Re: [PATCH] c++: Quash bogus -Wunused-value with new [PR107797]

[Marek Polacek <polacek@redhat.com>]

On Fri, Jan 20, 2023 at 02:37:19PM -0500, Jason Merrill wrote:
> On 1/19/23 21:03, Marek Polacek wrote:
> > We shouldn't emit "right operand of comma operator has no effect"
> > when that comma operator was created by the compiler for "new int{}".
> > convert_to_void/COMPOUND_EXPR already checks warning_suppressed_p so
> > we can just suppress -Wunused-value.
> > 
> > Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
> > 
> > 	PR c++/107797
> > 
> > gcc/cp/ChangeLog:
> > 
> > 	* cvt.cc (ocp_convert): copy_warning when creating a new
> > 	COMPOUND_EXPR.
> > 	* init.cc (build_new_1): Suppress -Wunused-value on
> > 	compiler-generated COMPOUND_EXPRs.
> > 
> > gcc/testsuite/ChangeLog:
> > 
> > 	* g++.dg/warn/Wunused-value-1.C: New test.
> > ---
> >   gcc/cp/cvt.cc                               |  6 ++++--
> >   gcc/cp/init.cc                              |  2 ++
> >   gcc/testsuite/g++.dg/warn/Wunused-value-1.C | 12 ++++++++++++
> >   3 files changed, 18 insertions(+), 2 deletions(-)
> >   create mode 100644 gcc/testsuite/g++.dg/warn/Wunused-value-1.C
> > 
> > diff --git a/gcc/cp/init.cc b/gcc/cp/init.cc
> > index f816c474cef..52e96fbe590 100644
> > --- a/gcc/cp/init.cc
> > +++ b/gcc/cp/init.cc
> > @@ -3800,6 +3800,8 @@ build_new_1 (vec<tree, va_gc> **placement, tree type, tree nelts,
> >     if (cookie_expr)
> >       rval = build2 (COMPOUND_EXPR, TREE_TYPE (rval), cookie_expr, rval);
> > +  suppress_warning (rval, OPT_Wunused_value);
> 
> This makes sense, but IIUC since rval is built with no location, this just
> sets nowarning_flag?

Correct (I didn't realize this when writing the patch).  There's another
suppress_warning on a location-less expr just a few lines above.
 
> >     if (rval == data_addr && TREE_CODE (alloc_expr) == TARGET_EXPR)
> >       /* If we don't have an initializer or a cookie, strip the TARGET_EXPR
> >          and return the call (which doesn't need to be adjusted).  */
> > diff --git a/gcc/cp/cvt.cc b/gcc/cp/cvt.cc
> > index 0cbfd8060cb..17827d06a4a 100644
> > --- a/gcc/cp/cvt.cc
> > +++ b/gcc/cp/cvt.cc
> > @@ -711,8 +711,10 @@ ocp_convert (tree type, tree expr, int convtype, int flags,
> >  	return error_mark_node;
> >        if (e == TREE_OPERAND (expr, 1))
> >  	return expr;
> > -      return build2_loc (EXPR_LOCATION (expr), COMPOUND_EXPR, TREE_TYPE (e),
> > -			 TREE_OPERAND (expr, 0), e);
> > +      e = build2_loc (EXPR_LOCATION (expr), COMPOUND_EXPR, TREE_TYPE (e),
> > +		      TREE_OPERAND (expr, 0), e);
> > +      copy_warning (e, expr);
> 
> And so I don't know what effect this would have; copy_warning doesn't seem
> to propagate nowarning_flag, which seems like a bug.

It actually does propagate nowarning_flag: in copy_warning we have

219   /* The no-warning bit might be set even if the map has not been consulted, or
220      otherwise if there's no entry in the map.  */
221   set_no_warning_bit (to, supp);

so 'e' ends up with TREE_NO_WARNING set.

I would not think we want to build 'rval' in build_new_1 with a location
or use TREE_NO_WARNING directly instead of suppress_warning so I'm leaving
the patch as-is, but please let me know if you disagree.

Marek