zig

fork of https://codeberg.org/ziglang/zig
Log | Files | Refs | README | LICENSE

mulf3.zig (8398B) - Raw


      1 const std = @import("std");
      2 const math = std.math;
      3 const builtin = @import("builtin");
      4 const common = @import("./common.zig");
      5 
      6 /// Ported from:
      7 /// https://github.com/llvm/llvm-project/blob/2ffb1b0413efa9a24eb3c49e710e36f92e2cb50b/compiler-rt/lib/builtins/fp_mul_impl.inc
      8 pub inline fn mulf3(comptime T: type, a: T, b: T) T {
      9     @setRuntimeSafety(common.test_safety);
     10     const typeWidth = @typeInfo(T).float.bits;
     11     const significandBits = math.floatMantissaBits(T);
     12     const fractionalBits = math.floatFractionalBits(T);
     13     const exponentBits = math.floatExponentBits(T);
     14 
     15     const Z = std.meta.Int(.unsigned, typeWidth);
     16 
     17     // ZSignificand is large enough to contain the significand, including an explicit integer bit
     18     const ZSignificand = PowerOfTwoSignificandZ(T);
     19     const ZSignificandBits = @typeInfo(ZSignificand).int.bits;
     20 
     21     const roundBit = (1 << (ZSignificandBits - 1));
     22     const signBit = (@as(Z, 1) << (significandBits + exponentBits));
     23     const maxExponent = ((1 << exponentBits) - 1);
     24     const exponentBias = (maxExponent >> 1);
     25 
     26     const integerBit = (@as(ZSignificand, 1) << fractionalBits);
     27     const quietBit = integerBit >> 1;
     28     const significandMask = (@as(Z, 1) << significandBits) - 1;
     29 
     30     const absMask = signBit - 1;
     31     const qnanRep = @as(Z, @bitCast(math.nan(T))) | quietBit;
     32     const infRep: Z = @bitCast(math.inf(T));
     33     const minNormalRep: Z = @bitCast(math.floatMin(T));
     34 
     35     const ZExp = if (typeWidth >= 32) u32 else Z;
     36     const aExponent: ZExp = @truncate((@as(Z, @bitCast(a)) >> significandBits) & maxExponent);
     37     const bExponent: ZExp = @truncate((@as(Z, @bitCast(b)) >> significandBits) & maxExponent);
     38     const productSign: Z = (@as(Z, @bitCast(a)) ^ @as(Z, @bitCast(b))) & signBit;
     39 
     40     var aSignificand: ZSignificand = @intCast(@as(Z, @bitCast(a)) & significandMask);
     41     var bSignificand: ZSignificand = @intCast(@as(Z, @bitCast(b)) & significandMask);
     42     var scale: i32 = 0;
     43 
     44     // Detect if a or b is zero, denormal, infinity, or NaN.
     45     if (aExponent -% 1 >= maxExponent - 1 or bExponent -% 1 >= maxExponent - 1) {
     46         const aAbs: Z = @as(Z, @bitCast(a)) & absMask;
     47         const bAbs: Z = @as(Z, @bitCast(b)) & absMask;
     48 
     49         // NaN * anything = qNaN
     50         if (aAbs > infRep) return @bitCast(@as(Z, @bitCast(a)) | quietBit);
     51         // anything * NaN = qNaN
     52         if (bAbs > infRep) return @bitCast(@as(Z, @bitCast(b)) | quietBit);
     53 
     54         if (aAbs == infRep) {
     55             // infinity * non-zero = +/- infinity
     56             if (bAbs != 0) {
     57                 return @bitCast(aAbs | productSign);
     58             } else {
     59                 // infinity * zero = NaN
     60                 return @bitCast(qnanRep);
     61             }
     62         }
     63 
     64         if (bAbs == infRep) {
     65             //? non-zero * infinity = +/- infinity
     66             if (aAbs != 0) {
     67                 return @bitCast(bAbs | productSign);
     68             } else {
     69                 // zero * infinity = NaN
     70                 return @bitCast(qnanRep);
     71             }
     72         }
     73 
     74         // zero * anything = +/- zero
     75         if (aAbs == 0) return @bitCast(productSign);
     76         // anything * zero = +/- zero
     77         if (bAbs == 0) return @bitCast(productSign);
     78 
     79         // one or both of a or b is denormal, the other (if applicable) is a
     80         // normal number.  Renormalize one or both of a and b, and set scale to
     81         // include the necessary exponent adjustment.
     82         if (aAbs < minNormalRep) scale += normalize(T, &aSignificand);
     83         if (bAbs < minNormalRep) scale += normalize(T, &bSignificand);
     84     }
     85 
     86     // Or in the implicit significand bit.  (If we fell through from the
     87     // denormal path it was already set by normalize( ), but setting it twice
     88     // won't hurt anything.)
     89     aSignificand |= integerBit;
     90     bSignificand |= integerBit;
     91 
     92     // Get the significand of a*b.  Before multiplying the significands, shift
     93     // one of them left to left-align it in the field.  Thus, the product will
     94     // have (exponentBits + 2) integral digits, all but two of which must be
     95     // zero.  Normalizing this result is just a conditional left-shift by one
     96     // and bumping the exponent accordingly.
     97     var productHi: ZSignificand = undefined;
     98     var productLo: ZSignificand = undefined;
     99     const left_align_shift = ZSignificandBits - fractionalBits - 1;
    100     common.wideMultiply(ZSignificand, aSignificand, bSignificand << left_align_shift, &productHi, &productLo);
    101 
    102     var productExponent: i32 = @as(i32, @intCast(aExponent + bExponent)) - exponentBias + scale;
    103 
    104     // Normalize the significand, adjust exponent if needed.
    105     if ((productHi & integerBit) != 0) {
    106         productExponent +%= 1;
    107     } else {
    108         productHi = (productHi << 1) | (productLo >> (ZSignificandBits - 1));
    109         productLo = productLo << 1;
    110     }
    111 
    112     // If we have overflowed the type, return +/- infinity.
    113     if (productExponent >= maxExponent) return @bitCast(infRep | productSign);
    114 
    115     var result: Z = undefined;
    116     if (productExponent <= 0) {
    117         // Result is denormal before rounding
    118         //
    119         // If the result is so small that it just underflows to zero, return
    120         // a zero of the appropriate sign.  Mathematically there is no need to
    121         // handle this case separately, but we make it a special case to
    122         // simplify the shift logic.
    123         const shift: u32 = @truncate(@as(Z, 1) -% @as(u32, @bitCast(productExponent)));
    124         if (shift >= ZSignificandBits) return @bitCast(productSign);
    125 
    126         // Otherwise, shift the significand of the result so that the round
    127         // bit is the high bit of productLo.
    128         const sticky = wideShrWithTruncation(ZSignificand, &productHi, &productLo, shift);
    129         productLo |= @intFromBool(sticky);
    130         result = productHi;
    131 
    132         // We include the integer bit so that rounding will carry to the exponent,
    133         // but it will be removed later if the result is still denormal
    134         if (significandBits != fractionalBits) result |= integerBit;
    135     } else {
    136         // Result is normal before rounding; insert the exponent.
    137         result = productHi & significandMask;
    138         result |= @as(Z, @intCast(productExponent)) << significandBits;
    139     }
    140 
    141     // Final rounding.  The final result may overflow to infinity, or underflow
    142     // to zero, but those are the correct results in those cases.  We use the
    143     // default IEEE-754 round-to-nearest, ties-to-even rounding mode.
    144     if (productLo > roundBit) result +%= 1;
    145     if (productLo == roundBit) result +%= result & 1;
    146 
    147     // Restore any explicit integer bit, if it was rounded off
    148     if (significandBits != fractionalBits) {
    149         if ((result >> significandBits) != 0) {
    150             result |= integerBit;
    151         } else {
    152             result &= ~integerBit;
    153         }
    154     }
    155 
    156     // Insert the sign of the result:
    157     result |= productSign;
    158 
    159     return @bitCast(result);
    160 }
    161 
    162 /// Returns `true` if the right shift is inexact (i.e. any bit shifted out is non-zero)
    163 ///
    164 /// This is analogous to an shr version of `@shlWithOverflow`
    165 fn wideShrWithTruncation(comptime Z: type, hi: *Z, lo: *Z, count: u32) bool {
    166     @setRuntimeSafety(common.test_safety);
    167     const typeWidth = @typeInfo(Z).int.bits;
    168     var inexact = false;
    169     if (count < typeWidth) {
    170         inexact = (lo.* << @intCast(typeWidth -% count)) != 0;
    171         lo.* = (hi.* << @intCast(typeWidth -% count)) | (lo.* >> @intCast(count));
    172         hi.* = hi.* >> @intCast(count);
    173     } else if (count < 2 * typeWidth) {
    174         inexact = (hi.* << @intCast(2 * typeWidth -% count) | lo.*) != 0;
    175         lo.* = hi.* >> @intCast(count -% typeWidth);
    176         hi.* = 0;
    177     } else {
    178         inexact = (hi.* | lo.*) != 0;
    179         lo.* = 0;
    180         hi.* = 0;
    181     }
    182     return inexact;
    183 }
    184 
    185 fn normalize(comptime T: type, significand: *PowerOfTwoSignificandZ(T)) i32 {
    186     const Z = PowerOfTwoSignificandZ(T);
    187     const integerBit = @as(Z, 1) << math.floatFractionalBits(T);
    188 
    189     const shift = @clz(significand.*) - @clz(integerBit);
    190     significand.* <<= @intCast(shift);
    191     return @as(i32, 1) - shift;
    192 }
    193 
    194 /// Returns a power-of-two integer type that is large enough to contain
    195 /// the significand of T, including an explicit integer bit
    196 fn PowerOfTwoSignificandZ(comptime T: type) type {
    197     const bits = math.ceilPowerOfTwoAssert(u16, math.floatFractionalBits(T) + 1);
    198     return std.meta.Int(.unsigned, bits);
    199 }
    200 
    201 test {
    202     _ = @import("mulf3_test.zig");
    203 }