On Wed, Jun 20, 2018 at 10:42:38AM -0400, Michael Meissner wrote:
> This patch prevents the special overriding of the complex float128
> multiply/divide functions from being run twice if there are clone or target
> attributes. I wasn't aware that the hook used to initialize the built-in
> functions is run each time you change the target options. The built-in
> function handling aborts if the name had already been set.
>
> I have tested this on a little endian power8 sysytem using two builds with
> long
> double set to IEEE and IBM 128-bit. This patch fixes testsuite errors from
> using the clone or target attributes. Can I install it in the trunk and GCC
> 8.x branches?
Next time, please mention what tests are fixed.
> --- gcc/config/rs6000/rs6000.c (revision 261574)
> +++ gcc/config/rs6000/rs6000.c (working copy)
> @@ -17892,9 +17892,14 @@ init_float128_ieee (machine_mode mode)
> {
> if (FLOAT128_VECTOR_P (mode))
> {
> - /* Set up to call __mulkc3 and __divkc3 under -mabi=ieeelongdouble. */
> - if (mode == TFmode && TARGET_IEEEQUAD)
> + static bool complex_muldiv_init_p = false;
> +
> + /* Set up to call __mulkc3 and __divkc3 under -mabi=ieeelongdouble. If
> + we have clone or target attributes, this will be called a second
> + time. We want to create the built-in function only once. */
> + if (mode == TFmode && TARGET_IEEEQUAD && !complex_muldiv_init_p)
> {
> + complex_muldiv_init_p = true;
> built_in_function fncode_mul =
> (built_in_function) (BUILT_IN_COMPLEX_MUL_MIN + TCmode
> - MIN_MODE_COMPLEX_FLOAT);
The indentation here is wrong (was before, too).
Is there no more elegant, less error-prone way to see things are already
initialised? Maybe use maybe_get_identifier?
But, okay for trunk.
Segher
