xref: /llvm-project/llvm/test/CodeGen/X86/avgceils-scalar.ll (revision 92715cf43b18d497cd034bdc7787b3a8eeb2edc5) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-- | FileCheck %s --check-prefixes=X86
3; RUN: llc < %s -mtriple=x86_64-- | FileCheck %s --check-prefixes=X64
4
5;
6; fixed avg(x,y) = sub(or(x,y),ashr(xor(x,y),1))
7;
8; ext avg(x,y) = trunc(ashr(add(sext(x),sext(y),1),1))
9;
10
11define i8 @test_fixed_i8(i8 %a0, i8 %a1) nounwind {
12; X86-LABEL: test_fixed_i8:
13; X86:       # %bb.0:
14; X86-NEXT:    movsbl {{[0-9]+}}(%esp), %eax
15; X86-NEXT:    movsbl {{[0-9]+}}(%esp), %ecx
16; X86-NEXT:    leal 1(%ecx,%eax), %eax
17; X86-NEXT:    shrl %eax
18; X86-NEXT:    # kill: def $al killed $al killed $eax
19; X86-NEXT:    retl
20;
21; X64-LABEL: test_fixed_i8:
22; X64:       # %bb.0:
23; X64-NEXT:    movsbl %sil, %eax
24; X64-NEXT:    movsbl %dil, %ecx
25; X64-NEXT:    leal 1(%rcx,%rax), %eax
26; X64-NEXT:    shrl %eax
27; X64-NEXT:    # kill: def $al killed $al killed $eax
28; X64-NEXT:    retq
29  %or = or i8 %a0, %a1
30  %xor = xor i8 %a0, %a1
31  %shift = ashr i8 %xor, 1
32  %res = sub i8 %or, %shift
33  ret i8 %res
34}
35
36define i8 @test_ext_i8(i8 %a0, i8 %a1) nounwind {
37; X86-LABEL: test_ext_i8:
38; X86:       # %bb.0:
39; X86-NEXT:    movsbl {{[0-9]+}}(%esp), %eax
40; X86-NEXT:    movsbl {{[0-9]+}}(%esp), %ecx
41; X86-NEXT:    leal 1(%ecx,%eax), %eax
42; X86-NEXT:    shrl %eax
43; X86-NEXT:    # kill: def $al killed $al killed $eax
44; X86-NEXT:    retl
45;
46; X64-LABEL: test_ext_i8:
47; X64:       # %bb.0:
48; X64-NEXT:    movsbl %sil, %eax
49; X64-NEXT:    movsbl %dil, %ecx
50; X64-NEXT:    leal 1(%rcx,%rax), %eax
51; X64-NEXT:    shrl %eax
52; X64-NEXT:    # kill: def $al killed $al killed $eax
53; X64-NEXT:    retq
54  %x0 = sext i8 %a0 to i16
55  %x1 = sext i8 %a1 to i16
56  %sum = add i16 %x0, %x1
57  %sum1 = add i16 %sum, 1
58  %shift = ashr i16 %sum1, 1
59  %res = trunc i16 %shift to i8
60  ret i8 %res
61}
62
63define i16 @test_fixed_i16(i16 %a0, i16 %a1) nounwind {
64; X86-LABEL: test_fixed_i16:
65; X86:       # %bb.0:
66; X86-NEXT:    movswl {{[0-9]+}}(%esp), %eax
67; X86-NEXT:    movswl {{[0-9]+}}(%esp), %ecx
68; X86-NEXT:    leal 1(%ecx,%eax), %eax
69; X86-NEXT:    shrl %eax
70; X86-NEXT:    # kill: def $ax killed $ax killed $eax
71; X86-NEXT:    retl
72;
73; X64-LABEL: test_fixed_i16:
74; X64:       # %bb.0:
75; X64-NEXT:    movswl %si, %eax
76; X64-NEXT:    movswl %di, %ecx
77; X64-NEXT:    leal 1(%rcx,%rax), %eax
78; X64-NEXT:    shrl %eax
79; X64-NEXT:    # kill: def $ax killed $ax killed $eax
80; X64-NEXT:    retq
81  %or = or i16 %a0, %a1
82  %xor = xor i16 %a0, %a1
83  %shift = ashr i16 %xor, 1
84  %res = sub i16 %or, %shift
85  ret i16 %res
86}
87
88define i16 @test_ext_i16(i16 %a0, i16 %a1) nounwind {
89; X86-LABEL: test_ext_i16:
90; X86:       # %bb.0:
91; X86-NEXT:    movswl {{[0-9]+}}(%esp), %eax
92; X86-NEXT:    movswl {{[0-9]+}}(%esp), %ecx
93; X86-NEXT:    leal 1(%ecx,%eax), %eax
94; X86-NEXT:    shrl %eax
95; X86-NEXT:    # kill: def $ax killed $ax killed $eax
96; X86-NEXT:    retl
97;
98; X64-LABEL: test_ext_i16:
99; X64:       # %bb.0:
100; X64-NEXT:    movswl %si, %eax
101; X64-NEXT:    movswl %di, %ecx
102; X64-NEXT:    leal 1(%rcx,%rax), %eax
103; X64-NEXT:    shrl %eax
104; X64-NEXT:    # kill: def $ax killed $ax killed $eax
105; X64-NEXT:    retq
106  %x0 = sext i16 %a0 to i32
107  %x1 = sext i16 %a1 to i32
108  %sum = add i32 %x0, %x1
109  %sum1 = add i32 %sum, 1
110  %shift = ashr i32 %sum1, 1
111  %res = trunc i32 %shift to i16
112  ret i16 %res
113}
114
115define i32 @test_fixed_i32(i32 %a0, i32 %a1) nounwind {
116; X86-LABEL: test_fixed_i32:
117; X86:       # %bb.0:
118; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
119; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
120; X86-NEXT:    movl %edx, %eax
121; X86-NEXT:    orl %ecx, %eax
122; X86-NEXT:    xorl %ecx, %edx
123; X86-NEXT:    sarl %edx
124; X86-NEXT:    subl %edx, %eax
125; X86-NEXT:    retl
126;
127; X64-LABEL: test_fixed_i32:
128; X64:       # %bb.0:
129; X64-NEXT:    movslq %esi, %rax
130; X64-NEXT:    movslq %edi, %rcx
131; X64-NEXT:    leaq 1(%rcx,%rax), %rax
132; X64-NEXT:    shrq %rax
133; X64-NEXT:    # kill: def $eax killed $eax killed $rax
134; X64-NEXT:    retq
135  %or = or i32 %a0, %a1
136  %xor = xor i32 %a1, %a0
137  %shift = ashr i32 %xor, 1
138  %res = sub i32 %or, %shift
139  ret i32 %res
140}
141
142define i32 @test_ext_i32(i32 %a0, i32 %a1) nounwind {
143; X86-LABEL: test_ext_i32:
144; X86:       # %bb.0:
145; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
146; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
147; X86-NEXT:    movl %edx, %eax
148; X86-NEXT:    orl %ecx, %eax
149; X86-NEXT:    xorl %ecx, %edx
150; X86-NEXT:    sarl %edx
151; X86-NEXT:    subl %edx, %eax
152; X86-NEXT:    retl
153;
154; X64-LABEL: test_ext_i32:
155; X64:       # %bb.0:
156; X64-NEXT:    movslq %esi, %rax
157; X64-NEXT:    movslq %edi, %rcx
158; X64-NEXT:    leaq 1(%rcx,%rax), %rax
159; X64-NEXT:    shrq %rax
160; X64-NEXT:    # kill: def $eax killed $eax killed $rax
161; X64-NEXT:    retq
162  %x0 = sext i32 %a0 to i64
163  %x1 = sext i32 %a1 to i64
164  %sum = add i64 %x0, %x1
165  %sum1 = add i64 %sum, 1
166  %shift = ashr i64 %sum1, 1
167  %res = trunc i64 %shift to i32
168  ret i32 %res
169}
170
171define i64 @test_fixed_i64(i64 %a0, i64 %a1) nounwind {
172; X86-LABEL: test_fixed_i64:
173; X86:       # %bb.0:
174; X86-NEXT:    pushl %ebx
175; X86-NEXT:    pushl %edi
176; X86-NEXT:    pushl %esi
177; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
178; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
179; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
180; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
181; X86-NEXT:    movl %eax, %edi
182; X86-NEXT:    xorl %ecx, %edi
183; X86-NEXT:    movl %edx, %ebx
184; X86-NEXT:    xorl %esi, %ebx
185; X86-NEXT:    shrdl $1, %ebx, %edi
186; X86-NEXT:    orl %esi, %edx
187; X86-NEXT:    sarl %ebx
188; X86-NEXT:    orl %ecx, %eax
189; X86-NEXT:    subl %edi, %eax
190; X86-NEXT:    sbbl %ebx, %edx
191; X86-NEXT:    popl %esi
192; X86-NEXT:    popl %edi
193; X86-NEXT:    popl %ebx
194; X86-NEXT:    retl
195;
196; X64-LABEL: test_fixed_i64:
197; X64:       # %bb.0:
198; X64-NEXT:    movq %rdi, %rax
199; X64-NEXT:    orq %rsi, %rax
200; X64-NEXT:    xorq %rsi, %rdi
201; X64-NEXT:    sarq %rdi
202; X64-NEXT:    subq %rdi, %rax
203; X64-NEXT:    retq
204  %or = or i64 %a0, %a1
205  %xor = xor i64 %a1, %a0
206  %shift = ashr i64 %xor, 1
207  %res = sub i64 %or, %shift
208  ret i64 %res
209}
210
211define i64 @test_ext_i64(i64 %a0, i64 %a1) nounwind {
212; X86-LABEL: test_ext_i64:
213; X86:       # %bb.0:
214; X86-NEXT:    pushl %ebx
215; X86-NEXT:    pushl %edi
216; X86-NEXT:    pushl %esi
217; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
218; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
219; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
220; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
221; X86-NEXT:    movl %eax, %edi
222; X86-NEXT:    xorl %ecx, %edi
223; X86-NEXT:    movl %edx, %ebx
224; X86-NEXT:    xorl %esi, %ebx
225; X86-NEXT:    shrdl $1, %ebx, %edi
226; X86-NEXT:    orl %esi, %edx
227; X86-NEXT:    sarl %ebx
228; X86-NEXT:    orl %ecx, %eax
229; X86-NEXT:    subl %edi, %eax
230; X86-NEXT:    sbbl %ebx, %edx
231; X86-NEXT:    popl %esi
232; X86-NEXT:    popl %edi
233; X86-NEXT:    popl %ebx
234; X86-NEXT:    retl
235;
236; X64-LABEL: test_ext_i64:
237; X64:       # %bb.0:
238; X64-NEXT:    movq %rdi, %rax
239; X64-NEXT:    orq %rsi, %rax
240; X64-NEXT:    xorq %rsi, %rdi
241; X64-NEXT:    sarq %rdi
242; X64-NEXT:    subq %rdi, %rax
243; X64-NEXT:    retq
244  %x0 = sext i64 %a0 to i128
245  %x1 = sext i64 %a1 to i128
246  %sum = add i128 %x0, %x1
247  %sum1 = add i128 %sum, 1
248  %shift = ashr i128 %sum1, 1
249  %res = trunc i128 %shift to i64
250  ret i64 %res
251}
252