printing float value causes fmt to panic

[kristoff-it]

Zig Version

0.10.0-dev.768+d811e03fb

Steps to Reproduce

const std = @import("std");

pub fn main() void {
    const float: f64 = 1.000000e+37;
    std.debug.print("value: {}\n", .{float});
}

Expected Behavior

Runs.

Actual Behavior

value: thread 934843 panic: integer overflow
/home/kristoff/zig-master/lib/std/fmt/errol.zig:339:26: 0x24f1af in std.fmt.errol.errolInt (float)
        buffer[buf_index - 1] += @boolToInt(buffer[buf_index] >= '5');
                         ^
...

Context

This error was encountered while trying to print values from a file in the standard library.

zig/lib/std/fmt/errol/lookup.zig

Line 277 in d298719

HP{ .val = 1.000000e+37, .off = 4.612373417978788577e+20 },

The highlighted line corresponds to this precise value.

It might be worth testing the fix against the whole contents of lookup_table.

[andrewrk]

#1299 would be a nice solution to this. I think @tiehuis got pretty far on it but didn't quite finish it.

[ehaas]

I looked into this a little bit. The underlying C library on which the Zig code is based performs an out-of-bounds write with the float value in this issue: this line writes to an address before the start of buf

#include <stdbool.h>
#include "errol.h"

int main(void) {
    char buf[100] = {};
    double val = 1.000000e+37;
    int exp = errol3_dtoa(val, buf);
}

Compile with -fsanitize=address or use zig cc.

This specific issue can be fixed by detecting the underflow and handling it, but the code is pretty opaque and I don't know if that would silently introduce incorrect results in some other cases.