On Sun, 15 May 2005, Richard Guenther wrote:
> > Exactly which optimization do we miss by changing:
> >
> > /* *&p => p */
> > - if (lang_hooks.types_compatible_p (type, optype))
> > + if (type == optype)
> > return op;
>
> I don't know - maybe stripping sign casts. But if we use equality
> comparison here we can as well use STRIP_TYPE_NOPS instead of
> STRIP_NOPS - but the patch doing so caused some optimization regressions.
You're confusing the types of the pointers vs. the types of what they
point to.
The call to STRIP_NOPS at the top of fold_indirect_ref is stripping
the conversions between pointer types. Both your recent change to
STRIP_TYPE_NOPS (and it's reversion) and your suggestion above about
STRIP_SIGN_NOPS are red herrings.
These calls are responsible for eliminating the casts between
"char*", "const char*", "int*", "unsigned int*", "float*" etc.
In the C front-end, for example, these are all equivalent types,
typically all with mode Pmode and with (irrelevant?) signedness.
The NOPS that you are trying to strip are for the pointed to
type, i.e. that you strip a conversion for T1* to T2*, if T1
and T2 have the same mode, not that the pointers T1* and T2*
have the same mode.
If you really know of a missed optimization, rather than just
your suspicion above about the potential optimization opportunity
for signedness changes, then you can try something more aggressive
such as:
/* *&p => p */
if (TYPE_MODE (type) == TYPE_MODE (optype)
&& TYPE_MODE (type) != BLKmode)
return fold_convert (type, op)
But this is only safe if the operand is known to be on the RHS
of an expression, and not used as an lvalue. Unfortunately, testing
"maybe_lvalue_p (op)" here doesn't help, as op has to have been an
lvalue to be the argument of an ADDR_EXPR. In gimplify.c, you might
be able to get away without the fold_convert, but I would like to
comment on what/when type conversions are useless in tree-ssa.
Hence, if there are any missed optimizations, for my original suggestion
these would have to be caught at gimplification (or in fold_stmt), with
a mode test like the one above, when the context of the indirect_ref
expression is known.
I hope this clears up the STRIP_NOPS confusion.
Roger
--
