Re: [PATCH] middle-end IFN_ASSUME support [PR106654]

[Martin Uecker <ma.uecker@gmail.com>]

Am Samstag, den 15.10.2022, 10:53 +0200 schrieb Jakub Jelinek:
> On Sat, Oct 15, 2022 at 10:07:46AM +0200, Martin Uecker wrote:
> > But why? Do we really want to encourage people to
> > write such code?
> 
> Of course these ++ cases inside of expressions are just obfuscation.
> But the point is to support using predicates that can be inlined and
> simplified into something really simple the optimizers can understand.

This makes sense,.

> The paper shows as useful e.g. being able to assert something is finite:
> [[assume (std::isfinite (x)]];
> and with the recent changes on the GCC side it is now or shortly will be
> possible to take advantage of such predicates.
> It is true that
> [[assume (__builtin_isfinite (x)]];
> could work if we check TREE_SIDE_EFFECTS on the GCC side because
> it is a const function, but that is a GNU extension, so the standard
> can't count with that.  std::isfinite isn't even marked const in libstdc++
> and one can figure that out during IPA propagation only.

Hm, that already seems to work with

if (!std::isfinite(x))
  __builtin_unreachable();

https://godbolt.org/z/hj3WrEhjb


> There are many similar predicates, or user could have some that are useful
> to his program.  And either in the end it wouldn't have side-effects
> but the compiler doesn't know, or would but those side-effects would be
> unimportant to the optimizations the compiler can derive from those.

I still have the feeling that relying on something
such as the pure and const attributes might then
be a better approach for this.

From the standards point of view, this is OK
as GCC can just set its own rules as long as it is
a subset of what the standard allows.

> As the spec defines it well what happens with the side-effects and it
> is an attribute, not a function and the languages have non-evaluated
> contexts in other places, I don't see where a user confusion could come.

The user confusion might come when somebody writes
something such as [[assume(1 == ++i)]] and I expect
that people will start doing this once this works.


But I am also a a bit worried about the slipperly slope
of exploiting this more because what "would evaluate to true"
implies in case of I/O, atomic accesses, volatile accesses 
etc.  does not seem clear to me.   But maybe I am worrying
too much.


> We don't warn for sizeof (i++) and similar either.

Which is also confusing and clang does indeed
warn about it outside of macros and I think GCC
should too.

> __builtin_assume (i++) is a bad choice because it looks like a function
> call (after all, the compilers have many similar builtins) and its argument
> looks like normal argument to the function, so it is certainly unexpected
> that the side-effects aren't evaluated.

I agree.

Best
Martin