tinycc version: 0.9.28rc, commit hash 12acbf3e92e4067155da2e6242cf4cd72b1e291e

The constant folding logic for cast in `tccgen.c:gen_cast` involves undefined behavior. When converting a double to signed integer, the double is first casted to `uint64_t`, then to the desired signed integer type. But if the double is negative, its conversion to unsigned integer is UB.

This bug result in different behavior of `tcc` on x86_64 and arm64. Consider the following program:

#include <stdio.h>
int main() {
  printf("%d\n", (int)-1.0);
}

`tcc` compiled with `gcc` on linux will output `-1` on this program, while `tcc` compiled with `clang` on OSX will output `0`.

The behavior difference boil down to different behavior of x86_64's `cvttsd2si` instruction and arm64's `fcvtzu` instruction. The two behave differently on negative input.