1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=cmov | FileCheck %s --check-prefixes=X86,X86-NOSSE
3 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse | FileCheck %s --check-prefixes=X86,SSE,X86-SSE1
4 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=X86,SSE,X86-SSE2
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefixes=X64,X64-SSE2
6 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx | FileCheck %s --check-prefixes=X64,X64-AVX,X64-AVX1
7 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2 | FileCheck %s --check-prefixes=X64,X64-AVX,X64-AVX2
8 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512f | FileCheck %s --check-prefixes=X64,X64-AVX512,X64-AVX512F
9 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512bw | FileCheck %s --check-prefixes=X64,X64-AVX512,X64-AVX512BW
11 ; This tests codegen time inlining/optimization of memcmp
14 @.str = private constant [65 x i8] c"0123456789012345678901234567890123456789012345678901234567890123\00", align 1
16 declare i32 @memcmp(i8*, i8*, i64)
18 define i32 @length0(i8* %X, i8* %Y) nounwind {
21 ; X86-NEXT: xorl %eax, %eax
26 ; X64-NEXT: xorl %eax, %eax
28 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 0) nounwind
32 define i1 @length0_eq(i8* %X, i8* %Y) nounwind {
33 ; X86-LABEL: length0_eq:
35 ; X86-NEXT: movb $1, %al
38 ; X64-LABEL: length0_eq:
40 ; X64-NEXT: movb $1, %al
42 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 0) nounwind
43 %c = icmp eq i32 %m, 0
47 define i1 @length0_lt(i8* %X, i8* %Y) nounwind {
48 ; X86-LABEL: length0_lt:
50 ; X86-NEXT: xorl %eax, %eax
53 ; X64-LABEL: length0_lt:
55 ; X64-NEXT: xorl %eax, %eax
57 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 0) nounwind
58 %c = icmp slt i32 %m, 0
62 define i32 @length2(i8* %X, i8* %Y) nounwind {
65 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
66 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
67 ; X86-NEXT: movzwl (%ecx), %ecx
68 ; X86-NEXT: movzwl (%eax), %edx
69 ; X86-NEXT: rolw $8, %cx
70 ; X86-NEXT: rolw $8, %dx
71 ; X86-NEXT: movzwl %cx, %eax
72 ; X86-NEXT: movzwl %dx, %ecx
73 ; X86-NEXT: subl %ecx, %eax
78 ; X64-NEXT: movzwl (%rdi), %eax
79 ; X64-NEXT: movzwl (%rsi), %ecx
80 ; X64-NEXT: rolw $8, %ax
81 ; X64-NEXT: rolw $8, %cx
82 ; X64-NEXT: movzwl %ax, %eax
83 ; X64-NEXT: movzwl %cx, %ecx
84 ; X64-NEXT: subl %ecx, %eax
86 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 2) nounwind
90 define i1 @length2_eq(i8* %X, i8* %Y) nounwind {
91 ; X86-LABEL: length2_eq:
93 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
94 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
95 ; X86-NEXT: movzwl (%ecx), %ecx
96 ; X86-NEXT: cmpw (%eax), %cx
100 ; X64-LABEL: length2_eq:
102 ; X64-NEXT: movzwl (%rdi), %eax
103 ; X64-NEXT: cmpw (%rsi), %ax
106 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 2) nounwind
107 %c = icmp eq i32 %m, 0
111 define i1 @length2_lt(i8* %X, i8* %Y) nounwind {
112 ; X86-LABEL: length2_lt:
114 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
115 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
116 ; X86-NEXT: movzwl (%ecx), %ecx
117 ; X86-NEXT: movzwl (%eax), %edx
118 ; X86-NEXT: rolw $8, %cx
119 ; X86-NEXT: rolw $8, %dx
120 ; X86-NEXT: movzwl %cx, %eax
121 ; X86-NEXT: movzwl %dx, %ecx
122 ; X86-NEXT: subl %ecx, %eax
123 ; X86-NEXT: shrl $31, %eax
124 ; X86-NEXT: # kill: def $al killed $al killed $eax
127 ; X64-LABEL: length2_lt:
129 ; X64-NEXT: movzwl (%rdi), %eax
130 ; X64-NEXT: movzwl (%rsi), %ecx
131 ; X64-NEXT: rolw $8, %ax
132 ; X64-NEXT: rolw $8, %cx
133 ; X64-NEXT: movzwl %ax, %eax
134 ; X64-NEXT: movzwl %cx, %ecx
135 ; X64-NEXT: subl %ecx, %eax
136 ; X64-NEXT: shrl $31, %eax
137 ; X64-NEXT: # kill: def $al killed $al killed $eax
139 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 2) nounwind
140 %c = icmp slt i32 %m, 0
144 define i1 @length2_gt(i8* %X, i8* %Y) nounwind {
145 ; X86-LABEL: length2_gt:
147 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
148 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
149 ; X86-NEXT: movzwl (%ecx), %ecx
150 ; X86-NEXT: movzwl (%eax), %eax
151 ; X86-NEXT: rolw $8, %cx
152 ; X86-NEXT: rolw $8, %ax
153 ; X86-NEXT: movzwl %cx, %ecx
154 ; X86-NEXT: movzwl %ax, %eax
155 ; X86-NEXT: subl %eax, %ecx
156 ; X86-NEXT: testl %ecx, %ecx
160 ; X64-LABEL: length2_gt:
162 ; X64-NEXT: movzwl (%rdi), %eax
163 ; X64-NEXT: movzwl (%rsi), %ecx
164 ; X64-NEXT: rolw $8, %ax
165 ; X64-NEXT: rolw $8, %cx
166 ; X64-NEXT: movzwl %ax, %eax
167 ; X64-NEXT: movzwl %cx, %ecx
168 ; X64-NEXT: subl %ecx, %eax
169 ; X64-NEXT: testl %eax, %eax
172 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 2) nounwind
173 %c = icmp sgt i32 %m, 0
177 define i1 @length2_eq_const(i8* %X) nounwind {
178 ; X86-LABEL: length2_eq_const:
180 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
181 ; X86-NEXT: movzwl (%eax), %eax
182 ; X86-NEXT: cmpl $12849, %eax # imm = 0x3231
183 ; X86-NEXT: setne %al
186 ; X64-LABEL: length2_eq_const:
188 ; X64-NEXT: movzwl (%rdi), %eax
189 ; X64-NEXT: cmpl $12849, %eax # imm = 0x3231
190 ; X64-NEXT: setne %al
192 %m = tail call i32 @memcmp(i8* %X, i8* getelementptr inbounds ([65 x i8], [65 x i8]* @.str, i32 0, i32 1), i64 2) nounwind
193 %c = icmp ne i32 %m, 0
197 define i1 @length2_eq_nobuiltin_attr(i8* %X, i8* %Y) nounwind {
198 ; X86-LABEL: length2_eq_nobuiltin_attr:
202 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
203 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
204 ; X86-NEXT: calll memcmp
205 ; X86-NEXT: addl $16, %esp
206 ; X86-NEXT: testl %eax, %eax
210 ; X64-LABEL: length2_eq_nobuiltin_attr:
212 ; X64-NEXT: pushq %rax
213 ; X64-NEXT: movl $2, %edx
214 ; X64-NEXT: callq memcmp
215 ; X64-NEXT: testl %eax, %eax
217 ; X64-NEXT: popq %rcx
219 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 2) nounwind nobuiltin
220 %c = icmp eq i32 %m, 0
224 define i32 @length3(i8* %X, i8* %Y) nounwind {
225 ; X86-LABEL: length3:
226 ; X86: # %bb.0: # %loadbb
227 ; X86-NEXT: pushl %esi
228 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
229 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
230 ; X86-NEXT: movzwl (%eax), %edx
231 ; X86-NEXT: movzwl (%ecx), %esi
232 ; X86-NEXT: rolw $8, %dx
233 ; X86-NEXT: rolw $8, %si
234 ; X86-NEXT: cmpw %si, %dx
235 ; X86-NEXT: jne .LBB9_1
236 ; X86-NEXT: # %bb.2: # %loadbb1
237 ; X86-NEXT: movzbl 2(%eax), %eax
238 ; X86-NEXT: movzbl 2(%ecx), %ecx
239 ; X86-NEXT: subl %ecx, %eax
240 ; X86-NEXT: popl %esi
242 ; X86-NEXT: .LBB9_1: # %res_block
243 ; X86-NEXT: setae %al
244 ; X86-NEXT: movzbl %al, %eax
245 ; X86-NEXT: leal -1(%eax,%eax), %eax
246 ; X86-NEXT: popl %esi
249 ; X64-LABEL: length3:
250 ; X64: # %bb.0: # %loadbb
251 ; X64-NEXT: movzwl (%rdi), %eax
252 ; X64-NEXT: movzwl (%rsi), %ecx
253 ; X64-NEXT: rolw $8, %ax
254 ; X64-NEXT: rolw $8, %cx
255 ; X64-NEXT: cmpw %cx, %ax
256 ; X64-NEXT: jne .LBB9_1
257 ; X64-NEXT: # %bb.2: # %loadbb1
258 ; X64-NEXT: movzbl 2(%rdi), %eax
259 ; X64-NEXT: movzbl 2(%rsi), %ecx
260 ; X64-NEXT: subl %ecx, %eax
262 ; X64-NEXT: .LBB9_1: # %res_block
263 ; X64-NEXT: setae %al
264 ; X64-NEXT: movzbl %al, %eax
265 ; X64-NEXT: leal -1(%rax,%rax), %eax
267 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 3) nounwind
271 define i1 @length3_eq(i8* %X, i8* %Y) nounwind {
272 ; X86-LABEL: length3_eq:
274 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
275 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
276 ; X86-NEXT: movzwl (%ecx), %edx
277 ; X86-NEXT: xorw (%eax), %dx
278 ; X86-NEXT: movb 2(%ecx), %cl
279 ; X86-NEXT: xorb 2(%eax), %cl
280 ; X86-NEXT: movzbl %cl, %eax
281 ; X86-NEXT: orw %dx, %ax
282 ; X86-NEXT: setne %al
285 ; X64-LABEL: length3_eq:
287 ; X64-NEXT: movzwl (%rdi), %eax
288 ; X64-NEXT: xorw (%rsi), %ax
289 ; X64-NEXT: movb 2(%rdi), %cl
290 ; X64-NEXT: xorb 2(%rsi), %cl
291 ; X64-NEXT: movzbl %cl, %ecx
292 ; X64-NEXT: orw %ax, %cx
293 ; X64-NEXT: setne %al
295 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 3) nounwind
296 %c = icmp ne i32 %m, 0
300 define i32 @length4(i8* %X, i8* %Y) nounwind {
301 ; X86-LABEL: length4:
303 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
304 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
305 ; X86-NEXT: movl (%ecx), %ecx
306 ; X86-NEXT: movl (%eax), %edx
307 ; X86-NEXT: bswapl %ecx
308 ; X86-NEXT: bswapl %edx
309 ; X86-NEXT: xorl %eax, %eax
310 ; X86-NEXT: cmpl %edx, %ecx
312 ; X86-NEXT: sbbl $0, %eax
315 ; X64-LABEL: length4:
317 ; X64-NEXT: movl (%rdi), %ecx
318 ; X64-NEXT: movl (%rsi), %edx
319 ; X64-NEXT: bswapl %ecx
320 ; X64-NEXT: bswapl %edx
321 ; X64-NEXT: xorl %eax, %eax
322 ; X64-NEXT: cmpl %edx, %ecx
324 ; X64-NEXT: sbbl $0, %eax
326 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 4) nounwind
330 define i1 @length4_eq(i8* %X, i8* %Y) nounwind {
331 ; X86-LABEL: length4_eq:
333 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
334 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
335 ; X86-NEXT: movl (%ecx), %ecx
336 ; X86-NEXT: cmpl (%eax), %ecx
337 ; X86-NEXT: setne %al
340 ; X64-LABEL: length4_eq:
342 ; X64-NEXT: movl (%rdi), %eax
343 ; X64-NEXT: cmpl (%rsi), %eax
344 ; X64-NEXT: setne %al
346 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 4) nounwind
347 %c = icmp ne i32 %m, 0
351 define i1 @length4_lt(i8* %X, i8* %Y) nounwind {
352 ; X86-LABEL: length4_lt:
354 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
355 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
356 ; X86-NEXT: movl (%ecx), %ecx
357 ; X86-NEXT: movl (%eax), %edx
358 ; X86-NEXT: bswapl %ecx
359 ; X86-NEXT: bswapl %edx
360 ; X86-NEXT: xorl %eax, %eax
361 ; X86-NEXT: cmpl %edx, %ecx
363 ; X86-NEXT: sbbl $0, %eax
364 ; X86-NEXT: shrl $31, %eax
365 ; X86-NEXT: # kill: def $al killed $al killed $eax
368 ; X64-LABEL: length4_lt:
370 ; X64-NEXT: movl (%rdi), %ecx
371 ; X64-NEXT: movl (%rsi), %edx
372 ; X64-NEXT: bswapl %ecx
373 ; X64-NEXT: bswapl %edx
374 ; X64-NEXT: xorl %eax, %eax
375 ; X64-NEXT: cmpl %edx, %ecx
377 ; X64-NEXT: sbbl $0, %eax
378 ; X64-NEXT: shrl $31, %eax
379 ; X64-NEXT: # kill: def $al killed $al killed $eax
381 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 4) nounwind
382 %c = icmp slt i32 %m, 0
386 define i1 @length4_gt(i8* %X, i8* %Y) nounwind {
387 ; X86-LABEL: length4_gt:
389 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
390 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
391 ; X86-NEXT: movl (%ecx), %ecx
392 ; X86-NEXT: movl (%eax), %eax
393 ; X86-NEXT: bswapl %ecx
394 ; X86-NEXT: bswapl %eax
395 ; X86-NEXT: xorl %edx, %edx
396 ; X86-NEXT: cmpl %eax, %ecx
398 ; X86-NEXT: sbbl $0, %edx
399 ; X86-NEXT: testl %edx, %edx
403 ; X64-LABEL: length4_gt:
405 ; X64-NEXT: movl (%rdi), %eax
406 ; X64-NEXT: movl (%rsi), %ecx
407 ; X64-NEXT: bswapl %eax
408 ; X64-NEXT: bswapl %ecx
409 ; X64-NEXT: xorl %edx, %edx
410 ; X64-NEXT: cmpl %ecx, %eax
412 ; X64-NEXT: sbbl $0, %edx
413 ; X64-NEXT: testl %edx, %edx
416 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 4) nounwind
417 %c = icmp sgt i32 %m, 0
421 define i1 @length4_eq_const(i8* %X) nounwind {
422 ; X86-LABEL: length4_eq_const:
424 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
425 ; X86-NEXT: cmpl $875770417, (%eax) # imm = 0x34333231
429 ; X64-LABEL: length4_eq_const:
431 ; X64-NEXT: cmpl $875770417, (%rdi) # imm = 0x34333231
434 %m = tail call i32 @memcmp(i8* %X, i8* getelementptr inbounds ([65 x i8], [65 x i8]* @.str, i32 0, i32 1), i64 4) nounwind
435 %c = icmp eq i32 %m, 0
439 define i32 @length5(i8* %X, i8* %Y) nounwind {
440 ; X86-LABEL: length5:
441 ; X86: # %bb.0: # %loadbb
442 ; X86-NEXT: pushl %esi
443 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
444 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
445 ; X86-NEXT: movl (%eax), %edx
446 ; X86-NEXT: movl (%ecx), %esi
447 ; X86-NEXT: bswapl %edx
448 ; X86-NEXT: bswapl %esi
449 ; X86-NEXT: cmpl %esi, %edx
450 ; X86-NEXT: jne .LBB16_1
451 ; X86-NEXT: # %bb.2: # %loadbb1
452 ; X86-NEXT: movzbl 4(%eax), %eax
453 ; X86-NEXT: movzbl 4(%ecx), %ecx
454 ; X86-NEXT: subl %ecx, %eax
455 ; X86-NEXT: popl %esi
457 ; X86-NEXT: .LBB16_1: # %res_block
458 ; X86-NEXT: setae %al
459 ; X86-NEXT: movzbl %al, %eax
460 ; X86-NEXT: leal -1(%eax,%eax), %eax
461 ; X86-NEXT: popl %esi
464 ; X64-LABEL: length5:
465 ; X64: # %bb.0: # %loadbb
466 ; X64-NEXT: movl (%rdi), %eax
467 ; X64-NEXT: movl (%rsi), %ecx
468 ; X64-NEXT: bswapl %eax
469 ; X64-NEXT: bswapl %ecx
470 ; X64-NEXT: cmpl %ecx, %eax
471 ; X64-NEXT: jne .LBB16_1
472 ; X64-NEXT: # %bb.2: # %loadbb1
473 ; X64-NEXT: movzbl 4(%rdi), %eax
474 ; X64-NEXT: movzbl 4(%rsi), %ecx
475 ; X64-NEXT: subl %ecx, %eax
477 ; X64-NEXT: .LBB16_1: # %res_block
478 ; X64-NEXT: setae %al
479 ; X64-NEXT: movzbl %al, %eax
480 ; X64-NEXT: leal -1(%rax,%rax), %eax
482 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 5) nounwind
486 define i1 @length5_eq(i8* %X, i8* %Y) nounwind {
487 ; X86-LABEL: length5_eq:
489 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
490 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
491 ; X86-NEXT: movl (%ecx), %edx
492 ; X86-NEXT: xorl (%eax), %edx
493 ; X86-NEXT: movb 4(%ecx), %cl
494 ; X86-NEXT: xorb 4(%eax), %cl
495 ; X86-NEXT: movzbl %cl, %eax
496 ; X86-NEXT: orl %edx, %eax
497 ; X86-NEXT: setne %al
500 ; X64-LABEL: length5_eq:
502 ; X64-NEXT: movl (%rdi), %eax
503 ; X64-NEXT: xorl (%rsi), %eax
504 ; X64-NEXT: movb 4(%rdi), %cl
505 ; X64-NEXT: xorb 4(%rsi), %cl
506 ; X64-NEXT: movzbl %cl, %ecx
507 ; X64-NEXT: orl %eax, %ecx
508 ; X64-NEXT: setne %al
510 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 5) nounwind
511 %c = icmp ne i32 %m, 0
515 define i1 @length5_lt(i8* %X, i8* %Y) nounwind {
516 ; X86-LABEL: length5_lt:
517 ; X86: # %bb.0: # %loadbb
518 ; X86-NEXT: pushl %esi
519 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
520 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
521 ; X86-NEXT: movl (%eax), %edx
522 ; X86-NEXT: movl (%ecx), %esi
523 ; X86-NEXT: bswapl %edx
524 ; X86-NEXT: bswapl %esi
525 ; X86-NEXT: cmpl %esi, %edx
526 ; X86-NEXT: jne .LBB18_1
527 ; X86-NEXT: # %bb.2: # %loadbb1
528 ; X86-NEXT: movzbl 4(%eax), %eax
529 ; X86-NEXT: movzbl 4(%ecx), %ecx
530 ; X86-NEXT: subl %ecx, %eax
531 ; X86-NEXT: jmp .LBB18_3
532 ; X86-NEXT: .LBB18_1: # %res_block
533 ; X86-NEXT: setae %al
534 ; X86-NEXT: movzbl %al, %eax
535 ; X86-NEXT: leal -1(%eax,%eax), %eax
536 ; X86-NEXT: .LBB18_3: # %endblock
537 ; X86-NEXT: shrl $31, %eax
538 ; X86-NEXT: # kill: def $al killed $al killed $eax
539 ; X86-NEXT: popl %esi
542 ; X64-LABEL: length5_lt:
543 ; X64: # %bb.0: # %loadbb
544 ; X64-NEXT: movl (%rdi), %eax
545 ; X64-NEXT: movl (%rsi), %ecx
546 ; X64-NEXT: bswapl %eax
547 ; X64-NEXT: bswapl %ecx
548 ; X64-NEXT: cmpl %ecx, %eax
549 ; X64-NEXT: jne .LBB18_1
550 ; X64-NEXT: # %bb.2: # %loadbb1
551 ; X64-NEXT: movzbl 4(%rdi), %eax
552 ; X64-NEXT: movzbl 4(%rsi), %ecx
553 ; X64-NEXT: subl %ecx, %eax
554 ; X64-NEXT: shrl $31, %eax
555 ; X64-NEXT: # kill: def $al killed $al killed $eax
557 ; X64-NEXT: .LBB18_1: # %res_block
558 ; X64-NEXT: setae %al
559 ; X64-NEXT: movzbl %al, %eax
560 ; X64-NEXT: leal -1(%rax,%rax), %eax
561 ; X64-NEXT: shrl $31, %eax
562 ; X64-NEXT: # kill: def $al killed $al killed $eax
564 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 5) nounwind
565 %c = icmp slt i32 %m, 0
569 define i1 @length7_eq(i8* %X, i8* %Y) nounwind {
570 ; X86-LABEL: length7_eq:
572 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
573 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
574 ; X86-NEXT: movl (%ecx), %edx
575 ; X86-NEXT: movl 3(%ecx), %ecx
576 ; X86-NEXT: xorl (%eax), %edx
577 ; X86-NEXT: xorl 3(%eax), %ecx
578 ; X86-NEXT: orl %edx, %ecx
579 ; X86-NEXT: setne %al
582 ; X64-LABEL: length7_eq:
584 ; X64-NEXT: movl (%rdi), %eax
585 ; X64-NEXT: movl 3(%rdi), %ecx
586 ; X64-NEXT: xorl (%rsi), %eax
587 ; X64-NEXT: xorl 3(%rsi), %ecx
588 ; X64-NEXT: orl %eax, %ecx
589 ; X64-NEXT: setne %al
591 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 7) nounwind
592 %c = icmp ne i32 %m, 0
596 define i32 @length8(i8* %X, i8* %Y) nounwind {
597 ; X86-LABEL: length8:
599 ; X86-NEXT: pushl %esi
600 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
601 ; X86-NEXT: movl {{[0-9]+}}(%esp), %esi
602 ; X86-NEXT: movl (%esi), %ecx
603 ; X86-NEXT: movl (%eax), %edx
604 ; X86-NEXT: bswapl %ecx
605 ; X86-NEXT: bswapl %edx
606 ; X86-NEXT: cmpl %edx, %ecx
607 ; X86-NEXT: jne .LBB20_2
608 ; X86-NEXT: # %bb.1: # %loadbb1
609 ; X86-NEXT: movl 4(%esi), %ecx
610 ; X86-NEXT: movl 4(%eax), %edx
611 ; X86-NEXT: bswapl %ecx
612 ; X86-NEXT: bswapl %edx
613 ; X86-NEXT: xorl %eax, %eax
614 ; X86-NEXT: cmpl %edx, %ecx
615 ; X86-NEXT: je .LBB20_3
616 ; X86-NEXT: .LBB20_2: # %res_block
617 ; X86-NEXT: xorl %eax, %eax
618 ; X86-NEXT: cmpl %edx, %ecx
619 ; X86-NEXT: setae %al
620 ; X86-NEXT: leal -1(%eax,%eax), %eax
621 ; X86-NEXT: .LBB20_3: # %endblock
622 ; X86-NEXT: popl %esi
625 ; X64-LABEL: length8:
627 ; X64-NEXT: movq (%rdi), %rcx
628 ; X64-NEXT: movq (%rsi), %rdx
629 ; X64-NEXT: bswapq %rcx
630 ; X64-NEXT: bswapq %rdx
631 ; X64-NEXT: xorl %eax, %eax
632 ; X64-NEXT: cmpq %rdx, %rcx
634 ; X64-NEXT: sbbl $0, %eax
636 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 8) nounwind
640 define i1 @length8_eq(i8* %X, i8* %Y) nounwind {
641 ; X86-LABEL: length8_eq:
643 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
644 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
645 ; X86-NEXT: movl (%ecx), %edx
646 ; X86-NEXT: movl 4(%ecx), %ecx
647 ; X86-NEXT: xorl (%eax), %edx
648 ; X86-NEXT: xorl 4(%eax), %ecx
649 ; X86-NEXT: orl %edx, %ecx
653 ; X64-LABEL: length8_eq:
655 ; X64-NEXT: movq (%rdi), %rax
656 ; X64-NEXT: cmpq (%rsi), %rax
659 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 8) nounwind
660 %c = icmp eq i32 %m, 0
664 define i1 @length8_eq_const(i8* %X) nounwind {
665 ; X86-LABEL: length8_eq_const:
667 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
668 ; X86-NEXT: movl $858927408, %ecx # imm = 0x33323130
669 ; X86-NEXT: xorl (%eax), %ecx
670 ; X86-NEXT: movl $926299444, %edx # imm = 0x37363534
671 ; X86-NEXT: xorl 4(%eax), %edx
672 ; X86-NEXT: orl %ecx, %edx
673 ; X86-NEXT: setne %al
676 ; X64-LABEL: length8_eq_const:
678 ; X64-NEXT: movabsq $3978425819141910832, %rax # imm = 0x3736353433323130
679 ; X64-NEXT: cmpq %rax, (%rdi)
680 ; X64-NEXT: setne %al
682 %m = tail call i32 @memcmp(i8* %X, i8* getelementptr inbounds ([65 x i8], [65 x i8]* @.str, i32 0, i32 0), i64 8) nounwind
683 %c = icmp ne i32 %m, 0
687 define i1 @length9_eq(i8* %X, i8* %Y) nounwind {
688 ; X86-LABEL: length9_eq:
692 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
693 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
694 ; X86-NEXT: calll memcmp
695 ; X86-NEXT: addl $16, %esp
696 ; X86-NEXT: testl %eax, %eax
700 ; X64-LABEL: length9_eq:
702 ; X64-NEXT: movq (%rdi), %rax
703 ; X64-NEXT: xorq (%rsi), %rax
704 ; X64-NEXT: movb 8(%rdi), %cl
705 ; X64-NEXT: xorb 8(%rsi), %cl
706 ; X64-NEXT: movzbl %cl, %ecx
707 ; X64-NEXT: orq %rax, %rcx
710 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 9) nounwind
711 %c = icmp eq i32 %m, 0
715 define i1 @length10_eq(i8* %X, i8* %Y) nounwind {
716 ; X86-LABEL: length10_eq:
719 ; X86-NEXT: pushl $10
720 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
721 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
722 ; X86-NEXT: calll memcmp
723 ; X86-NEXT: addl $16, %esp
724 ; X86-NEXT: testl %eax, %eax
728 ; X64-LABEL: length10_eq:
730 ; X64-NEXT: movq (%rdi), %rax
731 ; X64-NEXT: xorq (%rsi), %rax
732 ; X64-NEXT: movzwl 8(%rdi), %ecx
733 ; X64-NEXT: xorw 8(%rsi), %cx
734 ; X64-NEXT: movzwl %cx, %ecx
735 ; X64-NEXT: orq %rax, %rcx
738 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 10) nounwind
739 %c = icmp eq i32 %m, 0
743 define i1 @length11_eq(i8* %X, i8* %Y) nounwind {
744 ; X86-LABEL: length11_eq:
747 ; X86-NEXT: pushl $11
748 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
749 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
750 ; X86-NEXT: calll memcmp
751 ; X86-NEXT: addl $16, %esp
752 ; X86-NEXT: testl %eax, %eax
756 ; X64-LABEL: length11_eq:
758 ; X64-NEXT: movq (%rdi), %rax
759 ; X64-NEXT: movq 3(%rdi), %rcx
760 ; X64-NEXT: xorq (%rsi), %rax
761 ; X64-NEXT: xorq 3(%rsi), %rcx
762 ; X64-NEXT: orq %rax, %rcx
765 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 11) nounwind
766 %c = icmp eq i32 %m, 0
770 define i1 @length12_eq(i8* %X, i8* %Y) nounwind {
771 ; X86-LABEL: length12_eq:
774 ; X86-NEXT: pushl $12
775 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
776 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
777 ; X86-NEXT: calll memcmp
778 ; X86-NEXT: addl $16, %esp
779 ; X86-NEXT: testl %eax, %eax
780 ; X86-NEXT: setne %al
783 ; X64-LABEL: length12_eq:
785 ; X64-NEXT: movq (%rdi), %rax
786 ; X64-NEXT: xorq (%rsi), %rax
787 ; X64-NEXT: movl 8(%rdi), %ecx
788 ; X64-NEXT: xorl 8(%rsi), %ecx
789 ; X64-NEXT: orq %rax, %rcx
790 ; X64-NEXT: setne %al
792 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 12) nounwind
793 %c = icmp ne i32 %m, 0
797 define i32 @length12(i8* %X, i8* %Y) nounwind {
798 ; X86-LABEL: length12:
801 ; X86-NEXT: pushl $12
802 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
803 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
804 ; X86-NEXT: calll memcmp
805 ; X86-NEXT: addl $16, %esp
808 ; X64-LABEL: length12:
810 ; X64-NEXT: movq (%rdi), %rcx
811 ; X64-NEXT: movq (%rsi), %rdx
812 ; X64-NEXT: bswapq %rcx
813 ; X64-NEXT: bswapq %rdx
814 ; X64-NEXT: cmpq %rdx, %rcx
815 ; X64-NEXT: jne .LBB27_2
816 ; X64-NEXT: # %bb.1: # %loadbb1
817 ; X64-NEXT: movl 8(%rdi), %ecx
818 ; X64-NEXT: movl 8(%rsi), %edx
819 ; X64-NEXT: bswapl %ecx
820 ; X64-NEXT: bswapl %edx
821 ; X64-NEXT: xorl %eax, %eax
822 ; X64-NEXT: cmpq %rdx, %rcx
823 ; X64-NEXT: je .LBB27_3
824 ; X64-NEXT: .LBB27_2: # %res_block
825 ; X64-NEXT: xorl %eax, %eax
826 ; X64-NEXT: cmpq %rdx, %rcx
827 ; X64-NEXT: setae %al
828 ; X64-NEXT: leal -1(%rax,%rax), %eax
829 ; X64-NEXT: .LBB27_3: # %endblock
831 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 12) nounwind
835 define i1 @length13_eq(i8* %X, i8* %Y) nounwind {
836 ; X86-LABEL: length13_eq:
839 ; X86-NEXT: pushl $13
840 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
841 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
842 ; X86-NEXT: calll memcmp
843 ; X86-NEXT: addl $16, %esp
844 ; X86-NEXT: testl %eax, %eax
848 ; X64-LABEL: length13_eq:
850 ; X64-NEXT: movq (%rdi), %rax
851 ; X64-NEXT: movq 5(%rdi), %rcx
852 ; X64-NEXT: xorq (%rsi), %rax
853 ; X64-NEXT: xorq 5(%rsi), %rcx
854 ; X64-NEXT: orq %rax, %rcx
857 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 13) nounwind
858 %c = icmp eq i32 %m, 0
862 define i1 @length14_eq(i8* %X, i8* %Y) nounwind {
863 ; X86-LABEL: length14_eq:
866 ; X86-NEXT: pushl $14
867 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
868 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
869 ; X86-NEXT: calll memcmp
870 ; X86-NEXT: addl $16, %esp
871 ; X86-NEXT: testl %eax, %eax
875 ; X64-LABEL: length14_eq:
877 ; X64-NEXT: movq (%rdi), %rax
878 ; X64-NEXT: movq 6(%rdi), %rcx
879 ; X64-NEXT: xorq (%rsi), %rax
880 ; X64-NEXT: xorq 6(%rsi), %rcx
881 ; X64-NEXT: orq %rax, %rcx
884 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 14) nounwind
885 %c = icmp eq i32 %m, 0
889 define i1 @length15_eq(i8* %X, i8* %Y) nounwind {
890 ; X86-LABEL: length15_eq:
893 ; X86-NEXT: pushl $15
894 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
895 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
896 ; X86-NEXT: calll memcmp
897 ; X86-NEXT: addl $16, %esp
898 ; X86-NEXT: testl %eax, %eax
902 ; X64-LABEL: length15_eq:
904 ; X64-NEXT: movq (%rdi), %rax
905 ; X64-NEXT: movq 7(%rdi), %rcx
906 ; X64-NEXT: xorq (%rsi), %rax
907 ; X64-NEXT: xorq 7(%rsi), %rcx
908 ; X64-NEXT: orq %rax, %rcx
911 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 15) nounwind
912 %c = icmp eq i32 %m, 0
916 ; PR33329 - https://bugs.llvm.org/show_bug.cgi?id=33329
918 define i32 @length16(i8* %X, i8* %Y) nounwind {
919 ; X86-LABEL: length16:
922 ; X86-NEXT: pushl $16
923 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
924 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
925 ; X86-NEXT: calll memcmp
926 ; X86-NEXT: addl $16, %esp
929 ; X64-LABEL: length16:
931 ; X64-NEXT: movq (%rdi), %rcx
932 ; X64-NEXT: movq (%rsi), %rdx
933 ; X64-NEXT: bswapq %rcx
934 ; X64-NEXT: bswapq %rdx
935 ; X64-NEXT: cmpq %rdx, %rcx
936 ; X64-NEXT: jne .LBB31_2
937 ; X64-NEXT: # %bb.1: # %loadbb1
938 ; X64-NEXT: movq 8(%rdi), %rcx
939 ; X64-NEXT: movq 8(%rsi), %rdx
940 ; X64-NEXT: bswapq %rcx
941 ; X64-NEXT: bswapq %rdx
942 ; X64-NEXT: xorl %eax, %eax
943 ; X64-NEXT: cmpq %rdx, %rcx
944 ; X64-NEXT: je .LBB31_3
945 ; X64-NEXT: .LBB31_2: # %res_block
946 ; X64-NEXT: xorl %eax, %eax
947 ; X64-NEXT: cmpq %rdx, %rcx
948 ; X64-NEXT: setae %al
949 ; X64-NEXT: leal -1(%rax,%rax), %eax
950 ; X64-NEXT: .LBB31_3: # %endblock
952 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 16) nounwind
956 define i1 @length16_eq(i8* %x, i8* %y) nounwind {
957 ; X86-NOSSE-LABEL: length16_eq:
958 ; X86-NOSSE: # %bb.0:
959 ; X86-NOSSE-NEXT: pushl $0
960 ; X86-NOSSE-NEXT: pushl $16
961 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
962 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
963 ; X86-NOSSE-NEXT: calll memcmp
964 ; X86-NOSSE-NEXT: addl $16, %esp
965 ; X86-NOSSE-NEXT: testl %eax, %eax
966 ; X86-NOSSE-NEXT: setne %al
967 ; X86-NOSSE-NEXT: retl
969 ; X86-SSE1-LABEL: length16_eq:
971 ; X86-SSE1-NEXT: pushl $0
972 ; X86-SSE1-NEXT: pushl $16
973 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
974 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
975 ; X86-SSE1-NEXT: calll memcmp
976 ; X86-SSE1-NEXT: addl $16, %esp
977 ; X86-SSE1-NEXT: testl %eax, %eax
978 ; X86-SSE1-NEXT: setne %al
979 ; X86-SSE1-NEXT: retl
981 ; X86-SSE2-LABEL: length16_eq:
983 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %eax
984 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %ecx
985 ; X86-SSE2-NEXT: movdqu (%ecx), %xmm0
986 ; X86-SSE2-NEXT: movdqu (%eax), %xmm1
987 ; X86-SSE2-NEXT: pcmpeqb %xmm0, %xmm1
988 ; X86-SSE2-NEXT: pmovmskb %xmm1, %eax
989 ; X86-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
990 ; X86-SSE2-NEXT: setne %al
991 ; X86-SSE2-NEXT: retl
993 ; X64-SSE2-LABEL: length16_eq:
995 ; X64-SSE2-NEXT: movdqu (%rdi), %xmm0
996 ; X64-SSE2-NEXT: movdqu (%rsi), %xmm1
997 ; X64-SSE2-NEXT: pcmpeqb %xmm0, %xmm1
998 ; X64-SSE2-NEXT: pmovmskb %xmm1, %eax
999 ; X64-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1000 ; X64-SSE2-NEXT: setne %al
1001 ; X64-SSE2-NEXT: retq
1003 ; X64-AVX-LABEL: length16_eq:
1005 ; X64-AVX-NEXT: vmovdqu (%rdi), %xmm0
1006 ; X64-AVX-NEXT: vpxor (%rsi), %xmm0, %xmm0
1007 ; X64-AVX-NEXT: vptest %xmm0, %xmm0
1008 ; X64-AVX-NEXT: setne %al
1009 ; X64-AVX-NEXT: retq
1011 ; X64-AVX512-LABEL: length16_eq:
1012 ; X64-AVX512: # %bb.0:
1013 ; X64-AVX512-NEXT: vmovdqu (%rdi), %xmm0
1014 ; X64-AVX512-NEXT: vpxor (%rsi), %xmm0, %xmm0
1015 ; X64-AVX512-NEXT: vptest %xmm0, %xmm0
1016 ; X64-AVX512-NEXT: setne %al
1017 ; X64-AVX512-NEXT: retq
1018 %call = tail call i32 @memcmp(i8* %x, i8* %y, i64 16) nounwind
1019 %cmp = icmp ne i32 %call, 0
1023 define i1 @length16_eq_const(i8* %X) nounwind {
1024 ; X86-NOSSE-LABEL: length16_eq_const:
1025 ; X86-NOSSE: # %bb.0:
1026 ; X86-NOSSE-NEXT: pushl $0
1027 ; X86-NOSSE-NEXT: pushl $16
1028 ; X86-NOSSE-NEXT: pushl $.L.str
1029 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
1030 ; X86-NOSSE-NEXT: calll memcmp
1031 ; X86-NOSSE-NEXT: addl $16, %esp
1032 ; X86-NOSSE-NEXT: testl %eax, %eax
1033 ; X86-NOSSE-NEXT: sete %al
1034 ; X86-NOSSE-NEXT: retl
1036 ; X86-SSE1-LABEL: length16_eq_const:
1037 ; X86-SSE1: # %bb.0:
1038 ; X86-SSE1-NEXT: pushl $0
1039 ; X86-SSE1-NEXT: pushl $16
1040 ; X86-SSE1-NEXT: pushl $.L.str
1041 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
1042 ; X86-SSE1-NEXT: calll memcmp
1043 ; X86-SSE1-NEXT: addl $16, %esp
1044 ; X86-SSE1-NEXT: testl %eax, %eax
1045 ; X86-SSE1-NEXT: sete %al
1046 ; X86-SSE1-NEXT: retl
1048 ; X86-SSE2-LABEL: length16_eq_const:
1049 ; X86-SSE2: # %bb.0:
1050 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %eax
1051 ; X86-SSE2-NEXT: movdqu (%eax), %xmm0
1052 ; X86-SSE2-NEXT: pcmpeqb {{\.LCPI.*}}, %xmm0
1053 ; X86-SSE2-NEXT: pmovmskb %xmm0, %eax
1054 ; X86-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1055 ; X86-SSE2-NEXT: sete %al
1056 ; X86-SSE2-NEXT: retl
1058 ; X64-SSE2-LABEL: length16_eq_const:
1059 ; X64-SSE2: # %bb.0:
1060 ; X64-SSE2-NEXT: movdqu (%rdi), %xmm0
1061 ; X64-SSE2-NEXT: pcmpeqb {{.*}}(%rip), %xmm0
1062 ; X64-SSE2-NEXT: pmovmskb %xmm0, %eax
1063 ; X64-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1064 ; X64-SSE2-NEXT: sete %al
1065 ; X64-SSE2-NEXT: retq
1067 ; X64-AVX-LABEL: length16_eq_const:
1069 ; X64-AVX-NEXT: vmovdqu (%rdi), %xmm0
1070 ; X64-AVX-NEXT: vpxor {{.*}}(%rip), %xmm0, %xmm0
1071 ; X64-AVX-NEXT: vptest %xmm0, %xmm0
1072 ; X64-AVX-NEXT: sete %al
1073 ; X64-AVX-NEXT: retq
1075 ; X64-AVX512-LABEL: length16_eq_const:
1076 ; X64-AVX512: # %bb.0:
1077 ; X64-AVX512-NEXT: vmovdqu (%rdi), %xmm0
1078 ; X64-AVX512-NEXT: vpxor {{.*}}(%rip), %xmm0, %xmm0
1079 ; X64-AVX512-NEXT: vptest %xmm0, %xmm0
1080 ; X64-AVX512-NEXT: sete %al
1081 ; X64-AVX512-NEXT: retq
1082 %m = tail call i32 @memcmp(i8* %X, i8* getelementptr inbounds ([65 x i8], [65 x i8]* @.str, i32 0, i32 0), i64 16) nounwind
1083 %c = icmp eq i32 %m, 0
1087 ; PR33914 - https://bugs.llvm.org/show_bug.cgi?id=33914
1089 define i32 @length24(i8* %X, i8* %Y) nounwind {
1090 ; X86-LABEL: length24:
1092 ; X86-NEXT: pushl $0
1093 ; X86-NEXT: pushl $24
1094 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1095 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1096 ; X86-NEXT: calll memcmp
1097 ; X86-NEXT: addl $16, %esp
1100 ; X64-LABEL: length24:
1102 ; X64-NEXT: movl $24, %edx
1103 ; X64-NEXT: jmp memcmp # TAILCALL
1104 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 24) nounwind
1108 define i1 @length24_eq(i8* %x, i8* %y) nounwind {
1109 ; X86-NOSSE-LABEL: length24_eq:
1110 ; X86-NOSSE: # %bb.0:
1111 ; X86-NOSSE-NEXT: pushl $0
1112 ; X86-NOSSE-NEXT: pushl $24
1113 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
1114 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
1115 ; X86-NOSSE-NEXT: calll memcmp
1116 ; X86-NOSSE-NEXT: addl $16, %esp
1117 ; X86-NOSSE-NEXT: testl %eax, %eax
1118 ; X86-NOSSE-NEXT: sete %al
1119 ; X86-NOSSE-NEXT: retl
1121 ; X86-SSE1-LABEL: length24_eq:
1122 ; X86-SSE1: # %bb.0:
1123 ; X86-SSE1-NEXT: pushl $0
1124 ; X86-SSE1-NEXT: pushl $24
1125 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
1126 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
1127 ; X86-SSE1-NEXT: calll memcmp
1128 ; X86-SSE1-NEXT: addl $16, %esp
1129 ; X86-SSE1-NEXT: testl %eax, %eax
1130 ; X86-SSE1-NEXT: sete %al
1131 ; X86-SSE1-NEXT: retl
1133 ; X86-SSE2-LABEL: length24_eq:
1134 ; X86-SSE2: # %bb.0:
1135 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %eax
1136 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %ecx
1137 ; X86-SSE2-NEXT: movdqu (%ecx), %xmm0
1138 ; X86-SSE2-NEXT: movdqu 8(%ecx), %xmm1
1139 ; X86-SSE2-NEXT: movdqu (%eax), %xmm2
1140 ; X86-SSE2-NEXT: pcmpeqb %xmm0, %xmm2
1141 ; X86-SSE2-NEXT: movdqu 8(%eax), %xmm0
1142 ; X86-SSE2-NEXT: pcmpeqb %xmm1, %xmm0
1143 ; X86-SSE2-NEXT: pand %xmm2, %xmm0
1144 ; X86-SSE2-NEXT: pmovmskb %xmm0, %eax
1145 ; X86-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1146 ; X86-SSE2-NEXT: sete %al
1147 ; X86-SSE2-NEXT: retl
1149 ; X64-SSE2-LABEL: length24_eq:
1150 ; X64-SSE2: # %bb.0:
1151 ; X64-SSE2-NEXT: movdqu (%rdi), %xmm0
1152 ; X64-SSE2-NEXT: movdqu (%rsi), %xmm1
1153 ; X64-SSE2-NEXT: pcmpeqb %xmm0, %xmm1
1154 ; X64-SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
1155 ; X64-SSE2-NEXT: movq {{.*#+}} xmm2 = mem[0],zero
1156 ; X64-SSE2-NEXT: pcmpeqb %xmm0, %xmm2
1157 ; X64-SSE2-NEXT: pand %xmm1, %xmm2
1158 ; X64-SSE2-NEXT: pmovmskb %xmm2, %eax
1159 ; X64-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1160 ; X64-SSE2-NEXT: sete %al
1161 ; X64-SSE2-NEXT: retq
1163 ; X64-AVX-LABEL: length24_eq:
1165 ; X64-AVX-NEXT: vmovdqu (%rdi), %xmm0
1166 ; X64-AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
1167 ; X64-AVX-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
1168 ; X64-AVX-NEXT: vpxor %xmm2, %xmm1, %xmm1
1169 ; X64-AVX-NEXT: vpxor (%rsi), %xmm0, %xmm0
1170 ; X64-AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
1171 ; X64-AVX-NEXT: vptest %xmm0, %xmm0
1172 ; X64-AVX-NEXT: sete %al
1173 ; X64-AVX-NEXT: retq
1175 ; X64-AVX512-LABEL: length24_eq:
1176 ; X64-AVX512: # %bb.0:
1177 ; X64-AVX512-NEXT: vmovdqu (%rdi), %xmm0
1178 ; X64-AVX512-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
1179 ; X64-AVX512-NEXT: vmovq {{.*#+}} xmm2 = mem[0],zero
1180 ; X64-AVX512-NEXT: vpxor %xmm2, %xmm1, %xmm1
1181 ; X64-AVX512-NEXT: vpxor (%rsi), %xmm0, %xmm0
1182 ; X64-AVX512-NEXT: vpor %xmm1, %xmm0, %xmm0
1183 ; X64-AVX512-NEXT: vptest %xmm0, %xmm0
1184 ; X64-AVX512-NEXT: sete %al
1185 ; X64-AVX512-NEXT: retq
1186 %call = tail call i32 @memcmp(i8* %x, i8* %y, i64 24) nounwind
1187 %cmp = icmp eq i32 %call, 0
1191 define i1 @length24_eq_const(i8* %X) nounwind {
1192 ; X86-NOSSE-LABEL: length24_eq_const:
1193 ; X86-NOSSE: # %bb.0:
1194 ; X86-NOSSE-NEXT: pushl $0
1195 ; X86-NOSSE-NEXT: pushl $24
1196 ; X86-NOSSE-NEXT: pushl $.L.str
1197 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
1198 ; X86-NOSSE-NEXT: calll memcmp
1199 ; X86-NOSSE-NEXT: addl $16, %esp
1200 ; X86-NOSSE-NEXT: testl %eax, %eax
1201 ; X86-NOSSE-NEXT: setne %al
1202 ; X86-NOSSE-NEXT: retl
1204 ; X86-SSE1-LABEL: length24_eq_const:
1205 ; X86-SSE1: # %bb.0:
1206 ; X86-SSE1-NEXT: pushl $0
1207 ; X86-SSE1-NEXT: pushl $24
1208 ; X86-SSE1-NEXT: pushl $.L.str
1209 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
1210 ; X86-SSE1-NEXT: calll memcmp
1211 ; X86-SSE1-NEXT: addl $16, %esp
1212 ; X86-SSE1-NEXT: testl %eax, %eax
1213 ; X86-SSE1-NEXT: setne %al
1214 ; X86-SSE1-NEXT: retl
1216 ; X86-SSE2-LABEL: length24_eq_const:
1217 ; X86-SSE2: # %bb.0:
1218 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %eax
1219 ; X86-SSE2-NEXT: movdqu (%eax), %xmm0
1220 ; X86-SSE2-NEXT: movdqu 8(%eax), %xmm1
1221 ; X86-SSE2-NEXT: pcmpeqb {{\.LCPI.*}}, %xmm1
1222 ; X86-SSE2-NEXT: pcmpeqb {{\.LCPI.*}}, %xmm0
1223 ; X86-SSE2-NEXT: pand %xmm1, %xmm0
1224 ; X86-SSE2-NEXT: pmovmskb %xmm0, %eax
1225 ; X86-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1226 ; X86-SSE2-NEXT: setne %al
1227 ; X86-SSE2-NEXT: retl
1229 ; X64-SSE2-LABEL: length24_eq_const:
1230 ; X64-SSE2: # %bb.0:
1231 ; X64-SSE2-NEXT: movdqu (%rdi), %xmm0
1232 ; X64-SSE2-NEXT: movq {{.*#+}} xmm1 = mem[0],zero
1233 ; X64-SSE2-NEXT: pcmpeqb {{.*}}(%rip), %xmm1
1234 ; X64-SSE2-NEXT: pcmpeqb {{.*}}(%rip), %xmm0
1235 ; X64-SSE2-NEXT: pand %xmm1, %xmm0
1236 ; X64-SSE2-NEXT: pmovmskb %xmm0, %eax
1237 ; X64-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1238 ; X64-SSE2-NEXT: setne %al
1239 ; X64-SSE2-NEXT: retq
1241 ; X64-AVX-LABEL: length24_eq_const:
1243 ; X64-AVX-NEXT: vmovdqu (%rdi), %xmm0
1244 ; X64-AVX-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
1245 ; X64-AVX-NEXT: vpxor {{.*}}(%rip), %xmm1, %xmm1
1246 ; X64-AVX-NEXT: vpxor {{.*}}(%rip), %xmm0, %xmm0
1247 ; X64-AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
1248 ; X64-AVX-NEXT: vptest %xmm0, %xmm0
1249 ; X64-AVX-NEXT: setne %al
1250 ; X64-AVX-NEXT: retq
1252 ; X64-AVX512-LABEL: length24_eq_const:
1253 ; X64-AVX512: # %bb.0:
1254 ; X64-AVX512-NEXT: vmovdqu (%rdi), %xmm0
1255 ; X64-AVX512-NEXT: vmovq {{.*#+}} xmm1 = mem[0],zero
1256 ; X64-AVX512-NEXT: vpxor {{.*}}(%rip), %xmm1, %xmm1
1257 ; X64-AVX512-NEXT: vpxor {{.*}}(%rip), %xmm0, %xmm0
1258 ; X64-AVX512-NEXT: vpor %xmm1, %xmm0, %xmm0
1259 ; X64-AVX512-NEXT: vptest %xmm0, %xmm0
1260 ; X64-AVX512-NEXT: setne %al
1261 ; X64-AVX512-NEXT: retq
1262 %m = tail call i32 @memcmp(i8* %X, i8* getelementptr inbounds ([65 x i8], [65 x i8]* @.str, i32 0, i32 0), i64 24) nounwind
1263 %c = icmp ne i32 %m, 0
1267 define i32 @length32(i8* %X, i8* %Y) nounwind {
1268 ; X86-LABEL: length32:
1270 ; X86-NEXT: pushl $0
1271 ; X86-NEXT: pushl $32
1272 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1273 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1274 ; X86-NEXT: calll memcmp
1275 ; X86-NEXT: addl $16, %esp
1278 ; X64-LABEL: length32:
1280 ; X64-NEXT: movl $32, %edx
1281 ; X64-NEXT: jmp memcmp # TAILCALL
1282 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 32) nounwind
1286 ; PR33325 - https://bugs.llvm.org/show_bug.cgi?id=33325
1288 define i1 @length32_eq(i8* %x, i8* %y) nounwind {
1289 ; X86-NOSSE-LABEL: length32_eq:
1290 ; X86-NOSSE: # %bb.0:
1291 ; X86-NOSSE-NEXT: pushl $0
1292 ; X86-NOSSE-NEXT: pushl $32
1293 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
1294 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
1295 ; X86-NOSSE-NEXT: calll memcmp
1296 ; X86-NOSSE-NEXT: addl $16, %esp
1297 ; X86-NOSSE-NEXT: testl %eax, %eax
1298 ; X86-NOSSE-NEXT: sete %al
1299 ; X86-NOSSE-NEXT: retl
1301 ; X86-SSE1-LABEL: length32_eq:
1302 ; X86-SSE1: # %bb.0:
1303 ; X86-SSE1-NEXT: pushl $0
1304 ; X86-SSE1-NEXT: pushl $32
1305 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
1306 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
1307 ; X86-SSE1-NEXT: calll memcmp
1308 ; X86-SSE1-NEXT: addl $16, %esp
1309 ; X86-SSE1-NEXT: testl %eax, %eax
1310 ; X86-SSE1-NEXT: sete %al
1311 ; X86-SSE1-NEXT: retl
1313 ; X86-SSE2-LABEL: length32_eq:
1314 ; X86-SSE2: # %bb.0:
1315 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %eax
1316 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %ecx
1317 ; X86-SSE2-NEXT: movdqu (%ecx), %xmm0
1318 ; X86-SSE2-NEXT: movdqu 16(%ecx), %xmm1
1319 ; X86-SSE2-NEXT: movdqu (%eax), %xmm2
1320 ; X86-SSE2-NEXT: pcmpeqb %xmm0, %xmm2
1321 ; X86-SSE2-NEXT: movdqu 16(%eax), %xmm0
1322 ; X86-SSE2-NEXT: pcmpeqb %xmm1, %xmm0
1323 ; X86-SSE2-NEXT: pand %xmm2, %xmm0
1324 ; X86-SSE2-NEXT: pmovmskb %xmm0, %eax
1325 ; X86-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1326 ; X86-SSE2-NEXT: sete %al
1327 ; X86-SSE2-NEXT: retl
1329 ; X64-SSE2-LABEL: length32_eq:
1330 ; X64-SSE2: # %bb.0:
1331 ; X64-SSE2-NEXT: movdqu (%rdi), %xmm0
1332 ; X64-SSE2-NEXT: movdqu 16(%rdi), %xmm1
1333 ; X64-SSE2-NEXT: movdqu (%rsi), %xmm2
1334 ; X64-SSE2-NEXT: pcmpeqb %xmm0, %xmm2
1335 ; X64-SSE2-NEXT: movdqu 16(%rsi), %xmm0
1336 ; X64-SSE2-NEXT: pcmpeqb %xmm1, %xmm0
1337 ; X64-SSE2-NEXT: pand %xmm2, %xmm0
1338 ; X64-SSE2-NEXT: pmovmskb %xmm0, %eax
1339 ; X64-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1340 ; X64-SSE2-NEXT: sete %al
1341 ; X64-SSE2-NEXT: retq
1343 ; X64-AVX1-LABEL: length32_eq:
1344 ; X64-AVX1: # %bb.0:
1345 ; X64-AVX1-NEXT: vmovdqu (%rdi), %xmm0
1346 ; X64-AVX1-NEXT: vmovdqu 16(%rdi), %xmm1
1347 ; X64-AVX1-NEXT: vpxor 16(%rsi), %xmm1, %xmm1
1348 ; X64-AVX1-NEXT: vpxor (%rsi), %xmm0, %xmm0
1349 ; X64-AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
1350 ; X64-AVX1-NEXT: vptest %xmm0, %xmm0
1351 ; X64-AVX1-NEXT: sete %al
1352 ; X64-AVX1-NEXT: retq
1354 ; X64-AVX2-LABEL: length32_eq:
1355 ; X64-AVX2: # %bb.0:
1356 ; X64-AVX2-NEXT: vmovdqu (%rdi), %ymm0
1357 ; X64-AVX2-NEXT: vpxor (%rsi), %ymm0, %ymm0
1358 ; X64-AVX2-NEXT: vptest %ymm0, %ymm0
1359 ; X64-AVX2-NEXT: sete %al
1360 ; X64-AVX2-NEXT: vzeroupper
1361 ; X64-AVX2-NEXT: retq
1363 ; X64-AVX512-LABEL: length32_eq:
1364 ; X64-AVX512: # %bb.0:
1365 ; X64-AVX512-NEXT: vmovdqu (%rdi), %ymm0
1366 ; X64-AVX512-NEXT: vpxor (%rsi), %ymm0, %ymm0
1367 ; X64-AVX512-NEXT: vptest %ymm0, %ymm0
1368 ; X64-AVX512-NEXT: sete %al
1369 ; X64-AVX512-NEXT: vzeroupper
1370 ; X64-AVX512-NEXT: retq
1371 %call = tail call i32 @memcmp(i8* %x, i8* %y, i64 32) nounwind
1372 %cmp = icmp eq i32 %call, 0
1376 define i1 @length32_eq_prefer128(i8* %x, i8* %y) nounwind "prefer-vector-width"="128" {
1377 ; X86-NOSSE-LABEL: length32_eq_prefer128:
1378 ; X86-NOSSE: # %bb.0:
1379 ; X86-NOSSE-NEXT: pushl $0
1380 ; X86-NOSSE-NEXT: pushl $32
1381 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
1382 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
1383 ; X86-NOSSE-NEXT: calll memcmp
1384 ; X86-NOSSE-NEXT: addl $16, %esp
1385 ; X86-NOSSE-NEXT: testl %eax, %eax
1386 ; X86-NOSSE-NEXT: sete %al
1387 ; X86-NOSSE-NEXT: retl
1389 ; X86-SSE1-LABEL: length32_eq_prefer128:
1390 ; X86-SSE1: # %bb.0:
1391 ; X86-SSE1-NEXT: pushl $0
1392 ; X86-SSE1-NEXT: pushl $32
1393 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
1394 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
1395 ; X86-SSE1-NEXT: calll memcmp
1396 ; X86-SSE1-NEXT: addl $16, %esp
1397 ; X86-SSE1-NEXT: testl %eax, %eax
1398 ; X86-SSE1-NEXT: sete %al
1399 ; X86-SSE1-NEXT: retl
1401 ; X86-SSE2-LABEL: length32_eq_prefer128:
1402 ; X86-SSE2: # %bb.0:
1403 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %eax
1404 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %ecx
1405 ; X86-SSE2-NEXT: movdqu (%ecx), %xmm0
1406 ; X86-SSE2-NEXT: movdqu 16(%ecx), %xmm1
1407 ; X86-SSE2-NEXT: movdqu (%eax), %xmm2
1408 ; X86-SSE2-NEXT: pcmpeqb %xmm0, %xmm2
1409 ; X86-SSE2-NEXT: movdqu 16(%eax), %xmm0
1410 ; X86-SSE2-NEXT: pcmpeqb %xmm1, %xmm0
1411 ; X86-SSE2-NEXT: pand %xmm2, %xmm0
1412 ; X86-SSE2-NEXT: pmovmskb %xmm0, %eax
1413 ; X86-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1414 ; X86-SSE2-NEXT: sete %al
1415 ; X86-SSE2-NEXT: retl
1417 ; X64-SSE2-LABEL: length32_eq_prefer128:
1418 ; X64-SSE2: # %bb.0:
1419 ; X64-SSE2-NEXT: movdqu (%rdi), %xmm0
1420 ; X64-SSE2-NEXT: movdqu 16(%rdi), %xmm1
1421 ; X64-SSE2-NEXT: movdqu (%rsi), %xmm2
1422 ; X64-SSE2-NEXT: pcmpeqb %xmm0, %xmm2
1423 ; X64-SSE2-NEXT: movdqu 16(%rsi), %xmm0
1424 ; X64-SSE2-NEXT: pcmpeqb %xmm1, %xmm0
1425 ; X64-SSE2-NEXT: pand %xmm2, %xmm0
1426 ; X64-SSE2-NEXT: pmovmskb %xmm0, %eax
1427 ; X64-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1428 ; X64-SSE2-NEXT: sete %al
1429 ; X64-SSE2-NEXT: retq
1431 ; X64-AVX-LABEL: length32_eq_prefer128:
1433 ; X64-AVX-NEXT: vmovdqu (%rdi), %xmm0
1434 ; X64-AVX-NEXT: vmovdqu 16(%rdi), %xmm1
1435 ; X64-AVX-NEXT: vpxor 16(%rsi), %xmm1, %xmm1
1436 ; X64-AVX-NEXT: vpxor (%rsi), %xmm0, %xmm0
1437 ; X64-AVX-NEXT: vpor %xmm1, %xmm0, %xmm0
1438 ; X64-AVX-NEXT: vptest %xmm0, %xmm0
1439 ; X64-AVX-NEXT: sete %al
1440 ; X64-AVX-NEXT: retq
1442 ; X64-AVX512-LABEL: length32_eq_prefer128:
1443 ; X64-AVX512: # %bb.0:
1444 ; X64-AVX512-NEXT: vmovdqu (%rdi), %xmm0
1445 ; X64-AVX512-NEXT: vmovdqu 16(%rdi), %xmm1
1446 ; X64-AVX512-NEXT: vpxor 16(%rsi), %xmm1, %xmm1
1447 ; X64-AVX512-NEXT: vpxor (%rsi), %xmm0, %xmm0
1448 ; X64-AVX512-NEXT: vpor %xmm1, %xmm0, %xmm0
1449 ; X64-AVX512-NEXT: vptest %xmm0, %xmm0
1450 ; X64-AVX512-NEXT: sete %al
1451 ; X64-AVX512-NEXT: retq
1452 %call = tail call i32 @memcmp(i8* %x, i8* %y, i64 32) nounwind
1453 %cmp = icmp eq i32 %call, 0
1457 define i1 @length32_eq_const(i8* %X) nounwind {
1458 ; X86-NOSSE-LABEL: length32_eq_const:
1459 ; X86-NOSSE: # %bb.0:
1460 ; X86-NOSSE-NEXT: pushl $0
1461 ; X86-NOSSE-NEXT: pushl $32
1462 ; X86-NOSSE-NEXT: pushl $.L.str
1463 ; X86-NOSSE-NEXT: pushl {{[0-9]+}}(%esp)
1464 ; X86-NOSSE-NEXT: calll memcmp
1465 ; X86-NOSSE-NEXT: addl $16, %esp
1466 ; X86-NOSSE-NEXT: testl %eax, %eax
1467 ; X86-NOSSE-NEXT: setne %al
1468 ; X86-NOSSE-NEXT: retl
1470 ; X86-SSE1-LABEL: length32_eq_const:
1471 ; X86-SSE1: # %bb.0:
1472 ; X86-SSE1-NEXT: pushl $0
1473 ; X86-SSE1-NEXT: pushl $32
1474 ; X86-SSE1-NEXT: pushl $.L.str
1475 ; X86-SSE1-NEXT: pushl {{[0-9]+}}(%esp)
1476 ; X86-SSE1-NEXT: calll memcmp
1477 ; X86-SSE1-NEXT: addl $16, %esp
1478 ; X86-SSE1-NEXT: testl %eax, %eax
1479 ; X86-SSE1-NEXT: setne %al
1480 ; X86-SSE1-NEXT: retl
1482 ; X86-SSE2-LABEL: length32_eq_const:
1483 ; X86-SSE2: # %bb.0:
1484 ; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %eax
1485 ; X86-SSE2-NEXT: movdqu (%eax), %xmm0
1486 ; X86-SSE2-NEXT: movdqu 16(%eax), %xmm1
1487 ; X86-SSE2-NEXT: pcmpeqb {{\.LCPI.*}}, %xmm1
1488 ; X86-SSE2-NEXT: pcmpeqb {{\.LCPI.*}}, %xmm0
1489 ; X86-SSE2-NEXT: pand %xmm1, %xmm0
1490 ; X86-SSE2-NEXT: pmovmskb %xmm0, %eax
1491 ; X86-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1492 ; X86-SSE2-NEXT: setne %al
1493 ; X86-SSE2-NEXT: retl
1495 ; X64-SSE2-LABEL: length32_eq_const:
1496 ; X64-SSE2: # %bb.0:
1497 ; X64-SSE2-NEXT: movdqu (%rdi), %xmm0
1498 ; X64-SSE2-NEXT: movdqu 16(%rdi), %xmm1
1499 ; X64-SSE2-NEXT: pcmpeqb {{.*}}(%rip), %xmm1
1500 ; X64-SSE2-NEXT: pcmpeqb {{.*}}(%rip), %xmm0
1501 ; X64-SSE2-NEXT: pand %xmm1, %xmm0
1502 ; X64-SSE2-NEXT: pmovmskb %xmm0, %eax
1503 ; X64-SSE2-NEXT: cmpl $65535, %eax # imm = 0xFFFF
1504 ; X64-SSE2-NEXT: setne %al
1505 ; X64-SSE2-NEXT: retq
1507 ; X64-AVX1-LABEL: length32_eq_const:
1508 ; X64-AVX1: # %bb.0:
1509 ; X64-AVX1-NEXT: vmovdqu (%rdi), %xmm0
1510 ; X64-AVX1-NEXT: vmovdqu 16(%rdi), %xmm1
1511 ; X64-AVX1-NEXT: vpxor {{.*}}(%rip), %xmm1, %xmm1
1512 ; X64-AVX1-NEXT: vpxor {{.*}}(%rip), %xmm0, %xmm0
1513 ; X64-AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
1514 ; X64-AVX1-NEXT: vptest %xmm0, %xmm0
1515 ; X64-AVX1-NEXT: setne %al
1516 ; X64-AVX1-NEXT: retq
1518 ; X64-AVX2-LABEL: length32_eq_const:
1519 ; X64-AVX2: # %bb.0:
1520 ; X64-AVX2-NEXT: vmovdqu (%rdi), %ymm0
1521 ; X64-AVX2-NEXT: vpxor {{.*}}(%rip), %ymm0, %ymm0
1522 ; X64-AVX2-NEXT: vptest %ymm0, %ymm0
1523 ; X64-AVX2-NEXT: setne %al
1524 ; X64-AVX2-NEXT: vzeroupper
1525 ; X64-AVX2-NEXT: retq
1527 ; X64-AVX512-LABEL: length32_eq_const:
1528 ; X64-AVX512: # %bb.0:
1529 ; X64-AVX512-NEXT: vmovdqu (%rdi), %ymm0
1530 ; X64-AVX512-NEXT: vpxor {{.*}}(%rip), %ymm0, %ymm0
1531 ; X64-AVX512-NEXT: vptest %ymm0, %ymm0
1532 ; X64-AVX512-NEXT: setne %al
1533 ; X64-AVX512-NEXT: vzeroupper
1534 ; X64-AVX512-NEXT: retq
1535 %m = tail call i32 @memcmp(i8* %X, i8* getelementptr inbounds ([65 x i8], [65 x i8]* @.str, i32 0, i32 0), i64 32) nounwind
1536 %c = icmp ne i32 %m, 0
1540 define i32 @length64(i8* %X, i8* %Y) nounwind {
1541 ; X86-LABEL: length64:
1543 ; X86-NEXT: pushl $0
1544 ; X86-NEXT: pushl $64
1545 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1546 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1547 ; X86-NEXT: calll memcmp
1548 ; X86-NEXT: addl $16, %esp
1551 ; X64-LABEL: length64:
1553 ; X64-NEXT: movl $64, %edx
1554 ; X64-NEXT: jmp memcmp # TAILCALL
1555 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 64) nounwind
1559 define i1 @length64_eq(i8* %x, i8* %y) nounwind {
1560 ; X86-LABEL: length64_eq:
1562 ; X86-NEXT: pushl $0
1563 ; X86-NEXT: pushl $64
1564 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1565 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1566 ; X86-NEXT: calll memcmp
1567 ; X86-NEXT: addl $16, %esp
1568 ; X86-NEXT: testl %eax, %eax
1569 ; X86-NEXT: setne %al
1572 ; X64-SSE2-LABEL: length64_eq:
1573 ; X64-SSE2: # %bb.0:
1574 ; X64-SSE2-NEXT: pushq %rax
1575 ; X64-SSE2-NEXT: movl $64, %edx
1576 ; X64-SSE2-NEXT: callq memcmp
1577 ; X64-SSE2-NEXT: testl %eax, %eax
1578 ; X64-SSE2-NEXT: setne %al
1579 ; X64-SSE2-NEXT: popq %rcx
1580 ; X64-SSE2-NEXT: retq
1582 ; X64-AVX1-LABEL: length64_eq:
1583 ; X64-AVX1: # %bb.0:
1584 ; X64-AVX1-NEXT: pushq %rax
1585 ; X64-AVX1-NEXT: movl $64, %edx
1586 ; X64-AVX1-NEXT: callq memcmp
1587 ; X64-AVX1-NEXT: testl %eax, %eax
1588 ; X64-AVX1-NEXT: setne %al
1589 ; X64-AVX1-NEXT: popq %rcx
1590 ; X64-AVX1-NEXT: retq
1592 ; X64-AVX2-LABEL: length64_eq:
1593 ; X64-AVX2: # %bb.0:
1594 ; X64-AVX2-NEXT: vmovdqu (%rdi), %ymm0
1595 ; X64-AVX2-NEXT: vmovdqu 32(%rdi), %ymm1
1596 ; X64-AVX2-NEXT: vpxor 32(%rsi), %ymm1, %ymm1
1597 ; X64-AVX2-NEXT: vpxor (%rsi), %ymm0, %ymm0
1598 ; X64-AVX2-NEXT: vpor %ymm1, %ymm0, %ymm0
1599 ; X64-AVX2-NEXT: vptest %ymm0, %ymm0
1600 ; X64-AVX2-NEXT: setne %al
1601 ; X64-AVX2-NEXT: vzeroupper
1602 ; X64-AVX2-NEXT: retq
1604 ; X64-AVX512F-LABEL: length64_eq:
1605 ; X64-AVX512F: # %bb.0:
1606 ; X64-AVX512F-NEXT: vmovdqu64 (%rdi), %zmm0
1607 ; X64-AVX512F-NEXT: vpcmpeqd (%rsi), %zmm0, %k0
1608 ; X64-AVX512F-NEXT: kortestw %k0, %k0
1609 ; X64-AVX512F-NEXT: setae %al
1610 ; X64-AVX512F-NEXT: vzeroupper
1611 ; X64-AVX512F-NEXT: retq
1613 ; X64-AVX512BW-LABEL: length64_eq:
1614 ; X64-AVX512BW: # %bb.0:
1615 ; X64-AVX512BW-NEXT: vmovdqu64 (%rdi), %zmm0
1616 ; X64-AVX512BW-NEXT: vpcmpeqb (%rsi), %zmm0, %k0
1617 ; X64-AVX512BW-NEXT: kortestq %k0, %k0
1618 ; X64-AVX512BW-NEXT: setae %al
1619 ; X64-AVX512BW-NEXT: vzeroupper
1620 ; X64-AVX512BW-NEXT: retq
1621 %call = tail call i32 @memcmp(i8* %x, i8* %y, i64 64) nounwind
1622 %cmp = icmp ne i32 %call, 0
1626 define i1 @length64_eq_const(i8* %X) nounwind {
1627 ; X86-LABEL: length64_eq_const:
1629 ; X86-NEXT: pushl $0
1630 ; X86-NEXT: pushl $64
1631 ; X86-NEXT: pushl $.L.str
1632 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1633 ; X86-NEXT: calll memcmp
1634 ; X86-NEXT: addl $16, %esp
1635 ; X86-NEXT: testl %eax, %eax
1636 ; X86-NEXT: sete %al
1639 ; X64-SSE2-LABEL: length64_eq_const:
1640 ; X64-SSE2: # %bb.0:
1641 ; X64-SSE2-NEXT: pushq %rax
1642 ; X64-SSE2-NEXT: movl $.L.str, %esi
1643 ; X64-SSE2-NEXT: movl $64, %edx
1644 ; X64-SSE2-NEXT: callq memcmp
1645 ; X64-SSE2-NEXT: testl %eax, %eax
1646 ; X64-SSE2-NEXT: sete %al
1647 ; X64-SSE2-NEXT: popq %rcx
1648 ; X64-SSE2-NEXT: retq
1650 ; X64-AVX1-LABEL: length64_eq_const:
1651 ; X64-AVX1: # %bb.0:
1652 ; X64-AVX1-NEXT: pushq %rax
1653 ; X64-AVX1-NEXT: movl $.L.str, %esi
1654 ; X64-AVX1-NEXT: movl $64, %edx
1655 ; X64-AVX1-NEXT: callq memcmp
1656 ; X64-AVX1-NEXT: testl %eax, %eax
1657 ; X64-AVX1-NEXT: sete %al
1658 ; X64-AVX1-NEXT: popq %rcx
1659 ; X64-AVX1-NEXT: retq
1661 ; X64-AVX2-LABEL: length64_eq_const:
1662 ; X64-AVX2: # %bb.0:
1663 ; X64-AVX2-NEXT: vmovdqu (%rdi), %ymm0
1664 ; X64-AVX2-NEXT: vmovdqu 32(%rdi), %ymm1
1665 ; X64-AVX2-NEXT: vpxor {{.*}}(%rip), %ymm1, %ymm1
1666 ; X64-AVX2-NEXT: vpxor {{.*}}(%rip), %ymm0, %ymm0
1667 ; X64-AVX2-NEXT: vpor %ymm1, %ymm0, %ymm0
1668 ; X64-AVX2-NEXT: vptest %ymm0, %ymm0
1669 ; X64-AVX2-NEXT: sete %al
1670 ; X64-AVX2-NEXT: vzeroupper
1671 ; X64-AVX2-NEXT: retq
1673 ; X64-AVX512F-LABEL: length64_eq_const:
1674 ; X64-AVX512F: # %bb.0:
1675 ; X64-AVX512F-NEXT: vmovdqu64 (%rdi), %zmm0
1676 ; X64-AVX512F-NEXT: vpcmpeqd {{.*}}(%rip), %zmm0, %k0
1677 ; X64-AVX512F-NEXT: kortestw %k0, %k0
1678 ; X64-AVX512F-NEXT: setb %al
1679 ; X64-AVX512F-NEXT: vzeroupper
1680 ; X64-AVX512F-NEXT: retq
1682 ; X64-AVX512BW-LABEL: length64_eq_const:
1683 ; X64-AVX512BW: # %bb.0:
1684 ; X64-AVX512BW-NEXT: vmovdqu64 (%rdi), %zmm0
1685 ; X64-AVX512BW-NEXT: vpcmpeqb {{.*}}(%rip), %zmm0, %k0
1686 ; X64-AVX512BW-NEXT: kortestq %k0, %k0
1687 ; X64-AVX512BW-NEXT: setb %al
1688 ; X64-AVX512BW-NEXT: vzeroupper
1689 ; X64-AVX512BW-NEXT: retq
1690 %m = tail call i32 @memcmp(i8* %X, i8* getelementptr inbounds ([65 x i8], [65 x i8]* @.str, i32 0, i32 0), i64 64) nounwind
1691 %c = icmp eq i32 %m, 0
1695 ; This checks that we do not do stupid things with huge sizes.
1696 define i32 @huge_length(i8* %X, i8* %Y) nounwind {
1697 ; X86-LABEL: huge_length:
1699 ; X86-NEXT: pushl $2147483647 # imm = 0x7FFFFFFF
1700 ; X86-NEXT: pushl $-1
1701 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1702 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1703 ; X86-NEXT: calll memcmp
1704 ; X86-NEXT: addl $16, %esp
1707 ; X64-LABEL: huge_length:
1709 ; X64-NEXT: movabsq $9223372036854775807, %rdx # imm = 0x7FFFFFFFFFFFFFFF
1710 ; X64-NEXT: jmp memcmp # TAILCALL
1711 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 9223372036854775807) nounwind
1715 define i1 @huge_length_eq(i8* %X, i8* %Y) nounwind {
1716 ; X86-LABEL: huge_length_eq:
1718 ; X86-NEXT: pushl $2147483647 # imm = 0x7FFFFFFF
1719 ; X86-NEXT: pushl $-1
1720 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1721 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1722 ; X86-NEXT: calll memcmp
1723 ; X86-NEXT: addl $16, %esp
1724 ; X86-NEXT: testl %eax, %eax
1725 ; X86-NEXT: sete %al
1728 ; X64-LABEL: huge_length_eq:
1730 ; X64-NEXT: pushq %rax
1731 ; X64-NEXT: movabsq $9223372036854775807, %rdx # imm = 0x7FFFFFFFFFFFFFFF
1732 ; X64-NEXT: callq memcmp
1733 ; X64-NEXT: testl %eax, %eax
1734 ; X64-NEXT: sete %al
1735 ; X64-NEXT: popq %rcx
1737 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 9223372036854775807) nounwind
1738 %c = icmp eq i32 %m, 0
1742 ; This checks non-constant sizes.
1743 define i32 @nonconst_length(i8* %X, i8* %Y, i64 %size) nounwind {
1744 ; X86-LABEL: nonconst_length:
1746 ; X86-NEXT: jmp memcmp # TAILCALL
1748 ; X64-LABEL: nonconst_length:
1750 ; X64-NEXT: jmp memcmp # TAILCALL
1751 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 %size) nounwind
1755 define i1 @nonconst_length_eq(i8* %X, i8* %Y, i64 %size) nounwind {
1756 ; X86-LABEL: nonconst_length_eq:
1758 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1759 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1760 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1761 ; X86-NEXT: pushl {{[0-9]+}}(%esp)
1762 ; X86-NEXT: calll memcmp
1763 ; X86-NEXT: addl $16, %esp
1764 ; X86-NEXT: testl %eax, %eax
1765 ; X86-NEXT: sete %al
1768 ; X64-LABEL: nonconst_length_eq:
1770 ; X64-NEXT: pushq %rax
1771 ; X64-NEXT: callq memcmp
1772 ; X64-NEXT: testl %eax, %eax
1773 ; X64-NEXT: sete %al
1774 ; X64-NEXT: popq %rcx
1776 %m = tail call i32 @memcmp(i8* %X, i8* %Y, i64 %size) nounwind
1777 %c = icmp eq i32 %m, 0