1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=-sse2 | FileCheck %s --check-prefix=X32
3; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X32-SSE2
4; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=X32-AVX2
5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-sse2 | FileCheck %s --check-prefix=X64
6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X64-SSE2
7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=X64-AVX2
8
9define i32 @PR15215_bad(<4 x i32> %input) {
10; X32-LABEL: PR15215_bad:
11; X32:       # %bb.0: # %entry
12; X32-NEXT:    movb {{[0-9]+}}(%esp), %al
13; X32-NEXT:    movb {{[0-9]+}}(%esp), %cl
14; X32-NEXT:    movb {{[0-9]+}}(%esp), %dl
15; X32-NEXT:    movb {{[0-9]+}}(%esp), %ah
16; X32-NEXT:    addb %ah, %ah
17; X32-NEXT:    andb $1, %dl
18; X32-NEXT:    orb %ah, %dl
19; X32-NEXT:    shlb $2, %dl
20; X32-NEXT:    addb %cl, %cl
21; X32-NEXT:    andb $1, %al
22; X32-NEXT:    orb %cl, %al
23; X32-NEXT:    andb $3, %al
24; X32-NEXT:    orb %dl, %al
25; X32-NEXT:    movzbl %al, %eax
26; X32-NEXT:    andl $15, %eax
27; X32-NEXT:    retl
28;
29; X32-SSE2-LABEL: PR15215_bad:
30; X32-SSE2:       # %bb.0: # %entry
31; X32-SSE2-NEXT:    pslld $31, %xmm0
32; X32-SSE2-NEXT:    psrad $31, %xmm0
33; X32-SSE2-NEXT:    movmskps %xmm0, %eax
34; X32-SSE2-NEXT:    retl
35;
36; X32-AVX2-LABEL: PR15215_bad:
37; X32-AVX2:       # %bb.0: # %entry
38; X32-AVX2-NEXT:    vpslld $31, %xmm0, %xmm0
39; X32-AVX2-NEXT:    vpsrad $31, %xmm0, %xmm0
40; X32-AVX2-NEXT:    vmovmskps %xmm0, %eax
41; X32-AVX2-NEXT:    retl
42;
43; X64-LABEL: PR15215_bad:
44; X64:       # %bb.0: # %entry
45; X64-NEXT:    addb %cl, %cl
46; X64-NEXT:    andb $1, %dl
47; X64-NEXT:    orb %cl, %dl
48; X64-NEXT:    shlb $2, %dl
49; X64-NEXT:    addb %sil, %sil
50; X64-NEXT:    andb $1, %dil
51; X64-NEXT:    orb %sil, %dil
52; X64-NEXT:    andb $3, %dil
53; X64-NEXT:    orb %dl, %dil
54; X64-NEXT:    movzbl %dil, %eax
55; X64-NEXT:    andl $15, %eax
56; X64-NEXT:    retq
57;
58; X64-SSE2-LABEL: PR15215_bad:
59; X64-SSE2:       # %bb.0: # %entry
60; X64-SSE2-NEXT:    pslld $31, %xmm0
61; X64-SSE2-NEXT:    psrad $31, %xmm0
62; X64-SSE2-NEXT:    movmskps %xmm0, %eax
63; X64-SSE2-NEXT:    retq
64;
65; X64-AVX2-LABEL: PR15215_bad:
66; X64-AVX2:       # %bb.0: # %entry
67; X64-AVX2-NEXT:    vpslld $31, %xmm0, %xmm0
68; X64-AVX2-NEXT:    vpsrad $31, %xmm0, %xmm0
69; X64-AVX2-NEXT:    vmovmskps %xmm0, %eax
70; X64-AVX2-NEXT:    retq
71entry:
72  %0 = trunc <4 x i32> %input to <4 x i1>
73  %1 = bitcast <4 x i1> %0 to i4
74  %2 = zext i4 %1 to i32
75  ret i32 %2
76}
77
78define i32 @PR15215_good(<4 x i32> %input) {
79; X32-LABEL: PR15215_good:
80; X32:       # %bb.0: # %entry
81; X32-NEXT:    pushl %esi
82; X32-NEXT:    .cfi_def_cfa_offset 8
83; X32-NEXT:    .cfi_offset %esi, -8
84; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
85; X32-NEXT:    andl $1, %eax
86; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
87; X32-NEXT:    andl $1, %ecx
88; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %edx
89; X32-NEXT:    andl $1, %edx
90; X32-NEXT:    movzbl {{[0-9]+}}(%esp), %esi
91; X32-NEXT:    andl $1, %esi
92; X32-NEXT:    leal (%eax,%ecx,2), %eax
93; X32-NEXT:    leal (%eax,%edx,4), %eax
94; X32-NEXT:    leal (%eax,%esi,8), %eax
95; X32-NEXT:    popl %esi
96; X32-NEXT:    retl
97;
98; X32-SSE2-LABEL: PR15215_good:
99; X32-SSE2:       # %bb.0: # %entry
100; X32-SSE2-NEXT:    pushl %esi
101; X32-SSE2-NEXT:    .cfi_def_cfa_offset 8
102; X32-SSE2-NEXT:    .cfi_offset %esi, -8
103; X32-SSE2-NEXT:    movd %xmm0, %eax
104; X32-SSE2-NEXT:    andl $1, %eax
105; X32-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
106; X32-SSE2-NEXT:    movd %xmm1, %ecx
107; X32-SSE2-NEXT:    andl $1, %ecx
108; X32-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
109; X32-SSE2-NEXT:    movd %xmm1, %edx
110; X32-SSE2-NEXT:    andl $1, %edx
111; X32-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
112; X32-SSE2-NEXT:    movd %xmm0, %esi
113; X32-SSE2-NEXT:    andl $1, %esi
114; X32-SSE2-NEXT:    leal (%eax,%ecx,2), %eax
115; X32-SSE2-NEXT:    leal (%eax,%edx,4), %eax
116; X32-SSE2-NEXT:    leal (%eax,%esi,8), %eax
117; X32-SSE2-NEXT:    popl %esi
118; X32-SSE2-NEXT:    retl
119;
120; X32-AVX2-LABEL: PR15215_good:
121; X32-AVX2:       # %bb.0: # %entry
122; X32-AVX2-NEXT:    pushl %esi
123; X32-AVX2-NEXT:    .cfi_def_cfa_offset 8
124; X32-AVX2-NEXT:    .cfi_offset %esi, -8
125; X32-AVX2-NEXT:    vmovd %xmm0, %eax
126; X32-AVX2-NEXT:    andl $1, %eax
127; X32-AVX2-NEXT:    vpextrd $1, %xmm0, %ecx
128; X32-AVX2-NEXT:    andl $1, %ecx
129; X32-AVX2-NEXT:    vpextrd $2, %xmm0, %edx
130; X32-AVX2-NEXT:    andl $1, %edx
131; X32-AVX2-NEXT:    vpextrd $3, %xmm0, %esi
132; X32-AVX2-NEXT:    andl $1, %esi
133; X32-AVX2-NEXT:    leal (%eax,%ecx,2), %eax
134; X32-AVX2-NEXT:    leal (%eax,%edx,4), %eax
135; X32-AVX2-NEXT:    leal (%eax,%esi,8), %eax
136; X32-AVX2-NEXT:    popl %esi
137; X32-AVX2-NEXT:    retl
138;
139; X64-LABEL: PR15215_good:
140; X64:       # %bb.0: # %entry
141; X64-NEXT:    # kill: def $ecx killed $ecx def $rcx
142; X64-NEXT:    # kill: def $edx killed $edx def $rdx
143; X64-NEXT:    # kill: def $esi killed $esi def $rsi
144; X64-NEXT:    # kill: def $edi killed $edi def $rdi
145; X64-NEXT:    andl $1, %edi
146; X64-NEXT:    andl $1, %esi
147; X64-NEXT:    andl $1, %edx
148; X64-NEXT:    andl $1, %ecx
149; X64-NEXT:    leal (%rdi,%rsi,2), %eax
150; X64-NEXT:    leal (%rax,%rdx,4), %eax
151; X64-NEXT:    leal (%rax,%rcx,8), %eax
152; X64-NEXT:    retq
153;
154; X64-SSE2-LABEL: PR15215_good:
155; X64-SSE2:       # %bb.0: # %entry
156; X64-SSE2-NEXT:    movd %xmm0, %eax
157; X64-SSE2-NEXT:    andl $1, %eax
158; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
159; X64-SSE2-NEXT:    movd %xmm1, %ecx
160; X64-SSE2-NEXT:    andl $1, %ecx
161; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
162; X64-SSE2-NEXT:    movd %xmm1, %edx
163; X64-SSE2-NEXT:    andl $1, %edx
164; X64-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[3,1,2,3]
165; X64-SSE2-NEXT:    movd %xmm0, %esi
166; X64-SSE2-NEXT:    andl $1, %esi
167; X64-SSE2-NEXT:    leal (%rax,%rcx,2), %eax
168; X64-SSE2-NEXT:    leal (%rax,%rdx,4), %eax
169; X64-SSE2-NEXT:    leal (%rax,%rsi,8), %eax
170; X64-SSE2-NEXT:    retq
171;
172; X64-AVX2-LABEL: PR15215_good:
173; X64-AVX2:       # %bb.0: # %entry
174; X64-AVX2-NEXT:    vmovd %xmm0, %eax
175; X64-AVX2-NEXT:    andl $1, %eax
176; X64-AVX2-NEXT:    vpextrd $1, %xmm0, %ecx
177; X64-AVX2-NEXT:    andl $1, %ecx
178; X64-AVX2-NEXT:    vpextrd $2, %xmm0, %edx
179; X64-AVX2-NEXT:    andl $1, %edx
180; X64-AVX2-NEXT:    vpextrd $3, %xmm0, %esi
181; X64-AVX2-NEXT:    andl $1, %esi
182; X64-AVX2-NEXT:    leal (%rax,%rcx,2), %eax
183; X64-AVX2-NEXT:    leal (%rax,%rdx,4), %eax
184; X64-AVX2-NEXT:    leal (%rax,%rsi,8), %eax
185; X64-AVX2-NEXT:    retq
186entry:
187  %0 = trunc <4 x i32> %input to <4 x i1>
188  %1 = extractelement <4 x i1> %0, i32 0
189  %e1 = select i1 %1, i32 1, i32 0
190  %2 = extractelement <4 x i1> %0, i32 1
191  %e2 = select i1 %2, i32 2, i32 0
192  %3 = extractelement <4 x i1> %0, i32 2
193  %e3 = select i1 %3, i32 4, i32 0
194  %4 = extractelement <4 x i1> %0, i32 3
195  %e4 = select i1 %4, i32 8, i32 0
196  %5 = or i32 %e1, %e2
197  %6 = or i32 %5, %e3
198  %7 = or i32 %6, %e4
199  ret i32 %7
200}
201