Rainer Orth <[email protected]> writes:

> This patch
>
> commit cc12934b555625b130e242eb6199c60b353ab509
> Author: Rainer Orth <[email protected]>
> Date:   Tue Feb 3 20:41:40 2026 +0100
>
>     build: Only use gas_flag/gnu_ld_flag internally [PR123841]
>
> caused testsuite regressions on Linux/x86_64 in one configuration: if
> configured without --with-gnu-as/--with-gnu-ld, taking as and ld from
> PATH, the stage 1 gcc/auto-host.h differs:
>
> @@ -538 +544 @@
> -#define HAVE_AS_IX86_GOT32X 1
> +#define HAVE_AS_IX86_GOT32X 0
> @@ -586 +592 @@
> -#define HAVE_AS_IX86_TLSLDM 1
> +#define HAVE_AS_IX86_TLSLDM 0
> @@ -599 +605 @@
> -#define HAVE_AS_IX86_TLS_GET_ADDR_GOT 1
> +#define HAVE_AS_IX86_TLS_GET_ADDR_GOT 0
> @@ -1795 +1801 @@
> -#define HAVE_LD_DEMANGLE 1
> +/* #undef HAVE_LD_DEMANGLE */
>
> gas_flag and gnu_ld_flag weren't set correctly in this case:
>
> Without --with-gnu-ld, AC_ARG_WITH(gnu-ld) sets gnu_ld_flag to no.  The
> later check to determine gnu_ld_flag from $gcc_cv_ld --version isn't
> even run because it is guarded by test -z "$gnu_ld_flag".
>
> To avoid this, two things need to happen:
>
> * The ld --version check needs to be done in the ACTION_IF_NOT case of
>   AC_ARG_WITH(gnu-ld), not afterwards.
>
> * For this to work, gcc_cv_ld needs to be set at this point.
>
> Therefore this patch does the first.  In order for the second to work,
> AC_ARG_WITH(gnu-ld) is moved later when gcc_cv_ld has already been set.
>
> The same applies to gas_flag, too.
>
> Tested on x86_64-pc-linux-gnu and sparc64-pc-linux-gnu like so:

Belated but LGTM.

I'm happy to look at further patches for these as well
(hopefully mostly in stage1; unfortunately the timing for the previous issue
didn't interact well w/ FOSDEM timing).

Thanks for handling all of this - it's pretty grotty and it took me a
while to follow a bunch of it when I've had to look at it before.

> [...]

sam

Attachment: signature.asc
Description: PGP signature

Reply via email to