| blob | 7d1a3b70e1741c323665729964cd407be13dbe5a |
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(i256* %a, i256* %b, i256* %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), %eax
25 ; X32-NEXT: movl {{[0-9]+}}(%esp), %ecx
26 ; X32-NEXT: movl 12(%ecx), %esi
27 ; X32-NEXT: movl 8(%ecx), %ebx
28 ; X32-NEXT: movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
29 ; X32-NEXT: movl (%eax), %edi
30 ; X32-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
31 ; X32-NEXT: movl %ebx, %eax
32 ; X32-NEXT: mull %edi
33 ; X32-NEXT: movl %edx, %ebp
34 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
35 ; X32-NEXT: movl %esi, %eax
36 ; X32-NEXT: movl %esi, %ecx
37 ; X32-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
38 ; X32-NEXT: mull %edi
39 ; X32-NEXT: movl %edx, %esi
40 ; X32-NEXT: movl %eax, %edi
41 ; X32-NEXT: addl %ebp, %edi
42 ; X32-NEXT: adcl $0, %esi
43 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
44 ; X32-NEXT: movl 4(%eax), %ebp
45 ; X32-NEXT: movl %ebx, %eax
46 ; X32-NEXT: mull %ebp
47 ; X32-NEXT: movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
48 ; X32-NEXT: movl %edx, %ebx
49 ; X32-NEXT: addl %edi, %eax
50 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
51 ; X32-NEXT: adcl %esi, %ebx
52 ; X32-NEXT: setb {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Spill
53 ; X32-NEXT: movl %ecx, %eax
54 ; X32-NEXT: mull %ebp
55 ; X32-NEXT: addl %ebx, %eax
56 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
57 ; X32-NEXT: movzbl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 1-byte Folded Reload
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), %edi
61 ; X32-NEXT: movl (%edi), %eax
62 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
63 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
64 ; X32-NEXT: mull %ecx
65 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
66 ; X32-NEXT: movl %edx, %esi
67 ; X32-NEXT: movl 4(%edi), %eax
68 ; X32-NEXT: movl %eax, %edi
69 ; X32-NEXT: mull %ecx
70 ; X32-NEXT: movl %edx, %ebx
71 ; X32-NEXT: movl %eax, %ebp
72 ; X32-NEXT: addl %esi, %ebp
73 ; X32-NEXT: adcl $0, %ebx
74 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
75 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
76 ; X32-NEXT: mull %esi
77 ; X32-NEXT: movl %edx, %ecx
78 ; X32-NEXT: addl %ebp, %eax
79 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
80 ; X32-NEXT: adcl %ebx, %ecx
81 ; X32-NEXT: setb %bl
82 ; X32-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
83 ; X32-NEXT: movl %edi, %eax
84 ; X32-NEXT: mull %esi
85 ; X32-NEXT: movl %eax, %ebp
86 ; X32-NEXT: addl %ecx, %ebp
87 ; X32-NEXT: movzbl %bl, %eax
88 ; X32-NEXT: adcl %eax, %edx
89 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
90 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
91 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
92 ; X32-NEXT: adcl $0, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
93 ; X32-NEXT: adcl $0, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
94 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
95 ; X32-NEXT: movl 8(%eax), %ecx
96 ; X32-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
97 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
98 ; X32-NEXT: movl %esi, %eax
99 ; X32-NEXT: mull %ecx
100 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
101 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
102 ; X32-NEXT: movl %edi, %eax
103 ; X32-NEXT: mull %ecx
104 ; X32-NEXT: movl %edx, %ebx
105 ; X32-NEXT: movl %eax, %edi
106 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
107 ; X32-NEXT: adcl $0, %ebx
108 ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax
109 ; X32-NEXT: movl 12(%eax), %ecx
110 ; X32-NEXT: movl %esi, %eax
111 ; X32-NEXT: mull %ecx
112 ; X32-NEXT: movl %ecx, (%esp) # 4-byte Spill
113 ; X32-NEXT: movl %edx, %esi
114 ; X32-NEXT: addl %edi, %eax
115 ; X32-NEXT: movl %eax, %edi
116 ; X32-NEXT: adcl %ebx, %esi
117 ; X32-NEXT: setb %bl
118 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
119 ; X32-NEXT: mull %ecx
120 ; X32-NEXT: movl %eax, %ecx
121 ; X32-NEXT: addl %esi, %ecx
122 ; X32-NEXT: movzbl %bl, %eax
123 ; X32-NEXT: adcl %eax, %edx
124 ; X32-NEXT: addl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
125 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
126 ; X32-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
127 ; X32-NEXT: adcl $0, %ecx
128 ; X32-NEXT: adcl $0, %edx
129 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
130 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
131 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
132 ; X32-NEXT: setb {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Spill
133 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
134 ; X32-NEXT: movl %esi, %eax
135 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
136 ; X32-NEXT: mull %edi
137 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
138 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
139 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
140 ; X32-NEXT: mull %edi
141 ; X32-NEXT: movl %edx, %ebx
142 ; X32-NEXT: movl %eax, %ebp
143 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
144 ; X32-NEXT: adcl $0, %ebx
145 ; X32-NEXT: movl %esi, %eax
146 ; X32-NEXT: movl (%esp), %edi # 4-byte Reload
147 ; X32-NEXT: mull %edi
148 ; X32-NEXT: movl %edx, %esi
149 ; X32-NEXT: addl %ebp, %eax
150 ; X32-NEXT: movl %eax, %ebp
151 ; X32-NEXT: adcl %ebx, %esi
152 ; X32-NEXT: setb {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Spill
153 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
154 ; X32-NEXT: mull %edi
155 ; X32-NEXT: movl %edi, %ebx
156 ; X32-NEXT: addl %esi, %eax
157 ; X32-NEXT: movzbl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 1-byte Folded Reload
158 ; X32-NEXT: adcl %esi, %edx
159 ; X32-NEXT: addl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
160 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
161 ; X32-NEXT: movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
162 ; X32-NEXT: movzbl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 1-byte Folded Reload
163 ; X32-NEXT: adcl %ecx, %eax
164 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
165 ; X32-NEXT: adcl $0, %edx
166 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
167 ; X32-NEXT: movl {{[0-9]+}}(%esp), %ecx
168 ; X32-NEXT: movl 16(%ecx), %edi
169 ; X32-NEXT: imull %edi, %ebx
170 ; X32-NEXT: movl %edi, %eax
171 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
172 ; X32-NEXT: mull %ebp
173 ; X32-NEXT: movl %eax, (%esp) # 4-byte Spill
174 ; X32-NEXT: addl %ebx, %edx
175 ; X32-NEXT: movl 20(%ecx), %eax
176 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
177 ; X32-NEXT: imull %eax, %ebp
178 ; X32-NEXT: addl %edx, %ebp
179 ; X32-NEXT: movl 24(%ecx), %eax
180 ; X32-NEXT: movl %eax, %ecx
181 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebx # 4-byte Reload
182 ; X32-NEXT: imull %ebx, %ecx
183 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
184 ; X32-NEXT: mull %esi
185 ; X32-NEXT: addl %ecx, %edx
186 ; X32-NEXT: movl {{[0-9]+}}(%esp), %ecx
187 ; X32-NEXT: movl 28(%ecx), %ecx
188 ; X32-NEXT: imull %esi, %ecx
189 ; X32-NEXT: addl %edx, %ecx
190 ; X32-NEXT: addl (%esp), %eax # 4-byte Folded Reload
191 ; X32-NEXT: movl %eax, (%esp) # 4-byte Spill
192 ; X32-NEXT: adcl %ebp, %ecx
193 ; X32-NEXT: movl %esi, %eax
194 ; X32-NEXT: mull %edi
195 ; X32-NEXT: movl %edx, %ebp
196 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
197 ; X32-NEXT: movl %ebx, %eax
198 ; X32-NEXT: mull %edi
199 ; X32-NEXT: movl %edx, %esi
200 ; X32-NEXT: movl %eax, %ebx
201 ; X32-NEXT: addl %ebp, %ebx
202 ; X32-NEXT: adcl $0, %esi
203 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
204 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
205 ; X32-NEXT: mull %ebp
206 ; X32-NEXT: movl %edx, %edi
207 ; X32-NEXT: addl %ebx, %eax
208 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
209 ; X32-NEXT: adcl %esi, %edi
210 ; X32-NEXT: setb %bl
211 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
212 ; X32-NEXT: mull %ebp
213 ; X32-NEXT: addl %edi, %eax
214 ; X32-NEXT: movzbl %bl, %esi
215 ; X32-NEXT: adcl %esi, %edx
216 ; X32-NEXT: addl (%esp), %eax # 4-byte Folded Reload
217 ; X32-NEXT: movl %eax, (%esp) # 4-byte Spill
218 ; X32-NEXT: adcl %ecx, %edx
219 ; X32-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
220 ; X32-NEXT: movl {{[0-9]+}}(%esp), %ebx
221 ; X32-NEXT: movl 28(%ebx), %ecx
222 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload
223 ; X32-NEXT: imull %esi, %ecx
224 ; X32-NEXT: movl 24(%ebx), %edi
225 ; X32-NEXT: movl %esi, %eax
226 ; X32-NEXT: mull %edi
227 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
228 ; X32-NEXT: addl %ecx, %edx
229 ; X32-NEXT: imull {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
230 ; X32-NEXT: addl %edx, %edi
231 ; X32-NEXT: movl 16(%ebx), %ebp
232 ; X32-NEXT: movl 20(%ebx), %ebx
233 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
234 ; X32-NEXT: movl %eax, %ecx
235 ; X32-NEXT: imull %ebx, %ecx
236 ; X32-NEXT: movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
237 ; X32-NEXT: mull %ebp
238 ; X32-NEXT: addl %ecx, %edx
239 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
240 ; X32-NEXT: imull %ebp, %ecx
241 ; X32-NEXT: addl %edx, %ecx
242 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
243 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
244 ; X32-NEXT: adcl %edi, %ecx
245 ; X32-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
246 ; X32-NEXT: movl %ebp, %eax
247 ; X32-NEXT: mull %esi
248 ; X32-NEXT: movl %edx, %edi
249 ; X32-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
250 ; X32-NEXT: movl %ebx, %eax
251 ; X32-NEXT: mull %esi
252 ; X32-NEXT: movl %edx, %ecx
253 ; X32-NEXT: movl %eax, %ebx
254 ; X32-NEXT: addl %edi, %ebx
255 ; X32-NEXT: adcl $0, %ecx
256 ; X32-NEXT: movl %ebp, %eax
257 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
258 ; X32-NEXT: mull %ebp
259 ; X32-NEXT: movl %edx, %edi
260 ; X32-NEXT: movl %eax, %esi
261 ; X32-NEXT: addl %ebx, %esi
262 ; X32-NEXT: adcl %ecx, %edi
263 ; X32-NEXT: setb %cl
264 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
265 ; X32-NEXT: mull %ebp
266 ; X32-NEXT: addl %edi, %eax
267 ; X32-NEXT: movzbl %cl, %ecx
268 ; X32-NEXT: adcl %ecx, %edx
269 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
270 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
271 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
272 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
273 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
274 ; X32-NEXT: adcl (%esp), %eax # 4-byte Folded Reload
275 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
276 ; X32-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
277 ; X32-NEXT: movl %ecx, %ebx
278 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Folded Reload
279 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
280 ; X32-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Folded Reload
281 ; X32-NEXT: movl {{[0-9]+}}(%esp), %ecx
282 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
283 ; X32-NEXT: movl %edi, (%ecx)
284 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
285 ; X32-NEXT: movl %edi, 4(%ecx)
286 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
287 ; X32-NEXT: movl %edi, 8(%ecx)
288 ; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload
289 ; X32-NEXT: movl %edi, 12(%ecx)
290 ; X32-NEXT: movl %ebx, 16(%ecx)
291 ; X32-NEXT: movl %esi, 20(%ecx)
292 ; X32-NEXT: movl %eax, 24(%ecx)
293 ; X32-NEXT: movl %edx, 28(%ecx)
294 ; X32-NEXT: addl $72, %esp
295 ; X32-NEXT: .cfi_def_cfa_offset 20
296 ; X32-NEXT: popl %esi
297 ; X32-NEXT: .cfi_def_cfa_offset 16
298 ; X32-NEXT: popl %edi
299 ; X32-NEXT: .cfi_def_cfa_offset 12
300 ; X32-NEXT: popl %ebx
301 ; X32-NEXT: .cfi_def_cfa_offset 8
302 ; X32-NEXT: popl %ebp
303 ; X32-NEXT: .cfi_def_cfa_offset 4
304 ; X32-NEXT: retl
305 ;
306 ; X64-LABEL: test:
307 ; X64: # %bb.0: # %entry
308 ; X64-NEXT: pushq %r15
309 ; X64-NEXT: .cfi_def_cfa_offset 16
310 ; X64-NEXT: pushq %r14
311 ; X64-NEXT: .cfi_def_cfa_offset 24
312 ; X64-NEXT: pushq %rbx
313 ; X64-NEXT: .cfi_def_cfa_offset 32
314 ; X64-NEXT: .cfi_offset %rbx, -32
315 ; X64-NEXT: .cfi_offset %r14, -24
316 ; X64-NEXT: .cfi_offset %r15, -16
317 ; X64-NEXT: movq %rdx, %r9
318 ; X64-NEXT: movq (%rdi), %r11
319 ; X64-NEXT: movq 8(%rdi), %r8
320 ; X64-NEXT: movq 16(%rdi), %rbx
321 ; X64-NEXT: movq 16(%rsi), %r10
322 ; X64-NEXT: movq (%rsi), %rcx
323 ; X64-NEXT: movq 8(%rsi), %r15
324 ; X64-NEXT: movq 24(%rdi), %rdi
325 ; X64-NEXT: imulq %rcx, %rdi
326 ; X64-NEXT: movq %rcx, %rax
327 ; X64-NEXT: mulq %rbx
328 ; X64-NEXT: movq %rax, %r14
329 ; X64-NEXT: addq %rdi, %rdx
330 ; X64-NEXT: imulq %r15, %rbx
331 ; X64-NEXT: addq %rdx, %rbx
332 ; X64-NEXT: movq %r10, %rdi
333 ; X64-NEXT: imulq %r8, %rdi
334 ; X64-NEXT: movq %r10, %rax
335 ; X64-NEXT: mulq %r11
336 ; X64-NEXT: movq %rax, %r10
337 ; X64-NEXT: addq %rdi, %rdx
338 ; X64-NEXT: movq 24(%rsi), %rdi
339 ; X64-NEXT: imulq %r11, %rdi
340 ; X64-NEXT: addq %rdx, %rdi
341 ; X64-NEXT: addq %r14, %r10
342 ; X64-NEXT: adcq %rbx, %rdi
343 ; X64-NEXT: movq %r11, %rax
344 ; X64-NEXT: mulq %rcx
345 ; X64-NEXT: movq %rdx, %rsi
346 ; X64-NEXT: movq %rax, %r14
347 ; X64-NEXT: movq %r8, %rax
348 ; X64-NEXT: mulq %rcx
349 ; X64-NEXT: movq %rdx, %rcx
350 ; X64-NEXT: movq %rax, %rbx
351 ; X64-NEXT: addq %rsi, %rbx
352 ; X64-NEXT: adcq $0, %rcx
353 ; X64-NEXT: movq %r11, %rax
354 ; X64-NEXT: mulq %r15
355 ; X64-NEXT: movq %rdx, %rsi
356 ; X64-NEXT: movq %rax, %r11
357 ; X64-NEXT: addq %rbx, %r11
358 ; X64-NEXT: adcq %rcx, %rsi
359 ; X64-NEXT: setb %al
360 ; X64-NEXT: movzbl %al, %ecx
361 ; X64-NEXT: movq %r8, %rax
362 ; X64-NEXT: mulq %r15
363 ; X64-NEXT: addq %rsi, %rax
364 ; X64-NEXT: adcq %rcx, %rdx
365 ; X64-NEXT: addq %r10, %rax
366 ; X64-NEXT: adcq %rdi, %rdx
367 ; X64-NEXT: movq %r14, (%r9)
368 ; X64-NEXT: movq %r11, 8(%r9)
369 ; X64-NEXT: movq %rax, 16(%r9)
370 ; X64-NEXT: movq %rdx, 24(%r9)
371 ; X64-NEXT: popq %rbx
372 ; X64-NEXT: .cfi_def_cfa_offset 24
373 ; X64-NEXT: popq %r14
374 ; X64-NEXT: .cfi_def_cfa_offset 16
375 ; X64-NEXT: popq %r15
376 ; X64-NEXT: .cfi_def_cfa_offset 8
377 ; X64-NEXT: retq
378 entry:
379 %av = load i256, i256* %a
380 %bv = load i256, i256* %b
381 %r = mul i256 %av, %bv
382 store i256 %r, i256* %out
383 ret void
384 }
386 attributes #0 = { norecurse nounwind uwtable }