On Mon, 2024-10-21 at 19:59 -0700, Richard Henderson wrote:
> On 10/21/24 07:27, Ilya Leoshkevich wrote:
> > [1] deprecated -mpower8-vector, resulting in:
> >
> > powerpc64-linux-gnu-gcc: warning: switch '-mpower8-vector' is
> > no longer supported
> > qemu/tests/tcg/ppc64/vsx_f2i_nan.c:4:15: error: expected ';'
> > before 'float'
> > 4 | typedef vector float vsx_float32_vec_t;
> > | ^~~~~~
> >
> > Similar to how this was done for the GCC testcases, replace
> > -mpower8-vector with -mvsx.
> >
> > [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109987
> >
> > Signed-off-by: Ilya Leoshkevich <i...@linux.ibm.com>
> > ---
> > tests/tcg/ppc64/Makefile.target | 10 +++++-----
> > 1 file changed, 5 insertions(+), 5 deletions(-)
> >
> > diff --git a/tests/tcg/ppc64/Makefile.target
> > b/tests/tcg/ppc64/Makefile.target
> > index 1940886c737..d1b00d2bf09 100644
> > --- a/tests/tcg/ppc64/Makefile.target
> > +++ b/tests/tcg/ppc64/Makefile.target
> > @@ -6,7 +6,7 @@ VPATH += $(SRC_PATH)/tests/tcg/ppc64
> >
> > config-cc.mak: Makefile
> > $(quiet-@)( \
> > - $(call cc-option,-mpower8-vector,
> > CROSS_CC_HAS_POWER8_VECTOR); \
> > + $(call cc-option,-mvsx, CROSS_CC_HAS_VSX);
> > \
> > $(call cc-option,-mpower10,
> > CROSS_CC_HAS_POWER10)) 3> config-cc.mak
>
> I don't think this is quite right.
> I think you need -mpower8 to get OPTION_MASK_P8_VECTOR set.
Do you mean -mcpu=power8? -mpower8 is a GAS option.
>
> > +ifneq ($(CROSS_CC_HAS_VSX),)
> > PPC64_TESTS=bcdsub non_signalling_xscv
>
> bcdsub is not in the base VSX instruction set, for instance.
>
>
> r~
I was confused why -mvsx works in practice, so I spent some time and
managed to create a "hostile" gcc build, in which this is now a
problem, using:
./configure --target=powerpc64-linux-gnu --enable-checking \
--disable-bootstrap \
--with-as="$(which powerpc64-linux-gnu-as)"
The issue is masked by two things:
- GCC passes "-many" to GAS. GCC commit e154242724b0 ("[RS6000] Don't
pass -many to the assembler") stops --enable-checking builds from
doing this.
- binutils has the following line:
/* The minimum supported cpu for 64-bit little-endian is power8. */
ppc_cpu |= ppc_parse_cpu (ppc_cpu, &sticky, "power8");
so one needs to use the big-endian toolchain to see the problem.
I'll send a v2.
