src/gallium/drivers/nouveau/codegen/target_lib_nvf0.asm

   1 //
   2 // DIV U32
   3 //
   4 // UNR recurrence (q = a / b):
   5 // look for z such that 2^32 - b <= b * z < 2^32
   6 // then q - 1 <= (a * z) / 2^32 <= q
   7 //
   8 // INPUT:   $r0: dividend, $r1: divisor
   9 // OUTPUT:  $r0: result, $r1: modulus
  10 // CLOBBER: $r2 - $r3, $p0 - $p1
  11 // SIZE:    22 / 14 * 8 bytes
  12 //
  13 sched 0x28282804280428
  14 bfind u32 $r2 $r1
  15 xor b32 $r2 $r2 0x1f
  16 mov b32 $r3 0x1
  17 shl b32 $r2 $r3 clamp $r2
  18 cvt u32 $r1 neg u32 $r1
  19 mul $r3 u32 $r1 u32 $r2
  20 add $r2 (mul high u32 $r2 u32 $r3) $r2
  21 sched 0x28282828282828
  22 mul $r3 u32 $r1 u32 $r2
  23 add $r2 (mul high u32 $r2 u32 $r3) $r2
  24 mul $r3 u32 $r1 u32 $r2
  25 add $r2 (mul high u32 $r2 u32 $r3) $r2
  26 mul $r3 u32 $r1 u32 $r2
  27 add $r2 (mul high u32 $r2 u32 $r3) $r2
  28 mul $r3 u32 $r1 u32 $r2
  29 sched 0x042c2828042804
  30 add $r2 (mul high u32 $r2 u32 $r3) $r2
  31 mov b32 $r3 $r0
  32 mul high $r0 u32 $r0 u32 $r2
  33 cvt u32 $r2 neg u32 $r1
  34 add $r1 (mul u32 $r1 u32 $r0) $r3
  35 set $p0 0x1 ge u32 $r1 $r2
  36 $p0 sub b32 $r1 $r1 $r2
  37 sched 0x20282e20042c28
  38 $p0 add b32 $r0 $r0 0x1
  39 $p0 set $p0 0x1 ge u32 $r1 $r2
  40 $p0 sub b32 $r1 $r1 $r2
  41 $p0 add b32 $r0 $r0 0x1
  42 ret
  43 //
  44 // DIV S32, like DIV U32 after taking ABS(inputs)
  45 //
  46 // INPUT:   $r0: dividend, $r1: divisor
  47 // OUTPUT:  $r0: result, $r1: modulus
  48 // CLOBBER: $r2 - $r3, $p0 - $p3
  49 //
  50 set $p2 0x1 lt s32 $r0 0x0
  51 set $p3 0x1 lt s32 $r1 0x0 xor $p2
  52 sched 0x28042804282820
  53 cvt s32 $r0 abs s32 $r0
  54 cvt s32 $r1 abs s32 $r1
  55 bfind u32 $r2 $r1
  56 xor b32 $r2 $r2 0x1f
  57 mov b32 $r3 0x1
  58 shl b32 $r2 $r3 clamp $r2
  59 cvt u32 $r1 neg u32 $r1
  60 sched 0x28282828282828
  61 mul $r3 u32 $r1 u32 $r2
  62 add $r2 (mul high u32 $r2 u32 $r3) $r2
  63 mul $r3 u32 $r1 u32 $r2
  64 add $r2 (mul high u32 $r2 u32 $r3) $r2
  65 mul $r3 u32 $r1 u32 $r2
  66 add $r2 (mul high u32 $r2 u32 $r3) $r2
  67 mul $r3 u32 $r1 u32 $r2
  68 sched 0x28280428042828
  69 add $r2 (mul high u32 $r2 u32 $r3) $r2
  70 mul $r3 u32 $r1 u32 $r2
  71 add $r2 (mul high u32 $r2 u32 $r3) $r2
  72 mov b32 $r3 $r0
  73 mul high $r0 u32 $r0 u32 $r2
  74 cvt u32 $r2 neg u32 $r1
  75 add $r1 (mul u32 $r1 u32 $r0) $r3
  76 sched 0x2028042c28042c
  77 set $p0 0x1 ge u32 $r1 $r2
  78 $p0 sub b32 $r1 $r1 $r2
  79 $p0 add b32 $r0 $r0 0x1
  80 $p0 set $p0 0x1 ge u32 $r1 $r2
  81 $p0 sub b32 $r1 $r1 $r2
  82 $p0 add b32 $r0 $r0 0x1
  83 $p3 cvt s32 $r0 neg s32 $r0
  84 sched 0x2c200428042e04
  85 $p2 cvt s32 $r1 neg s32 $r1
  86 ret