Re: [PATCH] ipa: Fix throw in multi-versioned functions [PR106627]

["Simon Rainer" <gcc.gnu@vvalter.com>]

Hi,

Thanks for taking a look at my patch. I tested some combinations with pure/noreturn attributes. gcc seems to ignore those attributes on multiversion functions and generates sub-optimal assembly.
But I wasn't able to fix this by simply copying members like DECL_PURE_P. It's pretty hard for me to tell which members of tree are relevant for a function declaration and should be copied and which should not be copied.

Anyway, I think the TREE_NOTHROW change is the most important one, because it leads to correctness problems (and is what broke my original program :D ), so could you please commit my patch as I don't have write-access myself.

Should I open a new bug on bugzilla for the pure/noreturn issue?

Thanks
Simon Rainer


On Thu, Sep 1, 2022, at 08:37, Richard Biener wrote:
> On Wed, Aug 31, 2022 at 11:00 PM Simon Rainer <gcc.gnu@vvalter.com> wrote:
> >
> > Hi,
> >
> > This patch fixes PR106627. I ran the i386.exp tests on my x86_64-linux-gnu machine with a fully bootstrapped checkout. I also tested manually that no exception handling code is generated if none of the function versions throws an exception.
> > I don't have access to a machine to test the change to  rs6000.cc, but the code seems like an exact copy and I don't see a reason why it shouldn't work there the same way.
> >
> > Regards
> > Simon Rainer
> >
> > From 6fcb1c742fa1d61048f7d63243225a8d1931af4a Mon Sep 17 00:00:00 2001
> > From: Simon Rainer <gcc.gnu@vvalter.com>
> > Date: Wed, 31 Aug 2022 20:56:04 +0200
> > Subject: [PATCH] ipa: Fix throw in multi-versioned functions [PR106627]
> >
> > Any multi-versioned function was implicitly declared as noexcept, which
> > leads to an abort if an exception is thrown inside the function.
> > The reason for this is that the function declaration is replaced by a
> > newly created dispatcher declaration, which has TREE_NOTHROW always set
> > to 1. Instead we need to set TREE_NOTHROW to the value of the original
> > declaration.
> 
> Looks quite obvious.  The middle-end to target interface is a bit iffy
> since we have
> to duplicate this everywhere.  There's also other flags like
> pure/const and noreturn
> that do not impose correctness issues but may cause irritations if the IL gets
> a call to the dispatcher not marked noreturn but there's no code following.
> 
> That said, the fix looks good to me.
> 
> Thanks,
> Richard.
> 
> >         PR ipa/106627
> >
> > gcc/ChangeLog:
> >
> >         * config/i386/i386-features.cc (ix86_get_function_versions_dispatcher): Set TREE_NOTHROW
> >                 correctly for dispatcher declaration
> >         * config/rs6000/rs6000.cc (rs6000_get_function_versions_dispatcher): Likewise
> >
> > gcc/testsuite/ChangeLog:
> >
> >         * g++.target/i386/pr106627.C: New test.
> > ---
> >  gcc/config/i386/i386-features.cc         |  1 +
> >  gcc/config/rs6000/rs6000.cc              |  1 +
> >  gcc/testsuite/g++.target/i386/pr106627.C | 30 ++++++++++++++++++++++++
> >  3 files changed, 32 insertions(+)
> >  create mode 100644 gcc/testsuite/g++.target/i386/pr106627.C
> >
> > diff --git a/gcc/config/i386/i386-features.cc b/gcc/config/i386/i386-features.cc
> > index d6bb66cbe01..5b3b1aeff28 100644
> > --- a/gcc/config/i386/i386-features.cc
> > +++ b/gcc/config/i386/i386-features.cc
> > @@ -3268,6 +3268,7 @@ ix86_get_function_versions_dispatcher (void *decl)
> >
> >        /* Right now, the dispatching is done via ifunc.  */
> >        dispatch_decl = make_dispatcher_decl (default_node->decl);
> > +      TREE_NOTHROW(dispatch_decl) = TREE_NOTHROW(fn);
> >
> >        dispatcher_node = cgraph_node::get_create (dispatch_decl);
> >        gcc_assert (dispatcher_node != NULL);
> > diff --git a/gcc/config/rs6000/rs6000.cc b/gcc/config/rs6000/rs6000.cc
> > index 2f3146e56f8..9280da8a5c8 100644
> > --- a/gcc/config/rs6000/rs6000.cc
> > +++ b/gcc/config/rs6000/rs6000.cc
> > @@ -24861,6 +24861,7 @@ rs6000_get_function_versions_dispatcher (void *decl)
> >
> >        /* Right now, the dispatching is done via ifunc.  */
> >        dispatch_decl = make_dispatcher_decl (default_node->decl);
> > +      TREE_NOTHROW(dispatch_decl) = TREE_NOTHROW(fn);
> >
> >        dispatcher_node = cgraph_node::get_create (dispatch_decl);
> >        gcc_assert (dispatcher_node != NULL);
> > diff --git a/gcc/testsuite/g++.target/i386/pr106627.C b/gcc/testsuite/g++.target/i386/pr106627.C
> > new file mode 100644
> > index 00000000000..a67f5ae4813
> > --- /dev/null
> > +++ b/gcc/testsuite/g++.target/i386/pr106627.C
> > @@ -0,0 +1,30 @@
> > +/* PR c++/103012 Exception handling with multiversioned functions */
> > +/* { dg-do run } */
> > +/* { dg-require-ifunc "" }  */
> > +
> > +#include <assert.h>
> > +
> > +__attribute__((target("default")))
> > +void f() {
> > +    throw 1;
> > +}
> > +
> > +__attribute__((target("sse4.2,bmi")))
> > +void f() {
> > +    throw 2;
> > +}
> > +
> > +int main()
> > +{
> > +    try {
> > +        f();
> > +    }
> > +    catch(...)
> > +    {
> > +        return 0;
> > +    }
> > +
> > +    assert (false);
> > +    return 1;
> > +}
> > +
> > --
> > 2.34.1
> >
>