make test fails at 6.2 Numbers

[mbamac]

on macOS Sequoia, M1 chip:

2 failures (0.9%).
FAIL: (/ -4611686018427387904 -1)
    expected 4611686018427387904 but got -4611686018427387904
    on line 910 of file "tests/r7rs-tests.scm"
FAIL: (quotient -4611686018427387904 -1)
    expected 4611686018427387904 but got -4611686018427387904
    on line 911 of file "tests/r7rs-tests.scm"

[SwitchAxe]

Disabling compiler optimizations with make XCFLAGS="-O0" fixes it for me on a M3 macbook air

[ashinn]

I can't reproduce this on Sonoma with an M1 chip.

Note these two cases are testing division of the most negative fixnum by -1, resulting in the smallest positive number which needs to be represented as a bignum. The relevant code for this case is in SEXP_OP_DIV in vm.c:

if ((sexp_sint_t)tmp1 < 0 && (sexp_sint_t)tmp2 < 0 && (sexp_sint_t)_ARG1 < 0) {
    _ARG1 = sexp_quotient(ctx, tmp1=sexp_fixnum_to_bignum(ctx, tmp1), tmp2);
}

i.e. in the (/ fixnum fixnum) path we check if both inputs are negative and the result is also negative, we recompute by converting the first input to a bignum. I can't debug this if I can't reproduce though...

[SwitchAxe]

I fear some undefined behavior may be the culprit... if i try with the apple-provided gcc (the one based on clang 16) the tests fail, but with homebrew gcc (i.e. upstream gcc 14.2.0 i suppose) all tests pass successfully, so @ashinn you may be able to reproduce this by using clang 16 (or its gcc frontend) as the compiler