strchr.S 2.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172
  1. /* SPDX-License-Identifier: GPL-2.0 */
  2. /*
  3. * arch/alpha/lib/strchr.S
  4. * Contributed by Richard Henderson ([email protected])
  5. *
  6. * Return the address of a given character within a null-terminated
  7. * string, or null if it is not found.
  8. */
  9. #include <asm/export.h>
  10. #include <asm/regdef.h>
  11. .set noreorder
  12. .set noat
  13. .align 3
  14. .globl strchr
  15. .ent strchr
  16. strchr:
  17. .frame sp, 0, ra
  18. .prologue 0
  19. zapnot a1, 1, a1 # e0 : zero extend the search character
  20. ldq_u t0, 0(a0) # .. e1 : load first quadword
  21. sll a1, 8, t5 # e0 : replicate the search character
  22. andnot a0, 7, v0 # .. e1 : align our loop pointer
  23. or t5, a1, a1 # e0 :
  24. lda t4, -1 # .. e1 : build garbage mask
  25. sll a1, 16, t5 # e0 :
  26. cmpbge zero, t0, t2 # .. e1 : bits set iff byte == zero
  27. mskqh t4, a0, t4 # e0 :
  28. or t5, a1, a1 # .. e1 :
  29. sll a1, 32, t5 # e0 :
  30. cmpbge zero, t4, t4 # .. e1 : bits set iff byte is garbage
  31. or t5, a1, a1 # e0 :
  32. xor t0, a1, t1 # .. e1 : make bytes == c zero
  33. cmpbge zero, t1, t3 # e0 : bits set iff byte == c
  34. or t2, t3, t0 # e1 : bits set iff char match or zero match
  35. andnot t0, t4, t0 # e0 : clear garbage bits
  36. bne t0, $found # .. e1 (zdb)
  37. $loop: ldq t0, 8(v0) # e0 :
  38. addq v0, 8, v0 # .. e1 :
  39. nop # e0 :
  40. xor t0, a1, t1 # .. e1 (ev5 data stall)
  41. cmpbge zero, t0, t2 # e0 : bits set iff byte == 0
  42. cmpbge zero, t1, t3 # .. e1 : bits set iff byte == c
  43. or t2, t3, t0 # e0 :
  44. beq t0, $loop # .. e1 (zdb)
  45. $found: negq t0, t1 # e0 : clear all but least set bit
  46. and t0, t1, t0 # e1 (stall)
  47. and t0, t3, t1 # e0 : bit set iff byte was the char
  48. beq t1, $retnull # .. e1 (zdb)
  49. and t0, 0xf0, t2 # e0 : binary search for that set bit
  50. and t0, 0xcc, t3 # .. e1 :
  51. and t0, 0xaa, t4 # e0 :
  52. cmovne t2, 4, t2 # .. e1 :
  53. cmovne t3, 2, t3 # e0 :
  54. cmovne t4, 1, t4 # .. e1 :
  55. addq t2, t3, t2 # e0 :
  56. addq v0, t4, v0 # .. e1 :
  57. addq v0, t2, v0 # e0 :
  58. ret # .. e1 :
  59. $retnull:
  60. mov zero, v0 # e0 :
  61. ret # .. e1 :
  62. .end strchr
  63. EXPORT_SYMBOL(strchr)