xref: /llvm-project/llvm/test/CodeGen/X86/bmi-out-of-order.ll (revision 7b3bbd83c0c24087072ec5b22a76799ab31f87d5) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+cmov,+bmi | FileCheck %s --check-prefix=X86
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+cmov,+bmi | FileCheck %s --check-prefix=X64
4
5define i32 @blsmsk_used2(i32 %a) nounwind {
6; X86-LABEL: blsmsk_used2:
7; X86:       # %bb.0: # %entry
8; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
9; X86-NEXT:    leal -1(%eax), %ecx
10; X86-NEXT:    xorl %ecx, %eax
11; X86-NEXT:    imull %ecx, %eax
12; X86-NEXT:    retl
13;
14; X64-LABEL: blsmsk_used2:
15; X64:       # %bb.0: # %entry
16; X64-NEXT:    # kill: def $edi killed $edi def $rdi
17; X64-NEXT:    leal -1(%rdi), %eax
18; X64-NEXT:    xorl %eax, %edi
19; X64-NEXT:    imull %edi, %eax
20; X64-NEXT:    retq
21entry:
22  %sub = add i32 %a, -1
23  %xor = xor i32 %sub, %a
24  %mul = mul i32 %xor, %sub
25  ret i32 %mul
26}
27
28define i64 @blsmask_through1(i64 %a, i64 %b) nounwind {
29; X86-LABEL: blsmask_through1:
30; X86:       # %bb.0: # %entry
31; X86-NEXT:    pushl %esi
32; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
33; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
34; X86-NEXT:    movl %ecx, %eax
35; X86-NEXT:    addl $-1, %eax
36; X86-NEXT:    movl %esi, %edx
37; X86-NEXT:    adcl $-1, %edx
38; X86-NEXT:    xorl {{[0-9]+}}(%esp), %edx
39; X86-NEXT:    xorl {{[0-9]+}}(%esp), %eax
40; X86-NEXT:    xorl %ecx, %eax
41; X86-NEXT:    xorl %esi, %edx
42; X86-NEXT:    popl %esi
43; X86-NEXT:    retl
44;
45; X64-LABEL: blsmask_through1:
46; X64:       # %bb.0: # %entry
47; X64-NEXT:    blsmskq %rsi, %rax
48; X64-NEXT:    xorq %rdi, %rax
49; X64-NEXT:    retq
50entry:
51  %sub = add i64 %b, -1
52  %0 = xor i64 %sub, %a
53  %xor1 = xor i64 %0, %b
54  ret i64 %xor1
55}
56
57define i32 @blsmask_through2(i32 %a, i32 %b, i32 %c) nounwind {
58; X86-LABEL: blsmask_through2:
59; X86:       # %bb.0: # %entry
60; X86-NEXT:    blsmskl {{[0-9]+}}(%esp), %eax
61; X86-NEXT:    xorl {{[0-9]+}}(%esp), %eax
62; X86-NEXT:    xorl {{[0-9]+}}(%esp), %eax
63; X86-NEXT:    retl
64;
65; X64-LABEL: blsmask_through2:
66; X64:       # %bb.0: # %entry
67; X64-NEXT:    blsmskl %esi, %eax
68; X64-NEXT:    xorl %edx, %edi
69; X64-NEXT:    xorl %edi, %eax
70; X64-NEXT:    retq
71entry:
72  %sub = add nsw i32 %b, -1
73  %0 = xor i32 %sub, %a
74  %1 = xor i32 %0, %c
75  %xor2 = xor i32 %1, %b
76  ret i32 %xor2
77}
78
79define i64 @blsmask_through3(i64 %a, i64 %b, i64 %c, i64 %d) nounwind {
80; X86-LABEL: blsmask_through3:
81; X86:       # %bb.0: # %entry
82; X86-NEXT:    pushl %esi
83; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
84; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
85; X86-NEXT:    movl %esi, %eax
86; X86-NEXT:    addl $-1, %eax
87; X86-NEXT:    movl %ecx, %edx
88; X86-NEXT:    adcl $-1, %edx
89; X86-NEXT:    xorl {{[0-9]+}}(%esp), %edx
90; X86-NEXT:    xorl {{[0-9]+}}(%esp), %eax
91; X86-NEXT:    xorl {{[0-9]+}}(%esp), %edx
92; X86-NEXT:    xorl {{[0-9]+}}(%esp), %eax
93; X86-NEXT:    xorl {{[0-9]+}}(%esp), %edx
94; X86-NEXT:    xorl {{[0-9]+}}(%esp), %eax
95; X86-NEXT:    xorl %esi, %eax
96; X86-NEXT:    xorl %ecx, %edx
97; X86-NEXT:    popl %esi
98; X86-NEXT:    retl
99;
100; X64-LABEL: blsmask_through3:
101; X64:       # %bb.0: # %entry
102; X64-NEXT:    leaq -1(%rsi), %rax
103; X64-NEXT:    xorq %rdx, %rdi
104; X64-NEXT:    xorq %rdi, %rax
105; X64-NEXT:    xorq %rsi, %rcx
106; X64-NEXT:    xorq %rcx, %rax
107; X64-NEXT:    retq
108entry:
109  %sub = add nsw i64 %b, -1
110  %0 = xor i64 %sub, %a
111  %1 = xor i64 %0, %c
112  %2 = xor i64 %1, %d
113  %xor3 = xor i64 %2, %b
114  ret i64 %xor3
115}
116
117define i32 @blsmask_through1_used1(i32 %a, i32 %b) nounwind {
118; X86-LABEL: blsmask_through1_used1:
119; X86:       # %bb.0: # %entry
120; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
121; X86-NEXT:    leal -1(%ecx), %eax
122; X86-NEXT:    xorl {{[0-9]+}}(%esp), %eax
123; X86-NEXT:    xorl %eax, %ecx
124; X86-NEXT:    incl %eax
125; X86-NEXT:    orl %ecx, %eax
126; X86-NEXT:    retl
127;
128; X64-LABEL: blsmask_through1_used1:
129; X64:       # %bb.0: # %entry
130; X64-NEXT:    # kill: def $esi killed $esi def $rsi
131; X64-NEXT:    leal -1(%rsi), %eax
132; X64-NEXT:    xorl %edi, %eax
133; X64-NEXT:    xorl %eax, %esi
134; X64-NEXT:    incl %eax
135; X64-NEXT:    orl %esi, %eax
136; X64-NEXT:    retq
137entry:
138  %sub = add i32 %b, -1
139  %xor = xor i32 %sub, %a
140  %xor1 = xor i32 %xor, %b
141  %add = add i32 %xor, 1
142  %or = or i32 %add, %xor1
143  ret i32 %or
144}
145
146define i64 @blsmask_through1_used2(i64 %a, i64 %b) nounwind {
147; X86-LABEL: blsmask_through1_used2:
148; X86:       # %bb.0: # %entry
149; X86-NEXT:    pushl %ebp
150; X86-NEXT:    pushl %ebx
151; X86-NEXT:    pushl %edi
152; X86-NEXT:    pushl %esi
153; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
154; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
155; X86-NEXT:    movl %ecx, %edi
156; X86-NEXT:    addl $-1, %edi
157; X86-NEXT:    movl %esi, %ebp
158; X86-NEXT:    adcl $-1, %ebp
159; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
160; X86-NEXT:    xorl %ebp, %ebx
161; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
162; X86-NEXT:    xorl %edi, %eax
163; X86-NEXT:    xorl %ebx, %esi
164; X86-NEXT:    xorl %eax, %ecx
165; X86-NEXT:    imull %eax, %ebp
166; X86-NEXT:    mull %edi
167; X86-NEXT:    addl %ebp, %edx
168; X86-NEXT:    imull %edi, %ebx
169; X86-NEXT:    addl %ebx, %edx
170; X86-NEXT:    orl %esi, %edx
171; X86-NEXT:    orl %ecx, %eax
172; X86-NEXT:    popl %esi
173; X86-NEXT:    popl %edi
174; X86-NEXT:    popl %ebx
175; X86-NEXT:    popl %ebp
176; X86-NEXT:    retl
177;
178; X64-LABEL: blsmask_through1_used2:
179; X64:       # %bb.0: # %entry
180; X64-NEXT:    leaq -1(%rsi), %rax
181; X64-NEXT:    xorq %rax, %rdi
182; X64-NEXT:    xorq %rdi, %rsi
183; X64-NEXT:    imulq %rdi, %rax
184; X64-NEXT:    orq %rsi, %rax
185; X64-NEXT:    retq
186entry:
187  %sub = add i64 %b, -1
188  %xor = xor i64 %sub, %a
189  %xor1 = xor i64 %xor, %b
190  %mul = mul i64 %xor, %sub
191  %or = or i64 %mul, %xor1
192  ret i64 %or
193}
194
195define i32 @blsi_used2(i32 %a) nounwind {
196; X86-LABEL: blsi_used2:
197; X86:       # %bb.0: # %entry
198; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
199; X86-NEXT:    movl %eax, %ecx
200; X86-NEXT:    negl %ecx
201; X86-NEXT:    andl %ecx, %eax
202; X86-NEXT:    imull %ecx, %eax
203; X86-NEXT:    retl
204;
205; X64-LABEL: blsi_used2:
206; X64:       # %bb.0: # %entry
207; X64-NEXT:    movl %edi, %eax
208; X64-NEXT:    negl %eax
209; X64-NEXT:    andl %eax, %edi
210; X64-NEXT:    imull %edi, %eax
211; X64-NEXT:    retq
212entry:
213  %sub = sub nsw i32 0, %a
214  %and = and i32 %sub, %a
215  %mul = mul nsw i32 %and, %sub
216  ret i32 %mul
217}
218
219define i64 @blsi_through1(i64 %a, i64 %b) nounwind {
220; X86-LABEL: blsi_through1:
221; X86:       # %bb.0: # %entry
222; X86-NEXT:    pushl %esi
223; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
224; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
225; X86-NEXT:    xorl %edx, %edx
226; X86-NEXT:    movl %ecx, %eax
227; X86-NEXT:    negl %eax
228; X86-NEXT:    sbbl %esi, %edx
229; X86-NEXT:    andl {{[0-9]+}}(%esp), %edx
230; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
231; X86-NEXT:    andl %esi, %edx
232; X86-NEXT:    andl %ecx, %eax
233; X86-NEXT:    popl %esi
234; X86-NEXT:    retl
235;
236; X64-LABEL: blsi_through1:
237; X64:       # %bb.0: # %entry
238; X64-NEXT:    blsiq %rsi, %rax
239; X64-NEXT:    andq %rdi, %rax
240; X64-NEXT:    retq
241entry:
242  %sub = sub nsw i64 0, %b
243  %and = and i64 %sub, %a
244  %and1 = and i64 %and, %b
245  ret i64 %and1
246}
247
248define i32 @blsi_through2(i32 %a, i32 %b, i32 %c) nounwind {
249; X86-LABEL: blsi_through2:
250; X86:       # %bb.0: # %entry
251; X86-NEXT:    blsil {{[0-9]+}}(%esp), %eax
252; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
253; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
254; X86-NEXT:    retl
255;
256; X64-LABEL: blsi_through2:
257; X64:       # %bb.0: # %entry
258; X64-NEXT:    blsil %esi, %eax
259; X64-NEXT:    andl %edx, %edi
260; X64-NEXT:    andl %edi, %eax
261; X64-NEXT:    retq
262entry:
263  %sub = sub i32 0, %b
264  %and = and i32 %sub, %a
265  %0 = and i32 %and, %c
266  %and2 = and i32 %0, %b
267  ret i32 %and2
268}
269
270define i64 @blsi_through3(i64 %a, i64 %b, i64 %c) nounwind {
271; X86-LABEL: blsi_through3:
272; X86:       # %bb.0: # %entry
273; X86-NEXT:    pushl %esi
274; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
275; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
276; X86-NEXT:    xorl %edx, %edx
277; X86-NEXT:    movl %ecx, %eax
278; X86-NEXT:    negl %eax
279; X86-NEXT:    sbbl %esi, %edx
280; X86-NEXT:    andl {{[0-9]+}}(%esp), %edx
281; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
282; X86-NEXT:    andl {{[0-9]+}}(%esp), %edx
283; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
284; X86-NEXT:    andl %esi, %edx
285; X86-NEXT:    andl %ecx, %eax
286; X86-NEXT:    popl %esi
287; X86-NEXT:    retl
288;
289; X64-LABEL: blsi_through3:
290; X64:       # %bb.0: # %entry
291; X64-NEXT:    blsiq %rsi, %rax
292; X64-NEXT:    andq %rdx, %rdi
293; X64-NEXT:    andq %rdi, %rax
294; X64-NEXT:    retq
295entry:
296  %sub = sub i64 0, %b
297  %and = and i64 %sub, %a
298  %0 = and i64 %and, %c
299  %and3 = and i64 %0, %b
300  ret i64 %and3
301}
302
303define i32 @blsi_through1_used1(i32 %a, i32 %b) nounwind {
304; X86-LABEL: blsi_through1_used1:
305; X86:       # %bb.0: # %entry
306; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
307; X86-NEXT:    movl %ecx, %eax
308; X86-NEXT:    negl %eax
309; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
310; X86-NEXT:    andl %eax, %ecx
311; X86-NEXT:    incl %eax
312; X86-NEXT:    orl %ecx, %eax
313; X86-NEXT:    retl
314;
315; X64-LABEL: blsi_through1_used1:
316; X64:       # %bb.0: # %entry
317; X64-NEXT:    movl %esi, %eax
318; X64-NEXT:    negl %eax
319; X64-NEXT:    andl %edi, %eax
320; X64-NEXT:    andl %eax, %esi
321; X64-NEXT:    incl %eax
322; X64-NEXT:    orl %esi, %eax
323; X64-NEXT:    retq
324entry:
325  %sub = sub nsw i32 0, %b
326  %and = and i32 %sub, %a
327  %and1 = and i32 %and, %b
328  %add = add nsw i32 %and, 1
329  %or = or i32 %add, %and1
330  ret i32 %or
331}
332
333define i64 @blsi_through1_used2(i64 %a, i64 %b) nounwind {
334; X86-LABEL: blsi_through1_used2:
335; X86:       # %bb.0: # %entry
336; X86-NEXT:    pushl %ebx
337; X86-NEXT:    pushl %edi
338; X86-NEXT:    pushl %esi
339; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
340; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
341; X86-NEXT:    xorl %edi, %edi
342; X86-NEXT:    movl %ecx, %edx
343; X86-NEXT:    negl %edx
344; X86-NEXT:    sbbl %esi, %edi
345; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
346; X86-NEXT:    andl %edi, %ebx
347; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
348; X86-NEXT:    andl %edx, %eax
349; X86-NEXT:    andl %ebx, %esi
350; X86-NEXT:    andl %eax, %ecx
351; X86-NEXT:    imull %edx, %ebx
352; X86-NEXT:    imull %eax, %edi
353; X86-NEXT:    mull %edx
354; X86-NEXT:    addl %edi, %edx
355; X86-NEXT:    addl %ebx, %edx
356; X86-NEXT:    orl %esi, %edx
357; X86-NEXT:    orl %ecx, %eax
358; X86-NEXT:    popl %esi
359; X86-NEXT:    popl %edi
360; X86-NEXT:    popl %ebx
361; X86-NEXT:    retl
362;
363; X64-LABEL: blsi_through1_used2:
364; X64:       # %bb.0: # %entry
365; X64-NEXT:    movq %rsi, %rax
366; X64-NEXT:    negq %rax
367; X64-NEXT:    andq %rax, %rdi
368; X64-NEXT:    andq %rdi, %rsi
369; X64-NEXT:    imulq %rdi, %rax
370; X64-NEXT:    orq %rsi, %rax
371; X64-NEXT:    retq
372entry:
373  %sub = sub nsw i64 0, %b
374  %and = and i64 %sub, %a
375  %and1 = and i64 %and, %b
376  %mul = mul nsw i64 %and, %sub
377  %or = or i64 %mul, %and1
378  ret i64 %or
379}
380
381define i32 @blsr_used2(i32 %a) nounwind {
382; X86-LABEL: blsr_used2:
383; X86:       # %bb.0: # %entry
384; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
385; X86-NEXT:    leal -1(%eax), %ecx
386; X86-NEXT:    andl %ecx, %eax
387; X86-NEXT:    imull %ecx, %eax
388; X86-NEXT:    retl
389;
390; X64-LABEL: blsr_used2:
391; X64:       # %bb.0: # %entry
392; X64-NEXT:    # kill: def $edi killed $edi def $rdi
393; X64-NEXT:    leal -1(%rdi), %eax
394; X64-NEXT:    andl %eax, %edi
395; X64-NEXT:    imull %edi, %eax
396; X64-NEXT:    retq
397entry:
398  %sub = add i32 %a, -1
399  %and = and i32 %sub, %a
400  %mul = mul i32 %and, %sub
401  ret i32 %mul
402}
403
404define i64 @blsr_through1(i64 %a, i64 %b) nounwind {
405; X86-LABEL: blsr_through1:
406; X86:       # %bb.0: # %entry
407; X86-NEXT:    pushl %esi
408; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
409; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
410; X86-NEXT:    movl %ecx, %eax
411; X86-NEXT:    addl $-1, %eax
412; X86-NEXT:    movl %esi, %edx
413; X86-NEXT:    adcl $-1, %edx
414; X86-NEXT:    andl {{[0-9]+}}(%esp), %edx
415; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
416; X86-NEXT:    andl %ecx, %eax
417; X86-NEXT:    andl %esi, %edx
418; X86-NEXT:    popl %esi
419; X86-NEXT:    retl
420;
421; X64-LABEL: blsr_through1:
422; X64:       # %bb.0: # %entry
423; X64-NEXT:    blsrq %rsi, %rax
424; X64-NEXT:    andq %rdi, %rax
425; X64-NEXT:    retq
426entry:
427  %sub = add i64 %b, -1
428  %0 = and i64 %sub, %a
429  %and1 = and i64 %0, %b
430  ret i64 %and1
431}
432
433define i32 @blsr_through2(i32 %a, i32 %b, i32 %c) nounwind {
434; X86-LABEL: blsr_through2:
435; X86:       # %bb.0: # %entry
436; X86-NEXT:    blsrl {{[0-9]+}}(%esp), %eax
437; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
438; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
439; X86-NEXT:    retl
440;
441; X64-LABEL: blsr_through2:
442; X64:       # %bb.0: # %entry
443; X64-NEXT:    blsrl %esi, %eax
444; X64-NEXT:    andl %edx, %edi
445; X64-NEXT:    andl %edi, %eax
446; X64-NEXT:    retq
447entry:
448  %sub = add nsw i32 %b, -1
449  %0 = and i32 %sub, %a
450  %1 = and i32 %0, %c
451  %and2 = and i32 %1, %b
452  ret i32 %and2
453}
454
455define i64 @blsr_through3(i64 %a, i64 %b, i64 %c, i64 %d) nounwind {
456; X86-LABEL: blsr_through3:
457; X86:       # %bb.0: # %entry
458; X86-NEXT:    pushl %esi
459; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
460; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
461; X86-NEXT:    xorl %edx, %edx
462; X86-NEXT:    movl %ecx, %eax
463; X86-NEXT:    negl %eax
464; X86-NEXT:    sbbl %esi, %edx
465; X86-NEXT:    andl {{[0-9]+}}(%esp), %edx
466; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
467; X86-NEXT:    andl {{[0-9]+}}(%esp), %edx
468; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
469; X86-NEXT:    andl {{[0-9]+}}(%esp), %edx
470; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
471; X86-NEXT:    andl %esi, %edx
472; X86-NEXT:    andl %ecx, %eax
473; X86-NEXT:    popl %esi
474; X86-NEXT:    retl
475;
476; X64-LABEL: blsr_through3:
477; X64:       # %bb.0: # %entry
478; X64-NEXT:    movq %rdi, %rax
479; X64-NEXT:    andq %rsi, %rcx
480; X64-NEXT:    negq %rsi
481; X64-NEXT:    andq %rdx, %rax
482; X64-NEXT:    andq %rsi, %rax
483; X64-NEXT:    andq %rcx, %rax
484; X64-NEXT:    retq
485entry:
486  %sub = sub nsw i64 0, %b
487  %and = and i64 %sub, %a
488  %0 = and i64 %and, %c
489  %1 = and i64 %0, %d
490  %and4 = and i64 %1, %b
491  ret i64 %and4
492}
493
494define i32 @blsr_through1_used1(i32 %a, i32 %b) nounwind {
495; X86-LABEL: blsr_through1_used1:
496; X86:       # %bb.0: # %entry
497; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
498; X86-NEXT:    leal -1(%ecx), %eax
499; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
500; X86-NEXT:    andl %eax, %ecx
501; X86-NEXT:    incl %eax
502; X86-NEXT:    orl %ecx, %eax
503; X86-NEXT:    retl
504;
505; X64-LABEL: blsr_through1_used1:
506; X64:       # %bb.0: # %entry
507; X64-NEXT:    # kill: def $esi killed $esi def $rsi
508; X64-NEXT:    leal -1(%rsi), %eax
509; X64-NEXT:    andl %edi, %eax
510; X64-NEXT:    andl %eax, %esi
511; X64-NEXT:    incl %eax
512; X64-NEXT:    orl %esi, %eax
513; X64-NEXT:    retq
514entry:
515  %sub = add i32 %b, -1
516  %and = and i32 %sub, %a
517  %and1 = and i32 %and, %b
518  %add = add i32 %and, 1
519  %or = or i32 %add, %and1
520  ret i32 %or
521}
522
523define i64 @blsr_through1_used2(i64 %a, i64 %b) nounwind {
524; X86-LABEL: blsr_through1_used2:
525; X86:       # %bb.0: # %entry
526; X86-NEXT:    pushl %ebp
527; X86-NEXT:    pushl %ebx
528; X86-NEXT:    pushl %edi
529; X86-NEXT:    pushl %esi
530; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
531; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
532; X86-NEXT:    movl %ecx, %edi
533; X86-NEXT:    addl $-1, %edi
534; X86-NEXT:    movl %esi, %ebp
535; X86-NEXT:    adcl $-1, %ebp
536; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
537; X86-NEXT:    andl %ebp, %ebx
538; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
539; X86-NEXT:    andl %edi, %eax
540; X86-NEXT:    andl %ebx, %esi
541; X86-NEXT:    andl %eax, %ecx
542; X86-NEXT:    imull %eax, %ebp
543; X86-NEXT:    mull %edi
544; X86-NEXT:    addl %ebp, %edx
545; X86-NEXT:    imull %edi, %ebx
546; X86-NEXT:    addl %ebx, %edx
547; X86-NEXT:    orl %esi, %edx
548; X86-NEXT:    orl %ecx, %eax
549; X86-NEXT:    popl %esi
550; X86-NEXT:    popl %edi
551; X86-NEXT:    popl %ebx
552; X86-NEXT:    popl %ebp
553; X86-NEXT:    retl
554;
555; X64-LABEL: blsr_through1_used2:
556; X64:       # %bb.0: # %entry
557; X64-NEXT:    leaq -1(%rsi), %rax
558; X64-NEXT:    andq %rax, %rdi
559; X64-NEXT:    andq %rdi, %rsi
560; X64-NEXT:    imulq %rdi, %rax
561; X64-NEXT:    orq %rsi, %rax
562; X64-NEXT:    retq
563entry:
564  %sub = add i64 %b, -1
565  %and = and i64 %sub, %a
566  %and1 = and i64 %and, %b
567  %mul = mul i64 %and, %sub
568  %or = or i64 %mul, %and1
569  ret i64 %or
570}
571