| blob | 6f6dde3aa3cf4de9537aaf11ab67d974df3b3663 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=i386-unknown | FileCheck %s --check-prefix=X32
3 ; RUN: llc < %s -mtriple=x86_64-unknown | FileCheck %s --check-prefix=X64
5 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
7 define void @test(ptr %a, ptr %b, ptr %out) #0 {
8 ; X32-LABEL: test:
9 ; X32: # %bb.0: # %entry
10 ; X32-NEXT: pushl %ebp
11 ; X32-NEXT: .cfi_def_cfa_offset 8
12 ; X32-NEXT: pushl %ebx
13 ; X32-NEXT: .cfi_def_cfa_offset 12
14 ; X32-NEXT: pushl %edi
15 ; X32-NEXT: .cfi_def_cfa_offset 16
16 ; X32-NEXT: pushl %esi
17 ; X32-NEXT: .cfi_def_cfa_offset 20
18 ; X32-NEXT: subl $72, %esp
19 ; X32-NEXT: .cfi_def_cfa_offset 92
20 ; X32-NEXT: .cfi_offset %esi, -20
21 ; X32-NEXT: .cfi_offset %edi, -16
22 ; X32-NEXT: .cfi_offset %ebx, -12
23 ; X32-NEXT: .cfi_offset %ebp, -8
24 ; X32-NEXT: movl {{[0-9]+}}(%esp), %edx
25 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
26 ; X32-NEXT: movl 12(%eax), %ebx
27 ; X32-NEXT: movl 8(%eax), %ebp
28 ; X32-NEXT: movl (%edx), %edi
29 ; X32-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
30 ; X32-NEXT: movl %ebp, %eax
31 ; X32-NEXT: movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
32 ; X32-NEXT: mull %edi
33 ; X32-NEXT: movl %edx, %esi
34 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
35 ; X32-NEXT: movl %ebx, %eax
36 ; X32-NEXT: movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
37 ; X32-NEXT: mull %edi
38 ; X32-NEXT: movl %edx, %ecx
39 ; X32-NEXT: movl %eax, %edi
40 ; X32-NEXT: addl %esi, %edi
41 ; X32-NEXT: adcl $0, %ecx
42 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
43 ; X32-NEXT: movl 4(%eax), %esi
44 ; X32-NEXT: movl %ebp, %eax
45 ; X32-NEXT: mull %esi
46 ; X32-NEXT: movl %esi, %ebp
47 ; X32-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
48 ; X32-NEXT: movl %edx, %esi
49 ; X32-NEXT: addl %edi, %eax
50 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
51 ; X32-NEXT: adcl %ecx, %esi
52 ; X32-NEXT: setb %cl
53 ; X32-NEXT: movl %ebx, %eax
54 ; X32-NEXT: mull %ebp
55 ; X32-NEXT: addl %esi, %eax
56 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
57 ; X32-NEXT: movzbl %cl, %eax
58 ; X32-NEXT: adcl %eax, %edx
59 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
60 ; X32-NEXT: movl {{[0-9]+}}(%esp), %esi
61 ; X32-NEXT: movl (%esi), %ebx
62 ; X32-NEXT: movl %ebx, %eax
63 ; X32-NEXT: movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
64 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
65 ; X32-NEXT: mull %ecx
66 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
67 ; X32-NEXT: movl %edx, %edi
68 ; X32-NEXT: movl 4(%esi), %ebp
69 ; X32-NEXT: movl %ebp, %eax
70 ; X32-NEXT: mull %ecx
71 ; X32-NEXT: movl %edx, %ecx
72 ; X32-NEXT: movl %eax, %esi
73 ; X32-NEXT: addl %edi, %esi
74 ; X32-NEXT: adcl $0, %ecx
75 ; X32-NEXT: movl %ebx, %eax
76 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
77 ; X32-NEXT: mull %edi
78 ; X32-NEXT: movl %edx, %ebx
79 ; X32-NEXT: addl %esi, %eax
80 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
81 ; X32-NEXT: adcl %ecx, %ebx
82 ; X32-NEXT: setb %cl
83 ; X32-NEXT: movl %ebp, %eax
84 ; X32-NEXT: movl %ebp, %esi
85 ; X32-NEXT: mull %edi
86 ; X32-NEXT: addl %ebx, %eax
87 ; X32-NEXT: movzbl %cl, %ecx
88 ; X32-NEXT: adcl %ecx, %edx
89 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
90 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
91 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
92 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
93 ; X32-NEXT: adcl $0, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
94 ; X32-NEXT: adcl $0, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
95 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
96 ; X32-NEXT: movl 8(%eax), %ecx
97 ; X32-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
98 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
99 ; X32-NEXT: movl %edi, %eax
100 ; X32-NEXT: mull %ecx
101 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
102 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
103 ; X32-NEXT: movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
104 ; X32-NEXT: movl %ebp, %eax
105 ; X32-NEXT: mull %ecx
106 ; X32-NEXT: movl %edx, %ecx
107 ; X32-NEXT: movl %eax, %ebx
108 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
109 ; X32-NEXT: adcl $0, %ecx
110 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
111 ; X32-NEXT: movl 12(%eax), %ebp
112 ; X32-NEXT: movl %edi, %eax
113 ; X32-NEXT: mull %ebp
114 ; X32-NEXT: movl %ebp, (%esp) # 4-byte Spill
115 ; X32-NEXT: movl %edx, %edi
116 ; X32-NEXT: addl %ebx, %eax
117 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
118 ; X32-NEXT: adcl %ecx, %edi
119 ; X32-NEXT: setb %cl
120 ; X32-NEXT: movl %esi, %eax
121 ; X32-NEXT: mull %ebp
122 ; X32-NEXT: movl %eax, %ebx
123 ; X32-NEXT: addl %edi, %ebx
124 ; X32-NEXT: movzbl %cl, %eax
125 ; X32-NEXT: adcl %eax, %edx
126 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
127 ; X32-NEXT: addl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
128 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
129 ; X32-NEXT: adcl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
130 ; X32-NEXT: adcl $0, %ebx
131 ; X32-NEXT: adcl $0, %edx
132 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Folded Reload
133 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
134 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
135 ; X32-NEXT: setb {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Spill
136 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
137 ; X32-NEXT: movl %edi, %eax
138 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
139 ; X32-NEXT: mull %ecx
140 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
141 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
142 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
143 ; X32-NEXT: mull %ecx
144 ; X32-NEXT: movl %edx, %esi
145 ; X32-NEXT: movl %eax, %ebp
146 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
147 ; X32-NEXT: adcl $0, %esi
148 ; X32-NEXT: movl %edi, %eax
149 ; X32-NEXT: movl (%esp), %ecx # 4-byte Reload
150 ; X32-NEXT: mull %ecx
151 ; X32-NEXT: movl %edx, %edi
152 ; X32-NEXT: addl %ebp, %eax
153 ; X32-NEXT: movl %eax, %ebp
154 ; X32-NEXT: adcl %esi, %edi
155 ; X32-NEXT: setb {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Spill
156 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
157 ; X32-NEXT: mull %ecx
158 ; X32-NEXT: addl %edi, %eax
159 ; X32-NEXT: movl %eax, %esi
160 ; X32-NEXT: movzbl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 1-byte Folded Reload
161 ; X32-NEXT: adcl %eax, %edx
162 ; X32-NEXT: addl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
163 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
164 ; X32-NEXT: movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
165 ; X32-NEXT: movzbl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 1-byte Folded Reload
166 ; X32-NEXT: adcl %eax, %esi
167 ; X32-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
168 ; X32-NEXT: adcl $0, %edx
169 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
170 ; X32-NEXT: movl {{[0-9]+}}(%esp), %ecx
171 ; X32-NEXT: movl 16(%ecx), %esi
172 ; X32-NEXT: movl (%esp), %edi # 4-byte Reload
173 ; X32-NEXT: imull %esi, %edi
174 ; X32-NEXT: movl %esi, %eax
175 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
176 ; X32-NEXT: mull %ebx
177 ; X32-NEXT: movl %eax, (%esp) # 4-byte Spill
178 ; X32-NEXT: addl %edi, %edx
179 ; X32-NEXT: movl 20(%ecx), %eax
180 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
181 ; X32-NEXT: movl %ebx, %edi
182 ; X32-NEXT: imull %eax, %edi
183 ; X32-NEXT: addl %edx, %edi
184 ; X32-NEXT: movl 24(%ecx), %eax
185 ; X32-NEXT: movl %eax, %ecx
186 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
187 ; X32-NEXT: imull %ebp, %ecx
188 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
189 ; X32-NEXT: mull %ebx
190 ; X32-NEXT: addl %ecx, %edx
191 ; X32-NEXT: movl {{[0-9]+}}(%esp), %ecx
192 ; X32-NEXT: movl 28(%ecx), %ecx
193 ; X32-NEXT: imull %ebx, %ecx
194 ; X32-NEXT: addl %edx, %ecx
195 ; X32-NEXT: addl (%esp), %eax # 4-byte Folded Reload
196 ; X32-NEXT: movl %eax, (%esp) # 4-byte Spill
197 ; X32-NEXT: adcl %edi, %ecx
198 ; X32-NEXT: movl %ebx, %eax
199 ; X32-NEXT: mull %esi
200 ; X32-NEXT: movl %edx, %edi
201 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
202 ; X32-NEXT: movl %ebp, %eax
203 ; X32-NEXT: mull %esi
204 ; X32-NEXT: movl %edx, %esi
205 ; X32-NEXT: movl %eax, %ebx
206 ; X32-NEXT: addl %edi, %ebx
207 ; X32-NEXT: adcl $0, %esi
208 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
209 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
210 ; X32-NEXT: mull %ebp
211 ; X32-NEXT: movl %edx, %edi
212 ; X32-NEXT: addl %ebx, %eax
213 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
214 ; X32-NEXT: adcl %esi, %edi
215 ; X32-NEXT: setb %bl
216 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
217 ; X32-NEXT: mull %ebp
218 ; X32-NEXT: addl %edi, %eax
219 ; X32-NEXT: movzbl %bl, %esi
220 ; X32-NEXT: adcl %esi, %edx
221 ; X32-NEXT: addl (%esp), %eax # 4-byte Folded Reload
222 ; X32-NEXT: movl %eax, (%esp) # 4-byte Spill
223 ; X32-NEXT: adcl %ecx, %edx
224 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
225 ; X32-NEXT: movl {{[0-9]+}}(%esp), %edi
226 ; X32-NEXT: movl 24(%edi), %esi
227 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
228 ; X32-NEXT: movl %ecx, %eax
229 ; X32-NEXT: mull %esi
230 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
231 ; X32-NEXT: imull {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
232 ; X32-NEXT: addl %edx, %esi
233 ; X32-NEXT: movl %edi, %edx
234 ; X32-NEXT: movl 28(%edi), %eax
235 ; X32-NEXT: imull %ecx, %eax
236 ; X32-NEXT: addl %eax, %esi
237 ; X32-NEXT: movl 16(%edi), %edi
238 ; X32-NEXT: movl 20(%edx), %ebp
239 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
240 ; X32-NEXT: movl %eax, %ebx
241 ; X32-NEXT: imull %ebp, %ebx
242 ; X32-NEXT: movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
243 ; X32-NEXT: mull %edi
244 ; X32-NEXT: addl %ebx, %edx
245 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
246 ; X32-NEXT: imull %edi, %ecx
247 ; X32-NEXT: addl %edx, %ecx
248 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
249 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
250 ; X32-NEXT: adcl %esi, %ecx
251 ; X32-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
252 ; X32-NEXT: movl %edi, %eax
253 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
254 ; X32-NEXT: mull %ecx
255 ; X32-NEXT: movl %edx, %ebx
256 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
257 ; X32-NEXT: movl %ebp, %eax
258 ; X32-NEXT: mull %ecx
259 ; X32-NEXT: movl %edx, %esi
260 ; X32-NEXT: movl %eax, %ebp
261 ; X32-NEXT: addl %ebx, %ebp
262 ; X32-NEXT: adcl $0, %esi
263 ; X32-NEXT: movl %edi, %eax
264 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
265 ; X32-NEXT: mull %ecx
266 ; X32-NEXT: movl %edx, %ebx
267 ; X32-NEXT: movl %eax, %edi
268 ; X32-NEXT: addl %ebp, %edi
269 ; X32-NEXT: adcl %esi, %ebx
270 ; X32-NEXT: setb {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Spill
271 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
272 ; X32-NEXT: mull %ecx
273 ; X32-NEXT: addl %ebx, %eax
274 ; X32-NEXT: movzbl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 1-byte Folded Reload
275 ; X32-NEXT: adcl %esi, %edx
276 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
277 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
278 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
279 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
280 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
281 ; X32-NEXT: adcl (%esp), %eax # 4-byte Folded Reload
282 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
283 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
284 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
285 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
286 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
287 ; X32-NEXT: movl {{[0-9]+}}(%esp), %esi
288 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
289 ; X32-NEXT: movl %ebx, (%esi)
290 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
291 ; X32-NEXT: movl %ebx, 4(%esi)
292 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
293 ; X32-NEXT: movl %ebx, 8(%esi)
294 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
295 ; X32-NEXT: movl %ebx, 12(%esi)
296 ; X32-NEXT: movl %ecx, 16(%esi)
297 ; X32-NEXT: movl %edi, 20(%esi)
298 ; X32-NEXT: movl %eax, 24(%esi)
299 ; X32-NEXT: movl %edx, 28(%esi)
300 ; X32-NEXT: addl $72, %esp
301 ; X32-NEXT: .cfi_def_cfa_offset 20
302 ; X32-NEXT: popl %esi
303 ; X32-NEXT: .cfi_def_cfa_offset 16
304 ; X32-NEXT: popl %edi
305 ; X32-NEXT: .cfi_def_cfa_offset 12
306 ; X32-NEXT: popl %ebx
307 ; X32-NEXT: .cfi_def_cfa_offset 8
308 ; X32-NEXT: popl %ebp
309 ; X32-NEXT: .cfi_def_cfa_offset 4
310 ; X32-NEXT: retl
311 ;
312 ; X64-LABEL: test:
313 ; X64: # %bb.0: # %entry
314 ; X64-NEXT: pushq %r15
315 ; X64-NEXT: .cfi_def_cfa_offset 16
316 ; X64-NEXT: pushq %r14
317 ; X64-NEXT: .cfi_def_cfa_offset 24
318 ; X64-NEXT: pushq %rbx
319 ; X64-NEXT: .cfi_def_cfa_offset 32
320 ; X64-NEXT: .cfi_offset %rbx, -32
321 ; X64-NEXT: .cfi_offset %r14, -24
322 ; X64-NEXT: .cfi_offset %r15, -16
323 ; X64-NEXT: movq %rdx, %rcx
324 ; X64-NEXT: movq (%rdi), %rbx
325 ; X64-NEXT: movq 8(%rdi), %r11
326 ; X64-NEXT: movq 16(%rdi), %r10
327 ; X64-NEXT: movq 16(%rsi), %r8
328 ; X64-NEXT: movq (%rsi), %r9
329 ; X64-NEXT: movq 8(%rsi), %r14
330 ; X64-NEXT: movq 24(%rdi), %r15
331 ; X64-NEXT: imulq %r9, %r15
332 ; X64-NEXT: movq %r9, %rax
333 ; X64-NEXT: mulq %r10
334 ; X64-NEXT: movq %rax, %rdi
335 ; X64-NEXT: imulq %r14, %r10
336 ; X64-NEXT: addq %rdx, %r10
337 ; X64-NEXT: addq %r15, %r10
338 ; X64-NEXT: movq %r8, %r15
339 ; X64-NEXT: imulq %r11, %r15
340 ; X64-NEXT: movq %r8, %rax
341 ; X64-NEXT: mulq %rbx
342 ; X64-NEXT: movq %rax, %r8
343 ; X64-NEXT: addq %r15, %rdx
344 ; X64-NEXT: movq 24(%rsi), %r15
345 ; X64-NEXT: imulq %rbx, %r15
346 ; X64-NEXT: addq %rdx, %r15
347 ; X64-NEXT: addq %rdi, %r8
348 ; X64-NEXT: adcq %r10, %r15
349 ; X64-NEXT: movq %rbx, %rax
350 ; X64-NEXT: mulq %r9
351 ; X64-NEXT: movq %rdx, %rsi
352 ; X64-NEXT: movq %rax, %r10
353 ; X64-NEXT: movq %r11, %rax
354 ; X64-NEXT: mulq %r9
355 ; X64-NEXT: movq %rdx, %rdi
356 ; X64-NEXT: movq %rax, %r9
357 ; X64-NEXT: addq %rsi, %r9
358 ; X64-NEXT: adcq $0, %rdi
359 ; X64-NEXT: movq %rbx, %rax
360 ; X64-NEXT: mulq %r14
361 ; X64-NEXT: movq %rdx, %rsi
362 ; X64-NEXT: movq %rax, %rbx
363 ; X64-NEXT: addq %r9, %rbx
364 ; X64-NEXT: adcq %rdi, %rsi
365 ; X64-NEXT: setb %al
366 ; X64-NEXT: movzbl %al, %edi
367 ; X64-NEXT: movq %r11, %rax
368 ; X64-NEXT: mulq %r14
369 ; X64-NEXT: addq %rsi, %rax
370 ; X64-NEXT: adcq %rdi, %rdx
371 ; X64-NEXT: addq %r8, %rax
372 ; X64-NEXT: adcq %r15, %rdx
373 ; X64-NEXT: movq %r10, (%rcx)
374 ; X64-NEXT: movq %rbx, 8(%rcx)
375 ; X64-NEXT: movq %rax, 16(%rcx)
376 ; X64-NEXT: movq %rdx, 24(%rcx)
377 ; X64-NEXT: popq %rbx
378 ; X64-NEXT: .cfi_def_cfa_offset 24
379 ; X64-NEXT: popq %r14
380 ; X64-NEXT: .cfi_def_cfa_offset 16
381 ; X64-NEXT: popq %r15
382 ; X64-NEXT: .cfi_def_cfa_offset 8
383 ; X64-NEXT: retq
384 entry:
385 %av = load i256, ptr %a
386 %bv = load i256, ptr %b
387 %r = mul i256 %av, %bv
388 store i256 %r, ptr %out
389 ret void
390 }
392 attributes #0 = { norecurse nounwind uwtable }