// $G $D/$F.go && $L $F.$A && ./$A.out || echo BUG: bug120 // Copyright 2009 The Go Authors. All rights reserved. // Use of this source code is governed by a BSD-style // license that can be found in the LICENSE file. package main import "strconv"; type Test struct { f float64; in string; out string; } var tests = []Test { Test{ 123.5, "123.5", "123.5" }, Test{ 456.7, "456.7", "456.7" }, Test{ 1e23+8.5e6, "1e23+8.5e6", "1.0000000000000001e+23" }, Test{ 100000000000000008388608, "100000000000000008388608", "1.0000000000000001e+23" }, Test{ 1e23+8388609, "1e23+8388609", "1.0000000000000001e+23" }, // "x" = the floating point value from converting the string x. // These are exactly representable in 64-bit floating point: // 1e23-8388608 // 1e23+8388608 // The former has an even mantissa, so "1e23" rounds to 1e23-8388608. // If "1e23+8388608" is implemented as "1e23" + "8388608", // that ends up computing 1e23-8388608 + 8388608 = 1e23, // which rounds back to 1e23-8388608. // The correct answer, of course, would be "1e23+8388608" = 1e23+8388608. // This is not going to be correct until 6g has multiprecision floating point. // A simpler case is "1e23+1", which should also round to 1e23+8388608. Test{ 1e23+8.388608e6, "1e23+8.388608e6", "1.0000000000000001e+23" }, Test{ 1e23+1, "1e23+1", "1.0000000000000001e+23" }, } func main() { ok := true; for i := 0; i < len(tests); i++ { t := tests[i]; v := strconv.Ftoa64(t.f, 'g', -1); if v != t.out { println("Bad float64 const:", t.in, "want", t.out, "got", v); x, err := strconv.Atof64(t.out); if err != nil { panicln("bug120: strconv.Atof64", t.out); } println("\twant exact:", strconv.Ftoa64(x, 'g', 1000)); println("\tgot exact: ", strconv.Ftoa64(t.f, 'g', 1000)); ok = false; } } if !ok { sys.Exit(1); } }