fixed

Pure Haskell large fixed-width integers and Montgomery arithmetic (docs.ppad.tech/fixed).
git clone git://git.ppad.tech/fixed.git
Log | Files | Refs | README | LICENSE

Curve.hs (37816B)


      1 {-# LANGUAGE BangPatterns #-}
      2 {-# LANGUAGE MagicHash #-}
      3 {-# LANGUAGE NumericUnderscores #-}
      4 {-# LANGUAGE PatternSynonyms #-}
      5 {-# LANGUAGE ViewPatterns #-}
      6 {-# LANGUAGE UnboxedSums #-}
      7 {-# LANGUAGE UnboxedTuples #-}
      8 {-# LANGUAGE UnliftedNewtypes #-}
      9 
     10 -- |
     11 -- Module: Numeric.Montgomery.Secp256k1.Curve
     12 -- Copyright: (c) 2025 Jared Tobin
     13 -- License: MIT
     14 -- Maintainer: Jared Tobin <jared@ppad.tech>
     15 --
     16 -- Montgomery form 'Wider' words, as well as arithmetic operations, with
     17 -- domain derived from the secp256k1 elliptic curve field prime.
     18 
     19 module Numeric.Montgomery.Secp256k1.Curve (
     20   -- * Montgomery form, secp256k1 field prime modulus
     21     Montgomery(..)
     22   , render
     23   , to
     24   , from
     25   , zero
     26   , one
     27 
     28   -- * Comparison
     29   , eq
     30   , eq_vartime
     31 
     32   -- * Reduction and retrieval
     33   , redc
     34   , redc#
     35   , retr
     36   , retr#
     37 
     38   -- * Constant-time selection
     39   , select
     40   , select#
     41 
     42   -- * Montgomery arithmetic
     43   , add
     44   , add#
     45   , sub
     46   , sub#
     47   , mul
     48   , mul#
     49   , sqr
     50   , sqr#
     51   , neg
     52   , neg#
     53   , inv
     54   , inv#
     55   , sqrt_vartime
     56   , sqrt#
     57   , exp
     58   , exp#
     59   , odd#
     60   , odd_vartime
     61   ) where
     62 
     63 import Control.DeepSeq
     64 import qualified Data.Choice as C
     65 import Data.Word.Limb (Limb(..))
     66 import qualified Data.Word.Limb as L
     67 import qualified Data.Word.Wide as W
     68 import Data.Word.Wider (Wider(..))
     69 import qualified Data.Word.Wider as WW
     70 import GHC.Exts (Word(..), Word#)
     71 import Prelude hiding (or, and, not, sqrt, exp)
     72 
     73 -- montgomery arithmetic, specialized to the secp256k1 field prime modulus
     74 -- 0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFEFFFFFC2F
     75 
     76 -- | Montgomery-form 'Wider' words, on the Montgomery domain defined by
     77 --   the secp256k1 field prime.
     78 --
     79 --   >>> let one = 1 :: Montgomery
     80 --   >>> one
     81 --   1
     82 --   >>> putStrLn (render one)
     83 --   (4294968273, 0, 0, 0)
     84 data Montgomery = Montgomery !Limb4
     85 
     86 -- | Render a 'Montgomery' value as a 'String', showing its individual
     87 --   'Limb's.
     88 --
     89 --   >>> putStrLn (render 1)
     90 --   (4294968273, 0, 0, 0)
     91 render :: Montgomery -> String
     92 render (Montgomery (L4 a b c d)) =
     93      "(" <> show (W# a) <> ", " <> show (W# b) <> ", "
     94   <> show (W# c) <> ", " <> show (W# d) <> ")"
     95 
     96 instance Show Montgomery where
     97   show = show . from
     98 
     99 -- | Note that 'fromInteger' necessarily runs in variable time due
    100 --   to conversion from the variable-size, potentially heap-allocated
    101 --   'Integer' type.
    102 instance Num Montgomery where
    103   a + b = add a b
    104   a - b = sub a b
    105   a * b = mul a b
    106   negate a = neg a
    107   abs = id
    108   fromInteger = to . WW.to_vartime
    109   signum (Montgomery (# l0, l1, l2, l3 #)) =
    110     let !(Limb l) = l0 `L.or#` l1 `L.or#` l2 `L.or#` l3
    111         !n = C.from_word_nonzero# l
    112         !b = C.to_word# n
    113     in  Montgomery (# Limb b, Limb 0##, Limb 0##, Limb 0## #)
    114 
    115 instance NFData Montgomery where
    116   rnf (Montgomery a) = case a of (# _, _, _, _ #) -> ()
    117 
    118 -- utilities ------------------------------------------------------------------
    119 
    120 type Limb2 = (# Limb, Limb #)
    121 
    122 type Limb4 = (# Limb, Limb, Limb, Limb #)
    123 
    124 pattern L4 :: Word# -> Word# -> Word# -> Word# -> Limb4
    125 pattern L4 w0 w1 w2 w3 = (# Limb w0, Limb w1, Limb w2, Limb w3 #)
    126 {-# COMPLETE L4 #-}
    127 
    128 -- Wide wrapping addition, when addend is only a limb.
    129 wadd_w# :: Limb2 -> Limb -> Limb2
    130 wadd_w# (# x_lo, x_hi #) y_lo =
    131   let !(# s0, c0 #) = L.add_o# x_lo y_lo
    132       !(# s1, _ #) = L.add_o# x_hi c0
    133   in  (# s0, s1 #)
    134 {-# INLINE wadd_w# #-}
    135 
    136 -- Truncate a wide word to a 'Limb'.
    137 lo :: Limb2 -> Limb
    138 lo (# l, _ #) = l
    139 {-# INLINE lo #-}
    140 
    141 -- comparison -----------------------------------------------------------------
    142 
    143 -- | Constant-time equality comparison.
    144 eq :: Montgomery -> Montgomery -> C.Choice
    145 eq (Montgomery (L4 a0 a1 a2 a3)) (Montgomery (L4 b0 b1 b2 b3)) =
    146   C.eq_wider# (# a0, a1, a2, a3 #) (# b0, b1, b2, b3 #)
    147 {-# INLINE eq #-}
    148 
    149 -- | Variable-time equality comparison.
    150 eq_vartime :: Montgomery -> Montgomery -> Bool
    151 eq_vartime (Montgomery (Wider -> a)) (Montgomery (Wider -> b)) =
    152   WW.eq_vartime a b
    153 
    154 -- innards --------------------------------------------------------------------
    155 
    156 redc_inner#
    157   :: Limb4             -- ^ upper limbs
    158   -> Limb4             -- ^ lower limbs
    159   -> (# Limb4, Limb #) -- ^ upper limbs, meta-carry
    160 redc_inner# (# u0, u1, u2, u3 #) (# l0, l1, l2, l3 #) =
    161   let !(# m0, m1, m2, m3 #) =
    162         (# Limb 0xFFFFFFFEFFFFFC2F##, Limb 0xFFFFFFFFFFFFFFFF##
    163         ,  Limb 0xFFFFFFFFFFFFFFFF##, Limb 0xFFFFFFFFFFFFFFFF## #)
    164       !n                = Limb 0xD838091DD2253531##
    165       !w_0              = L.mul_w# l0 n
    166       !(# _, c_00 #)    = L.mac# w_0 m0 l0 (Limb 0##)
    167       !(# l0_1, c_01 #) = L.mac# w_0 m1 l1 c_00
    168       !(# l0_2, c_02 #) = L.mac# w_0 m2 l2 c_01
    169       !(# l0_3, c_03 #) = L.mac# w_0 m3 l3 c_02
    170       !(# u_0, mc_0 #)  = L.add_c# u0 c_03 (Limb 0##)
    171       !w_1              = L.mul_w# l0_1 n
    172       !(# _, c_10 #)    = L.mac# w_1 m0 l0_1 (Limb 0##)
    173       !(# l1_1, c_11 #) = L.mac# w_1 m1 l0_2 c_10
    174       !(# l1_2, c_12 #) = L.mac# w_1 m2 l0_3 c_11
    175       !(# u1_3, c_13 #) = L.mac# w_1 m3 u_0 c_12
    176       !(# u_1, mc_1 #)  = L.add_c# u1 c_13 mc_0
    177       !w_2              = L.mul_w# l1_1 n
    178       !(# _, c_20 #)    = L.mac# w_2 m0 l1_1 (Limb 0##)
    179       !(# l2_1, c_21 #) = L.mac# w_2 m1 l1_2 c_20
    180       !(# u2_2, c_22 #) = L.mac# w_2 m2 u1_3 c_21
    181       !(# u2_3, c_23 #) = L.mac# w_2 m3 u_1 c_22
    182       !(# u_2, mc_2 #)  = L.add_c# u2 c_23 mc_1
    183       !w_3              = L.mul_w# l2_1 n
    184       !(# _, c_30 #)    = L.mac# w_3 m0 l2_1 (Limb 0##)
    185       !(# u3_1, c_31 #) = L.mac# w_3 m1 u2_2 c_30
    186       !(# u3_2, c_32 #) = L.mac# w_3 m2 u2_3 c_31
    187       !(# u3_3, c_33 #) = L.mac# w_3 m3 u_2 c_32
    188       !(# u_3, mc_3 #)  = L.add_c# u3 c_33 mc_2
    189   in  (# (# u3_1, u3_2, u3_3, u_3 #), mc_3 #)
    190 {-# INLINE redc_inner# #-}
    191 
    192 -- | Montgomery reduction.
    193 redc#
    194   :: Limb4 -- ^ lower limbs
    195   -> Limb4 -- ^ upper limbs
    196   -> Limb4 -- ^ result
    197 redc# l u =
    198   let -- field prime
    199       !m = L4 0xFFFFFFFEFFFFFC2F## 0xFFFFFFFFFFFFFFFF##
    200               0xFFFFFFFFFFFFFFFF## 0xFFFFFFFFFFFFFFFF##
    201       !(# nu, mc #) = redc_inner# u l
    202   in  WW.sub_mod_c# nu mc m m
    203 {-# INLINE redc# #-}
    204 
    205 -- | Montgomery reduction.
    206 --
    207 --   The first argument represents the low words, and the second the
    208 --   high words, of an extra-large eight-limb word in Montgomery form.
    209 redc
    210   :: Montgomery -- ^ low wider-word, Montgomery form
    211   -> Montgomery -- ^ high wider-word, Montgomery form
    212   -> Montgomery -- ^ reduced value
    213 redc (Montgomery l) (Montgomery u) =
    214   let !res = redc# l u
    215   in  Montgomery res
    216 
    217 retr_inner#
    218   :: Limb4 -- ^ value in montgomery form
    219   -> Limb4 -- ^ retrieved value
    220 retr_inner# (# x0, x1, x2, x3 #) =
    221   let !(# m0, m1, m2, m3 #) =
    222         L4 0xFFFFFFFEFFFFFC2F## 0xFFFFFFFFFFFFFFFF##
    223            0xFFFFFFFFFFFFFFFF## 0xFFFFFFFFFFFFFFFF##
    224       !n                = Limb 0xD838091DD2253531##
    225       !u_0              = L.mul_w# x0 n
    226       !(# _, o0 #)      = L.mac# u_0 m0 x0 (Limb 0##)
    227       !(# o0_1, p0_1 #) = L.mac# u_0 m1 (Limb 0##) o0
    228       !(# p0_2, q0_2 #) = L.mac# u_0 m2 (Limb 0##) p0_1
    229       !(# q0_3, r0_3 #) = L.mac# u_0 m3 (Limb 0##) q0_2
    230       !u_1              = L.mul_w# (L.add_w# o0_1 x1) n
    231       !(# _, o1 #)      = L.mac# u_1 m0 x1 o0_1
    232       !(# o1_1, p1_1 #) = L.mac# u_1 m1 p0_2 o1
    233       !(# p1_2, q1_2 #) = L.mac# u_1 m2 q0_3 p1_1
    234       !(# q1_3, r1_3 #) = L.mac# u_1 m3 r0_3 q1_2
    235       !u_2              = L.mul_w# (L.add_w# o1_1 x2) n
    236       !(# _, o2 #)      = L.mac# u_2 m0 x2 o1_1
    237       !(# o2_1, p2_1 #) = L.mac# u_2 m1 p1_2 o2
    238       !(# p2_2, q2_2 #) = L.mac# u_2 m2 q1_3 p2_1
    239       !(# q2_3, r2_3 #) = L.mac# u_2 m3 r1_3 q2_2
    240       !u_3              = L.mul_w# (L.add_w# o2_1 x3) n
    241       !(# _, o3 #)      = L.mac# u_3 m0 x3 o2_1
    242       !(# o3_1, p3_1 #) = L.mac# u_3 m1 p2_2 o3
    243       !(# p3_2, q3_2 #) = L.mac# u_3 m2 q2_3 p3_1
    244       !(# q3_3, r3_3 #) = L.mac# u_3 m3 r2_3 q3_2
    245   in  (# o3_1, p3_2, q3_3, r3_3 #)
    246 {-# INLINE retr_inner# #-}
    247 
    248 retr#
    249   :: Limb4 -- montgomery form
    250   -> Limb4
    251 retr# f = retr_inner# f
    252 {-# INLINE retr# #-}
    253 
    254 -- | Retrieve a 'Montgomery' value from the Montgomery domain, producing
    255 --   a 'Wider' word.
    256 retr
    257   :: Montgomery -- ^ value in montgomery form
    258   -> Wider      -- ^ retrieved value
    259 retr (Montgomery f) =
    260   let !res = retr# f
    261   in  (Wider res)
    262 
    263 -- | Montgomery multiplication (FIOS), without conditional subtract.
    264 mul_inner#
    265   :: Limb4              -- ^ x
    266   -> Limb4              -- ^ y
    267   -> (# Limb4, Limb #)  -- ^ product, meta-carry
    268 mul_inner# (# x0, x1, x2, x3 #) (# y0, y1, y2, y3 #) =
    269   let !(# m0, m1, m2, m3 #) =
    270         L4 0xFFFFFFFEFFFFFC2F## 0xFFFFFFFFFFFFFFFF##
    271            0xFFFFFFFFFFFFFFFF## 0xFFFFFFFFFFFFFFFF##
    272       !n                           = Limb 0xD838091DD2253531##
    273       !axy0                        = L.mul_c# x0 y0
    274       !u0                          = L.mul_w# (lo axy0) n
    275       !(# (# _, a0 #), c0 #)       = W.add_o# (L.mul_c# u0 m0) axy0
    276       !carry0                      = (# a0, c0 #)
    277       !axy0_1                      = L.mul_c# x0 y1
    278       !umc0_1                      = W.add_w# (L.mul_c# u0 m1) carry0
    279       !(# (# o0, ab0_1 #), c0_1 #) = W.add_o# axy0_1 umc0_1
    280       !carry0_1                    = (# ab0_1, c0_1 #)
    281       !axy0_2                      = L.mul_c# x0 y2
    282       !umc0_2                      = W.add_w# (L.mul_c# u0 m2) carry0_1
    283       !(# (# p0, ab0_2 #), c0_2 #) = W.add_o# axy0_2 umc0_2
    284       !carry0_2                    = (# ab0_2, c0_2 #)
    285       !axy0_3                      = L.mul_c# x0 y3
    286       !umc0_3                      = W.add_w# (L.mul_c# u0 m3) carry0_2
    287       !(# (# q0, ab0_3 #), c0_3 #) = W.add_o# axy0_3 umc0_3
    288       !carry0_3                    = (# ab0_3, c0_3 #)
    289       !(# r0, mc0 #)               = carry0_3
    290       !axy1                        = wadd_w# (L.mul_c# x1 y0) o0
    291       !u1                          = L.mul_w# (lo axy1) n
    292       !(# (# _, a1 #), c1 #)       = W.add_o# (L.mul_c# u1 m0) axy1
    293       !carry1                      = (# a1, c1 #)
    294       !axy1_1                      = wadd_w# (L.mul_c# x1 y1) p0
    295       !umc1_1                      = W.add_w# (L.mul_c# u1 m1) carry1
    296       !(# (# o1, ab1_1 #), c1_1 #) = W.add_o# axy1_1 umc1_1
    297       !carry1_1                    = (# ab1_1, c1_1 #)
    298       !axy1_2                      = wadd_w# (L.mul_c# x1 y2) q0
    299       !umc1_2                      = W.add_w# (L.mul_c# u1 m2) carry1_1
    300       !(# (# p1, ab1_2 #), c1_2 #) = W.add_o# axy1_2 umc1_2
    301       !carry1_2                    = (# ab1_2, c1_2 #)
    302       !axy1_3                      = wadd_w# (L.mul_c# x1 y3) r0
    303       !umc1_3                      = W.add_w# (L.mul_c# u1 m3) carry1_2
    304       !(# (# q1, ab1_3 #), c1_3 #) = W.add_o# axy1_3 umc1_3
    305       !carry1_3                    = (# ab1_3, c1_3 #)
    306       !(# r1, mc1 #)               = wadd_w# carry1_3 mc0
    307       !axy2                        = wadd_w# (L.mul_c# x2 y0) o1
    308       !u2                          = L.mul_w# (lo axy2) n
    309       !(# (# _, a2 #), c2 #)       = W.add_o# (L.mul_c# u2 m0) axy2
    310       !carry2                      = (# a2, c2 #)
    311       !axy2_1                      = wadd_w# (L.mul_c# x2 y1) p1
    312       !umc2_1                      = W.add_w# (L.mul_c# u2 m1) carry2
    313       !(# (# o2, ab2_1 #), c2_1 #) = W.add_o# axy2_1 umc2_1
    314       !carry2_1                    = (# ab2_1, c2_1 #)
    315       !axy2_2                      = wadd_w# (L.mul_c# x2 y2) q1
    316       !umc2_2                      = W.add_w# (L.mul_c# u2 m2) carry2_1
    317       !(# (# p2, ab2_2 #), c2_2 #) = W.add_o# axy2_2 umc2_2
    318       !carry2_2                    = (# ab2_2, c2_2 #)
    319       !axy2_3                      = wadd_w# (L.mul_c# x2 y3) r1
    320       !umc2_3                      = W.add_w# (L.mul_c# u2 m3) carry2_2
    321       !(# (# q2, ab2_3 #), c2_3 #) = W.add_o# axy2_3 umc2_3
    322       !carry2_3                    = (# ab2_3, c2_3 #)
    323       !(# r2, mc2 #)               = wadd_w# carry2_3 mc1
    324       !axy3                        = wadd_w# (L.mul_c# x3 y0) o2
    325       !u3                          = L.mul_w# (lo axy3) n
    326       !(# (# _, a3 #), c3 #)       = W.add_o# (L.mul_c# u3 m0) axy3
    327       !carry3                      = (# a3, c3 #)
    328       !axy3_1                      = wadd_w# (L.mul_c# x3 y1) p2
    329       !umc3_1                      = W.add_w# (L.mul_c# u3 m1) carry3
    330       !(# (# o3, ab3_1 #), c3_1 #) = W.add_o# axy3_1 umc3_1
    331       !carry3_1                    = (# ab3_1, c3_1 #)
    332       !axy3_2                      = wadd_w# (L.mul_c# x3 y2) q2
    333       !umc3_2                      = W.add_w# (L.mul_c# u3 m2) carry3_1
    334       !(# (# p3, ab3_2 #), c3_2 #) = W.add_o# axy3_2 umc3_2
    335       !carry3_2                    = (# ab3_2, c3_2 #)
    336       !axy3_3                      = wadd_w# (L.mul_c# x3 y3) r2
    337       !umc3_3                      = W.add_w# (L.mul_c# u3 m3) carry3_2
    338       !(# (# q3, ab3_3 #), c3_3 #) = W.add_o# axy3_3 umc3_3
    339       !carry3_3                    = (# ab3_3, c3_3 #)
    340       !(# r3, mc3 #)               = wadd_w# carry3_3 mc2
    341   in  (# (# o3, p3, q3, r3 #), mc3 #)
    342 {-# INLINE mul_inner# #-}
    343 
    344 mul#
    345   :: Limb4
    346   -> Limb4
    347   -> Limb4
    348 mul# a b =
    349   let -- field prime
    350       !m = L4 0xFFFFFFFEFFFFFC2F## 0xFFFFFFFFFFFFFFFF##
    351               0xFFFFFFFFFFFFFFFF## 0xFFFFFFFFFFFFFFFF##
    352       !(# nu, mc #) = mul_inner# a b
    353   in  WW.sub_mod_c# nu mc m m
    354 {-# NOINLINE mul# #-} -- cannot be inlined without exploding comp time
    355 
    356 -- | Multiplication in the Montgomery domain.
    357 --
    358 --   Note that 'Montgomery' is an instance of 'Num', so you can use '*'
    359 --   to apply this function.
    360 --
    361 --   >>> 1 * 1 :: Montgomery
    362 --   1
    363 mul
    364   :: Montgomery -- ^ multiplicand in montgomery form
    365   -> Montgomery -- ^ multiplier in montgomery form
    366   -> Montgomery -- ^ montgomery product
    367 mul (Montgomery a) (Montgomery b) = Montgomery (mul# a b)
    368 
    369 to#
    370   :: Limb4 -- ^ integer
    371   -> Limb4
    372 to# x =
    373   let !r2 = L4 0x000007A2000E90A1## 0x1## 0## 0## -- r^2 mod m
    374   in  mul# x r2
    375 {-# INLINE to# #-}
    376 
    377 -- | Convert a 'Wider' word to the Montgomery domain.
    378 to :: Wider -> Montgomery
    379 to (Wider x) = Montgomery (to# x)
    380 
    381 -- | Retrieve a 'Montgomery' word from the Montgomery domain.
    382 --
    383 --   This function is a synonym for 'retr'.
    384 from :: Montgomery -> Wider
    385 from = retr
    386 
    387 add#
    388   :: Limb4 -- ^ augend
    389   -> Limb4 -- ^ addend
    390   -> Limb4 -- ^ sum
    391 add# a b =
    392   let -- field prime
    393       !m = L4 0xFFFFFFFEFFFFFC2F## 0xFFFFFFFFFFFFFFFF##
    394               0xFFFFFFFFFFFFFFFF## 0xFFFFFFFFFFFFFFFF##
    395   in  WW.add_mod# a b m
    396 {-# INLINE add# #-}
    397 
    398 -- | Addition in the Montgomery domain.
    399 --
    400 --   Note that 'Montgomery' is an instance of 'Num', so you can use '+'
    401 --   to apply this function.
    402 --
    403 --   >>> 1 + 1 :: Montgomery
    404 --   2
    405 add :: Montgomery -> Montgomery -> Montgomery
    406 add (Montgomery a) (Montgomery b) = Montgomery (add# a b)
    407 
    408 sub#
    409   :: Limb4 -- ^ minuend
    410   -> Limb4 -- ^ subtrahend
    411   -> Limb4 -- ^ difference
    412 sub# a b =
    413   let -- field prime
    414       !m = L4 0xFFFFFFFEFFFFFC2F## 0xFFFFFFFFFFFFFFFF##
    415               0xFFFFFFFFFFFFFFFF## 0xFFFFFFFFFFFFFFFF##
    416   in  WW.sub_mod# a b m
    417 {-# INLINE sub# #-}
    418 
    419 -- | Subtraction in the Montgomery domain.
    420 --
    421 --   Note that 'Montgomery' is an instance of 'Num', so you can use '-'
    422 --   to apply this function.
    423 --
    424 --   >>> 1 - 1 :: Montgomery
    425 --   0
    426 sub :: Montgomery -> Montgomery -> Montgomery
    427 sub (Montgomery a) (Montgomery b) = Montgomery (sub# a b)
    428 
    429 neg#
    430   :: Limb4 -- ^ argument
    431   -> Limb4 -- ^ modular negation
    432 neg# a = sub# (L4 0## 0## 0## 0##) a
    433 {-# INLINE neg# #-}
    434 
    435 -- | Additive inverse in the Montgomery domain.
    436 --
    437 --   Note that 'Montgomery' is an instance of 'Num', so you can use 'negate'
    438 --   to apply this function.
    439 --
    440 --   >>> negate 1 :: Montgomery
    441 --   115792089237316195423570985008687907853269984665640564039457584007908834671662
    442 --   >>> (negate 1 :: Montgomery) + 1
    443 --   0
    444 neg :: Montgomery -> Montgomery
    445 neg (Montgomery a) = Montgomery (neg# a)
    446 
    447 sqr# :: Limb4 -> Limb4
    448 sqr# a =
    449   let !(# l, h #) = WW.sqr# a
    450   in  redc# l h
    451 {-# NOINLINE sqr# #-} -- cannot be inlined without exploding comp time
    452 
    453 -- | Squaring in the Montgomery domain.
    454 --
    455 --   >>> sqr 1
    456 --   1
    457 --   >>> sqr 2
    458 --   4
    459 --   >>> sqr (negate 2)
    460 --   4
    461 sqr :: Montgomery -> Montgomery
    462 sqr (Montgomery a) = Montgomery (mul# a a)
    463 
    464 -- | Zero (the additive unit) in the Montgomery domain.
    465 zero :: Montgomery
    466 zero = Montgomery (L4 0## 0## 0## 0##)
    467 
    468 -- | One (the multiplicative unit) in the Montgomery domain.
    469 one :: Montgomery
    470 one = Montgomery (L4 0x1000003D1## 0## 0## 0##)
    471 
    472 -- generated by etc/generate_inv.sh
    473 inv#
    474   :: Limb4
    475   -> Limb4
    476 inv# a =
    477   let
    478       !t1 = sqr# a
    479       !t2 = mul# t1 a
    480       !t3 = sqr# t2
    481       !t4 = sqr# t3
    482       !t5 = mul# t4 t2
    483       !t6 = sqr# t5
    484       !t7 = sqr# t6
    485       !t8 = sqr# t7
    486       !t9 = sqr# t8
    487       !t10 = mul# t9 t5
    488       !t11 = sqr# t10
    489       !t12 = sqr# t11
    490       !t13 = sqr# t12
    491       !t14 = sqr# t13
    492       !t15 = sqr# t14
    493       !t16 = sqr# t15
    494       !t17 = sqr# t16
    495       !t18 = sqr# t17
    496       !t19 = mul# t18 t10
    497       !t20 = sqr# t19
    498       !t21 = sqr# t20
    499       !t22 = sqr# t21
    500       !t23 = sqr# t22
    501       !t24 = sqr# t23
    502       !t25 = sqr# t24
    503       !t26 = sqr# t25
    504       !t27 = sqr# t26
    505       !t28 = sqr# t27
    506       !t29 = sqr# t28
    507       !t30 = sqr# t29
    508       !t31 = sqr# t30
    509       !t32 = sqr# t31
    510       !t33 = sqr# t32
    511       !t34 = sqr# t33
    512       !t35 = sqr# t34
    513       !t36 = mul# t35 t19
    514       !t37 = sqr# t36
    515       !t38 = sqr# t37
    516       !t39 = sqr# t38
    517       !t40 = sqr# t39
    518       !t41 = sqr# t40
    519       !t42 = sqr# t41
    520       !t43 = sqr# t42
    521       !t44 = sqr# t43
    522       !t45 = sqr# t44
    523       !t46 = sqr# t45
    524       !t47 = sqr# t46
    525       !t48 = sqr# t47
    526       !t49 = sqr# t48
    527       !t50 = sqr# t49
    528       !t51 = sqr# t50
    529       !t52 = sqr# t51
    530       !t53 = sqr# t52
    531       !t54 = sqr# t53
    532       !t55 = sqr# t54
    533       !t56 = sqr# t55
    534       !t57 = sqr# t56
    535       !t58 = sqr# t57
    536       !t59 = sqr# t58
    537       !t60 = sqr# t59
    538       !t61 = sqr# t60
    539       !t62 = sqr# t61
    540       !t63 = sqr# t62
    541       !t64 = sqr# t63
    542       !t65 = sqr# t64
    543       !t66 = sqr# t65
    544       !t67 = sqr# t66
    545       !t68 = sqr# t67
    546       !t69 = mul# t68 t36
    547       !t70 = sqr# t69
    548       !t71 = sqr# t70
    549       !t72 = sqr# t71
    550       !t73 = sqr# t72
    551       !t74 = sqr# t73
    552       !t75 = sqr# t74
    553       !t76 = sqr# t75
    554       !t77 = sqr# t76
    555       !t78 = sqr# t77
    556       !t79 = sqr# t78
    557       !t80 = sqr# t79
    558       !t81 = sqr# t80
    559       !t82 = sqr# t81
    560       !t83 = sqr# t82
    561       !t84 = sqr# t83
    562       !t85 = sqr# t84
    563       !t86 = sqr# t85
    564       !t87 = sqr# t86
    565       !t88 = sqr# t87
    566       !t89 = sqr# t88
    567       !t90 = sqr# t89
    568       !t91 = sqr# t90
    569       !t92 = sqr# t91
    570       !t93 = sqr# t92
    571       !t94 = sqr# t93
    572       !t95 = sqr# t94
    573       !t96 = sqr# t95
    574       !t97 = sqr# t96
    575       !t98 = sqr# t97
    576       !t99 = sqr# t98
    577       !t100 = sqr# t99
    578       !t101 = sqr# t100
    579       !t102 = sqr# t101
    580       !t103 = sqr# t102
    581       !t104 = sqr# t103
    582       !t105 = sqr# t104
    583       !t106 = sqr# t105
    584       !t107 = sqr# t106
    585       !t108 = sqr# t107
    586       !t109 = sqr# t108
    587       !t110 = sqr# t109
    588       !t111 = sqr# t110
    589       !t112 = sqr# t111
    590       !t113 = sqr# t112
    591       !t114 = sqr# t113
    592       !t115 = sqr# t114
    593       !t116 = sqr# t115
    594       !t117 = sqr# t116
    595       !t118 = sqr# t117
    596       !t119 = sqr# t118
    597       !t120 = sqr# t119
    598       !t121 = sqr# t120
    599       !t122 = sqr# t121
    600       !t123 = sqr# t122
    601       !t124 = sqr# t123
    602       !t125 = sqr# t124
    603       !t126 = sqr# t125
    604       !t127 = sqr# t126
    605       !t128 = sqr# t127
    606       !t129 = sqr# t128
    607       !t130 = sqr# t129
    608       !t131 = sqr# t130
    609       !t132 = sqr# t131
    610       !t133 = sqr# t132
    611       !t134 = mul# t133 t69
    612       !t135 = sqr# t134
    613       !t136 = sqr# t135
    614       !t137 = sqr# t136
    615       !t138 = sqr# t137
    616       !t139 = sqr# t138
    617       !t140 = sqr# t139
    618       !t141 = sqr# t140
    619       !t142 = sqr# t141
    620       !t143 = sqr# t142
    621       !t144 = sqr# t143
    622       !t145 = sqr# t144
    623       !t146 = sqr# t145
    624       !t147 = sqr# t146
    625       !t148 = sqr# t147
    626       !t149 = sqr# t148
    627       !t150 = sqr# t149
    628       !t151 = sqr# t150
    629       !t152 = sqr# t151
    630       !t153 = sqr# t152
    631       !t154 = sqr# t153
    632       !t155 = sqr# t154
    633       !t156 = sqr# t155
    634       !t157 = sqr# t156
    635       !t158 = sqr# t157
    636       !t159 = sqr# t158
    637       !t160 = sqr# t159
    638       !t161 = sqr# t160
    639       !t162 = sqr# t161
    640       !t163 = sqr# t162
    641       !t164 = sqr# t163
    642       !t165 = sqr# t164
    643       !t166 = sqr# t165
    644       !t167 = sqr# t166
    645       !t168 = sqr# t167
    646       !t169 = sqr# t168
    647       !t170 = sqr# t169
    648       !t171 = sqr# t170
    649       !t172 = sqr# t171
    650       !t173 = sqr# t172
    651       !t174 = sqr# t173
    652       !t175 = sqr# t174
    653       !t176 = sqr# t175
    654       !t177 = sqr# t176
    655       !t178 = sqr# t177
    656       !t179 = sqr# t178
    657       !t180 = sqr# t179
    658       !t181 = sqr# t180
    659       !t182 = sqr# t181
    660       !t183 = sqr# t182
    661       !t184 = sqr# t183
    662       !t185 = sqr# t184
    663       !t186 = sqr# t185
    664       !t187 = sqr# t186
    665       !t188 = sqr# t187
    666       !t189 = sqr# t188
    667       !t190 = sqr# t189
    668       !t191 = sqr# t190
    669       !t192 = sqr# t191
    670       !t193 = sqr# t192
    671       !t194 = sqr# t193
    672       !t195 = sqr# t194
    673       !t196 = sqr# t195
    674       !t197 = sqr# t196
    675       !t198 = sqr# t197
    676       !t199 = mul# t198 t69
    677       !t200 = sqr# t199
    678       !t201 = sqr# t200
    679       !t202 = sqr# t201
    680       !t203 = sqr# t202
    681       !t204 = sqr# t203
    682       !t205 = sqr# t204
    683       !t206 = sqr# t205
    684       !t207 = sqr# t206
    685       !t208 = sqr# t207
    686       !t209 = sqr# t208
    687       !t210 = sqr# t209
    688       !t211 = sqr# t210
    689       !t212 = sqr# t211
    690       !t213 = sqr# t212
    691       !t214 = sqr# t213
    692       !t215 = sqr# t214
    693       !t216 = mul# t215 t19
    694       !t217 = sqr# t216
    695       !t218 = sqr# t217
    696       !t219 = sqr# t218
    697       !t220 = sqr# t219
    698       !t221 = sqr# t220
    699       !t222 = sqr# t221
    700       !t223 = sqr# t222
    701       !t224 = sqr# t223
    702       !t225 = mul# t224 t10
    703       !t226 = sqr# t225
    704       !t227 = sqr# t226
    705       !t228 = sqr# t227
    706       !t229 = sqr# t228
    707       !t230 = mul# t229 t5
    708       !t231 = sqr# t230
    709       !t232 = sqr# t231
    710       !t233 = mul# t232 t2
    711       !t234 = sqr# t233
    712       !t235 = mul# t234 a
    713       !t236 = sqr# t19
    714       !t237 = sqr# t236
    715       !t238 = sqr# t237
    716       !t239 = sqr# t238
    717       !t240 = mul# t239 t5
    718       !t241 = sqr# t240
    719       !t242 = sqr# t241
    720       !t243 = mul# t242 t2
    721       !t244 = sqr# t235
    722       !t245 = sqr# t244
    723       !t246 = sqr# t245
    724       !t247 = sqr# t246
    725       !t248 = sqr# t247
    726       !t249 = sqr# t248
    727       !t250 = sqr# t249
    728       !t251 = sqr# t250
    729       !t252 = sqr# t251
    730       !t253 = sqr# t252
    731       !t254 = sqr# t253
    732       !t255 = sqr# t254
    733       !t256 = sqr# t255
    734       !t257 = sqr# t256
    735       !t258 = sqr# t257
    736       !t259 = sqr# t258
    737       !t260 = sqr# t259
    738       !t261 = sqr# t260
    739       !t262 = sqr# t261
    740       !t263 = sqr# t262
    741       !t264 = sqr# t263
    742       !t265 = sqr# t264
    743       !t266 = sqr# t265
    744       !t267 = mul# t266 t243
    745       !t268 = sqr# t267
    746       !t269 = sqr# t268
    747       !t270 = sqr# t269
    748       !t271 = sqr# t270
    749       !t272 = sqr# t271
    750       !t273 = mul# t272 a
    751       !t274 = sqr# t273
    752       !t275 = sqr# t274
    753       !t276 = sqr# t275
    754       !t277 = mul# t276 t2
    755       !t278 = sqr# t277
    756       !t279 = sqr# t278
    757       !t280 = mul# t279 a
    758       !r = t280
    759   in  r
    760 {-# INLINE inv# #-}
    761 
    762 -- | Multiplicative inverse in the Montgomery domain.
    763 --
    764 --   >> inv 2
    765 --   57896044618658097711785492504343953926634992332820282019728792003954417335832
    766 --   >> inv 2 * 2
    767 --   1
    768 inv :: Montgomery -> Montgomery
    769 inv (Montgomery w) = Montgomery (inv# w)
    770 
    771 -- | Square root (Tonelli-Shanks) in the Montgomery domain.
    772 --
    773 --   Returns 'Nothing' if the square root doesn't exist.
    774 --
    775 --   Note that the square root calculation itself is performed in
    776 --   constant time; we branch only when casting to 'Maybe' at the end.
    777 --
    778 --   >>> sqrt_vartime 4
    779 --   Just 2
    780 --   >>> sqrt_vartime 15
    781 --   Just 69211104694897500952317515077652022726490027694212560352756646854116994689233
    782 --   >>> (*) <$> sqrt_vartime 15 <*> sqrt_vartime 15
    783 --   Just 15
    784 sqrt_vartime :: Montgomery -> Maybe Montgomery
    785 sqrt_vartime (Montgomery n) = case sqrt# n of
    786   (# a, c #)
    787     | C.decide c -> Just $! Montgomery a
    788     | otherwise  -> Nothing
    789 
    790 -- generated by etc/generate_sqrt.sh
    791 sqrt#
    792   :: Limb4
    793   -> (# Limb4, C.Choice #)
    794 sqrt# a =
    795   let !t0 = L4 0x1000003D1## 0## 0## 0##
    796       !t1 = sqr# t0
    797       !t2 = sqr# t1
    798       !t3 = sqr# t2
    799       !t4 = mul# a t3
    800       !t5 = sqr# t4
    801       !t6 = mul# a t5
    802       !t7 = sqr# t6
    803       !t8 = mul# a t7
    804       !t9 = sqr# t8
    805       !t10 = mul# a t9
    806       !t11 = sqr# t10
    807       !t12 = mul# a t11
    808       !t13 = sqr# t12
    809       !t14 = mul# a t13
    810       !t15 = sqr# t14
    811       !t16 = mul# a t15
    812       !t17 = sqr# t16
    813       !t18 = mul# a t17
    814       !t19 = sqr# t18
    815       !t20 = mul# a t19
    816       !t21 = sqr# t20
    817       !t22 = mul# a t21
    818       !t23 = sqr# t22
    819       !t24 = mul# a t23
    820       !t25 = sqr# t24
    821       !t26 = mul# a t25
    822       !t27 = sqr# t26
    823       !t28 = mul# a t27
    824       !t29 = sqr# t28
    825       !t30 = mul# a t29
    826       !t31 = sqr# t30
    827       !t32 = mul# a t31
    828       !t33 = sqr# t32
    829       !t34 = mul# a t33
    830       !t35 = sqr# t34
    831       !t36 = mul# a t35
    832       !t37 = sqr# t36
    833       !t38 = mul# a t37
    834       !t39 = sqr# t38
    835       !t40 = mul# a t39
    836       !t41 = sqr# t40
    837       !t42 = mul# a t41
    838       !t43 = sqr# t42
    839       !t44 = mul# a t43
    840       !t45 = sqr# t44
    841       !t46 = mul# a t45
    842       !t47 = sqr# t46
    843       !t48 = mul# a t47
    844       !t49 = sqr# t48
    845       !t50 = mul# a t49
    846       !t51 = sqr# t50
    847       !t52 = mul# a t51
    848       !t53 = sqr# t52
    849       !t54 = mul# a t53
    850       !t55 = sqr# t54
    851       !t56 = mul# a t55
    852       !t57 = sqr# t56
    853       !t58 = mul# a t57
    854       !t59 = sqr# t58
    855       !t60 = mul# a t59
    856       !t61 = sqr# t60
    857       !t62 = mul# a t61
    858       !t63 = sqr# t62
    859       !t64 = mul# a t63
    860       !t65 = sqr# t64
    861       !t66 = mul# a t65
    862       !t67 = sqr# t66
    863       !t68 = mul# a t67
    864       !t69 = sqr# t68
    865       !t70 = mul# a t69
    866       !t71 = sqr# t70
    867       !t72 = mul# a t71
    868       !t73 = sqr# t72
    869       !t74 = mul# a t73
    870       !t75 = sqr# t74
    871       !t76 = mul# a t75
    872       !t77 = sqr# t76
    873       !t78 = mul# a t77
    874       !t79 = sqr# t78
    875       !t80 = mul# a t79
    876       !t81 = sqr# t80
    877       !t82 = mul# a t81
    878       !t83 = sqr# t82
    879       !t84 = mul# a t83
    880       !t85 = sqr# t84
    881       !t86 = mul# a t85
    882       !t87 = sqr# t86
    883       !t88 = mul# a t87
    884       !t89 = sqr# t88
    885       !t90 = mul# a t89
    886       !t91 = sqr# t90
    887       !t92 = mul# a t91
    888       !t93 = sqr# t92
    889       !t94 = mul# a t93
    890       !t95 = sqr# t94
    891       !t96 = mul# a t95
    892       !t97 = sqr# t96
    893       !t98 = mul# a t97
    894       !t99 = sqr# t98
    895       !t100 = mul# a t99
    896       !t101 = sqr# t100
    897       !t102 = mul# a t101
    898       !t103 = sqr# t102
    899       !t104 = mul# a t103
    900       !t105 = sqr# t104
    901       !t106 = mul# a t105
    902       !t107 = sqr# t106
    903       !t108 = mul# a t107
    904       !t109 = sqr# t108
    905       !t110 = mul# a t109
    906       !t111 = sqr# t110
    907       !t112 = mul# a t111
    908       !t113 = sqr# t112
    909       !t114 = mul# a t113
    910       !t115 = sqr# t114
    911       !t116 = mul# a t115
    912       !t117 = sqr# t116
    913       !t118 = mul# a t117
    914       !t119 = sqr# t118
    915       !t120 = mul# a t119
    916       !t121 = sqr# t120
    917       !t122 = mul# a t121
    918       !t123 = sqr# t122
    919       !t124 = mul# a t123
    920       !t125 = sqr# t124
    921       !t126 = mul# a t125
    922       !t127 = sqr# t126
    923       !t128 = mul# a t127
    924       !t129 = sqr# t128
    925       !t130 = mul# a t129
    926       !t131 = sqr# t130
    927       !t132 = mul# a t131
    928       !t133 = sqr# t132
    929       !t134 = mul# a t133
    930       !t135 = sqr# t134
    931       !t136 = mul# a t135
    932       !t137 = sqr# t136
    933       !t138 = mul# a t137
    934       !t139 = sqr# t138
    935       !t140 = mul# a t139
    936       !t141 = sqr# t140
    937       !t142 = mul# a t141
    938       !t143 = sqr# t142
    939       !t144 = mul# a t143
    940       !t145 = sqr# t144
    941       !t146 = mul# a t145
    942       !t147 = sqr# t146
    943       !t148 = mul# a t147
    944       !t149 = sqr# t148
    945       !t150 = mul# a t149
    946       !t151 = sqr# t150
    947       !t152 = mul# a t151
    948       !t153 = sqr# t152
    949       !t154 = mul# a t153
    950       !t155 = sqr# t154
    951       !t156 = mul# a t155
    952       !t157 = sqr# t156
    953       !t158 = mul# a t157
    954       !t159 = sqr# t158
    955       !t160 = mul# a t159
    956       !t161 = sqr# t160
    957       !t162 = mul# a t161
    958       !t163 = sqr# t162
    959       !t164 = mul# a t163
    960       !t165 = sqr# t164
    961       !t166 = mul# a t165
    962       !t167 = sqr# t166
    963       !t168 = mul# a t167
    964       !t169 = sqr# t168
    965       !t170 = mul# a t169
    966       !t171 = sqr# t170
    967       !t172 = mul# a t171
    968       !t173 = sqr# t172
    969       !t174 = mul# a t173
    970       !t175 = sqr# t174
    971       !t176 = mul# a t175
    972       !t177 = sqr# t176
    973       !t178 = mul# a t177
    974       !t179 = sqr# t178
    975       !t180 = mul# a t179
    976       !t181 = sqr# t180
    977       !t182 = mul# a t181
    978       !t183 = sqr# t182
    979       !t184 = mul# a t183
    980       !t185 = sqr# t184
    981       !t186 = mul# a t185
    982       !t187 = sqr# t186
    983       !t188 = mul# a t187
    984       !t189 = sqr# t188
    985       !t190 = mul# a t189
    986       !t191 = sqr# t190
    987       !t192 = mul# a t191
    988       !t193 = sqr# t192
    989       !t194 = mul# a t193
    990       !t195 = sqr# t194
    991       !t196 = mul# a t195
    992       !t197 = sqr# t196
    993       !t198 = mul# a t197
    994       !t199 = sqr# t198
    995       !t200 = mul# a t199
    996       !t201 = sqr# t200
    997       !t202 = mul# a t201
    998       !t203 = sqr# t202
    999       !t204 = mul# a t203
   1000       !t205 = sqr# t204
   1001       !t206 = mul# a t205
   1002       !t207 = sqr# t206
   1003       !t208 = mul# a t207
   1004       !t209 = sqr# t208
   1005       !t210 = mul# a t209
   1006       !t211 = sqr# t210
   1007       !t212 = mul# a t211
   1008       !t213 = sqr# t212
   1009       !t214 = mul# a t213
   1010       !t215 = sqr# t214
   1011       !t216 = mul# a t215
   1012       !t217 = sqr# t216
   1013       !t218 = mul# a t217
   1014       !t219 = sqr# t218
   1015       !t220 = mul# a t219
   1016       !t221 = sqr# t220
   1017       !t222 = mul# a t221
   1018       !t223 = sqr# t222
   1019       !t224 = mul# a t223
   1020       !t225 = sqr# t224
   1021       !t226 = mul# a t225
   1022       !t227 = sqr# t226
   1023       !t228 = mul# a t227
   1024       !t229 = sqr# t228
   1025       !t230 = mul# a t229
   1026       !t231 = sqr# t230
   1027       !t232 = mul# a t231
   1028       !t233 = sqr# t232
   1029       !t234 = mul# a t233
   1030       !t235 = sqr# t234
   1031       !t236 = mul# a t235
   1032       !t237 = sqr# t236
   1033       !t238 = mul# a t237
   1034       !t239 = sqr# t238
   1035       !t240 = mul# a t239
   1036       !t241 = sqr# t240
   1037       !t242 = mul# a t241
   1038       !t243 = sqr# t242
   1039       !t244 = mul# a t243
   1040       !t245 = sqr# t244
   1041       !t246 = mul# a t245
   1042       !t247 = sqr# t246
   1043       !t248 = mul# a t247
   1044       !t249 = sqr# t248
   1045       !t250 = mul# a t249
   1046       !t251 = sqr# t250
   1047       !t252 = mul# a t251
   1048       !t253 = sqr# t252
   1049       !t254 = mul# a t253
   1050       !t255 = sqr# t254
   1051       !t256 = mul# a t255
   1052       !t257 = sqr# t256
   1053       !t258 = mul# a t257
   1054       !t259 = sqr# t258
   1055       !t260 = mul# a t259
   1056       !t261 = sqr# t260
   1057       !t262 = mul# a t261
   1058       !t263 = sqr# t262
   1059       !t264 = mul# a t263
   1060       !t265 = sqr# t264
   1061       !t266 = mul# a t265
   1062       !t267 = sqr# t266
   1063       !t268 = mul# a t267
   1064       !t269 = sqr# t268
   1065       !t270 = mul# a t269
   1066       !t271 = sqr# t270
   1067       !t272 = mul# a t271
   1068       !t273 = sqr# t272
   1069       !t274 = mul# a t273
   1070       !t275 = sqr# t274
   1071       !t276 = mul# a t275
   1072       !t277 = sqr# t276
   1073       !t278 = mul# a t277
   1074       !t279 = sqr# t278
   1075       !t280 = mul# a t279
   1076       !t281 = sqr# t280
   1077       !t282 = mul# a t281
   1078       !t283 = sqr# t282
   1079       !t284 = mul# a t283
   1080       !t285 = sqr# t284
   1081       !t286 = mul# a t285
   1082       !t287 = sqr# t286
   1083       !t288 = mul# a t287
   1084       !t289 = sqr# t288
   1085       !t290 = mul# a t289
   1086       !t291 = sqr# t290
   1087       !t292 = mul# a t291
   1088       !t293 = sqr# t292
   1089       !t294 = mul# a t293
   1090       !t295 = sqr# t294
   1091       !t296 = mul# a t295
   1092       !t297 = sqr# t296
   1093       !t298 = mul# a t297
   1094       !t299 = sqr# t298
   1095       !t300 = mul# a t299
   1096       !t301 = sqr# t300
   1097       !t302 = mul# a t301
   1098       !t303 = sqr# t302
   1099       !t304 = mul# a t303
   1100       !t305 = sqr# t304
   1101       !t306 = mul# a t305
   1102       !t307 = sqr# t306
   1103       !t308 = mul# a t307
   1104       !t309 = sqr# t308
   1105       !t310 = mul# a t309
   1106       !t311 = sqr# t310
   1107       !t312 = mul# a t311
   1108       !t313 = sqr# t312
   1109       !t314 = mul# a t313
   1110       !t315 = sqr# t314
   1111       !t316 = mul# a t315
   1112       !t317 = sqr# t316
   1113       !t318 = mul# a t317
   1114       !t319 = sqr# t318
   1115       !t320 = mul# a t319
   1116       !t321 = sqr# t320
   1117       !t322 = mul# a t321
   1118       !t323 = sqr# t322
   1119       !t324 = mul# a t323
   1120       !t325 = sqr# t324
   1121       !t326 = mul# a t325
   1122       !t327 = sqr# t326
   1123       !t328 = mul# a t327
   1124       !t329 = sqr# t328
   1125       !t330 = mul# a t329
   1126       !t331 = sqr# t330
   1127       !t332 = mul# a t331
   1128       !t333 = sqr# t332
   1129       !t334 = mul# a t333
   1130       !t335 = sqr# t334
   1131       !t336 = mul# a t335
   1132       !t337 = sqr# t336
   1133       !t338 = mul# a t337
   1134       !t339 = sqr# t338
   1135       !t340 = mul# a t339
   1136       !t341 = sqr# t340
   1137       !t342 = mul# a t341
   1138       !t343 = sqr# t342
   1139       !t344 = mul# a t343
   1140       !t345 = sqr# t344
   1141       !t346 = mul# a t345
   1142       !t347 = sqr# t346
   1143       !t348 = mul# a t347
   1144       !t349 = sqr# t348
   1145       !t350 = mul# a t349
   1146       !t351 = sqr# t350
   1147       !t352 = mul# a t351
   1148       !t353 = sqr# t352
   1149       !t354 = mul# a t353
   1150       !t355 = sqr# t354
   1151       !t356 = mul# a t355
   1152       !t357 = sqr# t356
   1153       !t358 = mul# a t357
   1154       !t359 = sqr# t358
   1155       !t360 = mul# a t359
   1156       !t361 = sqr# t360
   1157       !t362 = mul# a t361
   1158       !t363 = sqr# t362
   1159       !t364 = mul# a t363
   1160       !t365 = sqr# t364
   1161       !t366 = mul# a t365
   1162       !t367 = sqr# t366
   1163       !t368 = mul# a t367
   1164       !t369 = sqr# t368
   1165       !t370 = mul# a t369
   1166       !t371 = sqr# t370
   1167       !t372 = mul# a t371
   1168       !t373 = sqr# t372
   1169       !t374 = mul# a t373
   1170       !t375 = sqr# t374
   1171       !t376 = mul# a t375
   1172       !t377 = sqr# t376
   1173       !t378 = mul# a t377
   1174       !t379 = sqr# t378
   1175       !t380 = mul# a t379
   1176       !t381 = sqr# t380
   1177       !t382 = mul# a t381
   1178       !t383 = sqr# t382
   1179       !t384 = mul# a t383
   1180       !t385 = sqr# t384
   1181       !t386 = mul# a t385
   1182       !t387 = sqr# t386
   1183       !t388 = mul# a t387
   1184       !t389 = sqr# t388
   1185       !t390 = mul# a t389
   1186       !t391 = sqr# t390
   1187       !t392 = mul# a t391
   1188       !t393 = sqr# t392
   1189       !t394 = mul# a t393
   1190       !t395 = sqr# t394
   1191       !t396 = mul# a t395
   1192       !t397 = sqr# t396
   1193       !t398 = mul# a t397
   1194       !t399 = sqr# t398
   1195       !t400 = mul# a t399
   1196       !t401 = sqr# t400
   1197       !t402 = mul# a t401
   1198       !t403 = sqr# t402
   1199       !t404 = mul# a t403
   1200       !t405 = sqr# t404
   1201       !t406 = mul# a t405
   1202       !t407 = sqr# t406
   1203       !t408 = mul# a t407
   1204       !t409 = sqr# t408
   1205       !t410 = mul# a t409
   1206       !t411 = sqr# t410
   1207       !t412 = mul# a t411
   1208       !t413 = sqr# t412
   1209       !t414 = mul# a t413
   1210       !t415 = sqr# t414
   1211       !t416 = mul# a t415
   1212       !t417 = sqr# t416
   1213       !t418 = mul# a t417
   1214       !t419 = sqr# t418
   1215       !t420 = mul# a t419
   1216       !t421 = sqr# t420
   1217       !t422 = mul# a t421
   1218       !t423 = sqr# t422
   1219       !t424 = mul# a t423
   1220       !t425 = sqr# t424
   1221       !t426 = mul# a t425
   1222       !t427 = sqr# t426
   1223       !t428 = mul# a t427
   1224       !t429 = sqr# t428
   1225       !t430 = mul# a t429
   1226       !t431 = sqr# t430
   1227       !t432 = mul# a t431
   1228       !t433 = sqr# t432
   1229       !t434 = mul# a t433
   1230       !t435 = sqr# t434
   1231       !t436 = mul# a t435
   1232       !t437 = sqr# t436
   1233       !t438 = mul# a t437
   1234       !t439 = sqr# t438
   1235       !t440 = mul# a t439
   1236       !t441 = sqr# t440
   1237       !t442 = mul# a t441
   1238       !t443 = sqr# t442
   1239       !t444 = mul# a t443
   1240       !t445 = sqr# t444
   1241       !t446 = mul# a t445
   1242       !t447 = sqr# t446
   1243       !t448 = mul# a t447
   1244       !t449 = sqr# t448
   1245       !t450 = sqr# t449
   1246       !t451 = mul# a t450
   1247       !t452 = sqr# t451
   1248       !t453 = mul# a t452
   1249       !t454 = sqr# t453
   1250       !t455 = mul# a t454
   1251       !t456 = sqr# t455
   1252       !t457 = mul# a t456
   1253       !t458 = sqr# t457
   1254       !t459 = mul# a t458
   1255       !t460 = sqr# t459
   1256       !t461 = mul# a t460
   1257       !t462 = sqr# t461
   1258       !t463 = mul# a t462
   1259       !t464 = sqr# t463
   1260       !t465 = mul# a t464
   1261       !t466 = sqr# t465
   1262       !t467 = mul# a t466
   1263       !t468 = sqr# t467
   1264       !t469 = mul# a t468
   1265       !t470 = sqr# t469
   1266       !t471 = mul# a t470
   1267       !t472 = sqr# t471
   1268       !t473 = mul# a t472
   1269       !t474 = sqr# t473
   1270       !t475 = mul# a t474
   1271       !t476 = sqr# t475
   1272       !t477 = mul# a t476
   1273       !t478 = sqr# t477
   1274       !t479 = mul# a t478
   1275       !t480 = sqr# t479
   1276       !t481 = mul# a t480
   1277       !t482 = sqr# t481
   1278       !t483 = mul# a t482
   1279       !t484 = sqr# t483
   1280       !t485 = mul# a t484
   1281       !t486 = sqr# t485
   1282       !t487 = mul# a t486
   1283       !t488 = sqr# t487
   1284       !t489 = mul# a t488
   1285       !t490 = sqr# t489
   1286       !t491 = mul# a t490
   1287       !t492 = sqr# t491
   1288       !t493 = mul# a t492
   1289       !t494 = sqr# t493
   1290       !t495 = sqr# t494
   1291       !t496 = sqr# t495
   1292       !t497 = sqr# t496
   1293       !t498 = sqr# t497
   1294       !t499 = mul# a t498
   1295       !t500 = sqr# t499
   1296       !t501 = mul# a t500
   1297       !t502 = sqr# t501
   1298       !t503 = sqr# t502
   1299       !r = t503
   1300   in  (# r, WW.eq# (sqr# r) a #)
   1301 {-# INLINE sqrt# #-}
   1302 
   1303 -- | Exponentiation in the Montgomery domain.
   1304 --
   1305 --   >>> exp 2 3
   1306 --   8
   1307 --   >>> exp 2 10
   1308 --   1024
   1309 exp :: Montgomery -> Wider -> Montgomery
   1310 exp (Montgomery b) (Wider e) = Montgomery (exp# b e)
   1311 
   1312 exp#
   1313   :: Limb4
   1314   -> Limb4
   1315   -> Limb4
   1316 exp# b e =
   1317   let !o = L4 0x1000003D1## 0## 0## 0##
   1318       loop !r !m !ex n = case n of
   1319         0 -> r
   1320         _ ->
   1321           let !(# ne, bit #) = WW.shr1_c# ex
   1322               !candidate = mul# r m
   1323               !nr = select# r candidate bit
   1324               !nm = sqr# m
   1325           in  loop nr nm ne (n - 1)
   1326   in  loop o b e (256 :: Word)
   1327 {-# INLINE exp# #-}
   1328 
   1329 odd# :: Limb4 -> C.Choice
   1330 odd# = WW.odd#
   1331 {-# INLINE odd# #-}
   1332 
   1333 -- | Check if a 'Montgomery' value is odd.
   1334 --
   1335 --   Note that the comparison is performed in constant time, but we
   1336 --   branch when converting to 'Bool'.
   1337 --
   1338 --   >>> odd 1
   1339 --   True
   1340 --   >>> odd 2
   1341 --   False
   1342 --   >>> Data.Word.Wider.odd (retr 3) -- parity is preserved
   1343 --   True
   1344 odd_vartime :: Montgomery -> Bool
   1345 odd_vartime (Montgomery m) = C.decide (odd# m)
   1346 
   1347 -- constant-time selection ----------------------------------------------------
   1348 
   1349 select#
   1350   :: Limb4    -- ^ a
   1351   -> Limb4    -- ^ b
   1352   -> C.Choice -- ^ c
   1353   -> Limb4    -- ^ result
   1354 select# = WW.select#
   1355 {-# INLINE select# #-}
   1356 
   1357 -- | Return a if c is truthy, otherwise return b.
   1358 --
   1359 --   >>> import qualified Data.Choice as C
   1360 --   >>> select 0 1 (C.true# ())
   1361 --   1
   1362 select
   1363   :: Montgomery    -- ^ a
   1364   -> Montgomery    -- ^ b
   1365   -> C.Choice      -- ^ c
   1366   -> Montgomery    -- ^ result
   1367 select (Montgomery a) (Montgomery b) c = Montgomery (select# a b c)
   1368