Failure to match signed sub overflow test idiom

Checking for signed integer overflow after an addItion seems to work fine, but this subtraction test gives bad code on multiple architectures (every one I tried, both 32-bit and 64-bit):

bool idiomatic_check_sub(int x, int y) {
    int64_t result = int64_t(x) - int64_t(y);
    return !(INT_MIN <= result && result <= INT_MAX);
}

bool expected_check_sub(int x, int y) {
    int result;
    return __builtin_ssub_overflow(x, y, &result);
}

https://godbolt.org/z/a91sjPzGz

I haven't checked but this issue may generalise across overflow tests for multiple destination data types. And presumably there's another pattern which people would use when there's no larger temporary type, but I have tested any of those yet.

Start by reproducing the signed-subtraction examples from the issue using the linked Compiler Explorer case, comparing generated code across the reported 32-bit and 64-bit architectures. Trace how the idiomatic check differs from __builtin_ssub_overflow and determine the scope of the problem across destination types. Done means the behavior and a reproducible regression case are clearly established.