Re: [lmi] UBSAN flags

[Greg Chicares]

On 6/6/22 21:57, Vadim Zeitlin wrote:
> On Mon, 6 Jun 2022 18:59:13 +0000 Greg Chicares <gchicares@sbcglobal.net> 
> wrote:
[...]
> GC> Can you reproduce the 'actuarial_table_test' problem?
> 
>  I can't compile this file with -O3 because I get this error:

[...hearty snip...]

>  I'm not sure what's going on here and I'm not sure if this is the
> same as the "illegal memory operation" mentioned above or not.

Yes, that seems the same as the messages I get.

> FWIW, if I
> change -O3 to -O2 for this file compilation, then it compiles fine and the
> resulting test runs without any errors.

Same here.

> But this error definitely looks
> worrisome, so I'll look into it unless you already understand what is it
> about -- if so, please let me know.

I have no idea at all. In the stack traces, I see nothing that
pertains to lmi except for the inclusion of standard headers
and this one line:

inlined from ‘std::vector<double> {anonymous}::table_256(int, int)’ at 
actuarial_table_test.cpp:188:13:

where line 188 is:
      v.insert(v.end(), qult + age - 10, qult + nult);

In table_256(), 'qsel' indeed has 213 = 71 * 3 elements. Similarly,
121 - 13 + 1 equals 109, and 'qult' does have 109 elements. These
arrays can't be overflowing any reasonable implementation limits.

It seems wise to add this assertion:

    LMI_ASSERT(0 <= duration && duration <= 3);
+   LMI_ASSERT(10 <= age && age <= 80);

but it never fires. Because it never fires, this:

      v.insert(v.end(), qult + age - 10, qult + nult);

looks valid, and in particular

  qult + age - 10 // must be a valid "begin"
  qult + nult     // must be a valid (one-past-the-end) "end"

The test passes with the 'safestdlib' build type, which uses
  -O0 -D_GLIBCXX_DEBUG -D_GLIBCXX_DEBUG_PEDANTIC
so the "debug" version of libstdc++ considers this code valid.
It still passes if I change -O0 to -O2, and even if I change
it to -O3.

Dare we consider the possibility that UBSan is malfunctioning?