stdlib/rshift.shared.s

   1  .text
   2 .globl __mpn_rshift
   3 .type __mpn_rshift,@function
   4 .align 1<<4
   5 __mpn_rshift:
   6  mov %edx, %eax
   7  and $3, %eax
   8  jne .Lnb00
   9 .Lb00:
  10  mov (%rsi), %r10
  11  mov 8(%rsi), %r11
  12  xor %eax, %eax
  13  shrd %cl, %r10, %rax
  14  mov 16(%rsi), %r8
  15  lea 8(%rsi), %rsi
  16  lea -24(%rdi), %rdi
  17  sub $4, %rdx
  18  jmp .L00
  19 .Lnb00:
  20  cmp $2, %eax
  21  jae .Lnb01
  22 .Lb01: mov (%rsi), %r9
  23  xor %eax, %eax
  24  shrd %cl, %r9, %rax
  25  sub $2, %rdx
  26  jb .Lle1
  27  mov 8(%rsi), %r10
  28  mov 16(%rsi), %r11
  29  lea 16(%rsi), %rsi
  30  lea -16(%rdi), %rdi
  31  jmp .L01
  32 .Lle1: shr %cl, %r9
  33  mov %r9, (%rdi)
  34  ret
  35 .Lnb01:
  36  jne .Lb11
  37 .Lb10: mov (%rsi), %r8
  38  mov 8(%rsi), %r9
  39  xor %eax, %eax
  40  shrd %cl, %r8, %rax
  41  sub $3, %rdx
  42  jb .Lle2
  43  mov 16(%rsi), %r10
  44  lea 24(%rsi), %rsi
  45  lea -8(%rdi), %rdi
  46  jmp .L10
  47 .Lle2: shrd %cl, %r9, %r8
  48  mov %r8, (%rdi)
  49  shr %cl, %r9
  50  mov %r9, 8(%rdi)
  51  ret
  52  .p2align 4
  53 .Lb11:
  54  mov (%rsi), %r11
  55  mov 8(%rsi), %r8
  56  xor %eax, %eax
  57  shrd %cl, %r11, %rax
  58  mov 16(%rsi), %r9
  59  lea 32(%rsi), %rsi
  60  sub $4, %rdx
  61  jb .Lend
  62  .p2align 4
  63 .Ltop: shrd %cl, %r8, %r11
  64  mov -8(%rsi), %r10
  65  mov %r11, (%rdi)
  66 .L10: shrd %cl, %r9, %r8
  67  mov (%rsi), %r11
  68  mov %r8, 8(%rdi)
  69 .L01: shrd %cl, %r10, %r9
  70  mov 8(%rsi), %r8
  71  mov %r9, 16(%rdi)
  72 .L00: shrd %cl, %r11, %r10
  73  mov 16(%rsi), %r9
  74  mov %r10, 24(%rdi)
  75  add $32, %rsi
  76  lea 32(%rdi), %rdi
  77  sub $4, %rdx
  78  jnc .Ltop
  79 .Lend: shrd %cl, %r8, %r11
  80  mov %r11, (%rdi)
  81  shrd %cl, %r9, %r8
  82  mov %r8, 8(%rdi)
  83  shr %cl, %r9
  84  mov %r9, 16(%rdi)
  85  ret
  86 .size __mpn_rshift,.-__mpn_rshift