This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH, rs6000] Do not do gimple-folding of expressions that are missing their LHS
On Wed, Jul 12, 2017 at 11:24 PM, Segher Boessenkool
<segher@kernel.crashing.org> wrote:
> On Wed, Jul 12, 2017 at 01:34:01PM -0500, Bill Schmidt wrote:
>> Although I said this was invalid code, it really isn't -- it's legal code. It's more of an ice-on-stupid-code situation. :) So probably you should remove the "invalid" language from the commentary. Sorry for misleading you.
>
> We could fold this to nothing (if there are no side effects), but it
> would be better if we made stupid code slower instead of faster ;-)
Well, optimization opportunities are not always obvious to the writer.
Iff the builtins have no side-effects
besides the return value the backend should mark them PURE or CONST
and you wouldn't run into
this situation.
But yes, simply folding to GIMPLE_NOP is the appropriate thing when
you want to paper over the
above deficit in the folding routine.
gsi_replace (si_p, gimple_build_nop (), false);
note you'll eventually wreck virtual operands so before do
unlink_stmt_vdef (gsi_stmt (gsi));
which may have it's own share of issues (folding may not look at SSA
immediate uses...).
So better fixup builtin attributes!
Richard.
>
> Segher