@zigc/lib 0.16.0-test.1 → 0.17.0-dev.27

package/compiler_rt/tan.zig

@@ -3,6 +3,7 @@
 //!
 //! https://git.musl-libc.org/cgit/musl/tree/src/math/tanf.c
 //! https://git.musl-libc.org/cgit/musl/tree/src/math/tan.c
+//! https://git.musl-libc.org/cgit/musl/tree/src/math/tanl.c
 //! https://golang.org/src/math/tan.go
 
 const std = @import("std");
@@ -10,20 +11,23 @@ const builtin = @import("builtin");
 const math = std.math;
 const mem = std.mem;
 const expect = std.testing.expect;
+const expectApproxEqAbs = std.testing.expectApproxEqAbs;
 
 const kernel = @import("trig.zig");
 const rem_pio2 = @import("rem_pio2.zig").rem_pio2;
 const rem_pio2f = @import("rem_pio2f.zig").rem_pio2f;
+const rem_pio2l = @import("rem_pio2l.zig").rem_pio2l;
+const ld = @import("long_double.zig");
 
 const arch = builtin.cpu.arch;
 const compiler_rt = @import("../compiler_rt.zig");
 const symbol = @import("../compiler_rt.zig").symbol;
 
 comptime {
-    symbol(&__tanh, "__tanh");
+    symbol(&tanh, "__tanh");
     symbol(&tanf, "tanf");
     symbol(&tan, "tan");
-    symbol(&__tanx, "__tanx");
+    symbol(&tanx, "__tanx");
     if (compiler_rt.want_ppc_abi) {
         symbol(&tanq, "tanf128");
     }
@@ -31,7 +35,7 @@ comptime {
     symbol(&tanl, "tanl");
 }
 
-pub fn __tanh(x: f16) callconv(.c) f16 {
+pub fn tanh(x: f16) callconv(.c) f16 {
     // TODO: more efficient implementation
     return @floatCast(tanf(x));
 }
@@ -59,20 +63,20 @@ pub fn tanf(x: f32) callconv(.c) f32 {
             }
             return x;
         }
-        return kernel.__tandf(x, false);
+        return kernel.tandf(x, false);
     }
     if (ix <= 0x407b53d1) { // |x| ~<= 5*pi/4
         if (ix <= 0x4016cbe3) { // |x| ~<= 3pi/4
-            return kernel.__tandf((if (sign) x + t1pio2 else x - t1pio2), true);
+            return kernel.tandf((if (sign) x + t1pio2 else x - t1pio2), true);
         } else {
-            return kernel.__tandf((if (sign) x + t2pio2 else x - t2pio2), false);
+            return kernel.tandf((if (sign) x + t2pio2 else x - t2pio2), false);
         }
     }
     if (ix <= 0x40e231d5) { // |x| ~<= 9*pi/4
         if (ix <= 0x40afeddf) { // |x| ~<= 7*pi/4
-            return kernel.__tandf((if (sign) x + t3pio2 else x - t3pio2), true);
+            return kernel.tandf((if (sign) x + t3pio2 else x - t3pio2), true);
         } else {
-            return kernel.__tandf((if (sign) x + t4pio2 else x - t4pio2), false);
+            return kernel.tandf((if (sign) x + t4pio2 else x - t4pio2), false);
         }
     }
 
@@ -83,7 +87,7 @@ pub fn tanf(x: f32) callconv(.c) f32 {
 
     var y: f64 = undefined;
     const n = rem_pio2f(x, &y);
-    return kernel.__tandf(y, n & 1 != 0);
+    return kernel.tandf(y, n & 1 != 0);
 }
 
 pub fn tan(x: f64) callconv(.c) f64 {
@@ -103,7 +107,7 @@ pub fn tan(x: f64) callconv(.c) f64 {
             }
             return x;
         }
-        return kernel.__tan(x, 0.0, false);
+        return kernel.tan(x, 0.0, false);
     }
 
     // tan(Inf or NaN) is NaN
@@ -113,69 +117,136 @@ pub fn tan(x: f64) callconv(.c) f64 {
 
     var y: [2]f64 = undefined;
     const n = rem_pio2(x, &y);
-    return kernel.__tan(y[0], y[1], n & 1 != 0);
+    return kernel.tan(y[0], y[1], n & 1 != 0);
 }
 
-pub fn __tanx(x: f80) callconv(.c) f80 {
-    // TODO: more efficient implementation
-    return @floatCast(tanq(x));
+pub fn tanx(x: f80) callconv(.c) f80 {
+    const se = ld.signExponent(x) & 0x7fff;
+    if (se == 0x7fff) {
+        return x - x;
+    }
+
+    if (@abs(x) < kernel.pi_4) {
+        if (se < 0x3fff - math.floatMantissaBits(f80) / 2) {
+            if (compiler_rt.want_float_exceptions) {
+                mem.doNotOptimizeAway(if (se == 0) x * 0x1p-120 else x + 0x1p120);
+            }
+            return x;
+        }
+        return kernel.tanx(x, 0.0, 0);
+    }
+
+    var y: [2]f80 = undefined;
+    const n = rem_pio2l(f80, x, &y);
+    return kernel.tanx(y[0], y[1], n & 1);
 }
 
 pub fn tanq(x: f128) callconv(.c) f128 {
-    // TODO: more correct implementation
-    return tan(@floatCast(x));
+    const se = ld.signExponent(x) & 0x7fff;
+    if (se == 0x7fff) {
+        return x - x;
+    }
+
+    if (@abs(x) < kernel.pi_4) {
+        if (se < 0x3fff - math.floatMantissaBits(f128) / 2) {
+            if (compiler_rt.want_float_exceptions) {
+                mem.doNotOptimizeAway(if (se == 0) x * 0x1p-120 else x + 0x1p120);
+            }
+            return x;
+        }
+        return kernel.tanq(x, 0.0, 0);
+    }
+
+    var y: [2]f128 = undefined;
+    const n = rem_pio2l(f128, x, &y);
+    return kernel.tanq(y[0], y[1], n & 1);
 }
 
 pub fn tanl(x: c_longdouble) callconv(.c) c_longdouble {
     switch (@typeInfo(c_longdouble).float.bits) {
-        16 => return __tanh(x),
+        16 => return tanh(x),
         32 => return tanf(x),
         64 => return tan(x),
-        80 => return __tanx(x),
+        80 => return tanx(x),
         128 => return tanq(x),
         else => @compileError("unreachable"),
     }
 }
 
-test "tan" {
-    try expect(tan(@as(f32, 0.0)) == tanf(0.0));
-    try expect(tan(@as(f64, 0.0)) == tan(0.0));
+fn testTanNormal(comptime T: type) !void {
+    const f = switch (T) {
+        f32 => tanf,
+        f64 => tan,
+        else => @compileError("unimplemented"),
+    };
+    const epsilon = 0.00001;
+
+    try expectApproxEqAbs(@as(T, 0.0), f(0.0), epsilon);
+    try expectApproxEqAbs(@as(T, 0.202710), f(0.2), epsilon);
+    try expectApproxEqAbs(@as(T, 1.240422), f(0.8923), epsilon);
+    try expectApproxEqAbs(@as(T, 14.101420), f(1.5), epsilon);
+    try expectApproxEqAbs(@as(T, -0.254397), f(37.45), epsilon);
+    try expectApproxEqAbs(@as(T, 2.285837), f(89.123), epsilon);
+}
+
+fn testTanSpecial(comptime T: type) !void {
+    const f = switch (T) {
+        f32 => tanf,
+        f64 => tan,
+        f80 => tanx,
+        f128 => tanq,
+        else => @compileError("unimplemented"),
+    };
+
+    try expect(math.isPositiveZero(f(0.0)));
+    try expect(math.isNegativeZero(f(-0.0)));
+    try expect(math.isNan(f(math.inf(f32))));
+    try expect(math.isNan(f(-math.inf(f32))));
+    try expect(math.isNan(f(math.nan(f32))));
 }
 
-test "tan32" {
-    const epsilon = 0.00001;
+test "tan32.normal" {
+    try testTanNormal(f32);
+}
+
+test "tan64.normal" {
+    try testTanNormal(f64);
+}
+
+test "tan80.normal" {
+    const epsilon = math.floatEps(f80);
 
-    try expect(math.approxEqAbs(f32, tanf(0.0), 0.0, epsilon));
-    try expect(math.approxEqAbs(f32, tanf(0.2), 0.202710, epsilon));
-    try expect(math.approxEqAbs(f32, tanf(0.8923), 1.240422, epsilon));
-    try expect(math.approxEqAbs(f32, tanf(1.5), 14.101420, epsilon));
-    try expect(math.approxEqAbs(f32, tanf(37.45), -0.254397, epsilon));
-    try expect(math.approxEqAbs(f32, tanf(89.123), 2.285852, epsilon));
+    try expectApproxEqAbs(@as(f80, 0.0), tanx(0.0), epsilon);
+    try expectApproxEqAbs(@as(f80, 0.2027100355086724833213582716475345), tanx(0.2), epsilon);
+    try expectApproxEqAbs(@as(f80, 1.2404217445497097995561220131857544), tanx(0.8923), epsilon);
+    try expectApproxEqAbs(@as(f80, 14.10141994717171938764), tanx(1.5), epsilon);
+    try expectApproxEqAbs(@as(f80, -0.25439607116885656232), tanx(37.45), epsilon);
+    try expectApproxEqAbs(@as(f80, 2.2858376251355320963), tanx(89.123), epsilon);
 }
 
-test "tan64" {
-    const epsilon = 0.000001;
+test "tan128.normal" {
+    const epsilon = math.floatEps(f128);
 
-    try expect(math.approxEqAbs(f64, tan(0.0), 0.0, epsilon));
-    try expect(math.approxEqAbs(f64, tan(0.2), 0.202710, epsilon));
-    try expect(math.approxEqAbs(f64, tan(0.8923), 1.240422, epsilon));
-    try expect(math.approxEqAbs(f64, tan(1.5), 14.101420, epsilon));
-    try expect(math.approxEqAbs(f64, tan(37.45), -0.254397, epsilon));
-    try expect(math.approxEqAbs(f64, tan(89.123), 2.2858376, epsilon));
+    try expectApproxEqAbs(@as(f128, 0.0), tanq(0.0), epsilon);
+    try expectApproxEqAbs(@as(f128, 0.2027100355086724833213582716475345), tanq(0.2), epsilon);
+    try expectApproxEqAbs(@as(f128, 1.2404217445497097995561220131857544), tanq(0.8923), epsilon);
+    try expectApproxEqAbs(@as(f128, 14.101419947171719387646083651987755), tanq(1.5), epsilon);
+    try expectApproxEqAbs(@as(f128, -0.2543960711688565630469573224504774), tanq(37.45), epsilon);
+    try expectApproxEqAbs(@as(f128, 2.2858376251355321074066028114094292), tanq(89.123), epsilon);
 }
 
 test "tan32.special" {
-    try expect(tanf(0.0) == 0.0);
-    try expect(tanf(-0.0) == -0.0);
-    try expect(math.isNan(tanf(math.inf(f32))));
-    try expect(math.isNan(tanf(-math.inf(f32))));
-    try expect(math.isNan(tanf(math.nan(f32))));
+    try testTanSpecial(f32);
 }
 
 test "tan64.special" {
-    try expect(tan(0.0) == 0.0);
-    try expect(tan(-0.0) == -0.0);
-    try expect(math.isNan(tan(math.inf(f64))));
-    try expect(math.isNan(tan(-math.inf(f64))));
-    try expect(math.isNan(tan(math.nan(f64))));
+    try testTanSpecial(f64);
+}
+
+test "tan80.special" {
+    try testTanSpecial(f80);
+}
+
+test "tan128.special" {
+    try testTanSpecial(f128);
 }

package/compiler_rt/trig.zig

@@ -7,6 +7,13 @@
 // https://git.musl-libc.org/cgit/musl/tree/src/math/__sindf.c
 // https://git.musl-libc.org/cgit/musl/tree/src/math/__tand.c
 // https://git.musl-libc.org/cgit/musl/tree/src/math/__tandf.c
+// https://git.musl-libc.org/cgit/musl/tree/src/math/__sinl.c
+// https://git.musl-libc.org/cgit/musl/tree/src/math/__cosl.c
+// https://git.musl-libc.org/cgit/musl/tree/src/math/__tanl.c
+
+const std = @import("std");
+
+pub const pi_4 = std.math.pi / 4.0;
 
 /// kernel cos function on [-pi/4, pi/4], pi/4 ~ 0.785398164
 /// Input x is assumed to be bounded by ~pi/4 in magnitude.
@@ -43,7 +50,7 @@
 ///         expression for cos().  Retention happens in all cases tested
 ///         under FreeBSD, so don't pessimize things by forcibly clipping
 ///         any extra precision in w.
-pub fn __cos(x: f64, y: f64) f64 {
+pub fn cos(x: f64, y: f64) f64 {
     const C1 = 4.16666666666666019037e-02; // 0x3FA55555, 0x5555554C
     const C2 = -1.38888888888741095749e-03; // 0xBF56C16C, 0x16C15177
     const C3 = 2.48015872894767294178e-05; // 0x3EFA01A0, 0x19CB1590
@@ -59,7 +66,7 @@ pub fn __cos(x: f64, y: f64) f64 {
     return w + (((1.0 - w) - hz) + (z * r - x * y));
 }
 
-pub fn __cosdf(x: f64) f32 {
+pub fn cosdf(x: f64) f32 {
     // |cos(x) - c(x)| < 2**-34.1 (~[-5.37e-11, 5.295e-11]).
     const C0 = -0x1ffffffd0c5e81.0p-54; // -0.499999997251031003120
     const C1 = 0x155553e1053a42.0p-57; //  0.0416666233237390631894
@@ -73,6 +80,46 @@ pub fn __cosdf(x: f64) f32 {
     return @floatCast(((1.0 + z * C0) + w * C1) + (w * z) * r);
 }
 
+pub fn cosx(x: f80, y: f80) f80 {
+    const C1: f80 = 0.0416666666666666666136;
+    const C2: f64 = -0.0013888888888888874;
+    const C3: f64 = 0.000024801587301571716;
+    const C4: f64 = -0.00000027557319215507120;
+    const C5: f64 = 0.0000000020876754400407278;
+    const C6: f64 = -1.1470297442401303e-11;
+    const C7: f64 = 4.7383039476436467e-14;
+
+    const z = x * x;
+    const r = z * (C1 + z * (C2 + z * (C3 + z * (C4 +
+        z * (C5 + z * (C6 + z * C7))))));
+    const hz = 0.5 * z;
+    const w = 1.0 - hz;
+
+    return w + (((1.0 - w) - hz) + (z * r - x * y));
+}
+
+pub fn cosq(x: f128, y: f128) f128 {
+    const C1: f128 = 0.04166666666666666666666666666666658424671;
+    const C2: f128 = -0.001388888888888888888888888888863490893732;
+    const C3: f128 = 0.00002480158730158730158730158600795304914210;
+    const C4: f128 = -0.2755731922398589065255474947078934284324e-6;
+    const C5: f128 = 0.2087675698786809897659225313136400793948e-8;
+    const C6: f128 = -0.1147074559772972315817149986812031204775e-10;
+    const C7: f128 = 0.4779477332386808976875457937252120293400e-13;
+    const C8: f64 = -0.1561920696721507929516718307820958119868e-15;
+    const C9: f64 = 0.4110317413744594971475941557607804508039e-18;
+    const C10: f64 = -0.8896592467191938803288521958313920156409e-21;
+    const C11: f64 = 0.1601061435794535138244346256065192782581e-23;
+
+    const z = x * x;
+    const r = z * (C1 + z * (C2 + z * (C3 + z * (C4 + z * (C5 + z * (C6 +
+        z * (C7 + z * (C8 + z * (C9 + z * (C10 + z * C11))))))))));
+    const hz = 0.5 * z;
+    const w = 1.0 - hz;
+
+    return w + (((1.0 - w) - hz) + (z * r - x * y));
+}
+
 /// kernel sin function on ~[-pi/4, pi/4] (except on -0), pi/4 ~ 0.7854
 /// Input x is assumed to be bounded by ~pi/4 in magnitude.
 /// Input y is the tail of x.
@@ -100,7 +147,7 @@ pub fn __cosdf(x: f64) f32 {
 ///              r = x *(S2+x *(S3+x *(S4+x *(S5+x *S6))))
 ///         then                   3    2
 ///              sin(x) = x + (S1*x + (x *(r-y/2)+y))
-pub fn __sin(x: f64, y: f64, iy: i32) f64 {
+pub fn sin(x: f64, y: f64, iy: i32) f64 {
     const S1 = -1.66666666666666324348e-01; // 0xBFC55555, 0x55555549
     const S2 = 8.33333333332248946124e-03; // 0x3F811111, 0x1110F8A6
     const S3 = -1.98412698298579493134e-04; // 0xBF2A01A0, 0x19C161D5
@@ -119,7 +166,7 @@ pub fn __sin(x: f64, y: f64, iy: i32) f64 {
     }
 }
 
-pub fn __sindf(x: f64) f32 {
+pub fn sindf(x: f64) f32 {
     // |sin(x)/x - s(x)| < 2**-37.5 (~[-4.89e-12, 4.824e-12]).
     const S1 = -0x15555554cbac77.0p-55; // -0.166666666416265235595
     const S2 = 0x111110896efbb2.0p-59; //  0.0083333293858894631756
@@ -134,6 +181,52 @@ pub fn __sindf(x: f64) f32 {
     return @floatCast((x + s * (S1 + z * S2)) + s * w * r);
 }
 
+pub fn sinx(x: f80, y: f80, iy: i32) f80 {
+    const S1: f80 = -0.166666666666666666671;
+    const S2: f64 = 0.0083333333333333332;
+    const S3: f64 = -0.00019841269841269427;
+    const S4: f64 = 0.0000027557319223597490;
+    const S5: f64 = -0.000000025052108218074604;
+    const S6: f64 = 1.6059006598854211e-10;
+    const S7: f64 = -7.6429779983024564e-13;
+    const S8: f64 = 2.6174587166648325e-15;
+
+    const z = x * x;
+    const v = z * x;
+    const r = S2 + z * (S3 + z * (S4 + z * (S5 +
+        z * (S6 + z * (S7 + z * S8)))));
+
+    if (iy == 0)
+        return x + v * (S1 + z * r);
+
+    return x - ((z * (0.5 * y - v * r) - y) - v * S1);
+}
+
+pub fn sinq(x: f128, y: f128, iy: i32) f128 {
+    const S1: f128 = -0.16666666666666666666666666666666666606732416116558;
+    const S2: f128 = 0.0083333333333333333333333333333331135404851288270047;
+    const S3: f128 = -0.00019841269841269841269841269839935785325638310428717;
+    const S4: f128 = 0.27557319223985890652557316053039946268333231205686e-5;
+    const S5: f128 = -0.25052108385441718775048214826384312253862930064745e-7;
+    const S6: f128 = 0.16059043836821614596571832194524392581082444805729e-9;
+    const S7: f128 = -0.76471637318198151807063387954939213287488216303768e-12;
+    const S8: f128 = 0.28114572543451292625024967174638477283187397621303e-14;
+    const S9: f64 = -0.82206352458348947812512122163446202498005154296863e-17;
+    const S10: f64 = 0.19572940011906109418080609928334380560135358385256e-19;
+    const S11: f64 = -0.38680813379701966970673724299207480965452616911420e-22;
+    const S12: f64 = 0.64038150078671872796678569586315881020659912139412e-25;
+
+    const z = x * x;
+    const v = z * x;
+    const r = S2 + z * (S3 + z * (S4 + z * (S5 + z * (S6 + z * (S7 + z * (S8 +
+        z * (S9 + z * (S10 + z * (S11 + z * S12)))))))));
+
+    if (iy == 0)
+        return x + v * (S1 + z * r);
+
+    return x - ((z * (0.5 * y - v * r) - y) - v * S1);
+}
+
 /// kernel tan function on ~[-pi/4, pi/4] (except on -0), pi/4 ~ 0.7854
 /// Input x is assumed to be bounded by ~pi/4 in magnitude.
 /// Input y is the tail of x.
@@ -166,7 +259,7 @@ pub fn __sindf(x: f64) f32 {
 ///      4. For x in [0.67434,pi/4],  let y = pi/4 - x, then
 ///              tan(x) = tan(pi/4-y) = (1-tan(y))/(1+tan(y))
 ///                     = 1 - 2*(tan(y) - (tan(y)^2)/(1+tan(y)))
-pub fn __tan(x_: f64, y_: f64, odd: bool) f64 {
+pub fn tan(x_: f64, y_: f64, odd: bool) f64 {
     var x = x_;
     var y = y_;
 
@@ -239,7 +332,7 @@ pub fn __tan(x_: f64, y_: f64, odd: bool) f64 {
     return a0 + a * (1.0 + a0 * w0 + a0 * v);
 }
 
-pub fn __tandf(x: f64, odd: bool) f32 {
+pub fn tandf(x: f64, odd: bool) f32 {
     // |tan(x)/x - t(x)| < 2**-25.5 (~[-2e-08, 2e-08]).
     const T = [_]f64{
         0x15554d3418c99f.0p-54, // 0.333331395030791399758
@@ -271,3 +364,160 @@ pub fn __tandf(x: f64, odd: bool) f32 {
     const r0 = (x + s * u) + (s * w) * (t + w * r);
     return @floatCast(if (odd) -1.0 / r0 else r0);
 }
+
+pub fn tanx(x_: f80, y_: f80, odd: i32) f80 {
+    const pio4: f80 = 0.785398163397448309628;
+    const pio4lo: f80 = -1.25413940316708300586e-20;
+
+    const T3: f80 = 0.333333333333333333180;
+    const T5: f80 = 0.133333333333333372290;
+    const T7: f80 = 0.0539682539682504975744;
+    const T9: f64 = 0.021869488536312216;
+    const T11: f64 = 0.0088632355256619590;
+    const T13: f64 = 0.0035921281113786528;
+    const T15: f64 = 0.0014558334756312418;
+    const T17: f64 = 0.00059003538700862256;
+    const T19: f64 = 0.00023907843576635544;
+    const T21: f64 = 0.000097154625656538905;
+    const T23: f64 = 0.000038440165747303162;
+    const T25: f64 = 0.000018082171885432524;
+    const T27: f64 = 0.0000024196006108814377;
+    const T29: f64 = 0.0000078293456938132840;
+    const T31: f64 = -0.0000032609076735050182;
+    const T33: f64 = 0.0000023261313142559411;
+
+    var x = x_;
+    var y = y_;
+    const big = @abs(x) >= 0.67434;
+    var sign: i8 = 0;
+
+    if (big) {
+        if (x < 0) {
+            sign = -1;
+            x = -x;
+            y = -y;
+        }
+        x = (pio4 - x) + (pio4lo - y);
+        y = 0.0;
+    }
+
+    var z = x * x;
+    var w = z * z;
+
+    var r = T5 + w * (T9 + w * (T13 + w * (T17 + w * (T21 +
+        w * (T25 + w * (T29 + w * T33))))));
+
+    var v = z * (T7 + w * (T11 + w * (T15 + w * (T19 + w * (T23 +
+        w * (T27 + w * T31))))));
+
+    var s = z * x;
+    r = y + z * (s * (r + v) + y) + T3 * s;
+    w = x + r;
+
+    if (big) {
+        s = @as(f80, @floatFromInt(1 - 2 * odd));
+        v = s - 2.0 * (x + (r - w * w / (w + s)));
+        return if (sign == -1) -v else v;
+    }
+
+    if (odd == 0) {
+        return w;
+    }
+
+    // if allow error up to 2 ulp, simply return
+    // -1.0 / (x+r) here
+    //
+    // compute -1.0 / (x+r) accurately
+    z = w + 0x1p32 - 0x1p32;
+    v = r - (z - x);
+    const a = -1.0 / w;
+    const t = a + 0x1p32 - 0x1p32;
+    s = 1.0 + t * z;
+    return t + a * (s + t * v);
+}
+
+pub fn tanq(x_: f128, y_: f128, odd: i32) f128 {
+    const pio4: f128 = 0x1.921fb54442d18469898cc51701b8p-1;
+    const pio4lo: f128 = 0x1.cd129024e088a67cc74020bbea60p-116;
+
+    const T3: f128 = 0x1.5555555555555555555555555553p-2;
+    const T5: f128 = 0x1.1111111111111111111111111eb5p-3;
+    const T7: f128 = 0x1.ba1ba1ba1ba1ba1ba1ba1b694cd6p-5;
+    const T9: f128 = 0x1.664f4882c10f9f32d6bbe09d8bcdp-6;
+    const T11: f128 = 0x1.226e355e6c23c8f5b4f5762322eep-7;
+    const T13: f128 = 0x1.d6d3d0e157ddfb5fed8e84e27b37p-9;
+    const T15: f128 = 0x1.7da36452b75e2b5fce9ee7c2c92ep-10;
+    const T17: f128 = 0x1.355824803674477dfcf726649efep-11;
+    const T19: f128 = 0x1.f57d7734d1656e0aceb716f614c2p-13;
+    const T21: f128 = 0x1.967e18afcb180ed942dfdc518d6cp-14;
+    const T23: f128 = 0x1.497d8eea21e95bc7e2aa79b9f2cdp-15;
+    const T25: f128 = 0x1.0b132d39f055c81be49eff7afd50p-16;
+    const T27: f128 = 0x1.b0f72d33eff7bfa2fbc1059d90b6p-18;
+    const T29: f128 = 0x1.5ef2daf21d1113df38d0fbc00267p-19;
+    const T31: f128 = 0x1.1c77d6eac0234988cdaa04c96626p-20;
+    const T33: f128 = 0x1.cd2a5a292b180e0bdd701057dfe3p-22;
+    const T35: f128 = 0x1.75c7357d0298c01a31d0a6f7d518p-23;
+    const T37: f128 = 0x1.2f3190f4718a9a520f98f50081fcp-24;
+    const T39: f64 = 0.000000028443389121318352;
+    const T41: f64 = 0.000000011981013102001973;
+    const T43: f64 = 0.0000000038303578044958070;
+    const T45: f64 = 0.0000000034664378216909893;
+    const T47: f64 = -0.0000000015090641701997785;
+    const T49: f64 = 0.0000000029449552300483952;
+    const T51: f64 = -0.0000000022006995706097711;
+    const T53: f64 = 0.0000000015468200913196612;
+    const T55: f64 = -0.00000000061311613386849674;
+    const T57: f64 = 1.4912469681508012e-10;
+
+    var x = x_;
+    var y = y_;
+
+    const big = @abs(x) >= 0.67434;
+    var sign: i8 = 0;
+
+    if (big) {
+        if (x < 0) {
+            sign = -1;
+            x = -x;
+            y = -y;
+        }
+        x = (pio4 - x) + (pio4lo - y);
+        y = 0.0;
+    }
+
+    var z = x * x;
+    var w = z * z;
+
+    var r = T5 + w * (T9 + w * (T13 + w * (T17 + w * (T21 +
+        w * (T25 + w * (T29 + w * (T33 + w * (T37 + w * (T41 +
+            w * (T45 + w * (T49 + w * (T53 + w * T57))))))))))));
+
+    var v = z * (T7 + w * (T11 + w * (T15 + w * (T19 + w * (T23 +
+        w * (T27 + w * (T31 + w * (T35 + w * (T39 + w * (T43 +
+            w * (T47 + w * (T51 + w * T55))))))))))));
+
+    var s = z * x;
+    r = y + z * (s * (r + v) + y) + T3 * s;
+    w = x + r;
+
+    if (big) {
+        s = @as(f128, @floatFromInt(1 - 2 * odd));
+        v = s - 2.0 * (x + (r - w * w / (w + s)));
+        return if (sign == -1) -v else v;
+    }
+
+    if (odd == 0) {
+        return w;
+    }
+
+    // if allow error up to 2 ulp, simply return
+    // -1.0 / (x+r) here
+    //
+    // compute -1.0 / (x+r) accurately
+    z = w + 0x1p32 - 0x1p32;
+    v = r - (z - x);
+    const a = -1.0 / w;
+    const t = a + 0x1p32 - 0x1p32;
+    s = 1.0 + t * z;
+    return t + a * (s + t * v);
+}

package/compiler_rt/udivmodei4.zig

@@ -13,6 +13,8 @@ const max_limbs = std.math.divCeil(usize, 65535, 32) catch unreachable; // max s
 comptime {
     symbol(&__udivei4, "__udivei4");
     symbol(&__umodei4, "__umodei4");
+    symbol(&__udivei5, "__udivei5");
+    symbol(&__umodei5, "__umodei5");
 }
 
 /// Get the value of a limb.
@@ -132,6 +134,32 @@ pub fn __umodei4(r_p: [*]u8, u_p: [*]const u8, v_p: [*]const u8, bits: usize) ca
     @call(.always_inline, divmod, .{ null, r, u, v }) catch unreachable;
 }
 
+pub fn __udivei5(q_p: [*]u8, u_p: [*]const u8, v_p: [*]const u8, t_p: [*]u8, bits: usize) callconv(.c) void {
+    @setRuntimeSafety(compiler_rt.test_safety);
+    const byte_size = std.zig.target.intByteSize(&builtin.target, @intCast(bits));
+    const q: []u32 = @ptrCast(@alignCast(q_p[0..byte_size]));
+    const u: []const u32 = @ptrCast(@alignCast(u_p[0..byte_size]));
+    const v: []const u32 = @ptrCast(@alignCast(v_p[0..byte_size]));
+    const tu: []u32 = @ptrCast(@alignCast(t_p[0..byte_size]));
+    _ = tu;
+    const tv: []u32 = @ptrCast(@alignCast(t_p[byte_size..][0..byte_size]));
+    _ = tv;
+    @call(.always_inline, divmod, .{ q, null, u, v }) catch unreachable;
+}
+
+pub fn __umodei5(r_p: [*]u8, u_p: [*]const u8, v_p: [*]const u8, t_p: [*]u8, bits: usize) callconv(.c) void {
+    @setRuntimeSafety(compiler_rt.test_safety);
+    const byte_size = std.zig.target.intByteSize(&builtin.target, @intCast(bits));
+    const r: []u32 = @ptrCast(@alignCast(r_p[0..byte_size]));
+    const u: []const u32 = @ptrCast(@alignCast(u_p[0..byte_size]));
+    const v: []const u32 = @ptrCast(@alignCast(v_p[0..byte_size]));
+    const tu: []u32 = @ptrCast(@alignCast(t_p[0..byte_size]));
+    _ = tu;
+    const tv: []u32 = @ptrCast(@alignCast(t_p[byte_size..][0..byte_size]));
+    _ = tv;
+    @call(.always_inline, divmod, .{ null, r, u, v }) catch unreachable;
+}
+
 test "__udivei4/__umodei4" {
     if (builtin.zig_backend == .stage2_aarch64) return error.SkipZigTest;
     if (builtin.zig_backend == .stage2_c) return error.SkipZigTest;

package/compiler_rt.zig

@@ -279,6 +279,8 @@ comptime {
     _ = @import("compiler_rt/divmodei4.zig");
     _ = @import("compiler_rt/udivmodei4.zig");
 
+    _ = @import("compiler_rt/limb64.zig");
+
     // extra
     _ = @import("compiler_rt/os_version_check.zig");
     _ = @import("compiler_rt/emutls.zig");