[PATCH] x86: Compile CPUID functions with -mgeneral-regs-only

[H.J. Lu]

On Thu, Jun 24, 2021 at 5:47 AM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On Thu, Jun 24, 2021 at 2:42 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > On Thu, Jun 24, 2021 at 5:35 AM Richard Biener
> > <richard.guenther@gmail.com> wrote:
> > >
> > > On Thu, Jun 24, 2021 at 2:13 PM H.J. Lu via Gcc-patches
> > > <gcc-patches@gcc.gnu.org> wrote:
> > > >
> > > > CPUID functions are used to detect CPU features.  If vector ISAs
> > > > are enabled, compiler is free to use them in these functions.  Add
> > > > __attribute__ ((target("general-regs-only"))) to CPUID functions
> > > > to avoid vector instructions.
> > >
> > > But there are GPR instructions not in x86_64, so shouldn't
> > > we use target("march=x86_64") or so?  Note doing either will
> > > of course prevent inlining of those "inlines".
> >
> > Does -march=x86_64, which enables CMOV and other GPR
> > ISAs,  work for -m32?
>
> I don't think so.  I'm also not sure whether -march=xyz in a
> target attribute overrides -mavx512f on the command-line ;)
>
> > > So I'm not sure how much of a fix this is ... the error will almost
> > > always be visible in the caller as well.
> >
> > I think _attribute__ ((target("general-regs-only"))) is a step
> > forward.
>
> That I agree to, but then the cpuid code is likely written the
> way it is to allow inlining.  But code using CPUID should best compile
> functions under the check with additional target attribute
> (or in a separate TU) rather than compiling everything with
> extra -mXYZ and trying to "disable" things in the dispatching
> code (and the code leading to it!).

CPUID checks in GCC tests should be compiled with noinline, ...
plus minimum ISAs allowed.

> Richard.
>
> > > > gcc/
> > > >
> > > >         PR target/101185
> > > >         * config/i386/cpuid.h (__get_cpuid_max): Add
> > > >         __attribute__ ((target("general-regs-only"))).
> > > >         (__get_cpuid): Likewise.
> > > >         (__get_cpuid_count): Likewise.
> > > >         (__cpuidex): Likewise.
> > > >
> > > > gcc/testsuite/
> > > >
> > > >         PR target/101185
> > > >         * gcc.target/i386/avx512-check.h (check_osxsave): Add
> > > >         __attribute__ ((target("general-regs-only"))).
> > > >         (main): Likewise.
> > > > ---
> > > >  gcc/config/i386/cpuid.h                      | 4 ++++
> > > >  gcc/testsuite/gcc.target/i386/avx512-check.h | 2 ++
> > > >  2 files changed, 6 insertions(+)
> > > >
> > > > diff --git a/gcc/config/i386/cpuid.h b/gcc/config/i386/cpuid.h
> > > > index aebc17c6827..74881ee91e5 100644
> > > > --- a/gcc/config/i386/cpuid.h
> > > > +++ b/gcc/config/i386/cpuid.h
> > > > @@ -243,6 +243,7 @@
> > > >     pointer is non-null, then first four bytes of the signature
> > > >     (as found in ebx register) are returned in location pointed by sig.  */
> > > >
> > > > +__attribute__ ((target("general-regs-only")))
> > > >  static __inline unsigned int
> > > >  __get_cpuid_max (unsigned int __ext, unsigned int *__sig)
> > > >  {
> > > > @@ -298,6 +299,7 @@ __get_cpuid_max (unsigned int __ext, unsigned int *__sig)
> > > >     supported and returns 1 for valid cpuid information or 0 for
> > > >     unsupported cpuid leaf.  All pointers are required to be non-null.  */
> > > >
> > > > +__attribute__ ((target("general-regs-only")))
> > > >  static __inline int
> > > >  __get_cpuid (unsigned int __leaf,
> > > >              unsigned int *__eax, unsigned int *__ebx,
> > > > @@ -315,6 +317,7 @@ __get_cpuid (unsigned int __leaf,
> > > >
> > > >  /* Same as above, but sub-leaf can be specified.  */
> > > >
> > > > +__attribute__ ((target("general-regs-only")))
> > > >  static __inline int
> > > >  __get_cpuid_count (unsigned int __leaf, unsigned int __subleaf,
> > > >                    unsigned int *__eax, unsigned int *__ebx,
> > > > @@ -330,6 +333,7 @@ __get_cpuid_count (unsigned int __leaf, unsigned int __subleaf,
> > > >    return 1;
> > > >  }
> > > >
> > > > +__attribute__ ((target("general-regs-only")))
> > > >  static __inline void
> > > >  __cpuidex (int __cpuid_info[4], int __leaf, int __subleaf)
> > > >  {
> > > > diff --git a/gcc/testsuite/gcc.target/i386/avx512-check.h b/gcc/testsuite/gcc.target/i386/avx512-check.h
> > > > index 0a377dba1d5..406faf8fe03 100644
> > > > --- a/gcc/testsuite/gcc.target/i386/avx512-check.h
> > > > +++ b/gcc/testsuite/gcc.target/i386/avx512-check.h
> > > > @@ -25,6 +25,7 @@ do_test (void)
> > > >  }
> > > >  #endif
> > > >
> > > > +__attribute__ ((target("general-regs-only")))
> > > >  static int
> > > >  check_osxsave (void)
> > > >  {
> > > > @@ -34,6 +35,7 @@ check_osxsave (void)
> > > >    return (ecx & bit_OSXSAVE) != 0;
> > > >  }
> > > >
> > > > +__attribute__ ((target("general-regs-only")))
> > > >  int
> > > >  main ()
> > > >  {
> > > > --
> > > > 2.31.1
> > > >
> >
> >
> >
> > --
> > H.J.



-- 
H.J.