[PATCH] Fix ix86_split_long_move collision handling with TLS (PR target/66470)

[Uros Bizjak]

On Tue, Jun 9, 2015 at 3:39 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Tue, Jun 09, 2015 at 03:21:55PM +0200, Uros Bizjak wrote:
>> I'm afraid that simple scan loop won't work correctly on x32. There
>> are some issues with UNSPEC_TP for this target, so we have to generate
>> zero_extend of SImode UNSPEC, e.g.:
>>
>> (plus:DI (zero_extend:DI (unspec:SI [...] UNSPEC_TP) (reg:DI ...))
>>
>> as can be seen in get_thread_pointer to construct the address. It
>> looks that your loop won't find the UNSPEC_TP tag in the above case.
>
> You're right, for -m32 it would need to start with
>    rtx *x = &addr;
> +  while (GET_CODE (*x) == ZERO_EXTEND
> +        || GET_CODE (*x) == AND
> +        || GET_CODE (*x) == SUBREG)
> +    x = &XEXP (*x, 0);
> to get at the PLUS.  Now, with either the original patch with the above
> ammendment, or with the iterators, the question is what to do
> with the UNSPEC_TP for -m32.  Is it ok to just use addr32 on the
> lea and use normal Pmode (== DImode) addressing for the memory reads
> (if I read the code well, that is what it does right now for the non-TLS
> ones:
>           if (GET_MODE (base) != Pmode)
>             base = gen_rtx_REG (Pmode, REGNO (base));
> )?  Then we'd need to change tls_base mode to Pmode.

(I assume you referred to -mx32 above. Please also note that -mx32
uses Pmode == SImode by default).

Yes, please emit zero-extended address load to a DImode register, and
use this register with UNSPEC_TP to form final DImode address. The LEA
that will be generated from address load will use DImode inputs due to
%H operand modifier and will output to (implicitly zero-extended)
SImode register.

FYI, you just hit two special cases here:
- addresses that mention %fs and %gs are always in DImode
- LEA input operands are also always in DImode, but compiler already
takes care of it.

Uros.