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
5; These tests just check that the plumbing is in place for @llvm.bitreverse. The
6; actual output is massive at the moment as llvm.bitreverse is not yet legal.
7
8declare <2 x i16> @llvm.bitreverse.v2i16(<2 x i16>) readnone
9
10define <2 x i16> @test_bitreverse_v2i16(<2 x i16> %a) nounwind {
11; X86-LABEL: test_bitreverse_v2i16:
12; X86:       # %bb.0:
13; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
14; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
15; X86-NEXT:    rolw $8, %ax
16; X86-NEXT:    movl %eax, %edx
17; X86-NEXT:    andl $3855, %edx # imm = 0xF0F
18; X86-NEXT:    shll $4, %edx
19; X86-NEXT:    andl $61680, %eax # imm = 0xF0F0
20; X86-NEXT:    shrl $4, %eax
21; X86-NEXT:    orl %edx, %eax
22; X86-NEXT:    movl %eax, %edx
23; X86-NEXT:    andl $13107, %edx # imm = 0x3333
24; X86-NEXT:    andl $52428, %eax # imm = 0xCCCC
25; X86-NEXT:    shrl $2, %eax
26; X86-NEXT:    leal (%eax,%edx,4), %eax
27; X86-NEXT:    movl %eax, %edx
28; X86-NEXT:    andl $21845, %edx # imm = 0x5555
29; X86-NEXT:    andl $43690, %eax # imm = 0xAAAA
30; X86-NEXT:    shrl %eax
31; X86-NEXT:    leal (%eax,%edx,2), %eax
32; X86-NEXT:    rolw $8, %cx
33; X86-NEXT:    movl %ecx, %edx
34; X86-NEXT:    andl $3855, %edx # imm = 0xF0F
35; X86-NEXT:    shll $4, %edx
36; X86-NEXT:    andl $61680, %ecx # imm = 0xF0F0
37; X86-NEXT:    shrl $4, %ecx
38; X86-NEXT:    orl %edx, %ecx
39; X86-NEXT:    movl %ecx, %edx
40; X86-NEXT:    andl $13107, %edx # imm = 0x3333
41; X86-NEXT:    andl $52428, %ecx # imm = 0xCCCC
42; X86-NEXT:    shrl $2, %ecx
43; X86-NEXT:    leal (%ecx,%edx,4), %ecx
44; X86-NEXT:    movl %ecx, %edx
45; X86-NEXT:    andl $21845, %edx # imm = 0x5555
46; X86-NEXT:    andl $43690, %ecx # imm = 0xAAAA
47; X86-NEXT:    shrl %ecx
48; X86-NEXT:    leal (%ecx,%edx,2), %edx
49; X86-NEXT:    # kill: def $ax killed $ax killed $eax
50; X86-NEXT:    # kill: def $dx killed $dx killed $edx
51; X86-NEXT:    retl
52;
53; X64-LABEL: test_bitreverse_v2i16:
54; X64:       # %bb.0:
55; X64-NEXT:    movdqa %xmm0, %xmm1
56; X64-NEXT:    psrlw $8, %xmm1
57; X64-NEXT:    psllw $8, %xmm0
58; X64-NEXT:    por %xmm1, %xmm0
59; X64-NEXT:    movdqa %xmm0, %xmm1
60; X64-NEXT:    psllw $4, %xmm1
61; X64-NEXT:    pand {{.*}}(%rip), %xmm1
62; X64-NEXT:    psrlw $4, %xmm0
63; X64-NEXT:    pand {{.*}}(%rip), %xmm0
64; X64-NEXT:    por %xmm1, %xmm0
65; X64-NEXT:    movdqa {{.*#+}} xmm1 = [51,51,51,51,51,51,51,51,51,51,51,51,51,51,51,51]
66; X64-NEXT:    pand %xmm0, %xmm1
67; X64-NEXT:    psllw $2, %xmm1
68; X64-NEXT:    pand {{.*}}(%rip), %xmm0
69; X64-NEXT:    psrlw $2, %xmm0
70; X64-NEXT:    por %xmm1, %xmm0
71; X64-NEXT:    movdqa {{.*#+}} xmm1 = [85,85,85,85,85,85,85,85,85,85,85,85,85,85,85,85]
72; X64-NEXT:    pand %xmm0, %xmm1
73; X64-NEXT:    paddb %xmm1, %xmm1
74; X64-NEXT:    pand {{.*}}(%rip), %xmm0
75; X64-NEXT:    psrlw $1, %xmm0
76; X64-NEXT:    por %xmm1, %xmm0
77; X64-NEXT:    retq
78  %b = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> %a)
79  ret <2 x i16> %b
80}
81
82declare i64 @llvm.bitreverse.i64(i64) readnone
83
84define i64 @test_bitreverse_i64(i64 %a) nounwind {
85; X86-LABEL: test_bitreverse_i64:
86; X86:       # %bb.0:
87; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
88; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
89; X86-NEXT:    bswapl %eax
90; X86-NEXT:    movl %eax, %edx
91; X86-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
92; X86-NEXT:    shll $4, %edx
93; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
94; X86-NEXT:    shrl $4, %eax
95; X86-NEXT:    orl %edx, %eax
96; X86-NEXT:    movl %eax, %edx
97; X86-NEXT:    andl $858993459, %edx # imm = 0x33333333
98; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
99; X86-NEXT:    shrl $2, %eax
100; X86-NEXT:    leal (%eax,%edx,4), %eax
101; X86-NEXT:    movl %eax, %edx
102; X86-NEXT:    andl $1431655765, %edx # imm = 0x55555555
103; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
104; X86-NEXT:    shrl %eax
105; X86-NEXT:    leal (%eax,%edx,2), %eax
106; X86-NEXT:    bswapl %ecx
107; X86-NEXT:    movl %ecx, %edx
108; X86-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
109; X86-NEXT:    shll $4, %edx
110; X86-NEXT:    andl $-252645136, %ecx # imm = 0xF0F0F0F0
111; X86-NEXT:    shrl $4, %ecx
112; X86-NEXT:    orl %edx, %ecx
113; X86-NEXT:    movl %ecx, %edx
114; X86-NEXT:    andl $858993459, %edx # imm = 0x33333333
115; X86-NEXT:    andl $-858993460, %ecx # imm = 0xCCCCCCCC
116; X86-NEXT:    shrl $2, %ecx
117; X86-NEXT:    leal (%ecx,%edx,4), %ecx
118; X86-NEXT:    movl %ecx, %edx
119; X86-NEXT:    andl $1431655765, %edx # imm = 0x55555555
120; X86-NEXT:    andl $-1431655766, %ecx # imm = 0xAAAAAAAA
121; X86-NEXT:    shrl %ecx
122; X86-NEXT:    leal (%ecx,%edx,2), %edx
123; X86-NEXT:    retl
124;
125; X64-LABEL: test_bitreverse_i64:
126; X64:       # %bb.0:
127; X64-NEXT:    bswapq %rdi
128; X64-NEXT:    movabsq $1085102592571150095, %rax # imm = 0xF0F0F0F0F0F0F0F
129; X64-NEXT:    andq %rdi, %rax
130; X64-NEXT:    shlq $4, %rax
131; X64-NEXT:    movabsq $-1085102592571150096, %rcx # imm = 0xF0F0F0F0F0F0F0F0
132; X64-NEXT:    andq %rdi, %rcx
133; X64-NEXT:    shrq $4, %rcx
134; X64-NEXT:    orq %rax, %rcx
135; X64-NEXT:    movabsq $3689348814741910323, %rax # imm = 0x3333333333333333
136; X64-NEXT:    andq %rcx, %rax
137; X64-NEXT:    movabsq $-3689348814741910324, %rdx # imm = 0xCCCCCCCCCCCCCCCC
138; X64-NEXT:    andq %rcx, %rdx
139; X64-NEXT:    shrq $2, %rdx
140; X64-NEXT:    leaq (%rdx,%rax,4), %rax
141; X64-NEXT:    movabsq $6148914691236517205, %rcx # imm = 0x5555555555555555
142; X64-NEXT:    andq %rax, %rcx
143; X64-NEXT:    movabsq $-6148914691236517206, %rdx # imm = 0xAAAAAAAAAAAAAAAA
144; X64-NEXT:    andq %rax, %rdx
145; X64-NEXT:    shrq %rdx
146; X64-NEXT:    leaq (%rdx,%rcx,2), %rax
147; X64-NEXT:    retq
148  %b = call i64 @llvm.bitreverse.i64(i64 %a)
149  ret i64 %b
150}
151
152declare i32 @llvm.bitreverse.i32(i32) readnone
153
154define i32 @test_bitreverse_i32(i32 %a) nounwind {
155; X86-LABEL: test_bitreverse_i32:
156; X86:       # %bb.0:
157; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
158; X86-NEXT:    bswapl %eax
159; X86-NEXT:    movl %eax, %ecx
160; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
161; X86-NEXT:    shll $4, %ecx
162; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
163; X86-NEXT:    shrl $4, %eax
164; X86-NEXT:    orl %ecx, %eax
165; X86-NEXT:    movl %eax, %ecx
166; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
167; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
168; X86-NEXT:    shrl $2, %eax
169; X86-NEXT:    leal (%eax,%ecx,4), %eax
170; X86-NEXT:    movl %eax, %ecx
171; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
172; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
173; X86-NEXT:    shrl %eax
174; X86-NEXT:    leal (%eax,%ecx,2), %eax
175; X86-NEXT:    retl
176;
177; X64-LABEL: test_bitreverse_i32:
178; X64:       # %bb.0:
179; X64-NEXT:    # kill: def $edi killed $edi def $rdi
180; X64-NEXT:    bswapl %edi
181; X64-NEXT:    movl %edi, %eax
182; X64-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
183; X64-NEXT:    shll $4, %eax
184; X64-NEXT:    andl $-252645136, %edi # imm = 0xF0F0F0F0
185; X64-NEXT:    shrl $4, %edi
186; X64-NEXT:    orl %eax, %edi
187; X64-NEXT:    movl %edi, %eax
188; X64-NEXT:    andl $858993459, %eax # imm = 0x33333333
189; X64-NEXT:    andl $-858993460, %edi # imm = 0xCCCCCCCC
190; X64-NEXT:    shrl $2, %edi
191; X64-NEXT:    leal (%rdi,%rax,4), %eax
192; X64-NEXT:    movl %eax, %ecx
193; X64-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
194; X64-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
195; X64-NEXT:    shrl %eax
196; X64-NEXT:    leal (%rax,%rcx,2), %eax
197; X64-NEXT:    retq
198  %b = call i32 @llvm.bitreverse.i32(i32 %a)
199  ret i32 %b
200}
201
202declare i24 @llvm.bitreverse.i24(i24) readnone
203
204define i24 @test_bitreverse_i24(i24 %a) nounwind {
205; X86-LABEL: test_bitreverse_i24:
206; X86:       # %bb.0:
207; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
208; X86-NEXT:    bswapl %eax
209; X86-NEXT:    movl %eax, %ecx
210; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
211; X86-NEXT:    shll $4, %ecx
212; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
213; X86-NEXT:    shrl $4, %eax
214; X86-NEXT:    orl %ecx, %eax
215; X86-NEXT:    movl %eax, %ecx
216; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
217; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
218; X86-NEXT:    shrl $2, %eax
219; X86-NEXT:    leal (%eax,%ecx,4), %eax
220; X86-NEXT:    movl %eax, %ecx
221; X86-NEXT:    andl $1431655680, %ecx # imm = 0x55555500
222; X86-NEXT:    andl $-1431655936, %eax # imm = 0xAAAAAA00
223; X86-NEXT:    shrl %eax
224; X86-NEXT:    leal (%eax,%ecx,2), %eax
225; X86-NEXT:    shrl $8, %eax
226; X86-NEXT:    retl
227;
228; X64-LABEL: test_bitreverse_i24:
229; X64:       # %bb.0:
230; X64-NEXT:    # kill: def $edi killed $edi def $rdi
231; X64-NEXT:    bswapl %edi
232; X64-NEXT:    movl %edi, %eax
233; X64-NEXT:    andl $252645135, %eax # imm = 0xF0F0F0F
234; X64-NEXT:    shll $4, %eax
235; X64-NEXT:    andl $-252645136, %edi # imm = 0xF0F0F0F0
236; X64-NEXT:    shrl $4, %edi
237; X64-NEXT:    orl %eax, %edi
238; X64-NEXT:    movl %edi, %eax
239; X64-NEXT:    andl $858993459, %eax # imm = 0x33333333
240; X64-NEXT:    andl $-858993460, %edi # imm = 0xCCCCCCCC
241; X64-NEXT:    shrl $2, %edi
242; X64-NEXT:    leal (%rdi,%rax,4), %eax
243; X64-NEXT:    movl %eax, %ecx
244; X64-NEXT:    andl $1431655680, %ecx # imm = 0x55555500
245; X64-NEXT:    andl $-1431655936, %eax # imm = 0xAAAAAA00
246; X64-NEXT:    shrl %eax
247; X64-NEXT:    leal (%rax,%rcx,2), %eax
248; X64-NEXT:    shrl $8, %eax
249; X64-NEXT:    retq
250  %b = call i24 @llvm.bitreverse.i24(i24 %a)
251  ret i24 %b
252}
253
254declare i16 @llvm.bitreverse.i16(i16) readnone
255
256define i16 @test_bitreverse_i16(i16 %a) nounwind {
257; X86-LABEL: test_bitreverse_i16:
258; X86:       # %bb.0:
259; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
260; X86-NEXT:    rolw $8, %ax
261; X86-NEXT:    movl %eax, %ecx
262; X86-NEXT:    andl $3855, %ecx # imm = 0xF0F
263; X86-NEXT:    shll $4, %ecx
264; X86-NEXT:    andl $61680, %eax # imm = 0xF0F0
265; X86-NEXT:    shrl $4, %eax
266; X86-NEXT:    orl %ecx, %eax
267; X86-NEXT:    movl %eax, %ecx
268; X86-NEXT:    andl $13107, %ecx # imm = 0x3333
269; X86-NEXT:    andl $52428, %eax # imm = 0xCCCC
270; X86-NEXT:    shrl $2, %eax
271; X86-NEXT:    leal (%eax,%ecx,4), %eax
272; X86-NEXT:    movl %eax, %ecx
273; X86-NEXT:    andl $21845, %ecx # imm = 0x5555
274; X86-NEXT:    andl $43690, %eax # imm = 0xAAAA
275; X86-NEXT:    shrl %eax
276; X86-NEXT:    leal (%eax,%ecx,2), %eax
277; X86-NEXT:    # kill: def $ax killed $ax killed $eax
278; X86-NEXT:    retl
279;
280; X64-LABEL: test_bitreverse_i16:
281; X64:       # %bb.0:
282; X64-NEXT:    # kill: def $edi killed $edi def $rdi
283; X64-NEXT:    rolw $8, %di
284; X64-NEXT:    movl %edi, %eax
285; X64-NEXT:    andl $3855, %eax # imm = 0xF0F
286; X64-NEXT:    shll $4, %eax
287; X64-NEXT:    andl $61680, %edi # imm = 0xF0F0
288; X64-NEXT:    shrl $4, %edi
289; X64-NEXT:    orl %eax, %edi
290; X64-NEXT:    movl %edi, %eax
291; X64-NEXT:    andl $13107, %eax # imm = 0x3333
292; X64-NEXT:    andl $52428, %edi # imm = 0xCCCC
293; X64-NEXT:    shrl $2, %edi
294; X64-NEXT:    leal (%rdi,%rax,4), %eax
295; X64-NEXT:    movl %eax, %ecx
296; X64-NEXT:    andl $21845, %ecx # imm = 0x5555
297; X64-NEXT:    andl $43690, %eax # imm = 0xAAAA
298; X64-NEXT:    shrl %eax
299; X64-NEXT:    leal (%rax,%rcx,2), %eax
300; X64-NEXT:    # kill: def $ax killed $ax killed $eax
301; X64-NEXT:    retq
302  %b = call i16 @llvm.bitreverse.i16(i16 %a)
303  ret i16 %b
304}
305
306declare i8 @llvm.bitreverse.i8(i8) readnone
307
308define i8 @test_bitreverse_i8(i8 %a) {
309; X86-LABEL: test_bitreverse_i8:
310; X86:       # %bb.0:
311; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
312; X86-NEXT:    rolb $4, %al
313; X86-NEXT:    movl %eax, %ecx
314; X86-NEXT:    andb $51, %cl
315; X86-NEXT:    shlb $2, %cl
316; X86-NEXT:    andb $-52, %al
317; X86-NEXT:    shrb $2, %al
318; X86-NEXT:    orb %cl, %al
319; X86-NEXT:    movl %eax, %ecx
320; X86-NEXT:    andb $85, %cl
321; X86-NEXT:    addb %cl, %cl
322; X86-NEXT:    andb $-86, %al
323; X86-NEXT:    shrb %al
324; X86-NEXT:    orb %cl, %al
325; X86-NEXT:    retl
326;
327; X64-LABEL: test_bitreverse_i8:
328; X64:       # %bb.0:
329; X64-NEXT:    # kill: def $edi killed $edi def $rdi
330; X64-NEXT:    rolb $4, %dil
331; X64-NEXT:    movl %edi, %eax
332; X64-NEXT:    andb $51, %al
333; X64-NEXT:    shlb $2, %al
334; X64-NEXT:    andb $-52, %dil
335; X64-NEXT:    shrb $2, %dil
336; X64-NEXT:    orb %al, %dil
337; X64-NEXT:    movl %edi, %eax
338; X64-NEXT:    andb $85, %al
339; X64-NEXT:    addb %al, %al
340; X64-NEXT:    andb $-86, %dil
341; X64-NEXT:    shrb %dil
342; X64-NEXT:    addl %edi, %eax
343; X64-NEXT:    # kill: def $al killed $al killed $eax
344; X64-NEXT:    retq
345  %b = call i8 @llvm.bitreverse.i8(i8 %a)
346  ret i8 %b
347}
348
349declare i4 @llvm.bitreverse.i4(i4) readnone
350
351define i4 @test_bitreverse_i4(i4 %a) {
352; X86-LABEL: test_bitreverse_i4:
353; X86:       # %bb.0:
354; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
355; X86-NEXT:    rolb $4, %al
356; X86-NEXT:    movl %eax, %ecx
357; X86-NEXT:    andb $51, %cl
358; X86-NEXT:    shlb $2, %cl
359; X86-NEXT:    andb $-52, %al
360; X86-NEXT:    shrb $2, %al
361; X86-NEXT:    orb %cl, %al
362; X86-NEXT:    movl %eax, %ecx
363; X86-NEXT:    andb $80, %cl
364; X86-NEXT:    addb %cl, %cl
365; X86-NEXT:    andb $-96, %al
366; X86-NEXT:    shrb %al
367; X86-NEXT:    orb %cl, %al
368; X86-NEXT:    shrb $4, %al
369; X86-NEXT:    retl
370;
371; X64-LABEL: test_bitreverse_i4:
372; X64:       # %bb.0:
373; X64-NEXT:    # kill: def $edi killed $edi def $rdi
374; X64-NEXT:    rolb $4, %dil
375; X64-NEXT:    movl %edi, %eax
376; X64-NEXT:    andb $51, %al
377; X64-NEXT:    shlb $2, %al
378; X64-NEXT:    andb $-52, %dil
379; X64-NEXT:    shrb $2, %dil
380; X64-NEXT:    orb %al, %dil
381; X64-NEXT:    movl %edi, %eax
382; X64-NEXT:    andb $80, %al
383; X64-NEXT:    addb %al, %al
384; X64-NEXT:    andb $-96, %dil
385; X64-NEXT:    shrb %dil
386; X64-NEXT:    addl %edi, %eax
387; X64-NEXT:    shrb $4, %al
388; X64-NEXT:    # kill: def $al killed $al killed $eax
389; X64-NEXT:    retq
390  %b = call i4 @llvm.bitreverse.i4(i4 %a)
391  ret i4 %b
392}
393
394; These tests check that bitreverse(constant) calls are folded
395
396define <2 x i16> @fold_v2i16() {
397; X86-LABEL: fold_v2i16:
398; X86:       # %bb.0:
399; X86-NEXT:    movw $-4096, %ax # imm = 0xF000
400; X86-NEXT:    movw $240, %dx
401; X86-NEXT:    retl
402;
403; X64-LABEL: fold_v2i16:
404; X64:       # %bb.0:
405; X64-NEXT:    movaps {{.*#+}} xmm0 = <61440,240,u,u,u,u,u,u>
406; X64-NEXT:    retq
407  %b = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> <i16 15, i16 3840>)
408  ret <2 x i16> %b
409}
410
411define i24 @fold_i24() {
412; X86-LABEL: fold_i24:
413; X86:       # %bb.0:
414; X86-NEXT:    movl $2048, %eax # imm = 0x800
415; X86-NEXT:    retl
416;
417; X64-LABEL: fold_i24:
418; X64:       # %bb.0:
419; X64-NEXT:    movl $2048, %eax # imm = 0x800
420; X64-NEXT:    retq
421  %b = call i24 @llvm.bitreverse.i24(i24 4096)
422  ret i24 %b
423}
424
425define i8 @fold_i8() {
426; X86-LABEL: fold_i8:
427; X86:       # %bb.0:
428; X86-NEXT:    movb $-16, %al
429; X86-NEXT:    retl
430;
431; X64-LABEL: fold_i8:
432; X64:       # %bb.0:
433; X64-NEXT:    movb $-16, %al
434; X64-NEXT:    retq
435  %b = call i8 @llvm.bitreverse.i8(i8 15)
436  ret i8 %b
437}
438
439define i4 @fold_i4() {
440; X86-LABEL: fold_i4:
441; X86:       # %bb.0:
442; X86-NEXT:    movb $1, %al
443; X86-NEXT:    retl
444;
445; X64-LABEL: fold_i4:
446; X64:       # %bb.0:
447; X64-NEXT:    movb $1, %al
448; X64-NEXT:    retq
449  %b = call i4 @llvm.bitreverse.i4(i4 8)
450  ret i4 %b
451}
452
453; These tests check that bitreverse(bitreverse()) calls are removed
454
455define i8 @identity_i8(i8 %a) {
456; X86-LABEL: identity_i8:
457; X86:       # %bb.0:
458; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
459; X86-NEXT:    retl
460;
461; X64-LABEL: identity_i8:
462; X64:       # %bb.0:
463; X64-NEXT:    movl %edi, %eax
464; X64-NEXT:    # kill: def $al killed $al killed $eax
465; X64-NEXT:    retq
466  %b = call i8 @llvm.bitreverse.i8(i8 %a)
467  %c = call i8 @llvm.bitreverse.i8(i8 %b)
468  ret i8 %c
469}
470
471define <2 x i16> @identity_v2i16(<2 x i16> %a) {
472; X86-LABEL: identity_v2i16:
473; X86:       # %bb.0:
474; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
475; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %edx
476; X86-NEXT:    retl
477;
478; X64-LABEL: identity_v2i16:
479; X64:       # %bb.0:
480; X64-NEXT:    retq
481  %b = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> %a)
482  %c = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> %b)
483  ret <2 x i16> %c
484}
485
486; These tests check that bitreverse(undef) calls are removed
487
488define i8 @undef_i8() {
489; X86-LABEL: undef_i8:
490; X86:       # %bb.0:
491; X86-NEXT:    retl
492;
493; X64-LABEL: undef_i8:
494; X64:       # %bb.0:
495; X64-NEXT:    retq
496  %b = call i8 @llvm.bitreverse.i8(i8 undef)
497  ret i8 %b
498}
499
500define <2 x i16> @undef_v2i16() {
501; X86-LABEL: undef_v2i16:
502; X86:       # %bb.0:
503; X86-NEXT:    retl
504;
505; X64-LABEL: undef_v2i16:
506; X64:       # %bb.0:
507; X64-NEXT:    retq
508  %b = call <2 x i16> @llvm.bitreverse.v2i16(<2 x i16> undef)
509  ret <2 x i16> %b
510}
511
512; Make sure we don't assert during type legalization promoting a large
513; bitreverse due to the need for a large shift that won't fit in the i8 returned
514; from getShiftAmountTy.
515define i528 @large_promotion(i528 %A) nounwind {
516; X86-LABEL: large_promotion:
517; X86:       # %bb.0:
518; X86-NEXT:    pushl %ebp
519; X86-NEXT:    pushl %ebx
520; X86-NEXT:    pushl %edi
521; X86-NEXT:    pushl %esi
522; X86-NEXT:    subl $56, %esp
523; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
524; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
525; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
526; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
527; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
528; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
529; X86-NEXT:    bswapl %ebx
530; X86-NEXT:    movl %ebx, %ebp
531; X86-NEXT:    andl $252645135, %ebp # imm = 0xF0F0F0F
532; X86-NEXT:    shll $4, %ebp
533; X86-NEXT:    andl $-252645136, %ebx # imm = 0xF0F0F0F0
534; X86-NEXT:    shrl $4, %ebx
535; X86-NEXT:    orl %ebp, %ebx
536; X86-NEXT:    movl %ebx, %ebp
537; X86-NEXT:    andl $858993459, %ebp # imm = 0x33333333
538; X86-NEXT:    andl $-858993460, %ebx # imm = 0xCCCCCCCC
539; X86-NEXT:    shrl $2, %ebx
540; X86-NEXT:    leal (%ebx,%ebp,4), %ebx
541; X86-NEXT:    movl %ebx, %ebp
542; X86-NEXT:    andl $1431633920, %ebp # imm = 0x55550000
543; X86-NEXT:    andl $-1431699456, %ebx # imm = 0xAAAA0000
544; X86-NEXT:    shrl %ebx
545; X86-NEXT:    leal (%ebx,%ebp,2), %ebx
546; X86-NEXT:    movl %ebx, (%esp) # 4-byte Spill
547; X86-NEXT:    bswapl %edi
548; X86-NEXT:    movl %edi, %ebx
549; X86-NEXT:    andl $252645135, %ebx # imm = 0xF0F0F0F
550; X86-NEXT:    shll $4, %ebx
551; X86-NEXT:    andl $-252645136, %edi # imm = 0xF0F0F0F0
552; X86-NEXT:    shrl $4, %edi
553; X86-NEXT:    orl %ebx, %edi
554; X86-NEXT:    movl %edi, %ebx
555; X86-NEXT:    andl $858993459, %ebx # imm = 0x33333333
556; X86-NEXT:    andl $-858993460, %edi # imm = 0xCCCCCCCC
557; X86-NEXT:    shrl $2, %edi
558; X86-NEXT:    leal (%edi,%ebx,4), %edi
559; X86-NEXT:    movl %edi, %ebx
560; X86-NEXT:    andl $1431655765, %ebx # imm = 0x55555555
561; X86-NEXT:    andl $-1431655766, %edi # imm = 0xAAAAAAAA
562; X86-NEXT:    shrl %edi
563; X86-NEXT:    leal (%edi,%ebx,2), %edi
564; X86-NEXT:    movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
565; X86-NEXT:    bswapl %esi
566; X86-NEXT:    movl %esi, %edi
567; X86-NEXT:    andl $252645135, %edi # imm = 0xF0F0F0F
568; X86-NEXT:    shll $4, %edi
569; X86-NEXT:    andl $-252645136, %esi # imm = 0xF0F0F0F0
570; X86-NEXT:    shrl $4, %esi
571; X86-NEXT:    orl %edi, %esi
572; X86-NEXT:    movl %esi, %edi
573; X86-NEXT:    andl $858993459, %edi # imm = 0x33333333
574; X86-NEXT:    andl $-858993460, %esi # imm = 0xCCCCCCCC
575; X86-NEXT:    shrl $2, %esi
576; X86-NEXT:    leal (%esi,%edi,4), %esi
577; X86-NEXT:    movl %esi, %edi
578; X86-NEXT:    andl $1431655765, %edi # imm = 0x55555555
579; X86-NEXT:    andl $-1431655766, %esi # imm = 0xAAAAAAAA
580; X86-NEXT:    shrl %esi
581; X86-NEXT:    leal (%esi,%edi,2), %ebx
582; X86-NEXT:    bswapl %edx
583; X86-NEXT:    movl %edx, %esi
584; X86-NEXT:    andl $252645135, %esi # imm = 0xF0F0F0F
585; X86-NEXT:    shll $4, %esi
586; X86-NEXT:    andl $-252645136, %edx # imm = 0xF0F0F0F0
587; X86-NEXT:    shrl $4, %edx
588; X86-NEXT:    orl %esi, %edx
589; X86-NEXT:    movl %edx, %esi
590; X86-NEXT:    andl $858993459, %esi # imm = 0x33333333
591; X86-NEXT:    andl $-858993460, %edx # imm = 0xCCCCCCCC
592; X86-NEXT:    shrl $2, %edx
593; X86-NEXT:    leal (%edx,%esi,4), %edx
594; X86-NEXT:    movl %edx, %esi
595; X86-NEXT:    andl $1431655765, %esi # imm = 0x55555555
596; X86-NEXT:    andl $-1431655766, %edx # imm = 0xAAAAAAAA
597; X86-NEXT:    shrl %edx
598; X86-NEXT:    leal (%edx,%esi,2), %edx
599; X86-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
600; X86-NEXT:    bswapl %ecx
601; X86-NEXT:    movl %ecx, %edx
602; X86-NEXT:    andl $252645135, %edx # imm = 0xF0F0F0F
603; X86-NEXT:    shll $4, %edx
604; X86-NEXT:    andl $-252645136, %ecx # imm = 0xF0F0F0F0
605; X86-NEXT:    shrl $4, %ecx
606; X86-NEXT:    orl %edx, %ecx
607; X86-NEXT:    movl %ecx, %edx
608; X86-NEXT:    andl $858993459, %edx # imm = 0x33333333
609; X86-NEXT:    andl $-858993460, %ecx # imm = 0xCCCCCCCC
610; X86-NEXT:    shrl $2, %ecx
611; X86-NEXT:    leal (%ecx,%edx,4), %ecx
612; X86-NEXT:    movl %ecx, %edx
613; X86-NEXT:    andl $1431655765, %edx # imm = 0x55555555
614; X86-NEXT:    andl $-1431655766, %ecx # imm = 0xAAAAAAAA
615; X86-NEXT:    shrl %ecx
616; X86-NEXT:    leal (%ecx,%edx,2), %ecx
617; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
618; X86-NEXT:    bswapl %eax
619; X86-NEXT:    movl %eax, %ecx
620; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
621; X86-NEXT:    shll $4, %ecx
622; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
623; X86-NEXT:    shrl $4, %eax
624; X86-NEXT:    orl %ecx, %eax
625; X86-NEXT:    movl %eax, %ecx
626; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
627; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
628; X86-NEXT:    shrl $2, %eax
629; X86-NEXT:    leal (%eax,%ecx,4), %eax
630; X86-NEXT:    movl %eax, %ecx
631; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
632; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
633; X86-NEXT:    shrl %eax
634; X86-NEXT:    leal (%eax,%ecx,2), %eax
635; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
636; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
637; X86-NEXT:    bswapl %eax
638; X86-NEXT:    movl %eax, %ecx
639; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
640; X86-NEXT:    shll $4, %ecx
641; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
642; X86-NEXT:    shrl $4, %eax
643; X86-NEXT:    orl %ecx, %eax
644; X86-NEXT:    movl %eax, %ecx
645; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
646; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
647; X86-NEXT:    shrl $2, %eax
648; X86-NEXT:    leal (%eax,%ecx,4), %eax
649; X86-NEXT:    movl %eax, %ecx
650; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
651; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
652; X86-NEXT:    shrl %eax
653; X86-NEXT:    leal (%eax,%ecx,2), %eax
654; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
655; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
656; X86-NEXT:    bswapl %eax
657; X86-NEXT:    movl %eax, %ecx
658; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
659; X86-NEXT:    shll $4, %ecx
660; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
661; X86-NEXT:    shrl $4, %eax
662; X86-NEXT:    orl %ecx, %eax
663; X86-NEXT:    movl %eax, %ecx
664; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
665; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
666; X86-NEXT:    shrl $2, %eax
667; X86-NEXT:    leal (%eax,%ecx,4), %eax
668; X86-NEXT:    movl %eax, %ecx
669; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
670; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
671; X86-NEXT:    shrl %eax
672; X86-NEXT:    leal (%eax,%ecx,2), %eax
673; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
674; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
675; X86-NEXT:    bswapl %eax
676; X86-NEXT:    movl %eax, %ecx
677; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
678; X86-NEXT:    shll $4, %ecx
679; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
680; X86-NEXT:    shrl $4, %eax
681; X86-NEXT:    orl %ecx, %eax
682; X86-NEXT:    movl %eax, %ecx
683; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
684; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
685; X86-NEXT:    shrl $2, %eax
686; X86-NEXT:    leal (%eax,%ecx,4), %eax
687; X86-NEXT:    movl %eax, %ecx
688; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
689; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
690; X86-NEXT:    shrl %eax
691; X86-NEXT:    leal (%eax,%ecx,2), %eax
692; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
693; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
694; X86-NEXT:    bswapl %eax
695; X86-NEXT:    movl %eax, %ecx
696; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
697; X86-NEXT:    shll $4, %ecx
698; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
699; X86-NEXT:    shrl $4, %eax
700; X86-NEXT:    orl %ecx, %eax
701; X86-NEXT:    movl %eax, %ecx
702; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
703; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
704; X86-NEXT:    shrl $2, %eax
705; X86-NEXT:    leal (%eax,%ecx,4), %eax
706; X86-NEXT:    movl %eax, %ecx
707; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
708; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
709; X86-NEXT:    shrl %eax
710; X86-NEXT:    leal (%eax,%ecx,2), %eax
711; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
712; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
713; X86-NEXT:    bswapl %eax
714; X86-NEXT:    movl %eax, %ecx
715; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
716; X86-NEXT:    shll $4, %ecx
717; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
718; X86-NEXT:    shrl $4, %eax
719; X86-NEXT:    orl %ecx, %eax
720; X86-NEXT:    movl %eax, %ecx
721; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
722; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
723; X86-NEXT:    shrl $2, %eax
724; X86-NEXT:    leal (%eax,%ecx,4), %eax
725; X86-NEXT:    movl %eax, %ecx
726; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
727; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
728; X86-NEXT:    shrl %eax
729; X86-NEXT:    leal (%eax,%ecx,2), %eax
730; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
731; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
732; X86-NEXT:    bswapl %eax
733; X86-NEXT:    movl %eax, %ecx
734; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
735; X86-NEXT:    shll $4, %ecx
736; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
737; X86-NEXT:    shrl $4, %eax
738; X86-NEXT:    orl %ecx, %eax
739; X86-NEXT:    movl %eax, %ecx
740; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
741; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
742; X86-NEXT:    shrl $2, %eax
743; X86-NEXT:    leal (%eax,%ecx,4), %eax
744; X86-NEXT:    movl %eax, %ecx
745; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
746; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
747; X86-NEXT:    shrl %eax
748; X86-NEXT:    leal (%eax,%ecx,2), %eax
749; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
750; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
751; X86-NEXT:    bswapl %eax
752; X86-NEXT:    movl %eax, %ecx
753; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
754; X86-NEXT:    shll $4, %ecx
755; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
756; X86-NEXT:    shrl $4, %eax
757; X86-NEXT:    orl %ecx, %eax
758; X86-NEXT:    movl %eax, %ecx
759; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
760; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
761; X86-NEXT:    shrl $2, %eax
762; X86-NEXT:    leal (%eax,%ecx,4), %eax
763; X86-NEXT:    movl %eax, %ecx
764; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
765; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
766; X86-NEXT:    shrl %eax
767; X86-NEXT:    leal (%eax,%ecx,2), %eax
768; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
769; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
770; X86-NEXT:    bswapl %eax
771; X86-NEXT:    movl %eax, %ecx
772; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
773; X86-NEXT:    shll $4, %ecx
774; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
775; X86-NEXT:    shrl $4, %eax
776; X86-NEXT:    orl %ecx, %eax
777; X86-NEXT:    movl %eax, %ecx
778; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
779; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
780; X86-NEXT:    shrl $2, %eax
781; X86-NEXT:    leal (%eax,%ecx,4), %eax
782; X86-NEXT:    movl %eax, %ecx
783; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
784; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
785; X86-NEXT:    shrl %eax
786; X86-NEXT:    leal (%eax,%ecx,2), %eax
787; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
788; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
789; X86-NEXT:    bswapl %eax
790; X86-NEXT:    movl %eax, %ecx
791; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
792; X86-NEXT:    shll $4, %ecx
793; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
794; X86-NEXT:    shrl $4, %eax
795; X86-NEXT:    orl %ecx, %eax
796; X86-NEXT:    movl %eax, %ecx
797; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
798; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
799; X86-NEXT:    shrl $2, %eax
800; X86-NEXT:    leal (%eax,%ecx,4), %eax
801; X86-NEXT:    movl %eax, %ecx
802; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
803; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
804; X86-NEXT:    shrl %eax
805; X86-NEXT:    leal (%eax,%ecx,2), %eax
806; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
807; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
808; X86-NEXT:    bswapl %eax
809; X86-NEXT:    movl %eax, %ecx
810; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
811; X86-NEXT:    shll $4, %ecx
812; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
813; X86-NEXT:    shrl $4, %eax
814; X86-NEXT:    orl %ecx, %eax
815; X86-NEXT:    movl %eax, %ecx
816; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
817; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
818; X86-NEXT:    shrl $2, %eax
819; X86-NEXT:    leal (%eax,%ecx,4), %eax
820; X86-NEXT:    movl %eax, %ecx
821; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
822; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
823; X86-NEXT:    shrl %eax
824; X86-NEXT:    leal (%eax,%ecx,2), %edi
825; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
826; X86-NEXT:    bswapl %eax
827; X86-NEXT:    movl %eax, %ecx
828; X86-NEXT:    andl $252645135, %ecx # imm = 0xF0F0F0F
829; X86-NEXT:    shll $4, %ecx
830; X86-NEXT:    andl $-252645136, %eax # imm = 0xF0F0F0F0
831; X86-NEXT:    shrl $4, %eax
832; X86-NEXT:    orl %ecx, %eax
833; X86-NEXT:    movl %eax, %ecx
834; X86-NEXT:    andl $858993459, %ecx # imm = 0x33333333
835; X86-NEXT:    andl $-858993460, %eax # imm = 0xCCCCCCCC
836; X86-NEXT:    shrl $2, %eax
837; X86-NEXT:    leal (%eax,%ecx,4), %eax
838; X86-NEXT:    movl %eax, %ecx
839; X86-NEXT:    andl $1431655765, %ecx # imm = 0x55555555
840; X86-NEXT:    andl $-1431655766, %eax # imm = 0xAAAAAAAA
841; X86-NEXT:    shrl %eax
842; X86-NEXT:    leal (%eax,%ecx,2), %edx
843; X86-NEXT:    movl (%esp), %esi # 4-byte Reload
844; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
845; X86-NEXT:    shrdl $16, %eax, %esi
846; X86-NEXT:    shrdl $16, %ebx, %eax
847; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
848; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
849; X86-NEXT:    shrdl $16, %ecx, %ebx
850; X86-NEXT:    movl %ebx, (%esp) # 4-byte Spill
851; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
852; X86-NEXT:    shrdl $16, %eax, %ecx
853; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
854; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
855; X86-NEXT:    shrdl $16, %ecx, %eax
856; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
857; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
858; X86-NEXT:    shrdl $16, %eax, %ecx
859; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
860; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
861; X86-NEXT:    shrdl $16, %ecx, %eax
862; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
863; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
864; X86-NEXT:    shrdl $16, %eax, %ecx
865; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
866; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
867; X86-NEXT:    shrdl $16, %ecx, %eax
868; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
869; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
870; X86-NEXT:    shrdl $16, %eax, %ecx
871; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
872; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
873; X86-NEXT:    shrdl $16, %ecx, %eax
874; X86-NEXT:    movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
875; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
876; X86-NEXT:    shrdl $16, %ebp, %ecx
877; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
878; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
879; X86-NEXT:    shrdl $16, %ebx, %ebp
880; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
881; X86-NEXT:    shrdl $16, %eax, %ebx
882; X86-NEXT:    shrdl $16, %edi, %eax
883; X86-NEXT:    movl %eax, %ecx
884; X86-NEXT:    shrdl $16, %edx, %edi
885; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
886; X86-NEXT:    movl %edi, 60(%eax)
887; X86-NEXT:    movl %ecx, 56(%eax)
888; X86-NEXT:    movl %ebx, 52(%eax)
889; X86-NEXT:    movl %ebp, 48(%eax)
890; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
891; X86-NEXT:    movl %ecx, 44(%eax)
892; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
893; X86-NEXT:    movl %ecx, 40(%eax)
894; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
895; X86-NEXT:    movl %ecx, 36(%eax)
896; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
897; X86-NEXT:    movl %ecx, 32(%eax)
898; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
899; X86-NEXT:    movl %ecx, 28(%eax)
900; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
901; X86-NEXT:    movl %ecx, 24(%eax)
902; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
903; X86-NEXT:    movl %ecx, 20(%eax)
904; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
905; X86-NEXT:    movl %ecx, 16(%eax)
906; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
907; X86-NEXT:    movl %ecx, 12(%eax)
908; X86-NEXT:    movl (%esp), %ecx # 4-byte Reload
909; X86-NEXT:    movl %ecx, 8(%eax)
910; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
911; X86-NEXT:    movl %ecx, 4(%eax)
912; X86-NEXT:    movl %esi, (%eax)
913; X86-NEXT:    shrl $16, %edx
914; X86-NEXT:    movw %dx, 64(%eax)
915; X86-NEXT:    addl $56, %esp
916; X86-NEXT:    popl %esi
917; X86-NEXT:    popl %edi
918; X86-NEXT:    popl %ebx
919; X86-NEXT:    popl %ebp
920; X86-NEXT:    retl $4
921;
922; X64-LABEL: large_promotion:
923; X64:       # %bb.0:
924; X64-NEXT:    pushq %rbp
925; X64-NEXT:    pushq %r15
926; X64-NEXT:    pushq %r14
927; X64-NEXT:    pushq %r13
928; X64-NEXT:    pushq %r12
929; X64-NEXT:    pushq %rbx
930; X64-NEXT:    movq %rdi, %r12
931; X64-NEXT:    movq {{[0-9]+}}(%rsp), %rbp
932; X64-NEXT:    movq {{[0-9]+}}(%rsp), %rbx
933; X64-NEXT:    bswapq %rbx
934; X64-NEXT:    movabsq $1085102592571150095, %r13 # imm = 0xF0F0F0F0F0F0F0F
935; X64-NEXT:    movq %rbx, %r10
936; X64-NEXT:    andq %r13, %r10
937; X64-NEXT:    shlq $4, %r10
938; X64-NEXT:    movabsq $-1085102592571150096, %rax # imm = 0xF0F0F0F0F0F0F0F0
939; X64-NEXT:    andq %rax, %rbx
940; X64-NEXT:    shrq $4, %rbx
941; X64-NEXT:    orq %r10, %rbx
942; X64-NEXT:    movabsq $3689348814741910323, %r11 # imm = 0x3333333333333333
943; X64-NEXT:    movq %rbx, %r10
944; X64-NEXT:    andq %r11, %r10
945; X64-NEXT:    movabsq $-3689348814741910324, %r14 # imm = 0xCCCCCCCCCCCCCCCC
946; X64-NEXT:    andq %r14, %rbx
947; X64-NEXT:    shrq $2, %rbx
948; X64-NEXT:    leaq (%rbx,%r10,4), %r10
949; X64-NEXT:    movabsq $6148820866244280320, %rbx # imm = 0x5555000000000000
950; X64-NEXT:    andq %r10, %rbx
951; X64-NEXT:    movabsq $-6149102341220990976, %rdi # imm = 0xAAAA000000000000
952; X64-NEXT:    andq %r10, %rdi
953; X64-NEXT:    shrq %rdi
954; X64-NEXT:    leaq (%rdi,%rbx,2), %rdi
955; X64-NEXT:    movq %rdi, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
956; X64-NEXT:    bswapq %rbp
957; X64-NEXT:    movq %rbp, %rdi
958; X64-NEXT:    andq %r13, %rdi
959; X64-NEXT:    shlq $4, %rdi
960; X64-NEXT:    andq %rax, %rbp
961; X64-NEXT:    shrq $4, %rbp
962; X64-NEXT:    orq %rdi, %rbp
963; X64-NEXT:    movq %rbp, %rdi
964; X64-NEXT:    andq %r11, %rdi
965; X64-NEXT:    andq %r14, %rbp
966; X64-NEXT:    shrq $2, %rbp
967; X64-NEXT:    leaq (%rbp,%rdi,4), %rbp
968; X64-NEXT:    movabsq $6148914691236517205, %rbx # imm = 0x5555555555555555
969; X64-NEXT:    movq %rbp, %r10
970; X64-NEXT:    andq %rbx, %r10
971; X64-NEXT:    movabsq $-6148914691236517206, %rdi # imm = 0xAAAAAAAAAAAAAAAA
972; X64-NEXT:    andq %rdi, %rbp
973; X64-NEXT:    shrq %rbp
974; X64-NEXT:    leaq (%rbp,%r10,2), %rbp
975; X64-NEXT:    movq %rbp, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
976; X64-NEXT:    movq {{[0-9]+}}(%rsp), %rbp
977; X64-NEXT:    bswapq %rbp
978; X64-NEXT:    movq %rbp, %r10
979; X64-NEXT:    andq %r13, %r10
980; X64-NEXT:    shlq $4, %r10
981; X64-NEXT:    andq %rax, %rbp
982; X64-NEXT:    movq %rax, %r15
983; X64-NEXT:    shrq $4, %rbp
984; X64-NEXT:    orq %r10, %rbp
985; X64-NEXT:    movq %rbp, %r10
986; X64-NEXT:    andq %r11, %r10
987; X64-NEXT:    andq %r14, %rbp
988; X64-NEXT:    shrq $2, %rbp
989; X64-NEXT:    leaq (%rbp,%r10,4), %rbp
990; X64-NEXT:    movq %rbp, %r10
991; X64-NEXT:    andq %rbx, %r10
992; X64-NEXT:    andq %rdi, %rbp
993; X64-NEXT:    shrq %rbp
994; X64-NEXT:    leaq (%rbp,%r10,2), %rbp
995; X64-NEXT:    movq {{[0-9]+}}(%rsp), %r10
996; X64-NEXT:    bswapq %r10
997; X64-NEXT:    movq %r10, %rax
998; X64-NEXT:    andq %r13, %rax
999; X64-NEXT:    shlq $4, %rax
1000; X64-NEXT:    movq %r15, %rdi
1001; X64-NEXT:    andq %r15, %r10
1002; X64-NEXT:    shrq $4, %r10
1003; X64-NEXT:    orq %rax, %r10
1004; X64-NEXT:    movq %r10, %rax
1005; X64-NEXT:    andq %r11, %rax
1006; X64-NEXT:    andq %r14, %r10
1007; X64-NEXT:    shrq $2, %r10
1008; X64-NEXT:    leaq (%r10,%rax,4), %rax
1009; X64-NEXT:    movq %rax, %r10
1010; X64-NEXT:    andq %rbx, %r10
1011; X64-NEXT:    movabsq $-6148914691236517206, %r15 # imm = 0xAAAAAAAAAAAAAAAA
1012; X64-NEXT:    andq %r15, %rax
1013; X64-NEXT:    shrq %rax
1014; X64-NEXT:    leaq (%rax,%r10,2), %r10
1015; X64-NEXT:    bswapq %r9
1016; X64-NEXT:    movq %r9, %rax
1017; X64-NEXT:    andq %r13, %rax
1018; X64-NEXT:    shlq $4, %rax
1019; X64-NEXT:    andq %rdi, %r9
1020; X64-NEXT:    shrq $4, %r9
1021; X64-NEXT:    orq %rax, %r9
1022; X64-NEXT:    movq %r9, %rax
1023; X64-NEXT:    andq %r11, %rax
1024; X64-NEXT:    andq %r14, %r9
1025; X64-NEXT:    shrq $2, %r9
1026; X64-NEXT:    leaq (%r9,%rax,4), %rax
1027; X64-NEXT:    movq %rax, %r9
1028; X64-NEXT:    andq %rbx, %r9
1029; X64-NEXT:    andq %r15, %rax
1030; X64-NEXT:    shrq %rax
1031; X64-NEXT:    leaq (%rax,%r9,2), %r9
1032; X64-NEXT:    bswapq %r8
1033; X64-NEXT:    movq %r8, %rax
1034; X64-NEXT:    andq %r13, %rax
1035; X64-NEXT:    shlq $4, %rax
1036; X64-NEXT:    andq %rdi, %r8
1037; X64-NEXT:    shrq $4, %r8
1038; X64-NEXT:    orq %rax, %r8
1039; X64-NEXT:    movq %r8, %rax
1040; X64-NEXT:    andq %r11, %rax
1041; X64-NEXT:    andq %r14, %r8
1042; X64-NEXT:    shrq $2, %r8
1043; X64-NEXT:    leaq (%r8,%rax,4), %rax
1044; X64-NEXT:    movq %rax, %r8
1045; X64-NEXT:    andq %rbx, %r8
1046; X64-NEXT:    andq %r15, %rax
1047; X64-NEXT:    shrq %rax
1048; X64-NEXT:    leaq (%rax,%r8,2), %r8
1049; X64-NEXT:    bswapq %rcx
1050; X64-NEXT:    movq %rcx, %rax
1051; X64-NEXT:    andq %r13, %rax
1052; X64-NEXT:    shlq $4, %rax
1053; X64-NEXT:    andq %rdi, %rcx
1054; X64-NEXT:    shrq $4, %rcx
1055; X64-NEXT:    orq %rax, %rcx
1056; X64-NEXT:    movq %rcx, %rax
1057; X64-NEXT:    andq %r11, %rax
1058; X64-NEXT:    andq %r14, %rcx
1059; X64-NEXT:    shrq $2, %rcx
1060; X64-NEXT:    leaq (%rcx,%rax,4), %rax
1061; X64-NEXT:    movq %rax, %rcx
1062; X64-NEXT:    andq %rbx, %rcx
1063; X64-NEXT:    andq %r15, %rax
1064; X64-NEXT:    shrq %rax
1065; X64-NEXT:    leaq (%rax,%rcx,2), %rcx
1066; X64-NEXT:    bswapq %rdx
1067; X64-NEXT:    movq %rdx, %rax
1068; X64-NEXT:    andq %r13, %rax
1069; X64-NEXT:    shlq $4, %rax
1070; X64-NEXT:    andq %rdi, %rdx
1071; X64-NEXT:    shrq $4, %rdx
1072; X64-NEXT:    orq %rax, %rdx
1073; X64-NEXT:    movq %rdx, %rax
1074; X64-NEXT:    andq %r11, %rax
1075; X64-NEXT:    andq %r14, %rdx
1076; X64-NEXT:    shrq $2, %rdx
1077; X64-NEXT:    leaq (%rdx,%rax,4), %rax
1078; X64-NEXT:    movq %rax, %rdx
1079; X64-NEXT:    andq %rbx, %rdx
1080; X64-NEXT:    andq %r15, %rax
1081; X64-NEXT:    shrq %rax
1082; X64-NEXT:    leaq (%rax,%rdx,2), %rax
1083; X64-NEXT:    bswapq %rsi
1084; X64-NEXT:    andq %rsi, %r13
1085; X64-NEXT:    andq %rdi, %rsi
1086; X64-NEXT:    shlq $4, %r13
1087; X64-NEXT:    shrq $4, %rsi
1088; X64-NEXT:    orq %r13, %rsi
1089; X64-NEXT:    andq %rsi, %r11
1090; X64-NEXT:    andq %r14, %rsi
1091; X64-NEXT:    shrq $2, %rsi
1092; X64-NEXT:    leaq (%rsi,%r11,4), %rdx
1093; X64-NEXT:    andq %rdx, %rbx
1094; X64-NEXT:    andq %r15, %rdx
1095; X64-NEXT:    shrq %rdx
1096; X64-NEXT:    leaq (%rdx,%rbx,2), %rdx
1097; X64-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rsi # 8-byte Reload
1098; X64-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rdi # 8-byte Reload
1099; X64-NEXT:    shrdq $48, %rdi, %rsi
1100; X64-NEXT:    shrdq $48, %rbp, %rdi
1101; X64-NEXT:    shrdq $48, %r10, %rbp
1102; X64-NEXT:    shrdq $48, %r9, %r10
1103; X64-NEXT:    shrdq $48, %r8, %r9
1104; X64-NEXT:    shrdq $48, %rcx, %r8
1105; X64-NEXT:    shrdq $48, %rax, %rcx
1106; X64-NEXT:    shrdq $48, %rdx, %rax
1107; X64-NEXT:    movq %rax, 56(%r12)
1108; X64-NEXT:    movq %rcx, 48(%r12)
1109; X64-NEXT:    movq %r8, 40(%r12)
1110; X64-NEXT:    movq %r9, 32(%r12)
1111; X64-NEXT:    movq %r10, 24(%r12)
1112; X64-NEXT:    movq %rbp, 16(%r12)
1113; X64-NEXT:    movq %rdi, 8(%r12)
1114; X64-NEXT:    movq %rsi, (%r12)
1115; X64-NEXT:    shrq $48, %rdx
1116; X64-NEXT:    movw %dx, 64(%r12)
1117; X64-NEXT:    movq %r12, %rax
1118; X64-NEXT:    popq %rbx
1119; X64-NEXT:    popq %r12
1120; X64-NEXT:    popq %r13
1121; X64-NEXT:    popq %r14
1122; X64-NEXT:    popq %r15
1123; X64-NEXT:    popq %rbp
1124; X64-NEXT:    retq
1125  %Z = call i528 @llvm.bitreverse.i528(i528 %A)
1126  ret i528 %Z
1127}
1128declare i528 @llvm.bitreverse.i528(i528)
1129