xmm1
205; CHECK-NEXT:    paddq %xmm0, %xmm1
206; CHECK-NEXT:    pand {{.*}}(%rip), %xmm1
207; CHECK-NEXT:    psadbw %xmm2, %xmm1
208; CHECK-NEXT:    movdqa %xmm1, %xmm0
209; CHECK-NEXT:    retq
210  %c = call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> %a)
211  ret <2 x i32> %c
212}
213