Commit b2950866b1

Cody Tapscott <topolarity@tapscott.me>
2022-04-19 05:23:02
compiler_rt: Fix rounding/NaN handling for f80 add/sub
There were a few minor bugs in the rounding behavior and Inf/NaN handling for the f80 __addxf3 and __subtf3 functions. This change updates the original generic implementation to correctly handle f80 floats, including the explicit integer bit.
1 parent d760cae
Changed files (2)
lib
std
special
lib/std/special/compiler_rt/addXf3.zig
@@ -3,6 +3,7 @@
 // https://github.com/llvm/llvm-project/blob/02d85149a05cb1f6dc49f0ba7a2ceca53718ae17/compiler-rt/lib/builtins/fp_add_impl.inc
 
 const std = @import("std");
+const math = std.math;
 const builtin = @import("builtin");
 const compiler_rt = @import("../compiler_rt.zig");
 
@@ -14,6 +15,16 @@ pub fn __adddf3(a: f64, b: f64) callconv(.C) f64 {
     return addXf3(f64, a, b);
 }
 
+pub fn __addxf3(a: f80, b: f80) callconv(.C) f80 {
+    return addXf3(f80, a, b);
+}
+
+pub fn __subxf3(a: f80, b: f80) callconv(.C) f80 {
+    var b_rep = std.math.break_f80(b);
+    b_rep.exp ^= 0x8000;
+    return __addxf3(a, std.math.make_f80(b_rep));
+}
+
 pub fn __addtf3(a: f128, b: f128) callconv(.C) f128 {
     return addXf3(f128, a, b);
 }
@@ -58,10 +69,10 @@ fn normalize(comptime T: type, significand: *std.meta.Int(.unsigned, @typeInfo(T
     const bits = @typeInfo(T).Float.bits;
     const Z = std.meta.Int(.unsigned, bits);
     const S = std.meta.Int(.unsigned, bits - @clz(Z, @as(Z, bits) - 1));
-    const significandBits = std.math.floatMantissaBits(T);
-    const implicitBit = @as(Z, 1) << significandBits;
+    const fractionalBits = math.floatFractionalBits(T);
+    const integerBit = @as(Z, 1) << fractionalBits;
 
-    const shift = @clz(std.meta.Int(.unsigned, bits), significand.*) - @clz(Z, implicitBit);
+    const shift = @clz(std.meta.Int(.unsigned, bits), significand.*) - @clz(Z, integerBit);
     significand.* <<= @intCast(S, shift);
     return 1 - shift;
 }
@@ -73,26 +84,26 @@ fn addXf3(comptime T: type, a: T, b: T) T {
     const S = std.meta.Int(.unsigned, bits - @clz(Z, @as(Z, bits) - 1));
 
     const typeWidth = bits;
-    const significandBits = std.math.floatMantissaBits(T);
-    const exponentBits = std.math.floatExponentBits(T);
+    const significandBits = math.floatMantissaBits(T);
+    const fractionalBits = math.floatFractionalBits(T);
+    const exponentBits = math.floatExponentBits(T);
 
     const signBit = (@as(Z, 1) << (significandBits + exponentBits));
     const maxExponent = ((1 << exponentBits) - 1);
 
-    const implicitBit = (@as(Z, 1) << significandBits);
-    const quietBit = implicitBit >> 1;
-    const significandMask = implicitBit - 1;
+    const integerBit = (@as(Z, 1) << fractionalBits);
+    const quietBit = integerBit >> 1;
+    const significandMask = (@as(Z, 1) << significandBits) - 1;
 
     const absMask = signBit - 1;
-    const exponentMask = absMask ^ significandMask;
-    const qnanRep = exponentMask | quietBit;
+    const qnanRep = @bitCast(Z, math.nan(T)) | quietBit;
 
     var aRep = @bitCast(Z, a);
     var bRep = @bitCast(Z, b);
     const aAbs = aRep & absMask;
     const bAbs = bRep & absMask;
 
-    const infRep = @bitCast(Z, std.math.inf(T));
+    const infRep = @bitCast(Z, math.inf(T));
 
     // Detect if a or b is zero, infinity, or NaN.
     if (aAbs -% @as(Z, 1) >= infRep - @as(Z, 1) or
@@ -157,12 +168,12 @@ fn addXf3(comptime T: type, a: T, b: T) T {
     // implicit significand bit.  (If we fell through from the denormal path it
     // was already set by normalize( ), but setting it twice won't hurt
     // anything.)
-    aSignificand = (aSignificand | implicitBit) << 3;
-    bSignificand = (bSignificand | implicitBit) << 3;
+    aSignificand = (aSignificand | integerBit) << 3;
+    bSignificand = (bSignificand | integerBit) << 3;
 
     // Shift the significand of b by the difference in exponents, with a sticky
     // bottom bit to get rounding correct.
-    const @"align" = @intCast(Z, aExponent - bExponent);
+    const @"align" = @intCast(u32, aExponent - bExponent);
     if (@"align" != 0) {
         if (@"align" < typeWidth) {
             const sticky = if (bSignificand << @intCast(S, typeWidth - @"align") != 0) @as(Z, 1) else 0;
@@ -178,8 +189,8 @@ fn addXf3(comptime T: type, a: T, b: T) T {
 
         // If partial cancellation occured, we need to left-shift the result
         // and adjust the exponent:
-        if (aSignificand < implicitBit << 3) {
-            const shift = @intCast(i32, @clz(Z, aSignificand)) - @intCast(i32, @clz(std.meta.Int(.unsigned, bits), implicitBit << 3));
+        if (aSignificand < integerBit << 3) {
+            const shift = @intCast(i32, @clz(Z, aSignificand)) - @intCast(i32, @clz(std.meta.Int(.unsigned, bits), integerBit << 3));
             aSignificand <<= @intCast(S, shift);
             aExponent -= shift;
         }
@@ -188,7 +199,7 @@ fn addXf3(comptime T: type, a: T, b: T) T {
 
         // If the addition carried up, we need to right-shift the result and
         // adjust the exponent:
-        if (aSignificand & (implicitBit << 4) != 0) {
+        if (aSignificand & (integerBit << 4) != 0) {
             const sticky = aSignificand & 1;
             aSignificand = aSignificand >> 1 | sticky;
             aExponent += 1;
@@ -210,7 +221,7 @@ fn addXf3(comptime T: type, a: T, b: T) T {
     // Low three bits are round, guard, and sticky.
     const roundGuardSticky = aSignificand & 0x7;
 
-    // Shift the significand into place, and mask off the implicit bit.
+    // Shift the significand into place, and mask off the integer bit, if it's implicit.
     var result = (aSignificand >> 3) & significandMask;
 
     // Insert the exponent and sign.
@@ -222,178 +233,12 @@ fn addXf3(comptime T: type, a: T, b: T) T {
     if (roundGuardSticky > 0x4) result += 1;
     if (roundGuardSticky == 0x4) result += result & 1;
 
-    return @bitCast(T, result);
-}
-
-fn normalize_f80(exp: *i32, significand: *u80) void {
-    const shift = @clz(u64, @truncate(u64, significand.*));
-    significand.* = (significand.* << shift);
-    exp.* += -@as(i8, shift);
-}
-
-pub fn __addxf3(a: f80, b: f80) callconv(.C) f80 {
-    var a_rep = std.math.break_f80(a);
-    var b_rep = std.math.break_f80(b);
-    var a_exp: i32 = a_rep.exp & 0x7FFF;
-    var b_exp: i32 = b_rep.exp & 0x7FFF;
-
-    const significand_bits = std.math.floatMantissaBits(f80);
-    const int_bit = 0x8000000000000000;
-    const significand_mask = 0x7FFFFFFFFFFFFFFF;
-    const qnan_bit = 0xC000000000000000;
-    const max_exp = 0x7FFF;
-    const sign_bit = 0x8000;
-
-    // Detect if a or b is infinity, or NaN.
-    if (a_exp == max_exp) {
-        if (a_rep.fraction ^ int_bit == 0) {
-            if (b_exp == max_exp and (b_rep.fraction ^ int_bit == 0)) {
-                // +/-infinity + -/+infinity = qNaN
-                return std.math.qnan_f80;
-            }
-            // +/-infinity + anything = +/- infinity
-            return a;
-        } else {
-            std.debug.assert(a_rep.fraction & significand_mask != 0);
-            // NaN + anything = qNaN
-            a_rep.fraction |= qnan_bit;
-            return std.math.make_f80(a_rep);
-        }
-    }
-    if (b_exp == max_exp) {
-        if (b_rep.fraction ^ int_bit == 0) {
-            // anything + +/-infinity = +/-infinity
-            return b;
-        } else {
-            std.debug.assert(b_rep.fraction & significand_mask != 0);
-            // anything + NaN = qNaN
-            b_rep.fraction |= qnan_bit;
-            return std.math.make_f80(b_rep);
-        }
-    }
-
-    const a_zero = (a_rep.fraction | @bitCast(u32, a_exp)) == 0;
-    const b_zero = (b_rep.fraction | @bitCast(u32, b_exp)) == 0;
-    if (a_zero) {
-        // zero + anything = anything
-        if (b_zero) {
-            // but we need to get the sign right for zero + zero
-            a_rep.exp &= b_rep.exp;
-            return std.math.make_f80(a_rep);
-        } else {
-            return b;
-        }
-    } else if (b_zero) {
-        // anything + zero = anything
-        return a;
-    }
-
-    var a_int: u80 = a_rep.fraction | (@as(u80, a_rep.exp & max_exp) << significand_bits);
-    var b_int: u80 = b_rep.fraction | (@as(u80, b_rep.exp & max_exp) << significand_bits);
-
-    // Swap a and b if necessary so that a has the larger absolute value.
-    if (b_int > a_int) {
-        const temp = a_rep;
-        a_rep = b_rep;
-        b_rep = temp;
-    }
-
-    // Extract the exponent and significand from the (possibly swapped) a and b.
-    a_exp = a_rep.exp & max_exp;
-    b_exp = b_rep.exp & max_exp;
-    a_int = a_rep.fraction;
-    b_int = b_rep.fraction;
-
-    // Normalize any denormals, and adjust the exponent accordingly.
-    normalize_f80(&a_exp, &a_int);
-    normalize_f80(&b_exp, &b_int);
-
-    // The sign of the result is the sign of the larger operand, a.  If they
-    // have opposite signs, we are performing a subtraction; otherwise addition.
-    const result_sign = a_rep.exp & sign_bit;
-    const subtraction = (a_rep.exp ^ b_rep.exp) & sign_bit != 0;
-
-    // Shift the significands to give us round, guard and sticky, and or in the
-    // implicit significand bit.  (If we fell through from the denormal path it
-    // was already set by normalize( ), but setting it twice won't hurt
-    // anything.)
-    a_int = a_int << 3;
-    b_int = b_int << 3;
-
-    // Shift the significand of b by the difference in exponents, with a sticky
-    // bottom bit to get rounding correct.
-    const @"align" = @intCast(u80, a_exp - b_exp);
-    if (@"align" != 0) {
-        if (@"align" < 80) {
-            const sticky = if (b_int << @intCast(u7, 80 - @"align") != 0) @as(u80, 1) else 0;
-            b_int = (b_int >> @truncate(u7, @"align")) | sticky;
-        } else {
-            b_int = 1; // sticky; b is known to be non-zero.
-        }
-    }
-    if (subtraction) {
-        a_int -= b_int;
-        // If a == -b, return +zero.
-        if (a_int == 0) return 0.0;
-
-        // If partial cancellation occurred, we need to left-shift the result
-        // and adjust the exponent:
-        if (a_int < int_bit << 3) {
-            const shift = @intCast(i32, @clz(u80, a_int)) - @intCast(i32, @clz(u80, @as(u80, int_bit) << 3));
-            a_int <<= @intCast(u7, shift);
-            a_exp -= shift;
-        }
-    } else { // addition
-        a_int += b_int;
-
-        // If the addition carried up, we need to right-shift the result and
-        // adjust the exponent:
-        if (a_int & (int_bit << 4) != 0) {
-            const sticky = a_int & 1;
-            a_int = a_int >> 1 | sticky;
-            a_exp += 1;
-        }
+    // Restore any explicit integer bit, if it was rounded off
+    if (significandBits != fractionalBits) {
+        if ((result >> significandBits) != 0) result |= integerBit;
     }
 
-    // If we have overflowed the type, return +/- infinity:
-    if (a_exp >= max_exp) {
-        a_rep.exp = max_exp | result_sign;
-        a_rep.fraction = int_bit; // integer bit is set for +/-inf
-        return std.math.make_f80(a_rep);
-    }
-
-    if (a_exp <= 0) {
-        // Result is denormal before rounding; the exponent is zero and we
-        // need to shift the significand.
-        const shift = @intCast(u80, 1 - a_exp);
-        const sticky = if (a_int << @intCast(u7, 80 - shift) != 0) @as(u1, 1) else 0;
-        a_int = a_int >> @intCast(u7, shift | sticky);
-        a_exp = 0;
-    }
-
-    // Low three bits are round, guard, and sticky.
-    const round_guard_sticky = @truncate(u3, a_int);
-
-    // Shift the significand into place.
-    a_int = @truncate(u64, a_int >> 3);
-
-    // // Insert the exponent and sign.
-    a_int |= (@intCast(u80, a_exp) | result_sign) << significand_bits;
-
-    // Final rounding.  The result may overflow to infinity, but that is the
-    // correct result in that case.
-    if (round_guard_sticky > 0x4) a_int += 1;
-    if (round_guard_sticky == 0x4) a_int += a_int & 1;
-
-    a_rep.fraction = @truncate(u64, a_int);
-    a_rep.exp = @truncate(u16, a_int >> significand_bits);
-    return std.math.make_f80(a_rep);
-}
-
-pub fn __subxf3(a: f80, b: f80) callconv(.C) f80 {
-    var b_rep = std.math.break_f80(b);
-    b_rep.exp ^= 0x8000;
-    return __addxf3(a, std.math.make_f80(b_rep));
+    return @bitCast(T, result);
 }
 
 test {
lib/std/special/compiler_rt/addXf3_test.zig
@@ -3,8 +3,9 @@
 // https://github.com/llvm/llvm-project/blob/02d85149a05cb1f6dc49f0ba7a2ceca53718ae17/compiler-rt/test/builtins/Unit/addtf3_test.c
 // https://github.com/llvm/llvm-project/blob/02d85149a05cb1f6dc49f0ba7a2ceca53718ae17/compiler-rt/test/builtins/Unit/subtf3_test.c
 
+const std = @import("std");
+const math = std.math;
 const qnan128 = @bitCast(f128, @as(u128, 0x7fff800000000000) << 64);
-const inf128 = @bitCast(f128, @as(u128, 0x7fff000000000000) << 64);
 
 const __addtf3 = @import("addXf3.zig").__addtf3;
 
@@ -37,13 +38,14 @@ test "addtf3" {
     try test__addtf3(@bitCast(f128, (@as(u128, 0x7fff000000000000) << 64) | @as(u128, 0x800030000000)), 0x1.23456789abcdefp+5, 0x7fff800000000000, 0x0);
 
     // inf + inf = inf
-    try test__addtf3(inf128, inf128, 0x7fff000000000000, 0x0);
+    try test__addtf3(math.inf(f128), math.inf(f128), 0x7fff000000000000, 0x0);
 
     // inf + any = inf
-    try test__addtf3(inf128, 0x1.2335653452436234723489432abcdefp+5, 0x7fff000000000000, 0x0);
+    try test__addtf3(math.inf(f128), 0x1.2335653452436234723489432abcdefp+5, 0x7fff000000000000, 0x0);
 
     // any + any
     try test__addtf3(0x1.23456734245345543849abcdefp+5, 0x1.edcba52449872455634654321fp-1, 0x40042afc95c8b579, 0x61e58dd6c51eb77c);
+    try test__addtf3(0x1.edcba52449872455634654321fp-1, 0x1.23456734245345543849abcdefp+5, 0x40042afc95c8b579, 0x61e58dd6c51eb77c);
 }
 
 const __subtf3 = @import("addXf3.zig").__subtf3;
@@ -78,8 +80,76 @@ test "subtf3" {
     try test__subtf3(@bitCast(f128, (@as(u128, 0x7fff000000000000) << 64) | @as(u128, 0x800030000000)), 0x1.23456789abcdefp+5, 0x7fff800000000000, 0x0);
 
     // inf - any = inf
-    try test__subtf3(inf128, 0x1.23456789abcdefp+5, 0x7fff000000000000, 0x0);
+    try test__subtf3(math.inf(f128), 0x1.23456789abcdefp+5, 0x7fff000000000000, 0x0);
 
     // any + any
     try test__subtf3(0x1.234567829a3bcdef5678ade36734p+5, 0x1.ee9d7c52354a6936ab8d7654321fp-1, 0x40041b8af1915166, 0xa44a7bca780a166c);
+    try test__subtf3(0x1.ee9d7c52354a6936ab8d7654321fp-1, 0x1.234567829a3bcdef5678ade36734p+5, 0xc0041b8af1915166, 0xa44a7bca780a166c);
+}
+
+const __addxf3 = @import("addXf3.zig").__addxf3;
+const qnan80 = @bitCast(f80, @bitCast(u80, math.nan(f80)) | (1 << (math.floatFractionalBits(f80) - 1)));
+
+fn test__addxf3(a: f80, b: f80, expected: u80) !void {
+    const x = __addxf3(a, b);
+    const rep = @bitCast(u80, x);
+
+    if (rep == expected)
+        return;
+
+    if (math.isNan(@bitCast(f80, expected)) and math.isNan(x))
+        return; // We don't currently test NaN payload propagation
+
+    return error.TestFailed;
+}
+
+test "addxf3" {
+    // NaN + any = NaN
+    try test__addxf3(qnan80, 0x1.23456789abcdefp+5, @bitCast(u80, qnan80));
+    try test__addxf3(@bitCast(f80, @as(u80, 0x7fff_8000_8000_3000_0000)), 0x1.23456789abcdefp+5, @bitCast(u80, qnan80));
+
+    // any + NaN = NaN
+    try test__addxf3(0x1.23456789abcdefp+5, qnan80, @bitCast(u80, qnan80));
+    try test__addxf3(0x1.23456789abcdefp+5, @bitCast(f80, @as(u80, 0x7fff_8000_8000_3000_0000)), @bitCast(u80, qnan80));
+
+    // NaN + inf = NaN
+    try test__addxf3(qnan80, math.inf(f80), @bitCast(u80, qnan80));
+
+    // inf + NaN = NaN
+    try test__addxf3(math.inf(f80), qnan80, @bitCast(u80, qnan80));
+
+    // inf + inf = inf
+    try test__addxf3(math.inf(f80), math.inf(f80), @bitCast(u80, math.inf(f80)));
+
+    // inf + -inf = NaN
+    try test__addxf3(math.inf(f80), -math.inf(f80), @bitCast(u80, qnan80));
+
+    // -inf + inf = NaN
+    try test__addxf3(-math.inf(f80), math.inf(f80), @bitCast(u80, qnan80));
+
+    // inf + any = inf
+    try test__addxf3(math.inf(f80), 0x1.2335653452436234723489432abcdefp+5, @bitCast(u80, math.inf(f80)));
+
+    // any + inf = inf
+    try test__addxf3(0x1.2335653452436234723489432abcdefp+5, math.inf(f80), @bitCast(u80, math.inf(f80)));
+
+    // any + any
+    try test__addxf3(0x1.23456789abcdp+5, 0x1.dcba987654321p+5, 0x4005_BFFFFFFFFFFFC400);
+    try test__addxf3(0x1.23456734245345543849abcdefp+5, 0x1.edcba52449872455634654321fp-1, 0x4004_957E_4AE4_5ABC_B0F3);
+    try test__addxf3(0x1.ffff_ffff_ffff_fffcp+0, 0x1.0p-63, 0x3FFF_FFFFFFFFFFFFFFFF); // exact
+    try test__addxf3(0x1.ffff_ffff_ffff_fffep+0, 0x0.0p0, 0x3FFF_FFFFFFFFFFFFFFFF); // exact
+    try test__addxf3(0x1.ffff_ffff_ffff_fffcp+0, 0x1.4p-63, 0x3FFF_FFFFFFFFFFFFFFFF); // round down
+    try test__addxf3(0x1.ffff_ffff_ffff_fffcp+0, 0x1.8p-63, 0x4000_8000000000000000); // round up to even
+    try test__addxf3(0x1.ffff_ffff_ffff_fffcp+0, 0x1.cp-63, 0x4000_8000000000000000); // round up
+    try test__addxf3(0x1.ffff_ffff_ffff_fffcp+0, 0x2.0p-63, 0x4000_8000000000000000); // exact
+    try test__addxf3(0x1.ffff_ffff_ffff_fffcp+0, 0x2.1p-63, 0x4000_8000000000000000); // round down
+    try test__addxf3(0x1.ffff_ffff_ffff_fffcp+0, 0x3.0p-63, 0x4000_8000000000000000); // round down to even
+    try test__addxf3(0x1.ffff_ffff_ffff_fffcp+0, 0x3.1p-63, 0x4000_8000000000000001); // round up
+    try test__addxf3(0x1.ffff_ffff_ffff_fffcp+0, 0x4.0p-63, 0x4000_8000000000000001); // exact
+
+    try test__addxf3(0x1.0fff_ffff_ffff_fffep+0, 0x1.0p-63, 0x3FFF_8800000000000000); // exact
+    try test__addxf3(0x1.0fff_ffff_ffff_fffep+0, 0x1.7p-63, 0x3FFF_8800000000000000); // round down
+    try test__addxf3(0x1.0fff_ffff_ffff_fffep+0, 0x1.8p-63, 0x3FFF_8800000000000000); // round down to even
+    try test__addxf3(0x1.0fff_ffff_ffff_fffep+0, 0x1.9p-63, 0x3FFF_8800000000000001); // round up
+    try test__addxf3(0x1.0fff_ffff_ffff_fffep+0, 0x2.0p-63, 0x3FFF_8800000000000001); // exact
 }