| blob | 9f58ed084334857490c6bf2763803d1cd0d07638 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s --check-prefix=X86-NOBMI
3 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+bmi2 | FileCheck %s --check-prefix=X86-BMI
4 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefix=X64-NOBMI
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+bmi2 | FileCheck %s --check-prefix=X64-BMI
7 ; PR1198
9 define i64 @foo(i64 %x, i64 %y) nounwind {
10 ; X86-NOBMI-LABEL: foo:
11 ; X86-NOBMI: # %bb.0:
12 ; X86-NOBMI-NEXT: pushl %ebp
13 ; X86-NOBMI-NEXT: pushl %ebx
14 ; X86-NOBMI-NEXT: pushl %edi
15 ; X86-NOBMI-NEXT: pushl %esi
16 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edi
17 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %ebx
18 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
19 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %ecx
20 ; X86-NOBMI-NEXT: movl %edi, %eax
21 ; X86-NOBMI-NEXT: mull %esi
22 ; X86-NOBMI-NEXT: movl %edx, %ebp
23 ; X86-NOBMI-NEXT: movl %ebx, %eax
24 ; X86-NOBMI-NEXT: mull %esi
25 ; X86-NOBMI-NEXT: movl %edx, %esi
26 ; X86-NOBMI-NEXT: movl %eax, %ebx
27 ; X86-NOBMI-NEXT: addl %ebp, %ebx
28 ; X86-NOBMI-NEXT: adcl $0, %esi
29 ; X86-NOBMI-NEXT: movl %edi, %eax
30 ; X86-NOBMI-NEXT: mull %ecx
31 ; X86-NOBMI-NEXT: addl %ebx, %eax
32 ; X86-NOBMI-NEXT: adcl %edx, %esi
33 ; X86-NOBMI-NEXT: setb %al
34 ; X86-NOBMI-NEXT: movzbl %al, %edi
35 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
36 ; X86-NOBMI-NEXT: mull %ecx
37 ; X86-NOBMI-NEXT: addl %esi, %eax
38 ; X86-NOBMI-NEXT: adcl %edi, %edx
39 ; X86-NOBMI-NEXT: popl %esi
40 ; X86-NOBMI-NEXT: popl %edi
41 ; X86-NOBMI-NEXT: popl %ebx
42 ; X86-NOBMI-NEXT: popl %ebp
43 ; X86-NOBMI-NEXT: retl
44 ;
45 ; X86-BMI-LABEL: foo:
46 ; X86-BMI: # %bb.0:
47 ; X86-BMI-NEXT: pushl %ebp
48 ; X86-BMI-NEXT: pushl %ebx
49 ; X86-BMI-NEXT: pushl %edi
50 ; X86-BMI-NEXT: pushl %esi
51 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %ecx
52 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %eax
53 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %edi
54 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %esi
55 ; X86-BMI-NEXT: movl %ecx, %edx
56 ; X86-BMI-NEXT: mulxl %edi, %ebx, %ebx
57 ; X86-BMI-NEXT: movl %eax, %edx
58 ; X86-BMI-NEXT: mulxl %edi, %edi, %ebp
59 ; X86-BMI-NEXT: addl %ebx, %edi
60 ; X86-BMI-NEXT: adcl $0, %ebp
61 ; X86-BMI-NEXT: movl %ecx, %edx
62 ; X86-BMI-NEXT: mulxl %esi, %ecx, %ebx
63 ; X86-BMI-NEXT: addl %edi, %ecx
64 ; X86-BMI-NEXT: adcl %ebp, %ebx
65 ; X86-BMI-NEXT: setb %cl
66 ; X86-BMI-NEXT: movzbl %cl, %ecx
67 ; X86-BMI-NEXT: movl %eax, %edx
68 ; X86-BMI-NEXT: mulxl %esi, %eax, %edx
69 ; X86-BMI-NEXT: addl %ebx, %eax
70 ; X86-BMI-NEXT: adcl %ecx, %edx
71 ; X86-BMI-NEXT: popl %esi
72 ; X86-BMI-NEXT: popl %edi
73 ; X86-BMI-NEXT: popl %ebx
74 ; X86-BMI-NEXT: popl %ebp
75 ; X86-BMI-NEXT: retl
76 ;
77 ; X64-NOBMI-LABEL: foo:
78 ; X64-NOBMI: # %bb.0:
79 ; X64-NOBMI-NEXT: movq %rdi, %rax
80 ; X64-NOBMI-NEXT: mulq %rsi
81 ; X64-NOBMI-NEXT: movq %rdx, %rax
82 ; X64-NOBMI-NEXT: retq
83 ;
84 ; X64-BMI-LABEL: foo:
85 ; X64-BMI: # %bb.0:
86 ; X64-BMI-NEXT: movq %rdi, %rdx
87 ; X64-BMI-NEXT: mulxq %rsi, %rax, %rax
88 ; X64-BMI-NEXT: retq
89 %tmp0 = zext i64 %x to i128
90 %tmp1 = zext i64 %y to i128
91 %tmp2 = mul i128 %tmp0, %tmp1
92 %tmp7 = zext i32 64 to i128
93 %tmp3 = lshr i128 %tmp2, %tmp7
94 %tmp4 = trunc i128 %tmp3 to i64
95 ret i64 %tmp4
96 }
98 ; <rdar://problem/14096009> superfluous multiply by high part of
99 ; zero-extended value.
101 define i64 @mul1(i64 %n, ptr nocapture %z, ptr nocapture %x, i64 %y) nounwind {
102 ; X86-NOBMI-LABEL: mul1:
103 ; X86-NOBMI: # %bb.0: # %entry
104 ; X86-NOBMI-NEXT: pushl %ebp
105 ; X86-NOBMI-NEXT: pushl %ebx
106 ; X86-NOBMI-NEXT: pushl %edi
107 ; X86-NOBMI-NEXT: pushl %esi
108 ; X86-NOBMI-NEXT: subl $20, %esp
109 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %ecx
110 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
111 ; X86-NOBMI-NEXT: orl %ecx, %eax
112 ; X86-NOBMI-NEXT: je .LBB1_3
113 ; X86-NOBMI-NEXT: # %bb.1: # %for.body.preheader
114 ; X86-NOBMI-NEXT: xorl %eax, %eax
115 ; X86-NOBMI-NEXT: xorl %edx, %edx
116 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
117 ; X86-NOBMI-NEXT: movl $0, (%esp) # 4-byte Folded Spill
118 ; X86-NOBMI-NEXT: .p2align 4, 0x90
119 ; X86-NOBMI-NEXT: .LBB1_2: # %for.body
120 ; X86-NOBMI-NEXT: # =>This Inner Loop Header: Depth=1
121 ; X86-NOBMI-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
122 ; X86-NOBMI-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
123 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
124 ; X86-NOBMI-NEXT: movl (%eax,%ecx,8), %edi
125 ; X86-NOBMI-NEXT: movl 4(%eax,%ecx,8), %ebx
126 ; X86-NOBMI-NEXT: movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
127 ; X86-NOBMI-NEXT: movl %edi, %eax
128 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
129 ; X86-NOBMI-NEXT: mull %esi
130 ; X86-NOBMI-NEXT: movl %edx, %ebp
131 ; X86-NOBMI-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
132 ; X86-NOBMI-NEXT: movl %ebx, %eax
133 ; X86-NOBMI-NEXT: mull %esi
134 ; X86-NOBMI-NEXT: movl %edx, %ebx
135 ; X86-NOBMI-NEXT: movl %eax, %esi
136 ; X86-NOBMI-NEXT: addl %ebp, %esi
137 ; X86-NOBMI-NEXT: adcl $0, %ebx
138 ; X86-NOBMI-NEXT: movl %edi, %eax
139 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
140 ; X86-NOBMI-NEXT: mull %edx
141 ; X86-NOBMI-NEXT: movl %edx, %ebp
142 ; X86-NOBMI-NEXT: movl %eax, %edi
143 ; X86-NOBMI-NEXT: addl %esi, %edi
144 ; X86-NOBMI-NEXT: adcl %ebx, %ebp
145 ; X86-NOBMI-NEXT: setb %bl
146 ; X86-NOBMI-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
147 ; X86-NOBMI-NEXT: mull {{[0-9]+}}(%esp)
148 ; X86-NOBMI-NEXT: addl %ebp, %eax
149 ; X86-NOBMI-NEXT: movzbl %bl, %esi
150 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %ebp
151 ; X86-NOBMI-NEXT: adcl %esi, %edx
152 ; X86-NOBMI-NEXT: movl %ecx, %ebx
153 ; X86-NOBMI-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
154 ; X86-NOBMI-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload
155 ; X86-NOBMI-NEXT: adcl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Folded Reload
156 ; X86-NOBMI-NEXT: adcl $0, %eax
157 ; X86-NOBMI-NEXT: adcl $0, %edx
158 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
159 ; X86-NOBMI-NEXT: movl %ecx, (%esi,%ebx,8)
160 ; X86-NOBMI-NEXT: movl %ebx, %ecx
161 ; X86-NOBMI-NEXT: movl %edi, 4(%esi,%ebx,8)
162 ; X86-NOBMI-NEXT: addl $1, %ecx
163 ; X86-NOBMI-NEXT: movl (%esp), %edi # 4-byte Reload
164 ; X86-NOBMI-NEXT: adcl $0, %edi
165 ; X86-NOBMI-NEXT: movl %ecx, %esi
166 ; X86-NOBMI-NEXT: xorl {{[0-9]+}}(%esp), %esi
167 ; X86-NOBMI-NEXT: movl %edi, (%esp) # 4-byte Spill
168 ; X86-NOBMI-NEXT: xorl %ebp, %edi
169 ; X86-NOBMI-NEXT: orl %esi, %edi
170 ; X86-NOBMI-NEXT: jne .LBB1_2
171 ; X86-NOBMI-NEXT: .LBB1_3: # %for.end
172 ; X86-NOBMI-NEXT: xorl %eax, %eax
173 ; X86-NOBMI-NEXT: xorl %edx, %edx
174 ; X86-NOBMI-NEXT: addl $20, %esp
175 ; X86-NOBMI-NEXT: popl %esi
176 ; X86-NOBMI-NEXT: popl %edi
177 ; X86-NOBMI-NEXT: popl %ebx
178 ; X86-NOBMI-NEXT: popl %ebp
179 ; X86-NOBMI-NEXT: retl
180 ;
181 ; X86-BMI-LABEL: mul1:
182 ; X86-BMI: # %bb.0: # %entry
183 ; X86-BMI-NEXT: pushl %ebp
184 ; X86-BMI-NEXT: pushl %ebx
185 ; X86-BMI-NEXT: pushl %edi
186 ; X86-BMI-NEXT: pushl %esi
187 ; X86-BMI-NEXT: subl $20, %esp
188 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %ecx
189 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %eax
190 ; X86-BMI-NEXT: orl %ecx, %eax
191 ; X86-BMI-NEXT: je .LBB1_3
192 ; X86-BMI-NEXT: # %bb.1: # %for.body.preheader
193 ; X86-BMI-NEXT: xorl %ecx, %ecx
194 ; X86-BMI-NEXT: xorl %eax, %eax
195 ; X86-BMI-NEXT: xorl %ebx, %ebx
196 ; X86-BMI-NEXT: xorl %ebp, %ebp
197 ; X86-BMI-NEXT: .p2align 4, 0x90
198 ; X86-BMI-NEXT: .LBB1_2: # %for.body
199 ; X86-BMI-NEXT: # =>This Inner Loop Header: Depth=1
200 ; X86-BMI-NEXT: movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
201 ; X86-BMI-NEXT: movl %eax, (%esp) # 4-byte Spill
202 ; X86-BMI-NEXT: movl %ebp, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
203 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %eax
204 ; X86-BMI-NEXT: movl (%eax,%ebx,8), %ecx
205 ; X86-BMI-NEXT: movl 4(%eax,%ebx,8), %esi
206 ; X86-BMI-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
207 ; X86-BMI-NEXT: movl %ecx, %edx
208 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %eax
209 ; X86-BMI-NEXT: mulxl %eax, %edx, %edi
210 ; X86-BMI-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
211 ; X86-BMI-NEXT: movl %esi, %edx
212 ; X86-BMI-NEXT: mulxl %eax, %esi, %eax
213 ; X86-BMI-NEXT: addl %edi, %esi
214 ; X86-BMI-NEXT: adcl $0, %eax
215 ; X86-BMI-NEXT: movl %ecx, %edx
216 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %ecx
217 ; X86-BMI-NEXT: mulxl %ecx, %edi, %ebp
218 ; X86-BMI-NEXT: addl %esi, %edi
219 ; X86-BMI-NEXT: adcl %eax, %ebp
220 ; X86-BMI-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload
221 ; X86-BMI-NEXT: mulxl %ecx, %ecx, %eax
222 ; X86-BMI-NEXT: setb %dl
223 ; X86-BMI-NEXT: addl %ebp, %ecx
224 ; X86-BMI-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload
225 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %esi
226 ; X86-BMI-NEXT: movzbl %dl, %edx
227 ; X86-BMI-NEXT: adcl %edx, %eax
228 ; X86-BMI-NEXT: movl %eax, %edx
229 ; X86-BMI-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
230 ; X86-BMI-NEXT: addl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Folded Reload
231 ; X86-BMI-NEXT: adcl (%esp), %edi # 4-byte Folded Reload
232 ; X86-BMI-NEXT: adcl $0, %ecx
233 ; X86-BMI-NEXT: adcl $0, %edx
234 ; X86-BMI-NEXT: movl %edx, (%esp) # 4-byte Spill
235 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %edx
236 ; X86-BMI-NEXT: movl %eax, (%edx,%ebx,8)
237 ; X86-BMI-NEXT: movl %edi, 4(%edx,%ebx,8)
238 ; X86-BMI-NEXT: movl {{[0-9]+}}(%esp), %edi
239 ; X86-BMI-NEXT: addl $1, %ebx
240 ; X86-BMI-NEXT: adcl $0, %ebp
241 ; X86-BMI-NEXT: movl %ebx, %edx
242 ; X86-BMI-NEXT: xorl %esi, %edx
243 ; X86-BMI-NEXT: movl %ebp, %esi
244 ; X86-BMI-NEXT: xorl %edi, %esi
245 ; X86-BMI-NEXT: orl %edx, %esi
246 ; X86-BMI-NEXT: movl (%esp), %eax # 4-byte Reload
247 ; X86-BMI-NEXT: jne .LBB1_2
248 ; X86-BMI-NEXT: .LBB1_3: # %for.end
249 ; X86-BMI-NEXT: xorl %eax, %eax
250 ; X86-BMI-NEXT: xorl %edx, %edx
251 ; X86-BMI-NEXT: addl $20, %esp
252 ; X86-BMI-NEXT: popl %esi
253 ; X86-BMI-NEXT: popl %edi
254 ; X86-BMI-NEXT: popl %ebx
255 ; X86-BMI-NEXT: popl %ebp
256 ; X86-BMI-NEXT: retl
257 ;
258 ; X64-NOBMI-LABEL: mul1:
259 ; X64-NOBMI: # %bb.0: # %entry
260 ; X64-NOBMI-NEXT: testq %rdi, %rdi
261 ; X64-NOBMI-NEXT: je .LBB1_3
262 ; X64-NOBMI-NEXT: # %bb.1: # %for.body.preheader
263 ; X64-NOBMI-NEXT: movq %rdx, %r8
264 ; X64-NOBMI-NEXT: xorl %r10d, %r10d
265 ; X64-NOBMI-NEXT: xorl %r9d, %r9d
266 ; X64-NOBMI-NEXT: .p2align 4, 0x90
267 ; X64-NOBMI-NEXT: .LBB1_2: # %for.body
268 ; X64-NOBMI-NEXT: # =>This Inner Loop Header: Depth=1
269 ; X64-NOBMI-NEXT: movq %rcx, %rax
270 ; X64-NOBMI-NEXT: mulq (%r8,%r9,8)
271 ; X64-NOBMI-NEXT: addq %r10, %rax
272 ; X64-NOBMI-NEXT: adcq $0, %rdx
273 ; X64-NOBMI-NEXT: movq %rax, (%rsi,%r9,8)
274 ; X64-NOBMI-NEXT: incq %r9
275 ; X64-NOBMI-NEXT: cmpq %r9, %rdi
276 ; X64-NOBMI-NEXT: movq %rdx, %r10
277 ; X64-NOBMI-NEXT: jne .LBB1_2
278 ; X64-NOBMI-NEXT: .LBB1_3: # %for.end
279 ; X64-NOBMI-NEXT: xorl %eax, %eax
280 ; X64-NOBMI-NEXT: retq
281 ;
282 ; X64-BMI-LABEL: mul1:
283 ; X64-BMI: # %bb.0: # %entry
284 ; X64-BMI-NEXT: testq %rdi, %rdi
285 ; X64-BMI-NEXT: je .LBB1_3
286 ; X64-BMI-NEXT: # %bb.1: # %for.body.preheader
287 ; X64-BMI-NEXT: movq %rdx, %rax
288 ; X64-BMI-NEXT: xorl %r9d, %r9d
289 ; X64-BMI-NEXT: xorl %r8d, %r8d
290 ; X64-BMI-NEXT: .p2align 4, 0x90
291 ; X64-BMI-NEXT: .LBB1_2: # %for.body
292 ; X64-BMI-NEXT: # =>This Inner Loop Header: Depth=1
293 ; X64-BMI-NEXT: movq %rcx, %rdx
294 ; X64-BMI-NEXT: mulxq (%rax,%r8,8), %r10, %rdx
295 ; X64-BMI-NEXT: addq %r9, %r10
296 ; X64-BMI-NEXT: adcq $0, %rdx
297 ; X64-BMI-NEXT: movq %r10, (%rsi,%r8,8)
298 ; X64-BMI-NEXT: incq %r8
299 ; X64-BMI-NEXT: cmpq %r8, %rdi
300 ; X64-BMI-NEXT: movq %rdx, %r9
301 ; X64-BMI-NEXT: jne .LBB1_2
302 ; X64-BMI-NEXT: .LBB1_3: # %for.end
303 ; X64-BMI-NEXT: xorl %eax, %eax
304 ; X64-BMI-NEXT: retq
305 entry:
306 %conv = zext i64 %y to i128
307 %cmp11 = icmp eq i64 %n, 0
308 br i1 %cmp11, label %for.end, label %for.body
310 for.body: ; preds = %entry, %for.body
311 %carry.013 = phi i64 [ %conv6, %for.body ], [ 0, %entry ]
312 %i.012 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
313 %arrayidx = getelementptr inbounds i64, ptr %x, i64 %i.012
314 %0 = load i64, ptr %arrayidx, align 8
315 %conv2 = zext i64 %0 to i128
316 %mul = mul i128 %conv2, %conv
317 %conv3 = zext i64 %carry.013 to i128
318 %add = add i128 %mul, %conv3
319 %conv4 = trunc i128 %add to i64
320 %arrayidx5 = getelementptr inbounds i64, ptr %z, i64 %i.012
321 store i64 %conv4, ptr %arrayidx5, align 8
322 %shr = lshr i128 %add, 64
323 %conv6 = trunc i128 %shr to i64
324 %inc = add i64 %i.012, 1
325 %exitcond = icmp eq i64 %inc, %n
326 br i1 %exitcond, label %for.end, label %for.body
328 for.end: ; preds = %for.body, %entry
329 ret i64 0
330 }