[PATCH] don't presume undelegitimized UNSPEC_TLS SYMBOL_REF is a decl

[Roland McGrath]

ping?

On Wed, Jun 6, 2012 at 2:04 PM, Roland McGrath <mcgrathr@google.com> wrote:
> cf this change:
>
>        2010-11-19  Jakub Jelinek  <jakub@redhat.com>
>
>                PR target/45870
>                * dwarf2out.c (const_ok_for_output_1): Don't complain about
>                non-delegitimized TLS UNSPECs.
>
> This case hit me where the rtx was:
>
>        (unspec:SI [
>                (symbol_ref:SI ("*.LANCHOR0") [flags 0x1aa])
>                (const_int 4 [0x4])
>            ] UNSPEC_TLS)
>
> Note:
> 1. The UNSPEC has two operands, not one.
> 2. The SYMBOL_REF does not correspond to any decl.
>
> This corresponds to this ARM code:
>
>                ldr     r3, .L10+4
>        ...
>        .L10:
>                <something else>
>                .word   .LANCHOR0(tpoff)
>        ...
>                .section        .tdata,"awT",%progbits
>                .align  4
>        .LANCHOR0 = . + 0
>                .type   tdata1, %object
>                .size   tdata1, 4
>        tdata1:
>                .word   1
>
> The only way I know to reproduce this is using a variant ARM target that
> I'm still developing and is not yet ready to be submitted, so I don't have
> a proper test case to offer.  But I think the principle of the following
> change is fairly sound.
>
> What do you think?  (Recall that I am not a GCC committer, so if you like
> the change, please commit it for me.)
>
>
> Thanks,
> Roland
>
>
> 2012-06-06  Roland McGrath  <mcgrathr@google.com>
>
>        * dwarf2out.c (const_ok_for_output_1): Detect a TLS UNSPEC using
>        SYMBOL_REF_TLS_MODEL rather than DECL_THREAD_LOCAL_P, in case it's
>        not a VAR_DECL.  Also don't limit it to UNSPECs with exactly one
>        operand.
>
> diff --git a/gcc/dwarf2out.c b/gcc/dwarf2out.c
> index 6e4ab76..bc68205 100644
> --- a/gcc/dwarf2out.c
> +++ b/gcc/dwarf2out.c
> @@ -10129,12 +10129,12 @@ const_ok_for_output_1 (rtx *rtlp, void *data ATTRIBUTE_UNUSED)
>         we can't express it in the debug info.  */
>  #ifdef ENABLE_CHECKING
>       /* Don't complain about TLS UNSPECs, those are just too hard to
> -        delegitimize.  */
> -      if (XVECLEN (rtl, 0) != 1
> +        delegitimize.  Note this could be a non-decl SYMBOL_REF such as
> +        one in a constant pool entry, so testing SYMBOL_REF_TLS_MODEL
> +        rather than DECL_THREAD_LOCAL_P is not just an optimization.  */
> +      if (XVECLEN (rtl, 0) == 0
>          || GET_CODE (XVECEXP (rtl, 0, 0)) != SYMBOL_REF
> -         || SYMBOL_REF_DECL (XVECEXP (rtl, 0, 0)) == NULL
> -         || TREE_CODE (SYMBOL_REF_DECL (XVECEXP (rtl, 0, 0))) != VAR_DECL
> -         || !DECL_THREAD_LOCAL_P (SYMBOL_REF_DECL (XVECEXP (rtl, 0, 0))))
> +         || SYMBOL_REF_TLS_MODEL (XVECEXP (rtl, 0, 0)) == TLS_MODEL_NONE)
>        inform (current_function_decl
>                ? DECL_SOURCE_LOCATION (current_function_decl)
>                : UNKNOWN_LOCATION,