ev67-strrchr.S 3.3 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111
  1. /* SPDX-License-Identifier: GPL-2.0 */
  2. /*
  3. * arch/alpha/lib/ev67-strrchr.S
  4. * 21264 version by Rick Gorton <[email protected]>
  5. *
  6. * Finds length of a 0-terminated string. Optimized for the
  7. * Alpha architecture:
  8. *
  9. * - memory accessed as aligned quadwords only
  10. * - uses bcmpge to compare 8 bytes in parallel
  11. *
  12. * Much of the information about 21264 scheduling/coding comes from:
  13. * Compiler Writer's Guide for the Alpha 21264
  14. * abbreviated as 'CWG' in other comments here
  15. * ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
  16. * Scheduling notation:
  17. * E - either cluster
  18. * U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1
  19. * L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1
  20. */
  21. #include <asm/export.h>
  22. #include <asm/regdef.h>
  23. .set noreorder
  24. .set noat
  25. .align 4
  26. .ent strrchr
  27. .globl strrchr
  28. strrchr:
  29. .frame sp, 0, ra
  30. .prologue 0
  31. and a1, 0xff, t2 # E : 00000000000000ch
  32. insbl a1, 1, t4 # U : 000000000000ch00
  33. insbl a1, 2, t5 # U : 0000000000ch0000
  34. ldq_u t0, 0(a0) # L : load first quadword Latency=3
  35. mov zero, t6 # E : t6 is last match aligned addr
  36. or t2, t4, a1 # E : 000000000000chch
  37. sll t5, 8, t3 # U : 00000000ch000000
  38. mov zero, t8 # E : t8 is last match byte compare mask
  39. andnot a0, 7, v0 # E : align source addr
  40. or t5, t3, t3 # E : 00000000chch0000
  41. sll a1, 32, t2 # U : 0000chch00000000
  42. sll a1, 48, t4 # U : chch000000000000
  43. or t4, a1, a1 # E : chch00000000chch
  44. or t2, t3, t2 # E : 0000chchchch0000
  45. or a1, t2, a1 # E : chchchchchchchch
  46. lda t5, -1 # E : build garbage mask
  47. cmpbge zero, t0, t1 # E : bits set iff byte == zero
  48. mskqh t5, a0, t4 # E : Complete garbage mask
  49. xor t0, a1, t2 # E : make bytes == c zero
  50. cmpbge zero, t4, t4 # E : bits set iff byte is garbage
  51. cmpbge zero, t2, t3 # E : bits set iff byte == c
  52. andnot t1, t4, t1 # E : clear garbage from null test
  53. andnot t3, t4, t3 # E : clear garbage from char test
  54. bne t1, $eos # U : did we already hit the terminator?
  55. /* Character search main loop */
  56. $loop:
  57. ldq t0, 8(v0) # L : load next quadword
  58. cmovne t3, v0, t6 # E : save previous comparisons match
  59. nop # : Latency=2, extra map slot (keep nop with cmov)
  60. nop
  61. cmovne t3, t3, t8 # E : Latency=2, extra map slot
  62. nop # : keep with cmovne
  63. addq v0, 8, v0 # E :
  64. xor t0, a1, t2 # E :
  65. cmpbge zero, t0, t1 # E : bits set iff byte == zero
  66. cmpbge zero, t2, t3 # E : bits set iff byte == c
  67. beq t1, $loop # U : if we havnt seen a null, loop
  68. nop
  69. /* Mask out character matches after terminator */
  70. $eos:
  71. negq t1, t4 # E : isolate first null byte match
  72. and t1, t4, t4 # E :
  73. subq t4, 1, t5 # E : build a mask of the bytes up to...
  74. or t4, t5, t4 # E : ... and including the null
  75. and t3, t4, t3 # E : mask out char matches after null
  76. cmovne t3, t3, t8 # E : save it, if match found Latency=2, extra map slot
  77. nop # : Keep with cmovne
  78. nop
  79. cmovne t3, v0, t6 # E :
  80. nop # : Keep with cmovne
  81. /* Locate the address of the last matched character */
  82. ctlz t8, t2 # U0 : Latency=3 (0x40 for t8=0)
  83. nop
  84. cmoveq t8, 0x3f, t2 # E : Compensate for case when no match is seen
  85. nop # E : hide the cmov latency (2) behind ctlz latency
  86. lda t5, 0x3f($31) # E :
  87. subq t5, t2, t5 # E : Normalize leading zero count
  88. addq t6, t5, v0 # E : and add to quadword address
  89. ret # L0 : Latency=3
  90. nop
  91. nop
  92. .end strrchr
  93. EXPORT_SYMBOL(strrchr)