Re: [Tinycc-devel] Questions about commit 5b28165: Fix test 131 for 32 bits targets

[grischka]

On 07.07.2023 18:15, Ziyao via Tinycc-devel wrote:
> Hi,
>
> I noticed that test 131 (authored by me) has been modified
> in commit 5b28165: Fix test 131 for 32 bits targets:
> struct with two long int members are replaced by long long
> int ones.
>
> Test 131 (return struct in registers) is intended to test
> whether functions return small structs (maybe passed in
> registers in some ABIs) are handled properly.
>
> long long int on 32bit targets seem to be 8 bytes, right?
> And AFAIK, many ABIs only pass structs with 2 * WORDLENGTH
> size in registers (for 32-bit architectures, it is 8-bytes
> and is equal to the size of a struct with two long int
> members.)
>
> Is this fix necessary?

Hi,

well, problem is simply that with 32-bit longs your data
   0x1234567890abcd
would be truncated to
   0x7890abcd
and the output from %lx would not match the .expect file.

Btw is your fix incorrect.  The code that you "fixed" is about
alignment but there was no alignment problem in your case.

The real problem is below at "vtop->c.i += regsize;" which can
advance the location to load the 2nd register only for local or
global/static memory.

For pointers it would need to emit an 'add r, regsize' instruction
You can see such code in gv():

    gaddrof();
    vpushs(PTR_SIZE);
    gen_op('+');
    vtop->r |= VT_LVAL;

-- gr