On Fri, Dec 8, 2023 at 1:24 AM Jeff Law <jeffreya...@gmail.com> wrote:
>
>
>
> On 12/5/23 07:53, Richard Biener via Gcc wrote:
> > On Tue, Dec 5, 2023 at 3:54 PM Alexander Monakov via Gcc
> > <gcc@gcc.gnu.org> wrote:
> >>
> >> Greetings,
> >>
> >> the definitions for NOP_EXPR and CONVERT_EXPR in tree.def, having survived
> >> all the way from 1992, currently say:
> >>
> >> /* Represents a conversion of type of a value.
> >> All conversions, including implicit ones, must be
> >> represented by CONVERT_EXPR or NOP_EXPR nodes. */
> >> DEFTREECODE (CONVERT_EXPR, "convert_expr", tcc_unary, 1)
> >>
> >> /* Represents a conversion expected to require no code to be
> >> generated. */
> >> DEFTREECODE (NOP_EXPR, "nop_expr", tcc_unary, 1)
> >>
> >> Unfortunately, they are confusing, as in
> >>
> >> float f(double d)
> >> {
> >> return d;
> >> }
> >>
> >> the narrowing conversion is represented with NOP_EXPR, and it is definitely
> >> not a no-op.
> >>
> >> Does some clear distinction remain, and is it possible to clarify the
> >> definitions?
> >
> > {NOP,CONVERT}_EXPR are interchangeable in the middle-end but
> > frontends (IIRC the C++ FE mainly) distinguishes them. So a uniform
> > documentation might be difficult - in the end we could eventually
> > drop NOP_EXPR from the middle-end (during gimplification?) and
> > only use CONVERT_EXPR. All uses should use CASE_CONVERT
> > or CONVERT_EXPR_CODE_P which globs both.
> I thought someone looked at this a while ago (measured in years) and
> concluded it wasn't actually feasible. Perhaps because the middle end
> still hands things off to routines that are also used by the FE.
>
> I could see dropping/converting during gimplification with a checker
> that verifies they don't sneak back in. Then we can start to expunge
> them from gimple passes. Feels like a gcc-15+ problem to me.
It's not so long that I tried this (but really by removing NOP_EXPR) when
I figured the C++ FE at least won't be happy. The gimplification route
and IL checking so NOP_EXPR doesn't creep back in could work though.
Richard.
> jeff
