xref: /llvm-project/llvm/test/CodeGen/X86/divide-by-constant.ll (revision ed11f255a89f6eb3713a3d4f2b241317a41be40c) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-pc-linux-gnu | FileCheck %s --check-prefix=X86
3; RUN: llc < %s -mtriple=x86_64-pc-linux-gnu | FileCheck %s --check-prefixes=X64,X64-FAST
4; RUN: llc < %s -mtriple=x86_64-pc-linux-gnu -mattr=idivq-to-divl | FileCheck %s --check-prefixes=X64,X64-SLOW
5
6define zeroext i16 @test1(i16 zeroext %x) nounwind {
7; X86-LABEL: test1:
8; X86:       # %bb.0: # %entry
9; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
10; X86-NEXT:    imull $63551, %eax, %eax # imm = 0xF83F
11; X86-NEXT:    shrl $21, %eax
12; X86-NEXT:    # kill: def $ax killed $ax killed $eax
13; X86-NEXT:    retl
14;
15; X64-LABEL: test1:
16; X64:       # %bb.0: # %entry
17; X64-NEXT:    imull $63551, %edi, %eax # imm = 0xF83F
18; X64-NEXT:    shrl $21, %eax
19; X64-NEXT:    # kill: def $ax killed $ax killed $eax
20; X64-NEXT:    retq
21entry:
22  %div = udiv i16 %x, 33
23  ret i16 %div
24}
25
26define zeroext i16 @test2(i8 signext %x, i16 zeroext %c) nounwind readnone ssp noredzone {
27; X86-LABEL: test2:
28; X86:       # %bb.0: # %entry
29; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
30; X86-NEXT:    imull $43691, %eax, %eax # imm = 0xAAAB
31; X86-NEXT:    shrl $17, %eax
32; X86-NEXT:    # kill: def $ax killed $ax killed $eax
33; X86-NEXT:    retl
34;
35; X64-LABEL: test2:
36; X64:       # %bb.0: # %entry
37; X64-NEXT:    imull $43691, %esi, %eax # imm = 0xAAAB
38; X64-NEXT:    shrl $17, %eax
39; X64-NEXT:    # kill: def $ax killed $ax killed $eax
40; X64-NEXT:    retq
41entry:
42  %div = udiv i16 %c, 3
43  ret i16 %div
44
45}
46
47define zeroext i8 @test3(i8 zeroext %x, i8 zeroext %c) nounwind readnone ssp noredzone {
48; X86-LABEL: test3:
49; X86:       # %bb.0: # %entry
50; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
51; X86-NEXT:    imull $171, %eax, %eax
52; X86-NEXT:    shrl $9, %eax
53; X86-NEXT:    # kill: def $al killed $al killed $eax
54; X86-NEXT:    retl
55;
56; X64-LABEL: test3:
57; X64:       # %bb.0: # %entry
58; X64-NEXT:    imull $171, %esi, %eax
59; X64-NEXT:    shrl $9, %eax
60; X64-NEXT:    # kill: def $al killed $al killed $eax
61; X64-NEXT:    retq
62entry:
63  %div = udiv i8 %c, 3
64  ret i8 %div
65}
66
67define signext i16 @test4(i16 signext %x) nounwind {
68; X86-LABEL: test4:
69; X86:       # %bb.0: # %entry
70; X86-NEXT:    movswl {{[0-9]+}}(%esp), %eax
71; X86-NEXT:    imull $1986, %eax, %eax # imm = 0x7C2
72; X86-NEXT:    movl %eax, %ecx
73; X86-NEXT:    shrl $31, %ecx
74; X86-NEXT:    shrl $16, %eax
75; X86-NEXT:    addl %ecx, %eax
76; X86-NEXT:    # kill: def $ax killed $ax killed $eax
77; X86-NEXT:    retl
78;
79; X64-LABEL: test4:
80; X64:       # %bb.0: # %entry
81; X64-NEXT:    imull $1986, %edi, %eax # imm = 0x7C2
82; X64-NEXT:    movl %eax, %ecx
83; X64-NEXT:    shrl $31, %ecx
84; X64-NEXT:    shrl $16, %eax
85; X64-NEXT:    addl %ecx, %eax
86; X64-NEXT:    # kill: def $ax killed $ax killed $eax
87; X64-NEXT:    retq
88entry:
89  %div = sdiv i16 %x, 33		; <i32> [#uses=1]
90  ret i16 %div
91}
92
93define i32 @test5(i32 %A) nounwind {
94; X86-LABEL: test5:
95; X86:       # %bb.0:
96; X86-NEXT:    movl $365384439, %eax # imm = 0x15C752F7
97; X86-NEXT:    mull {{[0-9]+}}(%esp)
98; X86-NEXT:    movl %edx, %eax
99; X86-NEXT:    shrl $27, %eax
100; X86-NEXT:    retl
101;
102; X64-LABEL: test5:
103; X64:       # %bb.0:
104; X64-NEXT:    movl %edi, %eax
105; X64-NEXT:    imulq $365384439, %rax, %rax # imm = 0x15C752F7
106; X64-NEXT:    shrq $59, %rax
107; X64-NEXT:    # kill: def $eax killed $eax killed $rax
108; X64-NEXT:    retq
109  %tmp1 = udiv i32 %A, 1577682821         ; <i32> [#uses=1]
110  ret i32 %tmp1
111}
112
113define signext i16 @test6(i16 signext %x) nounwind {
114; X86-LABEL: test6:
115; X86:       # %bb.0: # %entry
116; X86-NEXT:    movswl {{[0-9]+}}(%esp), %eax
117; X86-NEXT:    imull $26215, %eax, %eax # imm = 0x6667
118; X86-NEXT:    movl %eax, %ecx
119; X86-NEXT:    shrl $31, %ecx
120; X86-NEXT:    sarl $18, %eax
121; X86-NEXT:    addl %ecx, %eax
122; X86-NEXT:    # kill: def $ax killed $ax killed $eax
123; X86-NEXT:    retl
124;
125; X64-LABEL: test6:
126; X64:       # %bb.0: # %entry
127; X64-NEXT:    imull $26215, %edi, %eax # imm = 0x6667
128; X64-NEXT:    movl %eax, %ecx
129; X64-NEXT:    shrl $31, %ecx
130; X64-NEXT:    sarl $18, %eax
131; X64-NEXT:    addl %ecx, %eax
132; X64-NEXT:    # kill: def $ax killed $ax killed $eax
133; X64-NEXT:    retq
134entry:
135  %div = sdiv i16 %x, 10
136  ret i16 %div
137}
138
139define i32 @test7(i32 %x) nounwind {
140; X86-LABEL: test7:
141; X86:       # %bb.0:
142; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
143; X86-NEXT:    shrl $2, %eax
144; X86-NEXT:    movl $613566757, %ecx # imm = 0x24924925
145; X86-NEXT:    mull %ecx
146; X86-NEXT:    movl %edx, %eax
147; X86-NEXT:    retl
148;
149; X64-LABEL: test7:
150; X64:       # %bb.0:
151; X64-NEXT:    # kill: def $edi killed $edi def $rdi
152; X64-NEXT:    shrl $2, %edi
153; X64-NEXT:    imulq $613566757, %rdi, %rax # imm = 0x24924925
154; X64-NEXT:    shrq $32, %rax
155; X64-NEXT:    # kill: def $eax killed $eax killed $rax
156; X64-NEXT:    retq
157  %div = udiv i32 %x, 28
158  ret i32 %div
159}
160
161; PR13326
162define i8 @test8(i8 %x) nounwind {
163; X86-LABEL: test8:
164; X86:       # %bb.0:
165; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
166; X86-NEXT:    shrb %al
167; X86-NEXT:    movzbl %al, %eax
168; X86-NEXT:    imull $211, %eax, %eax
169; X86-NEXT:    shrl $13, %eax
170; X86-NEXT:    # kill: def $al killed $al killed $eax
171; X86-NEXT:    retl
172;
173; X64-LABEL: test8:
174; X64:       # %bb.0:
175; X64-NEXT:    shrb %dil
176; X64-NEXT:    movzbl %dil, %eax
177; X64-NEXT:    imull $211, %eax, %eax
178; X64-NEXT:    shrl $13, %eax
179; X64-NEXT:    # kill: def $al killed $al killed $eax
180; X64-NEXT:    retq
181  %div = udiv i8 %x, 78
182  ret i8 %div
183}
184
185define i8 @test9(i8 %x) nounwind {
186; X86-LABEL: test9:
187; X86:       # %bb.0:
188; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
189; X86-NEXT:    shrb $2, %al
190; X86-NEXT:    movzbl %al, %eax
191; X86-NEXT:    imull $71, %eax, %eax
192; X86-NEXT:    shrl $11, %eax
193; X86-NEXT:    # kill: def $al killed $al killed $eax
194; X86-NEXT:    retl
195;
196; X64-LABEL: test9:
197; X64:       # %bb.0:
198; X64-NEXT:    shrb $2, %dil
199; X64-NEXT:    movzbl %dil, %eax
200; X64-NEXT:    imull $71, %eax, %eax
201; X64-NEXT:    shrl $11, %eax
202; X64-NEXT:    # kill: def $al killed $al killed $eax
203; X64-NEXT:    retq
204  %div = udiv i8 %x, 116
205  ret i8 %div
206}
207
208define i32 @testsize1(i32 %x) minsize nounwind {
209; X86-LABEL: testsize1:
210; X86:       # %bb.0: # %entry
211; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
212; X86-NEXT:    pushl $32
213; X86-NEXT:    popl %ecx
214; X86-NEXT:    cltd
215; X86-NEXT:    idivl %ecx
216; X86-NEXT:    retl
217;
218; X64-LABEL: testsize1:
219; X64:       # %bb.0: # %entry
220; X64-NEXT:    movl %edi, %eax
221; X64-NEXT:    pushq $32
222; X64-NEXT:    popq %rcx
223; X64-NEXT:    cltd
224; X64-NEXT:    idivl %ecx
225; X64-NEXT:    retq
226entry:
227  %div = sdiv i32 %x, 32
228  ret i32 %div
229}
230
231define i32 @testsize2(i32 %x) minsize nounwind {
232; X86-LABEL: testsize2:
233; X86:       # %bb.0: # %entry
234; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
235; X86-NEXT:    pushl $33
236; X86-NEXT:    popl %ecx
237; X86-NEXT:    cltd
238; X86-NEXT:    idivl %ecx
239; X86-NEXT:    retl
240;
241; X64-LABEL: testsize2:
242; X64:       # %bb.0: # %entry
243; X64-NEXT:    movl %edi, %eax
244; X64-NEXT:    pushq $33
245; X64-NEXT:    popq %rcx
246; X64-NEXT:    cltd
247; X64-NEXT:    idivl %ecx
248; X64-NEXT:    retq
249entry:
250  %div = sdiv i32 %x, 33
251  ret i32 %div
252}
253
254define i32 @testsize3(i32 %x) minsize nounwind {
255; X86-LABEL: testsize3:
256; X86:       # %bb.0: # %entry
257; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
258; X86-NEXT:    shrl $5, %eax
259; X86-NEXT:    retl
260;
261; X64-LABEL: testsize3:
262; X64:       # %bb.0: # %entry
263; X64-NEXT:    movl %edi, %eax
264; X64-NEXT:    shrl $5, %eax
265; X64-NEXT:    retq
266entry:
267  %div = udiv i32 %x, 32
268  ret i32 %div
269}
270
271define i32 @testsize4(i32 %x) minsize nounwind {
272; X86-LABEL: testsize4:
273; X86:       # %bb.0: # %entry
274; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
275; X86-NEXT:    pushl $33
276; X86-NEXT:    popl %ecx
277; X86-NEXT:    xorl %edx, %edx
278; X86-NEXT:    divl %ecx
279; X86-NEXT:    retl
280;
281; X64-LABEL: testsize4:
282; X64:       # %bb.0: # %entry
283; X64-NEXT:    movl %edi, %eax
284; X64-NEXT:    pushq $33
285; X64-NEXT:    popq %rcx
286; X64-NEXT:    xorl %edx, %edx
287; X64-NEXT:    divl %ecx
288; X64-NEXT:    retq
289entry:
290  %div = udiv i32 %x, 33
291  ret i32 %div
292}
293
294define i64 @PR23590(i64 %x) nounwind {
295; X86-LABEL: PR23590:
296; X86:       # %bb.0: # %entry
297; X86-NEXT:    subl $12, %esp
298; X86-NEXT:    pushl $0
299; X86-NEXT:    pushl $12345 # imm = 0x3039
300; X86-NEXT:    pushl {{[0-9]+}}(%esp)
301; X86-NEXT:    pushl {{[0-9]+}}(%esp)
302; X86-NEXT:    calll __umoddi3
303; X86-NEXT:    addl $16, %esp
304; X86-NEXT:    pushl $0
305; X86-NEXT:    pushl $7
306; X86-NEXT:    pushl %edx
307; X86-NEXT:    pushl %eax
308; X86-NEXT:    calll __udivdi3
309; X86-NEXT:    addl $28, %esp
310; X86-NEXT:    retl
311;
312; X64-FAST-LABEL: PR23590:
313; X64-FAST:       # %bb.0: # %entry
314; X64-FAST-NEXT:    movabsq $6120523590596543007, %rcx # imm = 0x54F077C718E7C21F
315; X64-FAST-NEXT:    movq %rdi, %rax
316; X64-FAST-NEXT:    mulq %rcx
317; X64-FAST-NEXT:    shrq $12, %rdx
318; X64-FAST-NEXT:    imulq $12345, %rdx, %rax # imm = 0x3039
319; X64-FAST-NEXT:    subq %rax, %rdi
320; X64-FAST-NEXT:    movabsq $2635249153387078803, %rcx # imm = 0x2492492492492493
321; X64-FAST-NEXT:    movq %rdi, %rax
322; X64-FAST-NEXT:    mulq %rcx
323; X64-FAST-NEXT:    movq %rdx, %rax
324; X64-FAST-NEXT:    retq
325;
326; X64-SLOW-LABEL: PR23590:
327; X64-SLOW:       # %bb.0: # %entry
328; X64-SLOW-NEXT:    movabsq $6120523590596543007, %rcx # imm = 0x54F077C718E7C21F
329; X64-SLOW-NEXT:    movq %rdi, %rax
330; X64-SLOW-NEXT:    mulq %rcx
331; X64-SLOW-NEXT:    shrq $12, %rdx
332; X64-SLOW-NEXT:    imulq $12345, %rdx, %rax # imm = 0x3039
333; X64-SLOW-NEXT:    subq %rax, %rdi
334; X64-SLOW-NEXT:    imulq $613566757, %rdi, %rax # imm = 0x24924925
335; X64-SLOW-NEXT:    shrq $32, %rax
336; X64-SLOW-NEXT:    retq
337entry:
338  %rem = urem i64 %x, 12345
339  %div = udiv i64 %rem, 7
340  ret i64 %div
341}
342
343define { i64, i32 } @PR38622(i64) nounwind {
344; X86-LABEL: PR38622:
345; X86:       # %bb.0:
346; X86-NEXT:    pushl %ebp
347; X86-NEXT:    pushl %ebx
348; X86-NEXT:    pushl %edi
349; X86-NEXT:    pushl %esi
350; X86-NEXT:    subl $12, %esp
351; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
352; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebp
353; X86-NEXT:    pushl $0
354; X86-NEXT:    pushl $-294967296 # imm = 0xEE6B2800
355; X86-NEXT:    pushl %ebp
356; X86-NEXT:    pushl %ebx
357; X86-NEXT:    calll __udivdi3
358; X86-NEXT:    addl $16, %esp
359; X86-NEXT:    movl %eax, %esi
360; X86-NEXT:    movl %edx, %edi
361; X86-NEXT:    pushl $0
362; X86-NEXT:    pushl $-294967296 # imm = 0xEE6B2800
363; X86-NEXT:    pushl %ebp
364; X86-NEXT:    pushl %ebx
365; X86-NEXT:    calll __umoddi3
366; X86-NEXT:    addl $16, %esp
367; X86-NEXT:    movl %eax, %ecx
368; X86-NEXT:    movl %esi, %eax
369; X86-NEXT:    movl %edi, %edx
370; X86-NEXT:    addl $12, %esp
371; X86-NEXT:    popl %esi
372; X86-NEXT:    popl %edi
373; X86-NEXT:    popl %ebx
374; X86-NEXT:    popl %ebp
375; X86-NEXT:    retl
376;
377; X64-LABEL: PR38622:
378; X64:       # %bb.0:
379; X64-NEXT:    movq %rdi, %rax
380; X64-NEXT:    shrq $11, %rax
381; X64-NEXT:    movabsq $4835703278458517, %rcx # imm = 0x112E0BE826D695
382; X64-NEXT:    mulq %rcx
383; X64-NEXT:    movq %rdx, %rax
384; X64-NEXT:    shrq $9, %rax
385; X64-NEXT:    imull $-294967296, %eax, %ecx # imm = 0xEE6B2800
386; X64-NEXT:    subl %ecx, %edi
387; X64-NEXT:    movl %edi, %edx
388; X64-NEXT:    retq
389  %2 = udiv i64 %0, 4000000000
390  %3 = urem i64 %0, 4000000000
391  %4 = trunc i64 %3 to i32
392  %5 = insertvalue { i64, i32 } undef, i64 %2, 0
393  %6 = insertvalue { i64, i32 } %5, i32 %4, 1
394  ret { i64, i32 } %6
395}
396
397define { i64, i32 } @PR38622_signed(i64) nounwind {
398; X86-LABEL: PR38622_signed:
399; X86:       # %bb.0:
400; X86-NEXT:    pushl %ebp
401; X86-NEXT:    pushl %ebx
402; X86-NEXT:    pushl %edi
403; X86-NEXT:    pushl %esi
404; X86-NEXT:    subl $12, %esp
405; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
406; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebp
407; X86-NEXT:    pushl $0
408; X86-NEXT:    pushl $-294967296 # imm = 0xEE6B2800
409; X86-NEXT:    pushl %ebp
410; X86-NEXT:    pushl %ebx
411; X86-NEXT:    calll __divdi3
412; X86-NEXT:    addl $16, %esp
413; X86-NEXT:    movl %eax, %esi
414; X86-NEXT:    movl %edx, %edi
415; X86-NEXT:    pushl $0
416; X86-NEXT:    pushl $-294967296 # imm = 0xEE6B2800
417; X86-NEXT:    pushl %ebp
418; X86-NEXT:    pushl %ebx
419; X86-NEXT:    calll __moddi3
420; X86-NEXT:    addl $16, %esp
421; X86-NEXT:    movl %eax, %ecx
422; X86-NEXT:    movl %esi, %eax
423; X86-NEXT:    movl %edi, %edx
424; X86-NEXT:    addl $12, %esp
425; X86-NEXT:    popl %esi
426; X86-NEXT:    popl %edi
427; X86-NEXT:    popl %ebx
428; X86-NEXT:    popl %ebp
429; X86-NEXT:    retl
430;
431; X64-LABEL: PR38622_signed:
432; X64:       # %bb.0:
433; X64-NEXT:    movabsq $1237940039285380275, %rcx # imm = 0x112E0BE826D694B3
434; X64-NEXT:    movq %rdi, %rax
435; X64-NEXT:    imulq %rcx
436; X64-NEXT:    movq %rdx, %rax
437; X64-NEXT:    shrq $63, %rax
438; X64-NEXT:    sarq $28, %rdx
439; X64-NEXT:    addq %rdx, %rax
440; X64-NEXT:    imull $-294967296, %eax, %ecx # imm = 0xEE6B2800
441; X64-NEXT:    subl %ecx, %edi
442; X64-NEXT:    movl %edi, %edx
443; X64-NEXT:    retq
444  %2 = sdiv i64 %0, 4000000000
445  %3 = srem i64 %0, 4000000000
446  %4 = trunc i64 %3 to i32
447  %5 = insertvalue { i64, i32 } undef, i64 %2, 0
448  %6 = insertvalue { i64, i32 } %5, i32 %4, 1
449  ret { i64, i32 } %6
450}
451
452define i64 @urem_i64_3(i64 %x) nounwind {
453; X86-LABEL: urem_i64_3:
454; X86:       # %bb.0: # %entry
455; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
456; X86-NEXT:    addl {{[0-9]+}}(%esp), %ecx
457; X86-NEXT:    adcl $0, %ecx
458; X86-NEXT:    movl $-1431655765, %edx # imm = 0xAAAAAAAB
459; X86-NEXT:    movl %ecx, %eax
460; X86-NEXT:    mull %edx
461; X86-NEXT:    shrl %edx
462; X86-NEXT:    leal (%edx,%edx,2), %eax
463; X86-NEXT:    subl %eax, %ecx
464; X86-NEXT:    movl %ecx, %eax
465; X86-NEXT:    xorl %edx, %edx
466; X86-NEXT:    retl
467;
468; X64-LABEL: urem_i64_3:
469; X64:       # %bb.0: # %entry
470; X64-NEXT:    movabsq $-6148914691236517205, %rcx # imm = 0xAAAAAAAAAAAAAAAB
471; X64-NEXT:    movq %rdi, %rax
472; X64-NEXT:    mulq %rcx
473; X64-NEXT:    shrq %rdx
474; X64-NEXT:    leaq (%rdx,%rdx,2), %rax
475; X64-NEXT:    subq %rax, %rdi
476; X64-NEXT:    movq %rdi, %rax
477; X64-NEXT:    retq
478entry:
479  %rem = urem i64 %x, 3
480  ret i64 %rem
481}
482
483define i64 @urem_i64_5(i64 %x) nounwind {
484; X86-LABEL: urem_i64_5:
485; X86:       # %bb.0: # %entry
486; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
487; X86-NEXT:    addl {{[0-9]+}}(%esp), %ecx
488; X86-NEXT:    adcl $0, %ecx
489; X86-NEXT:    movl $-858993459, %edx # imm = 0xCCCCCCCD
490; X86-NEXT:    movl %ecx, %eax
491; X86-NEXT:    mull %edx
492; X86-NEXT:    shrl $2, %edx
493; X86-NEXT:    leal (%edx,%edx,4), %eax
494; X86-NEXT:    subl %eax, %ecx
495; X86-NEXT:    movl %ecx, %eax
496; X86-NEXT:    xorl %edx, %edx
497; X86-NEXT:    retl
498;
499; X64-LABEL: urem_i64_5:
500; X64:       # %bb.0: # %entry
501; X64-NEXT:    movabsq $-3689348814741910323, %rcx # imm = 0xCCCCCCCCCCCCCCCD
502; X64-NEXT:    movq %rdi, %rax
503; X64-NEXT:    mulq %rcx
504; X64-NEXT:    shrq $2, %rdx
505; X64-NEXT:    leaq (%rdx,%rdx,4), %rax
506; X64-NEXT:    subq %rax, %rdi
507; X64-NEXT:    movq %rdi, %rax
508; X64-NEXT:    retq
509entry:
510  %rem = urem i64 %x, 5
511  ret i64 %rem
512}
513
514define i64 @urem_i64_15(i64 %x) nounwind {
515; X86-LABEL: urem_i64_15:
516; X86:       # %bb.0: # %entry
517; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
518; X86-NEXT:    addl {{[0-9]+}}(%esp), %ecx
519; X86-NEXT:    adcl $0, %ecx
520; X86-NEXT:    movl $-2004318071, %edx # imm = 0x88888889
521; X86-NEXT:    movl %ecx, %eax
522; X86-NEXT:    mull %edx
523; X86-NEXT:    shrl $3, %edx
524; X86-NEXT:    leal (%edx,%edx,4), %eax
525; X86-NEXT:    leal (%eax,%eax,2), %eax
526; X86-NEXT:    subl %eax, %ecx
527; X86-NEXT:    movl %ecx, %eax
528; X86-NEXT:    xorl %edx, %edx
529; X86-NEXT:    retl
530;
531; X64-LABEL: urem_i64_15:
532; X64:       # %bb.0: # %entry
533; X64-NEXT:    movabsq $-8608480567731124087, %rcx # imm = 0x8888888888888889
534; X64-NEXT:    movq %rdi, %rax
535; X64-NEXT:    mulq %rcx
536; X64-NEXT:    shrq $3, %rdx
537; X64-NEXT:    leaq (%rdx,%rdx,4), %rax
538; X64-NEXT:    leaq (%rax,%rax,2), %rax
539; X64-NEXT:    subq %rax, %rdi
540; X64-NEXT:    movq %rdi, %rax
541; X64-NEXT:    retq
542entry:
543  %rem = urem i64 %x, 15
544  ret i64 %rem
545}
546
547define i64 @urem_i64_17(i64 %x) nounwind {
548; X86-LABEL: urem_i64_17:
549; X86:       # %bb.0: # %entry
550; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
551; X86-NEXT:    addl {{[0-9]+}}(%esp), %ecx
552; X86-NEXT:    adcl $0, %ecx
553; X86-NEXT:    movl $-252645135, %edx # imm = 0xF0F0F0F1
554; X86-NEXT:    movl %ecx, %eax
555; X86-NEXT:    mull %edx
556; X86-NEXT:    movl %edx, %eax
557; X86-NEXT:    andl $-16, %eax
558; X86-NEXT:    shrl $4, %edx
559; X86-NEXT:    addl %eax, %edx
560; X86-NEXT:    subl %edx, %ecx
561; X86-NEXT:    movl %ecx, %eax
562; X86-NEXT:    xorl %edx, %edx
563; X86-NEXT:    retl
564;
565; X64-LABEL: urem_i64_17:
566; X64:       # %bb.0: # %entry
567; X64-NEXT:    movabsq $-1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F1
568; X64-NEXT:    movq %rdi, %rax
569; X64-NEXT:    mulq %rcx
570; X64-NEXT:    movq %rdx, %rax
571; X64-NEXT:    andq $-16, %rax
572; X64-NEXT:    shrq $4, %rdx
573; X64-NEXT:    addq %rax, %rdx
574; X64-NEXT:    subq %rdx, %rdi
575; X64-NEXT:    movq %rdi, %rax
576; X64-NEXT:    retq
577entry:
578  %rem = urem i64 %x, 17
579  ret i64 %rem
580}
581
582define i64 @urem_i64_255(i64 %x) nounwind {
583; X86-LABEL: urem_i64_255:
584; X86:       # %bb.0: # %entry
585; X86-NEXT:    pushl %esi
586; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
587; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
588; X86-NEXT:    movl %ecx, %eax
589; X86-NEXT:    addl %esi, %eax
590; X86-NEXT:    adcl $0, %eax
591; X86-NEXT:    movl $-2139062143, %edx # imm = 0x80808081
592; X86-NEXT:    mull %edx
593; X86-NEXT:    shrl $7, %edx
594; X86-NEXT:    movl %edx, %eax
595; X86-NEXT:    shll $8, %eax
596; X86-NEXT:    subl %eax, %edx
597; X86-NEXT:    addl %esi, %ecx
598; X86-NEXT:    adcl %edx, %ecx
599; X86-NEXT:    movl %ecx, %eax
600; X86-NEXT:    xorl %edx, %edx
601; X86-NEXT:    popl %esi
602; X86-NEXT:    retl
603;
604; X64-LABEL: urem_i64_255:
605; X64:       # %bb.0: # %entry
606; X64-NEXT:    movabsq $-9187201950435737471, %rcx # imm = 0x8080808080808081
607; X64-NEXT:    movq %rdi, %rax
608; X64-NEXT:    mulq %rcx
609; X64-NEXT:    shrq $7, %rdx
610; X64-NEXT:    movq %rdx, %rax
611; X64-NEXT:    shlq $8, %rax
612; X64-NEXT:    subq %rax, %rdx
613; X64-NEXT:    leaq (%rdx,%rdi), %rax
614; X64-NEXT:    retq
615entry:
616  %rem = urem i64 %x, 255
617  ret i64 %rem
618}
619
620define i64 @urem_i64_257(i64 %x) nounwind {
621; X86-LABEL: urem_i64_257:
622; X86:       # %bb.0: # %entry
623; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
624; X86-NEXT:    addl {{[0-9]+}}(%esp), %ecx
625; X86-NEXT:    adcl $0, %ecx
626; X86-NEXT:    movl $-16711935, %edx # imm = 0xFF00FF01
627; X86-NEXT:    movl %ecx, %eax
628; X86-NEXT:    mull %edx
629; X86-NEXT:    movl %edx, %eax
630; X86-NEXT:    andl $-256, %eax
631; X86-NEXT:    shrl $8, %edx
632; X86-NEXT:    addl %eax, %edx
633; X86-NEXT:    subl %edx, %ecx
634; X86-NEXT:    movl %ecx, %eax
635; X86-NEXT:    xorl %edx, %edx
636; X86-NEXT:    retl
637;
638; X64-LABEL: urem_i64_257:
639; X64:       # %bb.0: # %entry
640; X64-NEXT:    movabsq $-71777214294589695, %rcx # imm = 0xFF00FF00FF00FF01
641; X64-NEXT:    movq %rdi, %rax
642; X64-NEXT:    mulq %rcx
643; X64-NEXT:    movq %rdx, %rax
644; X64-NEXT:    andq $-256, %rax
645; X64-NEXT:    shrq $8, %rdx
646; X64-NEXT:    addq %rax, %rdx
647; X64-NEXT:    subq %rdx, %rdi
648; X64-NEXT:    movq %rdi, %rax
649; X64-NEXT:    retq
650entry:
651  %rem = urem i64 %x, 257
652  ret i64 %rem
653}
654
655define i64 @urem_i64_65535(i64 %x) nounwind {
656; X86-LABEL: urem_i64_65535:
657; X86:       # %bb.0: # %entry
658; X86-NEXT:    pushl %esi
659; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
660; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
661; X86-NEXT:    movl %ecx, %eax
662; X86-NEXT:    addl %esi, %eax
663; X86-NEXT:    adcl $0, %eax
664; X86-NEXT:    movl $-2147450879, %edx # imm = 0x80008001
665; X86-NEXT:    mull %edx
666; X86-NEXT:    shrl $15, %edx
667; X86-NEXT:    movl %edx, %eax
668; X86-NEXT:    shll $16, %eax
669; X86-NEXT:    subl %eax, %edx
670; X86-NEXT:    addl %esi, %ecx
671; X86-NEXT:    adcl %edx, %ecx
672; X86-NEXT:    movl %ecx, %eax
673; X86-NEXT:    xorl %edx, %edx
674; X86-NEXT:    popl %esi
675; X86-NEXT:    retl
676;
677; X64-LABEL: urem_i64_65535:
678; X64:       # %bb.0: # %entry
679; X64-NEXT:    movabsq $-9223231297218904063, %rcx # imm = 0x8000800080008001
680; X64-NEXT:    movq %rdi, %rax
681; X64-NEXT:    mulq %rcx
682; X64-NEXT:    shrq $15, %rdx
683; X64-NEXT:    movq %rdx, %rax
684; X64-NEXT:    shlq $16, %rax
685; X64-NEXT:    subq %rax, %rdx
686; X64-NEXT:    leaq (%rdx,%rdi), %rax
687; X64-NEXT:    retq
688entry:
689  %rem = urem i64 %x, 65535
690  ret i64 %rem
691}
692
693define i64 @urem_i64_65537(i64 %x) nounwind {
694; X86-LABEL: urem_i64_65537:
695; X86:       # %bb.0: # %entry
696; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
697; X86-NEXT:    addl {{[0-9]+}}(%esp), %ecx
698; X86-NEXT:    adcl $0, %ecx
699; X86-NEXT:    movl $-65535, %edx # imm = 0xFFFF0001
700; X86-NEXT:    movl %ecx, %eax
701; X86-NEXT:    mull %edx
702; X86-NEXT:    movl %edx, %eax
703; X86-NEXT:    shrl $16, %eax
704; X86-NEXT:    shldl $16, %edx, %eax
705; X86-NEXT:    subl %eax, %ecx
706; X86-NEXT:    movl %ecx, %eax
707; X86-NEXT:    xorl %edx, %edx
708; X86-NEXT:    retl
709;
710; X64-LABEL: urem_i64_65537:
711; X64:       # %bb.0: # %entry
712; X64-NEXT:    movabsq $-281470681808895, %rcx # imm = 0xFFFF0000FFFF0001
713; X64-NEXT:    movq %rdi, %rax
714; X64-NEXT:    mulq %rcx
715; X64-NEXT:    movq %rdx, %rax
716; X64-NEXT:    andq $-65536, %rax # imm = 0xFFFF0000
717; X64-NEXT:    shrq $16, %rdx
718; X64-NEXT:    addq %rax, %rdx
719; X64-NEXT:    subq %rdx, %rdi
720; X64-NEXT:    movq %rdi, %rax
721; X64-NEXT:    retq
722entry:
723  %rem = urem i64 %x, 65537
724  ret i64 %rem
725}
726
727define i64 @urem_i64_12(i64 %x) nounwind {
728; X86-LABEL: urem_i64_12:
729; X86:       # %bb.0: # %entry
730; X86-NEXT:    pushl %esi
731; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
732; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
733; X86-NEXT:    movl %ecx, %eax
734; X86-NEXT:    shrl $2, %eax
735; X86-NEXT:    shldl $30, %esi, %ecx
736; X86-NEXT:    addl %eax, %ecx
737; X86-NEXT:    adcl $0, %ecx
738; X86-NEXT:    movl $-1431655765, %edx # imm = 0xAAAAAAAB
739; X86-NEXT:    movl %ecx, %eax
740; X86-NEXT:    mull %edx
741; X86-NEXT:    shrl %edx
742; X86-NEXT:    leal (%edx,%edx,2), %eax
743; X86-NEXT:    subl %eax, %ecx
744; X86-NEXT:    andl $3, %esi
745; X86-NEXT:    leal (%esi,%ecx,4), %eax
746; X86-NEXT:    xorl %edx, %edx
747; X86-NEXT:    popl %esi
748; X86-NEXT:    retl
749;
750; X64-LABEL: urem_i64_12:
751; X64:       # %bb.0: # %entry
752; X64-NEXT:    movabsq $-6148914691236517205, %rcx # imm = 0xAAAAAAAAAAAAAAAB
753; X64-NEXT:    movq %rdi, %rax
754; X64-NEXT:    mulq %rcx
755; X64-NEXT:    shrq %rdx
756; X64-NEXT:    andq $-4, %rdx
757; X64-NEXT:    leaq (%rdx,%rdx,2), %rax
758; X64-NEXT:    subq %rax, %rdi
759; X64-NEXT:    movq %rdi, %rax
760; X64-NEXT:    retq
761entry:
762  %rem = urem i64 %x, 12
763  ret i64 %rem
764}
765
766define i64 @udiv_i64_3(i64 %x) nounwind {
767; X86-LABEL: udiv_i64_3:
768; X86:       # %bb.0: # %entry
769; X86-NEXT:    pushl %ebx
770; X86-NEXT:    pushl %edi
771; X86-NEXT:    pushl %esi
772; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
773; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
774; X86-NEXT:    movl %ecx, %esi
775; X86-NEXT:    addl %edi, %esi
776; X86-NEXT:    adcl $0, %esi
777; X86-NEXT:    movl $-1431655765, %ebx # imm = 0xAAAAAAAB
778; X86-NEXT:    movl %esi, %eax
779; X86-NEXT:    mull %ebx
780; X86-NEXT:    shrl %edx
781; X86-NEXT:    leal (%edx,%edx,2), %eax
782; X86-NEXT:    subl %eax, %esi
783; X86-NEXT:    subl %esi, %ecx
784; X86-NEXT:    sbbl $0, %edi
785; X86-NEXT:    movl %ecx, %eax
786; X86-NEXT:    mull %ebx
787; X86-NEXT:    imull $-1431655766, %ecx, %ecx # imm = 0xAAAAAAAA
788; X86-NEXT:    addl %ecx, %edx
789; X86-NEXT:    imull $-1431655765, %edi, %ecx # imm = 0xAAAAAAAB
790; X86-NEXT:    addl %ecx, %edx
791; X86-NEXT:    popl %esi
792; X86-NEXT:    popl %edi
793; X86-NEXT:    popl %ebx
794; X86-NEXT:    retl
795;
796; X64-LABEL: udiv_i64_3:
797; X64:       # %bb.0: # %entry
798; X64-NEXT:    movq %rdi, %rax
799; X64-NEXT:    movabsq $-6148914691236517205, %rcx # imm = 0xAAAAAAAAAAAAAAAB
800; X64-NEXT:    mulq %rcx
801; X64-NEXT:    movq %rdx, %rax
802; X64-NEXT:    shrq %rax
803; X64-NEXT:    retq
804entry:
805  %rem = udiv i64 %x, 3
806  ret i64 %rem
807}
808
809define i64 @udiv_i64_5(i64 %x) nounwind {
810; X86-LABEL: udiv_i64_5:
811; X86:       # %bb.0: # %entry
812; X86-NEXT:    pushl %ebx
813; X86-NEXT:    pushl %edi
814; X86-NEXT:    pushl %esi
815; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
816; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
817; X86-NEXT:    movl %ecx, %esi
818; X86-NEXT:    addl %edi, %esi
819; X86-NEXT:    adcl $0, %esi
820; X86-NEXT:    movl $-858993459, %ebx # imm = 0xCCCCCCCD
821; X86-NEXT:    movl %esi, %eax
822; X86-NEXT:    mull %ebx
823; X86-NEXT:    shrl $2, %edx
824; X86-NEXT:    leal (%edx,%edx,4), %eax
825; X86-NEXT:    subl %eax, %esi
826; X86-NEXT:    subl %esi, %ecx
827; X86-NEXT:    sbbl $0, %edi
828; X86-NEXT:    movl %ecx, %eax
829; X86-NEXT:    mull %ebx
830; X86-NEXT:    imull $-858993460, %ecx, %ecx # imm = 0xCCCCCCCC
831; X86-NEXT:    addl %ecx, %edx
832; X86-NEXT:    imull $-858993459, %edi, %ecx # imm = 0xCCCCCCCD
833; X86-NEXT:    addl %ecx, %edx
834; X86-NEXT:    popl %esi
835; X86-NEXT:    popl %edi
836; X86-NEXT:    popl %ebx
837; X86-NEXT:    retl
838;
839; X64-LABEL: udiv_i64_5:
840; X64:       # %bb.0: # %entry
841; X64-NEXT:    movq %rdi, %rax
842; X64-NEXT:    movabsq $-3689348814741910323, %rcx # imm = 0xCCCCCCCCCCCCCCCD
843; X64-NEXT:    mulq %rcx
844; X64-NEXT:    movq %rdx, %rax
845; X64-NEXT:    shrq $2, %rax
846; X64-NEXT:    retq
847entry:
848  %rem = udiv i64 %x, 5
849  ret i64 %rem
850}
851
852define i64 @udiv_i64_15(i64 %x) nounwind {
853; X86-LABEL: udiv_i64_15:
854; X86:       # %bb.0: # %entry
855; X86-NEXT:    pushl %edi
856; X86-NEXT:    pushl %esi
857; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
858; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
859; X86-NEXT:    movl %ecx, %esi
860; X86-NEXT:    addl %edi, %esi
861; X86-NEXT:    adcl $0, %esi
862; X86-NEXT:    movl $-2004318071, %edx # imm = 0x88888889
863; X86-NEXT:    movl %esi, %eax
864; X86-NEXT:    mull %edx
865; X86-NEXT:    shrl $3, %edx
866; X86-NEXT:    leal (%edx,%edx,4), %eax
867; X86-NEXT:    leal (%eax,%eax,2), %eax
868; X86-NEXT:    subl %eax, %esi
869; X86-NEXT:    subl %esi, %ecx
870; X86-NEXT:    sbbl $0, %edi
871; X86-NEXT:    movl $-286331153, %edx # imm = 0xEEEEEEEF
872; X86-NEXT:    movl %ecx, %eax
873; X86-NEXT:    mull %edx
874; X86-NEXT:    imull $-286331154, %ecx, %ecx # imm = 0xEEEEEEEE
875; X86-NEXT:    addl %ecx, %edx
876; X86-NEXT:    imull $-286331153, %edi, %ecx # imm = 0xEEEEEEEF
877; X86-NEXT:    addl %ecx, %edx
878; X86-NEXT:    popl %esi
879; X86-NEXT:    popl %edi
880; X86-NEXT:    retl
881;
882; X64-LABEL: udiv_i64_15:
883; X64:       # %bb.0: # %entry
884; X64-NEXT:    movq %rdi, %rax
885; X64-NEXT:    movabsq $-8608480567731124087, %rcx # imm = 0x8888888888888889
886; X64-NEXT:    mulq %rcx
887; X64-NEXT:    movq %rdx, %rax
888; X64-NEXT:    shrq $3, %rax
889; X64-NEXT:    retq
890entry:
891  %rem = udiv i64 %x, 15
892  ret i64 %rem
893}
894
895define i64 @udiv_i64_17(i64 %x) nounwind {
896; X86-LABEL: udiv_i64_17:
897; X86:       # %bb.0: # %entry
898; X86-NEXT:    pushl %ebx
899; X86-NEXT:    pushl %edi
900; X86-NEXT:    pushl %esi
901; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
902; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
903; X86-NEXT:    movl %ecx, %esi
904; X86-NEXT:    addl %edi, %esi
905; X86-NEXT:    adcl $0, %esi
906; X86-NEXT:    movl $-252645135, %ebx # imm = 0xF0F0F0F1
907; X86-NEXT:    movl %esi, %eax
908; X86-NEXT:    mull %ebx
909; X86-NEXT:    movl %edx, %eax
910; X86-NEXT:    andl $-16, %eax
911; X86-NEXT:    shrl $4, %edx
912; X86-NEXT:    addl %eax, %edx
913; X86-NEXT:    subl %edx, %esi
914; X86-NEXT:    subl %esi, %ecx
915; X86-NEXT:    sbbl $0, %edi
916; X86-NEXT:    movl %ecx, %eax
917; X86-NEXT:    mull %ebx
918; X86-NEXT:    imull $-252645136, %ecx, %ecx # imm = 0xF0F0F0F0
919; X86-NEXT:    addl %ecx, %edx
920; X86-NEXT:    imull $-252645135, %edi, %ecx # imm = 0xF0F0F0F1
921; X86-NEXT:    addl %ecx, %edx
922; X86-NEXT:    popl %esi
923; X86-NEXT:    popl %edi
924; X86-NEXT:    popl %ebx
925; X86-NEXT:    retl
926;
927; X64-LABEL: udiv_i64_17:
928; X64:       # %bb.0: # %entry
929; X64-NEXT:    movq %rdi, %rax
930; X64-NEXT:    movabsq $-1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F1
931; X64-NEXT:    mulq %rcx
932; X64-NEXT:    movq %rdx, %rax
933; X64-NEXT:    shrq $4, %rax
934; X64-NEXT:    retq
935entry:
936  %rem = udiv i64 %x, 17
937  ret i64 %rem
938}
939
940define i64 @udiv_i64_255(i64 %x) nounwind {
941; X86-LABEL: udiv_i64_255:
942; X86:       # %bb.0: # %entry
943; X86-NEXT:    pushl %esi
944; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
945; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
946; X86-NEXT:    movl %ecx, %eax
947; X86-NEXT:    addl %esi, %eax
948; X86-NEXT:    adcl $0, %eax
949; X86-NEXT:    movl $-2139062143, %edx # imm = 0x80808081
950; X86-NEXT:    mull %edx
951; X86-NEXT:    shrl $7, %edx
952; X86-NEXT:    movl %edx, %eax
953; X86-NEXT:    shll $8, %eax
954; X86-NEXT:    subl %eax, %edx
955; X86-NEXT:    movl %ecx, %eax
956; X86-NEXT:    addl %esi, %eax
957; X86-NEXT:    adcl %edx, %eax
958; X86-NEXT:    subl %eax, %ecx
959; X86-NEXT:    sbbl $0, %esi
960; X86-NEXT:    movl $-16843009, %edx # imm = 0xFEFEFEFF
961; X86-NEXT:    movl %ecx, %eax
962; X86-NEXT:    mull %edx
963; X86-NEXT:    imull $-16843010, %ecx, %ecx # imm = 0xFEFEFEFE
964; X86-NEXT:    addl %ecx, %edx
965; X86-NEXT:    imull $-16843009, %esi, %ecx # imm = 0xFEFEFEFF
966; X86-NEXT:    addl %ecx, %edx
967; X86-NEXT:    popl %esi
968; X86-NEXT:    retl
969;
970; X64-LABEL: udiv_i64_255:
971; X64:       # %bb.0: # %entry
972; X64-NEXT:    movq %rdi, %rax
973; X64-NEXT:    movabsq $-9187201950435737471, %rcx # imm = 0x8080808080808081
974; X64-NEXT:    mulq %rcx
975; X64-NEXT:    movq %rdx, %rax
976; X64-NEXT:    shrq $7, %rax
977; X64-NEXT:    retq
978entry:
979  %rem = udiv i64 %x, 255
980  ret i64 %rem
981}
982
983define i64 @udiv_i64_257(i64 %x) nounwind {
984; X86-LABEL: udiv_i64_257:
985; X86:       # %bb.0: # %entry
986; X86-NEXT:    pushl %ebx
987; X86-NEXT:    pushl %edi
988; X86-NEXT:    pushl %esi
989; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
990; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
991; X86-NEXT:    movl %ecx, %esi
992; X86-NEXT:    addl %edi, %esi
993; X86-NEXT:    adcl $0, %esi
994; X86-NEXT:    movl $-16711935, %ebx # imm = 0xFF00FF01
995; X86-NEXT:    movl %esi, %eax
996; X86-NEXT:    mull %ebx
997; X86-NEXT:    movl %edx, %eax
998; X86-NEXT:    andl $-256, %eax
999; X86-NEXT:    shrl $8, %edx
1000; X86-NEXT:    addl %eax, %edx
1001; X86-NEXT:    subl %edx, %esi
1002; X86-NEXT:    subl %esi, %ecx
1003; X86-NEXT:    sbbl $0, %edi
1004; X86-NEXT:    movl %ecx, %eax
1005; X86-NEXT:    mull %ebx
1006; X86-NEXT:    imull $-16711936, %ecx, %ecx # imm = 0xFF00FF00
1007; X86-NEXT:    addl %ecx, %edx
1008; X86-NEXT:    imull $-16711935, %edi, %ecx # imm = 0xFF00FF01
1009; X86-NEXT:    addl %ecx, %edx
1010; X86-NEXT:    popl %esi
1011; X86-NEXT:    popl %edi
1012; X86-NEXT:    popl %ebx
1013; X86-NEXT:    retl
1014;
1015; X64-LABEL: udiv_i64_257:
1016; X64:       # %bb.0: # %entry
1017; X64-NEXT:    movq %rdi, %rax
1018; X64-NEXT:    movabsq $-71777214294589695, %rcx # imm = 0xFF00FF00FF00FF01
1019; X64-NEXT:    mulq %rcx
1020; X64-NEXT:    movq %rdx, %rax
1021; X64-NEXT:    shrq $8, %rax
1022; X64-NEXT:    retq
1023entry:
1024  %rem = udiv i64 %x, 257
1025  ret i64 %rem
1026}
1027
1028define i64 @udiv_i64_65535(i64 %x) nounwind {
1029; X86-LABEL: udiv_i64_65535:
1030; X86:       # %bb.0: # %entry
1031; X86-NEXT:    pushl %esi
1032; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
1033; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
1034; X86-NEXT:    movl %ecx, %eax
1035; X86-NEXT:    addl %esi, %eax
1036; X86-NEXT:    adcl $0, %eax
1037; X86-NEXT:    movl $-2147450879, %edx # imm = 0x80008001
1038; X86-NEXT:    mull %edx
1039; X86-NEXT:    shrl $15, %edx
1040; X86-NEXT:    movl %edx, %eax
1041; X86-NEXT:    shll $16, %eax
1042; X86-NEXT:    subl %eax, %edx
1043; X86-NEXT:    movl %ecx, %eax
1044; X86-NEXT:    addl %esi, %eax
1045; X86-NEXT:    adcl %edx, %eax
1046; X86-NEXT:    subl %eax, %ecx
1047; X86-NEXT:    sbbl $0, %esi
1048; X86-NEXT:    movl $-65537, %edx # imm = 0xFFFEFFFF
1049; X86-NEXT:    movl %ecx, %eax
1050; X86-NEXT:    mull %edx
1051; X86-NEXT:    imull $-65538, %ecx, %ecx # imm = 0xFFFEFFFE
1052; X86-NEXT:    addl %ecx, %edx
1053; X86-NEXT:    movl %esi, %ecx
1054; X86-NEXT:    shll $16, %ecx
1055; X86-NEXT:    addl %esi, %ecx
1056; X86-NEXT:    subl %ecx, %edx
1057; X86-NEXT:    popl %esi
1058; X86-NEXT:    retl
1059;
1060; X64-LABEL: udiv_i64_65535:
1061; X64:       # %bb.0: # %entry
1062; X64-NEXT:    movq %rdi, %rax
1063; X64-NEXT:    movabsq $-9223231297218904063, %rcx # imm = 0x8000800080008001
1064; X64-NEXT:    mulq %rcx
1065; X64-NEXT:    movq %rdx, %rax
1066; X64-NEXT:    shrq $15, %rax
1067; X64-NEXT:    retq
1068entry:
1069  %rem = udiv i64 %x, 65535
1070  ret i64 %rem
1071}
1072
1073define i64 @udiv_i64_65537(i64 %x) nounwind {
1074; X86-LABEL: udiv_i64_65537:
1075; X86:       # %bb.0: # %entry
1076; X86-NEXT:    pushl %ebx
1077; X86-NEXT:    pushl %edi
1078; X86-NEXT:    pushl %esi
1079; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
1080; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
1081; X86-NEXT:    movl %ecx, %esi
1082; X86-NEXT:    addl %edi, %esi
1083; X86-NEXT:    adcl $0, %esi
1084; X86-NEXT:    movl $-65535, %ebx # imm = 0xFFFF0001
1085; X86-NEXT:    movl %esi, %eax
1086; X86-NEXT:    mull %ebx
1087; X86-NEXT:    movl %edx, %eax
1088; X86-NEXT:    shrl $16, %eax
1089; X86-NEXT:    shldl $16, %edx, %eax
1090; X86-NEXT:    subl %eax, %esi
1091; X86-NEXT:    subl %esi, %ecx
1092; X86-NEXT:    sbbl $0, %edi
1093; X86-NEXT:    movl %ecx, %eax
1094; X86-NEXT:    mull %ebx
1095; X86-NEXT:    shll $16, %ecx
1096; X86-NEXT:    subl %ecx, %edx
1097; X86-NEXT:    movl %edi, %ecx
1098; X86-NEXT:    shll $16, %ecx
1099; X86-NEXT:    subl %ecx, %edi
1100; X86-NEXT:    addl %edi, %edx
1101; X86-NEXT:    popl %esi
1102; X86-NEXT:    popl %edi
1103; X86-NEXT:    popl %ebx
1104; X86-NEXT:    retl
1105;
1106; X64-LABEL: udiv_i64_65537:
1107; X64:       # %bb.0: # %entry
1108; X64-NEXT:    movq %rdi, %rax
1109; X64-NEXT:    movabsq $-281470681808895, %rcx # imm = 0xFFFF0000FFFF0001
1110; X64-NEXT:    mulq %rcx
1111; X64-NEXT:    movq %rdx, %rax
1112; X64-NEXT:    shrq $16, %rax
1113; X64-NEXT:    retq
1114entry:
1115  %rem = udiv i64 %x, 65537
1116  ret i64 %rem
1117}
1118
1119define i64 @udiv_i64_12(i64 %x) nounwind {
1120; X86-LABEL: udiv_i64_12:
1121; X86:       # %bb.0: # %entry
1122; X86-NEXT:    pushl %ebx
1123; X86-NEXT:    pushl %edi
1124; X86-NEXT:    pushl %esi
1125; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
1126; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
1127; X86-NEXT:    shrdl $2, %edi, %ecx
1128; X86-NEXT:    shrl $2, %edi
1129; X86-NEXT:    movl %ecx, %esi
1130; X86-NEXT:    addl %edi, %esi
1131; X86-NEXT:    adcl $0, %esi
1132; X86-NEXT:    movl $-1431655765, %ebx # imm = 0xAAAAAAAB
1133; X86-NEXT:    movl %esi, %eax
1134; X86-NEXT:    mull %ebx
1135; X86-NEXT:    shrl %edx
1136; X86-NEXT:    leal (%edx,%edx,2), %eax
1137; X86-NEXT:    subl %eax, %esi
1138; X86-NEXT:    subl %esi, %ecx
1139; X86-NEXT:    sbbl $0, %edi
1140; X86-NEXT:    movl %ecx, %eax
1141; X86-NEXT:    mull %ebx
1142; X86-NEXT:    imull $-1431655766, %ecx, %ecx # imm = 0xAAAAAAAA
1143; X86-NEXT:    addl %ecx, %edx
1144; X86-NEXT:    imull $-1431655765, %edi, %ecx # imm = 0xAAAAAAAB
1145; X86-NEXT:    addl %ecx, %edx
1146; X86-NEXT:    popl %esi
1147; X86-NEXT:    popl %edi
1148; X86-NEXT:    popl %ebx
1149; X86-NEXT:    retl
1150;
1151; X64-LABEL: udiv_i64_12:
1152; X64:       # %bb.0: # %entry
1153; X64-NEXT:    movq %rdi, %rax
1154; X64-NEXT:    movabsq $-6148914691236517205, %rcx # imm = 0xAAAAAAAAAAAAAAAB
1155; X64-NEXT:    mulq %rcx
1156; X64-NEXT:    movq %rdx, %rax
1157; X64-NEXT:    shrq $3, %rax
1158; X64-NEXT:    retq
1159entry:
1160  %rem = udiv i64 %x, 12
1161  ret i64 %rem
1162}
1163
1164; Make sure we don't inline expand for optsize.
1165define i64 @urem_i64_3_optsize(i64 %x) nounwind optsize {
1166; X86-LABEL: urem_i64_3_optsize:
1167; X86:       # %bb.0: # %entry
1168; X86-NEXT:    subl $12, %esp
1169; X86-NEXT:    pushl $0
1170; X86-NEXT:    pushl $3
1171; X86-NEXT:    pushl {{[0-9]+}}(%esp)
1172; X86-NEXT:    pushl {{[0-9]+}}(%esp)
1173; X86-NEXT:    calll __umoddi3
1174; X86-NEXT:    addl $28, %esp
1175; X86-NEXT:    retl
1176;
1177; X64-LABEL: urem_i64_3_optsize:
1178; X64:       # %bb.0: # %entry
1179; X64-NEXT:    movabsq $-6148914691236517205, %rcx # imm = 0xAAAAAAAAAAAAAAAB
1180; X64-NEXT:    movq %rdi, %rax
1181; X64-NEXT:    mulq %rcx
1182; X64-NEXT:    shrq %rdx
1183; X64-NEXT:    leaq (%rdx,%rdx,2), %rax
1184; X64-NEXT:    subq %rax, %rdi
1185; X64-NEXT:    movq %rdi, %rax
1186; X64-NEXT:    retq
1187entry:
1188  %rem = urem i64 %x, 3
1189  ret i64 %rem
1190}
1191
1192