// Copyright 2026 The Wuffs Authors. // // Licensed under the Apache License, Version 2.0 or the MIT license // , at your // option. This file may not be copied, modified, or distributed // except according to those terms. // // SPDX-License-Identifier: Apache-2.0 OR MIT // Package softfloat implements IEEE 754 binary64 ("f64", C's double) and // binary32 ("f32", C's float) arithmetic using only integer operations. // // Values are passed around as their bit patterns: an f64 is a base.u64 and an // f32 is a base.u32. Every operation rounds to nearest, ties to even, and // handles signed zeroes, subnormals, infinities and NaNs. The results are // bit-for-bit identical to x86-64 SSE2 hardware (and any other conforming // IEEE 754 implementation), except that NaN payloads (and NaN signs) are not // guaranteed to match: callers should only rely on a result being "some NaN". // // See README.md for more details. pub struct arith?( util : base.utility, ) pri const F64_SIGN : base.u64 = 0x8000_0000_0000_0000 pri const F64_ABS : base.u64 = 0x7FFF_FFFF_FFFF_FFFF pri const F64_INF : base.u64 = 0x7FF0_0000_0000_0000 pri const F64_FRAC : base.u64 = 0x000F_FFFF_FFFF_FFFF pri const F64_HIDDEN : base.u64 = 0x0010_0000_0000_0000 pri const F64_QUIET : base.u64 = 0x0008_0000_0000_0000 // F64_DEFAULT_NAN is what x86 hardware produces for invalid operations such // as (0 / 0) or (inf - inf). pri const F64_DEFAULT_NAN : base.u64 = 0xFFF8_0000_0000_0000 // -------- Bit twiddling helpers. // clz64 returns the number of leading zero bits of a, which must be non-zero. // If a is zero, it returns 63. pri func arith.clz64(a: base.u64) base.u32[..= 63] { var x : base.u64 var n32 : base.u32[..= 32] var n16 : base.u32[..= 16] var n8 : base.u32[..= 8] x = args.a if (x >> 32) == 0 { x = x ~mod<< 32 n32 = 32 } if (x >> 48) == 0 { x = x ~mod<< 16 n16 = 16 } if (x >> 56) == 0 { x = x ~mod<< 8 n8 = 8 } return (n32 + n16 + n8 + ((CLZ8[x >> 56] as base.u32) & 7)) & 63 } // clz32 returns the number of leading zero bits of a, which must be non-zero. pri func arith.clz32(a: base.u32) base.u32[..= 31] { var x : base.u32 var n16 : base.u32[..= 16] var n8 : base.u32[..= 8] x = args.a if (x >> 16) == 0 { x = x ~mod<< 16 n16 = 16 } if (x >> 24) == 0 { x = x ~mod<< 8 n8 = 8 } return (n16 + n8 + ((CLZ8[x >> 24] as base.u32) & 7)) & 31 } // shift_right_jam64 returns (a >> dist), except that if any non-zero bits // were shifted out then the least significant bit of the result is set. This // "sticky bit" is needed to round correctly. pri func arith.shift_right_jam64(a: base.u64, dist: base.u32) base.u64 { var d : base.u32 if args.dist == 0 { return args.a } else if args.dist < 64 { d = args.dist if (args.a ~mod<< (64 - d)) <> 0 { return (args.a >> d) | 1 } return args.a >> d } else if args.a <> 0 { return 1 } return 0 } // jam128_hi and jam128_lo return the high and low 64-bit halves of the // 128-bit value ((hi << 64) | lo) shifted right by dist, with a sticky bit // (see shift_right_jam64). pri func arith.jam128_hi(hi: base.u64, lo: base.u64, dist: base.u32) base.u64 { if args.dist < 64 { return args.hi >> args.dist } return 0 } pri func arith.jam128_lo(hi: base.u64, lo: base.u64, dist: base.u32) base.u64 { var d : base.u32 var sticky : base.u64 if args.dist == 0 { return args.lo } else if args.dist < 64 { d = args.dist if (args.lo ~mod<< (64 - d)) <> 0 { sticky = 1 } return (args.hi ~mod<< (64 - d)) | (args.lo >> d) | sticky } else if args.dist < 128 { d = args.dist - 64 if d > 0 { if ((args.hi ~mod<< (64 - d)) | args.lo) <> 0 { sticky = 1 } } else if args.lo <> 0 { sticky = 1 } return (args.hi >> d) | sticky } else if (args.hi | args.lo) <> 0 { return 1 } return 0 } // mul_hi64 returns the high 64 bits of the 128-bit product (a * b). pri func arith.mul_hi64(a: base.u64, b: base.u64) base.u64 { var a0 : base.u64[..= 0xFFFF_FFFF] var a1 : base.u64[..= 0xFFFF_FFFF] var b0 : base.u64[..= 0xFFFF_FFFF] var b1 : base.u64[..= 0xFFFF_FFFF] var p00 : base.u64 var p01 : base.u64 var p10 : base.u64 var p11 : base.u64 var mid : base.u64 a0 = args.a & 0xFFFF_FFFF a1 = args.a >> 32 b0 = args.b & 0xFFFF_FFFF b1 = args.b >> 32 p00 = a0 * b0 p01 = a0 * b1 p10 = a1 * b0 p11 = a1 * b1 mid = (p00 >> 32) + (p01 & 0xFFFF_FFFF) + (p10 & 0xFFFF_FFFF) return ((p11 ~mod+ (p01 >> 32)) ~mod+ (p10 >> 32)) ~mod+ (mid >> 32) } // -------- f64 rounding and packing. // // Exponents are passed as base.u32 values holding two's complement signed // integers (so that 0xFFFF_FFFF means -1), manipulated with ~mod arithmetic. // // The round_pack_f64 sig argument holds a significand with its implicit // (hidden) bit at bit 62 and 10 extra bits of precision at the bottom (the // least significant of which is typically sticky). The value represented is // (sig * 2**(exp - 1084)), so exp is one less than the biased exponent. pri func arith.round_pack_f64(sign: base.u64, exp: base.u32, sig: base.u64) base.u64 { var e : base.u32 var s : base.u64 var round_bits : base.u64 e = args.exp s = args.sig round_bits = s & 0x3FF if e >= 0x8000_0000 { // Negative exponent: the result is subnormal (or zero). s = this.shift_right_jam64(a: s, dist: 0 ~mod- e) e = 0 round_bits = s & 0x3FF } else if e >= 0x7FD { if (e > 0x7FD) or (s >= 0x7FFF_FFFF_FFFF_FE00) { return args.sign | F64_INF } } s = (s ~mod+ 0x200) >> 10 if round_bits == 0x200 { // Ties to even. s = s & 0xFFFF_FFFF_FFFF_FFFE } if s == 0 { e = 0 } return args.sign | (((e as base.u64) ~mod<< 52) ~mod+ s) } // norm_round_pack_f64 is like round_pack_f64 but sig does not need to be // normalized (its most significant set bit can be anywhere). pri func arith.norm_round_pack_f64(sign: base.u64, exp: base.u32, sig: base.u64) base.u64 { var c : base.u32[..= 63] var shift : base.u32[..= 62] var e : base.u32 if args.sig == 0 { return args.sign } c = this.clz64(a: args.sig) if c == 0 { return this.round_pack_f64(sign: args.sign, exp: args.exp ~mod+ 1, sig: (args.sig >> 1) | (args.sig & 1)) } shift = c - 1 e = args.exp ~mod- shift if (shift >= 10) and (e < 0x7FD) { return args.sign | (((e as base.u64) ~mod<< 52) ~mod+ (args.sig ~mod<< (shift - 10))) } return this.round_pack_f64(sign: args.sign, exp: e, sig: args.sig ~mod<< shift) } // norm_sub_exp and norm_sub_sig normalize a subnormal's (non-zero) fraction // bits, returning the (two's complement) exponent and the significand such // that the significand's hidden bit is at bit 52. pri func arith.norm_sub_exp(frac: base.u64) base.u32 { return 12 ~mod- this.clz64(a: args.frac) } pri func arith.norm_sub_sig(frac: base.u64) base.u64 { return args.frac ~mod<< ((this.clz64(a: args.frac) ~mod- 11) & 63) } pri func arith.f64_propagate_nan(a: base.u64, b: base.u64) base.u64 { if (args.a & F64_ABS) > F64_INF { return args.a | F64_QUIET } return args.b | F64_QUIET } // -------- f64 classification, sign and comparison. // f64_is_nan returns 1 if a is a NaN (quiet or signaling) and 0 otherwise. pub func arith.f64_is_nan(a: base.u64) base.u32 { if (args.a & F64_ABS) > F64_INF { return 1 } return 0 } // f64_neg returns -a (flipping the sign bit, even for NaNs and zeroes). pub func arith.f64_neg(a: base.u64) base.u64 { return args.a ^ F64_SIGN } // f64_abs returns |a| (clearing the sign bit, even for NaNs). pub func arith.f64_abs(a: base.u64) base.u64 { return args.a & F64_ABS } // f64_cmp compares a and b, returning 0 if (a < b), 1 if (a == b), 2 if (a // > b) and 3 if they are unordered (either is a NaN). -0 and +0 compare equal. pub func arith.f64_cmp(a: base.u64, b: base.u64) base.u32 { if (args.a | args.b) <= F64_INF { // Fast path: neither is negative (the sign bits are clear) or a NaN // (both are at most +inf), so the bit patterns sort like the values. if args.a < args.b { return 0 } else if args.a == args.b { return 1 } return 2 } return this.f64_cmp_slow(a: args.a, b: args.b) } // f64_cmp_slow is f64_cmp without the fast path. pri func arith.f64_cmp_slow(a: base.u64, b: base.u64) base.u32 { var ka : base.u64 var kb : base.u64 if ((args.a & F64_ABS) > F64_INF) or ((args.b & F64_ABS) > F64_INF) { return 3 } else if ((args.a | args.b) & F64_ABS) == 0 { return 1 } // Map the bit patterns to unsigned integers that sort in the same order // as the f64 values they represent. ka = args.a | F64_SIGN if (args.a >> 63) <> 0 { ka = 0xFFFF_FFFF_FFFF_FFFF ^ args.a } kb = args.b | F64_SIGN if (args.b >> 63) <> 0 { kb = 0xFFFF_FFFF_FFFF_FFFF ^ args.b } if ka < kb { return 0 } else if ka == kb { return 1 } return 2 } // The f64_eq, f64_ne, f64_lt, f64_le, f64_gt and f64_ge functions implement // C's ==, !=, <, <=, > and >= operators. They return 1 for true and 0 for // false. Any comparison involving a NaN is false (except for !=, which is // true), and -0 equals +0. pub func arith.f64_eq(a: base.u64, b: base.u64) base.u32 { if (args.a | args.b) <= F64_INF { if args.a == args.b { return 1 } return 0 } if this.f64_cmp_slow(a: args.a, b: args.b) == 1 { return 1 } return 0 } pub func arith.f64_ne(a: base.u64, b: base.u64) base.u32 { if (args.a | args.b) <= F64_INF { if args.a <> args.b { return 1 } return 0 } if this.f64_cmp_slow(a: args.a, b: args.b) == 1 { return 0 } return 1 } pub func arith.f64_lt(a: base.u64, b: base.u64) base.u32 { if (args.a | args.b) <= F64_INF { if args.a < args.b { return 1 } return 0 } if this.f64_cmp_slow(a: args.a, b: args.b) == 0 { return 1 } return 0 } pub func arith.f64_le(a: base.u64, b: base.u64) base.u32 { if (args.a | args.b) <= F64_INF { if args.a <= args.b { return 1 } return 0 } if this.f64_cmp_slow(a: args.a, b: args.b) <= 1 { return 1 } return 0 } pub func arith.f64_gt(a: base.u64, b: base.u64) base.u32 { if (args.a | args.b) <= F64_INF { if args.a > args.b { return 1 } return 0 } if this.f64_cmp_slow(a: args.a, b: args.b) == 2 { return 1 } return 0 } pub func arith.f64_ge(a: base.u64, b: base.u64) base.u32 { var c : base.u32 if (args.a | args.b) <= F64_INF { if args.a >= args.b { return 1 } return 0 } c = this.f64_cmp_slow(a: args.a, b: args.b) if (c == 1) or (c == 2) { return 1 } return 0 } // f64_min returns ((a < b) ? a : b), like brotli's BROTLI_MIN(double, a, b). // Note that this is not IEEE 754 minNum: f64_min(NaN, x) and f64_min(x, NaN) // both return the second argument, and f64_min(-0, +0) returns +0. pub func arith.f64_min(a: base.u64, b: base.u64) base.u64 { if (args.a | args.b) <= F64_INF { if args.a < args.b { return args.a } return args.b } if this.f64_cmp_slow(a: args.a, b: args.b) == 0 { return args.a } return args.b } // f64_max returns ((a > b) ? a : b), like brotli's BROTLI_MAX(double, a, b). pub func arith.f64_max(a: base.u64, b: base.u64) base.u64 { if (args.a | args.b) <= F64_INF { if args.a > args.b { return args.a } return args.b } if this.f64_cmp_slow(a: args.a, b: args.b) == 2 { return args.a } return args.b } // -------- f64 arithmetic. // f64_add returns (a + b). pub func arith.f64_add(a: base.u64, b: base.u64) base.u64 { var z : base.u64 if ((((args.a >> 52) & 0x7FF) ~mod- 1) < 0x7FE) and ((((args.b >> 52) & 0x7FF) ~mod- 1) < 0x7FE) { // Both operands are normal: take the fast path. z = this.f64_add_normal(a: args.a, b: args.b) if z <> 0xFFFF_FFFF_FFFF_FFFF { return z } } else if ((args.b & F64_ABS) == 0) and (((args.a & F64_ABS) ~mod- 1) < (F64_INF - 1)) { // b is a zero and a is finite and non-zero, so (a + b) is a. return args.a } else if ((args.a & F64_ABS) == 0) and (((args.b & F64_ABS) ~mod- 1) < (F64_INF - 1)) { // a is a zero and b is finite and non-zero, so (a + b) is b. return args.b } if ((args.a ^ args.b) >> 63) == 0 { return this.f64_add_mags(a: args.a, b: args.b, sign: args.a & F64_SIGN) } return this.f64_sub_mags(a: args.a, b: args.b, sign: args.a & F64_SIGN) } // f64_sub returns (a - b). pub func arith.f64_sub(a: base.u64, b: base.u64) base.u64 { var z : base.u64 if ((((args.a >> 52) & 0x7FF) ~mod- 1) < 0x7FE) and ((((args.b >> 52) & 0x7FF) ~mod- 1) < 0x7FE) { // Both operands are normal: take the fast path. For non-NaN // operands, (a - b) is exactly (a + -b), including the sign of a // zero result. z = this.f64_add_normal(a: args.a, b: args.b ^ F64_SIGN) if z <> 0xFFFF_FFFF_FFFF_FFFF { return z } } else if ((args.b & F64_ABS) == 0) and (((args.a & F64_ABS) ~mod- 1) < (F64_INF - 1)) { // b is a zero and a is finite and non-zero, so (a - b) is a. return args.a } else if ((args.a & F64_ABS) == 0) and (((args.b & F64_ABS) ~mod- 1) < (F64_INF - 1)) { // a is a zero and b is finite and non-zero, so (a - b) is -b. return args.b ^ F64_SIGN } if ((args.a ^ args.b) >> 63) == 0 { return this.f64_sub_mags(a: args.a, b: args.b, sign: args.a & F64_SIGN) } return this.f64_add_mags(a: args.a, b: args.b, sign: args.a & F64_SIGN) } // f64_add_normal returns (a + b), for normal a and b, unless the result is // subnormal, in which case it returns 0xFFFF_FFFF_FFFF_FFFF (a NaN, which it // never otherwise returns) and the caller must take its slow path. It is a // faster (but less general) version of f64_add_mags and f64_sub_mags. pri func arith.f64_add_normal(a: base.u64, b: base.u64) base.u64 { var x : base.u64 var y : base.u64 var ex : base.u32[..= 0x800] var d : base.u32 var mx : base.u64[..= 0x7FFF_FFFF_FFFF_FFFF] var my : base.u64[..= 0x7FFF_FFFF_FFFF_FFFF] var z : base.u64 var s : base.u32[..= 63] var rb : base.u64 // Order the operands so that |x| >= |y|. The result has x's sign. x = args.a y = args.b if (y & F64_ABS) > (x & F64_ABS) { x = args.b y = args.a } ex = ((x >> 52) & 0x7FF) as base.u32 d = ex ~mod- (((y >> 52) & 0x7FF) as base.u32) if d > 63 { d = 63 } // The significands, with their implicit bit at bit 62 and 10 extra bits // of precision at the bottom. Align my with mx, keeping a sticky bit. mx = ((x & F64_FRAC) | F64_HIDDEN) << 10 my = ((y & F64_FRAC) | F64_HIDDEN) << 10 if (my & (((1 as base.u64) << (d & 63)) ~mod- 1)) <> 0 { my = (my >> (d & 63)) | 1 } else { my = my >> (d & 63) } if ((x ^ y) >> 63) == 0 { z = mx + my if z >= 0x8000_0000_0000_0000 { ex += 1 z = (z >> 1) | (z & 1) } } else { // mx >= my, since |x| >= |y|. z = mx ~mod- my if z == 0 { // Exact cancellation gives +0 when rounding to nearest. return 0 } else if z < 0x4000_0000_0000_0000 { // If (d >= 2) then z >= (1 << 61) and the normalization shift // is 1. Otherwise, no bits were shifted out of my. s = (this.clz64(a: z) ~mod- 1) & 63 if s >= ex { return 0xFFFF_FFFF_FFFF_FFFF } ex = (ex ~mod- s) & 0x7FF z = z ~mod<< s } } // Round to nearest, ties to even. z is now in [1 << 52 ..= 1 << 53] and // ex in [1 ..= 0x7FF]. Adding z (whose bit 52 is the implicit bit) to the // shifted exponent carries any rounding overflow into the exponent field. rb = z & 0x3FF z = (z ~mod+ 0x200) >> 10 if rb == 0x200 { z &= 0xFFFF_FFFF_FFFF_FFFE } z = (((ex ~mod- 1) as base.u64) ~mod<< 52) ~mod+ z if z >= F64_INF { z = F64_INF } return (x & F64_SIGN) | z } // f64_add_mags returns (sign * (|a| + |b|)). pri func arith.f64_add_mags(a: base.u64, b: base.u64, sign: base.u64) base.u64 { var exp_a : base.u32 var exp_b : base.u32 var exp_z : base.u32 var sig_a : base.u64 var sig_b : base.u64 var sig_z : base.u64 exp_a = ((args.a >> 52) & 0x7FF) as base.u32 exp_b = ((args.b >> 52) & 0x7FF) as base.u32 sig_a = args.a & F64_FRAC sig_b = args.b & F64_FRAC if exp_a == exp_b { if exp_a == 0 { // Both subnormal (or zero). The sum of the fractions might carry // into the exponent field, which is what we want. return args.sign | ((args.a & F64_ABS) ~mod+ sig_b) } else if exp_a == 0x7FF { if (sig_a | sig_b) <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } return args.a } exp_z = exp_a sig_z = ((0x0020_0000_0000_0000 ~mod+ sig_a) ~mod+ sig_b) ~mod<< 9 return this.round_pack_f64(sign: args.sign, exp: exp_z, sig: sig_z) } sig_a = sig_a ~mod<< 9 sig_b = sig_b ~mod<< 9 if exp_a < exp_b { if exp_b == 0x7FF { if sig_b <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } return args.sign | F64_INF } exp_z = exp_b if exp_a <> 0 { sig_a = sig_a ~mod+ 0x2000_0000_0000_0000 } else { sig_a = sig_a ~mod<< 1 } sig_a = this.shift_right_jam64(a: sig_a, dist: exp_b ~mod- exp_a) } else { if exp_a == 0x7FF { if sig_a <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } return args.sign | F64_INF } exp_z = exp_a if exp_b <> 0 { sig_b = sig_b ~mod+ 0x2000_0000_0000_0000 } else { sig_b = sig_b ~mod<< 1 } sig_b = this.shift_right_jam64(a: sig_b, dist: exp_a ~mod- exp_b) } sig_z = (0x2000_0000_0000_0000 ~mod+ sig_a) ~mod+ sig_b if sig_z < 0x4000_0000_0000_0000 { exp_z = exp_z ~mod- 1 sig_z = sig_z ~mod<< 1 } return this.round_pack_f64(sign: args.sign, exp: exp_z, sig: sig_z) } // f64_sub_mags returns (sign * (|a| - |b|)). pri func arith.f64_sub_mags(a: base.u64, b: base.u64, sign: base.u64) base.u64 { var sign_z : base.u64 var exp_a : base.u32 var exp_b : base.u32 var exp_z : base.u32 var sig_a : base.u64 var sig_b : base.u64 var sig_z : base.u64 var shift : base.u32 sign_z = args.sign exp_a = ((args.a >> 52) & 0x7FF) as base.u32 exp_b = ((args.b >> 52) & 0x7FF) as base.u32 sig_a = args.a & F64_FRAC sig_b = args.b & F64_FRAC if exp_a == exp_b { if exp_a == 0x7FF { if (sig_a | sig_b) <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } return F64_DEFAULT_NAN } else if sig_a == sig_b { // Exact cancellation gives +0 when rounding to nearest. return 0 } if exp_a <> 0 { exp_a = exp_a ~mod- 1 } if sig_a < sig_b { sign_z ^= F64_SIGN sig_z = sig_b ~mod- sig_a } else { sig_z = sig_a ~mod- sig_b } // sig_z is non-zero and less than (1 << 52), so shift is at least 1. shift = (this.clz64(a: sig_z) ~mod- 11) & 63 if shift > exp_a { shift = exp_a & 63 exp_z = 0 } else { exp_z = exp_a ~mod- shift } return sign_z | (((exp_z as base.u64) ~mod<< 52) ~mod+ (sig_z ~mod<< shift)) } sig_a = sig_a ~mod<< 10 sig_b = sig_b ~mod<< 10 if exp_a < exp_b { sign_z ^= F64_SIGN if exp_b == 0x7FF { if sig_b <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } return sign_z | F64_INF } if exp_a <> 0 { sig_a = sig_a ~mod+ 0x4000_0000_0000_0000 } else { sig_a = sig_a ~mod+ sig_a } sig_a = this.shift_right_jam64(a: sig_a, dist: exp_b ~mod- exp_a) sig_b = sig_b | 0x4000_0000_0000_0000 exp_z = exp_b sig_z = sig_b ~mod- sig_a } else { if exp_a == 0x7FF { if sig_a <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } return sign_z | F64_INF } if exp_b <> 0 { sig_b = sig_b ~mod+ 0x4000_0000_0000_0000 } else { sig_b = sig_b ~mod+ sig_b } sig_b = this.shift_right_jam64(a: sig_b, dist: exp_a ~mod- exp_b) sig_a = sig_a | 0x4000_0000_0000_0000 exp_z = exp_a sig_z = sig_a ~mod- sig_b } return this.norm_round_pack_f64(sign: sign_z, exp: exp_z ~mod- 1, sig: sig_z) } // f64_mul returns (a * b). pub func arith.f64_mul(a: base.u64, b: base.u64) base.u64 { var sign_z : base.u64 var exp_a : base.u32 var exp_b : base.u32 var exp_z : base.u32 var sig_a : base.u64 var sig_b : base.u64 var sig_z : base.u64 sign_z = (args.a ^ args.b) & F64_SIGN exp_a = ((args.a >> 52) & 0x7FF) as base.u32 exp_b = ((args.b >> 52) & 0x7FF) as base.u32 sig_a = args.a & F64_FRAC sig_b = args.b & F64_FRAC if exp_a == 0x7FF { if (sig_a <> 0) or ((exp_b == 0x7FF) and (sig_b <> 0)) { return this.f64_propagate_nan(a: args.a, b: args.b) } else if (args.b & F64_ABS) == 0 { return F64_DEFAULT_NAN } return sign_z | F64_INF } else if exp_b == 0x7FF { if sig_b <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } else if (args.a & F64_ABS) == 0 { return F64_DEFAULT_NAN } return sign_z | F64_INF } if exp_a == 0 { if sig_a == 0 { return sign_z } exp_a = this.norm_sub_exp(frac: sig_a) sig_a = this.norm_sub_sig(frac: sig_a) } if exp_b == 0 { if sig_b == 0 { return sign_z } exp_b = this.norm_sub_exp(frac: sig_b) sig_b = this.norm_sub_sig(frac: sig_b) } exp_z = (exp_a ~mod+ exp_b) ~mod- 0x3FF sig_a = (sig_a | F64_HIDDEN) ~mod<< 10 sig_b = (sig_b | F64_HIDDEN) ~mod<< 11 sig_z = this.mul_hi64(a: sig_a, b: sig_b) if (sig_a ~mod* sig_b) <> 0 { sig_z |= 1 } if sig_z < 0x4000_0000_0000_0000 { exp_z = exp_z ~mod- 1 sig_z = sig_z ~mod<< 1 } return this.round_pack_f64(sign: sign_z, exp: exp_z, sig: sig_z) } // f64_div returns (a / b). pub func arith.f64_div(a: base.u64, b: base.u64) base.u64 { var sign_z : base.u64 var exp_a : base.u32 var exp_b : base.u32 var exp_z : base.u32 var sig_a : base.u64 var sig_b : base.u64 var q : base.u64 var r : base.u64 var i : base.u32 sign_z = (args.a ^ args.b) & F64_SIGN exp_a = ((args.a >> 52) & 0x7FF) as base.u32 exp_b = ((args.b >> 52) & 0x7FF) as base.u32 sig_a = args.a & F64_FRAC sig_b = args.b & F64_FRAC if exp_a == 0x7FF { if sig_a <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } else if exp_b == 0x7FF { if sig_b <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } return F64_DEFAULT_NAN } return sign_z | F64_INF } else if exp_b == 0x7FF { if sig_b <> 0 { return this.f64_propagate_nan(a: args.a, b: args.b) } return sign_z } if exp_b == 0 { if sig_b == 0 { if (args.a & F64_ABS) == 0 { return F64_DEFAULT_NAN } return sign_z | F64_INF } exp_b = this.norm_sub_exp(frac: sig_b) sig_b = this.norm_sub_sig(frac: sig_b) } if exp_a == 0 { if sig_a == 0 { return sign_z } exp_a = this.norm_sub_exp(frac: sig_a) sig_a = this.norm_sub_sig(frac: sig_a) } exp_z = (exp_a ~mod- exp_b) ~mod+ 0x3FE sig_a = (sig_a & F64_FRAC) | F64_HIDDEN sig_b = (sig_b & F64_FRAC) | F64_HIDDEN if sig_a < sig_b { exp_z = exp_z ~mod- 1 sig_a = sig_a ~mod<< 1 } // Long division, 11 quotient bits at a time. The remainder r is always // less than sig_b (and so less than (1 << 53)), so (r << 11) never // overflows. The final quotient q has 56 bits: 1 integer bit and 55 // fraction bits. q = 1 r = sig_a ~mod- sig_b i = 0 while i < 5 { r = r ~mod<< 11 if sig_b > 0 { q = (q ~mod<< 11) | (r / sig_b) r = r % sig_b } i += 1 } q = q ~mod<< 7 if r <> 0 { q |= 1 } return this.round_pack_f64(sign: sign_z, exp: exp_z, sig: q) } // f64_fma returns ((a * b) + c), computed with a single rounding (a fused // multiply-add), like C's fma function. pub func arith.f64_fma(a: base.u64, b: base.u64, c: base.u64) base.u64 { var sign_p : base.u64 var sign_c : base.u64 var sign_z : base.u64 var exp_a : base.u32 var exp_b : base.u32 var exp_c : base.u32 var exp_z : base.u32 var sig_a : base.u64 var sig_b : base.u64 var sig_c : base.u64 var p_hi : base.u64 var p_lo : base.u64 var c_hi : base.u64 var c_lo : base.u64 var z_hi : base.u64 var z_lo : base.u64 var tmp : base.u64 var d : base.u32 var top : base.u32 if ((args.a & F64_ABS) > F64_INF) or ((args.b & F64_ABS) > F64_INF) { return this.f64_propagate_nan(a: args.a, b: args.b) } else if (args.c & F64_ABS) > F64_INF { return args.c | F64_QUIET } sign_p = (args.a ^ args.b) & F64_SIGN sign_c = args.c & F64_SIGN exp_a = ((args.a >> 52) & 0x7FF) as base.u32 exp_b = ((args.b >> 52) & 0x7FF) as base.u32 exp_c = ((args.c >> 52) & 0x7FF) as base.u32 sig_a = args.a & F64_FRAC sig_b = args.b & F64_FRAC sig_c = args.c & F64_FRAC if (exp_a == 0x7FF) or (exp_b == 0x7FF) { if ((args.a & F64_ABS) == 0) or ((args.b & F64_ABS) == 0) { // (inf * 0) is invalid. return F64_DEFAULT_NAN } else if (exp_c == 0x7FF) and (sign_c <> sign_p) { // (inf - inf) is invalid. return F64_DEFAULT_NAN } return sign_p | F64_INF } else if exp_c == 0x7FF { return args.c } if ((args.a & F64_ABS) == 0) or ((args.b & F64_ABS) == 0) { // The product is an exact (signed) zero. if (args.c & F64_ABS) == 0 { if sign_p == sign_c { return args.c } return 0 } return args.c } // Normalize a and b and compute their exact product, as a 128-bit // integer (p_hi, p_lo), whose value (in 106 bits) is shifted left by 20 // so that it is in [1 << 124, 1 << 126). if exp_a == 0 { exp_a = this.norm_sub_exp(frac: sig_a) sig_a = this.norm_sub_sig(frac: sig_a) } if exp_b == 0 { exp_b = this.norm_sub_exp(frac: sig_b) sig_b = this.norm_sub_sig(frac: sig_b) } sig_a = (sig_a & F64_FRAC) | F64_HIDDEN sig_b = (sig_b & F64_FRAC) | F64_HIDDEN p_hi = this.mul_hi64(a: sig_a, b: sig_b) p_lo = sig_a ~mod* sig_b p_hi = (p_hi ~mod<< 20) | (p_lo >> 44) p_lo = p_lo ~mod<< 20 // The product's value is ((p_hi, p_lo) * 2**(exp_a + exp_b - 2170)). if (args.c & F64_ABS) == 0 { // Find the top bit (in 0 ..= 127) and the round_pack_f64 exponent. top = 127 - this.clz64(a: p_hi) exp_z = ((exp_a ~mod+ exp_b) ~mod+ top) ~mod- 1148 return this.round_pack_f64(sign: sign_p, exp: exp_z, sig: this.jam128_lo(hi: p_hi, lo: p_lo, dist: top ~mod- 62)) } // Normalize c, as a 128-bit integer (c_hi, c_lo) in [1 << 125, 1 << 126) // whose value is ((c_hi, c_lo) * 2**(exp_c - 1148)). if exp_c == 0 { exp_c = this.norm_sub_exp(frac: sig_c) sig_c = this.norm_sub_sig(frac: sig_c) } sig_c = (sig_c & F64_FRAC) | F64_HIDDEN c_hi = sig_c ~mod<< 9 c_lo = 0 // Align the two operands to the larger exponent. d is the (two's // complement) difference in exponents. d = ((exp_a ~mod+ exp_b) ~mod- exp_c) ~mod- 1022 if d < 0x8000_0000 { tmp = this.jam128_hi(hi: c_hi, lo: c_lo, dist: d) c_lo = this.jam128_lo(hi: c_hi, lo: c_lo, dist: d) c_hi = tmp // The common exponent is (exp_a + exp_b - 2170). We will add (top + // 1022) to it, below. Set exp_z to their difference. exp_z = (exp_a ~mod+ exp_b) ~mod- 1148 } else { d = 0 ~mod- d tmp = this.jam128_hi(hi: p_hi, lo: p_lo, dist: d) p_lo = this.jam128_lo(hi: p_hi, lo: p_lo, dist: d) p_hi = tmp // The common exponent is (exp_c - 1148). exp_z = exp_c ~mod- 126 } // Add or subtract the 128-bit values. if sign_p == sign_c { sign_z = sign_p z_lo = p_lo ~mod+ c_lo z_hi = p_hi ~mod+ c_hi if z_lo < p_lo { z_hi = z_hi ~mod+ 1 } } else { if (p_hi > c_hi) or ((p_hi == c_hi) and (p_lo >= c_lo)) { sign_z = sign_p z_lo = p_lo ~mod- c_lo z_hi = p_hi ~mod- c_hi if p_lo < c_lo { z_hi = z_hi ~mod- 1 } } else { sign_z = sign_c z_lo = c_lo ~mod- p_lo z_hi = c_hi ~mod- p_hi if c_lo < p_lo { z_hi = z_hi ~mod- 1 } } if (z_hi | z_lo) == 0 { // Exact cancellation gives +0 when rounding to nearest. return 0 } } // Normalize and round. if z_hi <> 0 { top = 127 - this.clz64(a: z_hi) } else { top = 63 - this.clz64(a: z_lo) } exp_z = exp_z ~mod+ top if top >= 62 { return this.round_pack_f64(sign: sign_z, exp: exp_z, sig: this.jam128_lo(hi: z_hi, lo: z_lo, dist: top - 62)) } return this.round_pack_f64(sign: sign_z, exp: exp_z, sig: z_lo ~mod<< (62 - top)) } // -------- f64 conversions. // f64_from_u64 returns the f64 nearest to the integer a (C's (double)a). pub func arith.f64_from_u64(a: base.u64) base.u64 { return this.norm_round_pack_f64(sign: 0, exp: 0x43C, sig: args.a) } // f64_from_u32 returns the f64 equal to the integer a (C's (double)a). pub func arith.f64_from_u32(a: base.u32) base.u64 { return this.norm_round_pack_f64(sign: 0, exp: 0x43C, sig: args.a as base.u64) } // f64_from_i64 returns the f64 nearest to the integer a (C's (double)a). pub func arith.f64_from_i64(a: base.i64) base.u64 { var m : base.u64 if args.a >= 0 { return this.norm_round_pack_f64(sign: 0, exp: 0x43C, sig: args.a as base.u64) } m = ((0 - (args.a + 1)) as base.u64) + 1 return this.norm_round_pack_f64(sign: F64_SIGN, exp: 0x43C, sig: m) } // f64_to_u64_trunc returns the integer part of a, truncating towards zero, // like C's (uint64_t)a. That C conversion is undefined behavior unless a is // in the range (-1, 1 << 64): this function returns 0 for NaNs and for // negative values and returns 0xFFFF_FFFF_FFFF_FFFF for values at or above // (1 << 64), but callers should not rely on that. pub func arith.f64_to_u64_trunc(a: base.u64) base.u64 { var exp : base.u32 var sig : base.u64 if args.a >= F64_SIGN { return 0 } else if args.a > F64_INF { return 0 } exp = ((args.a >> 52) & 0x7FF) as base.u32 sig = (args.a & F64_FRAC) | F64_HIDDEN if exp < 0x3FF { return 0 } else if exp >= (0x3FF + 64) { return 0xFFFF_FFFF_FFFF_FFFF } else if exp >= 0x433 { return sig ~mod<< ((exp - 0x433) & 63) } return sig >> ((0x433 - exp) & 63) } // f64_to_f32 returns the f32 nearest to a (C's (float)a), rounding to // nearest, ties to even, overflowing to infinity and underflowing to // subnormals or zero. pub func arith.f64_to_f32(a: base.u64) base.u32 { var sign : base.u32 var exp : base.u32 var frac : base.u64 var sig : base.u32 sign = ((args.a >> 32) as base.u32) & 0x8000_0000 exp = ((args.a >> 52) & 0x7FF) as base.u32 frac = args.a & F64_FRAC if exp == 0x7FF { if frac <> 0 { return sign | 0x7FC0_0000 | ((frac >> 29) as base.u32) } return sign | 0x7F80_0000 } // Keep the top 30 bits of the 52 fraction bits, plus a sticky bit. sig = (this.shift_right_jam64(a: frac, dist: 22) & 0x3FFF_FFFF) as base.u32 if (exp == 0) and (sig == 0) { return sign } return this.round_pack_f32(sign: sign, exp: exp ~mod- 0x381, sig: sig | 0x4000_0000) } // round_pack_f32 is the f32 equivalent of round_pack_f64. The sig argument // has its implicit bit at bit 30 and 7 extra bits of precision at the // bottom. The value represented is (sig * 2**(exp - 156)). pri func arith.round_pack_f32(sign: base.u32, exp: base.u32, sig: base.u32) base.u32 { var e : base.u32 var s : base.u32 var round_bits : base.u32 e = args.exp s = args.sig round_bits = s & 0x7F if e >= 0x8000_0000 { s = (this.shift_right_jam64(a: s as base.u64, dist: 0 ~mod- e) & 0xFFFF_FFFF) as base.u32 e = 0 round_bits = s & 0x7F } else if e >= 0xFD { if (e > 0xFD) or (s >= 0x7FFF_FFC0) { return args.sign | 0x7F80_0000 } } s = (s ~mod+ 0x40) >> 7 if round_bits == 0x40 { s = s & 0xFFFF_FFFE } if s == 0 { e = 0 } return args.sign | ((e ~mod<< 23) ~mod+ s) } // -------- Tables. // CLZ8[i] is the number of leading zero bits in the 8-bit value i. pri const CLZ8 : roarray[256] base.u8 = [ 0x08, 0x07, 0x06, 0x06, 0x05, 0x05, 0x05, 0x05, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, ]