Re: [Tinycc-devel] Wrong __int64 computation

[grischka]

Christian Jullien wrote:
> In fact it was related to unsigned __int64 not signed __int64
>
> The following patch modeled after TOK_DIV in x86_64-gen fixes, at least for
> my case, this issue.

Yes, almost.  The signed div does sign-extend rax into rdx
   4:   48 99                   cqto
   6:   48 f7 f9                idiv   %rcx
while the unsigned case just clears rdx, like this:
   9:   48 31 d2                xor    %rdx,%rdx
   c:   48 f7 f1                div    %rcx

You'd notice the difference only with _very_ big numbers,
such as 0xBFFFFFFFFFFFFFFF / 16, though.

I hope this works: http://repo.or.cz/w/tinycc.git/commitdiff/a98caba0

--- grischka

> Could you please review the patch?
>
>         } else {
>             if (op == TOK_UDIV || op == TOK_UMOD) {
> #if     1
>                 if ((vtop->type.t & VT_BTYPE) & VT_LLONG) {
>                     o(0x9948); /* cqto */
>                     o(0x48 + REX_BASE(fr));
>                 } else {
>                     o(0x99); /* cltd */
>                 }
>                 o(0xf7); /* idiv fr, %eax */
>                 o(0xf0 + fr);
> #else
>                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
>                 o(0xf0 + fr);
> #endif
>
> Christian