On Mon, 2 Dec 2024 at 12:44, Richard Earnshaw (lists)
<[email protected]> wrote:
>
> On 02/12/2024 11:21, Christophe Lyon wrote:
> > If the target does not support floating-point, we register FP vector
> > types as 'void' (see register_vector_type).
> >
> > The leads to warnings about 'pure attribute on function returning
> > void' when we declare the various load intrinsics because their
> > call_properties say CP_READ_MEMORY (thus giving them the 'pure'
> > attribute), but their return type is void.
> >
> > To avoid such warnings, pretend the call_properties are empty when FP
> > is disabled and the function would return an FP value. If such
> > functions are incorrectly used in user code, a proper error is
> > emitted:
> > unknown type name ‘float16x8_t'; did you mean ‘int16x8_t’?
> >
> > gcc/ChangeLog:
> >
> > PR target/117814
> > * config/arm/arm-mve-builtins-base.cc (vld1_impl): Fix
> > call_properties.
> > (vld24_impl): Likewise.
> > * config/arm/arm-mve-builtins-functions.h (load_extending):
> > Likewise.
>
> Won't this lead to problems if the code is something like
>
> #include "arm_mve.h"
>
> #pragma gcc target ("arch=armv8.1-m.main+mve.fp")
>
> // Some use of an affected intrinsic
>
> and then compile with "-march=armv8.1-m.main+mve -mfpu=auto
> -mfloat-abi=softfp"?
>
Indeed.... actually the real cause is precisely that without FPU we do
not register FP types (and map them to "void" instead).
I had started to look at fixing that but it looked more invasive.
Actually if you take any intrinsic test (say vaddq_f16.c), remove
"_fp" from dg-require-effective-target and dg-add-options, and add the
pragma above, you'll get the same error:
vaddq_f16.c:20:1: error: unknown type name 'float16x8_t'; did you mean
'int16x8_t'?
Let me look at fixing this.
Thanks,
Christophe
> R.
>
> > ---
> > gcc/config/arm/arm-mve-builtins-base.cc | 22 +++++++++++++++++++--
> > gcc/config/arm/arm-mve-builtins-functions.h | 11 ++++++++++-
> > 2 files changed, 30 insertions(+), 3 deletions(-)
> >
> > diff --git a/gcc/config/arm/arm-mve-builtins-base.cc
> > b/gcc/config/arm/arm-mve-builtins-base.cc
> > index 723004b53d7..a322730eca8 100644
> > --- a/gcc/config/arm/arm-mve-builtins-base.cc
> > +++ b/gcc/config/arm/arm-mve-builtins-base.cc
> > @@ -141,8 +141,17 @@ class vld1_impl : public full_width_access
> > {
> > public:
> > unsigned int
> > - call_properties (const function_instance &) const override
> > + call_properties (const function_instance &instance) const override
> > {
> > + /* If the target does not support floating-point, we register FP vector
> > + types as 'void'. In this case, pretend we do not access memory to
> > avoid
> > + warnings about 'pure attribute on function returning void' when we
> > + declare the intrinsics. Such uses in user code are properly
> > + diagnosed. */
> > + if (!TARGET_HAVE_MVE_FLOAT
> > + && instance.type_suffix (0).float_p)
> > + return 0;
> > +
> > return CP_READ_MEMORY;
> > }
> >
> > @@ -1141,8 +1150,17 @@ public:
> > using full_width_access::full_width_access;
> >
> > unsigned int
> > - call_properties (const function_instance &) const override
> > + call_properties (const function_instance &instance) const override
> > {
> > + /* If the target does not support floating-point, we register FP vector
> > + types as 'void'. In this case, pretend we do not access memory to
> > avoid
> > + warnings about 'pure attribute on function returning void' when we
> > + declare the intrinsics. Such uses in user code are properly
> > + diagnosed. */
> > + if (!TARGET_HAVE_MVE_FLOAT
> > + && instance.type_suffix (0).float_p)
> > + return 0;
> > +
> > return CP_READ_MEMORY;
> > }
> >
> > diff --git a/gcc/config/arm/arm-mve-builtins-functions.h
> > b/gcc/config/arm/arm-mve-builtins-functions.h
> > index 0ade2157e4a..1a9a347805c 100644
> > --- a/gcc/config/arm/arm-mve-builtins-functions.h
> > +++ b/gcc/config/arm/arm-mve-builtins-functions.h
> > @@ -986,8 +986,17 @@ public:
> > m_float_memory_type (NUM_TYPE_SUFFIXES)
> > {}
> >
> > - unsigned int call_properties (const function_instance &) const override
> > + unsigned int call_properties (const function_instance &instance) const
> > override
> > {
> > + /* If the target does not support floating-point, we register FP vector
> > + types as 'void'. In this case, pretend we do not access memory to
> > avoid
> > + warnings about 'pure attribute on function returning void' when we
> > + declare the intrinsics. Such uses in user code are properly
> > + diagnosed. */
> > + if (!TARGET_HAVE_MVE_FLOAT
> > + && instance.type_suffix (0).float_p)
> > + return 0;
> > +
> > return CP_READ_MEMORY;
> > }
> >
>
