3 // Copyright 2018 The Go Authors. All rights reserved.
4 // Use of this source code is governed by a BSD-style
5 // license that can be found in the LICENSE file.
11 // ----------------------- //
12 // bits.LeadingZeros //
13 // ----------------------- //
15 func LeadingZeros(n uint) int {
18 // arm:"CLZ" arm64:"CLZ"
21 return bits.LeadingZeros(n)
24 func LeadingZeros64(n uint64) int {
27 // arm:"CLZ" arm64:"CLZ"
30 return bits.LeadingZeros64(n)
33 func LeadingZeros32(n uint32) int {
34 // amd64:"BSRQ","LEAQ",-"CMOVQEQ"
36 // arm:"CLZ" arm64:"CLZW"
39 return bits.LeadingZeros32(n)
42 func LeadingZeros16(n uint16) int {
43 // amd64:"BSRL","LEAL",-"CMOVQEQ"
45 // arm:"CLZ" arm64:"CLZ"
48 return bits.LeadingZeros16(n)
51 func LeadingZeros8(n uint8) int {
52 // amd64:"BSRL","LEAL",-"CMOVQEQ"
54 // arm:"CLZ" arm64:"CLZ"
57 return bits.LeadingZeros8(n)
64 func Len(n uint) int {
67 // arm:"CLZ" arm64:"CLZ"
73 func Len64(n uint64) int {
76 // arm:"CLZ" arm64:"CLZ"
79 // ppc64le:"SUBC","CNTLZD"
80 // ppc64:"SUBC","CNTLZD"
84 func SubFromLen64(n uint64) int {
85 // ppc64le:"CNTLZD",-"SUBC"
86 // ppc64:"CNTLZD",-"SUBC"
87 return 64 - bits.Len64(n)
90 func Len32(n uint32) int {
91 // amd64:"BSRQ","LEAQ",-"CMOVQEQ"
93 // arm:"CLZ" arm64:"CLZ"
99 func Len16(n uint16) int {
100 // amd64:"BSRL","LEAL",-"CMOVQEQ"
102 // arm:"CLZ" arm64:"CLZ"
108 func Len8(n uint8) int {
109 // amd64:"BSRL","LEAL",-"CMOVQEQ"
111 // arm:"CLZ" arm64:"CLZ"
117 // -------------------- //
119 // -------------------- //
121 // TODO(register args) Restore a m d 6 4 :.*x86HasPOPCNT when only one ABI is tested.
122 func OnesCount(n uint) int {
124 // arm64:"VCNT","VUADDLV"
129 return bits.OnesCount(n)
132 func OnesCount64(n uint64) int {
134 // arm64:"VCNT","VUADDLV"
139 return bits.OnesCount64(n)
142 func OnesCount32(n uint32) int {
144 // arm64:"VCNT","VUADDLV"
149 return bits.OnesCount32(n)
152 func OnesCount16(n uint16) int {
154 // arm64:"VCNT","VUADDLV"
159 return bits.OnesCount16(n)
162 func OnesCount8(n uint8) int {
167 return bits.OnesCount8(n)
170 // ----------------------- //
171 // bits.ReverseBytes //
172 // ----------------------- //
174 func ReverseBytes(n uint) uint {
178 return bits.ReverseBytes(n)
181 func ReverseBytes64(n uint64) uint64 {
185 return bits.ReverseBytes64(n)
188 func ReverseBytes32(n uint32) uint32 {
192 return bits.ReverseBytes32(n)
195 func ReverseBytes16(n uint16) uint16 {
197 // arm64:"REV16W",-"UBFX",-"ORR"
198 // arm/5:"SLL","SRL","ORR"
201 return bits.ReverseBytes16(n)
204 // --------------------- //
205 // bits.RotateLeft //
206 // --------------------- //
208 func RotateLeft64(n uint64) uint64 {
213 // s390x:"RISBGZ\t[$]0, [$]63, [$]37, "
215 return bits.RotateLeft64(n, 37)
218 func RotateLeft32(n uint32) uint32 {
219 // amd64:"ROLL" 386:"ROLL"
220 // arm:`MOVW\tR[0-9]+@>23`
226 return bits.RotateLeft32(n, 9)
229 func RotateLeft16(n uint16) uint16 {
230 // amd64:"ROLW" 386:"ROLW"
231 return bits.RotateLeft16(n, 5)
234 func RotateLeft8(n uint8) uint8 {
235 // amd64:"ROLB" 386:"ROLB"
236 return bits.RotateLeft8(n, 5)
239 func RotateLeftVariable(n uint, m int) uint {
246 return bits.RotateLeft(n, m)
249 func RotateLeftVariable64(n uint64, m int) uint64 {
256 return bits.RotateLeft64(n, m)
259 func RotateLeftVariable32(n uint32, m int) uint32 {
260 // arm:`MOVW\tR[0-9]+@>R[0-9]+`
267 return bits.RotateLeft32(n, m)
270 // ------------------------ //
271 // bits.TrailingZeros //
272 // ------------------------ //
274 func TrailingZeros(n uint) int {
275 // amd64:"BSFQ","MOVL\t\\$64","CMOVQEQ"
277 // arm64:"RBIT","CLZ"
279 // ppc64/power8:"ANDN","POPCNTD"
280 // ppc64le/power8:"ANDN","POPCNTD"
281 // ppc64/power9: "CNTTZD"
282 // ppc64le/power9: "CNTTZD"
284 return bits.TrailingZeros(n)
287 func TrailingZeros64(n uint64) int {
288 // amd64:"BSFQ","MOVL\t\\$64","CMOVQEQ"
289 // arm64:"RBIT","CLZ"
291 // ppc64/power8:"ANDN","POPCNTD"
292 // ppc64le/power8:"ANDN","POPCNTD"
293 // ppc64/power9: "CNTTZD"
294 // ppc64le/power9: "CNTTZD"
296 return bits.TrailingZeros64(n)
299 func TrailingZeros64Subtract(n uint64) int {
300 // ppc64le/power8:"NEG","SUBC","ANDN","POPCNTD"
301 // ppc64le/power9:"SUBC","CNTTZD"
302 return bits.TrailingZeros64(1 - n)
305 func TrailingZeros32(n uint32) int {
306 // amd64:"BTSQ\\t\\$32","BSFQ"
308 // arm64:"RBITW","CLZW"
309 // s390x:"FLOGR","MOVWZ"
310 // ppc64/power8:"ANDN","POPCNTW"
311 // ppc64le/power8:"ANDN","POPCNTW"
312 // ppc64/power9: "CNTTZW"
313 // ppc64le/power9: "CNTTZW"
315 return bits.TrailingZeros32(n)
318 func TrailingZeros16(n uint16) int {
319 // amd64:"BSFL","BTSL\\t\\$16"
321 // arm:"ORR\t\\$65536","CLZ",-"MOVHU\tR"
322 // arm64:"ORR\t\\$65536","RBITW","CLZW",-"MOVHU\tR",-"RBIT\t",-"CLZ\t"
323 // s390x:"FLOGR","OR\t\\$65536"
324 // ppc64/power8:"POPCNTD","OR\\t\\$65536"
325 // ppc64le/power8:"POPCNTD","OR\\t\\$65536"
326 // ppc64/power9:"CNTTZD","OR\\t\\$65536"
327 // ppc64le/power9:"CNTTZD","OR\\t\\$65536"
329 return bits.TrailingZeros16(n)
332 func TrailingZeros8(n uint8) int {
333 // amd64:"BSFL","BTSL\\t\\$8"
334 // arm:"ORR\t\\$256","CLZ",-"MOVBU\tR"
335 // arm64:"ORR\t\\$256","RBITW","CLZW",-"MOVBU\tR",-"RBIT\t",-"CLZ\t"
336 // s390x:"FLOGR","OR\t\\$256"
338 return bits.TrailingZeros8(n)
341 // IterateBitsNN checks special handling of TrailingZerosNN when the input is known to be non-zero.
343 func IterateBits(n uint) int {
346 // amd64:"BSFQ",-"CMOVEQ"
347 i += bits.TrailingZeros(n)
353 func IterateBits64(n uint64) int {
356 // amd64:"BSFQ",-"CMOVEQ"
357 i += bits.TrailingZeros64(n)
363 func IterateBits32(n uint32) int {
366 // amd64:"BSFL",-"BTSQ"
367 i += bits.TrailingZeros32(n)
373 func IterateBits16(n uint16) int {
376 // amd64:"BSFL",-"BTSL"
377 // arm64:"RBITW","CLZW",-"ORR"
378 i += bits.TrailingZeros16(n)
384 func IterateBits8(n uint8) int {
387 // amd64:"BSFL",-"BTSL"
388 // arm64:"RBITW","CLZW",-"ORR"
389 i += bits.TrailingZeros8(n)
395 // --------------- //
397 // --------------- //
399 func Add(x, y, ci uint) (r, co uint) {
400 // arm64:"ADDS","ADCS","ADC",-"ADD\t",-"CMP"
401 // amd64:"NEGL","ADCQ","SBBQ","NEGQ"
402 // s390x:"ADDE","ADDC\t[$]-1,"
403 return bits.Add(x, y, ci)
406 func AddC(x, ci uint) (r, co uint) {
407 // arm64:"ADDS","ADCS","ADC",-"ADD\t",-"CMP"
408 // amd64:"NEGL","ADCQ","SBBQ","NEGQ"
409 // s390x:"ADDE","ADDC\t[$]-1,"
410 return bits.Add(x, 7, ci)
413 func AddZ(x, y uint) (r, co uint) {
414 // arm64:"ADDS","ADC",-"ADCS",-"ADD\t",-"CMP"
415 // amd64:"ADDQ","SBBQ","NEGQ",-"NEGL",-"ADCQ"
416 // s390x:"ADDC",-"ADDC\t[$]-1,"
417 return bits.Add(x, y, 0)
420 func AddR(x, y, ci uint) uint {
421 // arm64:"ADDS","ADCS",-"ADD\t",-"CMP"
422 // amd64:"NEGL","ADCQ",-"SBBQ",-"NEGQ"
423 // s390x:"ADDE","ADDC\t[$]-1,"
424 r, _ := bits.Add(x, y, ci)
428 func AddM(p, q, r *[3]uint) {
430 r[0], c = bits.Add(p[0], q[0], c)
431 // arm64:"ADCS",-"ADD\t",-"CMP"
432 // amd64:"ADCQ",-"NEGL",-"SBBQ",-"NEGQ"
433 // s390x:"ADDE",-"ADDC\t[$]-1,"
434 r[1], c = bits.Add(p[1], q[1], c)
435 r[2], c = bits.Add(p[2], q[2], c)
438 func Add64(x, y, ci uint64) (r, co uint64) {
439 // arm64:"ADDS","ADCS","ADC",-"ADD\t",-"CMP"
440 // amd64:"NEGL","ADCQ","SBBQ","NEGQ"
441 // ppc64: "ADDC", "ADDE", "ADDZE"
442 // ppc64le: "ADDC", "ADDE", "ADDZE"
443 // s390x:"ADDE","ADDC\t[$]-1,"
444 return bits.Add64(x, y, ci)
447 func Add64C(x, ci uint64) (r, co uint64) {
448 // arm64:"ADDS","ADCS","ADC",-"ADD\t",-"CMP"
449 // amd64:"NEGL","ADCQ","SBBQ","NEGQ"
450 // ppc64: "ADDC", "ADDE", "ADDZE"
451 // ppc64le: "ADDC", "ADDE", "ADDZE"
452 // s390x:"ADDE","ADDC\t[$]-1,"
453 return bits.Add64(x, 7, ci)
456 func Add64Z(x, y uint64) (r, co uint64) {
457 // arm64:"ADDS","ADC",-"ADCS",-"ADD\t",-"CMP"
458 // amd64:"ADDQ","SBBQ","NEGQ",-"NEGL",-"ADCQ"
459 // ppc64: "ADDC", "ADDE", "ADDZE"
460 // ppc64le: "ADDC", "ADDE", "ADDZE"
461 // s390x:"ADDC",-"ADDC\t[$]-1,"
462 return bits.Add64(x, y, 0)
465 func Add64R(x, y, ci uint64) uint64 {
466 // arm64:"ADDS","ADCS",-"ADD\t",-"CMP"
467 // amd64:"NEGL","ADCQ",-"SBBQ",-"NEGQ"
468 // ppc64: "ADDC", "ADDE", "ADDZE"
469 // ppc64le: "ADDC", "ADDE", "ADDZE"
470 // s390x:"ADDE","ADDC\t[$]-1,"
471 r, _ := bits.Add64(x, y, ci)
474 func Add64M(p, q, r *[3]uint64) {
476 r[0], c = bits.Add64(p[0], q[0], c)
477 // arm64:"ADCS",-"ADD\t",-"CMP"
478 // amd64:"ADCQ",-"NEGL",-"SBBQ",-"NEGQ"
479 // ppc64: "ADDC", "ADDE", "ADDZE"
480 // ppc64le: "ADDC", "ADDE", "ADDZE"
481 // s390x:"ADDE",-"ADDC\t[$]-1,"
482 r[1], c = bits.Add64(p[1], q[1], c)
483 r[2], c = bits.Add64(p[2], q[2], c)
486 func Add64PanicOnOverflowEQ(a, b uint64) uint64 {
487 r, c := bits.Add64(a, b, 0)
488 // s390x:"BRC\t[$]3,",-"ADDE"
495 func Add64PanicOnOverflowNE(a, b uint64) uint64 {
496 r, c := bits.Add64(a, b, 0)
497 // s390x:"BRC\t[$]3,",-"ADDE"
504 func Add64PanicOnOverflowGT(a, b uint64) uint64 {
505 r, c := bits.Add64(a, b, 0)
506 // s390x:"BRC\t[$]3,",-"ADDE"
513 func Add64MPanicOnOverflowEQ(a, b [2]uint64) [2]uint64 {
516 r[0], c = bits.Add64(a[0], b[0], c)
517 r[1], c = bits.Add64(a[1], b[1], c)
518 // s390x:"BRC\t[$]3,"
525 func Add64MPanicOnOverflowNE(a, b [2]uint64) [2]uint64 {
528 r[0], c = bits.Add64(a[0], b[0], c)
529 r[1], c = bits.Add64(a[1], b[1], c)
530 // s390x:"BRC\t[$]3,"
537 func Add64MPanicOnOverflowGT(a, b [2]uint64) [2]uint64 {
540 r[0], c = bits.Add64(a[0], b[0], c)
541 r[1], c = bits.Add64(a[1], b[1], c)
542 // s390x:"BRC\t[$]3,"
549 // --------------- //
551 // --------------- //
553 func Sub(x, y, ci uint) (r, co uint) {
554 // amd64:"NEGL","SBBQ","NEGQ"
555 // arm64:"NEGS","SBCS","NGC","NEG",-"ADD",-"SUB",-"CMP"
557 return bits.Sub(x, y, ci)
560 func SubC(x, ci uint) (r, co uint) {
561 // amd64:"NEGL","SBBQ","NEGQ"
562 // arm64:"NEGS","SBCS","NGC","NEG",-"ADD",-"SUB",-"CMP"
564 return bits.Sub(x, 7, ci)
567 func SubZ(x, y uint) (r, co uint) {
568 // amd64:"SUBQ","SBBQ","NEGQ",-"NEGL"
569 // arm64:"SUBS","NGC","NEG",-"SBCS",-"ADD",-"SUB\t",-"CMP"
571 return bits.Sub(x, y, 0)
574 func SubR(x, y, ci uint) uint {
575 // amd64:"NEGL","SBBQ",-"NEGQ"
576 // arm64:"NEGS","SBCS",-"NGC",-"NEG\t",-"ADD",-"SUB",-"CMP"
578 r, _ := bits.Sub(x, y, ci)
581 func SubM(p, q, r *[3]uint) {
583 r[0], c = bits.Sub(p[0], q[0], c)
584 // amd64:"SBBQ",-"NEGL",-"NEGQ"
585 // arm64:"SBCS",-"NEGS",-"NGC",-"NEG",-"ADD",-"SUB",-"CMP"
587 r[1], c = bits.Sub(p[1], q[1], c)
588 r[2], c = bits.Sub(p[2], q[2], c)
591 func Sub64(x, y, ci uint64) (r, co uint64) {
592 // amd64:"NEGL","SBBQ","NEGQ"
593 // arm64:"NEGS","SBCS","NGC","NEG",-"ADD",-"SUB",-"CMP"
595 return bits.Sub64(x, y, ci)
598 func Sub64C(x, ci uint64) (r, co uint64) {
599 // amd64:"NEGL","SBBQ","NEGQ"
600 // arm64:"NEGS","SBCS","NGC","NEG",-"ADD",-"SUB",-"CMP"
602 return bits.Sub64(x, 7, ci)
605 func Sub64Z(x, y uint64) (r, co uint64) {
606 // amd64:"SUBQ","SBBQ","NEGQ",-"NEGL"
607 // arm64:"SUBS","NGC","NEG",-"SBCS",-"ADD",-"SUB\t",-"CMP"
609 return bits.Sub64(x, y, 0)
612 func Sub64R(x, y, ci uint64) uint64 {
613 // amd64:"NEGL","SBBQ",-"NEGQ"
614 // arm64:"NEGS","SBCS",-"NGC",-"NEG\t",-"ADD",-"SUB",-"CMP"
616 r, _ := bits.Sub64(x, y, ci)
619 func Sub64M(p, q, r *[3]uint64) {
621 r[0], c = bits.Sub64(p[0], q[0], c)
622 // amd64:"SBBQ",-"NEGL",-"NEGQ"
623 // arm64:"SBCS",-"NEGS",-"NGC",-"NEG",-"ADD",-"SUB",-"CMP"
625 r[1], c = bits.Sub64(p[1], q[1], c)
626 r[2], c = bits.Sub64(p[2], q[2], c)
629 func Sub64PanicOnOverflowEQ(a, b uint64) uint64 {
630 r, b := bits.Sub64(a, b, 0)
631 // s390x:"BRC\t[$]12,",-"ADDE",-"SUBE"
638 func Sub64PanicOnOverflowNE(a, b uint64) uint64 {
639 r, b := bits.Sub64(a, b, 0)
640 // s390x:"BRC\t[$]12,",-"ADDE",-"SUBE"
647 func Sub64PanicOnOverflowGT(a, b uint64) uint64 {
648 r, b := bits.Sub64(a, b, 0)
649 // s390x:"BRC\t[$]12,",-"ADDE",-"SUBE"
656 func Sub64MPanicOnOverflowEQ(a, b [2]uint64) [2]uint64 {
659 r[0], c = bits.Sub64(a[0], b[0], c)
660 r[1], c = bits.Sub64(a[1], b[1], c)
661 // s390x:"BRC\t[$]12,"
668 func Sub64MPanicOnOverflowNE(a, b [2]uint64) [2]uint64 {
671 r[0], c = bits.Sub64(a[0], b[0], c)
672 r[1], c = bits.Sub64(a[1], b[1], c)
673 // s390x:"BRC\t[$]12,"
680 func Sub64MPanicOnOverflowGT(a, b [2]uint64) [2]uint64 {
683 r[0], c = bits.Sub64(a[0], b[0], c)
684 r[1], c = bits.Sub64(a[1], b[1], c)
685 // s390x:"BRC\t[$]12,"
692 // --------------- //
694 // --------------- //
696 func Mul(x, y uint) (hi, lo uint) {
698 // arm64:"UMULH","MUL"
699 // ppc64:"MULHDU","MULLD"
700 // ppc64le:"MULHDU","MULLD"
703 return bits.Mul(x, y)
706 func Mul64(x, y uint64) (hi, lo uint64) {
708 // arm64:"UMULH","MUL"
709 // ppc64:"MULHDU","MULLD"
710 // ppc64le:"MULHDU","MULLD"
713 // riscv64:"MULHU","MUL"
714 return bits.Mul64(x, y)
717 // --------------- //
719 // --------------- //
721 func Div(hi, lo, x uint) (q, r uint) {
723 return bits.Div(hi, lo, x)
726 func Div32(hi, lo, x uint32) (q, r uint32) {
727 // arm64:"ORR","UDIV","MSUB",-"UREM"
728 return bits.Div32(hi, lo, x)
731 func Div64(hi, lo, x uint64) (q, r uint64) {
733 return bits.Div64(hi, lo, x)
736 func Div64degenerate(x uint64) (q, r uint64) {
738 return bits.Div64(0, x, 5)