std.fmt.parseFloat: add f80 formatFloat support

This commit is contained in:
Harrison McCarty 2024-06-30 04:12:52 -07:00 committed by Marc Tiehuis
parent 959d227d13
commit 8ff01f78f3
7 changed files with 79 additions and 33 deletions

View File

@ -1521,15 +1521,13 @@ fn check(comptime T: type, value: T, comptime expected: []const u8) !void {
const s = try formatFloat(&buf, value, .{});
try std.testing.expectEqualStrings(expected, s);
if (@bitSizeOf(T) != 80) {
const o = try std.fmt.parseFloat(T, s);
const o_bits: I = @bitCast(o);
const o = try std.fmt.parseFloat(T, s);
const o_bits: I = @bitCast(o);
if (std.math.isNan(value)) {
try std.testing.expect(std.math.isNan(o));
} else {
try std.testing.expectEqual(value_bits, o_bits);
}
if (std.math.isNan(value)) {
try std.testing.expect(std.math.isNan(o));
} else {
try std.testing.expectEqual(value_bits, o_bits);
}
}

View File

@ -21,10 +21,6 @@ pub fn parseFloat(comptime T: type, s: []const u8) ParseFloatError!T {
@compileError("Cannot parse a float into a non-floating point type.");
}
if (T == f80) {
@compileError("TODO support parsing float to f80");
}
if (s.len == 0) {
return error.InvalidCharacter;
}
@ -75,7 +71,7 @@ pub fn parseFloat(comptime T: type, s: []const u8) ParseFloatError!T {
// See https://github.com/tiehuis/parse-number-fxx-test-data for a wider-selection of test-data.
test parseFloat {
inline for ([_]type{ f16, f32, f64, f128 }) |T| {
inline for ([_]type{ f16, f32, f64, f80, f128 }) |T| {
try testing.expectError(error.InvalidCharacter, parseFloat(T, ""));
try testing.expectError(error.InvalidCharacter, parseFloat(T, " 1"));
try testing.expectError(error.InvalidCharacter, parseFloat(T, "1abc"));
@ -131,7 +127,7 @@ test parseFloat {
}
test "nan and inf" {
inline for ([_]type{ f16, f32, f64, f128 }) |T| {
inline for ([_]type{ f16, f32, f64, f80, f128 }) |T| {
const Z = std.meta.Int(.unsigned, @typeInfo(T).Float.bits);
try expectEqual(@as(Z, @bitCast(try parseFloat(T, "nAn"))), @as(Z, @bitCast(std.math.nan(T))));
@ -144,6 +140,7 @@ test "largest normals" {
try expectEqual(@as(u16, @bitCast(try parseFloat(f16, "65504"))), 0x7bff);
try expectEqual(@as(u32, @bitCast(try parseFloat(f32, "3.4028234664E38"))), 0x7f7f_ffff);
try expectEqual(@as(u64, @bitCast(try parseFloat(f64, "1.7976931348623157E308"))), 0x7fef_ffff_ffff_ffff);
try expectEqual(@as(u80, @bitCast(try parseFloat(f80, "1.189731495357231765E4932"))), 0x7ffe_ffff_ffff_ffff_ffff);
try expectEqual(@as(u128, @bitCast(try parseFloat(f128, "1.1897314953572317650857593266280070162E4932"))), 0x7ffe_ffff_ffff_ffff_ffff_ffff_ffff_ffff);
}
@ -152,8 +149,8 @@ test "#11169" {
}
test "many_digits hex" {
const a: f32 = try std.fmt.parseFloat(f32, "0xffffffffffffffff.0p0");
const b: f32 = @floatCast(try std.fmt.parseFloat(f128, "0xffffffffffffffff.0p0"));
const a: f32 = try parseFloat(f32, "0xffffffffffffffff.0p0");
const b: f32 = @floatCast(try parseFloat(f128, "0xffffffffffffffff.0p0"));
try std.testing.expectEqual(a, b);
}
@ -163,6 +160,7 @@ test "hex.special" {
try testing.expect(math.isPositiveInf(try parseFloat(f32, "+Inf")));
try testing.expect(math.isNegativeInf(try parseFloat(f32, "-iNf")));
}
test "hex.zero" {
try testing.expectEqual(@as(f32, 0.0), try parseFloat(f32, "0x0"));
try testing.expectEqual(@as(f32, 0.0), try parseFloat(f32, "-0x0"));
@ -221,6 +219,23 @@ test "hex.f64" {
try testing.expectEqual(try parseFloat(f64, "0x1p-1074"), math.floatTrueMin(f64));
try testing.expectEqual(try parseFloat(f64, "-0x1p-1074"), -math.floatTrueMin(f64));
}
test "hex.f80" {
try testing.expectEqual(try parseFloat(f80, "0x1p0"), 1.0);
try testing.expectEqual(try parseFloat(f80, "-0x1p-1"), -0.5);
try testing.expectEqual(try parseFloat(f80, "0x10p+10"), 16384.0);
try testing.expectEqual(try parseFloat(f80, "0x10p-10"), 0.015625);
// Max normalized value.
try testing.expectEqual(try parseFloat(f80, "0xf.fffffffffffffff7p+16380"), math.floatMax(f80));
try testing.expectEqual(try parseFloat(f80, "-0xf.fffffffffffffff7p+16380"), -math.floatMax(f80));
// Min normalized value.
try testing.expectEqual(try parseFloat(f80, "0x1p-16382"), math.floatMin(f80));
try testing.expectEqual(try parseFloat(f80, "-0x1p-16382"), -math.floatMin(f80));
// Min denormalized value.
try testing.expectEqual(try parseFloat(f80, "0x1p-16445"), math.floatTrueMin(f80));
try testing.expectEqual(try parseFloat(f80, "-0x1p-16445"), -math.floatTrueMin(f80));
}
test "hex.f128" {
try testing.expectEqual(try parseFloat(f128, "0x1p0"), 1.0);
try testing.expectEqual(try parseFloat(f128, "-0x1p-1"), -0.5);
@ -232,7 +247,7 @@ test "hex.f128" {
// Min normalized value.
try testing.expectEqual(try parseFloat(f128, "0x1p-16382"), math.floatMin(f128));
try testing.expectEqual(try parseFloat(f128, "-0x1p-16382"), -math.floatMin(f128));
// // Min denormalized value.
// Min denormalized value.
try testing.expectEqual(try parseFloat(f128, "0x1p-16494"), math.floatTrueMin(f128));
try testing.expectEqual(try parseFloat(f128, "-0x1p-16494"), -math.floatTrueMin(f128));
// ensure round-to-even

View File

@ -60,7 +60,7 @@ pub fn from(comptime T: type) Self {
.max_exponent_fast_path_disguised = 7,
.max_mantissa_fast_path = 2 << std.math.floatMantissaBits(T),
// Slow + Eisel-Lemire
.mantissa_explicit_bits = std.math.floatMantissaBits(T),
.mantissa_explicit_bits = std.math.floatFractionalBits(T),
.infinite_power = 0x1f,
// Eisel-Lemire
.smallest_power_of_ten = -26, // TODO: refine, fails one test
@ -81,7 +81,7 @@ pub fn from(comptime T: type) Self {
.max_exponent_fast_path_disguised = 17,
.max_mantissa_fast_path = 2 << std.math.floatMantissaBits(T),
// Slow + Eisel-Lemire
.mantissa_explicit_bits = std.math.floatMantissaBits(T),
.mantissa_explicit_bits = std.math.floatFractionalBits(T),
.infinite_power = 0xff,
// Eisel-Lemire
.smallest_power_of_ten = -65,
@ -106,6 +106,26 @@ pub fn from(comptime T: type) Self {
.min_exponent_round_to_even = -4,
.max_exponent_round_to_even = 23,
},
f80 => .{
// Fast-Path
.min_exponent_fast_path = -27,
.max_exponent_fast_path = 27,
.max_exponent_fast_path_disguised = 46,
.max_mantissa_fast_path = 2 << std.math.floatMantissaBits(T),
// Slow + Eisel-Lemire
.mantissa_explicit_bits = std.math.floatFractionalBits(T),
.infinite_power = 0x7fff,
// Eisel-Lemire.
// NOTE: Not yet tested (no f80 eisel-lemire implementation)
.smallest_power_of_ten = -4966,
.largest_power_of_ten = 4932,
.minimum_exponent = -16382,
// 2^65 * 5^-q < 2^80
// 5^-q < 2^15
// => q >= -6
.min_exponent_round_to_even = -6,
.max_exponent_round_to_even = 28,
},
f128 => .{
// Fast-Path
.min_exponent_fast_path = -48,
@ -113,7 +133,7 @@ pub fn from(comptime T: type) Self {
.max_exponent_fast_path_disguised = 82,
.max_mantissa_fast_path = 2 << std.math.floatMantissaBits(T),
// Slow + Eisel-Lemire
.mantissa_explicit_bits = std.math.floatMantissaBits(T),
.mantissa_explicit_bits = std.math.floatFractionalBits(T),
.infinite_power = 0x7fff,
// Eisel-Lemire.
// NOTE: Not yet tested (no f128 eisel-lemire implementation)

View File

@ -23,7 +23,11 @@ pub fn BiasedFp(comptime T: type) type {
}
pub fn inf(comptime FloatT: type) Self {
return .{ .f = 0, .e = (1 << std.math.floatExponentBits(FloatT)) - 1 };
const e = (1 << std.math.floatExponentBits(FloatT)) - 1;
return switch (FloatT) {
f80 => .{ .f = 0x8000000000000000, .e = e },
else => .{ .f = 0, .e = e },
};
}
pub fn eql(self: Self, other: Self) bool {
@ -45,6 +49,7 @@ pub fn floatFromUnsigned(comptime T: type, comptime MantissaT: type, v: Mantissa
f16 => @as(f16, @bitCast(@as(u16, @truncate(v)))),
f32 => @as(f32, @bitCast(@as(u32, @truncate(v)))),
f64 => @as(f64, @bitCast(@as(u64, @truncate(v)))),
f80 => @as(f80, @bitCast(@as(u80, @truncate(v)))),
f128 => @as(f128, @bitCast(v)),
else => unreachable,
};
@ -85,7 +90,7 @@ pub fn isDigit(c: u8, comptime base: u8) bool {
pub fn mantissaType(comptime T: type) type {
return switch (T) {
f16, f32, f64 => u64,
f128 => u128,
f80, f128 => u128,
else => unreachable,
};
}

View File

@ -46,6 +46,13 @@ fn fastPow10(comptime T: type, i: usize) T {
0, 0, 0, 0, 0, 0, 0, 0,
})[i & 31],
f80 => ([32]f80{
1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7,
1e8, 1e9, 1e10, 1e11, 1e12, 1e13, 1e14, 1e15,
1e16, 1e17, 1e18, 1e19, 1e20, 1e21, 1e22, 1e23,
1e24, 1e25, 1e26, 1e27, 0, 0, 0, 0,
})[i & 31],
f128 => ([64]f128{
1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7,
1e8, 1e9, 1e10, 1e11, 1e12, 1e13, 1e14, 1e15,

View File

@ -25,11 +25,12 @@ pub fn convertHex(comptime T: type, n_: Number(T)) T {
const max_exp = math.floatExponentMax(T);
const min_exp = math.floatExponentMin(T);
const mantissa_bits = math.floatMantissaBits(T);
const fractional_bits = math.floatFractionalBits(T);
const exp_bits = math.floatExponentBits(T);
const exp_bias = min_exp - 1;
// mantissa now implicitly divided by 2^mantissa_bits
n.exponent += mantissa_bits;
// mantissa now implicitly divided by 2^fractional_bits
n.exponent += fractional_bits;
// Shift mantissa and exponent to bring representation into float range.
// Eventually we want a mantissa with a leading 1-bit followed by mantbits other bits.
@ -44,7 +45,7 @@ pub fn convertHex(comptime T: type, n_: Number(T)) T {
if (n.many_digits) {
n.mantissa |= 1;
}
while (n.mantissa >> (1 + mantissa_bits + 2) != 0) {
while (n.mantissa >> (1 + fractional_bits + 2) != 0) {
n.mantissa = (n.mantissa >> 1) | (n.mantissa & 1);
n.exponent += 1;
}
@ -64,14 +65,14 @@ pub fn convertHex(comptime T: type, n_: Number(T)) T {
n.exponent += 2;
if (round == 3) {
n.mantissa += 1;
if (n.mantissa == 1 << (1 + mantissa_bits)) {
if (n.mantissa == 1 << (1 + fractional_bits)) {
n.mantissa >>= 1;
n.exponent += 1;
}
}
// Denormal or zero
if (n.mantissa >> mantissa_bits == 0) {
if (n.mantissa >> fractional_bits == 0) {
n.exponent = exp_bias;
}

View File

@ -41,7 +41,7 @@ pub fn convertSlow(comptime T: type, s: []const u8) BiasedFp(T) {
const MantissaT = mantissaType(T);
const min_exponent = -(1 << (math.floatExponentBits(T) - 1)) + 1;
const infinite_power = (1 << math.floatExponentBits(T)) - 1;
const mantissa_explicit_bits = math.floatMantissaBits(T);
const fractional_bits = math.floatFractionalBits(T);
var d = Decimal(T).parse(s); // no need to recheck underscores
if (d.num_digits == 0 or d.decimal_point < Decimal(T).min_exponent) {
@ -97,9 +97,9 @@ pub fn convertSlow(comptime T: type, s: []const u8) BiasedFp(T) {
// Shift the decimal to the hidden bit, and then round the value
// to get the high mantissa+1 bits.
d.leftShift(mantissa_explicit_bits + 1);
d.leftShift(fractional_bits + 1);
var mantissa = d.round();
if (mantissa >= (@as(MantissaT, 1) << (mantissa_explicit_bits + 1))) {
if (mantissa >= (@as(MantissaT, 1) << (fractional_bits + 1))) {
// Rounding up overflowed to the carry bit, need to
// shift back to the hidden bit.
d.rightShift(1);
@ -110,10 +110,10 @@ pub fn convertSlow(comptime T: type, s: []const u8) BiasedFp(T) {
}
}
var power2 = exp2 - min_exponent;
if (mantissa < (@as(MantissaT, 1) << mantissa_explicit_bits)) {
if (mantissa < (@as(MantissaT, 1) << fractional_bits)) {
power2 -= 1;
}
// Zero out all the bits above the explicit mantissa bits.
mantissa &= (@as(MantissaT, 1) << mantissa_explicit_bits) - 1;
// Zero out all the bits above the mantissa bits.
mantissa &= (@as(MantissaT, 1) << math.floatMantissaBits(T)) - 1;
return .{ .f = mantissa, .e = power2 };
}