| blob | 848b920490ab83b766ac56d9851007daab99390a |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=-bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-NOBMI
3 ; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-BMI1
4 ; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,+tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-BMI1
5 ; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-BMI2
6 ; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2 < %s | FileCheck %s --check-prefixes=X86,X86-BMI2
7 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=-bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-NOBMI
8 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,-tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-BMI1,X64-BMI1NOTBM
9 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,+tbm,-bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-BMI1,X64-BMI1TBM
10 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,+tbm,+bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-BMI2,X64-BMI2TBM
11 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+bmi,-tbm,+bmi2 < %s | FileCheck %s --check-prefixes=X64,X64-BMI2,X64-BMI2NOTBM
13 ; *Please* keep in sync with test/CodeGen/AArch64/extract-lowbits.ll
15 ; https://bugs.llvm.org/show_bug.cgi?id=36419
16 ; https://bugs.llvm.org/show_bug.cgi?id=37603
17 ; https://bugs.llvm.org/show_bug.cgi?id=37610
19 ; Patterns:
20 ; a) x & (1 << nbits) - 1
21 ; b) x & ~(-1 << nbits)
22 ; c) x & (-1 >> (32 - y))
23 ; d) x << (32 - y) >> (32 - y)
24 ; are equivalent.
26 ; ---------------------------------------------------------------------------- ;
27 ; Pattern a. 32-bit
28 ; ---------------------------------------------------------------------------- ;
30 define i32 @bzhi32_a0(i32 %val, i32 %numlowbits) nounwind {
31 ; X86-NOBMI-LABEL: bzhi32_a0:
32 ; X86-NOBMI: # %bb.0:
33 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
34 ; X86-NOBMI-NEXT: movl $1, %eax
35 ; X86-NOBMI-NEXT: shll %cl, %eax
36 ; X86-NOBMI-NEXT: decl %eax
37 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
38 ; X86-NOBMI-NEXT: retl
39 ;
40 ; X86-BMI1-LABEL: bzhi32_a0:
41 ; X86-BMI1: # %bb.0:
42 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
43 ; X86-BMI1-NEXT: shll $8, %eax
44 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
45 ; X86-BMI1-NEXT: retl
46 ;
47 ; X86-BMI2-LABEL: bzhi32_a0:
48 ; X86-BMI2: # %bb.0:
49 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
50 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
51 ; X86-BMI2-NEXT: retl
52 ;
53 ; X64-NOBMI-LABEL: bzhi32_a0:
54 ; X64-NOBMI: # %bb.0:
55 ; X64-NOBMI-NEXT: movl %esi, %ecx
56 ; X64-NOBMI-NEXT: movl $1, %eax
57 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
58 ; X64-NOBMI-NEXT: shll %cl, %eax
59 ; X64-NOBMI-NEXT: decl %eax
60 ; X64-NOBMI-NEXT: andl %edi, %eax
61 ; X64-NOBMI-NEXT: retq
62 ;
63 ; X64-BMI1-LABEL: bzhi32_a0:
64 ; X64-BMI1: # %bb.0:
65 ; X64-BMI1-NEXT: shll $8, %esi
66 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
67 ; X64-BMI1-NEXT: retq
68 ;
69 ; X64-BMI2-LABEL: bzhi32_a0:
70 ; X64-BMI2: # %bb.0:
71 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
72 ; X64-BMI2-NEXT: retq
73 %onebit = shl i32 1, %numlowbits
74 %mask = add nsw i32 %onebit, -1
75 %masked = and i32 %mask, %val
76 ret i32 %masked
77 }
79 define i32 @bzhi32_a1_indexzext(i32 %val, i8 zeroext %numlowbits) nounwind {
80 ; X86-NOBMI-LABEL: bzhi32_a1_indexzext:
81 ; X86-NOBMI: # %bb.0:
82 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
83 ; X86-NOBMI-NEXT: movl $1, %eax
84 ; X86-NOBMI-NEXT: shll %cl, %eax
85 ; X86-NOBMI-NEXT: decl %eax
86 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
87 ; X86-NOBMI-NEXT: retl
88 ;
89 ; X86-BMI1-LABEL: bzhi32_a1_indexzext:
90 ; X86-BMI1: # %bb.0:
91 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
92 ; X86-BMI1-NEXT: shll $8, %eax
93 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
94 ; X86-BMI1-NEXT: retl
95 ;
96 ; X86-BMI2-LABEL: bzhi32_a1_indexzext:
97 ; X86-BMI2: # %bb.0:
98 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
99 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
100 ; X86-BMI2-NEXT: retl
101 ;
102 ; X64-NOBMI-LABEL: bzhi32_a1_indexzext:
103 ; X64-NOBMI: # %bb.0:
104 ; X64-NOBMI-NEXT: movl %esi, %ecx
105 ; X64-NOBMI-NEXT: movl $1, %eax
106 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
107 ; X64-NOBMI-NEXT: shll %cl, %eax
108 ; X64-NOBMI-NEXT: decl %eax
109 ; X64-NOBMI-NEXT: andl %edi, %eax
110 ; X64-NOBMI-NEXT: retq
111 ;
112 ; X64-BMI1-LABEL: bzhi32_a1_indexzext:
113 ; X64-BMI1: # %bb.0:
114 ; X64-BMI1-NEXT: shll $8, %esi
115 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
116 ; X64-BMI1-NEXT: retq
117 ;
118 ; X64-BMI2-LABEL: bzhi32_a1_indexzext:
119 ; X64-BMI2: # %bb.0:
120 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
121 ; X64-BMI2-NEXT: retq
122 %conv = zext i8 %numlowbits to i32
123 %onebit = shl i32 1, %conv
124 %mask = add nsw i32 %onebit, -1
125 %masked = and i32 %mask, %val
126 ret i32 %masked
127 }
129 define i32 @bzhi32_a2_load(ptr %w, i32 %numlowbits) nounwind {
130 ; X86-NOBMI-LABEL: bzhi32_a2_load:
131 ; X86-NOBMI: # %bb.0:
132 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
133 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
134 ; X86-NOBMI-NEXT: movl $1, %eax
135 ; X86-NOBMI-NEXT: shll %cl, %eax
136 ; X86-NOBMI-NEXT: decl %eax
137 ; X86-NOBMI-NEXT: andl (%edx), %eax
138 ; X86-NOBMI-NEXT: retl
139 ;
140 ; X86-BMI1-LABEL: bzhi32_a2_load:
141 ; X86-BMI1: # %bb.0:
142 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
143 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
144 ; X86-BMI1-NEXT: shll $8, %ecx
145 ; X86-BMI1-NEXT: bextrl %ecx, (%eax), %eax
146 ; X86-BMI1-NEXT: retl
147 ;
148 ; X86-BMI2-LABEL: bzhi32_a2_load:
149 ; X86-BMI2: # %bb.0:
150 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
151 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
152 ; X86-BMI2-NEXT: bzhil %ecx, (%eax), %eax
153 ; X86-BMI2-NEXT: retl
154 ;
155 ; X64-NOBMI-LABEL: bzhi32_a2_load:
156 ; X64-NOBMI: # %bb.0:
157 ; X64-NOBMI-NEXT: movl %esi, %ecx
158 ; X64-NOBMI-NEXT: movl $1, %eax
159 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
160 ; X64-NOBMI-NEXT: shll %cl, %eax
161 ; X64-NOBMI-NEXT: decl %eax
162 ; X64-NOBMI-NEXT: andl (%rdi), %eax
163 ; X64-NOBMI-NEXT: retq
164 ;
165 ; X64-BMI1-LABEL: bzhi32_a2_load:
166 ; X64-BMI1: # %bb.0:
167 ; X64-BMI1-NEXT: shll $8, %esi
168 ; X64-BMI1-NEXT: bextrl %esi, (%rdi), %eax
169 ; X64-BMI1-NEXT: retq
170 ;
171 ; X64-BMI2-LABEL: bzhi32_a2_load:
172 ; X64-BMI2: # %bb.0:
173 ; X64-BMI2-NEXT: bzhil %esi, (%rdi), %eax
174 ; X64-BMI2-NEXT: retq
175 %val = load i32, ptr %w
176 %onebit = shl i32 1, %numlowbits
177 %mask = add nsw i32 %onebit, -1
178 %masked = and i32 %mask, %val
179 ret i32 %masked
180 }
182 define i32 @bzhi32_a3_load_indexzext(ptr %w, i8 zeroext %numlowbits) nounwind {
183 ; X86-NOBMI-LABEL: bzhi32_a3_load_indexzext:
184 ; X86-NOBMI: # %bb.0:
185 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
186 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
187 ; X86-NOBMI-NEXT: movl $1, %eax
188 ; X86-NOBMI-NEXT: shll %cl, %eax
189 ; X86-NOBMI-NEXT: decl %eax
190 ; X86-NOBMI-NEXT: andl (%edx), %eax
191 ; X86-NOBMI-NEXT: retl
192 ;
193 ; X86-BMI1-LABEL: bzhi32_a3_load_indexzext:
194 ; X86-BMI1: # %bb.0:
195 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
196 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
197 ; X86-BMI1-NEXT: shll $8, %ecx
198 ; X86-BMI1-NEXT: bextrl %ecx, (%eax), %eax
199 ; X86-BMI1-NEXT: retl
200 ;
201 ; X86-BMI2-LABEL: bzhi32_a3_load_indexzext:
202 ; X86-BMI2: # %bb.0:
203 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
204 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
205 ; X86-BMI2-NEXT: bzhil %ecx, (%eax), %eax
206 ; X86-BMI2-NEXT: retl
207 ;
208 ; X64-NOBMI-LABEL: bzhi32_a3_load_indexzext:
209 ; X64-NOBMI: # %bb.0:
210 ; X64-NOBMI-NEXT: movl %esi, %ecx
211 ; X64-NOBMI-NEXT: movl $1, %eax
212 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
213 ; X64-NOBMI-NEXT: shll %cl, %eax
214 ; X64-NOBMI-NEXT: decl %eax
215 ; X64-NOBMI-NEXT: andl (%rdi), %eax
216 ; X64-NOBMI-NEXT: retq
217 ;
218 ; X64-BMI1-LABEL: bzhi32_a3_load_indexzext:
219 ; X64-BMI1: # %bb.0:
220 ; X64-BMI1-NEXT: shll $8, %esi
221 ; X64-BMI1-NEXT: bextrl %esi, (%rdi), %eax
222 ; X64-BMI1-NEXT: retq
223 ;
224 ; X64-BMI2-LABEL: bzhi32_a3_load_indexzext:
225 ; X64-BMI2: # %bb.0:
226 ; X64-BMI2-NEXT: bzhil %esi, (%rdi), %eax
227 ; X64-BMI2-NEXT: retq
228 %val = load i32, ptr %w
229 %conv = zext i8 %numlowbits to i32
230 %onebit = shl i32 1, %conv
231 %mask = add nsw i32 %onebit, -1
232 %masked = and i32 %mask, %val
233 ret i32 %masked
234 }
236 define i32 @bzhi32_a4_commutative(i32 %val, i32 %numlowbits) nounwind {
237 ; X86-NOBMI-LABEL: bzhi32_a4_commutative:
238 ; X86-NOBMI: # %bb.0:
239 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
240 ; X86-NOBMI-NEXT: movl $1, %eax
241 ; X86-NOBMI-NEXT: shll %cl, %eax
242 ; X86-NOBMI-NEXT: decl %eax
243 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
244 ; X86-NOBMI-NEXT: retl
245 ;
246 ; X86-BMI1-LABEL: bzhi32_a4_commutative:
247 ; X86-BMI1: # %bb.0:
248 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
249 ; X86-BMI1-NEXT: shll $8, %eax
250 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
251 ; X86-BMI1-NEXT: retl
252 ;
253 ; X86-BMI2-LABEL: bzhi32_a4_commutative:
254 ; X86-BMI2: # %bb.0:
255 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
256 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
257 ; X86-BMI2-NEXT: retl
258 ;
259 ; X64-NOBMI-LABEL: bzhi32_a4_commutative:
260 ; X64-NOBMI: # %bb.0:
261 ; X64-NOBMI-NEXT: movl %esi, %ecx
262 ; X64-NOBMI-NEXT: movl $1, %eax
263 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
264 ; X64-NOBMI-NEXT: shll %cl, %eax
265 ; X64-NOBMI-NEXT: decl %eax
266 ; X64-NOBMI-NEXT: andl %edi, %eax
267 ; X64-NOBMI-NEXT: retq
268 ;
269 ; X64-BMI1-LABEL: bzhi32_a4_commutative:
270 ; X64-BMI1: # %bb.0:
271 ; X64-BMI1-NEXT: shll $8, %esi
272 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
273 ; X64-BMI1-NEXT: retq
274 ;
275 ; X64-BMI2-LABEL: bzhi32_a4_commutative:
276 ; X64-BMI2: # %bb.0:
277 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
278 ; X64-BMI2-NEXT: retq
279 %onebit = shl i32 1, %numlowbits
280 %mask = add nsw i32 %onebit, -1
281 %masked = and i32 %val, %mask ; swapped order
282 ret i32 %masked
283 }
285 ; 64-bit
287 define i64 @bzhi64_a0(i64 %val, i64 %numlowbits) nounwind {
288 ; X86-NOBMI-LABEL: bzhi64_a0:
289 ; X86-NOBMI: # %bb.0:
290 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
291 ; X86-NOBMI-NEXT: movl $1, %eax
292 ; X86-NOBMI-NEXT: xorl %edx, %edx
293 ; X86-NOBMI-NEXT: shldl %cl, %eax, %edx
294 ; X86-NOBMI-NEXT: shll %cl, %eax
295 ; X86-NOBMI-NEXT: testb $32, %cl
296 ; X86-NOBMI-NEXT: je .LBB5_2
297 ; X86-NOBMI-NEXT: # %bb.1:
298 ; X86-NOBMI-NEXT: movl %eax, %edx
299 ; X86-NOBMI-NEXT: xorl %eax, %eax
300 ; X86-NOBMI-NEXT: .LBB5_2:
301 ; X86-NOBMI-NEXT: addl $-1, %eax
302 ; X86-NOBMI-NEXT: adcl $-1, %edx
303 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
304 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
305 ; X86-NOBMI-NEXT: retl
306 ;
307 ; X86-BMI1-LABEL: bzhi64_a0:
308 ; X86-BMI1: # %bb.0:
309 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
310 ; X86-BMI1-NEXT: movl $1, %eax
311 ; X86-BMI1-NEXT: xorl %edx, %edx
312 ; X86-BMI1-NEXT: shldl %cl, %eax, %edx
313 ; X86-BMI1-NEXT: shll %cl, %eax
314 ; X86-BMI1-NEXT: testb $32, %cl
315 ; X86-BMI1-NEXT: je .LBB5_2
316 ; X86-BMI1-NEXT: # %bb.1:
317 ; X86-BMI1-NEXT: movl %eax, %edx
318 ; X86-BMI1-NEXT: xorl %eax, %eax
319 ; X86-BMI1-NEXT: .LBB5_2:
320 ; X86-BMI1-NEXT: addl $-1, %eax
321 ; X86-BMI1-NEXT: adcl $-1, %edx
322 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %edx
323 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
324 ; X86-BMI1-NEXT: retl
325 ;
326 ; X86-BMI2-LABEL: bzhi64_a0:
327 ; X86-BMI2: # %bb.0:
328 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
329 ; X86-BMI2-NEXT: movl $1, %eax
330 ; X86-BMI2-NEXT: xorl %edx, %edx
331 ; X86-BMI2-NEXT: shldl %cl, %eax, %edx
332 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %eax
333 ; X86-BMI2-NEXT: testb $32, %cl
334 ; X86-BMI2-NEXT: je .LBB5_2
335 ; X86-BMI2-NEXT: # %bb.1:
336 ; X86-BMI2-NEXT: movl %eax, %edx
337 ; X86-BMI2-NEXT: xorl %eax, %eax
338 ; X86-BMI2-NEXT: .LBB5_2:
339 ; X86-BMI2-NEXT: addl $-1, %eax
340 ; X86-BMI2-NEXT: adcl $-1, %edx
341 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx
342 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
343 ; X86-BMI2-NEXT: retl
344 ;
345 ; X64-NOBMI-LABEL: bzhi64_a0:
346 ; X64-NOBMI: # %bb.0:
347 ; X64-NOBMI-NEXT: movq %rsi, %rcx
348 ; X64-NOBMI-NEXT: movl $1, %eax
349 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
350 ; X64-NOBMI-NEXT: shlq %cl, %rax
351 ; X64-NOBMI-NEXT: decq %rax
352 ; X64-NOBMI-NEXT: andq %rdi, %rax
353 ; X64-NOBMI-NEXT: retq
354 ;
355 ; X64-BMI1-LABEL: bzhi64_a0:
356 ; X64-BMI1: # %bb.0:
357 ; X64-BMI1-NEXT: shll $8, %esi
358 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
359 ; X64-BMI1-NEXT: retq
360 ;
361 ; X64-BMI2-LABEL: bzhi64_a0:
362 ; X64-BMI2: # %bb.0:
363 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
364 ; X64-BMI2-NEXT: retq
365 %onebit = shl i64 1, %numlowbits
366 %mask = add nsw i64 %onebit, -1
367 %masked = and i64 %mask, %val
368 ret i64 %masked
369 }
371 ; Check that we don't throw away the vreg_width-1 mask if not using shifts
372 define i64 @bzhi64_a0_masked(i64 %val, i64 %numlowbits) nounwind {
373 ; X86-NOBMI-LABEL: bzhi64_a0_masked:
374 ; X86-NOBMI: # %bb.0:
375 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
376 ; X86-NOBMI-NEXT: movl $1, %eax
377 ; X86-NOBMI-NEXT: xorl %edx, %edx
378 ; X86-NOBMI-NEXT: shldl %cl, %eax, %edx
379 ; X86-NOBMI-NEXT: shll %cl, %eax
380 ; X86-NOBMI-NEXT: testb $32, %cl
381 ; X86-NOBMI-NEXT: je .LBB6_2
382 ; X86-NOBMI-NEXT: # %bb.1:
383 ; X86-NOBMI-NEXT: movl %eax, %edx
384 ; X86-NOBMI-NEXT: xorl %eax, %eax
385 ; X86-NOBMI-NEXT: .LBB6_2:
386 ; X86-NOBMI-NEXT: addl $-1, %eax
387 ; X86-NOBMI-NEXT: adcl $-1, %edx
388 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
389 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
390 ; X86-NOBMI-NEXT: retl
391 ;
392 ; X86-BMI1-LABEL: bzhi64_a0_masked:
393 ; X86-BMI1: # %bb.0:
394 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
395 ; X86-BMI1-NEXT: movl $1, %eax
396 ; X86-BMI1-NEXT: xorl %edx, %edx
397 ; X86-BMI1-NEXT: shldl %cl, %eax, %edx
398 ; X86-BMI1-NEXT: shll %cl, %eax
399 ; X86-BMI1-NEXT: testb $32, %cl
400 ; X86-BMI1-NEXT: je .LBB6_2
401 ; X86-BMI1-NEXT: # %bb.1:
402 ; X86-BMI1-NEXT: movl %eax, %edx
403 ; X86-BMI1-NEXT: xorl %eax, %eax
404 ; X86-BMI1-NEXT: .LBB6_2:
405 ; X86-BMI1-NEXT: addl $-1, %eax
406 ; X86-BMI1-NEXT: adcl $-1, %edx
407 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %edx
408 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
409 ; X86-BMI1-NEXT: retl
410 ;
411 ; X86-BMI2-LABEL: bzhi64_a0_masked:
412 ; X86-BMI2: # %bb.0:
413 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
414 ; X86-BMI2-NEXT: movl $1, %eax
415 ; X86-BMI2-NEXT: xorl %edx, %edx
416 ; X86-BMI2-NEXT: shldl %cl, %eax, %edx
417 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %eax
418 ; X86-BMI2-NEXT: testb $32, %cl
419 ; X86-BMI2-NEXT: je .LBB6_2
420 ; X86-BMI2-NEXT: # %bb.1:
421 ; X86-BMI2-NEXT: movl %eax, %edx
422 ; X86-BMI2-NEXT: xorl %eax, %eax
423 ; X86-BMI2-NEXT: .LBB6_2:
424 ; X86-BMI2-NEXT: addl $-1, %eax
425 ; X86-BMI2-NEXT: adcl $-1, %edx
426 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx
427 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
428 ; X86-BMI2-NEXT: retl
429 ;
430 ; X64-NOBMI-LABEL: bzhi64_a0_masked:
431 ; X64-NOBMI: # %bb.0:
432 ; X64-NOBMI-NEXT: movq %rsi, %rcx
433 ; X64-NOBMI-NEXT: movl $1, %eax
434 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
435 ; X64-NOBMI-NEXT: shlq %cl, %rax
436 ; X64-NOBMI-NEXT: decq %rax
437 ; X64-NOBMI-NEXT: andq %rdi, %rax
438 ; X64-NOBMI-NEXT: retq
439 ;
440 ; X64-BMI1-LABEL: bzhi64_a0_masked:
441 ; X64-BMI1: # %bb.0:
442 ; X64-BMI1-NEXT: andb $63, %sil
443 ; X64-BMI1-NEXT: shll $8, %esi
444 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
445 ; X64-BMI1-NEXT: retq
446 ;
447 ; X64-BMI2-LABEL: bzhi64_a0_masked:
448 ; X64-BMI2: # %bb.0:
449 ; X64-BMI2-NEXT: andb $63, %sil
450 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
451 ; X64-BMI2-NEXT: retq
452 %numlowbits.masked = and i64 %numlowbits, 63
453 %onebit = shl i64 1, %numlowbits.masked
454 %mask = add nsw i64 %onebit, -1
455 %masked = and i64 %mask, %val
456 ret i64 %masked
457 }
459 define i64 @bzhi64_a1_indexzext(i64 %val, i8 zeroext %numlowbits) nounwind {
460 ; X86-NOBMI-LABEL: bzhi64_a1_indexzext:
461 ; X86-NOBMI: # %bb.0:
462 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
463 ; X86-NOBMI-NEXT: movl $1, %eax
464 ; X86-NOBMI-NEXT: xorl %edx, %edx
465 ; X86-NOBMI-NEXT: shldl %cl, %eax, %edx
466 ; X86-NOBMI-NEXT: shll %cl, %eax
467 ; X86-NOBMI-NEXT: testb $32, %cl
468 ; X86-NOBMI-NEXT: je .LBB7_2
469 ; X86-NOBMI-NEXT: # %bb.1:
470 ; X86-NOBMI-NEXT: movl %eax, %edx
471 ; X86-NOBMI-NEXT: xorl %eax, %eax
472 ; X86-NOBMI-NEXT: .LBB7_2:
473 ; X86-NOBMI-NEXT: addl $-1, %eax
474 ; X86-NOBMI-NEXT: adcl $-1, %edx
475 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
476 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
477 ; X86-NOBMI-NEXT: retl
478 ;
479 ; X86-BMI1-LABEL: bzhi64_a1_indexzext:
480 ; X86-BMI1: # %bb.0:
481 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
482 ; X86-BMI1-NEXT: movl $1, %eax
483 ; X86-BMI1-NEXT: xorl %edx, %edx
484 ; X86-BMI1-NEXT: shldl %cl, %eax, %edx
485 ; X86-BMI1-NEXT: shll %cl, %eax
486 ; X86-BMI1-NEXT: testb $32, %cl
487 ; X86-BMI1-NEXT: je .LBB7_2
488 ; X86-BMI1-NEXT: # %bb.1:
489 ; X86-BMI1-NEXT: movl %eax, %edx
490 ; X86-BMI1-NEXT: xorl %eax, %eax
491 ; X86-BMI1-NEXT: .LBB7_2:
492 ; X86-BMI1-NEXT: addl $-1, %eax
493 ; X86-BMI1-NEXT: adcl $-1, %edx
494 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %edx
495 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
496 ; X86-BMI1-NEXT: retl
497 ;
498 ; X86-BMI2-LABEL: bzhi64_a1_indexzext:
499 ; X86-BMI2: # %bb.0:
500 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
501 ; X86-BMI2-NEXT: movl $1, %eax
502 ; X86-BMI2-NEXT: xorl %edx, %edx
503 ; X86-BMI2-NEXT: shldl %cl, %eax, %edx
504 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %eax
505 ; X86-BMI2-NEXT: testb $32, %cl
506 ; X86-BMI2-NEXT: je .LBB7_2
507 ; X86-BMI2-NEXT: # %bb.1:
508 ; X86-BMI2-NEXT: movl %eax, %edx
509 ; X86-BMI2-NEXT: xorl %eax, %eax
510 ; X86-BMI2-NEXT: .LBB7_2:
511 ; X86-BMI2-NEXT: addl $-1, %eax
512 ; X86-BMI2-NEXT: adcl $-1, %edx
513 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx
514 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
515 ; X86-BMI2-NEXT: retl
516 ;
517 ; X64-NOBMI-LABEL: bzhi64_a1_indexzext:
518 ; X64-NOBMI: # %bb.0:
519 ; X64-NOBMI-NEXT: movl %esi, %ecx
520 ; X64-NOBMI-NEXT: movl $1, %eax
521 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
522 ; X64-NOBMI-NEXT: shlq %cl, %rax
523 ; X64-NOBMI-NEXT: decq %rax
524 ; X64-NOBMI-NEXT: andq %rdi, %rax
525 ; X64-NOBMI-NEXT: retq
526 ;
527 ; X64-BMI1-LABEL: bzhi64_a1_indexzext:
528 ; X64-BMI1: # %bb.0:
529 ; X64-BMI1-NEXT: # kill: def $esi killed $esi def $rsi
530 ; X64-BMI1-NEXT: shll $8, %esi
531 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
532 ; X64-BMI1-NEXT: retq
533 ;
534 ; X64-BMI2-LABEL: bzhi64_a1_indexzext:
535 ; X64-BMI2: # %bb.0:
536 ; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi
537 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
538 ; X64-BMI2-NEXT: retq
539 %conv = zext i8 %numlowbits to i64
540 %onebit = shl i64 1, %conv
541 %mask = add nsw i64 %onebit, -1
542 %masked = and i64 %mask, %val
543 ret i64 %masked
544 }
546 define i64 @bzhi64_a2_load(ptr %w, i64 %numlowbits) nounwind {
547 ; X86-NOBMI-LABEL: bzhi64_a2_load:
548 ; X86-NOBMI: # %bb.0:
549 ; X86-NOBMI-NEXT: pushl %esi
550 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
551 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
552 ; X86-NOBMI-NEXT: movl $1, %eax
553 ; X86-NOBMI-NEXT: xorl %edx, %edx
554 ; X86-NOBMI-NEXT: shldl %cl, %eax, %edx
555 ; X86-NOBMI-NEXT: shll %cl, %eax
556 ; X86-NOBMI-NEXT: testb $32, %cl
557 ; X86-NOBMI-NEXT: je .LBB8_2
558 ; X86-NOBMI-NEXT: # %bb.1:
559 ; X86-NOBMI-NEXT: movl %eax, %edx
560 ; X86-NOBMI-NEXT: xorl %eax, %eax
561 ; X86-NOBMI-NEXT: .LBB8_2:
562 ; X86-NOBMI-NEXT: addl $-1, %eax
563 ; X86-NOBMI-NEXT: adcl $-1, %edx
564 ; X86-NOBMI-NEXT: andl 4(%esi), %edx
565 ; X86-NOBMI-NEXT: andl (%esi), %eax
566 ; X86-NOBMI-NEXT: popl %esi
567 ; X86-NOBMI-NEXT: retl
568 ;
569 ; X86-BMI1-LABEL: bzhi64_a2_load:
570 ; X86-BMI1: # %bb.0:
571 ; X86-BMI1-NEXT: pushl %esi
572 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %esi
573 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
574 ; X86-BMI1-NEXT: movl $1, %eax
575 ; X86-BMI1-NEXT: xorl %edx, %edx
576 ; X86-BMI1-NEXT: shldl %cl, %eax, %edx
577 ; X86-BMI1-NEXT: shll %cl, %eax
578 ; X86-BMI1-NEXT: testb $32, %cl
579 ; X86-BMI1-NEXT: je .LBB8_2
580 ; X86-BMI1-NEXT: # %bb.1:
581 ; X86-BMI1-NEXT: movl %eax, %edx
582 ; X86-BMI1-NEXT: xorl %eax, %eax
583 ; X86-BMI1-NEXT: .LBB8_2:
584 ; X86-BMI1-NEXT: addl $-1, %eax
585 ; X86-BMI1-NEXT: adcl $-1, %edx
586 ; X86-BMI1-NEXT: andl 4(%esi), %edx
587 ; X86-BMI1-NEXT: andl (%esi), %eax
588 ; X86-BMI1-NEXT: popl %esi
589 ; X86-BMI1-NEXT: retl
590 ;
591 ; X86-BMI2-LABEL: bzhi64_a2_load:
592 ; X86-BMI2: # %bb.0:
593 ; X86-BMI2-NEXT: pushl %esi
594 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %esi
595 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
596 ; X86-BMI2-NEXT: movl $1, %eax
597 ; X86-BMI2-NEXT: xorl %edx, %edx
598 ; X86-BMI2-NEXT: shldl %cl, %eax, %edx
599 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %eax
600 ; X86-BMI2-NEXT: testb $32, %cl
601 ; X86-BMI2-NEXT: je .LBB8_2
602 ; X86-BMI2-NEXT: # %bb.1:
603 ; X86-BMI2-NEXT: movl %eax, %edx
604 ; X86-BMI2-NEXT: xorl %eax, %eax
605 ; X86-BMI2-NEXT: .LBB8_2:
606 ; X86-BMI2-NEXT: addl $-1, %eax
607 ; X86-BMI2-NEXT: adcl $-1, %edx
608 ; X86-BMI2-NEXT: andl 4(%esi), %edx
609 ; X86-BMI2-NEXT: andl (%esi), %eax
610 ; X86-BMI2-NEXT: popl %esi
611 ; X86-BMI2-NEXT: retl
612 ;
613 ; X64-NOBMI-LABEL: bzhi64_a2_load:
614 ; X64-NOBMI: # %bb.0:
615 ; X64-NOBMI-NEXT: movq %rsi, %rcx
616 ; X64-NOBMI-NEXT: movl $1, %eax
617 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
618 ; X64-NOBMI-NEXT: shlq %cl, %rax
619 ; X64-NOBMI-NEXT: decq %rax
620 ; X64-NOBMI-NEXT: andq (%rdi), %rax
621 ; X64-NOBMI-NEXT: retq
622 ;
623 ; X64-BMI1-LABEL: bzhi64_a2_load:
624 ; X64-BMI1: # %bb.0:
625 ; X64-BMI1-NEXT: shll $8, %esi
626 ; X64-BMI1-NEXT: bextrq %rsi, (%rdi), %rax
627 ; X64-BMI1-NEXT: retq
628 ;
629 ; X64-BMI2-LABEL: bzhi64_a2_load:
630 ; X64-BMI2: # %bb.0:
631 ; X64-BMI2-NEXT: bzhiq %rsi, (%rdi), %rax
632 ; X64-BMI2-NEXT: retq
633 %val = load i64, ptr %w
634 %onebit = shl i64 1, %numlowbits
635 %mask = add nsw i64 %onebit, -1
636 %masked = and i64 %mask, %val
637 ret i64 %masked
638 }
640 define i64 @bzhi64_a3_load_indexzext(ptr %w, i8 zeroext %numlowbits) nounwind {
641 ; X86-NOBMI-LABEL: bzhi64_a3_load_indexzext:
642 ; X86-NOBMI: # %bb.0:
643 ; X86-NOBMI-NEXT: pushl %esi
644 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
645 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
646 ; X86-NOBMI-NEXT: movl $1, %eax
647 ; X86-NOBMI-NEXT: xorl %edx, %edx
648 ; X86-NOBMI-NEXT: shldl %cl, %eax, %edx
649 ; X86-NOBMI-NEXT: shll %cl, %eax
650 ; X86-NOBMI-NEXT: testb $32, %cl
651 ; X86-NOBMI-NEXT: je .LBB9_2
652 ; X86-NOBMI-NEXT: # %bb.1:
653 ; X86-NOBMI-NEXT: movl %eax, %edx
654 ; X86-NOBMI-NEXT: xorl %eax, %eax
655 ; X86-NOBMI-NEXT: .LBB9_2:
656 ; X86-NOBMI-NEXT: addl $-1, %eax
657 ; X86-NOBMI-NEXT: adcl $-1, %edx
658 ; X86-NOBMI-NEXT: andl 4(%esi), %edx
659 ; X86-NOBMI-NEXT: andl (%esi), %eax
660 ; X86-NOBMI-NEXT: popl %esi
661 ; X86-NOBMI-NEXT: retl
662 ;
663 ; X86-BMI1-LABEL: bzhi64_a3_load_indexzext:
664 ; X86-BMI1: # %bb.0:
665 ; X86-BMI1-NEXT: pushl %esi
666 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %esi
667 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
668 ; X86-BMI1-NEXT: movl $1, %eax
669 ; X86-BMI1-NEXT: xorl %edx, %edx
670 ; X86-BMI1-NEXT: shldl %cl, %eax, %edx
671 ; X86-BMI1-NEXT: shll %cl, %eax
672 ; X86-BMI1-NEXT: testb $32, %cl
673 ; X86-BMI1-NEXT: je .LBB9_2
674 ; X86-BMI1-NEXT: # %bb.1:
675 ; X86-BMI1-NEXT: movl %eax, %edx
676 ; X86-BMI1-NEXT: xorl %eax, %eax
677 ; X86-BMI1-NEXT: .LBB9_2:
678 ; X86-BMI1-NEXT: addl $-1, %eax
679 ; X86-BMI1-NEXT: adcl $-1, %edx
680 ; X86-BMI1-NEXT: andl 4(%esi), %edx
681 ; X86-BMI1-NEXT: andl (%esi), %eax
682 ; X86-BMI1-NEXT: popl %esi
683 ; X86-BMI1-NEXT: retl
684 ;
685 ; X86-BMI2-LABEL: bzhi64_a3_load_indexzext:
686 ; X86-BMI2: # %bb.0:
687 ; X86-BMI2-NEXT: pushl %esi
688 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %esi
689 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
690 ; X86-BMI2-NEXT: movl $1, %eax
691 ; X86-BMI2-NEXT: xorl %edx, %edx
692 ; X86-BMI2-NEXT: shldl %cl, %eax, %edx
693 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %eax
694 ; X86-BMI2-NEXT: testb $32, %cl
695 ; X86-BMI2-NEXT: je .LBB9_2
696 ; X86-BMI2-NEXT: # %bb.1:
697 ; X86-BMI2-NEXT: movl %eax, %edx
698 ; X86-BMI2-NEXT: xorl %eax, %eax
699 ; X86-BMI2-NEXT: .LBB9_2:
700 ; X86-BMI2-NEXT: addl $-1, %eax
701 ; X86-BMI2-NEXT: adcl $-1, %edx
702 ; X86-BMI2-NEXT: andl 4(%esi), %edx
703 ; X86-BMI2-NEXT: andl (%esi), %eax
704 ; X86-BMI2-NEXT: popl %esi
705 ; X86-BMI2-NEXT: retl
706 ;
707 ; X64-NOBMI-LABEL: bzhi64_a3_load_indexzext:
708 ; X64-NOBMI: # %bb.0:
709 ; X64-NOBMI-NEXT: movl %esi, %ecx
710 ; X64-NOBMI-NEXT: movl $1, %eax
711 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
712 ; X64-NOBMI-NEXT: shlq %cl, %rax
713 ; X64-NOBMI-NEXT: decq %rax
714 ; X64-NOBMI-NEXT: andq (%rdi), %rax
715 ; X64-NOBMI-NEXT: retq
716 ;
717 ; X64-BMI1-LABEL: bzhi64_a3_load_indexzext:
718 ; X64-BMI1: # %bb.0:
719 ; X64-BMI1-NEXT: # kill: def $esi killed $esi def $rsi
720 ; X64-BMI1-NEXT: shll $8, %esi
721 ; X64-BMI1-NEXT: bextrq %rsi, (%rdi), %rax
722 ; X64-BMI1-NEXT: retq
723 ;
724 ; X64-BMI2-LABEL: bzhi64_a3_load_indexzext:
725 ; X64-BMI2: # %bb.0:
726 ; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi
727 ; X64-BMI2-NEXT: bzhiq %rsi, (%rdi), %rax
728 ; X64-BMI2-NEXT: retq
729 %val = load i64, ptr %w
730 %conv = zext i8 %numlowbits to i64
731 %onebit = shl i64 1, %conv
732 %mask = add nsw i64 %onebit, -1
733 %masked = and i64 %mask, %val
734 ret i64 %masked
735 }
737 define i64 @bzhi64_a4_commutative(i64 %val, i64 %numlowbits) nounwind {
738 ; X86-NOBMI-LABEL: bzhi64_a4_commutative:
739 ; X86-NOBMI: # %bb.0:
740 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
741 ; X86-NOBMI-NEXT: movl $1, %eax
742 ; X86-NOBMI-NEXT: xorl %edx, %edx
743 ; X86-NOBMI-NEXT: shldl %cl, %eax, %edx
744 ; X86-NOBMI-NEXT: shll %cl, %eax
745 ; X86-NOBMI-NEXT: testb $32, %cl
746 ; X86-NOBMI-NEXT: je .LBB10_2
747 ; X86-NOBMI-NEXT: # %bb.1:
748 ; X86-NOBMI-NEXT: movl %eax, %edx
749 ; X86-NOBMI-NEXT: xorl %eax, %eax
750 ; X86-NOBMI-NEXT: .LBB10_2:
751 ; X86-NOBMI-NEXT: addl $-1, %eax
752 ; X86-NOBMI-NEXT: adcl $-1, %edx
753 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
754 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
755 ; X86-NOBMI-NEXT: retl
756 ;
757 ; X86-BMI1-LABEL: bzhi64_a4_commutative:
758 ; X86-BMI1: # %bb.0:
759 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
760 ; X86-BMI1-NEXT: movl $1, %eax
761 ; X86-BMI1-NEXT: xorl %edx, %edx
762 ; X86-BMI1-NEXT: shldl %cl, %eax, %edx
763 ; X86-BMI1-NEXT: shll %cl, %eax
764 ; X86-BMI1-NEXT: testb $32, %cl
765 ; X86-BMI1-NEXT: je .LBB10_2
766 ; X86-BMI1-NEXT: # %bb.1:
767 ; X86-BMI1-NEXT: movl %eax, %edx
768 ; X86-BMI1-NEXT: xorl %eax, %eax
769 ; X86-BMI1-NEXT: .LBB10_2:
770 ; X86-BMI1-NEXT: addl $-1, %eax
771 ; X86-BMI1-NEXT: adcl $-1, %edx
772 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %edx
773 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
774 ; X86-BMI1-NEXT: retl
775 ;
776 ; X86-BMI2-LABEL: bzhi64_a4_commutative:
777 ; X86-BMI2: # %bb.0:
778 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
779 ; X86-BMI2-NEXT: movl $1, %eax
780 ; X86-BMI2-NEXT: xorl %edx, %edx
781 ; X86-BMI2-NEXT: shldl %cl, %eax, %edx
782 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %eax
783 ; X86-BMI2-NEXT: testb $32, %cl
784 ; X86-BMI2-NEXT: je .LBB10_2
785 ; X86-BMI2-NEXT: # %bb.1:
786 ; X86-BMI2-NEXT: movl %eax, %edx
787 ; X86-BMI2-NEXT: xorl %eax, %eax
788 ; X86-BMI2-NEXT: .LBB10_2:
789 ; X86-BMI2-NEXT: addl $-1, %eax
790 ; X86-BMI2-NEXT: adcl $-1, %edx
791 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx
792 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
793 ; X86-BMI2-NEXT: retl
794 ;
795 ; X64-NOBMI-LABEL: bzhi64_a4_commutative:
796 ; X64-NOBMI: # %bb.0:
797 ; X64-NOBMI-NEXT: movq %rsi, %rcx
798 ; X64-NOBMI-NEXT: movl $1, %eax
799 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
800 ; X64-NOBMI-NEXT: shlq %cl, %rax
801 ; X64-NOBMI-NEXT: decq %rax
802 ; X64-NOBMI-NEXT: andq %rdi, %rax
803 ; X64-NOBMI-NEXT: retq
804 ;
805 ; X64-BMI1-LABEL: bzhi64_a4_commutative:
806 ; X64-BMI1: # %bb.0:
807 ; X64-BMI1-NEXT: shll $8, %esi
808 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
809 ; X64-BMI1-NEXT: retq
810 ;
811 ; X64-BMI2-LABEL: bzhi64_a4_commutative:
812 ; X64-BMI2: # %bb.0:
813 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
814 ; X64-BMI2-NEXT: retq
815 %onebit = shl i64 1, %numlowbits
816 %mask = add nsw i64 %onebit, -1
817 %masked = and i64 %val, %mask ; swapped order
818 ret i64 %masked
819 }
821 ; 64-bit, but with 32-bit output
823 ; Everything done in 64-bit, truncation happens last.
824 define i32 @bzhi64_32_a0(i64 %val, i64 %numlowbits) nounwind {
825 ; X86-NOBMI-LABEL: bzhi64_32_a0:
826 ; X86-NOBMI: # %bb.0:
827 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
828 ; X86-NOBMI-NEXT: movl $1, %edx
829 ; X86-NOBMI-NEXT: shll %cl, %edx
830 ; X86-NOBMI-NEXT: xorl %eax, %eax
831 ; X86-NOBMI-NEXT: testb $32, %cl
832 ; X86-NOBMI-NEXT: jne .LBB11_2
833 ; X86-NOBMI-NEXT: # %bb.1:
834 ; X86-NOBMI-NEXT: movl %edx, %eax
835 ; X86-NOBMI-NEXT: .LBB11_2:
836 ; X86-NOBMI-NEXT: decl %eax
837 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
838 ; X86-NOBMI-NEXT: retl
839 ;
840 ; X86-BMI1-LABEL: bzhi64_32_a0:
841 ; X86-BMI1: # %bb.0:
842 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
843 ; X86-BMI1-NEXT: movl $1, %edx
844 ; X86-BMI1-NEXT: shll %cl, %edx
845 ; X86-BMI1-NEXT: xorl %eax, %eax
846 ; X86-BMI1-NEXT: testb $32, %cl
847 ; X86-BMI1-NEXT: jne .LBB11_2
848 ; X86-BMI1-NEXT: # %bb.1:
849 ; X86-BMI1-NEXT: movl %edx, %eax
850 ; X86-BMI1-NEXT: .LBB11_2:
851 ; X86-BMI1-NEXT: decl %eax
852 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
853 ; X86-BMI1-NEXT: retl
854 ;
855 ; X86-BMI2-LABEL: bzhi64_32_a0:
856 ; X86-BMI2: # %bb.0:
857 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
858 ; X86-BMI2-NEXT: xorl %eax, %eax
859 ; X86-BMI2-NEXT: testb $32, %cl
860 ; X86-BMI2-NEXT: jne .LBB11_2
861 ; X86-BMI2-NEXT: # %bb.1:
862 ; X86-BMI2-NEXT: movl $1, %eax
863 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %eax
864 ; X86-BMI2-NEXT: .LBB11_2:
865 ; X86-BMI2-NEXT: decl %eax
866 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
867 ; X86-BMI2-NEXT: retl
868 ;
869 ; X64-NOBMI-LABEL: bzhi64_32_a0:
870 ; X64-NOBMI: # %bb.0:
871 ; X64-NOBMI-NEXT: movq %rsi, %rcx
872 ; X64-NOBMI-NEXT: movl $1, %eax
873 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
874 ; X64-NOBMI-NEXT: shlq %cl, %rax
875 ; X64-NOBMI-NEXT: decl %eax
876 ; X64-NOBMI-NEXT: andl %edi, %eax
877 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
878 ; X64-NOBMI-NEXT: retq
879 ;
880 ; X64-BMI1-LABEL: bzhi64_32_a0:
881 ; X64-BMI1: # %bb.0:
882 ; X64-BMI1-NEXT: shll $8, %esi
883 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
884 ; X64-BMI1-NEXT: retq
885 ;
886 ; X64-BMI2-LABEL: bzhi64_32_a0:
887 ; X64-BMI2: # %bb.0:
888 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
889 ; X64-BMI2-NEXT: retq
890 %onebit = shl i64 1, %numlowbits
891 %mask = add nsw i64 %onebit, -1
892 %masked = and i64 %mask, %val
893 %res = trunc i64 %masked to i32
894 ret i32 %res
895 }
897 ; Shifting happens in 64-bit, then truncation. Masking is 32-bit.
898 define i32 @bzhi64_32_a1(i64 %val, i32 %numlowbits) nounwind {
899 ; X86-NOBMI-LABEL: bzhi64_32_a1:
900 ; X86-NOBMI: # %bb.0:
901 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
902 ; X86-NOBMI-NEXT: movl $1, %eax
903 ; X86-NOBMI-NEXT: shll %cl, %eax
904 ; X86-NOBMI-NEXT: decl %eax
905 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
906 ; X86-NOBMI-NEXT: retl
907 ;
908 ; X86-BMI1-LABEL: bzhi64_32_a1:
909 ; X86-BMI1: # %bb.0:
910 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
911 ; X86-BMI1-NEXT: shll $8, %eax
912 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
913 ; X86-BMI1-NEXT: retl
914 ;
915 ; X86-BMI2-LABEL: bzhi64_32_a1:
916 ; X86-BMI2: # %bb.0:
917 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
918 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
919 ; X86-BMI2-NEXT: retl
920 ;
921 ; X64-NOBMI-LABEL: bzhi64_32_a1:
922 ; X64-NOBMI: # %bb.0:
923 ; X64-NOBMI-NEXT: movl %esi, %ecx
924 ; X64-NOBMI-NEXT: movl $1, %eax
925 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
926 ; X64-NOBMI-NEXT: shll %cl, %eax
927 ; X64-NOBMI-NEXT: decl %eax
928 ; X64-NOBMI-NEXT: andl %edi, %eax
929 ; X64-NOBMI-NEXT: retq
930 ;
931 ; X64-BMI1-LABEL: bzhi64_32_a1:
932 ; X64-BMI1: # %bb.0:
933 ; X64-BMI1-NEXT: shll $8, %esi
934 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
935 ; X64-BMI1-NEXT: retq
936 ;
937 ; X64-BMI2-LABEL: bzhi64_32_a1:
938 ; X64-BMI2: # %bb.0:
939 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
940 ; X64-BMI2-NEXT: retq
941 %truncval = trunc i64 %val to i32
942 %onebit = shl i32 1, %numlowbits
943 %mask = add nsw i32 %onebit, -1
944 %masked = and i32 %mask, %truncval
945 ret i32 %masked
946 }
948 ; Shifting happens in 64-bit, then truncation (with extra use).
949 ; Masking is 32-bit.
950 define i32 @bzhi64_32_a1_trunc_extrause(i64 %val, i32 %numlowbits, ptr %escape) nounwind {
951 ; X86-NOBMI-LABEL: bzhi64_32_a1_trunc_extrause:
952 ; X86-NOBMI: # %bb.0:
953 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
954 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
955 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
956 ; X86-NOBMI-NEXT: movl %edx, (%eax)
957 ; X86-NOBMI-NEXT: movl $1, %eax
958 ; X86-NOBMI-NEXT: shll %cl, %eax
959 ; X86-NOBMI-NEXT: decl %eax
960 ; X86-NOBMI-NEXT: andl %edx, %eax
961 ; X86-NOBMI-NEXT: retl
962 ;
963 ; X86-BMI1-LABEL: bzhi64_32_a1_trunc_extrause:
964 ; X86-BMI1: # %bb.0:
965 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
966 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %ecx
967 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
968 ; X86-BMI1-NEXT: movl %ecx, (%edx)
969 ; X86-BMI1-NEXT: shll $8, %eax
970 ; X86-BMI1-NEXT: bextrl %eax, %ecx, %eax
971 ; X86-BMI1-NEXT: retl
972 ;
973 ; X86-BMI2-LABEL: bzhi64_32_a1_trunc_extrause:
974 ; X86-BMI2: # %bb.0:
975 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
976 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
977 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %edx
978 ; X86-BMI2-NEXT: movl %ecx, (%edx)
979 ; X86-BMI2-NEXT: bzhil %eax, %ecx, %eax
980 ; X86-BMI2-NEXT: retl
981 ;
982 ; X64-NOBMI-LABEL: bzhi64_32_a1_trunc_extrause:
983 ; X64-NOBMI: # %bb.0:
984 ; X64-NOBMI-NEXT: movl %esi, %ecx
985 ; X64-NOBMI-NEXT: movl %edi, (%rdx)
986 ; X64-NOBMI-NEXT: movl $1, %eax
987 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
988 ; X64-NOBMI-NEXT: shll %cl, %eax
989 ; X64-NOBMI-NEXT: decl %eax
990 ; X64-NOBMI-NEXT: andl %edi, %eax
991 ; X64-NOBMI-NEXT: retq
992 ;
993 ; X64-BMI1-LABEL: bzhi64_32_a1_trunc_extrause:
994 ; X64-BMI1: # %bb.0:
995 ; X64-BMI1-NEXT: movl %edi, (%rdx)
996 ; X64-BMI1-NEXT: shll $8, %esi
997 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
998 ; X64-BMI1-NEXT: retq
999 ;
1000 ; X64-BMI2-LABEL: bzhi64_32_a1_trunc_extrause:
1001 ; X64-BMI2: # %bb.0:
1002 ; X64-BMI2-NEXT: movl %edi, (%rdx)
1003 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
1004 ; X64-BMI2-NEXT: retq
1005 %truncval = trunc i64 %val to i32
1006 store i32 %truncval, ptr %escape
1007 %onebit = shl i32 1, %numlowbits
1008 %mask = add nsw i32 %onebit, -1
1009 %masked = and i32 %mask, %truncval
1010 ret i32 %masked
1011 }
1013 ; Shifting happens in 64-bit. Mask is 32-bit, but extended to 64-bit.
1014 ; Masking is 64-bit. Then truncation.
1015 define i32 @bzhi64_32_a2(i64 %val, i32 %numlowbits) nounwind {
1016 ; X86-NOBMI-LABEL: bzhi64_32_a2:
1017 ; X86-NOBMI: # %bb.0:
1018 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1019 ; X86-NOBMI-NEXT: movl $1, %eax
1020 ; X86-NOBMI-NEXT: shll %cl, %eax
1021 ; X86-NOBMI-NEXT: decl %eax
1022 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1023 ; X86-NOBMI-NEXT: retl
1024 ;
1025 ; X86-BMI1-LABEL: bzhi64_32_a2:
1026 ; X86-BMI1: # %bb.0:
1027 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1028 ; X86-BMI1-NEXT: shll $8, %eax
1029 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
1030 ; X86-BMI1-NEXT: retl
1031 ;
1032 ; X86-BMI2-LABEL: bzhi64_32_a2:
1033 ; X86-BMI2: # %bb.0:
1034 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1035 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
1036 ; X86-BMI2-NEXT: retl
1037 ;
1038 ; X64-NOBMI-LABEL: bzhi64_32_a2:
1039 ; X64-NOBMI: # %bb.0:
1040 ; X64-NOBMI-NEXT: movl %esi, %ecx
1041 ; X64-NOBMI-NEXT: movl $1, %eax
1042 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1043 ; X64-NOBMI-NEXT: shll %cl, %eax
1044 ; X64-NOBMI-NEXT: decl %eax
1045 ; X64-NOBMI-NEXT: andl %edi, %eax
1046 ; X64-NOBMI-NEXT: retq
1047 ;
1048 ; X64-BMI1-LABEL: bzhi64_32_a2:
1049 ; X64-BMI1: # %bb.0:
1050 ; X64-BMI1-NEXT: shll $8, %esi
1051 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
1052 ; X64-BMI1-NEXT: retq
1053 ;
1054 ; X64-BMI2-LABEL: bzhi64_32_a2:
1055 ; X64-BMI2: # %bb.0:
1056 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
1057 ; X64-BMI2-NEXT: retq
1058 %onebit = shl i32 1, %numlowbits
1059 %mask = add nsw i32 %onebit, -1
1060 %zextmask = zext i32 %mask to i64
1061 %masked = and i64 %zextmask, %val
1062 %truncmasked = trunc i64 %masked to i32
1063 ret i32 %truncmasked
1064 }
1066 ; Shifting happens in 64-bit. Mask is 32-bit, but calculated in 64-bit.
1067 ; Masking is 64-bit. Then truncation.
1068 define i32 @bzhi64_32_a3(i64 %val, i64 %numlowbits) nounwind {
1069 ; X86-NOBMI-LABEL: bzhi64_32_a3:
1070 ; X86-NOBMI: # %bb.0:
1071 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1072 ; X86-NOBMI-NEXT: movl $1, %edx
1073 ; X86-NOBMI-NEXT: shll %cl, %edx
1074 ; X86-NOBMI-NEXT: xorl %eax, %eax
1075 ; X86-NOBMI-NEXT: testb $32, %cl
1076 ; X86-NOBMI-NEXT: jne .LBB15_2
1077 ; X86-NOBMI-NEXT: # %bb.1:
1078 ; X86-NOBMI-NEXT: movl %edx, %eax
1079 ; X86-NOBMI-NEXT: .LBB15_2:
1080 ; X86-NOBMI-NEXT: decl %eax
1081 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1082 ; X86-NOBMI-NEXT: retl
1083 ;
1084 ; X86-BMI1-LABEL: bzhi64_32_a3:
1085 ; X86-BMI1: # %bb.0:
1086 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1087 ; X86-BMI1-NEXT: movl $1, %edx
1088 ; X86-BMI1-NEXT: shll %cl, %edx
1089 ; X86-BMI1-NEXT: xorl %eax, %eax
1090 ; X86-BMI1-NEXT: testb $32, %cl
1091 ; X86-BMI1-NEXT: jne .LBB15_2
1092 ; X86-BMI1-NEXT: # %bb.1:
1093 ; X86-BMI1-NEXT: movl %edx, %eax
1094 ; X86-BMI1-NEXT: .LBB15_2:
1095 ; X86-BMI1-NEXT: decl %eax
1096 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
1097 ; X86-BMI1-NEXT: retl
1098 ;
1099 ; X86-BMI2-LABEL: bzhi64_32_a3:
1100 ; X86-BMI2: # %bb.0:
1101 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1102 ; X86-BMI2-NEXT: xorl %eax, %eax
1103 ; X86-BMI2-NEXT: testb $32, %cl
1104 ; X86-BMI2-NEXT: jne .LBB15_2
1105 ; X86-BMI2-NEXT: # %bb.1:
1106 ; X86-BMI2-NEXT: movl $1, %eax
1107 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %eax
1108 ; X86-BMI2-NEXT: .LBB15_2:
1109 ; X86-BMI2-NEXT: decl %eax
1110 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
1111 ; X86-BMI2-NEXT: retl
1112 ;
1113 ; X64-NOBMI-LABEL: bzhi64_32_a3:
1114 ; X64-NOBMI: # %bb.0:
1115 ; X64-NOBMI-NEXT: movq %rsi, %rcx
1116 ; X64-NOBMI-NEXT: movl $1, %eax
1117 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
1118 ; X64-NOBMI-NEXT: shlq %cl, %rax
1119 ; X64-NOBMI-NEXT: decl %eax
1120 ; X64-NOBMI-NEXT: andl %edi, %eax
1121 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
1122 ; X64-NOBMI-NEXT: retq
1123 ;
1124 ; X64-BMI1-LABEL: bzhi64_32_a3:
1125 ; X64-BMI1: # %bb.0:
1126 ; X64-BMI1-NEXT: shll $8, %esi
1127 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
1128 ; X64-BMI1-NEXT: retq
1129 ;
1130 ; X64-BMI2-LABEL: bzhi64_32_a3:
1131 ; X64-BMI2: # %bb.0:
1132 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
1133 ; X64-BMI2-NEXT: retq
1134 %onebit = shl i64 1, %numlowbits
1135 %mask = add nsw i64 %onebit, 4294967295
1136 %masked = and i64 %mask, %val
1137 %truncmasked = trunc i64 %masked to i32
1138 ret i32 %truncmasked
1139 }
1141 ; ---------------------------------------------------------------------------- ;
1142 ; Pattern b. 32-bit
1143 ; ---------------------------------------------------------------------------- ;
1145 define i32 @bzhi32_b0(i32 %val, i32 %numlowbits) nounwind {
1146 ; X86-NOBMI-LABEL: bzhi32_b0:
1147 ; X86-NOBMI: # %bb.0:
1148 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1149 ; X86-NOBMI-NEXT: movl $-1, %eax
1150 ; X86-NOBMI-NEXT: shll %cl, %eax
1151 ; X86-NOBMI-NEXT: notl %eax
1152 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1153 ; X86-NOBMI-NEXT: retl
1154 ;
1155 ; X86-BMI1-LABEL: bzhi32_b0:
1156 ; X86-BMI1: # %bb.0:
1157 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1158 ; X86-BMI1-NEXT: shll $8, %eax
1159 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
1160 ; X86-BMI1-NEXT: retl
1161 ;
1162 ; X86-BMI2-LABEL: bzhi32_b0:
1163 ; X86-BMI2: # %bb.0:
1164 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1165 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
1166 ; X86-BMI2-NEXT: retl
1167 ;
1168 ; X64-NOBMI-LABEL: bzhi32_b0:
1169 ; X64-NOBMI: # %bb.0:
1170 ; X64-NOBMI-NEXT: movl %esi, %ecx
1171 ; X64-NOBMI-NEXT: movl $-1, %eax
1172 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1173 ; X64-NOBMI-NEXT: shll %cl, %eax
1174 ; X64-NOBMI-NEXT: notl %eax
1175 ; X64-NOBMI-NEXT: andl %edi, %eax
1176 ; X64-NOBMI-NEXT: retq
1177 ;
1178 ; X64-BMI1-LABEL: bzhi32_b0:
1179 ; X64-BMI1: # %bb.0:
1180 ; X64-BMI1-NEXT: shll $8, %esi
1181 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
1182 ; X64-BMI1-NEXT: retq
1183 ;
1184 ; X64-BMI2-LABEL: bzhi32_b0:
1185 ; X64-BMI2: # %bb.0:
1186 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
1187 ; X64-BMI2-NEXT: retq
1188 %notmask = shl i32 -1, %numlowbits
1189 %mask = xor i32 %notmask, -1
1190 %masked = and i32 %mask, %val
1191 ret i32 %masked
1192 }
1194 define i32 @bzhi32_b1_indexzext(i32 %val, i8 zeroext %numlowbits) nounwind {
1195 ; X86-NOBMI-LABEL: bzhi32_b1_indexzext:
1196 ; X86-NOBMI: # %bb.0:
1197 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1198 ; X86-NOBMI-NEXT: movl $-1, %eax
1199 ; X86-NOBMI-NEXT: shll %cl, %eax
1200 ; X86-NOBMI-NEXT: notl %eax
1201 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1202 ; X86-NOBMI-NEXT: retl
1203 ;
1204 ; X86-BMI1-LABEL: bzhi32_b1_indexzext:
1205 ; X86-BMI1: # %bb.0:
1206 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1207 ; X86-BMI1-NEXT: shll $8, %eax
1208 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
1209 ; X86-BMI1-NEXT: retl
1210 ;
1211 ; X86-BMI2-LABEL: bzhi32_b1_indexzext:
1212 ; X86-BMI2: # %bb.0:
1213 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1214 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
1215 ; X86-BMI2-NEXT: retl
1216 ;
1217 ; X64-NOBMI-LABEL: bzhi32_b1_indexzext:
1218 ; X64-NOBMI: # %bb.0:
1219 ; X64-NOBMI-NEXT: movl %esi, %ecx
1220 ; X64-NOBMI-NEXT: movl $-1, %eax
1221 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1222 ; X64-NOBMI-NEXT: shll %cl, %eax
1223 ; X64-NOBMI-NEXT: notl %eax
1224 ; X64-NOBMI-NEXT: andl %edi, %eax
1225 ; X64-NOBMI-NEXT: retq
1226 ;
1227 ; X64-BMI1-LABEL: bzhi32_b1_indexzext:
1228 ; X64-BMI1: # %bb.0:
1229 ; X64-BMI1-NEXT: shll $8, %esi
1230 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
1231 ; X64-BMI1-NEXT: retq
1232 ;
1233 ; X64-BMI2-LABEL: bzhi32_b1_indexzext:
1234 ; X64-BMI2: # %bb.0:
1235 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
1236 ; X64-BMI2-NEXT: retq
1237 %conv = zext i8 %numlowbits to i32
1238 %notmask = shl i32 -1, %conv
1239 %mask = xor i32 %notmask, -1
1240 %masked = and i32 %mask, %val
1241 ret i32 %masked
1242 }
1244 define i32 @bzhi32_b2_load(ptr %w, i32 %numlowbits) nounwind {
1245 ; X86-NOBMI-LABEL: bzhi32_b2_load:
1246 ; X86-NOBMI: # %bb.0:
1247 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
1248 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1249 ; X86-NOBMI-NEXT: movl $-1, %eax
1250 ; X86-NOBMI-NEXT: shll %cl, %eax
1251 ; X86-NOBMI-NEXT: notl %eax
1252 ; X86-NOBMI-NEXT: andl (%edx), %eax
1253 ; X86-NOBMI-NEXT: retl
1254 ;
1255 ; X86-BMI1-LABEL: bzhi32_b2_load:
1256 ; X86-BMI1: # %bb.0:
1257 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
1258 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1259 ; X86-BMI1-NEXT: shll $8, %ecx
1260 ; X86-BMI1-NEXT: bextrl %ecx, (%eax), %eax
1261 ; X86-BMI1-NEXT: retl
1262 ;
1263 ; X86-BMI2-LABEL: bzhi32_b2_load:
1264 ; X86-BMI2: # %bb.0:
1265 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
1266 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1267 ; X86-BMI2-NEXT: bzhil %ecx, (%eax), %eax
1268 ; X86-BMI2-NEXT: retl
1269 ;
1270 ; X64-NOBMI-LABEL: bzhi32_b2_load:
1271 ; X64-NOBMI: # %bb.0:
1272 ; X64-NOBMI-NEXT: movl %esi, %ecx
1273 ; X64-NOBMI-NEXT: movl $-1, %eax
1274 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1275 ; X64-NOBMI-NEXT: shll %cl, %eax
1276 ; X64-NOBMI-NEXT: notl %eax
1277 ; X64-NOBMI-NEXT: andl (%rdi), %eax
1278 ; X64-NOBMI-NEXT: retq
1279 ;
1280 ; X64-BMI1-LABEL: bzhi32_b2_load:
1281 ; X64-BMI1: # %bb.0:
1282 ; X64-BMI1-NEXT: shll $8, %esi
1283 ; X64-BMI1-NEXT: bextrl %esi, (%rdi), %eax
1284 ; X64-BMI1-NEXT: retq
1285 ;
1286 ; X64-BMI2-LABEL: bzhi32_b2_load:
1287 ; X64-BMI2: # %bb.0:
1288 ; X64-BMI2-NEXT: bzhil %esi, (%rdi), %eax
1289 ; X64-BMI2-NEXT: retq
1290 %val = load i32, ptr %w
1291 %notmask = shl i32 -1, %numlowbits
1292 %mask = xor i32 %notmask, -1
1293 %masked = and i32 %mask, %val
1294 ret i32 %masked
1295 }
1297 define i32 @bzhi32_b3_load_indexzext(ptr %w, i8 zeroext %numlowbits) nounwind {
1298 ; X86-NOBMI-LABEL: bzhi32_b3_load_indexzext:
1299 ; X86-NOBMI: # %bb.0:
1300 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
1301 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1302 ; X86-NOBMI-NEXT: movl $-1, %eax
1303 ; X86-NOBMI-NEXT: shll %cl, %eax
1304 ; X86-NOBMI-NEXT: notl %eax
1305 ; X86-NOBMI-NEXT: andl (%edx), %eax
1306 ; X86-NOBMI-NEXT: retl
1307 ;
1308 ; X86-BMI1-LABEL: bzhi32_b3_load_indexzext:
1309 ; X86-BMI1: # %bb.0:
1310 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
1311 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1312 ; X86-BMI1-NEXT: shll $8, %ecx
1313 ; X86-BMI1-NEXT: bextrl %ecx, (%eax), %eax
1314 ; X86-BMI1-NEXT: retl
1315 ;
1316 ; X86-BMI2-LABEL: bzhi32_b3_load_indexzext:
1317 ; X86-BMI2: # %bb.0:
1318 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
1319 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1320 ; X86-BMI2-NEXT: bzhil %ecx, (%eax), %eax
1321 ; X86-BMI2-NEXT: retl
1322 ;
1323 ; X64-NOBMI-LABEL: bzhi32_b3_load_indexzext:
1324 ; X64-NOBMI: # %bb.0:
1325 ; X64-NOBMI-NEXT: movl %esi, %ecx
1326 ; X64-NOBMI-NEXT: movl $-1, %eax
1327 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1328 ; X64-NOBMI-NEXT: shll %cl, %eax
1329 ; X64-NOBMI-NEXT: notl %eax
1330 ; X64-NOBMI-NEXT: andl (%rdi), %eax
1331 ; X64-NOBMI-NEXT: retq
1332 ;
1333 ; X64-BMI1-LABEL: bzhi32_b3_load_indexzext:
1334 ; X64-BMI1: # %bb.0:
1335 ; X64-BMI1-NEXT: shll $8, %esi
1336 ; X64-BMI1-NEXT: bextrl %esi, (%rdi), %eax
1337 ; X64-BMI1-NEXT: retq
1338 ;
1339 ; X64-BMI2-LABEL: bzhi32_b3_load_indexzext:
1340 ; X64-BMI2: # %bb.0:
1341 ; X64-BMI2-NEXT: bzhil %esi, (%rdi), %eax
1342 ; X64-BMI2-NEXT: retq
1343 %val = load i32, ptr %w
1344 %conv = zext i8 %numlowbits to i32
1345 %notmask = shl i32 -1, %conv
1346 %mask = xor i32 %notmask, -1
1347 %masked = and i32 %mask, %val
1348 ret i32 %masked
1349 }
1351 define i32 @bzhi32_b4_commutative(i32 %val, i32 %numlowbits) nounwind {
1352 ; X86-NOBMI-LABEL: bzhi32_b4_commutative:
1353 ; X86-NOBMI: # %bb.0:
1354 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1355 ; X86-NOBMI-NEXT: movl $-1, %eax
1356 ; X86-NOBMI-NEXT: shll %cl, %eax
1357 ; X86-NOBMI-NEXT: notl %eax
1358 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1359 ; X86-NOBMI-NEXT: retl
1360 ;
1361 ; X86-BMI1-LABEL: bzhi32_b4_commutative:
1362 ; X86-BMI1: # %bb.0:
1363 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1364 ; X86-BMI1-NEXT: shll $8, %eax
1365 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
1366 ; X86-BMI1-NEXT: retl
1367 ;
1368 ; X86-BMI2-LABEL: bzhi32_b4_commutative:
1369 ; X86-BMI2: # %bb.0:
1370 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1371 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
1372 ; X86-BMI2-NEXT: retl
1373 ;
1374 ; X64-NOBMI-LABEL: bzhi32_b4_commutative:
1375 ; X64-NOBMI: # %bb.0:
1376 ; X64-NOBMI-NEXT: movl %esi, %ecx
1377 ; X64-NOBMI-NEXT: movl $-1, %eax
1378 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1379 ; X64-NOBMI-NEXT: shll %cl, %eax
1380 ; X64-NOBMI-NEXT: notl %eax
1381 ; X64-NOBMI-NEXT: andl %edi, %eax
1382 ; X64-NOBMI-NEXT: retq
1383 ;
1384 ; X64-BMI1-LABEL: bzhi32_b4_commutative:
1385 ; X64-BMI1: # %bb.0:
1386 ; X64-BMI1-NEXT: shll $8, %esi
1387 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
1388 ; X64-BMI1-NEXT: retq
1389 ;
1390 ; X64-BMI2-LABEL: bzhi32_b4_commutative:
1391 ; X64-BMI2: # %bb.0:
1392 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
1393 ; X64-BMI2-NEXT: retq
1394 %notmask = shl i32 -1, %numlowbits
1395 %mask = xor i32 %notmask, -1
1396 %masked = and i32 %val, %mask ; swapped order
1397 ret i32 %masked
1398 }
1400 ; 64-bit
1402 define i64 @bzhi64_b0(i64 %val, i64 %numlowbits) nounwind {
1403 ; X86-NOBMI-LABEL: bzhi64_b0:
1404 ; X86-NOBMI: # %bb.0:
1405 ; X86-NOBMI-NEXT: pushl %esi
1406 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1407 ; X86-NOBMI-NEXT: movl $-1, %edx
1408 ; X86-NOBMI-NEXT: movl $-1, %esi
1409 ; X86-NOBMI-NEXT: shll %cl, %esi
1410 ; X86-NOBMI-NEXT: xorl %eax, %eax
1411 ; X86-NOBMI-NEXT: testb $32, %cl
1412 ; X86-NOBMI-NEXT: jne .LBB21_1
1413 ; X86-NOBMI-NEXT: # %bb.2:
1414 ; X86-NOBMI-NEXT: movl %esi, %eax
1415 ; X86-NOBMI-NEXT: jmp .LBB21_3
1416 ; X86-NOBMI-NEXT: .LBB21_1:
1417 ; X86-NOBMI-NEXT: movl %esi, %edx
1418 ; X86-NOBMI-NEXT: .LBB21_3:
1419 ; X86-NOBMI-NEXT: notl %edx
1420 ; X86-NOBMI-NEXT: notl %eax
1421 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1422 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
1423 ; X86-NOBMI-NEXT: popl %esi
1424 ; X86-NOBMI-NEXT: retl
1425 ;
1426 ; X86-BMI1-LABEL: bzhi64_b0:
1427 ; X86-BMI1: # %bb.0:
1428 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1429 ; X86-BMI1-NEXT: movl $-1, %edx
1430 ; X86-BMI1-NEXT: movl $-1, %eax
1431 ; X86-BMI1-NEXT: shll %cl, %eax
1432 ; X86-BMI1-NEXT: testb $32, %cl
1433 ; X86-BMI1-NEXT: je .LBB21_2
1434 ; X86-BMI1-NEXT: # %bb.1:
1435 ; X86-BMI1-NEXT: movl %eax, %edx
1436 ; X86-BMI1-NEXT: xorl %eax, %eax
1437 ; X86-BMI1-NEXT: .LBB21_2:
1438 ; X86-BMI1-NEXT: andnl {{[0-9]+}}(%esp), %eax, %eax
1439 ; X86-BMI1-NEXT: andnl {{[0-9]+}}(%esp), %edx, %edx
1440 ; X86-BMI1-NEXT: retl
1441 ;
1442 ; X86-BMI2-LABEL: bzhi64_b0:
1443 ; X86-BMI2: # %bb.0:
1444 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %edx
1445 ; X86-BMI2-NEXT: movl $-1, %ecx
1446 ; X86-BMI2-NEXT: shlxl %edx, %ecx, %eax
1447 ; X86-BMI2-NEXT: testb $32, %dl
1448 ; X86-BMI2-NEXT: je .LBB21_2
1449 ; X86-BMI2-NEXT: # %bb.1:
1450 ; X86-BMI2-NEXT: movl %eax, %ecx
1451 ; X86-BMI2-NEXT: xorl %eax, %eax
1452 ; X86-BMI2-NEXT: .LBB21_2:
1453 ; X86-BMI2-NEXT: andnl {{[0-9]+}}(%esp), %eax, %eax
1454 ; X86-BMI2-NEXT: andnl {{[0-9]+}}(%esp), %ecx, %edx
1455 ; X86-BMI2-NEXT: retl
1456 ;
1457 ; X64-NOBMI-LABEL: bzhi64_b0:
1458 ; X64-NOBMI: # %bb.0:
1459 ; X64-NOBMI-NEXT: movq %rsi, %rcx
1460 ; X64-NOBMI-NEXT: movq $-1, %rax
1461 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
1462 ; X64-NOBMI-NEXT: shlq %cl, %rax
1463 ; X64-NOBMI-NEXT: notq %rax
1464 ; X64-NOBMI-NEXT: andq %rdi, %rax
1465 ; X64-NOBMI-NEXT: retq
1466 ;
1467 ; X64-BMI1-LABEL: bzhi64_b0:
1468 ; X64-BMI1: # %bb.0:
1469 ; X64-BMI1-NEXT: shll $8, %esi
1470 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
1471 ; X64-BMI1-NEXT: retq
1472 ;
1473 ; X64-BMI2-LABEL: bzhi64_b0:
1474 ; X64-BMI2: # %bb.0:
1475 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
1476 ; X64-BMI2-NEXT: retq
1477 %notmask = shl i64 -1, %numlowbits
1478 %mask = xor i64 %notmask, -1
1479 %masked = and i64 %mask, %val
1480 ret i64 %masked
1481 }
1483 define i64 @bzhi64_b1_indexzext(i64 %val, i8 zeroext %numlowbits) nounwind {
1484 ; X86-NOBMI-LABEL: bzhi64_b1_indexzext:
1485 ; X86-NOBMI: # %bb.0:
1486 ; X86-NOBMI-NEXT: pushl %esi
1487 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1488 ; X86-NOBMI-NEXT: movl $-1, %edx
1489 ; X86-NOBMI-NEXT: movl $-1, %esi
1490 ; X86-NOBMI-NEXT: shll %cl, %esi
1491 ; X86-NOBMI-NEXT: xorl %eax, %eax
1492 ; X86-NOBMI-NEXT: testb $32, %cl
1493 ; X86-NOBMI-NEXT: jne .LBB22_1
1494 ; X86-NOBMI-NEXT: # %bb.2:
1495 ; X86-NOBMI-NEXT: movl %esi, %eax
1496 ; X86-NOBMI-NEXT: jmp .LBB22_3
1497 ; X86-NOBMI-NEXT: .LBB22_1:
1498 ; X86-NOBMI-NEXT: movl %esi, %edx
1499 ; X86-NOBMI-NEXT: .LBB22_3:
1500 ; X86-NOBMI-NEXT: notl %edx
1501 ; X86-NOBMI-NEXT: notl %eax
1502 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1503 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
1504 ; X86-NOBMI-NEXT: popl %esi
1505 ; X86-NOBMI-NEXT: retl
1506 ;
1507 ; X86-BMI1-LABEL: bzhi64_b1_indexzext:
1508 ; X86-BMI1: # %bb.0:
1509 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1510 ; X86-BMI1-NEXT: movl $-1, %edx
1511 ; X86-BMI1-NEXT: movl $-1, %eax
1512 ; X86-BMI1-NEXT: shll %cl, %eax
1513 ; X86-BMI1-NEXT: testb $32, %cl
1514 ; X86-BMI1-NEXT: je .LBB22_2
1515 ; X86-BMI1-NEXT: # %bb.1:
1516 ; X86-BMI1-NEXT: movl %eax, %edx
1517 ; X86-BMI1-NEXT: xorl %eax, %eax
1518 ; X86-BMI1-NEXT: .LBB22_2:
1519 ; X86-BMI1-NEXT: andnl {{[0-9]+}}(%esp), %eax, %eax
1520 ; X86-BMI1-NEXT: andnl {{[0-9]+}}(%esp), %edx, %edx
1521 ; X86-BMI1-NEXT: retl
1522 ;
1523 ; X86-BMI2-LABEL: bzhi64_b1_indexzext:
1524 ; X86-BMI2: # %bb.0:
1525 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %edx
1526 ; X86-BMI2-NEXT: movl $-1, %ecx
1527 ; X86-BMI2-NEXT: shlxl %edx, %ecx, %eax
1528 ; X86-BMI2-NEXT: testb $32, %dl
1529 ; X86-BMI2-NEXT: je .LBB22_2
1530 ; X86-BMI2-NEXT: # %bb.1:
1531 ; X86-BMI2-NEXT: movl %eax, %ecx
1532 ; X86-BMI2-NEXT: xorl %eax, %eax
1533 ; X86-BMI2-NEXT: .LBB22_2:
1534 ; X86-BMI2-NEXT: andnl {{[0-9]+}}(%esp), %eax, %eax
1535 ; X86-BMI2-NEXT: andnl {{[0-9]+}}(%esp), %ecx, %edx
1536 ; X86-BMI2-NEXT: retl
1537 ;
1538 ; X64-NOBMI-LABEL: bzhi64_b1_indexzext:
1539 ; X64-NOBMI: # %bb.0:
1540 ; X64-NOBMI-NEXT: movl %esi, %ecx
1541 ; X64-NOBMI-NEXT: movq $-1, %rax
1542 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1543 ; X64-NOBMI-NEXT: shlq %cl, %rax
1544 ; X64-NOBMI-NEXT: notq %rax
1545 ; X64-NOBMI-NEXT: andq %rdi, %rax
1546 ; X64-NOBMI-NEXT: retq
1547 ;
1548 ; X64-BMI1-LABEL: bzhi64_b1_indexzext:
1549 ; X64-BMI1: # %bb.0:
1550 ; X64-BMI1-NEXT: # kill: def $esi killed $esi def $rsi
1551 ; X64-BMI1-NEXT: shll $8, %esi
1552 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
1553 ; X64-BMI1-NEXT: retq
1554 ;
1555 ; X64-BMI2-LABEL: bzhi64_b1_indexzext:
1556 ; X64-BMI2: # %bb.0:
1557 ; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi
1558 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
1559 ; X64-BMI2-NEXT: retq
1560 %conv = zext i8 %numlowbits to i64
1561 %notmask = shl i64 -1, %conv
1562 %mask = xor i64 %notmask, -1
1563 %masked = and i64 %mask, %val
1564 ret i64 %masked
1565 }
1567 define i64 @bzhi64_b2_load(ptr %w, i64 %numlowbits) nounwind {
1568 ; X86-NOBMI-LABEL: bzhi64_b2_load:
1569 ; X86-NOBMI: # %bb.0:
1570 ; X86-NOBMI-NEXT: pushl %edi
1571 ; X86-NOBMI-NEXT: pushl %esi
1572 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
1573 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1574 ; X86-NOBMI-NEXT: movl $-1, %edx
1575 ; X86-NOBMI-NEXT: movl $-1, %edi
1576 ; X86-NOBMI-NEXT: shll %cl, %edi
1577 ; X86-NOBMI-NEXT: xorl %eax, %eax
1578 ; X86-NOBMI-NEXT: testb $32, %cl
1579 ; X86-NOBMI-NEXT: jne .LBB23_1
1580 ; X86-NOBMI-NEXT: # %bb.2:
1581 ; X86-NOBMI-NEXT: movl %edi, %eax
1582 ; X86-NOBMI-NEXT: jmp .LBB23_3
1583 ; X86-NOBMI-NEXT: .LBB23_1:
1584 ; X86-NOBMI-NEXT: movl %edi, %edx
1585 ; X86-NOBMI-NEXT: .LBB23_3:
1586 ; X86-NOBMI-NEXT: notl %edx
1587 ; X86-NOBMI-NEXT: notl %eax
1588 ; X86-NOBMI-NEXT: andl (%esi), %eax
1589 ; X86-NOBMI-NEXT: andl 4(%esi), %edx
1590 ; X86-NOBMI-NEXT: popl %esi
1591 ; X86-NOBMI-NEXT: popl %edi
1592 ; X86-NOBMI-NEXT: retl
1593 ;
1594 ; X86-BMI1-LABEL: bzhi64_b2_load:
1595 ; X86-BMI1: # %bb.0:
1596 ; X86-BMI1-NEXT: pushl %esi
1597 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
1598 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1599 ; X86-BMI1-NEXT: movl $-1, %esi
1600 ; X86-BMI1-NEXT: movl $-1, %eax
1601 ; X86-BMI1-NEXT: shll %cl, %eax
1602 ; X86-BMI1-NEXT: testb $32, %cl
1603 ; X86-BMI1-NEXT: je .LBB23_2
1604 ; X86-BMI1-NEXT: # %bb.1:
1605 ; X86-BMI1-NEXT: movl %eax, %esi
1606 ; X86-BMI1-NEXT: xorl %eax, %eax
1607 ; X86-BMI1-NEXT: .LBB23_2:
1608 ; X86-BMI1-NEXT: andnl (%edx), %eax, %eax
1609 ; X86-BMI1-NEXT: andnl 4(%edx), %esi, %edx
1610 ; X86-BMI1-NEXT: popl %esi
1611 ; X86-BMI1-NEXT: retl
1612 ;
1613 ; X86-BMI2-LABEL: bzhi64_b2_load:
1614 ; X86-BMI2: # %bb.0:
1615 ; X86-BMI2-NEXT: pushl %ebx
1616 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
1617 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ebx
1618 ; X86-BMI2-NEXT: movl $-1, %edx
1619 ; X86-BMI2-NEXT: shlxl %ebx, %edx, %eax
1620 ; X86-BMI2-NEXT: testb $32, %bl
1621 ; X86-BMI2-NEXT: je .LBB23_2
1622 ; X86-BMI2-NEXT: # %bb.1:
1623 ; X86-BMI2-NEXT: movl %eax, %edx
1624 ; X86-BMI2-NEXT: xorl %eax, %eax
1625 ; X86-BMI2-NEXT: .LBB23_2:
1626 ; X86-BMI2-NEXT: andnl (%ecx), %eax, %eax
1627 ; X86-BMI2-NEXT: andnl 4(%ecx), %edx, %edx
1628 ; X86-BMI2-NEXT: popl %ebx
1629 ; X86-BMI2-NEXT: retl
1630 ;
1631 ; X64-NOBMI-LABEL: bzhi64_b2_load:
1632 ; X64-NOBMI: # %bb.0:
1633 ; X64-NOBMI-NEXT: movq %rsi, %rcx
1634 ; X64-NOBMI-NEXT: movq $-1, %rax
1635 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
1636 ; X64-NOBMI-NEXT: shlq %cl, %rax
1637 ; X64-NOBMI-NEXT: notq %rax
1638 ; X64-NOBMI-NEXT: andq (%rdi), %rax
1639 ; X64-NOBMI-NEXT: retq
1640 ;
1641 ; X64-BMI1-LABEL: bzhi64_b2_load:
1642 ; X64-BMI1: # %bb.0:
1643 ; X64-BMI1-NEXT: shll $8, %esi
1644 ; X64-BMI1-NEXT: bextrq %rsi, (%rdi), %rax
1645 ; X64-BMI1-NEXT: retq
1646 ;
1647 ; X64-BMI2-LABEL: bzhi64_b2_load:
1648 ; X64-BMI2: # %bb.0:
1649 ; X64-BMI2-NEXT: bzhiq %rsi, (%rdi), %rax
1650 ; X64-BMI2-NEXT: retq
1651 %val = load i64, ptr %w
1652 %notmask = shl i64 -1, %numlowbits
1653 %mask = xor i64 %notmask, -1
1654 %masked = and i64 %mask, %val
1655 ret i64 %masked
1656 }
1658 define i64 @bzhi64_b3_load_indexzext(ptr %w, i8 zeroext %numlowbits) nounwind {
1659 ; X86-NOBMI-LABEL: bzhi64_b3_load_indexzext:
1660 ; X86-NOBMI: # %bb.0:
1661 ; X86-NOBMI-NEXT: pushl %edi
1662 ; X86-NOBMI-NEXT: pushl %esi
1663 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
1664 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1665 ; X86-NOBMI-NEXT: movl $-1, %edx
1666 ; X86-NOBMI-NEXT: movl $-1, %edi
1667 ; X86-NOBMI-NEXT: shll %cl, %edi
1668 ; X86-NOBMI-NEXT: xorl %eax, %eax
1669 ; X86-NOBMI-NEXT: testb $32, %cl
1670 ; X86-NOBMI-NEXT: jne .LBB24_1
1671 ; X86-NOBMI-NEXT: # %bb.2:
1672 ; X86-NOBMI-NEXT: movl %edi, %eax
1673 ; X86-NOBMI-NEXT: jmp .LBB24_3
1674 ; X86-NOBMI-NEXT: .LBB24_1:
1675 ; X86-NOBMI-NEXT: movl %edi, %edx
1676 ; X86-NOBMI-NEXT: .LBB24_3:
1677 ; X86-NOBMI-NEXT: notl %edx
1678 ; X86-NOBMI-NEXT: notl %eax
1679 ; X86-NOBMI-NEXT: andl (%esi), %eax
1680 ; X86-NOBMI-NEXT: andl 4(%esi), %edx
1681 ; X86-NOBMI-NEXT: popl %esi
1682 ; X86-NOBMI-NEXT: popl %edi
1683 ; X86-NOBMI-NEXT: retl
1684 ;
1685 ; X86-BMI1-LABEL: bzhi64_b3_load_indexzext:
1686 ; X86-BMI1: # %bb.0:
1687 ; X86-BMI1-NEXT: pushl %esi
1688 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
1689 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1690 ; X86-BMI1-NEXT: movl $-1, %esi
1691 ; X86-BMI1-NEXT: movl $-1, %eax
1692 ; X86-BMI1-NEXT: shll %cl, %eax
1693 ; X86-BMI1-NEXT: testb $32, %cl
1694 ; X86-BMI1-NEXT: je .LBB24_2
1695 ; X86-BMI1-NEXT: # %bb.1:
1696 ; X86-BMI1-NEXT: movl %eax, %esi
1697 ; X86-BMI1-NEXT: xorl %eax, %eax
1698 ; X86-BMI1-NEXT: .LBB24_2:
1699 ; X86-BMI1-NEXT: andnl (%edx), %eax, %eax
1700 ; X86-BMI1-NEXT: andnl 4(%edx), %esi, %edx
1701 ; X86-BMI1-NEXT: popl %esi
1702 ; X86-BMI1-NEXT: retl
1703 ;
1704 ; X86-BMI2-LABEL: bzhi64_b3_load_indexzext:
1705 ; X86-BMI2: # %bb.0:
1706 ; X86-BMI2-NEXT: pushl %ebx
1707 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
1708 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ebx
1709 ; X86-BMI2-NEXT: movl $-1, %edx
1710 ; X86-BMI2-NEXT: shlxl %ebx, %edx, %eax
1711 ; X86-BMI2-NEXT: testb $32, %bl
1712 ; X86-BMI2-NEXT: je .LBB24_2
1713 ; X86-BMI2-NEXT: # %bb.1:
1714 ; X86-BMI2-NEXT: movl %eax, %edx
1715 ; X86-BMI2-NEXT: xorl %eax, %eax
1716 ; X86-BMI2-NEXT: .LBB24_2:
1717 ; X86-BMI2-NEXT: andnl (%ecx), %eax, %eax
1718 ; X86-BMI2-NEXT: andnl 4(%ecx), %edx, %edx
1719 ; X86-BMI2-NEXT: popl %ebx
1720 ; X86-BMI2-NEXT: retl
1721 ;
1722 ; X64-NOBMI-LABEL: bzhi64_b3_load_indexzext:
1723 ; X64-NOBMI: # %bb.0:
1724 ; X64-NOBMI-NEXT: movl %esi, %ecx
1725 ; X64-NOBMI-NEXT: movq $-1, %rax
1726 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1727 ; X64-NOBMI-NEXT: shlq %cl, %rax
1728 ; X64-NOBMI-NEXT: notq %rax
1729 ; X64-NOBMI-NEXT: andq (%rdi), %rax
1730 ; X64-NOBMI-NEXT: retq
1731 ;
1732 ; X64-BMI1-LABEL: bzhi64_b3_load_indexzext:
1733 ; X64-BMI1: # %bb.0:
1734 ; X64-BMI1-NEXT: # kill: def $esi killed $esi def $rsi
1735 ; X64-BMI1-NEXT: shll $8, %esi
1736 ; X64-BMI1-NEXT: bextrq %rsi, (%rdi), %rax
1737 ; X64-BMI1-NEXT: retq
1738 ;
1739 ; X64-BMI2-LABEL: bzhi64_b3_load_indexzext:
1740 ; X64-BMI2: # %bb.0:
1741 ; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi
1742 ; X64-BMI2-NEXT: bzhiq %rsi, (%rdi), %rax
1743 ; X64-BMI2-NEXT: retq
1744 %val = load i64, ptr %w
1745 %conv = zext i8 %numlowbits to i64
1746 %notmask = shl i64 -1, %conv
1747 %mask = xor i64 %notmask, -1
1748 %masked = and i64 %mask, %val
1749 ret i64 %masked
1750 }
1752 define i64 @bzhi64_b4_commutative(i64 %val, i64 %numlowbits) nounwind {
1753 ; X86-NOBMI-LABEL: bzhi64_b4_commutative:
1754 ; X86-NOBMI: # %bb.0:
1755 ; X86-NOBMI-NEXT: pushl %esi
1756 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1757 ; X86-NOBMI-NEXT: movl $-1, %edx
1758 ; X86-NOBMI-NEXT: movl $-1, %esi
1759 ; X86-NOBMI-NEXT: shll %cl, %esi
1760 ; X86-NOBMI-NEXT: xorl %eax, %eax
1761 ; X86-NOBMI-NEXT: testb $32, %cl
1762 ; X86-NOBMI-NEXT: jne .LBB25_1
1763 ; X86-NOBMI-NEXT: # %bb.2:
1764 ; X86-NOBMI-NEXT: movl %esi, %eax
1765 ; X86-NOBMI-NEXT: jmp .LBB25_3
1766 ; X86-NOBMI-NEXT: .LBB25_1:
1767 ; X86-NOBMI-NEXT: movl %esi, %edx
1768 ; X86-NOBMI-NEXT: .LBB25_3:
1769 ; X86-NOBMI-NEXT: notl %edx
1770 ; X86-NOBMI-NEXT: notl %eax
1771 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1772 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
1773 ; X86-NOBMI-NEXT: popl %esi
1774 ; X86-NOBMI-NEXT: retl
1775 ;
1776 ; X86-BMI1-LABEL: bzhi64_b4_commutative:
1777 ; X86-BMI1: # %bb.0:
1778 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1779 ; X86-BMI1-NEXT: movl $-1, %edx
1780 ; X86-BMI1-NEXT: movl $-1, %eax
1781 ; X86-BMI1-NEXT: shll %cl, %eax
1782 ; X86-BMI1-NEXT: testb $32, %cl
1783 ; X86-BMI1-NEXT: je .LBB25_2
1784 ; X86-BMI1-NEXT: # %bb.1:
1785 ; X86-BMI1-NEXT: movl %eax, %edx
1786 ; X86-BMI1-NEXT: xorl %eax, %eax
1787 ; X86-BMI1-NEXT: .LBB25_2:
1788 ; X86-BMI1-NEXT: andnl {{[0-9]+}}(%esp), %eax, %eax
1789 ; X86-BMI1-NEXT: andnl {{[0-9]+}}(%esp), %edx, %edx
1790 ; X86-BMI1-NEXT: retl
1791 ;
1792 ; X86-BMI2-LABEL: bzhi64_b4_commutative:
1793 ; X86-BMI2: # %bb.0:
1794 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %edx
1795 ; X86-BMI2-NEXT: movl $-1, %ecx
1796 ; X86-BMI2-NEXT: shlxl %edx, %ecx, %eax
1797 ; X86-BMI2-NEXT: testb $32, %dl
1798 ; X86-BMI2-NEXT: je .LBB25_2
1799 ; X86-BMI2-NEXT: # %bb.1:
1800 ; X86-BMI2-NEXT: movl %eax, %ecx
1801 ; X86-BMI2-NEXT: xorl %eax, %eax
1802 ; X86-BMI2-NEXT: .LBB25_2:
1803 ; X86-BMI2-NEXT: andnl {{[0-9]+}}(%esp), %eax, %eax
1804 ; X86-BMI2-NEXT: andnl {{[0-9]+}}(%esp), %ecx, %edx
1805 ; X86-BMI2-NEXT: retl
1806 ;
1807 ; X64-NOBMI-LABEL: bzhi64_b4_commutative:
1808 ; X64-NOBMI: # %bb.0:
1809 ; X64-NOBMI-NEXT: movq %rsi, %rcx
1810 ; X64-NOBMI-NEXT: movq $-1, %rax
1811 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
1812 ; X64-NOBMI-NEXT: shlq %cl, %rax
1813 ; X64-NOBMI-NEXT: notq %rax
1814 ; X64-NOBMI-NEXT: andq %rdi, %rax
1815 ; X64-NOBMI-NEXT: retq
1816 ;
1817 ; X64-BMI1-LABEL: bzhi64_b4_commutative:
1818 ; X64-BMI1: # %bb.0:
1819 ; X64-BMI1-NEXT: shll $8, %esi
1820 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
1821 ; X64-BMI1-NEXT: retq
1822 ;
1823 ; X64-BMI2-LABEL: bzhi64_b4_commutative:
1824 ; X64-BMI2: # %bb.0:
1825 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
1826 ; X64-BMI2-NEXT: retq
1827 %notmask = shl i64 -1, %numlowbits
1828 %mask = xor i64 %notmask, -1
1829 %masked = and i64 %val, %mask ; swapped order
1830 ret i64 %masked
1831 }
1833 ; 64-bit, but with 32-bit output
1835 ; Everything done in 64-bit, truncation happens last.
1836 define i32 @bzhi64_32_b0(i64 %val, i8 %numlowbits) nounwind {
1837 ; X86-NOBMI-LABEL: bzhi64_32_b0:
1838 ; X86-NOBMI: # %bb.0:
1839 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1840 ; X86-NOBMI-NEXT: movl $-1, %edx
1841 ; X86-NOBMI-NEXT: shll %cl, %edx
1842 ; X86-NOBMI-NEXT: xorl %eax, %eax
1843 ; X86-NOBMI-NEXT: testb $32, %cl
1844 ; X86-NOBMI-NEXT: jne .LBB26_2
1845 ; X86-NOBMI-NEXT: # %bb.1:
1846 ; X86-NOBMI-NEXT: movl %edx, %eax
1847 ; X86-NOBMI-NEXT: .LBB26_2:
1848 ; X86-NOBMI-NEXT: notl %eax
1849 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1850 ; X86-NOBMI-NEXT: retl
1851 ;
1852 ; X86-BMI1-LABEL: bzhi64_32_b0:
1853 ; X86-BMI1: # %bb.0:
1854 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1855 ; X86-BMI1-NEXT: movl $-1, %eax
1856 ; X86-BMI1-NEXT: shll %cl, %eax
1857 ; X86-BMI1-NEXT: xorl %edx, %edx
1858 ; X86-BMI1-NEXT: testb $32, %cl
1859 ; X86-BMI1-NEXT: jne .LBB26_2
1860 ; X86-BMI1-NEXT: # %bb.1:
1861 ; X86-BMI1-NEXT: movl %eax, %edx
1862 ; X86-BMI1-NEXT: .LBB26_2:
1863 ; X86-BMI1-NEXT: andnl {{[0-9]+}}(%esp), %edx, %eax
1864 ; X86-BMI1-NEXT: retl
1865 ;
1866 ; X86-BMI2-LABEL: bzhi64_32_b0:
1867 ; X86-BMI2: # %bb.0:
1868 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1869 ; X86-BMI2-NEXT: xorl %ecx, %ecx
1870 ; X86-BMI2-NEXT: testb $32, %al
1871 ; X86-BMI2-NEXT: jne .LBB26_2
1872 ; X86-BMI2-NEXT: # %bb.1:
1873 ; X86-BMI2-NEXT: movl $-1, %ecx
1874 ; X86-BMI2-NEXT: shlxl %eax, %ecx, %ecx
1875 ; X86-BMI2-NEXT: .LBB26_2:
1876 ; X86-BMI2-NEXT: andnl {{[0-9]+}}(%esp), %ecx, %eax
1877 ; X86-BMI2-NEXT: retl
1878 ;
1879 ; X64-NOBMI-LABEL: bzhi64_32_b0:
1880 ; X64-NOBMI: # %bb.0:
1881 ; X64-NOBMI-NEXT: movl %esi, %ecx
1882 ; X64-NOBMI-NEXT: movq $-1, %rax
1883 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1884 ; X64-NOBMI-NEXT: shlq %cl, %rax
1885 ; X64-NOBMI-NEXT: notl %eax
1886 ; X64-NOBMI-NEXT: andl %edi, %eax
1887 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
1888 ; X64-NOBMI-NEXT: retq
1889 ;
1890 ; X64-BMI1-LABEL: bzhi64_32_b0:
1891 ; X64-BMI1: # %bb.0:
1892 ; X64-BMI1-NEXT: shll $8, %esi
1893 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
1894 ; X64-BMI1-NEXT: retq
1895 ;
1896 ; X64-BMI2-LABEL: bzhi64_32_b0:
1897 ; X64-BMI2: # %bb.0:
1898 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
1899 ; X64-BMI2-NEXT: retq
1900 %widenumlowbits = zext i8 %numlowbits to i64
1901 %notmask = shl nsw i64 -1, %widenumlowbits
1902 %mask = xor i64 %notmask, -1
1903 %wideres = and i64 %val, %mask
1904 %res = trunc i64 %wideres to i32
1905 ret i32 %res
1906 }
1908 ; Shifting happens in 64-bit, then truncation. Masking is 32-bit.
1909 define i32 @bzhi64_32_b1(i64 %val, i8 %numlowbits) nounwind {
1910 ; X86-NOBMI-LABEL: bzhi64_32_b1:
1911 ; X86-NOBMI: # %bb.0:
1912 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1913 ; X86-NOBMI-NEXT: movl $-1, %eax
1914 ; X86-NOBMI-NEXT: shll %cl, %eax
1915 ; X86-NOBMI-NEXT: notl %eax
1916 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1917 ; X86-NOBMI-NEXT: retl
1918 ;
1919 ; X86-BMI1-LABEL: bzhi64_32_b1:
1920 ; X86-BMI1: # %bb.0:
1921 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1922 ; X86-BMI1-NEXT: shll $8, %eax
1923 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
1924 ; X86-BMI1-NEXT: retl
1925 ;
1926 ; X86-BMI2-LABEL: bzhi64_32_b1:
1927 ; X86-BMI2: # %bb.0:
1928 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1929 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
1930 ; X86-BMI2-NEXT: retl
1931 ;
1932 ; X64-NOBMI-LABEL: bzhi64_32_b1:
1933 ; X64-NOBMI: # %bb.0:
1934 ; X64-NOBMI-NEXT: movl %esi, %ecx
1935 ; X64-NOBMI-NEXT: movl $-1, %eax
1936 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1937 ; X64-NOBMI-NEXT: shll %cl, %eax
1938 ; X64-NOBMI-NEXT: notl %eax
1939 ; X64-NOBMI-NEXT: andl %edi, %eax
1940 ; X64-NOBMI-NEXT: retq
1941 ;
1942 ; X64-BMI1-LABEL: bzhi64_32_b1:
1943 ; X64-BMI1: # %bb.0:
1944 ; X64-BMI1-NEXT: shll $8, %esi
1945 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
1946 ; X64-BMI1-NEXT: retq
1947 ;
1948 ; X64-BMI2-LABEL: bzhi64_32_b1:
1949 ; X64-BMI2: # %bb.0:
1950 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
1951 ; X64-BMI2-NEXT: retq
1952 %truncval = trunc i64 %val to i32
1953 %widenumlowbits = zext i8 %numlowbits to i32
1954 %notmask = shl nsw i32 -1, %widenumlowbits
1955 %mask = xor i32 %notmask, -1
1956 %res = and i32 %truncval, %mask
1957 ret i32 %res
1958 }
1960 ; Shifting happens in 64-bit. Mask is 32-bit, but extended to 64-bit.
1961 ; Masking is 64-bit. Then truncation.
1962 define i32 @bzhi64_32_b2(i64 %val, i8 %numlowbits) nounwind {
1963 ; X86-NOBMI-LABEL: bzhi64_32_b2:
1964 ; X86-NOBMI: # %bb.0:
1965 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
1966 ; X86-NOBMI-NEXT: movl $-1, %eax
1967 ; X86-NOBMI-NEXT: shll %cl, %eax
1968 ; X86-NOBMI-NEXT: notl %eax
1969 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
1970 ; X86-NOBMI-NEXT: retl
1971 ;
1972 ; X86-BMI1-LABEL: bzhi64_32_b2:
1973 ; X86-BMI1: # %bb.0:
1974 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1975 ; X86-BMI1-NEXT: shll $8, %eax
1976 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
1977 ; X86-BMI1-NEXT: retl
1978 ;
1979 ; X86-BMI2-LABEL: bzhi64_32_b2:
1980 ; X86-BMI2: # %bb.0:
1981 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
1982 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
1983 ; X86-BMI2-NEXT: retl
1984 ;
1985 ; X64-NOBMI-LABEL: bzhi64_32_b2:
1986 ; X64-NOBMI: # %bb.0:
1987 ; X64-NOBMI-NEXT: movl %esi, %ecx
1988 ; X64-NOBMI-NEXT: movl $-1, %eax
1989 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
1990 ; X64-NOBMI-NEXT: shll %cl, %eax
1991 ; X64-NOBMI-NEXT: notl %eax
1992 ; X64-NOBMI-NEXT: andl %edi, %eax
1993 ; X64-NOBMI-NEXT: retq
1994 ;
1995 ; X64-BMI1-LABEL: bzhi64_32_b2:
1996 ; X64-BMI1: # %bb.0:
1997 ; X64-BMI1-NEXT: shll $8, %esi
1998 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
1999 ; X64-BMI1-NEXT: retq
2000 ;
2001 ; X64-BMI2-LABEL: bzhi64_32_b2:
2002 ; X64-BMI2: # %bb.0:
2003 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
2004 ; X64-BMI2-NEXT: retq
2005 %widenumlowbits = zext i8 %numlowbits to i32
2006 %notmask = shl nsw i32 -1, %widenumlowbits
2007 %mask = xor i32 %notmask, -1
2008 %zextmask = zext i32 %mask to i64
2009 %wideres = and i64 %val, %zextmask
2010 %res = trunc i64 %wideres to i32
2011 ret i32 %res
2012 }
2014 ; Shifting happens in 64-bit. Mask is 32-bit, but calculated in 64-bit.
2015 ; Masking is 64-bit. Then truncation.
2016 define i32 @bzhi64_32_b3(i64 %val, i8 %numlowbits) nounwind {
2017 ; X86-NOBMI-LABEL: bzhi64_32_b3:
2018 ; X86-NOBMI: # %bb.0:
2019 ; X86-NOBMI-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
2020 ; X86-NOBMI-NEXT: movl $-1, %edx
2021 ; X86-NOBMI-NEXT: shll %cl, %edx
2022 ; X86-NOBMI-NEXT: xorl %eax, %eax
2023 ; X86-NOBMI-NEXT: testb $32, %cl
2024 ; X86-NOBMI-NEXT: jne .LBB29_2
2025 ; X86-NOBMI-NEXT: # %bb.1:
2026 ; X86-NOBMI-NEXT: movl %edx, %eax
2027 ; X86-NOBMI-NEXT: .LBB29_2:
2028 ; X86-NOBMI-NEXT: notl %eax
2029 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
2030 ; X86-NOBMI-NEXT: retl
2031 ;
2032 ; X86-BMI1-LABEL: bzhi64_32_b3:
2033 ; X86-BMI1: # %bb.0:
2034 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
2035 ; X86-BMI1-NEXT: movl $-1, %eax
2036 ; X86-BMI1-NEXT: shll %cl, %eax
2037 ; X86-BMI1-NEXT: xorl %edx, %edx
2038 ; X86-BMI1-NEXT: testb $32, %cl
2039 ; X86-BMI1-NEXT: jne .LBB29_2
2040 ; X86-BMI1-NEXT: # %bb.1:
2041 ; X86-BMI1-NEXT: movl %eax, %edx
2042 ; X86-BMI1-NEXT: .LBB29_2:
2043 ; X86-BMI1-NEXT: andnl {{[0-9]+}}(%esp), %edx, %eax
2044 ; X86-BMI1-NEXT: retl
2045 ;
2046 ; X86-BMI2-LABEL: bzhi64_32_b3:
2047 ; X86-BMI2: # %bb.0:
2048 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
2049 ; X86-BMI2-NEXT: xorl %ecx, %ecx
2050 ; X86-BMI2-NEXT: testb $32, %al
2051 ; X86-BMI2-NEXT: jne .LBB29_2
2052 ; X86-BMI2-NEXT: # %bb.1:
2053 ; X86-BMI2-NEXT: movl $-1, %ecx
2054 ; X86-BMI2-NEXT: shlxl %eax, %ecx, %ecx
2055 ; X86-BMI2-NEXT: .LBB29_2:
2056 ; X86-BMI2-NEXT: andnl {{[0-9]+}}(%esp), %ecx, %eax
2057 ; X86-BMI2-NEXT: retl
2058 ;
2059 ; X64-NOBMI-LABEL: bzhi64_32_b3:
2060 ; X64-NOBMI: # %bb.0:
2061 ; X64-NOBMI-NEXT: movl %esi, %ecx
2062 ; X64-NOBMI-NEXT: movl $4294967295, %eax # imm = 0xFFFFFFFF
2063 ; X64-NOBMI-NEXT: movl $4294967295, %edx # imm = 0xFFFFFFFF
2064 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2065 ; X64-NOBMI-NEXT: shlq %cl, %rdx
2066 ; X64-NOBMI-NEXT: xorl %edx, %eax
2067 ; X64-NOBMI-NEXT: andl %edi, %eax
2068 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
2069 ; X64-NOBMI-NEXT: retq
2070 ;
2071 ; X64-BMI1-LABEL: bzhi64_32_b3:
2072 ; X64-BMI1: # %bb.0:
2073 ; X64-BMI1-NEXT: shll $8, %esi
2074 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
2075 ; X64-BMI1-NEXT: retq
2076 ;
2077 ; X64-BMI2-LABEL: bzhi64_32_b3:
2078 ; X64-BMI2: # %bb.0:
2079 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
2080 ; X64-BMI2-NEXT: retq
2081 %widenumlowbits = zext i8 %numlowbits to i64
2082 %notmask = shl nsw i64 4294967295, %widenumlowbits
2083 %mask = xor i64 %notmask, 4294967295
2084 %wideres = and i64 %val, %mask
2085 %res = trunc i64 %wideres to i32
2086 ret i32 %res
2087 }
2089 ; ---------------------------------------------------------------------------- ;
2090 ; Pattern c. 32-bit
2091 ; ---------------------------------------------------------------------------- ;
2093 define i32 @bzhi32_c0(i32 %val, i32 %numlowbits, ptr %escape) nounwind {
2094 ; X86-NOBMI-LABEL: bzhi32_c0:
2095 ; X86-NOBMI: # %bb.0:
2096 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
2097 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
2098 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2099 ; X86-NOBMI-NEXT: movl $-1, %eax
2100 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2101 ; X86-NOBMI-NEXT: shrl %cl, %eax
2102 ; X86-NOBMI-NEXT: movl %eax, (%edx)
2103 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
2104 ; X86-NOBMI-NEXT: retl
2105 ;
2106 ; X86-BMI1-LABEL: bzhi32_c0:
2107 ; X86-BMI1: # %bb.0:
2108 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
2109 ; X86-BMI1-NEXT: xorl %ecx, %ecx
2110 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2111 ; X86-BMI1-NEXT: movl $-1, %eax
2112 ; X86-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2113 ; X86-BMI1-NEXT: shrl %cl, %eax
2114 ; X86-BMI1-NEXT: movl %eax, (%edx)
2115 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
2116 ; X86-BMI1-NEXT: retl
2117 ;
2118 ; X86-BMI2-LABEL: bzhi32_c0:
2119 ; X86-BMI2: # %bb.0:
2120 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
2121 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
2122 ; X86-BMI2-NEXT: movl $-1, %edx
2123 ; X86-BMI2-NEXT: bzhil %ecx, %edx, %edx
2124 ; X86-BMI2-NEXT: movl %edx, (%eax)
2125 ; X86-BMI2-NEXT: bzhil %ecx, {{[0-9]+}}(%esp), %eax
2126 ; X86-BMI2-NEXT: retl
2127 ;
2128 ; X64-NOBMI-LABEL: bzhi32_c0:
2129 ; X64-NOBMI: # %bb.0:
2130 ; X64-NOBMI-NEXT: movl %esi, %ecx
2131 ; X64-NOBMI-NEXT: negb %cl
2132 ; X64-NOBMI-NEXT: movl $-1, %eax
2133 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2134 ; X64-NOBMI-NEXT: shrl %cl, %eax
2135 ; X64-NOBMI-NEXT: movl %eax, (%rdx)
2136 ; X64-NOBMI-NEXT: andl %edi, %eax
2137 ; X64-NOBMI-NEXT: retq
2138 ;
2139 ; X64-BMI1-LABEL: bzhi32_c0:
2140 ; X64-BMI1: # %bb.0:
2141 ; X64-BMI1-NEXT: movl %esi, %ecx
2142 ; X64-BMI1-NEXT: negb %cl
2143 ; X64-BMI1-NEXT: movl $-1, %eax
2144 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2145 ; X64-BMI1-NEXT: shrl %cl, %eax
2146 ; X64-BMI1-NEXT: movl %eax, (%rdx)
2147 ; X64-BMI1-NEXT: andl %edi, %eax
2148 ; X64-BMI1-NEXT: retq
2149 ;
2150 ; X64-BMI2-LABEL: bzhi32_c0:
2151 ; X64-BMI2: # %bb.0:
2152 ; X64-BMI2-NEXT: movl $-1, %eax
2153 ; X64-BMI2-NEXT: bzhil %esi, %eax, %eax
2154 ; X64-BMI2-NEXT: movl %eax, (%rdx)
2155 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
2156 ; X64-BMI2-NEXT: retq
2157 %numhighbits = sub i32 32, %numlowbits
2158 %mask = lshr i32 -1, %numhighbits
2159 store i32 %mask, ptr %escape
2160 %masked = and i32 %mask, %val
2161 ret i32 %masked
2162 }
2164 define i32 @bzhi32_c1_indexzext(i32 %val, i8 %numlowbits, ptr %escape) nounwind {
2165 ; X86-NOBMI-LABEL: bzhi32_c1_indexzext:
2166 ; X86-NOBMI: # %bb.0:
2167 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
2168 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
2169 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2170 ; X86-NOBMI-NEXT: movl $-1, %eax
2171 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2172 ; X86-NOBMI-NEXT: shrl %cl, %eax
2173 ; X86-NOBMI-NEXT: movl %eax, (%edx)
2174 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
2175 ; X86-NOBMI-NEXT: retl
2176 ;
2177 ; X86-BMI1-LABEL: bzhi32_c1_indexzext:
2178 ; X86-BMI1: # %bb.0:
2179 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
2180 ; X86-BMI1-NEXT: xorl %ecx, %ecx
2181 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2182 ; X86-BMI1-NEXT: movl $-1, %eax
2183 ; X86-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2184 ; X86-BMI1-NEXT: shrl %cl, %eax
2185 ; X86-BMI1-NEXT: movl %eax, (%edx)
2186 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
2187 ; X86-BMI1-NEXT: retl
2188 ;
2189 ; X86-BMI2-LABEL: bzhi32_c1_indexzext:
2190 ; X86-BMI2: # %bb.0:
2191 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
2192 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
2193 ; X86-BMI2-NEXT: movl $-1, %edx
2194 ; X86-BMI2-NEXT: bzhil %ecx, %edx, %edx
2195 ; X86-BMI2-NEXT: movl %edx, (%eax)
2196 ; X86-BMI2-NEXT: bzhil %ecx, {{[0-9]+}}(%esp), %eax
2197 ; X86-BMI2-NEXT: retl
2198 ;
2199 ; X64-NOBMI-LABEL: bzhi32_c1_indexzext:
2200 ; X64-NOBMI: # %bb.0:
2201 ; X64-NOBMI-NEXT: movl %esi, %ecx
2202 ; X64-NOBMI-NEXT: negb %cl
2203 ; X64-NOBMI-NEXT: movl $-1, %eax
2204 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2205 ; X64-NOBMI-NEXT: shrl %cl, %eax
2206 ; X64-NOBMI-NEXT: movl %eax, (%rdx)
2207 ; X64-NOBMI-NEXT: andl %edi, %eax
2208 ; X64-NOBMI-NEXT: retq
2209 ;
2210 ; X64-BMI1-LABEL: bzhi32_c1_indexzext:
2211 ; X64-BMI1: # %bb.0:
2212 ; X64-BMI1-NEXT: movl %esi, %ecx
2213 ; X64-BMI1-NEXT: negb %cl
2214 ; X64-BMI1-NEXT: movl $-1, %eax
2215 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2216 ; X64-BMI1-NEXT: shrl %cl, %eax
2217 ; X64-BMI1-NEXT: movl %eax, (%rdx)
2218 ; X64-BMI1-NEXT: andl %edi, %eax
2219 ; X64-BMI1-NEXT: retq
2220 ;
2221 ; X64-BMI2-LABEL: bzhi32_c1_indexzext:
2222 ; X64-BMI2: # %bb.0:
2223 ; X64-BMI2-NEXT: movl $-1, %eax
2224 ; X64-BMI2-NEXT: bzhil %esi, %eax, %eax
2225 ; X64-BMI2-NEXT: movl %eax, (%rdx)
2226 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
2227 ; X64-BMI2-NEXT: retq
2228 %numhighbits = sub i8 32, %numlowbits
2229 %sh_prom = zext i8 %numhighbits to i32
2230 %mask = lshr i32 -1, %sh_prom
2231 store i32 %mask, ptr %escape
2232 %masked = and i32 %mask, %val
2233 ret i32 %masked
2234 }
2236 define i32 @bzhi32_c2_load(ptr %w, i32 %numlowbits, ptr %escape) nounwind {
2237 ; X86-NOBMI-LABEL: bzhi32_c2_load:
2238 ; X86-NOBMI: # %bb.0:
2239 ; X86-NOBMI-NEXT: pushl %esi
2240 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
2241 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
2242 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
2243 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2244 ; X86-NOBMI-NEXT: movl $-1, %esi
2245 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2246 ; X86-NOBMI-NEXT: shrl %cl, %esi
2247 ; X86-NOBMI-NEXT: movl (%eax), %eax
2248 ; X86-NOBMI-NEXT: andl %esi, %eax
2249 ; X86-NOBMI-NEXT: movl %esi, (%edx)
2250 ; X86-NOBMI-NEXT: popl %esi
2251 ; X86-NOBMI-NEXT: retl
2252 ;
2253 ; X86-BMI1-LABEL: bzhi32_c2_load:
2254 ; X86-BMI1: # %bb.0:
2255 ; X86-BMI1-NEXT: pushl %esi
2256 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
2257 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
2258 ; X86-BMI1-NEXT: xorl %ecx, %ecx
2259 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2260 ; X86-BMI1-NEXT: movl $-1, %esi
2261 ; X86-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2262 ; X86-BMI1-NEXT: shrl %cl, %esi
2263 ; X86-BMI1-NEXT: movl (%eax), %eax
2264 ; X86-BMI1-NEXT: andl %esi, %eax
2265 ; X86-BMI1-NEXT: movl %esi, (%edx)
2266 ; X86-BMI1-NEXT: popl %esi
2267 ; X86-BMI1-NEXT: retl
2268 ;
2269 ; X86-BMI2-LABEL: bzhi32_c2_load:
2270 ; X86-BMI2: # %bb.0:
2271 ; X86-BMI2-NEXT: pushl %esi
2272 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
2273 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
2274 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %edx
2275 ; X86-BMI2-NEXT: movl $-1, %esi
2276 ; X86-BMI2-NEXT: bzhil %edx, %esi, %esi
2277 ; X86-BMI2-NEXT: bzhil %edx, (%eax), %eax
2278 ; X86-BMI2-NEXT: movl %esi, (%ecx)
2279 ; X86-BMI2-NEXT: popl %esi
2280 ; X86-BMI2-NEXT: retl
2281 ;
2282 ; X64-NOBMI-LABEL: bzhi32_c2_load:
2283 ; X64-NOBMI: # %bb.0:
2284 ; X64-NOBMI-NEXT: movl %esi, %ecx
2285 ; X64-NOBMI-NEXT: negb %cl
2286 ; X64-NOBMI-NEXT: movl $-1, %esi
2287 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2288 ; X64-NOBMI-NEXT: shrl %cl, %esi
2289 ; X64-NOBMI-NEXT: movl (%rdi), %eax
2290 ; X64-NOBMI-NEXT: andl %esi, %eax
2291 ; X64-NOBMI-NEXT: movl %esi, (%rdx)
2292 ; X64-NOBMI-NEXT: retq
2293 ;
2294 ; X64-BMI1-LABEL: bzhi32_c2_load:
2295 ; X64-BMI1: # %bb.0:
2296 ; X64-BMI1-NEXT: movl %esi, %ecx
2297 ; X64-BMI1-NEXT: negb %cl
2298 ; X64-BMI1-NEXT: movl $-1, %esi
2299 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2300 ; X64-BMI1-NEXT: shrl %cl, %esi
2301 ; X64-BMI1-NEXT: movl (%rdi), %eax
2302 ; X64-BMI1-NEXT: andl %esi, %eax
2303 ; X64-BMI1-NEXT: movl %esi, (%rdx)
2304 ; X64-BMI1-NEXT: retq
2305 ;
2306 ; X64-BMI2-LABEL: bzhi32_c2_load:
2307 ; X64-BMI2: # %bb.0:
2308 ; X64-BMI2-NEXT: movl $-1, %eax
2309 ; X64-BMI2-NEXT: bzhil %esi, %eax, %ecx
2310 ; X64-BMI2-NEXT: bzhil %esi, (%rdi), %eax
2311 ; X64-BMI2-NEXT: movl %ecx, (%rdx)
2312 ; X64-BMI2-NEXT: retq
2313 %val = load i32, ptr %w
2314 %numhighbits = sub i32 32, %numlowbits
2315 %mask = lshr i32 -1, %numhighbits
2316 store i32 %mask, ptr %escape
2317 %masked = and i32 %mask, %val
2318 ret i32 %masked
2319 }
2321 define i32 @bzhi32_c3_load_indexzext(ptr %w, i8 %numlowbits, ptr %escape) nounwind {
2322 ; X86-NOBMI-LABEL: bzhi32_c3_load_indexzext:
2323 ; X86-NOBMI: # %bb.0:
2324 ; X86-NOBMI-NEXT: pushl %esi
2325 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
2326 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
2327 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
2328 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2329 ; X86-NOBMI-NEXT: movl $-1, %esi
2330 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2331 ; X86-NOBMI-NEXT: shrl %cl, %esi
2332 ; X86-NOBMI-NEXT: movl (%eax), %eax
2333 ; X86-NOBMI-NEXT: andl %esi, %eax
2334 ; X86-NOBMI-NEXT: movl %esi, (%edx)
2335 ; X86-NOBMI-NEXT: popl %esi
2336 ; X86-NOBMI-NEXT: retl
2337 ;
2338 ; X86-BMI1-LABEL: bzhi32_c3_load_indexzext:
2339 ; X86-BMI1: # %bb.0:
2340 ; X86-BMI1-NEXT: pushl %esi
2341 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
2342 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
2343 ; X86-BMI1-NEXT: xorl %ecx, %ecx
2344 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2345 ; X86-BMI1-NEXT: movl $-1, %esi
2346 ; X86-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2347 ; X86-BMI1-NEXT: shrl %cl, %esi
2348 ; X86-BMI1-NEXT: movl (%eax), %eax
2349 ; X86-BMI1-NEXT: andl %esi, %eax
2350 ; X86-BMI1-NEXT: movl %esi, (%edx)
2351 ; X86-BMI1-NEXT: popl %esi
2352 ; X86-BMI1-NEXT: retl
2353 ;
2354 ; X86-BMI2-LABEL: bzhi32_c3_load_indexzext:
2355 ; X86-BMI2: # %bb.0:
2356 ; X86-BMI2-NEXT: pushl %esi
2357 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
2358 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
2359 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %edx
2360 ; X86-BMI2-NEXT: movl $-1, %esi
2361 ; X86-BMI2-NEXT: bzhil %edx, %esi, %esi
2362 ; X86-BMI2-NEXT: bzhil %edx, (%eax), %eax
2363 ; X86-BMI2-NEXT: movl %esi, (%ecx)
2364 ; X86-BMI2-NEXT: popl %esi
2365 ; X86-BMI2-NEXT: retl
2366 ;
2367 ; X64-NOBMI-LABEL: bzhi32_c3_load_indexzext:
2368 ; X64-NOBMI: # %bb.0:
2369 ; X64-NOBMI-NEXT: movl %esi, %ecx
2370 ; X64-NOBMI-NEXT: negb %cl
2371 ; X64-NOBMI-NEXT: movl $-1, %esi
2372 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2373 ; X64-NOBMI-NEXT: shrl %cl, %esi
2374 ; X64-NOBMI-NEXT: movl (%rdi), %eax
2375 ; X64-NOBMI-NEXT: andl %esi, %eax
2376 ; X64-NOBMI-NEXT: movl %esi, (%rdx)
2377 ; X64-NOBMI-NEXT: retq
2378 ;
2379 ; X64-BMI1-LABEL: bzhi32_c3_load_indexzext:
2380 ; X64-BMI1: # %bb.0:
2381 ; X64-BMI1-NEXT: movl %esi, %ecx
2382 ; X64-BMI1-NEXT: negb %cl
2383 ; X64-BMI1-NEXT: movl $-1, %esi
2384 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2385 ; X64-BMI1-NEXT: shrl %cl, %esi
2386 ; X64-BMI1-NEXT: movl (%rdi), %eax
2387 ; X64-BMI1-NEXT: andl %esi, %eax
2388 ; X64-BMI1-NEXT: movl %esi, (%rdx)
2389 ; X64-BMI1-NEXT: retq
2390 ;
2391 ; X64-BMI2-LABEL: bzhi32_c3_load_indexzext:
2392 ; X64-BMI2: # %bb.0:
2393 ; X64-BMI2-NEXT: movl $-1, %eax
2394 ; X64-BMI2-NEXT: bzhil %esi, %eax, %ecx
2395 ; X64-BMI2-NEXT: bzhil %esi, (%rdi), %eax
2396 ; X64-BMI2-NEXT: movl %ecx, (%rdx)
2397 ; X64-BMI2-NEXT: retq
2398 %val = load i32, ptr %w
2399 %numhighbits = sub i8 32, %numlowbits
2400 %sh_prom = zext i8 %numhighbits to i32
2401 %mask = lshr i32 -1, %sh_prom
2402 store i32 %mask, ptr %escape
2403 %masked = and i32 %mask, %val
2404 ret i32 %masked
2405 }
2407 define i32 @bzhi32_c4_commutative(i32 %val, i32 %numlowbits, ptr %escape) nounwind {
2408 ; X86-NOBMI-LABEL: bzhi32_c4_commutative:
2409 ; X86-NOBMI: # %bb.0:
2410 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
2411 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
2412 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2413 ; X86-NOBMI-NEXT: movl $-1, %eax
2414 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2415 ; X86-NOBMI-NEXT: shrl %cl, %eax
2416 ; X86-NOBMI-NEXT: movl %eax, (%edx)
2417 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
2418 ; X86-NOBMI-NEXT: retl
2419 ;
2420 ; X86-BMI1-LABEL: bzhi32_c4_commutative:
2421 ; X86-BMI1: # %bb.0:
2422 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
2423 ; X86-BMI1-NEXT: xorl %ecx, %ecx
2424 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2425 ; X86-BMI1-NEXT: movl $-1, %eax
2426 ; X86-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2427 ; X86-BMI1-NEXT: shrl %cl, %eax
2428 ; X86-BMI1-NEXT: movl %eax, (%edx)
2429 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
2430 ; X86-BMI1-NEXT: retl
2431 ;
2432 ; X86-BMI2-LABEL: bzhi32_c4_commutative:
2433 ; X86-BMI2: # %bb.0:
2434 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
2435 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
2436 ; X86-BMI2-NEXT: movl $-1, %edx
2437 ; X86-BMI2-NEXT: bzhil %ecx, %edx, %edx
2438 ; X86-BMI2-NEXT: movl %edx, (%eax)
2439 ; X86-BMI2-NEXT: bzhil %ecx, {{[0-9]+}}(%esp), %eax
2440 ; X86-BMI2-NEXT: retl
2441 ;
2442 ; X64-NOBMI-LABEL: bzhi32_c4_commutative:
2443 ; X64-NOBMI: # %bb.0:
2444 ; X64-NOBMI-NEXT: movl %esi, %ecx
2445 ; X64-NOBMI-NEXT: negb %cl
2446 ; X64-NOBMI-NEXT: movl $-1, %eax
2447 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2448 ; X64-NOBMI-NEXT: shrl %cl, %eax
2449 ; X64-NOBMI-NEXT: movl %eax, (%rdx)
2450 ; X64-NOBMI-NEXT: andl %edi, %eax
2451 ; X64-NOBMI-NEXT: retq
2452 ;
2453 ; X64-BMI1-LABEL: bzhi32_c4_commutative:
2454 ; X64-BMI1: # %bb.0:
2455 ; X64-BMI1-NEXT: movl %esi, %ecx
2456 ; X64-BMI1-NEXT: negb %cl
2457 ; X64-BMI1-NEXT: movl $-1, %eax
2458 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2459 ; X64-BMI1-NEXT: shrl %cl, %eax
2460 ; X64-BMI1-NEXT: movl %eax, (%rdx)
2461 ; X64-BMI1-NEXT: andl %edi, %eax
2462 ; X64-BMI1-NEXT: retq
2463 ;
2464 ; X64-BMI2-LABEL: bzhi32_c4_commutative:
2465 ; X64-BMI2: # %bb.0:
2466 ; X64-BMI2-NEXT: movl $-1, %eax
2467 ; X64-BMI2-NEXT: bzhil %esi, %eax, %eax
2468 ; X64-BMI2-NEXT: movl %eax, (%rdx)
2469 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
2470 ; X64-BMI2-NEXT: retq
2471 %numhighbits = sub i32 32, %numlowbits
2472 %mask = lshr i32 -1, %numhighbits
2473 store i32 %mask, ptr %escape
2474 %masked = and i32 %val, %mask ; swapped order
2475 ret i32 %masked
2476 }
2478 ; 64-bit
2480 define i64 @bzhi64_c0(i64 %val, i64 %numlowbits, ptr %escape) nounwind {
2481 ; X86-NOBMI-LABEL: bzhi64_c0:
2482 ; X86-NOBMI: # %bb.0:
2483 ; X86-NOBMI-NEXT: pushl %esi
2484 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
2485 ; X86-NOBMI-NEXT: movb $64, %cl
2486 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2487 ; X86-NOBMI-NEXT: movl $-1, %eax
2488 ; X86-NOBMI-NEXT: movl $-1, %edx
2489 ; X86-NOBMI-NEXT: shrl %cl, %edx
2490 ; X86-NOBMI-NEXT: testb $32, %cl
2491 ; X86-NOBMI-NEXT: je .LBB35_2
2492 ; X86-NOBMI-NEXT: # %bb.1:
2493 ; X86-NOBMI-NEXT: movl %edx, %eax
2494 ; X86-NOBMI-NEXT: xorl %edx, %edx
2495 ; X86-NOBMI-NEXT: .LBB35_2:
2496 ; X86-NOBMI-NEXT: movl %edx, 4(%esi)
2497 ; X86-NOBMI-NEXT: movl %eax, (%esi)
2498 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
2499 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
2500 ; X86-NOBMI-NEXT: popl %esi
2501 ; X86-NOBMI-NEXT: retl
2502 ;
2503 ; X86-BMI1-LABEL: bzhi64_c0:
2504 ; X86-BMI1: # %bb.0:
2505 ; X86-BMI1-NEXT: pushl %esi
2506 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %esi
2507 ; X86-BMI1-NEXT: movb $64, %cl
2508 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2509 ; X86-BMI1-NEXT: movl $-1, %eax
2510 ; X86-BMI1-NEXT: movl $-1, %edx
2511 ; X86-BMI1-NEXT: shrl %cl, %edx
2512 ; X86-BMI1-NEXT: testb $32, %cl
2513 ; X86-BMI1-NEXT: je .LBB35_2
2514 ; X86-BMI1-NEXT: # %bb.1:
2515 ; X86-BMI1-NEXT: movl %edx, %eax
2516 ; X86-BMI1-NEXT: xorl %edx, %edx
2517 ; X86-BMI1-NEXT: .LBB35_2:
2518 ; X86-BMI1-NEXT: movl %edx, 4(%esi)
2519 ; X86-BMI1-NEXT: movl %eax, (%esi)
2520 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
2521 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %edx
2522 ; X86-BMI1-NEXT: popl %esi
2523 ; X86-BMI1-NEXT: retl
2524 ;
2525 ; X86-BMI2-LABEL: bzhi64_c0:
2526 ; X86-BMI2: # %bb.0:
2527 ; X86-BMI2-NEXT: pushl %ebx
2528 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
2529 ; X86-BMI2-NEXT: movb $64, %bl
2530 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %bl
2531 ; X86-BMI2-NEXT: movl $-1, %eax
2532 ; X86-BMI2-NEXT: shrxl %ebx, %eax, %edx
2533 ; X86-BMI2-NEXT: testb $32, %bl
2534 ; X86-BMI2-NEXT: je .LBB35_2
2535 ; X86-BMI2-NEXT: # %bb.1:
2536 ; X86-BMI2-NEXT: movl %edx, %eax
2537 ; X86-BMI2-NEXT: xorl %edx, %edx
2538 ; X86-BMI2-NEXT: .LBB35_2:
2539 ; X86-BMI2-NEXT: movl %edx, 4(%ecx)
2540 ; X86-BMI2-NEXT: movl %eax, (%ecx)
2541 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
2542 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx
2543 ; X86-BMI2-NEXT: popl %ebx
2544 ; X86-BMI2-NEXT: retl
2545 ;
2546 ; X64-NOBMI-LABEL: bzhi64_c0:
2547 ; X64-NOBMI: # %bb.0:
2548 ; X64-NOBMI-NEXT: movq %rsi, %rcx
2549 ; X64-NOBMI-NEXT: negb %cl
2550 ; X64-NOBMI-NEXT: movq $-1, %rax
2551 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
2552 ; X64-NOBMI-NEXT: shrq %cl, %rax
2553 ; X64-NOBMI-NEXT: movq %rax, (%rdx)
2554 ; X64-NOBMI-NEXT: andq %rdi, %rax
2555 ; X64-NOBMI-NEXT: retq
2556 ;
2557 ; X64-BMI1-LABEL: bzhi64_c0:
2558 ; X64-BMI1: # %bb.0:
2559 ; X64-BMI1-NEXT: movq %rsi, %rcx
2560 ; X64-BMI1-NEXT: negb %cl
2561 ; X64-BMI1-NEXT: movq $-1, %rax
2562 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $rcx
2563 ; X64-BMI1-NEXT: shrq %cl, %rax
2564 ; X64-BMI1-NEXT: movq %rax, (%rdx)
2565 ; X64-BMI1-NEXT: andq %rdi, %rax
2566 ; X64-BMI1-NEXT: retq
2567 ;
2568 ; X64-BMI2-LABEL: bzhi64_c0:
2569 ; X64-BMI2: # %bb.0:
2570 ; X64-BMI2-NEXT: movq $-1, %rax
2571 ; X64-BMI2-NEXT: bzhiq %rsi, %rax, %rax
2572 ; X64-BMI2-NEXT: movq %rax, (%rdx)
2573 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
2574 ; X64-BMI2-NEXT: retq
2575 %numhighbits = sub i64 64, %numlowbits
2576 %mask = lshr i64 -1, %numhighbits
2577 store i64 %mask, ptr %escape
2578 %masked = and i64 %mask, %val
2579 ret i64 %masked
2580 }
2582 define i64 @bzhi64_c1_indexzext(i64 %val, i8 %numlowbits, ptr %escape) nounwind {
2583 ; X86-NOBMI-LABEL: bzhi64_c1_indexzext:
2584 ; X86-NOBMI: # %bb.0:
2585 ; X86-NOBMI-NEXT: pushl %esi
2586 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
2587 ; X86-NOBMI-NEXT: movb $64, %cl
2588 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2589 ; X86-NOBMI-NEXT: movl $-1, %eax
2590 ; X86-NOBMI-NEXT: movl $-1, %edx
2591 ; X86-NOBMI-NEXT: shrl %cl, %edx
2592 ; X86-NOBMI-NEXT: testb $32, %cl
2593 ; X86-NOBMI-NEXT: je .LBB36_2
2594 ; X86-NOBMI-NEXT: # %bb.1:
2595 ; X86-NOBMI-NEXT: movl %edx, %eax
2596 ; X86-NOBMI-NEXT: xorl %edx, %edx
2597 ; X86-NOBMI-NEXT: .LBB36_2:
2598 ; X86-NOBMI-NEXT: movl %edx, 4(%esi)
2599 ; X86-NOBMI-NEXT: movl %eax, (%esi)
2600 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
2601 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
2602 ; X86-NOBMI-NEXT: popl %esi
2603 ; X86-NOBMI-NEXT: retl
2604 ;
2605 ; X86-BMI1-LABEL: bzhi64_c1_indexzext:
2606 ; X86-BMI1: # %bb.0:
2607 ; X86-BMI1-NEXT: pushl %esi
2608 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %esi
2609 ; X86-BMI1-NEXT: movb $64, %cl
2610 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2611 ; X86-BMI1-NEXT: movl $-1, %eax
2612 ; X86-BMI1-NEXT: movl $-1, %edx
2613 ; X86-BMI1-NEXT: shrl %cl, %edx
2614 ; X86-BMI1-NEXT: testb $32, %cl
2615 ; X86-BMI1-NEXT: je .LBB36_2
2616 ; X86-BMI1-NEXT: # %bb.1:
2617 ; X86-BMI1-NEXT: movl %edx, %eax
2618 ; X86-BMI1-NEXT: xorl %edx, %edx
2619 ; X86-BMI1-NEXT: .LBB36_2:
2620 ; X86-BMI1-NEXT: movl %edx, 4(%esi)
2621 ; X86-BMI1-NEXT: movl %eax, (%esi)
2622 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
2623 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %edx
2624 ; X86-BMI1-NEXT: popl %esi
2625 ; X86-BMI1-NEXT: retl
2626 ;
2627 ; X86-BMI2-LABEL: bzhi64_c1_indexzext:
2628 ; X86-BMI2: # %bb.0:
2629 ; X86-BMI2-NEXT: pushl %ebx
2630 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
2631 ; X86-BMI2-NEXT: movb $64, %bl
2632 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %bl
2633 ; X86-BMI2-NEXT: movl $-1, %eax
2634 ; X86-BMI2-NEXT: shrxl %ebx, %eax, %edx
2635 ; X86-BMI2-NEXT: testb $32, %bl
2636 ; X86-BMI2-NEXT: je .LBB36_2
2637 ; X86-BMI2-NEXT: # %bb.1:
2638 ; X86-BMI2-NEXT: movl %edx, %eax
2639 ; X86-BMI2-NEXT: xorl %edx, %edx
2640 ; X86-BMI2-NEXT: .LBB36_2:
2641 ; X86-BMI2-NEXT: movl %edx, 4(%ecx)
2642 ; X86-BMI2-NEXT: movl %eax, (%ecx)
2643 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
2644 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx
2645 ; X86-BMI2-NEXT: popl %ebx
2646 ; X86-BMI2-NEXT: retl
2647 ;
2648 ; X64-NOBMI-LABEL: bzhi64_c1_indexzext:
2649 ; X64-NOBMI: # %bb.0:
2650 ; X64-NOBMI-NEXT: movl %esi, %ecx
2651 ; X64-NOBMI-NEXT: negb %cl
2652 ; X64-NOBMI-NEXT: movq $-1, %rax
2653 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2654 ; X64-NOBMI-NEXT: shrq %cl, %rax
2655 ; X64-NOBMI-NEXT: movq %rax, (%rdx)
2656 ; X64-NOBMI-NEXT: andq %rdi, %rax
2657 ; X64-NOBMI-NEXT: retq
2658 ;
2659 ; X64-BMI1-LABEL: bzhi64_c1_indexzext:
2660 ; X64-BMI1: # %bb.0:
2661 ; X64-BMI1-NEXT: movl %esi, %ecx
2662 ; X64-BMI1-NEXT: negb %cl
2663 ; X64-BMI1-NEXT: movq $-1, %rax
2664 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2665 ; X64-BMI1-NEXT: shrq %cl, %rax
2666 ; X64-BMI1-NEXT: movq %rax, (%rdx)
2667 ; X64-BMI1-NEXT: andq %rdi, %rax
2668 ; X64-BMI1-NEXT: retq
2669 ;
2670 ; X64-BMI2-LABEL: bzhi64_c1_indexzext:
2671 ; X64-BMI2: # %bb.0:
2672 ; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi
2673 ; X64-BMI2-NEXT: movq $-1, %rax
2674 ; X64-BMI2-NEXT: bzhiq %rsi, %rax, %rax
2675 ; X64-BMI2-NEXT: movq %rax, (%rdx)
2676 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
2677 ; X64-BMI2-NEXT: retq
2678 %numhighbits = sub i8 64, %numlowbits
2679 %sh_prom = zext i8 %numhighbits to i64
2680 %mask = lshr i64 -1, %sh_prom
2681 store i64 %mask, ptr %escape
2682 %masked = and i64 %mask, %val
2683 ret i64 %masked
2684 }
2686 define i64 @bzhi64_c2_load(ptr %w, i64 %numlowbits, ptr %escape) nounwind {
2687 ; X86-NOBMI-LABEL: bzhi64_c2_load:
2688 ; X86-NOBMI: # %bb.0:
2689 ; X86-NOBMI-NEXT: pushl %ebx
2690 ; X86-NOBMI-NEXT: pushl %edi
2691 ; X86-NOBMI-NEXT: pushl %esi
2692 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
2693 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
2694 ; X86-NOBMI-NEXT: movb $64, %cl
2695 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2696 ; X86-NOBMI-NEXT: movl $-1, %edi
2697 ; X86-NOBMI-NEXT: movl $-1, %ebx
2698 ; X86-NOBMI-NEXT: shrl %cl, %ebx
2699 ; X86-NOBMI-NEXT: testb $32, %cl
2700 ; X86-NOBMI-NEXT: je .LBB37_2
2701 ; X86-NOBMI-NEXT: # %bb.1:
2702 ; X86-NOBMI-NEXT: movl %ebx, %edi
2703 ; X86-NOBMI-NEXT: xorl %ebx, %ebx
2704 ; X86-NOBMI-NEXT: .LBB37_2:
2705 ; X86-NOBMI-NEXT: movl 4(%eax), %edx
2706 ; X86-NOBMI-NEXT: andl %ebx, %edx
2707 ; X86-NOBMI-NEXT: movl (%eax), %eax
2708 ; X86-NOBMI-NEXT: andl %edi, %eax
2709 ; X86-NOBMI-NEXT: movl %ebx, 4(%esi)
2710 ; X86-NOBMI-NEXT: movl %edi, (%esi)
2711 ; X86-NOBMI-NEXT: popl %esi
2712 ; X86-NOBMI-NEXT: popl %edi
2713 ; X86-NOBMI-NEXT: popl %ebx
2714 ; X86-NOBMI-NEXT: retl
2715 ;
2716 ; X86-BMI1-LABEL: bzhi64_c2_load:
2717 ; X86-BMI1: # %bb.0:
2718 ; X86-BMI1-NEXT: pushl %ebx
2719 ; X86-BMI1-NEXT: pushl %edi
2720 ; X86-BMI1-NEXT: pushl %esi
2721 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %esi
2722 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
2723 ; X86-BMI1-NEXT: movb $64, %cl
2724 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2725 ; X86-BMI1-NEXT: movl $-1, %edi
2726 ; X86-BMI1-NEXT: movl $-1, %ebx
2727 ; X86-BMI1-NEXT: shrl %cl, %ebx
2728 ; X86-BMI1-NEXT: testb $32, %cl
2729 ; X86-BMI1-NEXT: je .LBB37_2
2730 ; X86-BMI1-NEXT: # %bb.1:
2731 ; X86-BMI1-NEXT: movl %ebx, %edi
2732 ; X86-BMI1-NEXT: xorl %ebx, %ebx
2733 ; X86-BMI1-NEXT: .LBB37_2:
2734 ; X86-BMI1-NEXT: movl 4(%eax), %edx
2735 ; X86-BMI1-NEXT: andl %ebx, %edx
2736 ; X86-BMI1-NEXT: movl (%eax), %eax
2737 ; X86-BMI1-NEXT: andl %edi, %eax
2738 ; X86-BMI1-NEXT: movl %ebx, 4(%esi)
2739 ; X86-BMI1-NEXT: movl %edi, (%esi)
2740 ; X86-BMI1-NEXT: popl %esi
2741 ; X86-BMI1-NEXT: popl %edi
2742 ; X86-BMI1-NEXT: popl %ebx
2743 ; X86-BMI1-NEXT: retl
2744 ;
2745 ; X86-BMI2-LABEL: bzhi64_c2_load:
2746 ; X86-BMI2: # %bb.0:
2747 ; X86-BMI2-NEXT: pushl %edi
2748 ; X86-BMI2-NEXT: pushl %esi
2749 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
2750 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
2751 ; X86-BMI2-NEXT: movb $64, %dl
2752 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %dl
2753 ; X86-BMI2-NEXT: movl $-1, %esi
2754 ; X86-BMI2-NEXT: shrxl %edx, %esi, %edi
2755 ; X86-BMI2-NEXT: testb $32, %dl
2756 ; X86-BMI2-NEXT: je .LBB37_2
2757 ; X86-BMI2-NEXT: # %bb.1:
2758 ; X86-BMI2-NEXT: movl %edi, %esi
2759 ; X86-BMI2-NEXT: xorl %edi, %edi
2760 ; X86-BMI2-NEXT: .LBB37_2:
2761 ; X86-BMI2-NEXT: movl 4(%eax), %edx
2762 ; X86-BMI2-NEXT: andl %edi, %edx
2763 ; X86-BMI2-NEXT: movl (%eax), %eax
2764 ; X86-BMI2-NEXT: andl %esi, %eax
2765 ; X86-BMI2-NEXT: movl %edi, 4(%ecx)
2766 ; X86-BMI2-NEXT: movl %esi, (%ecx)
2767 ; X86-BMI2-NEXT: popl %esi
2768 ; X86-BMI2-NEXT: popl %edi
2769 ; X86-BMI2-NEXT: retl
2770 ;
2771 ; X64-NOBMI-LABEL: bzhi64_c2_load:
2772 ; X64-NOBMI: # %bb.0:
2773 ; X64-NOBMI-NEXT: movq %rsi, %rcx
2774 ; X64-NOBMI-NEXT: negb %cl
2775 ; X64-NOBMI-NEXT: movq $-1, %rsi
2776 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
2777 ; X64-NOBMI-NEXT: shrq %cl, %rsi
2778 ; X64-NOBMI-NEXT: movq (%rdi), %rax
2779 ; X64-NOBMI-NEXT: andq %rsi, %rax
2780 ; X64-NOBMI-NEXT: movq %rsi, (%rdx)
2781 ; X64-NOBMI-NEXT: retq
2782 ;
2783 ; X64-BMI1-LABEL: bzhi64_c2_load:
2784 ; X64-BMI1: # %bb.0:
2785 ; X64-BMI1-NEXT: movq %rsi, %rcx
2786 ; X64-BMI1-NEXT: negb %cl
2787 ; X64-BMI1-NEXT: movq $-1, %rsi
2788 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $rcx
2789 ; X64-BMI1-NEXT: shrq %cl, %rsi
2790 ; X64-BMI1-NEXT: movq (%rdi), %rax
2791 ; X64-BMI1-NEXT: andq %rsi, %rax
2792 ; X64-BMI1-NEXT: movq %rsi, (%rdx)
2793 ; X64-BMI1-NEXT: retq
2794 ;
2795 ; X64-BMI2-LABEL: bzhi64_c2_load:
2796 ; X64-BMI2: # %bb.0:
2797 ; X64-BMI2-NEXT: movq $-1, %rax
2798 ; X64-BMI2-NEXT: bzhiq %rsi, %rax, %rcx
2799 ; X64-BMI2-NEXT: bzhiq %rsi, (%rdi), %rax
2800 ; X64-BMI2-NEXT: movq %rcx, (%rdx)
2801 ; X64-BMI2-NEXT: retq
2802 %val = load i64, ptr %w
2803 %numhighbits = sub i64 64, %numlowbits
2804 %mask = lshr i64 -1, %numhighbits
2805 store i64 %mask, ptr %escape
2806 %masked = and i64 %mask, %val
2807 ret i64 %masked
2808 }
2810 define i64 @bzhi64_c3_load_indexzext(ptr %w, i8 %numlowbits, ptr %escape) nounwind {
2811 ; X86-NOBMI-LABEL: bzhi64_c3_load_indexzext:
2812 ; X86-NOBMI: # %bb.0:
2813 ; X86-NOBMI-NEXT: pushl %ebx
2814 ; X86-NOBMI-NEXT: pushl %edi
2815 ; X86-NOBMI-NEXT: pushl %esi
2816 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
2817 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
2818 ; X86-NOBMI-NEXT: movb $64, %cl
2819 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2820 ; X86-NOBMI-NEXT: movl $-1, %edi
2821 ; X86-NOBMI-NEXT: movl $-1, %ebx
2822 ; X86-NOBMI-NEXT: shrl %cl, %ebx
2823 ; X86-NOBMI-NEXT: testb $32, %cl
2824 ; X86-NOBMI-NEXT: je .LBB38_2
2825 ; X86-NOBMI-NEXT: # %bb.1:
2826 ; X86-NOBMI-NEXT: movl %ebx, %edi
2827 ; X86-NOBMI-NEXT: xorl %ebx, %ebx
2828 ; X86-NOBMI-NEXT: .LBB38_2:
2829 ; X86-NOBMI-NEXT: movl 4(%eax), %edx
2830 ; X86-NOBMI-NEXT: andl %ebx, %edx
2831 ; X86-NOBMI-NEXT: movl (%eax), %eax
2832 ; X86-NOBMI-NEXT: andl %edi, %eax
2833 ; X86-NOBMI-NEXT: movl %ebx, 4(%esi)
2834 ; X86-NOBMI-NEXT: movl %edi, (%esi)
2835 ; X86-NOBMI-NEXT: popl %esi
2836 ; X86-NOBMI-NEXT: popl %edi
2837 ; X86-NOBMI-NEXT: popl %ebx
2838 ; X86-NOBMI-NEXT: retl
2839 ;
2840 ; X86-BMI1-LABEL: bzhi64_c3_load_indexzext:
2841 ; X86-BMI1: # %bb.0:
2842 ; X86-BMI1-NEXT: pushl %ebx
2843 ; X86-BMI1-NEXT: pushl %edi
2844 ; X86-BMI1-NEXT: pushl %esi
2845 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %esi
2846 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
2847 ; X86-BMI1-NEXT: movb $64, %cl
2848 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2849 ; X86-BMI1-NEXT: movl $-1, %edi
2850 ; X86-BMI1-NEXT: movl $-1, %ebx
2851 ; X86-BMI1-NEXT: shrl %cl, %ebx
2852 ; X86-BMI1-NEXT: testb $32, %cl
2853 ; X86-BMI1-NEXT: je .LBB38_2
2854 ; X86-BMI1-NEXT: # %bb.1:
2855 ; X86-BMI1-NEXT: movl %ebx, %edi
2856 ; X86-BMI1-NEXT: xorl %ebx, %ebx
2857 ; X86-BMI1-NEXT: .LBB38_2:
2858 ; X86-BMI1-NEXT: movl 4(%eax), %edx
2859 ; X86-BMI1-NEXT: andl %ebx, %edx
2860 ; X86-BMI1-NEXT: movl (%eax), %eax
2861 ; X86-BMI1-NEXT: andl %edi, %eax
2862 ; X86-BMI1-NEXT: movl %ebx, 4(%esi)
2863 ; X86-BMI1-NEXT: movl %edi, (%esi)
2864 ; X86-BMI1-NEXT: popl %esi
2865 ; X86-BMI1-NEXT: popl %edi
2866 ; X86-BMI1-NEXT: popl %ebx
2867 ; X86-BMI1-NEXT: retl
2868 ;
2869 ; X86-BMI2-LABEL: bzhi64_c3_load_indexzext:
2870 ; X86-BMI2: # %bb.0:
2871 ; X86-BMI2-NEXT: pushl %edi
2872 ; X86-BMI2-NEXT: pushl %esi
2873 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
2874 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
2875 ; X86-BMI2-NEXT: movb $64, %dl
2876 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %dl
2877 ; X86-BMI2-NEXT: movl $-1, %esi
2878 ; X86-BMI2-NEXT: shrxl %edx, %esi, %edi
2879 ; X86-BMI2-NEXT: testb $32, %dl
2880 ; X86-BMI2-NEXT: je .LBB38_2
2881 ; X86-BMI2-NEXT: # %bb.1:
2882 ; X86-BMI2-NEXT: movl %edi, %esi
2883 ; X86-BMI2-NEXT: xorl %edi, %edi
2884 ; X86-BMI2-NEXT: .LBB38_2:
2885 ; X86-BMI2-NEXT: movl 4(%eax), %edx
2886 ; X86-BMI2-NEXT: andl %edi, %edx
2887 ; X86-BMI2-NEXT: movl (%eax), %eax
2888 ; X86-BMI2-NEXT: andl %esi, %eax
2889 ; X86-BMI2-NEXT: movl %edi, 4(%ecx)
2890 ; X86-BMI2-NEXT: movl %esi, (%ecx)
2891 ; X86-BMI2-NEXT: popl %esi
2892 ; X86-BMI2-NEXT: popl %edi
2893 ; X86-BMI2-NEXT: retl
2894 ;
2895 ; X64-NOBMI-LABEL: bzhi64_c3_load_indexzext:
2896 ; X64-NOBMI: # %bb.0:
2897 ; X64-NOBMI-NEXT: movl %esi, %ecx
2898 ; X64-NOBMI-NEXT: negb %cl
2899 ; X64-NOBMI-NEXT: movq $-1, %rsi
2900 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
2901 ; X64-NOBMI-NEXT: shrq %cl, %rsi
2902 ; X64-NOBMI-NEXT: movq (%rdi), %rax
2903 ; X64-NOBMI-NEXT: andq %rsi, %rax
2904 ; X64-NOBMI-NEXT: movq %rsi, (%rdx)
2905 ; X64-NOBMI-NEXT: retq
2906 ;
2907 ; X64-BMI1-LABEL: bzhi64_c3_load_indexzext:
2908 ; X64-BMI1: # %bb.0:
2909 ; X64-BMI1-NEXT: movl %esi, %ecx
2910 ; X64-BMI1-NEXT: negb %cl
2911 ; X64-BMI1-NEXT: movq $-1, %rsi
2912 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $ecx
2913 ; X64-BMI1-NEXT: shrq %cl, %rsi
2914 ; X64-BMI1-NEXT: movq (%rdi), %rax
2915 ; X64-BMI1-NEXT: andq %rsi, %rax
2916 ; X64-BMI1-NEXT: movq %rsi, (%rdx)
2917 ; X64-BMI1-NEXT: retq
2918 ;
2919 ; X64-BMI2-LABEL: bzhi64_c3_load_indexzext:
2920 ; X64-BMI2: # %bb.0:
2921 ; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi
2922 ; X64-BMI2-NEXT: movq $-1, %rax
2923 ; X64-BMI2-NEXT: bzhiq %rsi, %rax, %rcx
2924 ; X64-BMI2-NEXT: bzhiq %rsi, (%rdi), %rax
2925 ; X64-BMI2-NEXT: movq %rcx, (%rdx)
2926 ; X64-BMI2-NEXT: retq
2927 %val = load i64, ptr %w
2928 %numhighbits = sub i8 64, %numlowbits
2929 %sh_prom = zext i8 %numhighbits to i64
2930 %mask = lshr i64 -1, %sh_prom
2931 store i64 %mask, ptr %escape
2932 %masked = and i64 %mask, %val
2933 ret i64 %masked
2934 }
2936 define i64 @bzhi64_c4_commutative(i64 %val, i64 %numlowbits, ptr %escape) nounwind {
2937 ; X86-NOBMI-LABEL: bzhi64_c4_commutative:
2938 ; X86-NOBMI: # %bb.0:
2939 ; X86-NOBMI-NEXT: pushl %esi
2940 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
2941 ; X86-NOBMI-NEXT: movb $64, %cl
2942 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
2943 ; X86-NOBMI-NEXT: movl $-1, %eax
2944 ; X86-NOBMI-NEXT: movl $-1, %edx
2945 ; X86-NOBMI-NEXT: shrl %cl, %edx
2946 ; X86-NOBMI-NEXT: testb $32, %cl
2947 ; X86-NOBMI-NEXT: je .LBB39_2
2948 ; X86-NOBMI-NEXT: # %bb.1:
2949 ; X86-NOBMI-NEXT: movl %edx, %eax
2950 ; X86-NOBMI-NEXT: xorl %edx, %edx
2951 ; X86-NOBMI-NEXT: .LBB39_2:
2952 ; X86-NOBMI-NEXT: movl %edx, 4(%esi)
2953 ; X86-NOBMI-NEXT: movl %eax, (%esi)
2954 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
2955 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %edx
2956 ; X86-NOBMI-NEXT: popl %esi
2957 ; X86-NOBMI-NEXT: retl
2958 ;
2959 ; X86-BMI1-LABEL: bzhi64_c4_commutative:
2960 ; X86-BMI1: # %bb.0:
2961 ; X86-BMI1-NEXT: pushl %esi
2962 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %esi
2963 ; X86-BMI1-NEXT: movb $64, %cl
2964 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
2965 ; X86-BMI1-NEXT: movl $-1, %eax
2966 ; X86-BMI1-NEXT: movl $-1, %edx
2967 ; X86-BMI1-NEXT: shrl %cl, %edx
2968 ; X86-BMI1-NEXT: testb $32, %cl
2969 ; X86-BMI1-NEXT: je .LBB39_2
2970 ; X86-BMI1-NEXT: # %bb.1:
2971 ; X86-BMI1-NEXT: movl %edx, %eax
2972 ; X86-BMI1-NEXT: xorl %edx, %edx
2973 ; X86-BMI1-NEXT: .LBB39_2:
2974 ; X86-BMI1-NEXT: movl %edx, 4(%esi)
2975 ; X86-BMI1-NEXT: movl %eax, (%esi)
2976 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
2977 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %edx
2978 ; X86-BMI1-NEXT: popl %esi
2979 ; X86-BMI1-NEXT: retl
2980 ;
2981 ; X86-BMI2-LABEL: bzhi64_c4_commutative:
2982 ; X86-BMI2: # %bb.0:
2983 ; X86-BMI2-NEXT: pushl %ebx
2984 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %ecx
2985 ; X86-BMI2-NEXT: movb $64, %bl
2986 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %bl
2987 ; X86-BMI2-NEXT: movl $-1, %eax
2988 ; X86-BMI2-NEXT: shrxl %ebx, %eax, %edx
2989 ; X86-BMI2-NEXT: testb $32, %bl
2990 ; X86-BMI2-NEXT: je .LBB39_2
2991 ; X86-BMI2-NEXT: # %bb.1:
2992 ; X86-BMI2-NEXT: movl %edx, %eax
2993 ; X86-BMI2-NEXT: xorl %edx, %edx
2994 ; X86-BMI2-NEXT: .LBB39_2:
2995 ; X86-BMI2-NEXT: movl %edx, 4(%ecx)
2996 ; X86-BMI2-NEXT: movl %eax, (%ecx)
2997 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
2998 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %edx
2999 ; X86-BMI2-NEXT: popl %ebx
3000 ; X86-BMI2-NEXT: retl
3001 ;
3002 ; X64-NOBMI-LABEL: bzhi64_c4_commutative:
3003 ; X64-NOBMI: # %bb.0:
3004 ; X64-NOBMI-NEXT: movq %rsi, %rcx
3005 ; X64-NOBMI-NEXT: negb %cl
3006 ; X64-NOBMI-NEXT: movq $-1, %rax
3007 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
3008 ; X64-NOBMI-NEXT: shrq %cl, %rax
3009 ; X64-NOBMI-NEXT: movq %rax, (%rdx)
3010 ; X64-NOBMI-NEXT: andq %rdi, %rax
3011 ; X64-NOBMI-NEXT: retq
3012 ;
3013 ; X64-BMI1-LABEL: bzhi64_c4_commutative:
3014 ; X64-BMI1: # %bb.0:
3015 ; X64-BMI1-NEXT: movq %rsi, %rcx
3016 ; X64-BMI1-NEXT: negb %cl
3017 ; X64-BMI1-NEXT: movq $-1, %rax
3018 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $rcx
3019 ; X64-BMI1-NEXT: shrq %cl, %rax
3020 ; X64-BMI1-NEXT: movq %rax, (%rdx)
3021 ; X64-BMI1-NEXT: andq %rdi, %rax
3022 ; X64-BMI1-NEXT: retq
3023 ;
3024 ; X64-BMI2-LABEL: bzhi64_c4_commutative:
3025 ; X64-BMI2: # %bb.0:
3026 ; X64-BMI2-NEXT: movq $-1, %rax
3027 ; X64-BMI2-NEXT: bzhiq %rsi, %rax, %rax
3028 ; X64-BMI2-NEXT: movq %rax, (%rdx)
3029 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
3030 ; X64-BMI2-NEXT: retq
3031 %numhighbits = sub i64 64, %numlowbits
3032 %mask = lshr i64 -1, %numhighbits
3033 store i64 %mask, ptr %escape
3034 %masked = and i64 %val, %mask ; swapped order
3035 ret i64 %masked
3036 }
3038 ; 64-bit, but with 32-bit output
3040 ; Everything done in 64-bit, truncation happens last.
3041 define i32 @bzhi64_32_c0(i64 %val, i64 %numlowbits) nounwind {
3042 ; X86-NOBMI-LABEL: bzhi64_32_c0:
3043 ; X86-NOBMI: # %bb.0:
3044 ; X86-NOBMI-NEXT: movb $64, %cl
3045 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3046 ; X86-NOBMI-NEXT: movl $-1, %eax
3047 ; X86-NOBMI-NEXT: shrl %cl, %eax
3048 ; X86-NOBMI-NEXT: testb $32, %cl
3049 ; X86-NOBMI-NEXT: jne .LBB40_2
3050 ; X86-NOBMI-NEXT: # %bb.1:
3051 ; X86-NOBMI-NEXT: movl $-1, %eax
3052 ; X86-NOBMI-NEXT: .LBB40_2:
3053 ; X86-NOBMI-NEXT: andl {{[0-9]+}}(%esp), %eax
3054 ; X86-NOBMI-NEXT: retl
3055 ;
3056 ; X86-BMI1-LABEL: bzhi64_32_c0:
3057 ; X86-BMI1: # %bb.0:
3058 ; X86-BMI1-NEXT: movb $64, %cl
3059 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
3060 ; X86-BMI1-NEXT: movl $-1, %eax
3061 ; X86-BMI1-NEXT: shrl %cl, %eax
3062 ; X86-BMI1-NEXT: testb $32, %cl
3063 ; X86-BMI1-NEXT: jne .LBB40_2
3064 ; X86-BMI1-NEXT: # %bb.1:
3065 ; X86-BMI1-NEXT: movl $-1, %eax
3066 ; X86-BMI1-NEXT: .LBB40_2:
3067 ; X86-BMI1-NEXT: andl {{[0-9]+}}(%esp), %eax
3068 ; X86-BMI1-NEXT: retl
3069 ;
3070 ; X86-BMI2-LABEL: bzhi64_32_c0:
3071 ; X86-BMI2: # %bb.0:
3072 ; X86-BMI2-NEXT: movb $64, %cl
3073 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %cl
3074 ; X86-BMI2-NEXT: movl $-1, %eax
3075 ; X86-BMI2-NEXT: testb $32, %cl
3076 ; X86-BMI2-NEXT: je .LBB40_2
3077 ; X86-BMI2-NEXT: # %bb.1:
3078 ; X86-BMI2-NEXT: shrxl %ecx, %eax, %eax
3079 ; X86-BMI2-NEXT: .LBB40_2:
3080 ; X86-BMI2-NEXT: andl {{[0-9]+}}(%esp), %eax
3081 ; X86-BMI2-NEXT: retl
3082 ;
3083 ; X64-NOBMI-LABEL: bzhi64_32_c0:
3084 ; X64-NOBMI: # %bb.0:
3085 ; X64-NOBMI-NEXT: movq %rsi, %rcx
3086 ; X64-NOBMI-NEXT: negb %cl
3087 ; X64-NOBMI-NEXT: movq $-1, %rax
3088 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
3089 ; X64-NOBMI-NEXT: shrq %cl, %rax
3090 ; X64-NOBMI-NEXT: andl %edi, %eax
3091 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
3092 ; X64-NOBMI-NEXT: retq
3093 ;
3094 ; X64-BMI1-LABEL: bzhi64_32_c0:
3095 ; X64-BMI1: # %bb.0:
3096 ; X64-BMI1-NEXT: shll $8, %esi
3097 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
3098 ; X64-BMI1-NEXT: retq
3099 ;
3100 ; X64-BMI2-LABEL: bzhi64_32_c0:
3101 ; X64-BMI2: # %bb.0:
3102 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
3103 ; X64-BMI2-NEXT: retq
3104 %numhighbits = sub i64 64, %numlowbits
3105 %mask = lshr i64 -1, %numhighbits
3106 %masked = and i64 %mask, %val
3107 %res = trunc i64 %masked to i32
3108 ret i32 %res
3109 }
3111 ; Shifting happens in 64-bit, then truncation. Masking is 32-bit.
3112 define i32 @bzhi64_32_c1(i64 %val, i32 %numlowbits) nounwind {
3113 ; X86-NOBMI-LABEL: bzhi64_32_c1:
3114 ; X86-NOBMI: # %bb.0:
3115 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3116 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
3117 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3118 ; X86-NOBMI-NEXT: shll %cl, %eax
3119 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3120 ; X86-NOBMI-NEXT: shrl %cl, %eax
3121 ; X86-NOBMI-NEXT: retl
3122 ;
3123 ; X86-BMI1-LABEL: bzhi64_32_c1:
3124 ; X86-BMI1: # %bb.0:
3125 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
3126 ; X86-BMI1-NEXT: shll $8, %eax
3127 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
3128 ; X86-BMI1-NEXT: retl
3129 ;
3130 ; X86-BMI2-LABEL: bzhi64_32_c1:
3131 ; X86-BMI2: # %bb.0:
3132 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
3133 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
3134 ; X86-BMI2-NEXT: retl
3135 ;
3136 ; X64-NOBMI-LABEL: bzhi64_32_c1:
3137 ; X64-NOBMI: # %bb.0:
3138 ; X64-NOBMI-NEXT: movl %esi, %ecx
3139 ; X64-NOBMI-NEXT: movq %rdi, %rax
3140 ; X64-NOBMI-NEXT: negb %cl
3141 ; X64-NOBMI-NEXT: shll %cl, %eax
3142 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3143 ; X64-NOBMI-NEXT: shrl %cl, %eax
3144 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
3145 ; X64-NOBMI-NEXT: retq
3146 ;
3147 ; X64-BMI1-LABEL: bzhi64_32_c1:
3148 ; X64-BMI1: # %bb.0:
3149 ; X64-BMI1-NEXT: shll $8, %esi
3150 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
3151 ; X64-BMI1-NEXT: retq
3152 ;
3153 ; X64-BMI2-LABEL: bzhi64_32_c1:
3154 ; X64-BMI2: # %bb.0:
3155 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
3156 ; X64-BMI2-NEXT: retq
3157 %truncval = trunc i64 %val to i32
3158 %numhighbits = sub i32 32, %numlowbits
3159 %mask = lshr i32 -1, %numhighbits
3160 %masked = and i32 %mask, %truncval
3161 ret i32 %masked
3162 }
3164 ; Shifting happens in 64-bit. Mask is 32-bit, but extended to 64-bit.
3165 ; Masking is 64-bit. Then truncation.
3166 define i32 @bzhi64_32_c2(i64 %val, i32 %numlowbits) nounwind {
3167 ; X86-NOBMI-LABEL: bzhi64_32_c2:
3168 ; X86-NOBMI: # %bb.0:
3169 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3170 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
3171 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3172 ; X86-NOBMI-NEXT: shll %cl, %eax
3173 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3174 ; X86-NOBMI-NEXT: shrl %cl, %eax
3175 ; X86-NOBMI-NEXT: retl
3176 ;
3177 ; X86-BMI1-LABEL: bzhi64_32_c2:
3178 ; X86-BMI1: # %bb.0:
3179 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
3180 ; X86-BMI1-NEXT: shll $8, %eax
3181 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
3182 ; X86-BMI1-NEXT: retl
3183 ;
3184 ; X86-BMI2-LABEL: bzhi64_32_c2:
3185 ; X86-BMI2: # %bb.0:
3186 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
3187 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
3188 ; X86-BMI2-NEXT: retl
3189 ;
3190 ; X64-NOBMI-LABEL: bzhi64_32_c2:
3191 ; X64-NOBMI: # %bb.0:
3192 ; X64-NOBMI-NEXT: movl %esi, %ecx
3193 ; X64-NOBMI-NEXT: movq %rdi, %rax
3194 ; X64-NOBMI-NEXT: negb %cl
3195 ; X64-NOBMI-NEXT: shll %cl, %eax
3196 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3197 ; X64-NOBMI-NEXT: shrl %cl, %eax
3198 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
3199 ; X64-NOBMI-NEXT: retq
3200 ;
3201 ; X64-BMI1-LABEL: bzhi64_32_c2:
3202 ; X64-BMI1: # %bb.0:
3203 ; X64-BMI1-NEXT: shll $8, %esi
3204 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
3205 ; X64-BMI1-NEXT: retq
3206 ;
3207 ; X64-BMI2-LABEL: bzhi64_32_c2:
3208 ; X64-BMI2: # %bb.0:
3209 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
3210 ; X64-BMI2-NEXT: retq
3211 %numhighbits = sub i32 32, %numlowbits
3212 %mask = lshr i32 -1, %numhighbits
3213 %zextmask = zext i32 %mask to i64
3214 %masked = and i64 %zextmask, %val
3215 %truncmasked = trunc i64 %masked to i32
3216 ret i32 %truncmasked
3217 }
3219 ; Shifting happens in 64-bit. Mask is 32-bit, but calculated in 64-bit.
3220 ; Masking is 64-bit. Then truncation.
3221 define i32 @bzhi64_32_c3(i64 %val, i64 %numlowbits) nounwind {
3222 ; X86-LABEL: bzhi64_32_c3:
3223 ; X86: # %bb.0:
3224 ; X86-NEXT: movb $64, %cl
3225 ; X86-NEXT: subb {{[0-9]+}}(%esp), %cl
3226 ; X86-NEXT: xorl %eax, %eax
3227 ; X86-NEXT: movl $-1, %edx
3228 ; X86-NEXT: shrdl %cl, %eax, %edx
3229 ; X86-NEXT: testb $32, %cl
3230 ; X86-NEXT: jne .LBB43_2
3231 ; X86-NEXT: # %bb.1:
3232 ; X86-NEXT: movl %edx, %eax
3233 ; X86-NEXT: .LBB43_2:
3234 ; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
3235 ; X86-NEXT: retl
3236 ;
3237 ; X64-NOBMI-LABEL: bzhi64_32_c3:
3238 ; X64-NOBMI: # %bb.0:
3239 ; X64-NOBMI-NEXT: movq %rsi, %rcx
3240 ; X64-NOBMI-NEXT: negb %cl
3241 ; X64-NOBMI-NEXT: movl $4294967295, %eax # imm = 0xFFFFFFFF
3242 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
3243 ; X64-NOBMI-NEXT: shrq %cl, %rax
3244 ; X64-NOBMI-NEXT: andl %edi, %eax
3245 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
3246 ; X64-NOBMI-NEXT: retq
3247 ;
3248 ; X64-BMI1-LABEL: bzhi64_32_c3:
3249 ; X64-BMI1: # %bb.0:
3250 ; X64-BMI1-NEXT: movq %rsi, %rcx
3251 ; X64-BMI1-NEXT: negb %cl
3252 ; X64-BMI1-NEXT: movl $4294967295, %eax # imm = 0xFFFFFFFF
3253 ; X64-BMI1-NEXT: # kill: def $cl killed $cl killed $rcx
3254 ; X64-BMI1-NEXT: shrq %cl, %rax
3255 ; X64-BMI1-NEXT: andl %edi, %eax
3256 ; X64-BMI1-NEXT: # kill: def $eax killed $eax killed $rax
3257 ; X64-BMI1-NEXT: retq
3258 ;
3259 ; X64-BMI2-LABEL: bzhi64_32_c3:
3260 ; X64-BMI2: # %bb.0:
3261 ; X64-BMI2-NEXT: negb %sil
3262 ; X64-BMI2-NEXT: movl $4294967295, %eax # imm = 0xFFFFFFFF
3263 ; X64-BMI2-NEXT: shrxq %rsi, %rax, %rax
3264 ; X64-BMI2-NEXT: andl %edi, %eax
3265 ; X64-BMI2-NEXT: # kill: def $eax killed $eax killed $rax
3266 ; X64-BMI2-NEXT: retq
3267 %numhighbits = sub i64 64, %numlowbits
3268 %mask = lshr i64 4294967295, %numhighbits
3269 %masked = and i64 %mask, %val
3270 %truncmasked = trunc i64 %masked to i32
3271 ret i32 %truncmasked
3272 }
3274 ; ---------------------------------------------------------------------------- ;
3275 ; Pattern d. 32-bit.
3276 ; ---------------------------------------------------------------------------- ;
3278 define i32 @bzhi32_d0(i32 %val, i32 %numlowbits) nounwind {
3279 ; X86-NOBMI-LABEL: bzhi32_d0:
3280 ; X86-NOBMI: # %bb.0:
3281 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3282 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
3283 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3284 ; X86-NOBMI-NEXT: shll %cl, %eax
3285 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3286 ; X86-NOBMI-NEXT: shrl %cl, %eax
3287 ; X86-NOBMI-NEXT: retl
3288 ;
3289 ; X86-BMI1-LABEL: bzhi32_d0:
3290 ; X86-BMI1: # %bb.0:
3291 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
3292 ; X86-BMI1-NEXT: shll $8, %eax
3293 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
3294 ; X86-BMI1-NEXT: retl
3295 ;
3296 ; X86-BMI2-LABEL: bzhi32_d0:
3297 ; X86-BMI2: # %bb.0:
3298 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
3299 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
3300 ; X86-BMI2-NEXT: retl
3301 ;
3302 ; X64-NOBMI-LABEL: bzhi32_d0:
3303 ; X64-NOBMI: # %bb.0:
3304 ; X64-NOBMI-NEXT: movl %esi, %ecx
3305 ; X64-NOBMI-NEXT: movl %edi, %eax
3306 ; X64-NOBMI-NEXT: negb %cl
3307 ; X64-NOBMI-NEXT: shll %cl, %eax
3308 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3309 ; X64-NOBMI-NEXT: shrl %cl, %eax
3310 ; X64-NOBMI-NEXT: retq
3311 ;
3312 ; X64-BMI1-LABEL: bzhi32_d0:
3313 ; X64-BMI1: # %bb.0:
3314 ; X64-BMI1-NEXT: shll $8, %esi
3315 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
3316 ; X64-BMI1-NEXT: retq
3317 ;
3318 ; X64-BMI2-LABEL: bzhi32_d0:
3319 ; X64-BMI2: # %bb.0:
3320 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
3321 ; X64-BMI2-NEXT: retq
3322 %numhighbits = sub i32 32, %numlowbits
3323 %highbitscleared = shl i32 %val, %numhighbits
3324 %masked = lshr i32 %highbitscleared, %numhighbits
3325 ret i32 %masked
3326 }
3328 define i32 @bzhi32_d1_indexzext(i32 %val, i8 %numlowbits) nounwind {
3329 ; X86-NOBMI-LABEL: bzhi32_d1_indexzext:
3330 ; X86-NOBMI: # %bb.0:
3331 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3332 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
3333 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3334 ; X86-NOBMI-NEXT: shll %cl, %eax
3335 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3336 ; X86-NOBMI-NEXT: shrl %cl, %eax
3337 ; X86-NOBMI-NEXT: retl
3338 ;
3339 ; X86-BMI1-LABEL: bzhi32_d1_indexzext:
3340 ; X86-BMI1: # %bb.0:
3341 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
3342 ; X86-BMI1-NEXT: shll $8, %eax
3343 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
3344 ; X86-BMI1-NEXT: retl
3345 ;
3346 ; X86-BMI2-LABEL: bzhi32_d1_indexzext:
3347 ; X86-BMI2: # %bb.0:
3348 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
3349 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
3350 ; X86-BMI2-NEXT: retl
3351 ;
3352 ; X64-NOBMI-LABEL: bzhi32_d1_indexzext:
3353 ; X64-NOBMI: # %bb.0:
3354 ; X64-NOBMI-NEXT: movl %esi, %ecx
3355 ; X64-NOBMI-NEXT: movl %edi, %eax
3356 ; X64-NOBMI-NEXT: negb %cl
3357 ; X64-NOBMI-NEXT: shll %cl, %eax
3358 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3359 ; X64-NOBMI-NEXT: shrl %cl, %eax
3360 ; X64-NOBMI-NEXT: retq
3361 ;
3362 ; X64-BMI1-LABEL: bzhi32_d1_indexzext:
3363 ; X64-BMI1: # %bb.0:
3364 ; X64-BMI1-NEXT: shll $8, %esi
3365 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
3366 ; X64-BMI1-NEXT: retq
3367 ;
3368 ; X64-BMI2-LABEL: bzhi32_d1_indexzext:
3369 ; X64-BMI2: # %bb.0:
3370 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
3371 ; X64-BMI2-NEXT: retq
3372 %numhighbits = sub i8 32, %numlowbits
3373 %sh_prom = zext i8 %numhighbits to i32
3374 %highbitscleared = shl i32 %val, %sh_prom
3375 %masked = lshr i32 %highbitscleared, %sh_prom
3376 ret i32 %masked
3377 }
3379 define i32 @bzhi32_d2_load(ptr %w, i32 %numlowbits) nounwind {
3380 ; X86-NOBMI-LABEL: bzhi32_d2_load:
3381 ; X86-NOBMI: # %bb.0:
3382 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3383 ; X86-NOBMI-NEXT: movl (%eax), %eax
3384 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
3385 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3386 ; X86-NOBMI-NEXT: shll %cl, %eax
3387 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3388 ; X86-NOBMI-NEXT: shrl %cl, %eax
3389 ; X86-NOBMI-NEXT: retl
3390 ;
3391 ; X86-BMI1-LABEL: bzhi32_d2_load:
3392 ; X86-BMI1: # %bb.0:
3393 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
3394 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
3395 ; X86-BMI1-NEXT: shll $8, %ecx
3396 ; X86-BMI1-NEXT: bextrl %ecx, (%eax), %eax
3397 ; X86-BMI1-NEXT: retl
3398 ;
3399 ; X86-BMI2-LABEL: bzhi32_d2_load:
3400 ; X86-BMI2: # %bb.0:
3401 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
3402 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
3403 ; X86-BMI2-NEXT: bzhil %ecx, (%eax), %eax
3404 ; X86-BMI2-NEXT: retl
3405 ;
3406 ; X64-NOBMI-LABEL: bzhi32_d2_load:
3407 ; X64-NOBMI: # %bb.0:
3408 ; X64-NOBMI-NEXT: movl %esi, %ecx
3409 ; X64-NOBMI-NEXT: movl (%rdi), %eax
3410 ; X64-NOBMI-NEXT: negb %cl
3411 ; X64-NOBMI-NEXT: shll %cl, %eax
3412 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3413 ; X64-NOBMI-NEXT: shrl %cl, %eax
3414 ; X64-NOBMI-NEXT: retq
3415 ;
3416 ; X64-BMI1-LABEL: bzhi32_d2_load:
3417 ; X64-BMI1: # %bb.0:
3418 ; X64-BMI1-NEXT: shll $8, %esi
3419 ; X64-BMI1-NEXT: bextrl %esi, (%rdi), %eax
3420 ; X64-BMI1-NEXT: retq
3421 ;
3422 ; X64-BMI2-LABEL: bzhi32_d2_load:
3423 ; X64-BMI2: # %bb.0:
3424 ; X64-BMI2-NEXT: bzhil %esi, (%rdi), %eax
3425 ; X64-BMI2-NEXT: retq
3426 %val = load i32, ptr %w
3427 %numhighbits = sub i32 32, %numlowbits
3428 %highbitscleared = shl i32 %val, %numhighbits
3429 %masked = lshr i32 %highbitscleared, %numhighbits
3430 ret i32 %masked
3431 }
3433 define i32 @bzhi32_d3_load_indexzext(ptr %w, i8 %numlowbits) nounwind {
3434 ; X86-NOBMI-LABEL: bzhi32_d3_load_indexzext:
3435 ; X86-NOBMI: # %bb.0:
3436 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3437 ; X86-NOBMI-NEXT: movl (%eax), %eax
3438 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
3439 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3440 ; X86-NOBMI-NEXT: shll %cl, %eax
3441 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3442 ; X86-NOBMI-NEXT: shrl %cl, %eax
3443 ; X86-NOBMI-NEXT: retl
3444 ;
3445 ; X86-BMI1-LABEL: bzhi32_d3_load_indexzext:
3446 ; X86-BMI1: # %bb.0:
3447 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
3448 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
3449 ; X86-BMI1-NEXT: shll $8, %ecx
3450 ; X86-BMI1-NEXT: bextrl %ecx, (%eax), %eax
3451 ; X86-BMI1-NEXT: retl
3452 ;
3453 ; X86-BMI2-LABEL: bzhi32_d3_load_indexzext:
3454 ; X86-BMI2: # %bb.0:
3455 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
3456 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %ecx
3457 ; X86-BMI2-NEXT: bzhil %ecx, (%eax), %eax
3458 ; X86-BMI2-NEXT: retl
3459 ;
3460 ; X64-NOBMI-LABEL: bzhi32_d3_load_indexzext:
3461 ; X64-NOBMI: # %bb.0:
3462 ; X64-NOBMI-NEXT: movl %esi, %ecx
3463 ; X64-NOBMI-NEXT: movl (%rdi), %eax
3464 ; X64-NOBMI-NEXT: negb %cl
3465 ; X64-NOBMI-NEXT: shll %cl, %eax
3466 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3467 ; X64-NOBMI-NEXT: shrl %cl, %eax
3468 ; X64-NOBMI-NEXT: retq
3469 ;
3470 ; X64-BMI1-LABEL: bzhi32_d3_load_indexzext:
3471 ; X64-BMI1: # %bb.0:
3472 ; X64-BMI1-NEXT: shll $8, %esi
3473 ; X64-BMI1-NEXT: bextrl %esi, (%rdi), %eax
3474 ; X64-BMI1-NEXT: retq
3475 ;
3476 ; X64-BMI2-LABEL: bzhi32_d3_load_indexzext:
3477 ; X64-BMI2: # %bb.0:
3478 ; X64-BMI2-NEXT: bzhil %esi, (%rdi), %eax
3479 ; X64-BMI2-NEXT: retq
3480 %val = load i32, ptr %w
3481 %numhighbits = sub i8 32, %numlowbits
3482 %sh_prom = zext i8 %numhighbits to i32
3483 %highbitscleared = shl i32 %val, %sh_prom
3484 %masked = lshr i32 %highbitscleared, %sh_prom
3485 ret i32 %masked
3486 }
3488 ; 64-bit.
3490 define i64 @bzhi64_d0(i64 %val, i64 %numlowbits) nounwind {
3491 ; X86-NOBMI-LABEL: bzhi64_d0:
3492 ; X86-NOBMI: # %bb.0:
3493 ; X86-NOBMI-NEXT: pushl %ebx
3494 ; X86-NOBMI-NEXT: pushl %edi
3495 ; X86-NOBMI-NEXT: pushl %esi
3496 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
3497 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3498 ; X86-NOBMI-NEXT: movb $64, %cl
3499 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3500 ; X86-NOBMI-NEXT: movl %edx, %esi
3501 ; X86-NOBMI-NEXT: shll %cl, %esi
3502 ; X86-NOBMI-NEXT: shldl %cl, %edx, %eax
3503 ; X86-NOBMI-NEXT: testb $32, %cl
3504 ; X86-NOBMI-NEXT: movl %esi, %edi
3505 ; X86-NOBMI-NEXT: jne .LBB48_2
3506 ; X86-NOBMI-NEXT: # %bb.1:
3507 ; X86-NOBMI-NEXT: movl %eax, %edi
3508 ; X86-NOBMI-NEXT: .LBB48_2:
3509 ; X86-NOBMI-NEXT: movl %edi, %eax
3510 ; X86-NOBMI-NEXT: shrl %cl, %eax
3511 ; X86-NOBMI-NEXT: xorl %ebx, %ebx
3512 ; X86-NOBMI-NEXT: testb $32, %cl
3513 ; X86-NOBMI-NEXT: movl $0, %edx
3514 ; X86-NOBMI-NEXT: jne .LBB48_4
3515 ; X86-NOBMI-NEXT: # %bb.3:
3516 ; X86-NOBMI-NEXT: movl %esi, %ebx
3517 ; X86-NOBMI-NEXT: movl %eax, %edx
3518 ; X86-NOBMI-NEXT: .LBB48_4:
3519 ; X86-NOBMI-NEXT: shrdl %cl, %edi, %ebx
3520 ; X86-NOBMI-NEXT: testb $32, %cl
3521 ; X86-NOBMI-NEXT: jne .LBB48_6
3522 ; X86-NOBMI-NEXT: # %bb.5:
3523 ; X86-NOBMI-NEXT: movl %ebx, %eax
3524 ; X86-NOBMI-NEXT: .LBB48_6:
3525 ; X86-NOBMI-NEXT: popl %esi
3526 ; X86-NOBMI-NEXT: popl %edi
3527 ; X86-NOBMI-NEXT: popl %ebx
3528 ; X86-NOBMI-NEXT: retl
3529 ;
3530 ; X86-BMI1-LABEL: bzhi64_d0:
3531 ; X86-BMI1: # %bb.0:
3532 ; X86-BMI1-NEXT: pushl %ebx
3533 ; X86-BMI1-NEXT: pushl %edi
3534 ; X86-BMI1-NEXT: pushl %esi
3535 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
3536 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
3537 ; X86-BMI1-NEXT: movb $64, %cl
3538 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
3539 ; X86-BMI1-NEXT: movl %edx, %esi
3540 ; X86-BMI1-NEXT: shll %cl, %esi
3541 ; X86-BMI1-NEXT: shldl %cl, %edx, %eax
3542 ; X86-BMI1-NEXT: testb $32, %cl
3543 ; X86-BMI1-NEXT: movl %esi, %edi
3544 ; X86-BMI1-NEXT: jne .LBB48_2
3545 ; X86-BMI1-NEXT: # %bb.1:
3546 ; X86-BMI1-NEXT: movl %eax, %edi
3547 ; X86-BMI1-NEXT: .LBB48_2:
3548 ; X86-BMI1-NEXT: movl %edi, %eax
3549 ; X86-BMI1-NEXT: shrl %cl, %eax
3550 ; X86-BMI1-NEXT: xorl %ebx, %ebx
3551 ; X86-BMI1-NEXT: testb $32, %cl
3552 ; X86-BMI1-NEXT: movl $0, %edx
3553 ; X86-BMI1-NEXT: jne .LBB48_4
3554 ; X86-BMI1-NEXT: # %bb.3:
3555 ; X86-BMI1-NEXT: movl %esi, %ebx
3556 ; X86-BMI1-NEXT: movl %eax, %edx
3557 ; X86-BMI1-NEXT: .LBB48_4:
3558 ; X86-BMI1-NEXT: shrdl %cl, %edi, %ebx
3559 ; X86-BMI1-NEXT: testb $32, %cl
3560 ; X86-BMI1-NEXT: jne .LBB48_6
3561 ; X86-BMI1-NEXT: # %bb.5:
3562 ; X86-BMI1-NEXT: movl %ebx, %eax
3563 ; X86-BMI1-NEXT: .LBB48_6:
3564 ; X86-BMI1-NEXT: popl %esi
3565 ; X86-BMI1-NEXT: popl %edi
3566 ; X86-BMI1-NEXT: popl %ebx
3567 ; X86-BMI1-NEXT: retl
3568 ;
3569 ; X86-BMI2-LABEL: bzhi64_d0:
3570 ; X86-BMI2: # %bb.0:
3571 ; X86-BMI2-NEXT: pushl %edi
3572 ; X86-BMI2-NEXT: pushl %esi
3573 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
3574 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %esi
3575 ; X86-BMI2-NEXT: movb $64, %cl
3576 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %cl
3577 ; X86-BMI2-NEXT: shldl %cl, %eax, %esi
3578 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %edi
3579 ; X86-BMI2-NEXT: xorl %edx, %edx
3580 ; X86-BMI2-NEXT: testb $32, %cl
3581 ; X86-BMI2-NEXT: je .LBB48_2
3582 ; X86-BMI2-NEXT: # %bb.1:
3583 ; X86-BMI2-NEXT: movl %edi, %esi
3584 ; X86-BMI2-NEXT: movl $0, %edi
3585 ; X86-BMI2-NEXT: .LBB48_2:
3586 ; X86-BMI2-NEXT: shrxl %ecx, %esi, %eax
3587 ; X86-BMI2-NEXT: jne .LBB48_4
3588 ; X86-BMI2-NEXT: # %bb.3:
3589 ; X86-BMI2-NEXT: movl %eax, %edx
3590 ; X86-BMI2-NEXT: .LBB48_4:
3591 ; X86-BMI2-NEXT: shrdl %cl, %esi, %edi
3592 ; X86-BMI2-NEXT: testb $32, %cl
3593 ; X86-BMI2-NEXT: jne .LBB48_6
3594 ; X86-BMI2-NEXT: # %bb.5:
3595 ; X86-BMI2-NEXT: movl %edi, %eax
3596 ; X86-BMI2-NEXT: .LBB48_6:
3597 ; X86-BMI2-NEXT: popl %esi
3598 ; X86-BMI2-NEXT: popl %edi
3599 ; X86-BMI2-NEXT: retl
3600 ;
3601 ; X64-NOBMI-LABEL: bzhi64_d0:
3602 ; X64-NOBMI: # %bb.0:
3603 ; X64-NOBMI-NEXT: movq %rsi, %rcx
3604 ; X64-NOBMI-NEXT: movq %rdi, %rax
3605 ; X64-NOBMI-NEXT: negb %cl
3606 ; X64-NOBMI-NEXT: shlq %cl, %rax
3607 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
3608 ; X64-NOBMI-NEXT: shrq %cl, %rax
3609 ; X64-NOBMI-NEXT: retq
3610 ;
3611 ; X64-BMI1-LABEL: bzhi64_d0:
3612 ; X64-BMI1: # %bb.0:
3613 ; X64-BMI1-NEXT: shll $8, %esi
3614 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
3615 ; X64-BMI1-NEXT: retq
3616 ;
3617 ; X64-BMI2-LABEL: bzhi64_d0:
3618 ; X64-BMI2: # %bb.0:
3619 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
3620 ; X64-BMI2-NEXT: retq
3621 %numhighbits = sub i64 64, %numlowbits
3622 %highbitscleared = shl i64 %val, %numhighbits
3623 %masked = lshr i64 %highbitscleared, %numhighbits
3624 ret i64 %masked
3625 }
3627 define i64 @bzhi64_d1_indexzext(i64 %val, i8 %numlowbits) nounwind {
3628 ; X86-NOBMI-LABEL: bzhi64_d1_indexzext:
3629 ; X86-NOBMI: # %bb.0:
3630 ; X86-NOBMI-NEXT: pushl %ebx
3631 ; X86-NOBMI-NEXT: pushl %edi
3632 ; X86-NOBMI-NEXT: pushl %esi
3633 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %edx
3634 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3635 ; X86-NOBMI-NEXT: movb $64, %cl
3636 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3637 ; X86-NOBMI-NEXT: movl %edx, %esi
3638 ; X86-NOBMI-NEXT: shll %cl, %esi
3639 ; X86-NOBMI-NEXT: shldl %cl, %edx, %eax
3640 ; X86-NOBMI-NEXT: testb $32, %cl
3641 ; X86-NOBMI-NEXT: movl %esi, %edi
3642 ; X86-NOBMI-NEXT: jne .LBB49_2
3643 ; X86-NOBMI-NEXT: # %bb.1:
3644 ; X86-NOBMI-NEXT: movl %eax, %edi
3645 ; X86-NOBMI-NEXT: .LBB49_2:
3646 ; X86-NOBMI-NEXT: movl %edi, %eax
3647 ; X86-NOBMI-NEXT: shrl %cl, %eax
3648 ; X86-NOBMI-NEXT: xorl %ebx, %ebx
3649 ; X86-NOBMI-NEXT: testb $32, %cl
3650 ; X86-NOBMI-NEXT: movl $0, %edx
3651 ; X86-NOBMI-NEXT: jne .LBB49_4
3652 ; X86-NOBMI-NEXT: # %bb.3:
3653 ; X86-NOBMI-NEXT: movl %esi, %ebx
3654 ; X86-NOBMI-NEXT: movl %eax, %edx
3655 ; X86-NOBMI-NEXT: .LBB49_4:
3656 ; X86-NOBMI-NEXT: shrdl %cl, %edi, %ebx
3657 ; X86-NOBMI-NEXT: testb $32, %cl
3658 ; X86-NOBMI-NEXT: jne .LBB49_6
3659 ; X86-NOBMI-NEXT: # %bb.5:
3660 ; X86-NOBMI-NEXT: movl %ebx, %eax
3661 ; X86-NOBMI-NEXT: .LBB49_6:
3662 ; X86-NOBMI-NEXT: popl %esi
3663 ; X86-NOBMI-NEXT: popl %edi
3664 ; X86-NOBMI-NEXT: popl %ebx
3665 ; X86-NOBMI-NEXT: retl
3666 ;
3667 ; X86-BMI1-LABEL: bzhi64_d1_indexzext:
3668 ; X86-BMI1: # %bb.0:
3669 ; X86-BMI1-NEXT: pushl %ebx
3670 ; X86-BMI1-NEXT: pushl %edi
3671 ; X86-BMI1-NEXT: pushl %esi
3672 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %edx
3673 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
3674 ; X86-BMI1-NEXT: movb $64, %cl
3675 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
3676 ; X86-BMI1-NEXT: movl %edx, %esi
3677 ; X86-BMI1-NEXT: shll %cl, %esi
3678 ; X86-BMI1-NEXT: shldl %cl, %edx, %eax
3679 ; X86-BMI1-NEXT: testb $32, %cl
3680 ; X86-BMI1-NEXT: movl %esi, %edi
3681 ; X86-BMI1-NEXT: jne .LBB49_2
3682 ; X86-BMI1-NEXT: # %bb.1:
3683 ; X86-BMI1-NEXT: movl %eax, %edi
3684 ; X86-BMI1-NEXT: .LBB49_2:
3685 ; X86-BMI1-NEXT: movl %edi, %eax
3686 ; X86-BMI1-NEXT: shrl %cl, %eax
3687 ; X86-BMI1-NEXT: xorl %ebx, %ebx
3688 ; X86-BMI1-NEXT: testb $32, %cl
3689 ; X86-BMI1-NEXT: movl $0, %edx
3690 ; X86-BMI1-NEXT: jne .LBB49_4
3691 ; X86-BMI1-NEXT: # %bb.3:
3692 ; X86-BMI1-NEXT: movl %esi, %ebx
3693 ; X86-BMI1-NEXT: movl %eax, %edx
3694 ; X86-BMI1-NEXT: .LBB49_4:
3695 ; X86-BMI1-NEXT: shrdl %cl, %edi, %ebx
3696 ; X86-BMI1-NEXT: testb $32, %cl
3697 ; X86-BMI1-NEXT: jne .LBB49_6
3698 ; X86-BMI1-NEXT: # %bb.5:
3699 ; X86-BMI1-NEXT: movl %ebx, %eax
3700 ; X86-BMI1-NEXT: .LBB49_6:
3701 ; X86-BMI1-NEXT: popl %esi
3702 ; X86-BMI1-NEXT: popl %edi
3703 ; X86-BMI1-NEXT: popl %ebx
3704 ; X86-BMI1-NEXT: retl
3705 ;
3706 ; X86-BMI2-LABEL: bzhi64_d1_indexzext:
3707 ; X86-BMI2: # %bb.0:
3708 ; X86-BMI2-NEXT: pushl %edi
3709 ; X86-BMI2-NEXT: pushl %esi
3710 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
3711 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %esi
3712 ; X86-BMI2-NEXT: movb $64, %cl
3713 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %cl
3714 ; X86-BMI2-NEXT: shldl %cl, %eax, %esi
3715 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %edi
3716 ; X86-BMI2-NEXT: xorl %edx, %edx
3717 ; X86-BMI2-NEXT: testb $32, %cl
3718 ; X86-BMI2-NEXT: je .LBB49_2
3719 ; X86-BMI2-NEXT: # %bb.1:
3720 ; X86-BMI2-NEXT: movl %edi, %esi
3721 ; X86-BMI2-NEXT: movl $0, %edi
3722 ; X86-BMI2-NEXT: .LBB49_2:
3723 ; X86-BMI2-NEXT: shrxl %ecx, %esi, %eax
3724 ; X86-BMI2-NEXT: jne .LBB49_4
3725 ; X86-BMI2-NEXT: # %bb.3:
3726 ; X86-BMI2-NEXT: movl %eax, %edx
3727 ; X86-BMI2-NEXT: .LBB49_4:
3728 ; X86-BMI2-NEXT: shrdl %cl, %esi, %edi
3729 ; X86-BMI2-NEXT: testb $32, %cl
3730 ; X86-BMI2-NEXT: jne .LBB49_6
3731 ; X86-BMI2-NEXT: # %bb.5:
3732 ; X86-BMI2-NEXT: movl %edi, %eax
3733 ; X86-BMI2-NEXT: .LBB49_6:
3734 ; X86-BMI2-NEXT: popl %esi
3735 ; X86-BMI2-NEXT: popl %edi
3736 ; X86-BMI2-NEXT: retl
3737 ;
3738 ; X64-NOBMI-LABEL: bzhi64_d1_indexzext:
3739 ; X64-NOBMI: # %bb.0:
3740 ; X64-NOBMI-NEXT: movl %esi, %ecx
3741 ; X64-NOBMI-NEXT: movq %rdi, %rax
3742 ; X64-NOBMI-NEXT: negb %cl
3743 ; X64-NOBMI-NEXT: shlq %cl, %rax
3744 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
3745 ; X64-NOBMI-NEXT: shrq %cl, %rax
3746 ; X64-NOBMI-NEXT: retq
3747 ;
3748 ; X64-BMI1-LABEL: bzhi64_d1_indexzext:
3749 ; X64-BMI1: # %bb.0:
3750 ; X64-BMI1-NEXT: # kill: def $esi killed $esi def $rsi
3751 ; X64-BMI1-NEXT: shll $8, %esi
3752 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
3753 ; X64-BMI1-NEXT: retq
3754 ;
3755 ; X64-BMI2-LABEL: bzhi64_d1_indexzext:
3756 ; X64-BMI2: # %bb.0:
3757 ; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi
3758 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
3759 ; X64-BMI2-NEXT: retq
3760 %numhighbits = sub i8 64, %numlowbits
3761 %sh_prom = zext i8 %numhighbits to i64
3762 %highbitscleared = shl i64 %val, %sh_prom
3763 %masked = lshr i64 %highbitscleared, %sh_prom
3764 ret i64 %masked
3765 }
3767 define i64 @bzhi64_d2_load(ptr %w, i64 %numlowbits) nounwind {
3768 ; X86-NOBMI-LABEL: bzhi64_d2_load:
3769 ; X86-NOBMI: # %bb.0:
3770 ; X86-NOBMI-NEXT: pushl %ebx
3771 ; X86-NOBMI-NEXT: pushl %edi
3772 ; X86-NOBMI-NEXT: pushl %esi
3773 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3774 ; X86-NOBMI-NEXT: movl (%eax), %edx
3775 ; X86-NOBMI-NEXT: movl 4(%eax), %eax
3776 ; X86-NOBMI-NEXT: movb $64, %cl
3777 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3778 ; X86-NOBMI-NEXT: movl %edx, %esi
3779 ; X86-NOBMI-NEXT: shll %cl, %esi
3780 ; X86-NOBMI-NEXT: shldl %cl, %edx, %eax
3781 ; X86-NOBMI-NEXT: testb $32, %cl
3782 ; X86-NOBMI-NEXT: movl %esi, %edi
3783 ; X86-NOBMI-NEXT: jne .LBB50_2
3784 ; X86-NOBMI-NEXT: # %bb.1:
3785 ; X86-NOBMI-NEXT: movl %eax, %edi
3786 ; X86-NOBMI-NEXT: .LBB50_2:
3787 ; X86-NOBMI-NEXT: movl %edi, %eax
3788 ; X86-NOBMI-NEXT: shrl %cl, %eax
3789 ; X86-NOBMI-NEXT: xorl %ebx, %ebx
3790 ; X86-NOBMI-NEXT: testb $32, %cl
3791 ; X86-NOBMI-NEXT: movl $0, %edx
3792 ; X86-NOBMI-NEXT: jne .LBB50_4
3793 ; X86-NOBMI-NEXT: # %bb.3:
3794 ; X86-NOBMI-NEXT: movl %esi, %ebx
3795 ; X86-NOBMI-NEXT: movl %eax, %edx
3796 ; X86-NOBMI-NEXT: .LBB50_4:
3797 ; X86-NOBMI-NEXT: shrdl %cl, %edi, %ebx
3798 ; X86-NOBMI-NEXT: testb $32, %cl
3799 ; X86-NOBMI-NEXT: jne .LBB50_6
3800 ; X86-NOBMI-NEXT: # %bb.5:
3801 ; X86-NOBMI-NEXT: movl %ebx, %eax
3802 ; X86-NOBMI-NEXT: .LBB50_6:
3803 ; X86-NOBMI-NEXT: popl %esi
3804 ; X86-NOBMI-NEXT: popl %edi
3805 ; X86-NOBMI-NEXT: popl %ebx
3806 ; X86-NOBMI-NEXT: retl
3807 ;
3808 ; X86-BMI1-LABEL: bzhi64_d2_load:
3809 ; X86-BMI1: # %bb.0:
3810 ; X86-BMI1-NEXT: pushl %ebx
3811 ; X86-BMI1-NEXT: pushl %edi
3812 ; X86-BMI1-NEXT: pushl %esi
3813 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
3814 ; X86-BMI1-NEXT: movl (%eax), %edx
3815 ; X86-BMI1-NEXT: movl 4(%eax), %eax
3816 ; X86-BMI1-NEXT: movb $64, %cl
3817 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
3818 ; X86-BMI1-NEXT: movl %edx, %esi
3819 ; X86-BMI1-NEXT: shll %cl, %esi
3820 ; X86-BMI1-NEXT: shldl %cl, %edx, %eax
3821 ; X86-BMI1-NEXT: testb $32, %cl
3822 ; X86-BMI1-NEXT: movl %esi, %edi
3823 ; X86-BMI1-NEXT: jne .LBB50_2
3824 ; X86-BMI1-NEXT: # %bb.1:
3825 ; X86-BMI1-NEXT: movl %eax, %edi
3826 ; X86-BMI1-NEXT: .LBB50_2:
3827 ; X86-BMI1-NEXT: movl %edi, %eax
3828 ; X86-BMI1-NEXT: shrl %cl, %eax
3829 ; X86-BMI1-NEXT: xorl %ebx, %ebx
3830 ; X86-BMI1-NEXT: testb $32, %cl
3831 ; X86-BMI1-NEXT: movl $0, %edx
3832 ; X86-BMI1-NEXT: jne .LBB50_4
3833 ; X86-BMI1-NEXT: # %bb.3:
3834 ; X86-BMI1-NEXT: movl %esi, %ebx
3835 ; X86-BMI1-NEXT: movl %eax, %edx
3836 ; X86-BMI1-NEXT: .LBB50_4:
3837 ; X86-BMI1-NEXT: shrdl %cl, %edi, %ebx
3838 ; X86-BMI1-NEXT: testb $32, %cl
3839 ; X86-BMI1-NEXT: jne .LBB50_6
3840 ; X86-BMI1-NEXT: # %bb.5:
3841 ; X86-BMI1-NEXT: movl %ebx, %eax
3842 ; X86-BMI1-NEXT: .LBB50_6:
3843 ; X86-BMI1-NEXT: popl %esi
3844 ; X86-BMI1-NEXT: popl %edi
3845 ; X86-BMI1-NEXT: popl %ebx
3846 ; X86-BMI1-NEXT: retl
3847 ;
3848 ; X86-BMI2-LABEL: bzhi64_d2_load:
3849 ; X86-BMI2: # %bb.0:
3850 ; X86-BMI2-NEXT: pushl %edi
3851 ; X86-BMI2-NEXT: pushl %esi
3852 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
3853 ; X86-BMI2-NEXT: movl (%eax), %edx
3854 ; X86-BMI2-NEXT: movl 4(%eax), %esi
3855 ; X86-BMI2-NEXT: movb $64, %cl
3856 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %cl
3857 ; X86-BMI2-NEXT: shldl %cl, %edx, %esi
3858 ; X86-BMI2-NEXT: shlxl %ecx, %edx, %edi
3859 ; X86-BMI2-NEXT: xorl %edx, %edx
3860 ; X86-BMI2-NEXT: testb $32, %cl
3861 ; X86-BMI2-NEXT: je .LBB50_2
3862 ; X86-BMI2-NEXT: # %bb.1:
3863 ; X86-BMI2-NEXT: movl %edi, %esi
3864 ; X86-BMI2-NEXT: movl $0, %edi
3865 ; X86-BMI2-NEXT: .LBB50_2:
3866 ; X86-BMI2-NEXT: shrxl %ecx, %esi, %eax
3867 ; X86-BMI2-NEXT: jne .LBB50_4
3868 ; X86-BMI2-NEXT: # %bb.3:
3869 ; X86-BMI2-NEXT: movl %eax, %edx
3870 ; X86-BMI2-NEXT: .LBB50_4:
3871 ; X86-BMI2-NEXT: shrdl %cl, %esi, %edi
3872 ; X86-BMI2-NEXT: testb $32, %cl
3873 ; X86-BMI2-NEXT: jne .LBB50_6
3874 ; X86-BMI2-NEXT: # %bb.5:
3875 ; X86-BMI2-NEXT: movl %edi, %eax
3876 ; X86-BMI2-NEXT: .LBB50_6:
3877 ; X86-BMI2-NEXT: popl %esi
3878 ; X86-BMI2-NEXT: popl %edi
3879 ; X86-BMI2-NEXT: retl
3880 ;
3881 ; X64-NOBMI-LABEL: bzhi64_d2_load:
3882 ; X64-NOBMI: # %bb.0:
3883 ; X64-NOBMI-NEXT: movq %rsi, %rcx
3884 ; X64-NOBMI-NEXT: movq (%rdi), %rax
3885 ; X64-NOBMI-NEXT: negb %cl
3886 ; X64-NOBMI-NEXT: shlq %cl, %rax
3887 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
3888 ; X64-NOBMI-NEXT: shrq %cl, %rax
3889 ; X64-NOBMI-NEXT: retq
3890 ;
3891 ; X64-BMI1-LABEL: bzhi64_d2_load:
3892 ; X64-BMI1: # %bb.0:
3893 ; X64-BMI1-NEXT: shll $8, %esi
3894 ; X64-BMI1-NEXT: bextrq %rsi, (%rdi), %rax
3895 ; X64-BMI1-NEXT: retq
3896 ;
3897 ; X64-BMI2-LABEL: bzhi64_d2_load:
3898 ; X64-BMI2: # %bb.0:
3899 ; X64-BMI2-NEXT: bzhiq %rsi, (%rdi), %rax
3900 ; X64-BMI2-NEXT: retq
3901 %val = load i64, ptr %w
3902 %numhighbits = sub i64 64, %numlowbits
3903 %highbitscleared = shl i64 %val, %numhighbits
3904 %masked = lshr i64 %highbitscleared, %numhighbits
3905 ret i64 %masked
3906 }
3908 define i64 @bzhi64_d3_load_indexzext(ptr %w, i8 %numlowbits) nounwind {
3909 ; X86-NOBMI-LABEL: bzhi64_d3_load_indexzext:
3910 ; X86-NOBMI: # %bb.0:
3911 ; X86-NOBMI-NEXT: pushl %ebx
3912 ; X86-NOBMI-NEXT: pushl %edi
3913 ; X86-NOBMI-NEXT: pushl %esi
3914 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
3915 ; X86-NOBMI-NEXT: movl (%eax), %edx
3916 ; X86-NOBMI-NEXT: movl 4(%eax), %eax
3917 ; X86-NOBMI-NEXT: movb $64, %cl
3918 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
3919 ; X86-NOBMI-NEXT: movl %edx, %esi
3920 ; X86-NOBMI-NEXT: shll %cl, %esi
3921 ; X86-NOBMI-NEXT: shldl %cl, %edx, %eax
3922 ; X86-NOBMI-NEXT: testb $32, %cl
3923 ; X86-NOBMI-NEXT: movl %esi, %edi
3924 ; X86-NOBMI-NEXT: jne .LBB51_2
3925 ; X86-NOBMI-NEXT: # %bb.1:
3926 ; X86-NOBMI-NEXT: movl %eax, %edi
3927 ; X86-NOBMI-NEXT: .LBB51_2:
3928 ; X86-NOBMI-NEXT: movl %edi, %eax
3929 ; X86-NOBMI-NEXT: shrl %cl, %eax
3930 ; X86-NOBMI-NEXT: xorl %ebx, %ebx
3931 ; X86-NOBMI-NEXT: testb $32, %cl
3932 ; X86-NOBMI-NEXT: movl $0, %edx
3933 ; X86-NOBMI-NEXT: jne .LBB51_4
3934 ; X86-NOBMI-NEXT: # %bb.3:
3935 ; X86-NOBMI-NEXT: movl %esi, %ebx
3936 ; X86-NOBMI-NEXT: movl %eax, %edx
3937 ; X86-NOBMI-NEXT: .LBB51_4:
3938 ; X86-NOBMI-NEXT: shrdl %cl, %edi, %ebx
3939 ; X86-NOBMI-NEXT: testb $32, %cl
3940 ; X86-NOBMI-NEXT: jne .LBB51_6
3941 ; X86-NOBMI-NEXT: # %bb.5:
3942 ; X86-NOBMI-NEXT: movl %ebx, %eax
3943 ; X86-NOBMI-NEXT: .LBB51_6:
3944 ; X86-NOBMI-NEXT: popl %esi
3945 ; X86-NOBMI-NEXT: popl %edi
3946 ; X86-NOBMI-NEXT: popl %ebx
3947 ; X86-NOBMI-NEXT: retl
3948 ;
3949 ; X86-BMI1-LABEL: bzhi64_d3_load_indexzext:
3950 ; X86-BMI1: # %bb.0:
3951 ; X86-BMI1-NEXT: pushl %ebx
3952 ; X86-BMI1-NEXT: pushl %edi
3953 ; X86-BMI1-NEXT: pushl %esi
3954 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
3955 ; X86-BMI1-NEXT: movl (%eax), %edx
3956 ; X86-BMI1-NEXT: movl 4(%eax), %eax
3957 ; X86-BMI1-NEXT: movb $64, %cl
3958 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
3959 ; X86-BMI1-NEXT: movl %edx, %esi
3960 ; X86-BMI1-NEXT: shll %cl, %esi
3961 ; X86-BMI1-NEXT: shldl %cl, %edx, %eax
3962 ; X86-BMI1-NEXT: testb $32, %cl
3963 ; X86-BMI1-NEXT: movl %esi, %edi
3964 ; X86-BMI1-NEXT: jne .LBB51_2
3965 ; X86-BMI1-NEXT: # %bb.1:
3966 ; X86-BMI1-NEXT: movl %eax, %edi
3967 ; X86-BMI1-NEXT: .LBB51_2:
3968 ; X86-BMI1-NEXT: movl %edi, %eax
3969 ; X86-BMI1-NEXT: shrl %cl, %eax
3970 ; X86-BMI1-NEXT: xorl %ebx, %ebx
3971 ; X86-BMI1-NEXT: testb $32, %cl
3972 ; X86-BMI1-NEXT: movl $0, %edx
3973 ; X86-BMI1-NEXT: jne .LBB51_4
3974 ; X86-BMI1-NEXT: # %bb.3:
3975 ; X86-BMI1-NEXT: movl %esi, %ebx
3976 ; X86-BMI1-NEXT: movl %eax, %edx
3977 ; X86-BMI1-NEXT: .LBB51_4:
3978 ; X86-BMI1-NEXT: shrdl %cl, %edi, %ebx
3979 ; X86-BMI1-NEXT: testb $32, %cl
3980 ; X86-BMI1-NEXT: jne .LBB51_6
3981 ; X86-BMI1-NEXT: # %bb.5:
3982 ; X86-BMI1-NEXT: movl %ebx, %eax
3983 ; X86-BMI1-NEXT: .LBB51_6:
3984 ; X86-BMI1-NEXT: popl %esi
3985 ; X86-BMI1-NEXT: popl %edi
3986 ; X86-BMI1-NEXT: popl %ebx
3987 ; X86-BMI1-NEXT: retl
3988 ;
3989 ; X86-BMI2-LABEL: bzhi64_d3_load_indexzext:
3990 ; X86-BMI2: # %bb.0:
3991 ; X86-BMI2-NEXT: pushl %edi
3992 ; X86-BMI2-NEXT: pushl %esi
3993 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
3994 ; X86-BMI2-NEXT: movl (%eax), %edx
3995 ; X86-BMI2-NEXT: movl 4(%eax), %esi
3996 ; X86-BMI2-NEXT: movb $64, %cl
3997 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %cl
3998 ; X86-BMI2-NEXT: shldl %cl, %edx, %esi
3999 ; X86-BMI2-NEXT: shlxl %ecx, %edx, %edi
4000 ; X86-BMI2-NEXT: xorl %edx, %edx
4001 ; X86-BMI2-NEXT: testb $32, %cl
4002 ; X86-BMI2-NEXT: je .LBB51_2
4003 ; X86-BMI2-NEXT: # %bb.1:
4004 ; X86-BMI2-NEXT: movl %edi, %esi
4005 ; X86-BMI2-NEXT: movl $0, %edi
4006 ; X86-BMI2-NEXT: .LBB51_2:
4007 ; X86-BMI2-NEXT: shrxl %ecx, %esi, %eax
4008 ; X86-BMI2-NEXT: jne .LBB51_4
4009 ; X86-BMI2-NEXT: # %bb.3:
4010 ; X86-BMI2-NEXT: movl %eax, %edx
4011 ; X86-BMI2-NEXT: .LBB51_4:
4012 ; X86-BMI2-NEXT: shrdl %cl, %esi, %edi
4013 ; X86-BMI2-NEXT: testb $32, %cl
4014 ; X86-BMI2-NEXT: jne .LBB51_6
4015 ; X86-BMI2-NEXT: # %bb.5:
4016 ; X86-BMI2-NEXT: movl %edi, %eax
4017 ; X86-BMI2-NEXT: .LBB51_6:
4018 ; X86-BMI2-NEXT: popl %esi
4019 ; X86-BMI2-NEXT: popl %edi
4020 ; X86-BMI2-NEXT: retl
4021 ;
4022 ; X64-NOBMI-LABEL: bzhi64_d3_load_indexzext:
4023 ; X64-NOBMI: # %bb.0:
4024 ; X64-NOBMI-NEXT: movl %esi, %ecx
4025 ; X64-NOBMI-NEXT: movq (%rdi), %rax
4026 ; X64-NOBMI-NEXT: negb %cl
4027 ; X64-NOBMI-NEXT: shlq %cl, %rax
4028 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
4029 ; X64-NOBMI-NEXT: shrq %cl, %rax
4030 ; X64-NOBMI-NEXT: retq
4031 ;
4032 ; X64-BMI1-LABEL: bzhi64_d3_load_indexzext:
4033 ; X64-BMI1: # %bb.0:
4034 ; X64-BMI1-NEXT: # kill: def $esi killed $esi def $rsi
4035 ; X64-BMI1-NEXT: shll $8, %esi
4036 ; X64-BMI1-NEXT: bextrq %rsi, (%rdi), %rax
4037 ; X64-BMI1-NEXT: retq
4038 ;
4039 ; X64-BMI2-LABEL: bzhi64_d3_load_indexzext:
4040 ; X64-BMI2: # %bb.0:
4041 ; X64-BMI2-NEXT: # kill: def $esi killed $esi def $rsi
4042 ; X64-BMI2-NEXT: bzhiq %rsi, (%rdi), %rax
4043 ; X64-BMI2-NEXT: retq
4044 %val = load i64, ptr %w
4045 %numhighbits = sub i8 64, %numlowbits
4046 %sh_prom = zext i8 %numhighbits to i64
4047 %highbitscleared = shl i64 %val, %sh_prom
4048 %masked = lshr i64 %highbitscleared, %sh_prom
4049 ret i64 %masked
4050 }
4052 ; 64-bit, but with 32-bit output
4054 ; Everything done in 64-bit, truncation happens last.
4055 define i32 @bzhi64_32_d0(i64 %val, i64 %numlowbits) nounwind {
4056 ; X86-NOBMI-LABEL: bzhi64_32_d0:
4057 ; X86-NOBMI: # %bb.0:
4058 ; X86-NOBMI-NEXT: pushl %esi
4059 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %esi
4060 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
4061 ; X86-NOBMI-NEXT: movb $64, %cl
4062 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
4063 ; X86-NOBMI-NEXT: movl %esi, %edx
4064 ; X86-NOBMI-NEXT: shll %cl, %edx
4065 ; X86-NOBMI-NEXT: shldl %cl, %esi, %eax
4066 ; X86-NOBMI-NEXT: testb $32, %cl
4067 ; X86-NOBMI-NEXT: je .LBB52_2
4068 ; X86-NOBMI-NEXT: # %bb.1:
4069 ; X86-NOBMI-NEXT: movl %edx, %eax
4070 ; X86-NOBMI-NEXT: xorl %edx, %edx
4071 ; X86-NOBMI-NEXT: .LBB52_2:
4072 ; X86-NOBMI-NEXT: shrdl %cl, %eax, %edx
4073 ; X86-NOBMI-NEXT: shrl %cl, %eax
4074 ; X86-NOBMI-NEXT: testb $32, %cl
4075 ; X86-NOBMI-NEXT: jne .LBB52_4
4076 ; X86-NOBMI-NEXT: # %bb.3:
4077 ; X86-NOBMI-NEXT: movl %edx, %eax
4078 ; X86-NOBMI-NEXT: .LBB52_4:
4079 ; X86-NOBMI-NEXT: popl %esi
4080 ; X86-NOBMI-NEXT: retl
4081 ;
4082 ; X86-BMI1-LABEL: bzhi64_32_d0:
4083 ; X86-BMI1: # %bb.0:
4084 ; X86-BMI1-NEXT: pushl %esi
4085 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %esi
4086 ; X86-BMI1-NEXT: movl {{[0-9]+}}(%esp), %eax
4087 ; X86-BMI1-NEXT: movb $64, %cl
4088 ; X86-BMI1-NEXT: subb {{[0-9]+}}(%esp), %cl
4089 ; X86-BMI1-NEXT: movl %esi, %edx
4090 ; X86-BMI1-NEXT: shll %cl, %edx
4091 ; X86-BMI1-NEXT: shldl %cl, %esi, %eax
4092 ; X86-BMI1-NEXT: testb $32, %cl
4093 ; X86-BMI1-NEXT: je .LBB52_2
4094 ; X86-BMI1-NEXT: # %bb.1:
4095 ; X86-BMI1-NEXT: movl %edx, %eax
4096 ; X86-BMI1-NEXT: xorl %edx, %edx
4097 ; X86-BMI1-NEXT: .LBB52_2:
4098 ; X86-BMI1-NEXT: shrdl %cl, %eax, %edx
4099 ; X86-BMI1-NEXT: shrl %cl, %eax
4100 ; X86-BMI1-NEXT: testb $32, %cl
4101 ; X86-BMI1-NEXT: jne .LBB52_4
4102 ; X86-BMI1-NEXT: # %bb.3:
4103 ; X86-BMI1-NEXT: movl %edx, %eax
4104 ; X86-BMI1-NEXT: .LBB52_4:
4105 ; X86-BMI1-NEXT: popl %esi
4106 ; X86-BMI1-NEXT: retl
4107 ;
4108 ; X86-BMI2-LABEL: bzhi64_32_d0:
4109 ; X86-BMI2: # %bb.0:
4110 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %eax
4111 ; X86-BMI2-NEXT: movl {{[0-9]+}}(%esp), %edx
4112 ; X86-BMI2-NEXT: movb $64, %cl
4113 ; X86-BMI2-NEXT: subb {{[0-9]+}}(%esp), %cl
4114 ; X86-BMI2-NEXT: shldl %cl, %eax, %edx
4115 ; X86-BMI2-NEXT: shlxl %ecx, %eax, %eax
4116 ; X86-BMI2-NEXT: testb $32, %cl
4117 ; X86-BMI2-NEXT: je .LBB52_2
4118 ; X86-BMI2-NEXT: # %bb.1:
4119 ; X86-BMI2-NEXT: movl %eax, %edx
4120 ; X86-BMI2-NEXT: xorl %eax, %eax
4121 ; X86-BMI2-NEXT: .LBB52_2:
4122 ; X86-BMI2-NEXT: shrdl %cl, %edx, %eax
4123 ; X86-BMI2-NEXT: testb $32, %cl
4124 ; X86-BMI2-NEXT: je .LBB52_4
4125 ; X86-BMI2-NEXT: # %bb.3:
4126 ; X86-BMI2-NEXT: shrxl %ecx, %edx, %eax
4127 ; X86-BMI2-NEXT: .LBB52_4:
4128 ; X86-BMI2-NEXT: retl
4129 ;
4130 ; X64-NOBMI-LABEL: bzhi64_32_d0:
4131 ; X64-NOBMI: # %bb.0:
4132 ; X64-NOBMI-NEXT: movq %rsi, %rcx
4133 ; X64-NOBMI-NEXT: movq %rdi, %rax
4134 ; X64-NOBMI-NEXT: negb %cl
4135 ; X64-NOBMI-NEXT: shlq %cl, %rax
4136 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $rcx
4137 ; X64-NOBMI-NEXT: shrq %cl, %rax
4138 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
4139 ; X64-NOBMI-NEXT: retq
4140 ;
4141 ; X64-BMI1-LABEL: bzhi64_32_d0:
4142 ; X64-BMI1: # %bb.0:
4143 ; X64-BMI1-NEXT: shll $8, %esi
4144 ; X64-BMI1-NEXT: bextrq %rsi, %rdi, %rax
4145 ; X64-BMI1-NEXT: # kill: def $eax killed $eax killed $rax
4146 ; X64-BMI1-NEXT: retq
4147 ;
4148 ; X64-BMI2-LABEL: bzhi64_32_d0:
4149 ; X64-BMI2: # %bb.0:
4150 ; X64-BMI2-NEXT: bzhiq %rsi, %rdi, %rax
4151 ; X64-BMI2-NEXT: # kill: def $eax killed $eax killed $rax
4152 ; X64-BMI2-NEXT: retq
4153 %numhighbits = sub i64 64, %numlowbits
4154 %highbitscleared = shl i64 %val, %numhighbits
4155 %masked = lshr i64 %highbitscleared, %numhighbits
4156 %res = trunc i64 %masked to i32
4157 ret i32 %res
4158 }
4160 ; Shifting happens in 64-bit, then truncation. Masking is 32-bit.
4161 define i32 @bzhi64_32_d1(i64 %val, i32 %numlowbits) nounwind {
4162 ; X86-NOBMI-LABEL: bzhi64_32_d1:
4163 ; X86-NOBMI: # %bb.0:
4164 ; X86-NOBMI-NEXT: movl {{[0-9]+}}(%esp), %eax
4165 ; X86-NOBMI-NEXT: xorl %ecx, %ecx
4166 ; X86-NOBMI-NEXT: subb {{[0-9]+}}(%esp), %cl
4167 ; X86-NOBMI-NEXT: shll %cl, %eax
4168 ; X86-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
4169 ; X86-NOBMI-NEXT: shrl %cl, %eax
4170 ; X86-NOBMI-NEXT: retl
4171 ;
4172 ; X86-BMI1-LABEL: bzhi64_32_d1:
4173 ; X86-BMI1: # %bb.0:
4174 ; X86-BMI1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
4175 ; X86-BMI1-NEXT: shll $8, %eax
4176 ; X86-BMI1-NEXT: bextrl %eax, {{[0-9]+}}(%esp), %eax
4177 ; X86-BMI1-NEXT: retl
4178 ;
4179 ; X86-BMI2-LABEL: bzhi64_32_d1:
4180 ; X86-BMI2: # %bb.0:
4181 ; X86-BMI2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
4182 ; X86-BMI2-NEXT: bzhil %eax, {{[0-9]+}}(%esp), %eax
4183 ; X86-BMI2-NEXT: retl
4184 ;
4185 ; X64-NOBMI-LABEL: bzhi64_32_d1:
4186 ; X64-NOBMI: # %bb.0:
4187 ; X64-NOBMI-NEXT: movl %esi, %ecx
4188 ; X64-NOBMI-NEXT: movq %rdi, %rax
4189 ; X64-NOBMI-NEXT: negb %cl
4190 ; X64-NOBMI-NEXT: shll %cl, %eax
4191 ; X64-NOBMI-NEXT: # kill: def $cl killed $cl killed $ecx
4192 ; X64-NOBMI-NEXT: shrl %cl, %eax
4193 ; X64-NOBMI-NEXT: # kill: def $eax killed $eax killed $rax
4194 ; X64-NOBMI-NEXT: retq
4195 ;
4196 ; X64-BMI1-LABEL: bzhi64_32_d1:
4197 ; X64-BMI1: # %bb.0:
4198 ; X64-BMI1-NEXT: shll $8, %esi
4199 ; X64-BMI1-NEXT: bextrl %esi, %edi, %eax
4200 ; X64-BMI1-NEXT: retq
4201 ;
4202 ; X64-BMI2-LABEL: bzhi64_32_d1:
4203 ; X64-BMI2: # %bb.0:
4204 ; X64-BMI2-NEXT: bzhil %esi, %edi, %eax
4205 ; X64-BMI2-NEXT: retq
4206 %truncval = trunc i64 %val to i32
4207 %numhighbits = sub i32 32, %numlowbits
4208 %highbitscleared = shl i32 %truncval, %numhighbits
4209 %masked = lshr i32 %highbitscleared, %numhighbits
4210 ret i32 %masked
4211 }
4213 ; ---------------------------------------------------------------------------- ;
4214 ; Constant mask
4215 ; ---------------------------------------------------------------------------- ;
4217 ; 32-bit
4219 define i32 @bzhi32_constant_mask32(i32 %val) nounwind {
4220 ; X86-LABEL: bzhi32_constant_mask32:
4221 ; X86: # %bb.0:
4222 ; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
4223 ; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
4224 ; X86-NEXT: retl
4225 ;
4226 ; X64-LABEL: bzhi32_constant_mask32:
4227 ; X64: # %bb.0:
4228 ; X64-NEXT: movl %edi, %eax
4229 ; X64-NEXT: andl $2147483647, %eax # imm = 0x7FFFFFFF
4230 ; X64-NEXT: retq
4231 %masked = and i32 %val, 2147483647
4232 ret i32 %masked
4233 }
4235 define i32 @bzhi32_constant_mask32_load(ptr %val) nounwind {
4236 ; X86-LABEL: bzhi32_constant_mask32_load:
4237 ; X86: # %bb.0:
4238 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
4239 ; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
4240 ; X86-NEXT: andl (%ecx), %eax
4241 ; X86-NEXT: retl
4242 ;
4243 ; X64-LABEL: bzhi32_constant_mask32_load:
4244 ; X64: # %bb.0:
4245 ; X64-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
4246 ; X64-NEXT: andl (%rdi), %eax
4247 ; X64-NEXT: retq
4248 %val1 = load i32, ptr %val
4249 %masked = and i32 %val1, 2147483647
4250 ret i32 %masked
4251 }
4253 define i32 @bzhi32_constant_mask16(i32 %val) nounwind {
4254 ; X86-LABEL: bzhi32_constant_mask16:
4255 ; X86: # %bb.0:
4256 ; X86-NEXT: movl $32767, %eax # imm = 0x7FFF
4257 ; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
4258 ; X86-NEXT: retl
4259 ;
4260 ; X64-LABEL: bzhi32_constant_mask16:
4261 ; X64: # %bb.0:
4262 ; X64-NEXT: movl %edi, %eax
4263 ; X64-NEXT: andl $32767, %eax # imm = 0x7FFF
4264 ; X64-NEXT: retq
4265 %masked = and i32 %val, 32767
4266 ret i32 %masked
4267 }
4269 define i32 @bzhi32_constant_mask16_load(ptr %val) nounwind {
4270 ; X86-LABEL: bzhi32_constant_mask16_load:
4271 ; X86: # %bb.0:
4272 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
4273 ; X86-NEXT: movl $32767, %eax # imm = 0x7FFF
4274 ; X86-NEXT: andl (%ecx), %eax
4275 ; X86-NEXT: retl
4276 ;
4277 ; X64-LABEL: bzhi32_constant_mask16_load:
4278 ; X64: # %bb.0:
4279 ; X64-NEXT: movl $32767, %eax # imm = 0x7FFF
4280 ; X64-NEXT: andl (%rdi), %eax
4281 ; X64-NEXT: retq
4282 %val1 = load i32, ptr %val
4283 %masked = and i32 %val1, 32767
4284 ret i32 %masked
4285 }
4287 define i32 @bzhi32_constant_mask8(i32 %val) nounwind {
4288 ; X86-LABEL: bzhi32_constant_mask8:
4289 ; X86: # %bb.0:
4290 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4291 ; X86-NEXT: andl $127, %eax
4292 ; X86-NEXT: retl
4293 ;
4294 ; X64-LABEL: bzhi32_constant_mask8:
4295 ; X64: # %bb.0:
4296 ; X64-NEXT: movl %edi, %eax
4297 ; X64-NEXT: andl $127, %eax
4298 ; X64-NEXT: retq
4299 %masked = and i32 %val, 127
4300 ret i32 %masked
4301 }
4303 define i32 @bzhi32_constant_mask8_load(ptr %val) nounwind {
4304 ; X86-LABEL: bzhi32_constant_mask8_load:
4305 ; X86: # %bb.0:
4306 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4307 ; X86-NEXT: movl (%eax), %eax
4308 ; X86-NEXT: andl $127, %eax
4309 ; X86-NEXT: retl
4310 ;
4311 ; X64-LABEL: bzhi32_constant_mask8_load:
4312 ; X64: # %bb.0:
4313 ; X64-NEXT: movl (%rdi), %eax
4314 ; X64-NEXT: andl $127, %eax
4315 ; X64-NEXT: retq
4316 %val1 = load i32, ptr %val
4317 %masked = and i32 %val1, 127
4318 ret i32 %masked
4319 }
4321 ; 64-bit
4323 define i64 @bzhi64_constant_mask64(i64 %val) nounwind {
4324 ; X86-LABEL: bzhi64_constant_mask64:
4325 ; X86: # %bb.0:
4326 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4327 ; X86-NEXT: movl $1073741823, %edx # imm = 0x3FFFFFFF
4328 ; X86-NEXT: andl {{[0-9]+}}(%esp), %edx
4329 ; X86-NEXT: retl
4330 ;
4331 ; X64-NOBMI-LABEL: bzhi64_constant_mask64:
4332 ; X64-NOBMI: # %bb.0:
4333 ; X64-NOBMI-NEXT: movabsq $4611686018427387903, %rax # imm = 0x3FFFFFFFFFFFFFFF
4334 ; X64-NOBMI-NEXT: andq %rdi, %rax
4335 ; X64-NOBMI-NEXT: retq
4336 ;
4337 ; X64-BMI1NOTBM-LABEL: bzhi64_constant_mask64:
4338 ; X64-BMI1NOTBM: # %bb.0:
4339 ; X64-BMI1NOTBM-NEXT: movl $15872, %eax # imm = 0x3E00
4340 ; X64-BMI1NOTBM-NEXT: bextrq %rax, %rdi, %rax
4341 ; X64-BMI1NOTBM-NEXT: retq
4342 ;
4343 ; X64-BMI1TBM-LABEL: bzhi64_constant_mask64:
4344 ; X64-BMI1TBM: # %bb.0:
4345 ; X64-BMI1TBM-NEXT: bextrq $15872, %rdi, %rax # imm = 0x3E00
4346 ; X64-BMI1TBM-NEXT: retq
4347 ;
4348 ; X64-BMI2TBM-LABEL: bzhi64_constant_mask64:
4349 ; X64-BMI2TBM: # %bb.0:
4350 ; X64-BMI2TBM-NEXT: bextrq $15872, %rdi, %rax # imm = 0x3E00
4351 ; X64-BMI2TBM-NEXT: retq
4352 ;
4353 ; X64-BMI2NOTBM-LABEL: bzhi64_constant_mask64:
4354 ; X64-BMI2NOTBM: # %bb.0:
4355 ; X64-BMI2NOTBM-NEXT: movb $62, %al
4356 ; X64-BMI2NOTBM-NEXT: bzhiq %rax, %rdi, %rax
4357 ; X64-BMI2NOTBM-NEXT: retq
4358 %masked = and i64 %val, 4611686018427387903
4359 ret i64 %masked
4360 }
4362 define i64 @bzhi64_constant_mask64_load(ptr %val) nounwind {
4363 ; X86-LABEL: bzhi64_constant_mask64_load:
4364 ; X86: # %bb.0:
4365 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
4366 ; X86-NEXT: movl (%ecx), %eax
4367 ; X86-NEXT: movl $1073741823, %edx # imm = 0x3FFFFFFF
4368 ; X86-NEXT: andl 4(%ecx), %edx
4369 ; X86-NEXT: retl
4370 ;
4371 ; X64-NOBMI-LABEL: bzhi64_constant_mask64_load:
4372 ; X64-NOBMI: # %bb.0:
4373 ; X64-NOBMI-NEXT: movabsq $4611686018427387903, %rax # imm = 0x3FFFFFFFFFFFFFFF
4374 ; X64-NOBMI-NEXT: andq (%rdi), %rax
4375 ; X64-NOBMI-NEXT: retq
4376 ;
4377 ; X64-BMI1NOTBM-LABEL: bzhi64_constant_mask64_load:
4378 ; X64-BMI1NOTBM: # %bb.0:
4379 ; X64-BMI1NOTBM-NEXT: movl $15872, %eax # imm = 0x3E00
4380 ; X64-BMI1NOTBM-NEXT: bextrq %rax, (%rdi), %rax
4381 ; X64-BMI1NOTBM-NEXT: retq
4382 ;
4383 ; X64-BMI1TBM-LABEL: bzhi64_constant_mask64_load:
4384 ; X64-BMI1TBM: # %bb.0:
4385 ; X64-BMI1TBM-NEXT: bextrq $15872, (%rdi), %rax # imm = 0x3E00
4386 ; X64-BMI1TBM-NEXT: retq
4387 ;
4388 ; X64-BMI2TBM-LABEL: bzhi64_constant_mask64_load:
4389 ; X64-BMI2TBM: # %bb.0:
4390 ; X64-BMI2TBM-NEXT: bextrq $15872, (%rdi), %rax # imm = 0x3E00
4391 ; X64-BMI2TBM-NEXT: retq
4392 ;
4393 ; X64-BMI2NOTBM-LABEL: bzhi64_constant_mask64_load:
4394 ; X64-BMI2NOTBM: # %bb.0:
4395 ; X64-BMI2NOTBM-NEXT: movb $62, %al
4396 ; X64-BMI2NOTBM-NEXT: bzhiq %rax, (%rdi), %rax
4397 ; X64-BMI2NOTBM-NEXT: retq
4398 %val1 = load i64, ptr %val
4399 %masked = and i64 %val1, 4611686018427387903
4400 ret i64 %masked
4401 }
4403 define i64 @bzhi64_constant_mask32(i64 %val) nounwind {
4404 ; X86-LABEL: bzhi64_constant_mask32:
4405 ; X86: # %bb.0:
4406 ; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
4407 ; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
4408 ; X86-NEXT: xorl %edx, %edx
4409 ; X86-NEXT: retl
4410 ;
4411 ; X64-LABEL: bzhi64_constant_mask32:
4412 ; X64: # %bb.0:
4413 ; X64-NEXT: movq %rdi, %rax
4414 ; X64-NEXT: andl $2147483647, %eax # imm = 0x7FFFFFFF
4415 ; X64-NEXT: retq
4416 %masked = and i64 %val, 2147483647
4417 ret i64 %masked
4418 }
4420 define i64 @bzhi64_constant_mask32_load(ptr %val) nounwind {
4421 ; X86-LABEL: bzhi64_constant_mask32_load:
4422 ; X86: # %bb.0:
4423 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
4424 ; X86-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
4425 ; X86-NEXT: andl (%ecx), %eax
4426 ; X86-NEXT: xorl %edx, %edx
4427 ; X86-NEXT: retl
4428 ;
4429 ; X64-LABEL: bzhi64_constant_mask32_load:
4430 ; X64: # %bb.0:
4431 ; X64-NEXT: movq (%rdi), %rax
4432 ; X64-NEXT: andl $2147483647, %eax # imm = 0x7FFFFFFF
4433 ; X64-NEXT: retq
4434 %val1 = load i64, ptr %val
4435 %masked = and i64 %val1, 2147483647
4436 ret i64 %masked
4437 }
4439 define i64 @bzhi64_constant_mask16(i64 %val) nounwind {
4440 ; X86-LABEL: bzhi64_constant_mask16:
4441 ; X86: # %bb.0:
4442 ; X86-NEXT: movl $32767, %eax # imm = 0x7FFF
4443 ; X86-NEXT: andl {{[0-9]+}}(%esp), %eax
4444 ; X86-NEXT: xorl %edx, %edx
4445 ; X86-NEXT: retl
4446 ;
4447 ; X64-LABEL: bzhi64_constant_mask16:
4448 ; X64: # %bb.0:
4449 ; X64-NEXT: movq %rdi, %rax
4450 ; X64-NEXT: andl $32767, %eax # imm = 0x7FFF
4451 ; X64-NEXT: retq
4452 %masked = and i64 %val, 32767
4453 ret i64 %masked
4454 }
4456 define i64 @bzhi64_constant_mask16_load(ptr %val) nounwind {
4457 ; X86-LABEL: bzhi64_constant_mask16_load:
4458 ; X86: # %bb.0:
4459 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
4460 ; X86-NEXT: movl $32767, %eax # imm = 0x7FFF
4461 ; X86-NEXT: andl (%ecx), %eax
4462 ; X86-NEXT: xorl %edx, %edx
4463 ; X86-NEXT: retl
4464 ;
4465 ; X64-LABEL: bzhi64_constant_mask16_load:
4466 ; X64: # %bb.0:
4467 ; X64-NEXT: movq (%rdi), %rax
4468 ; X64-NEXT: andl $32767, %eax # imm = 0x7FFF
4469 ; X64-NEXT: retq
4470 %val1 = load i64, ptr %val
4471 %masked = and i64 %val1, 32767
4472 ret i64 %masked
4473 }
4475 define i64 @bzhi64_constant_mask8(i64 %val) nounwind {
4476 ; X86-LABEL: bzhi64_constant_mask8:
4477 ; X86: # %bb.0:
4478 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4479 ; X86-NEXT: andl $127, %eax
4480 ; X86-NEXT: xorl %edx, %edx
4481 ; X86-NEXT: retl
4482 ;
4483 ; X64-LABEL: bzhi64_constant_mask8:
4484 ; X64: # %bb.0:
4485 ; X64-NEXT: movq %rdi, %rax
4486 ; X64-NEXT: andl $127, %eax
4487 ; X64-NEXT: retq
4488 %masked = and i64 %val, 127
4489 ret i64 %masked
4490 }
4492 define i64 @bzhi64_constant_mask8_load(ptr %val) nounwind {
4493 ; X86-LABEL: bzhi64_constant_mask8_load:
4494 ; X86: # %bb.0:
4495 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4496 ; X86-NEXT: movl (%eax), %eax
4497 ; X86-NEXT: andl $127, %eax
4498 ; X86-NEXT: xorl %edx, %edx
4499 ; X86-NEXT: retl
4500 ;
4501 ; X64-LABEL: bzhi64_constant_mask8_load:
4502 ; X64: # %bb.0:
4503 ; X64-NEXT: movq (%rdi), %rax
4504 ; X64-NEXT: andl $127, %eax
4505 ; X64-NEXT: retq
4506 %val1 = load i64, ptr %val
4507 %masked = and i64 %val1, 127
4508 ret i64 %masked
4509 }