Re: [PATCH v4 2/2] x86: Add general_regs_only function attribute

["H.J. Lu" <hjl.tools@gmail.com>]

On Thu, Apr 22, 2021 at 4:23 AM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On Thu, Apr 22, 2021 at 12:30 PM Jakub Jelinek via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > On Wed, Apr 21, 2021 at 06:01:07PM -0700, H.J. Lu via Gcc-patches wrote:
> > > How about this?
> > >
> > > @item general_regs_only
> > > @cindex @code{general_regs_only} function attribute, x86
> > > The @code{general_regs_only} function attribute informs the compiler
> > > that the function uses only general purpose registers.  When the
> > > compiler inlines a function with the @code{always_inline} attribute,
> > > target-specific compilation options may lead to inline failures.
> > > The @code{general_regs_only} attribute, if applicable, can be used
> > > together with the @code{always_inline} attribute to reduce inlining
> > > failure.
> >
> > I don't really like this attribute.
> > It is very specific to what you want to solve and doesn't address the
> > general problem, that always_inline means different things in different
> > code, and that is a problem for many targets, not just one.
> >
> > As has been written, in some cases it means inline always, error
> > whenever it is called indirectly which can't be optimized into a direct
> > call that can be inlined and error whenever the inlining fails for other
> > reasons.
> >
> > Another case, e.g. in the glibc fortify wrappers, is inline always
> > when the call is direct or an indirect call can be optimized into a direct
> > call and error when the inlining fails, but support indirect calls without
> > errors.
> >
> > Another case, which is most of the x86/aarch64/arm etc. intrinsics, is
> > inline always unless there is a target mismatch (roughly what is
> > actually implemented).
> >
> > Because from the always_inline attribute it is impossible to determine which
> > one of those it is (for the indirect calls the rule could be
> > gnu_inline extern inline means indirect calls are ok, anything else means
> > indirect calls are bad), we need new syntax to distinguish those cases.
> >
> > general_regs_only attribute doesn't seem to be it, e.g. for the glibc
> > fortify wrappers cases I don't see why we should forbid using floating point
> > in such inlines.
> >
> > So IMHO we need some new attribute for one of those, or optional parameter
> > to always_inline.
> >
> > For the intrinsic case, ideal would be if we could record which ISA flags
> > (or more generally which options) are required and which are not.  Either
> > have some syntax where those would be explicitly specified in attribute (but
> > frankly that would be a maintainance nightmare), or derive those from
> > surrounding pragmas.  Right now we have those wrapped in
> > #ifndef __AVX2__
> > #pragma GCC push_options
> > #pragma GCC target("avx2")
> > #define __DISABLE_AVX2__
> > #endif /* __AVX2__ */
> >
> > ...
> >
> > #ifdef __DISABLE_AVX2__
> > #undef __DISABLE_AVX2__
> > #pragma GCC pop_options
> > #endif /* __DISABLE_AVX2__ */
> >
> > The question is if the pragma GCC target right now behaves incrementally
> > or not, whether
> > #pragma GCC target("avx2")
> > adds -mavx2 to options if it was missing before and nothing otherwise, or if
> > it switches other options off.  If it is incremental, we could e.g. try to
> > use the second least significant bit of global_options_set.x_* to mean
> > this option has been set explicitly by some surrounding #pragma GCC target.
> > The normal tests - global_options_set.x_flag_whatever could still work
> > fine because they wouldn't care if the option was explicit from anywhere
> > (command line or GCC target or target attribute) and just & 2 would mean
> > it was explicit from pragma GCC target; though there is the case of
> > bitfields... And then the inlining decision could check the & 2 flags to
> > see what is required and what is just from command line.
> > Or we can have some other pragma GCC that would be like target but would
> > have flags that are explicit (and could e.g. be more restricted, to ISA
> > options only, and let those use in addition to #pragma GCC target.
>
> I'm still curious as to what you think will break if always-inline does what
> it is documented to do.

No wrong code.  But the compiler will generate a different error message
at the later stage if the ISA for the intrinsic isn't enabled.

-- 
H.J.