Bug #7371: Fix undefined overflow checking in bigdecimal

• Status: Closed
• Assignee: mrkn (Kenta Murata)
• Target version: -
• ruby -v: 1.9.x

[xi (Xi Wang)]

In AddExponent() at ext/bigdecimal/bigdecimal.c:3677, the overflow checks rely on signed integer overflow, which is undefined behavior in C.

SIGNED_VALUE m = e+n;
SIGNED_VALUE eb, mb;
if(e>0) {
    if(n>0) {
        mb = m*(SIGNED_VALUE)BASE_FIG;
        eb = e*(SIGNED_VALUE)BASE_FIG;
        if(mb<eb) goto overflow;
    }

Some compilers (e.g., gcc 4.8) will optimize away such overflow checks due to undefined behavior. Ruby currently uses "-fno-strict-overflow" to disable such offending optimizations in gcc, but this workaround option is not supported by other compilers, thus not portable.

The attached patch uses unsigned multiplication for overflow checking, which is well defined in C.

[mrkn (Kenta Murata)]

• Category set to ext
• Assignee set to mrkn (Kenta Murata)

[usa (Usaku NAKAMURA)]

• Status changed from Open to Assigned

[xi (Xi Wang)]

To see how it works, try to compile the following (simplified) code with gcc 4.8. The entire function will be optimized away with "gcc -O2" (just grep "bar" in the resulting assembly code); gcc 4.7 or earlier doesn't do that.

#define SIGNED_VALUE long
#define BASE_FIG 9

void bar(void);

static void AddExponent(SIGNED_VALUE e, SIGNED_VALUE n)
{
SIGNED_VALUE m = e+n;
SIGNED_VALUE eb, mb;
if(e>0) {
if(n>0) {
mb = m*(SIGNED_VALUE)BASE_FIG;
eb = e*(SIGNED_VALUE)BASE_FIG;
if(mb<eb) goto overflow;
}
}
return;
overflow:
bar();
}

void foo(SIGNED_VALUE e)
{
AddExponent(e, 1);
}

[mrkn (Kenta Murata)]

• Status changed from Assigned to Closed

I think this issue had been fixed by akr in r40214.