| blob | 4340f8fd484aeba6682a3242730148021e7e82db |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s --check-prefixes=X86,X86-FAST
3 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+slow-shld | FileCheck %s --check-prefixes=X86,X86-SLOW
4 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefixes=X64,X64-FAST
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+slow-shld | FileCheck %s --check-prefixes=X64,X64-SLOW
7 declare i8 @llvm.fshr.i8(i8, i8, i8) nounwind readnone
8 declare i16 @llvm.fshr.i16(i16, i16, i16) nounwind readnone
9 declare i32 @llvm.fshr.i32(i32, i32, i32) nounwind readnone
10 declare i64 @llvm.fshr.i64(i64, i64, i64) nounwind readnone
11 declare i128 @llvm.fshr.i128(i128, i128, i128) nounwind readnone
13 ;
14 ; Variable Funnel Shift
15 ;
17 define i8 @var_shift_i8(i8 %x, i8 %y, i8 %z) nounwind {
18 ; X86-LABEL: var_shift_i8:
19 ; X86: # %bb.0:
20 ; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
21 ; X86-NEXT: movzbl {{[0-9]+}}(%esp), %edx
22 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
23 ; X86-NEXT: shll $8, %eax
24 ; X86-NEXT: orl %edx, %eax
25 ; X86-NEXT: andb $7, %cl
26 ; X86-NEXT: shrl %cl, %eax
27 ; X86-NEXT: # kill: def $al killed $al killed $eax
28 ; X86-NEXT: retl
29 ;
30 ; X64-LABEL: var_shift_i8:
31 ; X64: # %bb.0:
32 ; X64-NEXT: movl %edx, %ecx
33 ; X64-NEXT: shll $8, %edi
34 ; X64-NEXT: movzbl %sil, %eax
35 ; X64-NEXT: orl %edi, %eax
36 ; X64-NEXT: andb $7, %cl
37 ; X64-NEXT: # kill: def $cl killed $cl killed $ecx
38 ; X64-NEXT: shrl %cl, %eax
39 ; X64-NEXT: # kill: def $al killed $al killed $eax
40 ; X64-NEXT: retq
41 %tmp = tail call i8 @llvm.fshr.i8(i8 %x, i8 %y, i8 %z)
42 ret i8 %tmp
43 }
45 define i16 @var_shift_i16(i16 %x, i16 %y, i16 %z) nounwind {
46 ; X86-FAST-LABEL: var_shift_i16:
47 ; X86-FAST: # %bb.0:
48 ; X86-FAST-NEXT: movzwl {{[0-9]+}}(%esp), %edx
49 ; X86-FAST-NEXT: movzwl {{[0-9]+}}(%esp), %eax
50 ; X86-FAST-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
51 ; X86-FAST-NEXT: andb $15, %cl
52 ; X86-FAST-NEXT: shrdw %cl, %dx, %ax
53 ; X86-FAST-NEXT: retl
54 ;
55 ; X86-SLOW-LABEL: var_shift_i16:
56 ; X86-SLOW: # %bb.0:
57 ; X86-SLOW-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
58 ; X86-SLOW-NEXT: movzwl {{[0-9]+}}(%esp), %edx
59 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %eax
60 ; X86-SLOW-NEXT: shll $16, %eax
61 ; X86-SLOW-NEXT: orl %edx, %eax
62 ; X86-SLOW-NEXT: andb $15, %cl
63 ; X86-SLOW-NEXT: shrl %cl, %eax
64 ; X86-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
65 ; X86-SLOW-NEXT: retl
66 ;
67 ; X64-FAST-LABEL: var_shift_i16:
68 ; X64-FAST: # %bb.0:
69 ; X64-FAST-NEXT: movl %edx, %ecx
70 ; X64-FAST-NEXT: movl %esi, %eax
71 ; X64-FAST-NEXT: andb $15, %cl
72 ; X64-FAST-NEXT: # kill: def $cl killed $cl killed $ecx
73 ; X64-FAST-NEXT: shrdw %cl, %di, %ax
74 ; X64-FAST-NEXT: # kill: def $ax killed $ax killed $eax
75 ; X64-FAST-NEXT: retq
76 ;
77 ; X64-SLOW-LABEL: var_shift_i16:
78 ; X64-SLOW: # %bb.0:
79 ; X64-SLOW-NEXT: movl %edx, %ecx
80 ; X64-SLOW-NEXT: shll $16, %edi
81 ; X64-SLOW-NEXT: movzwl %si, %eax
82 ; X64-SLOW-NEXT: orl %edi, %eax
83 ; X64-SLOW-NEXT: andb $15, %cl
84 ; X64-SLOW-NEXT: # kill: def $cl killed $cl killed $ecx
85 ; X64-SLOW-NEXT: shrl %cl, %eax
86 ; X64-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
87 ; X64-SLOW-NEXT: retq
88 %tmp = tail call i16 @llvm.fshr.i16(i16 %x, i16 %y, i16 %z)
89 ret i16 %tmp
90 }
92 define i32 @var_shift_i32(i32 %x, i32 %y, i32 %z) nounwind {
93 ; X86-FAST-LABEL: var_shift_i32:
94 ; X86-FAST: # %bb.0:
95 ; X86-FAST-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
96 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %edx
97 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %eax
98 ; X86-FAST-NEXT: shrdl %cl, %edx, %eax
99 ; X86-FAST-NEXT: retl
100 ;
101 ; X86-SLOW-LABEL: var_shift_i32:
102 ; X86-SLOW: # %bb.0:
103 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %eax
104 ; X86-SLOW-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
105 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %edx
106 ; X86-SLOW-NEXT: shrl %cl, %edx
107 ; X86-SLOW-NEXT: notb %cl
108 ; X86-SLOW-NEXT: addl %eax, %eax
109 ; X86-SLOW-NEXT: shll %cl, %eax
110 ; X86-SLOW-NEXT: orl %edx, %eax
111 ; X86-SLOW-NEXT: retl
112 ;
113 ; X64-FAST-LABEL: var_shift_i32:
114 ; X64-FAST: # %bb.0:
115 ; X64-FAST-NEXT: movl %edx, %ecx
116 ; X64-FAST-NEXT: movl %esi, %eax
117 ; X64-FAST-NEXT: # kill: def $cl killed $cl killed $ecx
118 ; X64-FAST-NEXT: shrdl %cl, %edi, %eax
119 ; X64-FAST-NEXT: retq
120 ;
121 ; X64-SLOW-LABEL: var_shift_i32:
122 ; X64-SLOW: # %bb.0:
123 ; X64-SLOW-NEXT: movl %edx, %ecx
124 ; X64-SLOW-NEXT: # kill: def $edi killed $edi def $rdi
125 ; X64-SLOW-NEXT: shrl %cl, %esi
126 ; X64-SLOW-NEXT: leal (%rdi,%rdi), %eax
127 ; X64-SLOW-NEXT: notb %cl
128 ; X64-SLOW-NEXT: # kill: def $cl killed $cl killed $ecx
129 ; X64-SLOW-NEXT: shll %cl, %eax
130 ; X64-SLOW-NEXT: orl %esi, %eax
131 ; X64-SLOW-NEXT: retq
132 %tmp = tail call i32 @llvm.fshr.i32(i32 %x, i32 %y, i32 %z)
133 ret i32 %tmp
134 }
136 define i32 @var_shift_i32_optsize(i32 %x, i32 %y, i32 %z) nounwind optsize {
137 ; X86-LABEL: var_shift_i32_optsize:
138 ; X86: # %bb.0:
139 ; X86-NEXT: movb {{[0-9]+}}(%esp), %cl
140 ; X86-NEXT: movl {{[0-9]+}}(%esp), %edx
141 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
142 ; X86-NEXT: shrdl %cl, %edx, %eax
143 ; X86-NEXT: retl
144 ;
145 ; X64-LABEL: var_shift_i32_optsize:
146 ; X64: # %bb.0:
147 ; X64-NEXT: movl %edx, %ecx
148 ; X64-NEXT: movl %esi, %eax
149 ; X64-NEXT: # kill: def $cl killed $cl killed $ecx
150 ; X64-NEXT: shrdl %cl, %edi, %eax
151 ; X64-NEXT: retq
152 %tmp = tail call i32 @llvm.fshr.i32(i32 %x, i32 %y, i32 %z)
153 ret i32 %tmp
154 }
156 define i32 @var_shift_i32_pgso(i32 %x, i32 %y, i32 %z) nounwind !prof !14 {
157 ; X86-LABEL: var_shift_i32_pgso:
158 ; X86: # %bb.0:
159 ; X86-NEXT: movb {{[0-9]+}}(%esp), %cl
160 ; X86-NEXT: movl {{[0-9]+}}(%esp), %edx
161 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
162 ; X86-NEXT: shrdl %cl, %edx, %eax
163 ; X86-NEXT: retl
164 ;
165 ; X64-LABEL: var_shift_i32_pgso:
166 ; X64: # %bb.0:
167 ; X64-NEXT: movl %edx, %ecx
168 ; X64-NEXT: movl %esi, %eax
169 ; X64-NEXT: # kill: def $cl killed $cl killed $ecx
170 ; X64-NEXT: shrdl %cl, %edi, %eax
171 ; X64-NEXT: retq
172 %tmp = tail call i32 @llvm.fshr.i32(i32 %x, i32 %y, i32 %z)
173 ret i32 %tmp
174 }
176 define i64 @var_shift_i64(i64 %x, i64 %y, i64 %z) nounwind {
177 ; X86-FAST-LABEL: var_shift_i64:
178 ; X86-FAST: # %bb.0:
179 ; X86-FAST-NEXT: pushl %esi
180 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %eax
181 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %esi
182 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %ecx
183 ; X86-FAST-NEXT: testb $32, %cl
184 ; X86-FAST-NEXT: je .LBB5_1
185 ; X86-FAST-NEXT: # %bb.2:
186 ; X86-FAST-NEXT: movl %esi, %edx
187 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %esi
188 ; X86-FAST-NEXT: jmp .LBB5_3
189 ; X86-FAST-NEXT: .LBB5_1:
190 ; X86-FAST-NEXT: movl %eax, %edx
191 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %eax
192 ; X86-FAST-NEXT: .LBB5_3:
193 ; X86-FAST-NEXT: shrdl %cl, %edx, %eax
194 ; X86-FAST-NEXT: # kill: def $cl killed $cl killed $ecx
195 ; X86-FAST-NEXT: shrdl %cl, %esi, %edx
196 ; X86-FAST-NEXT: popl %esi
197 ; X86-FAST-NEXT: retl
198 ;
199 ; X86-SLOW-LABEL: var_shift_i64:
200 ; X86-SLOW: # %bb.0:
201 ; X86-SLOW-NEXT: pushl %ebx
202 ; X86-SLOW-NEXT: pushl %edi
203 ; X86-SLOW-NEXT: pushl %esi
204 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %eax
205 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %edx
206 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ebx
207 ; X86-SLOW-NEXT: testb $32, %bl
208 ; X86-SLOW-NEXT: je .LBB5_1
209 ; X86-SLOW-NEXT: # %bb.2:
210 ; X86-SLOW-NEXT: movl %edx, %esi
211 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %edx
212 ; X86-SLOW-NEXT: jmp .LBB5_3
213 ; X86-SLOW-NEXT: .LBB5_1:
214 ; X86-SLOW-NEXT: movl %eax, %esi
215 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %eax
216 ; X86-SLOW-NEXT: .LBB5_3:
217 ; X86-SLOW-NEXT: leal (%esi,%esi), %edi
218 ; X86-SLOW-NEXT: movb %bl, %ch
219 ; X86-SLOW-NEXT: notb %ch
220 ; X86-SLOW-NEXT: movb %ch, %cl
221 ; X86-SLOW-NEXT: shll %cl, %edi
222 ; X86-SLOW-NEXT: movb %bl, %cl
223 ; X86-SLOW-NEXT: shrl %cl, %eax
224 ; X86-SLOW-NEXT: orl %edi, %eax
225 ; X86-SLOW-NEXT: shrl %cl, %esi
226 ; X86-SLOW-NEXT: addl %edx, %edx
227 ; X86-SLOW-NEXT: movb %ch, %cl
228 ; X86-SLOW-NEXT: shll %cl, %edx
229 ; X86-SLOW-NEXT: orl %esi, %edx
230 ; X86-SLOW-NEXT: popl %esi
231 ; X86-SLOW-NEXT: popl %edi
232 ; X86-SLOW-NEXT: popl %ebx
233 ; X86-SLOW-NEXT: retl
234 ;
235 ; X64-FAST-LABEL: var_shift_i64:
236 ; X64-FAST: # %bb.0:
237 ; X64-FAST-NEXT: movq %rdx, %rcx
238 ; X64-FAST-NEXT: movq %rsi, %rax
239 ; X64-FAST-NEXT: # kill: def $cl killed $cl killed $rcx
240 ; X64-FAST-NEXT: shrdq %cl, %rdi, %rax
241 ; X64-FAST-NEXT: retq
242 ;
243 ; X64-SLOW-LABEL: var_shift_i64:
244 ; X64-SLOW: # %bb.0:
245 ; X64-SLOW-NEXT: movq %rdx, %rcx
246 ; X64-SLOW-NEXT: shrq %cl, %rsi
247 ; X64-SLOW-NEXT: leaq (%rdi,%rdi), %rax
248 ; X64-SLOW-NEXT: notb %cl
249 ; X64-SLOW-NEXT: # kill: def $cl killed $cl killed $rcx
250 ; X64-SLOW-NEXT: shlq %cl, %rax
251 ; X64-SLOW-NEXT: orq %rsi, %rax
252 ; X64-SLOW-NEXT: retq
253 %tmp = tail call i64 @llvm.fshr.i64(i64 %x, i64 %y, i64 %z)
254 ret i64 %tmp
255 }
257 define i128 @var_shift_i128(i128 %x, i128 %y, i128 %z) nounwind {
258 ; X86-FAST-LABEL: var_shift_i128:
259 ; X86-FAST: # %bb.0:
260 ; X86-FAST-NEXT: pushl %ebp
261 ; X86-FAST-NEXT: pushl %ebx
262 ; X86-FAST-NEXT: pushl %edi
263 ; X86-FAST-NEXT: pushl %esi
264 ; X86-FAST-NEXT: pushl %eax
265 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %eax
266 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %edi
267 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %ebx
268 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %edx
269 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %esi
270 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %ecx
271 ; X86-FAST-NEXT: testb $64, %cl
272 ; X86-FAST-NEXT: je .LBB6_1
273 ; X86-FAST-NEXT: # %bb.2:
274 ; X86-FAST-NEXT: movl %edx, (%esp) # 4-byte Spill
275 ; X86-FAST-NEXT: movl %edi, %edx
276 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %edi
277 ; X86-FAST-NEXT: movl %esi, %ebp
278 ; X86-FAST-NEXT: movl %ebx, %esi
279 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %ebx
280 ; X86-FAST-NEXT: testb $32, %cl
281 ; X86-FAST-NEXT: je .LBB6_4
282 ; X86-FAST-NEXT: jmp .LBB6_5
283 ; X86-FAST-NEXT: .LBB6_1:
284 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %ebp
285 ; X86-FAST-NEXT: movl %ebp, (%esp) # 4-byte Spill
286 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %ebp
287 ; X86-FAST-NEXT: testb $32, %cl
288 ; X86-FAST-NEXT: jne .LBB6_5
289 ; X86-FAST-NEXT: .LBB6_4:
290 ; X86-FAST-NEXT: movl %edi, %ebx
291 ; X86-FAST-NEXT: movl %esi, %edi
292 ; X86-FAST-NEXT: movl %edx, %esi
293 ; X86-FAST-NEXT: movl %ebp, %edx
294 ; X86-FAST-NEXT: movl (%esp), %ebp # 4-byte Reload
295 ; X86-FAST-NEXT: .LBB6_5:
296 ; X86-FAST-NEXT: shrdl %cl, %edx, %ebp
297 ; X86-FAST-NEXT: shrdl %cl, %esi, %edx
298 ; X86-FAST-NEXT: shrdl %cl, %edi, %esi
299 ; X86-FAST-NEXT: # kill: def $cl killed $cl killed $ecx
300 ; X86-FAST-NEXT: shrdl %cl, %ebx, %edi
301 ; X86-FAST-NEXT: movl %edi, 12(%eax)
302 ; X86-FAST-NEXT: movl %esi, 8(%eax)
303 ; X86-FAST-NEXT: movl %edx, 4(%eax)
304 ; X86-FAST-NEXT: movl %ebp, (%eax)
305 ; X86-FAST-NEXT: addl $4, %esp
306 ; X86-FAST-NEXT: popl %esi
307 ; X86-FAST-NEXT: popl %edi
308 ; X86-FAST-NEXT: popl %ebx
309 ; X86-FAST-NEXT: popl %ebp
310 ; X86-FAST-NEXT: retl $4
311 ;
312 ; X86-SLOW-LABEL: var_shift_i128:
313 ; X86-SLOW: # %bb.0:
314 ; X86-SLOW-NEXT: pushl %ebp
315 ; X86-SLOW-NEXT: pushl %ebx
316 ; X86-SLOW-NEXT: pushl %edi
317 ; X86-SLOW-NEXT: pushl %esi
318 ; X86-SLOW-NEXT: subl $8, %esp
319 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ebx
320 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %esi
321 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ebp
322 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %edi
323 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ecx
324 ; X86-SLOW-NEXT: testb $64, %cl
325 ; X86-SLOW-NEXT: je .LBB6_1
326 ; X86-SLOW-NEXT: # %bb.2:
327 ; X86-SLOW-NEXT: movl %ebp, %eax
328 ; X86-SLOW-NEXT: movl %ebx, %ebp
329 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ebx
330 ; X86-SLOW-NEXT: movl %edi, %edx
331 ; X86-SLOW-NEXT: movl %esi, %edi
332 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %esi
333 ; X86-SLOW-NEXT: testb $32, %cl
334 ; X86-SLOW-NEXT: jne .LBB6_5
335 ; X86-SLOW-NEXT: .LBB6_4:
336 ; X86-SLOW-NEXT: movl %ebx, %esi
337 ; X86-SLOW-NEXT: movl %edi, (%esp) # 4-byte Spill
338 ; X86-SLOW-NEXT: movl %ebp, %edi
339 ; X86-SLOW-NEXT: movl %edx, %ebp
340 ; X86-SLOW-NEXT: movl %eax, %edx
341 ; X86-SLOW-NEXT: jmp .LBB6_6
342 ; X86-SLOW-NEXT: .LBB6_1:
343 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %eax
344 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %edx
345 ; X86-SLOW-NEXT: testb $32, %cl
346 ; X86-SLOW-NEXT: je .LBB6_4
347 ; X86-SLOW-NEXT: .LBB6_5:
348 ; X86-SLOW-NEXT: movl %ebx, (%esp) # 4-byte Spill
349 ; X86-SLOW-NEXT: .LBB6_6:
350 ; X86-SLOW-NEXT: shrl %cl, %edx
351 ; X86-SLOW-NEXT: movl %ecx, %ebx
352 ; X86-SLOW-NEXT: notb %bl
353 ; X86-SLOW-NEXT: leal (%ebp,%ebp), %eax
354 ; X86-SLOW-NEXT: movl %ebx, %ecx
355 ; X86-SLOW-NEXT: shll %cl, %eax
356 ; X86-SLOW-NEXT: orl %edx, %eax
357 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ecx
358 ; X86-SLOW-NEXT: # kill: def $cl killed $cl killed $ecx
359 ; X86-SLOW-NEXT: shrl %cl, %ebp
360 ; X86-SLOW-NEXT: leal (%edi,%edi), %edx
361 ; X86-SLOW-NEXT: movl %ebx, %ecx
362 ; X86-SLOW-NEXT: shll %cl, %edx
363 ; X86-SLOW-NEXT: orl %ebp, %edx
364 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ecx
365 ; X86-SLOW-NEXT: # kill: def $cl killed $cl killed $ecx
366 ; X86-SLOW-NEXT: shrl %cl, %edi
367 ; X86-SLOW-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
368 ; X86-SLOW-NEXT: movl (%esp), %edi # 4-byte Reload
369 ; X86-SLOW-NEXT: leal (%edi,%edi), %ebp
370 ; X86-SLOW-NEXT: movl %ebx, %ecx
371 ; X86-SLOW-NEXT: shll %cl, %ebp
372 ; X86-SLOW-NEXT: orl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Folded Reload
373 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ecx
374 ; X86-SLOW-NEXT: # kill: def $cl killed $cl killed $ecx
375 ; X86-SLOW-NEXT: shrl %cl, %edi
376 ; X86-SLOW-NEXT: addl %esi, %esi
377 ; X86-SLOW-NEXT: movl %ebx, %ecx
378 ; X86-SLOW-NEXT: shll %cl, %esi
379 ; X86-SLOW-NEXT: orl %edi, %esi
380 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ecx
381 ; X86-SLOW-NEXT: movl %esi, 12(%ecx)
382 ; X86-SLOW-NEXT: movl %ebp, 8(%ecx)
383 ; X86-SLOW-NEXT: movl %edx, 4(%ecx)
384 ; X86-SLOW-NEXT: movl %eax, (%ecx)
385 ; X86-SLOW-NEXT: movl %ecx, %eax
386 ; X86-SLOW-NEXT: addl $8, %esp
387 ; X86-SLOW-NEXT: popl %esi
388 ; X86-SLOW-NEXT: popl %edi
389 ; X86-SLOW-NEXT: popl %ebx
390 ; X86-SLOW-NEXT: popl %ebp
391 ; X86-SLOW-NEXT: retl $4
392 ;
393 ; X64-FAST-LABEL: var_shift_i128:
394 ; X64-FAST: # %bb.0:
395 ; X64-FAST-NEXT: movq %rdx, %rax
396 ; X64-FAST-NEXT: testb $64, %r8b
397 ; X64-FAST-NEXT: cmoveq %rdi, %rsi
398 ; X64-FAST-NEXT: cmoveq %rcx, %rdi
399 ; X64-FAST-NEXT: cmovneq %rcx, %rax
400 ; X64-FAST-NEXT: movl %r8d, %ecx
401 ; X64-FAST-NEXT: shrdq %cl, %rdi, %rax
402 ; X64-FAST-NEXT: shrdq %cl, %rsi, %rdi
403 ; X64-FAST-NEXT: movq %rdi, %rdx
404 ; X64-FAST-NEXT: retq
405 ;
406 ; X64-SLOW-LABEL: var_shift_i128:
407 ; X64-SLOW: # %bb.0:
408 ; X64-SLOW-NEXT: testb $64, %r8b
409 ; X64-SLOW-NEXT: cmoveq %rdi, %rsi
410 ; X64-SLOW-NEXT: cmoveq %rcx, %rdi
411 ; X64-SLOW-NEXT: cmovneq %rcx, %rdx
412 ; X64-SLOW-NEXT: movl %r8d, %ecx
413 ; X64-SLOW-NEXT: shrq %cl, %rdx
414 ; X64-SLOW-NEXT: leaq (%rdi,%rdi), %rax
415 ; X64-SLOW-NEXT: movl %r8d, %r9d
416 ; X64-SLOW-NEXT: notb %r9b
417 ; X64-SLOW-NEXT: movl %r9d, %ecx
418 ; X64-SLOW-NEXT: shlq %cl, %rax
419 ; X64-SLOW-NEXT: orq %rdx, %rax
420 ; X64-SLOW-NEXT: movl %r8d, %ecx
421 ; X64-SLOW-NEXT: shrq %cl, %rdi
422 ; X64-SLOW-NEXT: leaq (%rsi,%rsi), %rdx
423 ; X64-SLOW-NEXT: movl %r9d, %ecx
424 ; X64-SLOW-NEXT: shlq %cl, %rdx
425 ; X64-SLOW-NEXT: orq %rdi, %rdx
426 ; X64-SLOW-NEXT: retq
427 %tmp = tail call i128 @llvm.fshr.i128(i128 %x, i128 %y, i128 %z)
428 ret i128 %tmp
429 }
431 ;
432 ; Const Funnel Shift
433 ;
435 define i8 @const_shift_i8(i8 %x, i8 %y) nounwind {
436 ; X86-LABEL: const_shift_i8:
437 ; X86: # %bb.0:
438 ; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax
439 ; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
440 ; X86-NEXT: shrb $7, %cl
441 ; X86-NEXT: addb %al, %al
442 ; X86-NEXT: orb %cl, %al
443 ; X86-NEXT: retl
444 ;
445 ; X64-LABEL: const_shift_i8:
446 ; X64: # %bb.0:
447 ; X64-NEXT: # kill: def $edi killed $edi def $rdi
448 ; X64-NEXT: shrb $7, %sil
449 ; X64-NEXT: leal (%rdi,%rdi), %eax
450 ; X64-NEXT: orb %sil, %al
451 ; X64-NEXT: # kill: def $al killed $al killed $eax
452 ; X64-NEXT: retq
453 %tmp = tail call i8 @llvm.fshr.i8(i8 %x, i8 %y, i8 7)
454 ret i8 %tmp
455 }
457 define i16 @const_shift_i16(i16 %x, i16 %y) nounwind {
458 ; X86-FAST-LABEL: const_shift_i16:
459 ; X86-FAST: # %bb.0:
460 ; X86-FAST-NEXT: movzwl {{[0-9]+}}(%esp), %ecx
461 ; X86-FAST-NEXT: movzwl {{[0-9]+}}(%esp), %eax
462 ; X86-FAST-NEXT: shrdw $7, %cx, %ax
463 ; X86-FAST-NEXT: retl
464 ;
465 ; X86-SLOW-LABEL: const_shift_i16:
466 ; X86-SLOW: # %bb.0:
467 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %eax
468 ; X86-SLOW-NEXT: movzwl {{[0-9]+}}(%esp), %ecx
469 ; X86-SLOW-NEXT: shrl $7, %ecx
470 ; X86-SLOW-NEXT: shll $9, %eax
471 ; X86-SLOW-NEXT: orl %ecx, %eax
472 ; X86-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
473 ; X86-SLOW-NEXT: retl
474 ;
475 ; X64-FAST-LABEL: const_shift_i16:
476 ; X64-FAST: # %bb.0:
477 ; X64-FAST-NEXT: movl %esi, %eax
478 ; X64-FAST-NEXT: shrdw $7, %di, %ax
479 ; X64-FAST-NEXT: # kill: def $ax killed $ax killed $eax
480 ; X64-FAST-NEXT: retq
481 ;
482 ; X64-SLOW-LABEL: const_shift_i16:
483 ; X64-SLOW: # %bb.0:
484 ; X64-SLOW-NEXT: movzwl %si, %eax
485 ; X64-SLOW-NEXT: shll $9, %edi
486 ; X64-SLOW-NEXT: shrl $7, %eax
487 ; X64-SLOW-NEXT: orl %edi, %eax
488 ; X64-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
489 ; X64-SLOW-NEXT: retq
490 %tmp = tail call i16 @llvm.fshr.i16(i16 %x, i16 %y, i16 7)
491 ret i16 %tmp
492 }
494 define i32 @const_shift_i32(i32 %x, i32 %y) nounwind {
495 ; X86-FAST-LABEL: const_shift_i32:
496 ; X86-FAST: # %bb.0:
497 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %ecx
498 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %eax
499 ; X86-FAST-NEXT: shrdl $7, %ecx, %eax
500 ; X86-FAST-NEXT: retl
501 ;
502 ; X86-SLOW-LABEL: const_shift_i32:
503 ; X86-SLOW: # %bb.0:
504 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %eax
505 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ecx
506 ; X86-SLOW-NEXT: shrl $7, %ecx
507 ; X86-SLOW-NEXT: shll $25, %eax
508 ; X86-SLOW-NEXT: orl %ecx, %eax
509 ; X86-SLOW-NEXT: retl
510 ;
511 ; X64-FAST-LABEL: const_shift_i32:
512 ; X64-FAST: # %bb.0:
513 ; X64-FAST-NEXT: movl %edi, %eax
514 ; X64-FAST-NEXT: shldl $25, %esi, %eax
515 ; X64-FAST-NEXT: retq
516 ;
517 ; X64-SLOW-LABEL: const_shift_i32:
518 ; X64-SLOW: # %bb.0:
519 ; X64-SLOW-NEXT: # kill: def $esi killed $esi def $rsi
520 ; X64-SLOW-NEXT: # kill: def $edi killed $edi def $rdi
521 ; X64-SLOW-NEXT: shrl $7, %esi
522 ; X64-SLOW-NEXT: shll $25, %edi
523 ; X64-SLOW-NEXT: leal (%rdi,%rsi), %eax
524 ; X64-SLOW-NEXT: retq
525 %tmp = tail call i32 @llvm.fshr.i32(i32 %x, i32 %y, i32 7)
526 ret i32 %tmp
527 }
529 define i64 @const_shift_i64(i64 %x, i64 %y) nounwind {
530 ; X86-FAST-LABEL: const_shift_i64:
531 ; X86-FAST: # %bb.0:
532 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %edx
533 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %eax
534 ; X86-FAST-NEXT: movl {{[0-9]+}}(%esp), %ecx
535 ; X86-FAST-NEXT: shldl $25, %ecx, %edx
536 ; X86-FAST-NEXT: shrdl $7, %ecx, %eax
537 ; X86-FAST-NEXT: retl
538 ;
539 ; X86-SLOW-LABEL: const_shift_i64:
540 ; X86-SLOW: # %bb.0:
541 ; X86-SLOW-NEXT: pushl %esi
542 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %edx
543 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %ecx
544 ; X86-SLOW-NEXT: movl {{[0-9]+}}(%esp), %esi
545 ; X86-SLOW-NEXT: shrl $7, %ecx
546 ; X86-SLOW-NEXT: movl %esi, %eax
547 ; X86-SLOW-NEXT: shll $25, %eax
548 ; X86-SLOW-NEXT: orl %ecx, %eax
549 ; X86-SLOW-NEXT: shrl $7, %esi
550 ; X86-SLOW-NEXT: shll $25, %edx
551 ; X86-SLOW-NEXT: orl %esi, %edx
552 ; X86-SLOW-NEXT: popl %esi
553 ; X86-SLOW-NEXT: retl
554 ;
555 ; X64-FAST-LABEL: const_shift_i64:
556 ; X64-FAST: # %bb.0:
557 ; X64-FAST-NEXT: movq %rdi, %rax
558 ; X64-FAST-NEXT: shldq $57, %rsi, %rax
559 ; X64-FAST-NEXT: retq
560 ;
561 ; X64-SLOW-LABEL: const_shift_i64:
562 ; X64-SLOW: # %bb.0:
563 ; X64-SLOW-NEXT: shrq $7, %rsi
564 ; X64-SLOW-NEXT: shlq $57, %rdi
565 ; X64-SLOW-NEXT: leaq (%rdi,%rsi), %rax
566 ; X64-SLOW-NEXT: retq
567 %tmp = tail call i64 @llvm.fshr.i64(i64 %x, i64 %y, i64 7)
568 ret i64 %tmp
569 }
571 ;
572 ; Combine Consecutive Loads
573 ;
575 define i8 @combine_fshr_load_i8(ptr %p) nounwind {
576 ; X86-LABEL: combine_fshr_load_i8:
577 ; X86: # %bb.0:
578 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
579 ; X86-NEXT: movzbl (%eax), %eax
580 ; X86-NEXT: retl
581 ;
582 ; X64-LABEL: combine_fshr_load_i8:
583 ; X64: # %bb.0:
584 ; X64-NEXT: movzbl (%rdi), %eax
585 ; X64-NEXT: retq
586 %p1 = getelementptr i8, ptr %p, i32 1
587 %ld0 = load i8, ptr%p
588 %ld1 = load i8, ptr%p1
589 %res = call i8 @llvm.fshr.i8(i8 %ld1, i8 %ld0, i8 8)
590 ret i8 %res
591 }
593 define i16 @combine_fshr_load_i16(ptr %p) nounwind {
594 ; X86-LABEL: combine_fshr_load_i16:
595 ; X86: # %bb.0:
596 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
597 ; X86-NEXT: movzwl 1(%eax), %eax
598 ; X86-NEXT: retl
599 ;
600 ; X64-LABEL: combine_fshr_load_i16:
601 ; X64: # %bb.0:
602 ; X64-NEXT: movzwl 1(%rdi), %eax
603 ; X64-NEXT: retq
604 %p1 = getelementptr i16, ptr %p, i32 1
605 %ld0 = load i16, ptr%p
606 %ld1 = load i16, ptr%p1
607 %res = call i16 @llvm.fshr.i16(i16 %ld1, i16 %ld0, i16 8)
608 ret i16 %res
609 }
611 define i32 @combine_fshr_load_i32(ptr %p) nounwind {
612 ; X86-LABEL: combine_fshr_load_i32:
613 ; X86: # %bb.0:
614 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
615 ; X86-NEXT: movl 9(%eax), %eax
616 ; X86-NEXT: retl
617 ;
618 ; X64-LABEL: combine_fshr_load_i32:
619 ; X64: # %bb.0:
620 ; X64-NEXT: movl 9(%rdi), %eax
621 ; X64-NEXT: retq
622 %p0 = getelementptr i32, ptr %p, i32 2
623 %p1 = getelementptr i32, ptr %p, i32 3
624 %ld0 = load i32, ptr%p0
625 %ld1 = load i32, ptr%p1
626 %res = call i32 @llvm.fshr.i32(i32 %ld1, i32 %ld0, i32 8)
627 ret i32 %res
628 }
630 define i64 @combine_fshr_load_i64(ptr %p) nounwind {
631 ; X86-LABEL: combine_fshr_load_i64:
632 ; X86: # %bb.0:
633 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
634 ; X86-NEXT: movl 11(%ecx), %eax
635 ; X86-NEXT: movl 15(%ecx), %edx
636 ; X86-NEXT: retl
637 ;
638 ; X64-LABEL: combine_fshr_load_i64:
639 ; X64: # %bb.0:
640 ; X64-NEXT: movq 11(%rdi), %rax
641 ; X64-NEXT: retq
642 %p0 = getelementptr i64, ptr %p, i64 1
643 %p1 = getelementptr i64, ptr %p, i64 2
644 %ld0 = load i64, ptr%p0
645 %ld1 = load i64, ptr%p1
646 %res = call i64 @llvm.fshr.i64(i64 %ld1, i64 %ld0, i64 24)
647 ret i64 %res
648 }
650 !llvm.module.flags = !{!0}
651 !0 = !{i32 1, !"ProfileSummary", !1}
652 !1 = !{!2, !3, !4, !5, !6, !7, !8, !9}
653 !2 = !{!"ProfileFormat", !"InstrProf"}
654 !3 = !{!"TotalCount", i64 10000}
655 !4 = !{!"MaxCount", i64 10}
656 !5 = !{!"MaxInternalCount", i64 1}
657 !6 = !{!"MaxFunctionCount", i64 1000}
658 !7 = !{!"NumCounts", i64 3}
659 !8 = !{!"NumFunctions", i64 3}
660 !9 = !{!"DetailedSummary", !10}
661 !10 = !{!11, !12, !13}
662 !11 = !{i32 10000, i64 100, i32 1}
663 !12 = !{i32 999000, i64 100, i32 1}
664 !13 = !{i32 999999, i64 1, i32 2}
665 !14 = !{!"function_entry_count", i64 0}