1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; Test sequences that can use RISBG with a zeroed first operand.
3 ; The tests here assume that RISBLG is available.
5 ; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z196 | FileCheck %s
7 ; Test an extraction of bit 0 from a right-shifted value.
8 define i32 @f1(i32 %foo) {
11 ; CHECK-NEXT: risblg %r2, %r2, 31, 159, 54
13 %shr = lshr i32 %foo, 10
14 %and = and i32 %shr, 1
18 ; ...and again with i64.
19 define i64 @f2(i64 %foo) {
22 ; CHECK-NEXT: risbg %r2, %r2, 63, 191, 54
24 %shr = lshr i64 %foo, 10
25 %and = and i64 %shr, 1
29 ; Test an extraction of other bits from a right-shifted value.
30 define i32 @f3(i32 %foo) {
33 ; CHECK-NEXT: risblg %r2, %r2, 28, 157, 42
35 %shr = lshr i32 %foo, 22
36 %and = and i32 %shr, 12
40 ; ...and again with i64.
41 define i64 @f4(i64 %foo) {
44 ; CHECK-NEXT: risbg %r2, %r2, 60, 189, 42
46 %shr = lshr i64 %foo, 22
47 %and = and i64 %shr, 12
51 ; Test an extraction of most bits from a right-shifted value.
52 ; The range should be reduced to exclude the zeroed high bits.
53 define i32 @f5(i32 %foo) {
56 ; CHECK-NEXT: risblg %r2, %r2, 2, 156, 62
58 %shr = lshr i32 %foo, 2
59 %and = and i32 %shr, -8
63 ; ...and again with i64.
64 define i64 @f6(i64 %foo) {
67 ; CHECK-NEXT: risbg %r2, %r2, 2, 188, 62
69 %shr = lshr i64 %foo, 2
70 %and = and i64 %shr, -8
74 ; Try the next value up (mask ....1111001). This needs a separate shift
76 define i32 @f7(i32 %foo) {
79 ; CHECK-NEXT: srl %r2, 2
80 ; CHECK-NEXT: nill %r2, 65529
82 %shr = lshr i32 %foo, 2
83 %and = and i32 %shr, -7
87 ; ...and again with i64.
88 define i64 @f8(i64 %foo) {
91 ; CHECK-NEXT: srlg %r2, %r2, 2
92 ; CHECK-NEXT: nill %r2, 65529
94 %shr = lshr i64 %foo, 2
95 %and = and i64 %shr, -7
99 ; Test an extraction of bits from a left-shifted value. The range should
100 ; be reduced to exclude the zeroed low bits.
101 define i32 @f9(i32 %foo) {
104 ; CHECK-NEXT: risblg %r2, %r2, 24, 157, 2
105 ; CHECK-NEXT: br %r14
106 %shr = shl i32 %foo, 2
107 %and = and i32 %shr, 255
111 ; ...and again with i64.
112 define i64 @f10(i64 %foo) {
115 ; CHECK-NEXT: risbg %r2, %r2, 56, 189, 2
116 ; CHECK-NEXT: br %r14
117 %shr = shl i64 %foo, 2
118 %and = and i64 %shr, 255
122 ; Try a wrap-around mask (mask ....111100001111). This needs a separate shift
124 define i32 @f11(i32 %foo) {
127 ; CHECK-NEXT: sll %r2, 2
128 ; CHECK-NEXT: nill %r2, 65295
129 ; CHECK-NEXT: br %r14
130 %shr = shl i32 %foo, 2
131 %and = and i32 %shr, -241
135 ; ...and again with i64.
136 define i64 @f12(i64 %foo) {
139 ; CHECK-NEXT: sllg %r2, %r2, 2
140 ; CHECK-NEXT: nill %r2, 65295
141 ; CHECK-NEXT: br %r14
142 %shr = shl i64 %foo, 2
143 %and = and i64 %shr, -241
147 ; Test an extraction from a rotated value, no mask wraparound.
148 ; This is equivalent to the lshr case, because the bits from the
150 define i32 @f13(i32 %foo) {
153 ; CHECK-NEXT: risblg %r2, %r2, 24, 156, 46
154 ; CHECK-NEXT: br %r14
155 %parta = shl i32 %foo, 14
156 %partb = lshr i32 %foo, 18
157 %rotl = or i32 %parta, %partb
158 %and = and i32 %rotl, 248
162 ; ...and again with i64.
163 define i64 @f14(i64 %foo) {
166 ; CHECK-NEXT: risbg %r2, %r2, 56, 188, 14
167 ; CHECK-NEXT: br %r14
168 %parta = shl i64 %foo, 14
169 %partb = lshr i64 %foo, 50
170 %rotl = or i64 %parta, %partb
171 %and = and i64 %rotl, 248
175 ; Try a case in which only the bits from the shl are used.
176 define i32 @f15(i32 %foo) {
179 ; CHECK-NEXT: risblg %r2, %r2, 15, 145, 14
180 ; CHECK-NEXT: br %r14
181 %parta = shl i32 %foo, 14
182 %partb = lshr i32 %foo, 18
183 %rotl = or i32 %parta, %partb
184 %and = and i32 %rotl, 114688
188 ; ...and again with i64.
189 define i64 @f16(i64 %foo) {
192 ; CHECK-NEXT: risbg %r2, %r2, 47, 177, 14
193 ; CHECK-NEXT: br %r14
194 %parta = shl i64 %foo, 14
195 %partb = lshr i64 %foo, 50
196 %rotl = or i64 %parta, %partb
197 %and = and i64 %rotl, 114688
201 ; Test a 32-bit rotate in which both parts of the OR are needed.
202 ; This needs a separate shift and mask.
203 define i32 @f17(i32 %foo) {
206 ; CHECK-NEXT: rll %r2, %r2, 4
207 ; CHECK-NEXT: nilf %r2, 126
208 ; CHECK-NEXT: br %r14
209 %parta = shl i32 %foo, 4
210 %partb = lshr i32 %foo, 28
211 %rotl = or i32 %parta, %partb
212 %and = and i32 %rotl, 126
216 ; ...and for i64, where RISBG should do the rotate too.
217 define i64 @f18(i64 %foo) {
220 ; CHECK-NEXT: risbg %r2, %r2, 57, 190, 4
221 ; CHECK-NEXT: br %r14
222 %parta = shl i64 %foo, 4
223 %partb = lshr i64 %foo, 60
224 %rotl = or i64 %parta, %partb
225 %and = and i64 %rotl, 126
229 ; Test an arithmetic shift right in which some of the sign bits are kept.
230 ; This needs a separate shift and mask.
231 define i32 @f19(i32 %foo) {
234 ; CHECK-NEXT: sra %r2, 28
235 ; CHECK-NEXT: nilf %r2, 30
236 ; CHECK-NEXT: br %r14
237 %shr = ashr i32 %foo, 28
238 %and = and i32 %shr, 30
242 ; ...and again with i64. In this case RISBG is the best way of doing the AND.
243 define i64 @f20(i64 %foo) {
246 ; CHECK-NEXT: srag %r0, %r2, 60
247 ; CHECK-NEXT: risbg %r2, %r0, 59, 190, 0
248 ; CHECK-NEXT: br %r14
249 %shr = ashr i64 %foo, 60
250 %and = and i64 %shr, 30
254 ; Now try an arithmetic right shift in which the sign bits aren't needed.
255 ; Introduce a second use of %shr so that the ashr doesn't decompose to
257 define i32 @f21(i32 %foo, ptr %dest) {
260 ; CHECK-NEXT: srak %r0, %r2, 28
261 ; CHECK-NEXT: risblg %r2, %r2, 28, 158, 36
262 ; CHECK-NEXT: st %r0, 0(%r3)
263 ; CHECK-NEXT: br %r14
264 %shr = ashr i32 %foo, 28
265 store i32 %shr, ptr %dest
266 %and = and i32 %shr, 14
270 ; ...and again with i64.
271 define i64 @f22(i64 %foo, ptr %dest) {
274 ; CHECK-NEXT: srag %r0, %r2, 60
275 ; CHECK-NEXT: risbg %r2, %r2, 60, 190, 4
276 ; CHECK-NEXT: stg %r0, 0(%r3)
277 ; CHECK-NEXT: br %r14
278 %shr = ashr i64 %foo, 60
279 store i64 %shr, ptr %dest
280 %and = and i64 %shr, 14
284 ; Check that we use RISBG for shifted values even if the AND is a
285 ; natural zero extension.
286 define i64 @f23(i64 %foo) {
289 ; CHECK-NEXT: risbg %r2, %r2, 56, 191, 62
290 ; CHECK-NEXT: br %r14
291 %shr = lshr i64 %foo, 2
292 %and = and i64 %shr, 255
296 ; Test a case where the AND comes before a rotate. This needs a separate
298 define i32 @f24(i32 %foo) {
301 ; CHECK-NEXT: nilf %r2, 254
302 ; CHECK-NEXT: rll %r2, %r2, 29
303 ; CHECK-NEXT: br %r14
304 %and = and i32 %foo, 254
305 %parta = lshr i32 %and, 3
306 %partb = shl i32 %and, 29
307 %rotl = or i32 %parta, %partb
311 ; ...and again with i64, where a single RISBG is enough.
312 define i64 @f25(i64 %foo) {
315 ; CHECK-NEXT: risbg %r2, %r2, 57, 187, 3
316 ; CHECK-NEXT: br %r14
317 %and = and i64 %foo, 14
318 %parta = shl i64 %and, 3
319 %partb = lshr i64 %and, 61
320 %rotl = or i64 %parta, %partb
324 ; Test a wrap-around case in which the AND comes before a rotate.
325 ; This again needs a separate mask and rotate.
326 define i32 @f26(i32 %foo) {
329 ; CHECK-NEXT: nill %r2, 65487
330 ; CHECK-NEXT: rll %r2, %r2, 5
331 ; CHECK-NEXT: br %r14
332 %and = and i32 %foo, -49
333 %parta = shl i32 %and, 5
334 %partb = lshr i32 %and, 27
335 %rotl = or i32 %parta, %partb
339 ; ...and again with i64, where a single RISBG is OK.
340 define i64 @f27(i64 %foo) {
343 ; CHECK-NEXT: risbg %r2, %r2, 55, 180, 5
344 ; CHECK-NEXT: br %r14
345 %and = and i64 %foo, -49
346 %parta = shl i64 %and, 5
347 %partb = lshr i64 %and, 59
348 %rotl = or i64 %parta, %partb
352 ; Test a case where the AND comes before a shift left.
353 define i32 @f28(i32 %foo) {
356 ; CHECK-NEXT: risblg %r2, %r2, 0, 141, 17
357 ; CHECK-NEXT: br %r14
358 %and = and i32 %foo, 32766
359 %shl = shl i32 %and, 17
363 ; ...and again with i64.
364 define i64 @f29(i64 %foo) {
367 ; CHECK-NEXT: risbg %r2, %r2, 0, 141, 49
368 ; CHECK-NEXT: br %r14
369 %and = and i64 %foo, 32766
370 %shl = shl i64 %and, 49
374 ; Test the next shift up from f28, in which the mask should get shortened.
375 define i32 @f30(i32 %foo) {
378 ; CHECK-NEXT: risblg %r2, %r2, 0, 140, 18
379 ; CHECK-NEXT: br %r14
380 %and = and i32 %foo, 32766
381 %shl = shl i32 %and, 18
385 ; ...and again with i64.
386 define i64 @f31(i64 %foo) {
389 ; CHECK-NEXT: risbg %r2, %r2, 0, 140, 50
390 ; CHECK-NEXT: br %r14
391 %and = and i64 %foo, 32766
392 %shl = shl i64 %and, 50
396 ; Test a wrap-around case in which the shift left comes after the AND.
397 ; We can't use RISBG for the shift in that case.
398 define i32 @f32(i32 %foo) {
401 ; CHECK-NEXT: nilf %r2, 4194297
402 ; CHECK-NEXT: sll %r2, 10
403 ; CHECK-NEXT: br %r14
404 %and = and i32 %foo, -7
405 %shl = shl i32 %and, 10
409 ; ...and again with i64.
410 define i64 @f33(i64 %foo) {
413 ; CHECK-NEXT: llihf %r0, 4194303
414 ; CHECK-NEXT: oilf %r0, 4294967289
415 ; CHECK-NEXT: ngr %r0, %r2
416 ; CHECK-NEXT: sllg %r2, %r0, 10
417 ; CHECK-NEXT: br %r14
418 %and = and i64 %foo, -7
419 %shl = shl i64 %and, 10
423 ; Test a case where the AND comes before a shift right.
424 define i32 @f34(i32 %foo) {
427 ; CHECK-NEXT: risblg %r2, %r2, 25, 159, 55
428 ; CHECK-NEXT: br %r14
429 %and = and i32 %foo, 65535
430 %shl = lshr i32 %and, 9
434 ; ...and again with i64.
435 define i64 @f35(i64 %foo) {
438 ; CHECK-NEXT: risbg %r2, %r2, 57, 191, 55
439 ; CHECK-NEXT: br %r14
440 %and = and i64 %foo, 65535
441 %shl = lshr i64 %and, 9
445 ; Test a wrap-around case where the AND comes before a shift right.
446 ; We can't use RISBG for the shift in that case.
447 define i32 @f36(i32 %foo) {
450 ; CHECK-NEXT: nill %r2, 65510
451 ; CHECK-NEXT: srl %r2, 1
452 ; CHECK-NEXT: br %r14
453 %and = and i32 %foo, -25
454 %shl = lshr i32 %and, 1
458 ; ...and again with i64.
459 define i64 @f37(i64 %foo) {
462 ; CHECK-NEXT: nill %r2, 65510
463 ; CHECK-NEXT: srlg %r2, %r2, 1
464 ; CHECK-NEXT: br %r14
465 %and = and i64 %foo, -25
466 %shl = lshr i64 %and, 1
470 ; Test a combination involving a large ASHR and a shift left. We can't
472 define i64 @f38(i64 %foo) {
475 ; CHECK-NEXT: srag %r0, %r2, 32
476 ; CHECK-NEXT: sllg %r2, %r0, 5
477 ; CHECK-NEXT: br %r14
478 %ashr = ashr i64 %foo, 32
479 %shl = shl i64 %ashr, 5
483 ; Try a similar thing in which no shifted sign bits are kept.
484 define i64 @f39(i64 %foo, ptr %dest) {
487 ; CHECK-NEXT: srag %r0, %r2, 35
488 ; CHECK-NEXT: risbg %r2, %r2, 33, 189, 31
489 ; CHECK-NEXT: stg %r0, 0(%r3)
490 ; CHECK-NEXT: br %r14
491 %ashr = ashr i64 %foo, 35
492 store i64 %ashr, ptr %dest
493 %shl = shl i64 %ashr, 2
494 %and = and i64 %shl, 2147483647
498 ; ...and again with the next highest shift value, where one sign bit is kept.
499 define i64 @f40(i64 %foo, ptr %dest) {
502 ; CHECK-NEXT: srag %r0, %r2, 36
503 ; CHECK-NEXT: risbg %r2, %r0, 33, 189, 2
504 ; CHECK-NEXT: stg %r0, 0(%r3)
505 ; CHECK-NEXT: br %r14
506 %ashr = ashr i64 %foo, 36
507 store i64 %ashr, ptr %dest
508 %shl = shl i64 %ashr, 2
509 %and = and i64 %shl, 2147483647
513 ; Check a case where the result is zero-extended.
514 define i64 @f41(i32 %a) {
517 ; CHECK-NEXT: # kill: def $r2l killed $r2l def $r2d
518 ; CHECK-NEXT: risbg %r2, %r2, 36, 191, 62
519 ; CHECK-NEXT: br %r14
521 %shr = lshr i32 %shl, 4
522 %ext = zext i32 %shr to i64
526 ; In this case the sign extension is converted to a pair of 32-bit shifts,
527 ; which is then extended to 64 bits. We previously used the wrong bit size
528 ; when testing whether the shifted-in bits of the shift right were significant.
529 define i64 @f42(i1 %x) {
532 ; CHECK-NEXT: nilf %r2, 1
533 ; CHECK-NEXT: lcr %r0, %r2
534 ; CHECK-NEXT: llgcr %r2, %r0
535 ; CHECK-NEXT: br %r14
536 %ext = sext i1 %x to i8
537 %ext2 = zext i8 %ext to i64
541 ; Check that we get the case where a 64-bit shift is used by a 32-bit and.
542 ; Note that this cannot use RISBLG, but should use RISBG.
543 define signext i32 @f43(i64 %x) {
546 ; CHECK-NEXT: risbg %r0, %r2, 32, 189, 52
547 ; CHECK-NEXT: lgfr %r2, %r0
548 ; CHECK-NEXT: br %r14
549 %shr3 = lshr i64 %x, 12
550 %shr3.tr = trunc i64 %shr3 to i32
551 %conv = and i32 %shr3.tr, -4
555 ; Check that we don't get the case where the 32-bit and mask is not contiguous
556 define signext i32 @f44(i64 %x) {
559 ; CHECK-NEXT: srlg %r2, %r2, 12
560 ; CHECK-NEXT: lghi %r0, 10
561 ; CHECK-NEXT: ngr %r2, %r0
562 ; CHECK-NEXT: br %r14
563 %shr4 = lshr i64 %x, 12
564 %conv = trunc i64 %shr4 to i32
565 %and = and i32 %conv, 10