1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-unknown | FileCheck %s --check-prefix=X86
3; RUN: llc < %s -mtriple=x86_64-unknown | FileCheck %s --check-prefix=X64
4; RUN: llc < %s -mtriple=i686-unknown -mattr=+xop | FileCheck %s --check-prefixes=X86XOP
5
6; These tests just check that the plumbing is in place for @llvm.bitreverse. The
7; actual output is massive at the moment as llvm.bitreverse is not yet legal.
8
9declare <2 x i16> @llvm.bitreverse.v2i16(<2 x i16>) readnone
10
11define <2 x i16> @test_bitreverse_v2i16(<2 x i16> %a) nounwind {
12; X86-LABEL: test_bitreverse_v2i16:
13; X86:       # %bb.0:
14; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
15; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
16; X86-NEXT:    rolw $8, %ax
17; X86-NEXT:    movl %eax, %edx
18; X86-NEXT:    andl $3855, %edx # imm = 0xF0F
19; X86-NEXT:    shll $4, %edx
20; X86-NEXT:    shrl $4, %eax
21; X86-NEXT:    andl $3855, %eax # imm = 0xF0F
22; X86-NEXT:    orl %edx, %eax
23; X86-NEXT:    movl %eax, %edx
24; X86-NEXT:    andl $13107, %edx # imm = 0x3333
25; X86-NEXT:    shrl $2, %eax
26; X86-NEXT:    andl $13107, %eax # imm = 0x3333
27; X86-NEXT:    leal (%eax,%edx,4), %eax
28; X86-NEXT:    movl %eax, %edx
29; X86-NEXT:    andl $21845, %edx # imm = 0x5555
30; X86-NEXT:    shrl %eax
31; X86-NEXT:    andl $21845, %eax # imm = 0x5555
32; X86-NEXT:    leal (%eax,%edx,2), %eax
33; X86-NEXT:    rolw $8, %cx
34; X86-NEXT:    movl %ecx, %edx
35; X86-NEXT:    andl $3855, %edx # imm = 0xF0F
36; X86-NEXT:    shll $4, %edx
37; X86-NEXT:    shrl $4, %ecx
38; X86-NEXT:    andl $3855, %ecx # imm = 0xF0F
39; X86-NEXT:    orl %edx, %ecx
40; X86-NEXT:    movl %ecx, %edx
41; X86-NEXT:    andl $13107, %edx # imm = 0x3333
42; X86-NEXT:    shrl $2, %ecx
43; X86-NEXT:    andl $13107, %ecx # imm = 0x3333
44; X86-NEXT:    leal (%ecx,%edx,4), %ecx
45; X86-NEXT:    movl %ecx, %edx
46; X86-NEXT:    andl $21845, %edx # imm = 0x5555
47; X86-NEXT:    shrl %ecx
48; X86-NEXT:    andl $21845, %ecx # imm = 0x5555
49; X86-NEXT:    leal (%ecx,%edx,2), %edx
50; X86-NEXT:    # kill: def $ax killed $ax killed $eax
51; X86-NEXT:    # kill: def $dx killed $dx killed $edx
52; X86-NEXT:    retl
53;
54; X64-LABEL: test_bitreverse_v2i16:
55; X64:       # %bb.0:
56; X64-NEXT:    movdqa %xmm0, %xmm1
57; X64-NEXT:    psrlw $8, %xmm1
58; X64-NEXT:    psllw $8, %xmm0
59; X64-NEXT:    por %xmm1, %xmm0
60; X64-NEXT:    movdqa %xmm0, %xmm1
61; X64-NEXT:    psllw $4, %xmm1
62; X64-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
63; X64-NEXT:    psrlw $4, %xmm0
64; X64-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
65; X64-NEXT:    por %xmm1, %xmm0
66; X64-NEXT:    movdqa %xmm0, %xmm1
67; X64-NEXT:    psrlw $2, %xmm1
68; X64-NEXT:    movdqa {{.*#+}} xmm2 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
69; X64-NEXT:    pand %xmm2, %xmm1
70; X64-NEXT:    pand %xmm2, %xmm0
71; X64-NEXT:    psllw $2, %xmm0
72; X64-NEXT:    por %xmm1, %xmm0
73; X64-NEXT:    movdqa %xmm0, %xmm1
74; X64-NEXT:    psrlw $1, %xmm1
75; X64-NEXT:    movdqa {{.*#+}} xmm2 = [85,85,85,85,85,85,85,85,85,85,85,85,85,85,85,85]
76; X64-NEXT:    pand %xmm2, %xmm1
77; X64-NEXT:    pand %xmm2, %xmm0
78; X64-NEXT:    paddb %xmm0, %xmm0
79; X64-NEXT:    por %xmm1, %xmm0
80; X64-NEXT:    retq
81;
82; X86XOP-LABEL: test_bitreverse_v2i16:
83; X86XOP:       # %bb.0:
84; X86XOP-NEXT:    vpperm {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0, %xmm0
85; X86XOP-NEXT:    retl
86  %b = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> %a)
87  ret <2 x i16> %b
88}
89
90declare i64 @llvm.bitreverse.i64(i64) readnone
91
92define i64 @test_bitreverse_i64(i64 %a) nounwind {
93; X86-LABEL: test_bitreverse_i64:
94; X86:       # %bb.0:
95; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
96; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
97; X86-NEXT:    bswapl %eax
98; X86-NEXT:    movl %eax, %edx
99; X86-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
100; X86-NEXT:    shll $4, %edx
101; X86-NEXT:    shrl $4, %eax
102; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
103; X86-NEXT:    orl %edx, %eax
104; X86-NEXT:    movl %eax, %edx
105; X86-NEXT:    andl $858993459, %edx # imm = 0x33333333
106; X86-NEXT:    shrl $2, %eax
107; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
108; X86-NEXT:    leal (%eax,%edx,4), %eax
109; X86-NEXT:    movl %eax, %edx
110; X86-NEXT:    andl $1431655765, %edx # imm = 0x55555555
111; X86-NEXT:    shrl %eax
112; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
113; X86-NEXT:    leal (%eax,%edx,2), %eax
114; X86-NEXT:    bswapl %ecx
115; X86-NEXT:    movl %ecx, %edx
116; X86-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
117; X86-NEXT:    shll $4, %edx
118; X86-NEXT:    shrl $4, %ecx
119; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
120; X86-NEXT:    orl %edx, %ecx
121; X86-NEXT:    movl %ecx, %edx
122; X86-NEXT:    andl $858993459, %edx # imm = 0x33333333
123; X86-NEXT:    shrl $2, %ecx
124; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
125; X86-NEXT:    leal (%ecx,%edx,4), %ecx
126; X86-NEXT:    movl %ecx, %edx
127; X86-NEXT:    andl $1431655765, %edx # imm = 0x55555555
128; X86-NEXT:    shrl %ecx
129; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
130; X86-NEXT:    leal (%ecx,%edx,2), %edx
131; X86-NEXT:    retl
132;
133; X64-LABEL: test_bitreverse_i64:
134; X64:       # %bb.0:
135; X64-NEXT:    bswapq %rdi
136; X64-NEXT:    movq %rdi, %rax
137; X64-NEXT:    shrq $4, %rax
138; X64-NEXT:    movabsq $1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F
139; X64-NEXT:    andq %rcx, %rax
140; X64-NEXT:    andq %rcx, %rdi
141; X64-NEXT:    shlq $4, %rdi
142; X64-NEXT:    orq %rax, %rdi
143; X64-NEXT:    movabsq $3689348814741910323, %rax # imm = 0x3333333333333333
144; X64-NEXT:    movq %rdi, %rcx
145; X64-NEXT:    andq %rax, %rcx
146; X64-NEXT:    shrq $2, %rdi
147; X64-NEXT:    andq %rax, %rdi
148; X64-NEXT:    leaq (%rdi,%rcx,4), %rax
149; X64-NEXT:    movabsq $6148914691236517205, %rcx # imm = 0x5555555555555555
150; X64-NEXT:    movq %rax, %rdx
151; X64-NEXT:    andq %rcx, %rdx
152; X64-NEXT:    shrq %rax
153; X64-NEXT:    andq %rcx, %rax
154; X64-NEXT:    leaq (%rax,%rdx,2), %rax
155; X64-NEXT:    retq
156;
157; X86XOP-LABEL: test_bitreverse_i64:
158; X86XOP:       # %bb.0:
159; X86XOP-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
160; X86XOP-NEXT:    vpperm {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0, %xmm0
161; X86XOP-NEXT:    vmovd %xmm0, %eax
162; X86XOP-NEXT:    vpextrd $1, %xmm0, %edx
163; X86XOP-NEXT:    retl
164  %b = call i64 @llvm.bitreverse.i64(i64 %a)
165  ret i64 %b
166}
167
168declare i32 @llvm.bitreverse.i32(i32) readnone
169
170define i32 @test_bitreverse_i32(i32 %a) nounwind {
171; X86-LABEL: test_bitreverse_i32:
172; X86:       # %bb.0:
173; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
174; X86-NEXT:    bswapl %eax
175; X86-NEXT:    movl %eax, %ecx
176; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
177; X86-NEXT:    shll $4, %ecx
178; X86-NEXT:    shrl $4, %eax
179; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
180; X86-NEXT:    orl %ecx, %eax
181; X86-NEXT:    movl %eax, %ecx
182; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
183; X86-NEXT:    shrl $2, %eax
184; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
185; X86-NEXT:    leal (%eax,%ecx,4), %eax
186; X86-NEXT:    movl %eax, %ecx
187; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
188; X86-NEXT:    shrl %eax
189; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
190; X86-NEXT:    leal (%eax,%ecx,2), %eax
191; X86-NEXT:    retl
192;
193; X64-LABEL: test_bitreverse_i32:
194; X64:       # %bb.0:
195; X64-NEXT:    # kill: def $edi killed $edi def $rdi
196; X64-NEXT:    bswapl %edi
197; X64-NEXT:    movl %edi, %eax
198; X64-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
199; X64-NEXT:    shll $4, %eax
200; X64-NEXT:    shrl $4, %edi
201; X64-NEXT:    andl $252645135, %edi # imm = 0xF0F0F0F
202; X64-NEXT:    orl %eax, %edi
203; X64-NEXT:    movl %edi, %eax
204; X64-NEXT:    andl $858993459, %eax # imm = 0x33333333
205; X64-NEXT:    shrl $2, %edi
206; X64-NEXT:    andl $858993459, %edi # imm = 0x33333333
207; X64-NEXT:    leal (%rdi,%rax,4), %eax
208; X64-NEXT:    movl %eax, %ecx
209; X64-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
210; X64-NEXT:    shrl %eax
211; X64-NEXT:    andl $1431655765, %eax # imm = 0x55555555
212; X64-NEXT:    leal (%rax,%rcx,2), %eax
213; X64-NEXT:    retq
214;
215; X86XOP-LABEL: test_bitreverse_i32:
216; X86XOP:       # %bb.0:
217; X86XOP-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
218; X86XOP-NEXT:    vpperm {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0, %xmm0
219; X86XOP-NEXT:    vmovd %xmm0, %eax
220; X86XOP-NEXT:    retl
221  %b = call i32 @llvm.bitreverse.i32(i32 %a)
222  ret i32 %b
223}
224
225declare i24 @llvm.bitreverse.i24(i24) readnone
226
227define i24 @test_bitreverse_i24(i24 %a) nounwind {
228; X86-LABEL: test_bitreverse_i24:
229; X86:       # %bb.0:
230; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
231; X86-NEXT:    bswapl %eax
232; X86-NEXT:    movl %eax, %ecx
233; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
234; X86-NEXT:    shll $4, %ecx
235; X86-NEXT:    shrl $4, %eax
236; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
237; X86-NEXT:    orl %ecx, %eax
238; X86-NEXT:    movl %eax, %ecx
239; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
240; X86-NEXT:    shrl $2, %eax
241; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
242; X86-NEXT:    leal (%eax,%ecx,4), %eax
243; X86-NEXT:    movl %eax, %ecx
244; X86-NEXT:    andl $1431655680, %ecx # imm = 0x55555500
245; X86-NEXT:    shrl %eax
246; X86-NEXT:    andl $1431655680, %eax # imm = 0x55555500
247; X86-NEXT:    leal (%eax,%ecx,2), %eax
248; X86-NEXT:    shrl $8, %eax
249; X86-NEXT:    retl
250;
251; X64-LABEL: test_bitreverse_i24:
252; X64:       # %bb.0:
253; X64-NEXT:    # kill: def $edi killed $edi def $rdi
254; X64-NEXT:    bswapl %edi
255; X64-NEXT:    movl %edi, %eax
256; X64-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
257; X64-NEXT:    shll $4, %eax
258; X64-NEXT:    shrl $4, %edi
259; X64-NEXT:    andl $252645135, %edi # imm = 0xF0F0F0F
260; X64-NEXT:    orl %eax, %edi
261; X64-NEXT:    movl %edi, %eax
262; X64-NEXT:    andl $858993459, %eax # imm = 0x33333333
263; X64-NEXT:    shrl $2, %edi
264; X64-NEXT:    andl $858993459, %edi # imm = 0x33333333
265; X64-NEXT:    leal (%rdi,%rax,4), %eax
266; X64-NEXT:    movl %eax, %ecx
267; X64-NEXT:    andl $1431655680, %ecx # imm = 0x55555500
268; X64-NEXT:    shrl %eax
269; X64-NEXT:    andl $1431655680, %eax # imm = 0x55555500
270; X64-NEXT:    leal (%rax,%rcx,2), %eax
271; X64-NEXT:    shrl $8, %eax
272; X64-NEXT:    retq
273;
274; X86XOP-LABEL: test_bitreverse_i24:
275; X86XOP:       # %bb.0:
276; X86XOP-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
277; X86XOP-NEXT:    vpperm {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0, %xmm0
278; X86XOP-NEXT:    vmovd %xmm0, %eax
279; X86XOP-NEXT:    shrl $8, %eax
280; X86XOP-NEXT:    retl
281  %b = call i24 @llvm.bitreverse.i24(i24 %a)
282  ret i24 %b
283}
284
285declare i16 @llvm.bitreverse.i16(i16) readnone
286
287define i16 @test_bitreverse_i16(i16 %a) nounwind {
288; X86-LABEL: test_bitreverse_i16:
289; X86:       # %bb.0:
290; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
291; X86-NEXT:    rolw $8, %ax
292; X86-NEXT:    movl %eax, %ecx
293; X86-NEXT:    andl $3855, %ecx # imm = 0xF0F
294; X86-NEXT:    shll $4, %ecx
295; X86-NEXT:    shrl $4, %eax
296; X86-NEXT:    andl $3855, %eax # imm = 0xF0F
297; X86-NEXT:    orl %ecx, %eax
298; X86-NEXT:    movl %eax, %ecx
299; X86-NEXT:    andl $13107, %ecx # imm = 0x3333
300; X86-NEXT:    shrl $2, %eax
301; X86-NEXT:    andl $13107, %eax # imm = 0x3333
302; X86-NEXT:    leal (%eax,%ecx,4), %eax
303; X86-NEXT:    movl %eax, %ecx
304; X86-NEXT:    andl $21845, %ecx # imm = 0x5555
305; X86-NEXT:    shrl %eax
306; X86-NEXT:    andl $21845, %eax # imm = 0x5555
307; X86-NEXT:    leal (%eax,%ecx,2), %eax
308; X86-NEXT:    # kill: def $ax killed $ax killed $eax
309; X86-NEXT:    retl
310;
311; X64-LABEL: test_bitreverse_i16:
312; X64:       # %bb.0:
313; X64-NEXT:    # kill: def $edi killed $edi def $rdi
314; X64-NEXT:    rolw $8, %di
315; X64-NEXT:    movl %edi, %eax
316; X64-NEXT:    andl $3855, %eax # imm = 0xF0F
317; X64-NEXT:    shll $4, %eax
318; X64-NEXT:    shrl $4, %edi
319; X64-NEXT:    andl $3855, %edi # imm = 0xF0F
320; X64-NEXT:    orl %eax, %edi
321; X64-NEXT:    movl %edi, %eax
322; X64-NEXT:    andl $13107, %eax # imm = 0x3333
323; X64-NEXT:    shrl $2, %edi
324; X64-NEXT:    andl $13107, %edi # imm = 0x3333
325; X64-NEXT:    leal (%rdi,%rax,4), %eax
326; X64-NEXT:    movl %eax, %ecx
327; X64-NEXT:    andl $21845, %ecx # imm = 0x5555
328; X64-NEXT:    shrl %eax
329; X64-NEXT:    andl $21845, %eax # imm = 0x5555
330; X64-NEXT:    leal (%rax,%rcx,2), %eax
331; X64-NEXT:    # kill: def $ax killed $ax killed $eax
332; X64-NEXT:    retq
333;
334; X86XOP-LABEL: test_bitreverse_i16:
335; X86XOP:       # %bb.0:
336; X86XOP-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
337; X86XOP-NEXT:    vpperm {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0, %xmm0
338; X86XOP-NEXT:    vmovd %xmm0, %eax
339; X86XOP-NEXT:    # kill: def $ax killed $ax killed $eax
340; X86XOP-NEXT:    retl
341  %b = call i16 @llvm.bitreverse.i16(i16 %a)
342  ret i16 %b
343}
344
345declare i8 @llvm.bitreverse.i8(i8) readnone
346
347define i8 @test_bitreverse_i8(i8 %a) {
348; X86-LABEL: test_bitreverse_i8:
349; X86:       # %bb.0:
350; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
351; X86-NEXT:    rolb $4, %al
352; X86-NEXT:    movl %eax, %ecx
353; X86-NEXT:    andb $51, %cl
354; X86-NEXT:    shlb $2, %cl
355; X86-NEXT:    shrb $2, %al
356; X86-NEXT:    andb $51, %al
357; X86-NEXT:    orb %cl, %al
358; X86-NEXT:    movl %eax, %ecx
359; X86-NEXT:    andb $85, %cl
360; X86-NEXT:    addb %cl, %cl
361; X86-NEXT:    shrb %al
362; X86-NEXT:    andb $85, %al
363; X86-NEXT:    orb %cl, %al
364; X86-NEXT:    retl
365;
366; X64-LABEL: test_bitreverse_i8:
367; X64:       # %bb.0:
368; X64-NEXT:    rolb $4, %dil
369; X64-NEXT:    movl %edi, %eax
370; X64-NEXT:    andb $51, %al
371; X64-NEXT:    shlb $2, %al
372; X64-NEXT:    shrb $2, %dil
373; X64-NEXT:    andb $51, %dil
374; X64-NEXT:    orb %al, %dil
375; X64-NEXT:    movl %edi, %eax
376; X64-NEXT:    andb $85, %al
377; X64-NEXT:    addb %al, %al
378; X64-NEXT:    shrb %dil
379; X64-NEXT:    andb $85, %dil
380; X64-NEXT:    orb %dil, %al
381; X64-NEXT:    retq
382;
383; X86XOP-LABEL: test_bitreverse_i8:
384; X86XOP:       # %bb.0:
385; X86XOP-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
386; X86XOP-NEXT:    vpperm {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0, %xmm0
387; X86XOP-NEXT:    vmovd %xmm0, %eax
388; X86XOP-NEXT:    # kill: def $al killed $al killed $eax
389; X86XOP-NEXT:    retl
390  %b = call i8 @llvm.bitreverse.i8(i8 %a)
391  ret i8 %b
392}
393
394declare i4 @llvm.bitreverse.i4(i4) readnone
395
396define i4 @test_bitreverse_i4(i4 %a) {
397; X86-LABEL: test_bitreverse_i4:
398; X86:       # %bb.0:
399; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
400; X86-NEXT:    rolb $4, %al
401; X86-NEXT:    movl %eax, %ecx
402; X86-NEXT:    andb $51, %cl
403; X86-NEXT:    shlb $2, %cl
404; X86-NEXT:    shrb $2, %al
405; X86-NEXT:    andb $51, %al
406; X86-NEXT:    orb %cl, %al
407; X86-NEXT:    movl %eax, %ecx
408; X86-NEXT:    andb $80, %cl
409; X86-NEXT:    addb %cl, %cl
410; X86-NEXT:    shrb %al
411; X86-NEXT:    andb $80, %al
412; X86-NEXT:    orb %cl, %al
413; X86-NEXT:    shrb $4, %al
414; X86-NEXT:    retl
415;
416; X64-LABEL: test_bitreverse_i4:
417; X64:       # %bb.0:
418; X64-NEXT:    rolb $4, %dil
419; X64-NEXT:    movl %edi, %eax
420; X64-NEXT:    andb $51, %al
421; X64-NEXT:    shlb $2, %al
422; X64-NEXT:    shrb $2, %dil
423; X64-NEXT:    andb $51, %dil
424; X64-NEXT:    orb %al, %dil
425; X64-NEXT:    movl %edi, %eax
426; X64-NEXT:    andb $80, %al
427; X64-NEXT:    addb %al, %al
428; X64-NEXT:    shrb %dil
429; X64-NEXT:    andb $80, %dil
430; X64-NEXT:    orb %dil, %al
431; X64-NEXT:    shrb $4, %al
432; X64-NEXT:    retq
433;
434; X86XOP-LABEL: test_bitreverse_i4:
435; X86XOP:       # %bb.0:
436; X86XOP-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
437; X86XOP-NEXT:    vpperm {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0, %xmm0
438; X86XOP-NEXT:    vmovd %xmm0, %eax
439; X86XOP-NEXT:    shrb $4, %al
440; X86XOP-NEXT:    # kill: def $al killed $al killed $eax
441; X86XOP-NEXT:    retl
442  %b = call i4 @llvm.bitreverse.i4(i4 %a)
443  ret i4 %b
444}
445
446; These tests check that bitreverse(constant) calls are folded
447
448define <2 x i16> @fold_v2i16() {
449; X86-LABEL: fold_v2i16:
450; X86:       # %bb.0:
451; X86-NEXT:    movw $-4096, %ax # imm = 0xF000
452; X86-NEXT:    movw $240, %dx
453; X86-NEXT:    retl
454;
455; X64-LABEL: fold_v2i16:
456; X64:       # %bb.0:
457; X64-NEXT:    movaps {{.*#+}} xmm0 = <61440,240,u,u,u,u,u,u>
458; X64-NEXT:    retq
459;
460; X86XOP-LABEL: fold_v2i16:
461; X86XOP:       # %bb.0:
462; X86XOP-NEXT:    vmovaps {{.*#+}} xmm0 = <61440,240,u,u,u,u,u,u>
463; X86XOP-NEXT:    retl
464  %b = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> <i16 15, i16 3840>)
465  ret <2 x i16> %b
466}
467
468define i24 @fold_i24() {
469; X86-LABEL: fold_i24:
470; X86:       # %bb.0:
471; X86-NEXT:    movl $2048, %eax # imm = 0x800
472; X86-NEXT:    retl
473;
474; X64-LABEL: fold_i24:
475; X64:       # %bb.0:
476; X64-NEXT:    movl $2048, %eax # imm = 0x800
477; X64-NEXT:    retq
478;
479; X86XOP-LABEL: fold_i24:
480; X86XOP:       # %bb.0:
481; X86XOP-NEXT:    movl $2048, %eax # imm = 0x800
482; X86XOP-NEXT:    retl
483  %b = call i24 @llvm.bitreverse.i24(i24 4096)
484  ret i24 %b
485}
486
487define i8 @fold_i8() {
488; X86-LABEL: fold_i8:
489; X86:       # %bb.0:
490; X86-NEXT:    movb $-16, %al
491; X86-NEXT:    retl
492;
493; X64-LABEL: fold_i8:
494; X64:       # %bb.0:
495; X64-NEXT:    movb $-16, %al
496; X64-NEXT:    retq
497;
498; X86XOP-LABEL: fold_i8:
499; X86XOP:       # %bb.0:
500; X86XOP-NEXT:    movb $-16, %al
501; X86XOP-NEXT:    retl
502  %b = call i8 @llvm.bitreverse.i8(i8 15)
503  ret i8 %b
504}
505
506define i4 @fold_i4() {
507; X86-LABEL: fold_i4:
508; X86:       # %bb.0:
509; X86-NEXT:    movb $1, %al
510; X86-NEXT:    retl
511;
512; X64-LABEL: fold_i4:
513; X64:       # %bb.0:
514; X64-NEXT:    movb $1, %al
515; X64-NEXT:    retq
516;
517; X86XOP-LABEL: fold_i4:
518; X86XOP:       # %bb.0:
519; X86XOP-NEXT:    movb $1, %al
520; X86XOP-NEXT:    retl
521  %b = call i4 @llvm.bitreverse.i4(i4 8)
522  ret i4 %b
523}
524
525; These tests check that bitreverse(bitreverse()) calls are removed
526
527define i8 @identity_i8(i8 %a) {
528; X86-LABEL: identity_i8:
529; X86:       # %bb.0:
530; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
531; X86-NEXT:    retl
532;
533; X64-LABEL: identity_i8:
534; X64:       # %bb.0:
535; X64-NEXT:    movl %edi, %eax
536; X64-NEXT:    # kill: def $al killed $al killed $eax
537; X64-NEXT:    retq
538;
539; X86XOP-LABEL: identity_i8:
540; X86XOP:       # %bb.0:
541; X86XOP-NEXT:    movb {{[0-9]+}}(%esp), %al
542; X86XOP-NEXT:    retl
543  %b = call i8 @llvm.bitreverse.i8(i8 %a)
544  %c = call i8 @llvm.bitreverse.i8(i8 %b)
545  ret i8 %c
546}
547
548define <2 x i16> @identity_v2i16(<2 x i16> %a) {
549; X86-LABEL: identity_v2i16:
550; X86:       # %bb.0:
551; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
552; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %edx
553; X86-NEXT:    retl
554;
555; X64-LABEL: identity_v2i16:
556; X64:       # %bb.0:
557; X64-NEXT:    retq
558;
559; X86XOP-LABEL: identity_v2i16:
560; X86XOP:       # %bb.0:
561; X86XOP-NEXT:    retl
562  %b = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> %a)
563  %c = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> %b)
564  ret <2 x i16> %c
565}
566
567; These tests check that bitreverse(undef) calls are removed
568
569define i8 @undef_i8() {
570; X86-LABEL: undef_i8:
571; X86:       # %bb.0:
572; X86-NEXT:    retl
573;
574; X64-LABEL: undef_i8:
575; X64:       # %bb.0:
576; X64-NEXT:    retq
577;
578; X86XOP-LABEL: undef_i8:
579; X86XOP:       # %bb.0:
580; X86XOP-NEXT:    retl
581  %b = call i8 @llvm.bitreverse.i8(i8 undef)
582  ret i8 %b
583}
584
585define <2 x i16> @undef_v2i16() {
586; X86-LABEL: undef_v2i16:
587; X86:       # %bb.0:
588; X86-NEXT:    retl
589;
590; X64-LABEL: undef_v2i16:
591; X64:       # %bb.0:
592; X64-NEXT:    retq
593;
594; X86XOP-LABEL: undef_v2i16:
595; X86XOP:       # %bb.0:
596; X86XOP-NEXT:    retl
597  %b = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> undef)
598  ret <2 x i16> %b
599}
600
601; Make sure we don't assert during type legalization promoting a large
602; bitreverse due to the need for a large shift that won't fit in the i8 returned
603; from getShiftAmountTy.
604define i528 @large_promotion(i528 %A) nounwind {
605; X86-LABEL: large_promotion:
606; X86:       # %bb.0:
607; X86-NEXT:    pushl %ebp
608; X86-NEXT:    pushl %ebx
609; X86-NEXT:    pushl %edi
610; X86-NEXT:    pushl %esi
611; X86-NEXT:    subl $56, %esp
612; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
613; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
614; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
615; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
616; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
617; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
618; X86-NEXT:    bswapl %ebx
619; X86-NEXT:    movl %ebx, %ebp
620; X86-NEXT:    andl $252645135, %ebp # imm = 0xF0F0F0F
621; X86-NEXT:    shll $4, %ebp
622; X86-NEXT:    shrl $4, %ebx
623; X86-NEXT:    andl $252645135, %ebx # imm = 0xF0F0F0F
624; X86-NEXT:    orl %ebp, %ebx
625; X86-NEXT:    movl %ebx, %ebp
626; X86-NEXT:    andl $858993459, %ebp # imm = 0x33333333
627; X86-NEXT:    shrl $2, %ebx
628; X86-NEXT:    andl $858993459, %ebx # imm = 0x33333333
629; X86-NEXT:    leal (%ebx,%ebp,4), %ebx
630; X86-NEXT:    movl %ebx, %ebp
631; X86-NEXT:    andl $1431633920, %ebp # imm = 0x55550000
632; X86-NEXT:    shrl %ebx
633; X86-NEXT:    andl $1431633920, %ebx # imm = 0x55550000
634; X86-NEXT:    leal (%ebx,%ebp,2), %ebx
635; X86-NEXT:    movl %ebx, (%esp) # 4-byte Spill
636; X86-NEXT:    bswapl %edi
637; X86-NEXT:    movl %edi, %ebx
638; X86-NEXT:    andl $252645135, %ebx # imm = 0xF0F0F0F
639; X86-NEXT:    shll $4, %ebx
640; X86-NEXT:    shrl $4, %edi
641; X86-NEXT:    andl $252645135, %edi # imm = 0xF0F0F0F
642; X86-NEXT:    orl %ebx, %edi
643; X86-NEXT:    movl %edi, %ebx
644; X86-NEXT:    andl $858993459, %ebx # imm = 0x33333333
645; X86-NEXT:    shrl $2, %edi
646; X86-NEXT:    andl $858993459, %edi # imm = 0x33333333
647; X86-NEXT:    leal (%edi,%ebx,4), %edi
648; X86-NEXT:    movl %edi, %ebx
649; X86-NEXT:    andl $1431655765, %ebx # imm = 0x55555555
650; X86-NEXT:    shrl %edi
651; X86-NEXT:    andl $1431655765, %edi # imm = 0x55555555
652; X86-NEXT:    leal (%edi,%ebx,2), %ebx
653; X86-NEXT:    bswapl %esi
654; X86-NEXT:    movl %esi, %edi
655; X86-NEXT:    andl $252645135, %edi # imm = 0xF0F0F0F
656; X86-NEXT:    shll $4, %edi
657; X86-NEXT:    shrl $4, %esi
658; X86-NEXT:    andl $252645135, %esi # imm = 0xF0F0F0F
659; X86-NEXT:    orl %edi, %esi
660; X86-NEXT:    movl %esi, %edi
661; X86-NEXT:    andl $858993459, %edi # imm = 0x33333333
662; X86-NEXT:    shrl $2, %esi
663; X86-NEXT:    andl $858993459, %esi # imm = 0x33333333
664; X86-NEXT:    leal (%esi,%edi,4), %esi
665; X86-NEXT:    movl %esi, %edi
666; X86-NEXT:    andl $1431655765, %edi # imm = 0x55555555
667; X86-NEXT:    shrl %esi
668; X86-NEXT:    andl $1431655765, %esi # imm = 0x55555555
669; X86-NEXT:    leal (%esi,%edi,2), %esi
670; X86-NEXT:    movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
671; X86-NEXT:    bswapl %edx
672; X86-NEXT:    movl %edx, %esi
673; X86-NEXT:    andl $252645135, %esi # imm = 0xF0F0F0F
674; X86-NEXT:    shll $4, %esi
675; X86-NEXT:    shrl $4, %edx
676; X86-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
677; X86-NEXT:    orl %esi, %edx
678; X86-NEXT:    movl %edx, %esi
679; X86-NEXT:    andl $858993459, %esi # imm = 0x33333333
680; X86-NEXT:    shrl $2, %edx
681; X86-NEXT:    andl $858993459, %edx # imm = 0x33333333
682; X86-NEXT:    leal (%edx,%esi,4), %edx
683; X86-NEXT:    movl %edx, %esi
684; X86-NEXT:    andl $1431655765, %esi # imm = 0x55555555
685; X86-NEXT:    shrl %edx
686; X86-NEXT:    andl $1431655765, %edx # imm = 0x55555555
687; X86-NEXT:    leal (%edx,%esi,2), %edx
688; X86-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
689; X86-NEXT:    bswapl %ecx
690; X86-NEXT:    movl %ecx, %edx
691; X86-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
692; X86-NEXT:    shll $4, %edx
693; X86-NEXT:    shrl $4, %ecx
694; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
695; X86-NEXT:    orl %edx, %ecx
696; X86-NEXT:    movl %ecx, %edx
697; X86-NEXT:    andl $858993459, %edx # imm = 0x33333333
698; X86-NEXT:    shrl $2, %ecx
699; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
700; X86-NEXT:    leal (%ecx,%edx,4), %ecx
701; X86-NEXT:    movl %ecx, %edx
702; X86-NEXT:    andl $1431655765, %edx # imm = 0x55555555
703; X86-NEXT:    shrl %ecx
704; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
705; X86-NEXT:    leal (%ecx,%edx,2), %ecx
706; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
707; X86-NEXT:    bswapl %eax
708; X86-NEXT:    movl %eax, %ecx
709; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
710; X86-NEXT:    shll $4, %ecx
711; X86-NEXT:    shrl $4, %eax
712; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
713; X86-NEXT:    orl %ecx, %eax
714; X86-NEXT:    movl %eax, %ecx
715; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
716; X86-NEXT:    shrl $2, %eax
717; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
718; X86-NEXT:    leal (%eax,%ecx,4), %eax
719; X86-NEXT:    movl %eax, %ecx
720; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
721; X86-NEXT:    shrl %eax
722; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
723; X86-NEXT:    leal (%eax,%ecx,2), %eax
724; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
725; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
726; X86-NEXT:    bswapl %eax
727; X86-NEXT:    movl %eax, %ecx
728; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
729; X86-NEXT:    shll $4, %ecx
730; X86-NEXT:    shrl $4, %eax
731; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
732; X86-NEXT:    orl %ecx, %eax
733; X86-NEXT:    movl %eax, %ecx
734; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
735; X86-NEXT:    shrl $2, %eax
736; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
737; X86-NEXT:    leal (%eax,%ecx,4), %eax
738; X86-NEXT:    movl %eax, %ecx
739; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
740; X86-NEXT:    shrl %eax
741; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
742; X86-NEXT:    leal (%eax,%ecx,2), %eax
743; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
744; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
745; X86-NEXT:    bswapl %eax
746; X86-NEXT:    movl %eax, %ecx
747; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
748; X86-NEXT:    shll $4, %ecx
749; X86-NEXT:    shrl $4, %eax
750; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
751; X86-NEXT:    orl %ecx, %eax
752; X86-NEXT:    movl %eax, %ecx
753; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
754; X86-NEXT:    shrl $2, %eax
755; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
756; X86-NEXT:    leal (%eax,%ecx,4), %eax
757; X86-NEXT:    movl %eax, %ecx
758; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
759; X86-NEXT:    shrl %eax
760; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
761; X86-NEXT:    leal (%eax,%ecx,2), %eax
762; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
763; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
764; X86-NEXT:    bswapl %eax
765; X86-NEXT:    movl %eax, %ecx
766; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
767; X86-NEXT:    shll $4, %ecx
768; X86-NEXT:    shrl $4, %eax
769; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
770; X86-NEXT:    orl %ecx, %eax
771; X86-NEXT:    movl %eax, %ecx
772; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
773; X86-NEXT:    shrl $2, %eax
774; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
775; X86-NEXT:    leal (%eax,%ecx,4), %eax
776; X86-NEXT:    movl %eax, %ecx
777; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
778; X86-NEXT:    shrl %eax
779; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
780; X86-NEXT:    leal (%eax,%ecx,2), %eax
781; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
782; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
783; X86-NEXT:    bswapl %eax
784; X86-NEXT:    movl %eax, %ecx
785; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
786; X86-NEXT:    shll $4, %ecx
787; X86-NEXT:    shrl $4, %eax
788; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
789; X86-NEXT:    orl %ecx, %eax
790; X86-NEXT:    movl %eax, %ecx
791; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
792; X86-NEXT:    shrl $2, %eax
793; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
794; X86-NEXT:    leal (%eax,%ecx,4), %eax
795; X86-NEXT:    movl %eax, %ecx
796; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
797; X86-NEXT:    shrl %eax
798; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
799; X86-NEXT:    leal (%eax,%ecx,2), %eax
800; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
801; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
802; X86-NEXT:    bswapl %eax
803; X86-NEXT:    movl %eax, %ecx
804; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
805; X86-NEXT:    shll $4, %ecx
806; X86-NEXT:    shrl $4, %eax
807; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
808; X86-NEXT:    orl %ecx, %eax
809; X86-NEXT:    movl %eax, %ecx
810; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
811; X86-NEXT:    shrl $2, %eax
812; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
813; X86-NEXT:    leal (%eax,%ecx,4), %eax
814; X86-NEXT:    movl %eax, %ecx
815; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
816; X86-NEXT:    shrl %eax
817; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
818; X86-NEXT:    leal (%eax,%ecx,2), %eax
819; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
820; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
821; X86-NEXT:    bswapl %eax
822; X86-NEXT:    movl %eax, %ecx
823; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
824; X86-NEXT:    shll $4, %ecx
825; X86-NEXT:    shrl $4, %eax
826; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
827; X86-NEXT:    orl %ecx, %eax
828; X86-NEXT:    movl %eax, %ecx
829; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
830; X86-NEXT:    shrl $2, %eax
831; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
832; X86-NEXT:    leal (%eax,%ecx,4), %eax
833; X86-NEXT:    movl %eax, %ecx
834; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
835; X86-NEXT:    shrl %eax
836; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
837; X86-NEXT:    leal (%eax,%ecx,2), %eax
838; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
839; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
840; X86-NEXT:    bswapl %eax
841; X86-NEXT:    movl %eax, %ecx
842; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
843; X86-NEXT:    shll $4, %ecx
844; X86-NEXT:    shrl $4, %eax
845; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
846; X86-NEXT:    orl %ecx, %eax
847; X86-NEXT:    movl %eax, %ecx
848; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
849; X86-NEXT:    shrl $2, %eax
850; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
851; X86-NEXT:    leal (%eax,%ecx,4), %eax
852; X86-NEXT:    movl %eax, %ecx
853; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
854; X86-NEXT:    shrl %eax
855; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
856; X86-NEXT:    leal (%eax,%ecx,2), %eax
857; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
858; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
859; X86-NEXT:    bswapl %eax
860; X86-NEXT:    movl %eax, %ecx
861; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
862; X86-NEXT:    shll $4, %ecx
863; X86-NEXT:    shrl $4, %eax
864; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
865; X86-NEXT:    orl %ecx, %eax
866; X86-NEXT:    movl %eax, %ecx
867; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
868; X86-NEXT:    shrl $2, %eax
869; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
870; X86-NEXT:    leal (%eax,%ecx,4), %eax
871; X86-NEXT:    movl %eax, %ecx
872; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
873; X86-NEXT:    shrl %eax
874; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
875; X86-NEXT:    leal (%eax,%ecx,2), %eax
876; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
877; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
878; X86-NEXT:    bswapl %eax
879; X86-NEXT:    movl %eax, %ecx
880; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
881; X86-NEXT:    shll $4, %ecx
882; X86-NEXT:    shrl $4, %eax
883; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
884; X86-NEXT:    orl %ecx, %eax
885; X86-NEXT:    movl %eax, %ecx
886; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
887; X86-NEXT:    shrl $2, %eax
888; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
889; X86-NEXT:    leal (%eax,%ecx,4), %eax
890; X86-NEXT:    movl %eax, %ecx
891; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
892; X86-NEXT:    shrl %eax
893; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
894; X86-NEXT:    leal (%eax,%ecx,2), %eax
895; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
896; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
897; X86-NEXT:    bswapl %eax
898; X86-NEXT:    movl %eax, %ecx
899; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
900; X86-NEXT:    shll $4, %ecx
901; X86-NEXT:    shrl $4, %eax
902; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
903; X86-NEXT:    orl %ecx, %eax
904; X86-NEXT:    movl %eax, %ecx
905; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
906; X86-NEXT:    shrl $2, %eax
907; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
908; X86-NEXT:    leal (%eax,%ecx,4), %eax
909; X86-NEXT:    movl %eax, %ecx
910; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
911; X86-NEXT:    shrl %eax
912; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
913; X86-NEXT:    leal (%eax,%ecx,2), %edi
914; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
915; X86-NEXT:    bswapl %eax
916; X86-NEXT:    movl %eax, %ecx
917; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
918; X86-NEXT:    shll $4, %ecx
919; X86-NEXT:    shrl $4, %eax
920; X86-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
921; X86-NEXT:    orl %ecx, %eax
922; X86-NEXT:    movl %eax, %ecx
923; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
924; X86-NEXT:    shrl $2, %eax
925; X86-NEXT:    andl $858993459, %eax # imm = 0x33333333
926; X86-NEXT:    leal (%eax,%ecx,4), %eax
927; X86-NEXT:    movl %eax, %ecx
928; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
929; X86-NEXT:    shrl %eax
930; X86-NEXT:    andl $1431655765, %eax # imm = 0x55555555
931; X86-NEXT:    leal (%eax,%ecx,2), %edx
932; X86-NEXT:    movl (%esp), %esi # 4-byte Reload
933; X86-NEXT:    shrdl $16, %ebx, %esi
934; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
935; X86-NEXT:    shrdl $16, %eax, %ebx
936; X86-NEXT:    movl %ebx, (%esp) # 4-byte Spill
937; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
938; X86-NEXT:    shrdl $16, %ecx, %eax
939; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
940; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
941; X86-NEXT:    shrdl $16, %eax, %ecx
942; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
943; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
944; X86-NEXT:    shrdl $16, %ecx, %eax
945; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
946; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
947; X86-NEXT:    shrdl $16, %eax, %ecx
948; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
949; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
950; X86-NEXT:    shrdl $16, %ecx, %eax
951; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
952; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
953; X86-NEXT:    shrdl $16, %eax, %ecx
954; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
955; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
956; X86-NEXT:    shrdl $16, %ecx, %eax
957; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
958; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
959; X86-NEXT:    shrdl $16, %eax, %ecx
960; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
961; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
962; X86-NEXT:    shrdl $16, %ecx, %eax
963; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
964; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
965; X86-NEXT:    shrdl $16, %ebp, %ecx
966; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
967; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
968; X86-NEXT:    shrdl $16, %ebx, %ebp
969; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
970; X86-NEXT:    shrdl $16, %eax, %ebx
971; X86-NEXT:    shrdl $16, %edi, %eax
972; X86-NEXT:    movl %eax, %ecx
973; X86-NEXT:    shrdl $16, %edx, %edi
974; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
975; X86-NEXT:    movl %edi, 60(%eax)
976; X86-NEXT:    movl %ecx, 56(%eax)
977; X86-NEXT:    movl %ebx, 52(%eax)
978; X86-NEXT:    movl %ebp, 48(%eax)
979; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
980; X86-NEXT:    movl %ecx, 44(%eax)
981; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
982; X86-NEXT:    movl %ecx, 40(%eax)
983; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
984; X86-NEXT:    movl %ecx, 36(%eax)
985; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
986; X86-NEXT:    movl %ecx, 32(%eax)
987; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
988; X86-NEXT:    movl %ecx, 28(%eax)
989; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
990; X86-NEXT:    movl %ecx, 24(%eax)
991; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
992; X86-NEXT:    movl %ecx, 20(%eax)
993; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
994; X86-NEXT:    movl %ecx, 16(%eax)
995; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
996; X86-NEXT:    movl %ecx, 12(%eax)
997; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
998; X86-NEXT:    movl %ecx, 8(%eax)
999; X86-NEXT:    movl (%esp), %ecx # 4-byte Reload
1000; X86-NEXT:    movl %ecx, 4(%eax)
1001; X86-NEXT:    movl %esi, (%eax)
1002; X86-NEXT:    shrl $16, %edx
1003; X86-NEXT:    movw %dx, 64(%eax)
1004; X86-NEXT:    addl $56, %esp
1005; X86-NEXT:    popl %esi
1006; X86-NEXT:    popl %edi
1007; X86-NEXT:    popl %ebx
1008; X86-NEXT:    popl %ebp
1009; X86-NEXT:    retl $4
1010;
1011; X64-LABEL: large_promotion:
1012; X64:       # %bb.0:
1013; X64-NEXT:    pushq %rbp
1014; X64-NEXT:    pushq %r15
1015; X64-NEXT:    pushq %r14
1016; X64-NEXT:    pushq %r13
1017; X64-NEXT:    pushq %r12
1018; X64-NEXT:    pushq %rbx
1019; X64-NEXT:    movq %rdi, %rax
1020; X64-NEXT:    movq {{[0-9]+}}(%rsp), %r12
1021; X64-NEXT:    movq {{[0-9]+}}(%rsp), %r15
1022; X64-NEXT:    movq {{[0-9]+}}(%rsp), %rbp
1023; X64-NEXT:    movq {{[0-9]+}}(%rsp), %rdi
1024; X64-NEXT:    bswapq %rdi
1025; X64-NEXT:    movq %rdi, %rbx
1026; X64-NEXT:    shrq $4, %rbx
1027; X64-NEXT:    movabsq $1085102592571150095, %r13 # imm = 0xF0F0F0F0F0F0F0F
1028; X64-NEXT:    andq %r13, %rbx
1029; X64-NEXT:    andq %r13, %rdi
1030; X64-NEXT:    shlq $4, %rdi
1031; X64-NEXT:    orq %rbx, %rdi
1032; X64-NEXT:    movabsq $3689348814741910323, %r11 # imm = 0x3333333333333333
1033; X64-NEXT:    movq %rdi, %rbx
1034; X64-NEXT:    andq %r11, %rbx
1035; X64-NEXT:    shrq $2, %rdi
1036; X64-NEXT:    andq %r11, %rdi
1037; X64-NEXT:    leaq (%rdi,%rbx,4), %rdi
1038; X64-NEXT:    movabsq $6148820866244280320, %r10 # imm = 0x5555000000000000
1039; X64-NEXT:    movq %rdi, %rbx
1040; X64-NEXT:    andq %r10, %rbx
1041; X64-NEXT:    shrq %rdi
1042; X64-NEXT:    andq %r10, %rdi
1043; X64-NEXT:    leaq (%rdi,%rbx,2), %r10
1044; X64-NEXT:    bswapq %rbp
1045; X64-NEXT:    movq %rbp, %rdi
1046; X64-NEXT:    shrq $4, %rdi
1047; X64-NEXT:    andq %r13, %rdi
1048; X64-NEXT:    andq %r13, %rbp
1049; X64-NEXT:    shlq $4, %rbp
1050; X64-NEXT:    orq %rdi, %rbp
1051; X64-NEXT:    movq %rbp, %rdi
1052; X64-NEXT:    andq %r11, %rdi
1053; X64-NEXT:    shrq $2, %rbp
1054; X64-NEXT:    andq %r11, %rbp
1055; X64-NEXT:    leaq (%rbp,%rdi,4), %rdi
1056; X64-NEXT:    movabsq $6148914691236517205, %rbp # imm = 0x5555555555555555
1057; X64-NEXT:    movq %rdi, %rbx
1058; X64-NEXT:    andq %rbp, %rbx
1059; X64-NEXT:    shrq %rdi
1060; X64-NEXT:    andq %rbp, %rdi
1061; X64-NEXT:    leaq (%rdi,%rbx,2), %r14
1062; X64-NEXT:    shrdq $48, %r14, %r10
1063; X64-NEXT:    bswapq %r15
1064; X64-NEXT:    movq %r15, %rdi
1065; X64-NEXT:    shrq $4, %rdi
1066; X64-NEXT:    andq %r13, %rdi
1067; X64-NEXT:    andq %r13, %r15
1068; X64-NEXT:    shlq $4, %r15
1069; X64-NEXT:    orq %rdi, %r15
1070; X64-NEXT:    movq %r15, %rdi
1071; X64-NEXT:    andq %r11, %rdi
1072; X64-NEXT:    shrq $2, %r15
1073; X64-NEXT:    andq %r11, %r15
1074; X64-NEXT:    leaq (%r15,%rdi,4), %rdi
1075; X64-NEXT:    movq %rdi, %rbx
1076; X64-NEXT:    andq %rbp, %rbx
1077; X64-NEXT:    shrq %rdi
1078; X64-NEXT:    andq %rbp, %rdi
1079; X64-NEXT:    leaq (%rdi,%rbx,2), %r15
1080; X64-NEXT:    shrdq $48, %r15, %r14
1081; X64-NEXT:    bswapq %r12
1082; X64-NEXT:    movq %r12, %rdi
1083; X64-NEXT:    shrq $4, %rdi
1084; X64-NEXT:    andq %r13, %rdi
1085; X64-NEXT:    andq %r13, %r12
1086; X64-NEXT:    shlq $4, %r12
1087; X64-NEXT:    orq %rdi, %r12
1088; X64-NEXT:    movq %r12, %rdi
1089; X64-NEXT:    andq %r11, %rdi
1090; X64-NEXT:    shrq $2, %r12
1091; X64-NEXT:    andq %r11, %r12
1092; X64-NEXT:    leaq (%r12,%rdi,4), %rdi
1093; X64-NEXT:    movq %rdi, %rbx
1094; X64-NEXT:    andq %rbp, %rbx
1095; X64-NEXT:    shrq %rdi
1096; X64-NEXT:    andq %rbp, %rdi
1097; X64-NEXT:    leaq (%rdi,%rbx,2), %r12
1098; X64-NEXT:    shrdq $48, %r12, %r15
1099; X64-NEXT:    bswapq %r9
1100; X64-NEXT:    movq %r9, %rdi
1101; X64-NEXT:    shrq $4, %rdi
1102; X64-NEXT:    andq %r13, %rdi
1103; X64-NEXT:    andq %r13, %r9
1104; X64-NEXT:    shlq $4, %r9
1105; X64-NEXT:    orq %rdi, %r9
1106; X64-NEXT:    movq %r9, %rdi
1107; X64-NEXT:    andq %r11, %rdi
1108; X64-NEXT:    shrq $2, %r9
1109; X64-NEXT:    andq %r11, %r9
1110; X64-NEXT:    leaq (%r9,%rdi,4), %rdi
1111; X64-NEXT:    movq %rdi, %rbx
1112; X64-NEXT:    andq %rbp, %rbx
1113; X64-NEXT:    shrq %rdi
1114; X64-NEXT:    andq %rbp, %rdi
1115; X64-NEXT:    leaq (%rdi,%rbx,2), %r9
1116; X64-NEXT:    shrdq $48, %r9, %r12
1117; X64-NEXT:    bswapq %r8
1118; X64-NEXT:    movq %r8, %rdi
1119; X64-NEXT:    shrq $4, %rdi
1120; X64-NEXT:    andq %r13, %rdi
1121; X64-NEXT:    andq %r13, %r8
1122; X64-NEXT:    shlq $4, %r8
1123; X64-NEXT:    orq %rdi, %r8
1124; X64-NEXT:    movq %r8, %rdi
1125; X64-NEXT:    andq %r11, %rdi
1126; X64-NEXT:    shrq $2, %r8
1127; X64-NEXT:    andq %r11, %r8
1128; X64-NEXT:    leaq (%r8,%rdi,4), %rdi
1129; X64-NEXT:    movq %rdi, %rbx
1130; X64-NEXT:    andq %rbp, %rbx
1131; X64-NEXT:    shrq %rdi
1132; X64-NEXT:    andq %rbp, %rdi
1133; X64-NEXT:    leaq (%rdi,%rbx,2), %rdi
1134; X64-NEXT:    shrdq $48, %rdi, %r9
1135; X64-NEXT:    bswapq %rcx
1136; X64-NEXT:    movq %rcx, %rbx
1137; X64-NEXT:    shrq $4, %rbx
1138; X64-NEXT:    andq %r13, %rbx
1139; X64-NEXT:    andq %r13, %rcx
1140; X64-NEXT:    shlq $4, %rcx
1141; X64-NEXT:    orq %rbx, %rcx
1142; X64-NEXT:    movq %rcx, %rbx
1143; X64-NEXT:    andq %r11, %rbx
1144; X64-NEXT:    shrq $2, %rcx
1145; X64-NEXT:    andq %r11, %rcx
1146; X64-NEXT:    leaq (%rcx,%rbx,4), %rcx
1147; X64-NEXT:    movq %rcx, %rbx
1148; X64-NEXT:    andq %rbp, %rbx
1149; X64-NEXT:    shrq %rcx
1150; X64-NEXT:    andq %rbp, %rcx
1151; X64-NEXT:    leaq (%rcx,%rbx,2), %rcx
1152; X64-NEXT:    shrdq $48, %rcx, %rdi
1153; X64-NEXT:    bswapq %rdx
1154; X64-NEXT:    movq %rdx, %rbx
1155; X64-NEXT:    shrq $4, %rbx
1156; X64-NEXT:    andq %r13, %rbx
1157; X64-NEXT:    andq %r13, %rdx
1158; X64-NEXT:    shlq $4, %rdx
1159; X64-NEXT:    orq %rbx, %rdx
1160; X64-NEXT:    movq %rdx, %rbx
1161; X64-NEXT:    andq %r11, %rbx
1162; X64-NEXT:    shrq $2, %rdx
1163; X64-NEXT:    andq %r11, %rdx
1164; X64-NEXT:    leaq (%rdx,%rbx,4), %rdx
1165; X64-NEXT:    movq %rdx, %rbx
1166; X64-NEXT:    andq %rbp, %rbx
1167; X64-NEXT:    shrq %rdx
1168; X64-NEXT:    andq %rbp, %rdx
1169; X64-NEXT:    leaq (%rdx,%rbx,2), %rdx
1170; X64-NEXT:    shrdq $48, %rdx, %rcx
1171; X64-NEXT:    bswapq %rsi
1172; X64-NEXT:    movq %rsi, %rbx
1173; X64-NEXT:    shrq $4, %rbx
1174; X64-NEXT:    andq %r13, %rbx
1175; X64-NEXT:    andq %r13, %rsi
1176; X64-NEXT:    shlq $4, %rsi
1177; X64-NEXT:    orq %rbx, %rsi
1178; X64-NEXT:    movq %rsi, %rbx
1179; X64-NEXT:    andq %r11, %rbx
1180; X64-NEXT:    shrq $2, %rsi
1181; X64-NEXT:    andq %r11, %rsi
1182; X64-NEXT:    leaq (%rsi,%rbx,4), %rsi
1183; X64-NEXT:    movq %rsi, %rbx
1184; X64-NEXT:    andq %rbp, %rbx
1185; X64-NEXT:    shrq %rsi
1186; X64-NEXT:    andq %rbp, %rsi
1187; X64-NEXT:    leaq (%rsi,%rbx,2), %rsi
1188; X64-NEXT:    shrdq $48, %rsi, %rdx
1189; X64-NEXT:    shrq $48, %rsi
1190; X64-NEXT:    movq %rdx, 56(%rax)
1191; X64-NEXT:    movq %rcx, 48(%rax)
1192; X64-NEXT:    movq %rdi, 40(%rax)
1193; X64-NEXT:    movq %r9, 32(%rax)
1194; X64-NEXT:    movq %r12, 24(%rax)
1195; X64-NEXT:    movq %r15, 16(%rax)
1196; X64-NEXT:    movq %r14, 8(%rax)
1197; X64-NEXT:    movq %r10, (%rax)
1198; X64-NEXT:    movw %si, 64(%rax)
1199; X64-NEXT:    popq %rbx
1200; X64-NEXT:    popq %r12
1201; X64-NEXT:    popq %r13
1202; X64-NEXT:    popq %r14
1203; X64-NEXT:    popq %r15
1204; X64-NEXT:    popq %rbp
1205; X64-NEXT:    retq
1206;
1207; X86XOP-LABEL: large_promotion:
1208; X86XOP:       # %bb.0:
1209; X86XOP-NEXT:    pushl %ebp
1210; X86XOP-NEXT:    pushl %ebx
1211; X86XOP-NEXT:    pushl %edi
1212; X86XOP-NEXT:    pushl %esi
1213; X86XOP-NEXT:    subl $44, %esp
1214; X86XOP-NEXT:    vmovdqa {{.*#+}} xmm0 = [87,86,85,84,83,82,81,80,95,94,93,92,91,90,89,88]
1215; X86XOP-NEXT:    vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
1216; X86XOP-NEXT:    vpperm %xmm0, %xmm1, %xmm0, %xmm1
1217; X86XOP-NEXT:    vpextrd $1, %xmm1, %eax
1218; X86XOP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
1219; X86XOP-NEXT:    vpperm %xmm0, %xmm1, %xmm0, %xmm1
1220; X86XOP-NEXT:    vmovd %xmm1, %ecx
1221; X86XOP-NEXT:    shrdl $16, %ecx, %eax
1222; X86XOP-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1223; X86XOP-NEXT:    vpextrd $1, %xmm1, %eax
1224; X86XOP-NEXT:    shrdl $16, %eax, %ecx
1225; X86XOP-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1226; X86XOP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
1227; X86XOP-NEXT:    vpperm %xmm0, %xmm1, %xmm0, %xmm1
1228; X86XOP-NEXT:    vmovd %xmm1, %ecx
1229; X86XOP-NEXT:    shrdl $16, %ecx, %eax
1230; X86XOP-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1231; X86XOP-NEXT:    vpextrd $1, %xmm1, %eax
1232; X86XOP-NEXT:    shrdl $16, %eax, %ecx
1233; X86XOP-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1234; X86XOP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
1235; X86XOP-NEXT:    vpperm %xmm0, %xmm1, %xmm0, %xmm1
1236; X86XOP-NEXT:    vmovd %xmm1, %ecx
1237; X86XOP-NEXT:    shrdl $16, %ecx, %eax
1238; X86XOP-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1239; X86XOP-NEXT:    vpextrd $1, %xmm1, %eax
1240; X86XOP-NEXT:    shrdl $16, %eax, %ecx
1241; X86XOP-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1242; X86XOP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
1243; X86XOP-NEXT:    vpperm %xmm0, %xmm1, %xmm0, %xmm1
1244; X86XOP-NEXT:    vmovd %xmm1, %ecx
1245; X86XOP-NEXT:    shrdl $16, %ecx, %eax
1246; X86XOP-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1247; X86XOP-NEXT:    vpextrd $1, %xmm1, %eax
1248; X86XOP-NEXT:    shrdl $16, %eax, %ecx
1249; X86XOP-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1250; X86XOP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
1251; X86XOP-NEXT:    vpperm %xmm0, %xmm1, %xmm0, %xmm1
1252; X86XOP-NEXT:    vmovd %xmm1, %ecx
1253; X86XOP-NEXT:    shrdl $16, %ecx, %eax
1254; X86XOP-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1255; X86XOP-NEXT:    vpextrd $1, %xmm1, %eax
1256; X86XOP-NEXT:    shrdl $16, %eax, %ecx
1257; X86XOP-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
1258; X86XOP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
1259; X86XOP-NEXT:    vpperm %xmm0, %xmm1, %xmm0, %xmm1
1260; X86XOP-NEXT:    vmovd %xmm1, %ebp
1261; X86XOP-NEXT:    shrdl $16, %ebp, %eax
1262; X86XOP-NEXT:    movl %eax, (%esp) # 4-byte Spill
1263; X86XOP-NEXT:    vpextrd $1, %xmm1, %ebx
1264; X86XOP-NEXT:    shrdl $16, %ebx, %ebp
1265; X86XOP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
1266; X86XOP-NEXT:    vpperm %xmm0, %xmm1, %xmm0, %xmm1
1267; X86XOP-NEXT:    vmovd %xmm1, %esi
1268; X86XOP-NEXT:    shrdl $16, %esi, %ebx
1269; X86XOP-NEXT:    vpextrd $1, %xmm1, %edx
1270; X86XOP-NEXT:    shrdl $16, %edx, %esi
1271; X86XOP-NEXT:    vmovq {{.*#+}} xmm1 = mem[0],zero
1272; X86XOP-NEXT:    vpperm %xmm0, %xmm1, %xmm0, %xmm0
1273; X86XOP-NEXT:    vmovd %xmm0, %ecx
1274; X86XOP-NEXT:    shrdl $16, %ecx, %edx
1275; X86XOP-NEXT:    vpextrd $1, %xmm0, %edi
1276; X86XOP-NEXT:    shrdl $16, %edi, %ecx
1277; X86XOP-NEXT:    movl {{[0-9]+}}(%esp), %eax
1278; X86XOP-NEXT:    movl %ecx, 60(%eax)
1279; X86XOP-NEXT:    movl %edx, 56(%eax)
1280; X86XOP-NEXT:    movl %esi, 52(%eax)
1281; X86XOP-NEXT:    movl %ebx, 48(%eax)
1282; X86XOP-NEXT:    movl %ebp, 44(%eax)
1283; X86XOP-NEXT:    movl (%esp), %ecx # 4-byte Reload
1284; X86XOP-NEXT:    movl %ecx, 40(%eax)
1285; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1286; X86XOP-NEXT:    movl %ecx, 36(%eax)
1287; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1288; X86XOP-NEXT:    movl %ecx, 32(%eax)
1289; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1290; X86XOP-NEXT:    movl %ecx, 28(%eax)
1291; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1292; X86XOP-NEXT:    movl %ecx, 24(%eax)
1293; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1294; X86XOP-NEXT:    movl %ecx, 20(%eax)
1295; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1296; X86XOP-NEXT:    movl %ecx, 16(%eax)
1297; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1298; X86XOP-NEXT:    movl %ecx, 12(%eax)
1299; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1300; X86XOP-NEXT:    movl %ecx, 8(%eax)
1301; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1302; X86XOP-NEXT:    movl %ecx, 4(%eax)
1303; X86XOP-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
1304; X86XOP-NEXT:    movl %ecx, (%eax)
1305; X86XOP-NEXT:    shrl $16, %edi
1306; X86XOP-NEXT:    movw %di, 64(%eax)
1307; X86XOP-NEXT:    addl $44, %esp
1308; X86XOP-NEXT:    popl %esi
1309; X86XOP-NEXT:    popl %edi
1310; X86XOP-NEXT:    popl %ebx
1311; X86XOP-NEXT:    popl %ebp
1312; X86XOP-NEXT:    retl $4
1313  %Z = call i528 @llvm.bitreverse.i528(i528 %A)
1314  ret i528 %Z
1315}
1316declare i528 @llvm.bitreverse.i528(i528)
1317