1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-unknown | FileCheck %s --check-prefixes=X32,X32-NOSSE
3; RUN: llc < %s -mtriple=x86_64-unknown | FileCheck %s --check-prefix=X64
4; RUN: llc < %s -mtriple=i686-unknown -mattr=+popcnt | FileCheck %s --check-prefix=X32-POPCNT
5; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+popcnt | FileCheck %s --check-prefix=X64-POPCNT
6; RUN: llc < %s -mtriple=i686-unknown -mattr=sse2 | FileCheck %s --check-prefixes=X32,X32-SSE2
7; RUN: llc < %s -mtriple=i686-unknown -mattr=ssse3 | FileCheck %s --check-prefixes=X32,X32-SSSE3
8
9define i8 @cnt8(i8 %x) nounwind readnone {
10; X32-LABEL: cnt8:
11; X32:       # %bb.0:
12; X32-NEXT:    movb {{[0-9]+}}(%esp), %cl
13; X32-NEXT:    movl %ecx, %eax
14; X32-NEXT:    shrb %al
15; X32-NEXT:    andb $85, %al
16; X32-NEXT:    subb %al, %cl
17; X32-NEXT:    movl %ecx, %eax
18; X32-NEXT:    andb $51, %al
19; X32-NEXT:    shrb $2, %cl
20; X32-NEXT:    andb $51, %cl
21; X32-NEXT:    addb %al, %cl
22; X32-NEXT:    movl %ecx, %eax
23; X32-NEXT:    shrb $4, %al
24; X32-NEXT:    addb %cl, %al
25; X32-NEXT:    andb $15, %al
26; X32-NEXT:    retl
27;
28; X64-LABEL: cnt8:
29; X64:       # %bb.0:
30; X64-NEXT:    # kill: def $edi killed $edi def $rdi
31; X64-NEXT:    movl %edi, %eax
32; X64-NEXT:    shrb %al
33; X64-NEXT:    andb $85, %al
34; X64-NEXT:    subb %al, %dil
35; X64-NEXT:    movl %edi, %eax
36; X64-NEXT:    andb $51, %al
37; X64-NEXT:    shrb $2, %dil
38; X64-NEXT:    andb $51, %dil
39; X64-NEXT:    addb %al, %dil
40; X64-NEXT:    movl %edi, %eax
41; X64-NEXT:    shrb $4, %al
42; X64-NEXT:    leal (%rax,%rdi), %eax
43; X64-NEXT:    andb $15, %al
44; X64-NEXT:    # kill: def $al killed $al killed $eax
45; X64-NEXT:    retq
46;
47; X32-POPCNT-LABEL: cnt8:
48; X32-POPCNT:       # %bb.0:
49; X32-POPCNT-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
50; X32-POPCNT-NEXT:    popcntl %eax, %eax
51; X32-POPCNT-NEXT:    # kill: def $al killed $al killed $eax
52; X32-POPCNT-NEXT:    retl
53;
54; X64-POPCNT-LABEL: cnt8:
55; X64-POPCNT:       # %bb.0:
56; X64-POPCNT-NEXT:    movzbl %dil, %eax
57; X64-POPCNT-NEXT:    popcntl %eax, %eax
58; X64-POPCNT-NEXT:    # kill: def $al killed $al killed $eax
59; X64-POPCNT-NEXT:    retq
60  %cnt = tail call i8 @llvm.ctpop.i8(i8 %x)
61  ret i8 %cnt
62}
63
64define i16 @cnt16(i16 %x) nounwind readnone {
65; X32-LABEL: cnt16:
66; X32:       # %bb.0:
67; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
68; X32-NEXT:    movl %eax, %ecx
69; X32-NEXT:    shrl %ecx
70; X32-NEXT:    andl $21845, %ecx # imm = 0x5555
71; X32-NEXT:    subl %ecx, %eax
72; X32-NEXT:    movl %eax, %ecx
73; X32-NEXT:    andl $13107, %ecx # imm = 0x3333
74; X32-NEXT:    shrl $2, %eax
75; X32-NEXT:    andl $13107, %eax # imm = 0x3333
76; X32-NEXT:    addl %ecx, %eax
77; X32-NEXT:    movl %eax, %ecx
78; X32-NEXT:    shrl $4, %ecx
79; X32-NEXT:    addl %eax, %ecx
80; X32-NEXT:    andl $3855, %ecx # imm = 0xF0F
81; X32-NEXT:    movl %ecx, %eax
82; X32-NEXT:    shll $8, %eax
83; X32-NEXT:    addl %ecx, %eax
84; X32-NEXT:    movzbl %ah, %eax
85; X32-NEXT:    # kill: def $ax killed $ax killed $eax
86; X32-NEXT:    retl
87;
88; X64-LABEL: cnt16:
89; X64:       # %bb.0:
90; X64-NEXT:    movl %edi, %eax
91; X64-NEXT:    shrl %eax
92; X64-NEXT:    andl $21845, %eax # imm = 0x5555
93; X64-NEXT:    subl %eax, %edi
94; X64-NEXT:    movl %edi, %eax
95; X64-NEXT:    andl $13107, %eax # imm = 0x3333
96; X64-NEXT:    shrl $2, %edi
97; X64-NEXT:    andl $13107, %edi # imm = 0x3333
98; X64-NEXT:    addl %eax, %edi
99; X64-NEXT:    movl %edi, %eax
100; X64-NEXT:    shrl $4, %eax
101; X64-NEXT:    addl %edi, %eax
102; X64-NEXT:    andl $3855, %eax # imm = 0xF0F
103; X64-NEXT:    movl %eax, %ecx
104; X64-NEXT:    shll $8, %ecx
105; X64-NEXT:    addl %eax, %ecx
106; X64-NEXT:    movzbl %ch, %eax
107; X64-NEXT:    # kill: def $ax killed $ax killed $eax
108; X64-NEXT:    retq
109;
110; X32-POPCNT-LABEL: cnt16:
111; X32-POPCNT:       # %bb.0:
112; X32-POPCNT-NEXT:    popcntw {{[0-9]+}}(%esp), %ax
113; X32-POPCNT-NEXT:    retl
114;
115; X64-POPCNT-LABEL: cnt16:
116; X64-POPCNT:       # %bb.0:
117; X64-POPCNT-NEXT:    popcntw %di, %ax
118; X64-POPCNT-NEXT:    retq
119  %cnt = tail call i16 @llvm.ctpop.i16(i16 %x)
120  ret i16 %cnt
121}
122
123define i32 @cnt32(i32 %x) nounwind readnone {
124; X32-LABEL: cnt32:
125; X32:       # %bb.0:
126; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
127; X32-NEXT:    movl %eax, %ecx
128; X32-NEXT:    shrl %ecx
129; X32-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
130; X32-NEXT:    subl %ecx, %eax
131; X32-NEXT:    movl %eax, %ecx
132; X32-NEXT:    andl $858993459, %ecx # imm = 0x33333333
133; X32-NEXT:    shrl $2, %eax
134; X32-NEXT:    andl $858993459, %eax # imm = 0x33333333
135; X32-NEXT:    addl %ecx, %eax
136; X32-NEXT:    movl %eax, %ecx
137; X32-NEXT:    shrl $4, %ecx
138; X32-NEXT:    addl %eax, %ecx
139; X32-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
140; X32-NEXT:    imull $16843009, %ecx, %eax # imm = 0x1010101
141; X32-NEXT:    shrl $24, %eax
142; X32-NEXT:    retl
143;
144; X64-LABEL: cnt32:
145; X64:       # %bb.0:
146; X64-NEXT:    movl %edi, %eax
147; X64-NEXT:    shrl %eax
148; X64-NEXT:    andl $1431655765, %eax # imm = 0x55555555
149; X64-NEXT:    subl %eax, %edi
150; X64-NEXT:    movl %edi, %eax
151; X64-NEXT:    andl $858993459, %eax # imm = 0x33333333
152; X64-NEXT:    shrl $2, %edi
153; X64-NEXT:    andl $858993459, %edi # imm = 0x33333333
154; X64-NEXT:    addl %eax, %edi
155; X64-NEXT:    movl %edi, %eax
156; X64-NEXT:    shrl $4, %eax
157; X64-NEXT:    addl %edi, %eax
158; X64-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
159; X64-NEXT:    imull $16843009, %eax, %eax # imm = 0x1010101
160; X64-NEXT:    shrl $24, %eax
161; X64-NEXT:    retq
162;
163; X32-POPCNT-LABEL: cnt32:
164; X32-POPCNT:       # %bb.0:
165; X32-POPCNT-NEXT:    popcntl {{[0-9]+}}(%esp), %eax
166; X32-POPCNT-NEXT:    retl
167;
168; X64-POPCNT-LABEL: cnt32:
169; X64-POPCNT:       # %bb.0:
170; X64-POPCNT-NEXT:    popcntl %edi, %eax
171; X64-POPCNT-NEXT:    retq
172  %cnt = tail call i32 @llvm.ctpop.i32(i32 %x)
173  ret i32 %cnt
174}
175
176define i64 @cnt64(i64 %x) nounwind readnone {
177; X32-NOSSE-LABEL: cnt64:
178; X32-NOSSE:       # %bb.0:
179; X32-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
180; X32-NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %ecx
181; X32-NOSSE-NEXT:    movl %ecx, %edx
182; X32-NOSSE-NEXT:    shrl %edx
183; X32-NOSSE-NEXT:    andl $1431655765, %edx # imm = 0x55555555
184; X32-NOSSE-NEXT:    subl %edx, %ecx
185; X32-NOSSE-NEXT:    movl %ecx, %edx
186; X32-NOSSE-NEXT:    andl $858993459, %edx # imm = 0x33333333
187; X32-NOSSE-NEXT:    shrl $2, %ecx
188; X32-NOSSE-NEXT:    andl $858993459, %ecx # imm = 0x33333333
189; X32-NOSSE-NEXT:    addl %edx, %ecx
190; X32-NOSSE-NEXT:    movl %ecx, %edx
191; X32-NOSSE-NEXT:    shrl $4, %edx
192; X32-NOSSE-NEXT:    addl %ecx, %edx
193; X32-NOSSE-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
194; X32-NOSSE-NEXT:    imull $16843009, %edx, %ecx # imm = 0x1010101
195; X32-NOSSE-NEXT:    shrl $24, %ecx
196; X32-NOSSE-NEXT:    movl %eax, %edx
197; X32-NOSSE-NEXT:    shrl %edx
198; X32-NOSSE-NEXT:    andl $1431655765, %edx # imm = 0x55555555
199; X32-NOSSE-NEXT:    subl %edx, %eax
200; X32-NOSSE-NEXT:    movl %eax, %edx
201; X32-NOSSE-NEXT:    andl $858993459, %edx # imm = 0x33333333
202; X32-NOSSE-NEXT:    shrl $2, %eax
203; X32-NOSSE-NEXT:    andl $858993459, %eax # imm = 0x33333333
204; X32-NOSSE-NEXT:    addl %edx, %eax
205; X32-NOSSE-NEXT:    movl %eax, %edx
206; X32-NOSSE-NEXT:    shrl $4, %edx
207; X32-NOSSE-NEXT:    addl %eax, %edx
208; X32-NOSSE-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
209; X32-NOSSE-NEXT:    imull $16843009, %edx, %eax # imm = 0x1010101
210; X32-NOSSE-NEXT:    shrl $24, %eax
211; X32-NOSSE-NEXT:    addl %ecx, %eax
212; X32-NOSSE-NEXT:    xorl %edx, %edx
213; X32-NOSSE-NEXT:    retl
214;
215; X64-LABEL: cnt64:
216; X64:       # %bb.0:
217; X64-NEXT:    movq %rdi, %rax
218; X64-NEXT:    shrq %rax
219; X64-NEXT:    movabsq $6148914691236517205, %rcx # imm = 0x5555555555555555
220; X64-NEXT:    andq %rax, %rcx
221; X64-NEXT:    subq %rcx, %rdi
222; X64-NEXT:    movabsq $3689348814741910323, %rax # imm = 0x3333333333333333
223; X64-NEXT:    movq %rdi, %rcx
224; X64-NEXT:    andq %rax, %rcx
225; X64-NEXT:    shrq $2, %rdi
226; X64-NEXT:    andq %rax, %rdi
227; X64-NEXT:    addq %rcx, %rdi
228; X64-NEXT:    movq %rdi, %rax
229; X64-NEXT:    shrq $4, %rax
230; X64-NEXT:    leaq (%rax,%rdi), %rax
231; X64-NEXT:    movabsq $1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F
232; X64-NEXT:    andq %rax, %rcx
233; X64-NEXT:    movabsq $72340172838076673, %rax # imm = 0x101010101010101
234; X64-NEXT:    imulq %rcx, %rax
235; X64-NEXT:    shrq $56, %rax
236; X64-NEXT:    retq
237;
238; X32-POPCNT-LABEL: cnt64:
239; X32-POPCNT:       # %bb.0:
240; X32-POPCNT-NEXT:    popcntl {{[0-9]+}}(%esp), %ecx
241; X32-POPCNT-NEXT:    popcntl {{[0-9]+}}(%esp), %eax
242; X32-POPCNT-NEXT:    addl %ecx, %eax
243; X32-POPCNT-NEXT:    xorl %edx, %edx
244; X32-POPCNT-NEXT:    retl
245;
246; X64-POPCNT-LABEL: cnt64:
247; X64-POPCNT:       # %bb.0:
248; X64-POPCNT-NEXT:    popcntq %rdi, %rax
249; X64-POPCNT-NEXT:    retq
250;
251; X32-SSE2-LABEL: cnt64:
252; X32-SSE2:       # %bb.0:
253; X32-SSE2-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
254; X32-SSE2-NEXT:    movdqa %xmm0, %xmm1
255; X32-SSE2-NEXT:    psrlw $1, %xmm1
256; X32-SSE2-NEXT:    pand {{\.LCPI.*}}, %xmm1
257; X32-SSE2-NEXT:    psubb %xmm1, %xmm0
258; X32-SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
259; X32-SSE2-NEXT:    movdqa %xmm0, %xmm2
260; X32-SSE2-NEXT:    pand %xmm1, %xmm2
261; X32-SSE2-NEXT:    psrlw $2, %xmm0
262; X32-SSE2-NEXT:    pand %xmm1, %xmm0
263; X32-SSE2-NEXT:    paddb %xmm2, %xmm0
264; X32-SSE2-NEXT:    movdqa %xmm0, %xmm1
265; X32-SSE2-NEXT:    psrlw $4, %xmm1
266; X32-SSE2-NEXT:    paddb %xmm0, %xmm1
267; X32-SSE2-NEXT:    pand {{\.LCPI.*}}, %xmm1
268; X32-SSE2-NEXT:    pxor %xmm0, %xmm0
269; X32-SSE2-NEXT:    psadbw %xmm1, %xmm0
270; X32-SSE2-NEXT:    movd %xmm0, %eax
271; X32-SSE2-NEXT:    xorl %edx, %edx
272; X32-SSE2-NEXT:    retl
273;
274; X32-SSSE3-LABEL: cnt64:
275; X32-SSSE3:       # %bb.0:
276; X32-SSSE3-NEXT:    movdqa {{.*#+}} xmm0 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
277; X32-SSSE3-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
278; X32-SSSE3-NEXT:    movdqa %xmm1, %xmm2
279; X32-SSSE3-NEXT:    pand %xmm0, %xmm2
280; X32-SSSE3-NEXT:    movdqa {{.*#+}} xmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
281; X32-SSSE3-NEXT:    movdqa %xmm3, %xmm4
282; X32-SSSE3-NEXT:    pshufb %xmm2, %xmm4
283; X32-SSSE3-NEXT:    psrlw $4, %xmm1
284; X32-SSSE3-NEXT:    pand %xmm0, %xmm1
285; X32-SSSE3-NEXT:    pshufb %xmm1, %xmm3
286; X32-SSSE3-NEXT:    paddb %xmm4, %xmm3
287; X32-SSSE3-NEXT:    pxor %xmm0, %xmm0
288; X32-SSSE3-NEXT:    psadbw %xmm3, %xmm0
289; X32-SSSE3-NEXT:    movd %xmm0, %eax
290; X32-SSSE3-NEXT:    xorl %edx, %edx
291; X32-SSSE3-NEXT:    retl
292  %cnt = tail call i64 @llvm.ctpop.i64(i64 %x)
293  ret i64 %cnt
294}
295
296define i64 @cnt64_noimplicitfloat(i64 %x) nounwind readnone noimplicitfloat  {
297; X32-LABEL: cnt64_noimplicitfloat:
298; X32:       # %bb.0:
299; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
300; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
301; X32-NEXT:    movl %ecx, %edx
302; X32-NEXT:    shrl %edx
303; X32-NEXT:    andl $1431655765, %edx # imm = 0x55555555
304; X32-NEXT:    subl %edx, %ecx
305; X32-NEXT:    movl %ecx, %edx
306; X32-NEXT:    andl $858993459, %edx # imm = 0x33333333
307; X32-NEXT:    shrl $2, %ecx
308; X32-NEXT:    andl $858993459, %ecx # imm = 0x33333333
309; X32-NEXT:    addl %edx, %ecx
310; X32-NEXT:    movl %ecx, %edx
311; X32-NEXT:    shrl $4, %edx
312; X32-NEXT:    addl %ecx, %edx
313; X32-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
314; X32-NEXT:    imull $16843009, %edx, %ecx # imm = 0x1010101
315; X32-NEXT:    shrl $24, %ecx
316; X32-NEXT:    movl %eax, %edx
317; X32-NEXT:    shrl %edx
318; X32-NEXT:    andl $1431655765, %edx # imm = 0x55555555
319; X32-NEXT:    subl %edx, %eax
320; X32-NEXT:    movl %eax, %edx
321; X32-NEXT:    andl $858993459, %edx # imm = 0x33333333
322; X32-NEXT:    shrl $2, %eax
323; X32-NEXT:    andl $858993459, %eax # imm = 0x33333333
324; X32-NEXT:    addl %edx, %eax
325; X32-NEXT:    movl %eax, %edx
326; X32-NEXT:    shrl $4, %edx
327; X32-NEXT:    addl %eax, %edx
328; X32-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
329; X32-NEXT:    imull $16843009, %edx, %eax # imm = 0x1010101
330; X32-NEXT:    shrl $24, %eax
331; X32-NEXT:    addl %ecx, %eax
332; X32-NEXT:    xorl %edx, %edx
333; X32-NEXT:    retl
334;
335; X64-LABEL: cnt64_noimplicitfloat:
336; X64:       # %bb.0:
337; X64-NEXT:    movq %rdi, %rax
338; X64-NEXT:    shrq %rax
339; X64-NEXT:    movabsq $6148914691236517205, %rcx # imm = 0x5555555555555555
340; X64-NEXT:    andq %rax, %rcx
341; X64-NEXT:    subq %rcx, %rdi
342; X64-NEXT:    movabsq $3689348814741910323, %rax # imm = 0x3333333333333333
343; X64-NEXT:    movq %rdi, %rcx
344; X64-NEXT:    andq %rax, %rcx
345; X64-NEXT:    shrq $2, %rdi
346; X64-NEXT:    andq %rax, %rdi
347; X64-NEXT:    addq %rcx, %rdi
348; X64-NEXT:    movq %rdi, %rax
349; X64-NEXT:    shrq $4, %rax
350; X64-NEXT:    leaq (%rax,%rdi), %rax
351; X64-NEXT:    movabsq $1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F
352; X64-NEXT:    andq %rax, %rcx
353; X64-NEXT:    movabsq $72340172838076673, %rax # imm = 0x101010101010101
354; X64-NEXT:    imulq %rcx, %rax
355; X64-NEXT:    shrq $56, %rax
356; X64-NEXT:    retq
357;
358; X32-POPCNT-LABEL: cnt64_noimplicitfloat:
359; X32-POPCNT:       # %bb.0:
360; X32-POPCNT-NEXT:    popcntl {{[0-9]+}}(%esp), %ecx
361; X32-POPCNT-NEXT:    popcntl {{[0-9]+}}(%esp), %eax
362; X32-POPCNT-NEXT:    addl %ecx, %eax
363; X32-POPCNT-NEXT:    xorl %edx, %edx
364; X32-POPCNT-NEXT:    retl
365;
366; X64-POPCNT-LABEL: cnt64_noimplicitfloat:
367; X64-POPCNT:       # %bb.0:
368; X64-POPCNT-NEXT:    popcntq %rdi, %rax
369; X64-POPCNT-NEXT:    retq
370  %cnt = tail call i64 @llvm.ctpop.i64(i64 %x)
371  ret i64 %cnt
372}
373
374declare i8 @llvm.ctpop.i8(i8) nounwind readnone
375declare i16 @llvm.ctpop.i16(i16) nounwind readnone
376declare i32 @llvm.ctpop.i32(i32) nounwind readnone
377declare i64 @llvm.ctpop.i64(i64) nounwind readnone
378