cmd/compile: truncates constants

[dvyukov]

The following program:

package main
import "fmt"
func main() {
    var g = 1e81391777742999
    fmt.Println(g)
}

prints:

1e+151

That's a wrong answer.

go version devel +514014c Thu Jun 18 15:54:35 2015 +0200 linux/amd64

[griesemer]

This is related to a math/big bug. See issue #11341 .

Once issue #11341 is fixed, re-vendor math/big to src/cmd/compile/internal. The respective method call in the compiler is in mparith3.go:156 (function mpatoflt).

Depending on the math/big fix, the value may become +Inf (and then we're all done), or we may need to add a range check.

[rsc]

I am a bit worried about deferring to math/big on this. For example, the value should not become +Inf, since that does not mean anything inside the compiler. More generally, math/big's job is to do what is asked, and the compiler has the added requirement of not running forever. The two are different, so it makes sense for the compiler to have some checks of its own. I will send a CL rejecting numbers with too many exponent digits. That should defend against a whole class of problems.

[gopherbot]

CL https://golang.org/cl/11673 mentions this issue.

[rsc]

I'm going to cut the size of the accepted exponents in the gc compiler for Go 1.5 to no more than 8 digits. It's not perfect but it eliminates the simplest problems. We can revisit both correctness and performance for Go 1.6.

In particular, if I write:

var h = 1e2147483647

then I should get an error that that doesn't fit in a float64, but instead the error says a completely different number doesn't fit in the float64:

x.go:13: constant 0.93342e+536870911 overflows float64