Re: Is there really only one symbol_ref object referring to eachsymbolic label?

[Geoff Keating <geoffk at geoffk dot org>]

> Date: Sat, 28 Jun 2003 12:09:26 -0400 (EDT)
> Cc: gcc@gcc.gnu.org
> From: Kazu Hirata <kazu@cs.umass.edu>

> Hi Geoff,
> 
> > > GCC Internals Manual states that
> > > 
> > >   For any symbolic label, there is only one symbol_ref object
> > >   referring to it.
> > > 
> > > Does this mean x == y iff XSTR (x, 0) == XSTR (y, 0), assuming both x
> > > and y are SYMBOL_REF?  If so, looking at gen_rtx_SYMBOL_REF, I think
> > > we can get two different rtx objects just by saying
> > > 
> > >   x = gen_rtx_SYMBOL_REF (Pmode, "hello");
> > >   y = gen_rtx_SYMBOL_REF (Pmode, strdup ("hello"));
> > > 
> > > Or are we not supposed to do this?
> > 
> > Strings in SYMBOL_REFs are supposed to live in the identifier hash
> > table, and so should be unique...
> 
> I see.  Then I guess it could potentially be dangerous to have
> multiple places creating the same SYMBOL_REFs like
> 
> sh/sh.md:8059:  operands[2] = gen_rtx_SYMBOL_REF (SImode, \"__fpscr_values\");
> sh/sh.md:8073:  operands[2] = gen_rtx_SYMBOL_REF (SImode, \"__fpscr_values\");
> 
> and
> 
> sparc/sparc.md:1937:  operands[2] = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
> sparc/sparc.md:2178:  operands[2] = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_");
> x
> 
> because rtx_equal_p and a *lot* of its variants use
> 
>   XSTR (x, 0) == XSTR (y, 0)
> 
> to compare two SYMBOL_REFs.  If the host compiler assigns two
> identical string constants two different addresses, the comparison may
> fail.

Yup.  They should look like

gen_rtx_SYMBOL_REF (xxx, ggc_strdup ("yyy"))

You don't want to make the ggc_strdup be part of gen_rtx_* because 99%
of the time it's not necessary and the extra hash-table lookup is slow...

Perhaps there should be a check in gen_rtx_fmt_* when debugging is on
so that when it does

  XSTR (rt, 0) = arg0;

it also does:

#if ENABLE_CHECKING_XXX
 if (ggc_strdup (arg0) != arg0)
   abort ();
#endif

-- 
- Geoffrey Keating <geoffk@geoffk.org>