Re: [PATCH][C] Fix PR37985

[Richard Guenther <rguenther at suse dot de>]

On Mon, 16 Jan 2012, Richard Guenther wrote:

> On Mon, 16 Jan 2012, Richard Guenther wrote:
> 
> > On Fri, 13 Jan 2012, Joseph S. Myers wrote:
> > 
> > > On Fri, 13 Jan 2012, Richard Guenther wrote:
> > > 
> > > > This fixes PR37985 where since 
> > > > http://gcc.gnu.org/ml/gcc-patches/2006-08/msg01041.html we
> > > > mark conversions produced by convert_to_integer with TREE_NO_WARNING.
> > > > This may shadow "real" stmts with no effects, thus we should
> > > > simply strip them again before checking for TREE_SIDE_EFFECTS.
> > > > 
> > > > Bootstrap & regtest pending on x86_64-unknown-linux-gnu.
> > > > 
> > > > Ok if that passes?
> > > 
> > > OK.
> > 
> > It FAILs gcc.dg/20040202-1.c (we warn about a folded memcpy).  It
> > looks like wrapping things in TREE_NO_WARNING NOP_EXPRs is
> > fragile as soon as you consider multiple warning kinds (thus, of
> > course the patch causing this regression is at fault).
> > 
> > OTOH for the regression warning on folded stmts I really wonder
> > why c_process_expr_stmt folds at all before emitting warnings - why
> > intentionally divert from the source AST here?  I'm testing
> > the following before giving up on this regression (which solves
> > the gcc.dg/20040202-1.c FAIL at least).
> 
> Which works besides now FAILing gcc.dg/strncpy-fix-1.c.  We get
> a "value computed is not used" warning because we are still folding
> the strncpy to a memcpy because we fold all builtin calls at the very
> point they are created (build_function_call_vec):
> 
>   if (name != NULL_TREE
>       && !strncmp (IDENTIFIER_POINTER (name), "__builtin_", 10))
>     {
>       if (require_constant_value)
>         result =
>           fold_build_call_array_initializer_loc (loc, TREE_TYPE (fntype),
>                                                  function, nargs, 
> argarray);
>       else
>         result = fold_build_call_array_loc (loc, TREE_TYPE (fntype),
>                                             function, nargs, argarray);
> 
> "Fixed" by not doing that if !require_constant_value.  We'll fold
> it during the c_fully_fold call anyway (and during gimplification again).
> 
> Re-testing as follows.

Which does not work either - we seem to rely on that early folding
at least for the type-generic math stuff (gcc.dg/c99-tgmath-*.c).

I'm dropping all these patches.  There doesn't seem to be a good
way to make everything well-behaved with the patch causing the
regression installed.

Richard.