Re: read_only access attribute as optimizer hint

[Richard Biener <richard.guenther@gmail.com>]

On Wed, Sep 7, 2022 at 1:37 PM Henrik Holst
<henrik.holst@millistream.com> wrote:
>
>
>
> Den ons 7 sep. 2022 kl 09:48 skrev Richard Biener <richard.guenther@gmail.com>:
>>
>> On Tue, Sep 6, 2022 at 5:19 PM Henrik Holst
>> <henrik.holst@millistream.com> wrote:
>> >
>> >
>> >
>> > Den tis 6 sep. 2022 kl 16:47 skrev Richard Biener <richard.guenther@gmail.com>:
>> >>
>> >>
>> >>
>> >> > Am 06.09.2022 um 16:23 schrieb Henrik Holst <henrik.holst@millistream.com>:
>> >> >
>> >> > Hi all,
>> >> >
>> >> >  is there any reason why the access attribute is not used as hints to the
>> >> > optimizer?
>> >> >
>> >> > If we take this ancient example:
>> >> >
>> >> > void foo(const int *);
>> >> >
>> >> > int bar(void)
>> >> > {
>> >> >    int x = 0;
>> >> >    int y = 0;
>> >> >
>> >> >    for (int i = 0; i < 10; i++) {
>> >> >        foo(&x);
>> >> >        y += x;  // this load not optimized out
>> >> >    }
>> >> >    return y;
>> >> > }
>> >> >
>> >> > The load of X is not optimized out in the loop since the compiler does not
>> >> > know if the external function foo() will cast away the const internally.
>> >> > However changing the x variable to const as in:
>> >> >
>> >> > void foo(const int *);
>> >> >
>> >> > int bar(void)
>> >> > {
>> >> >    const int x = 0;
>> >> >    int y = 0;
>> >> >
>> >> >    for (int i = 0; i < 10; i++) {
>> >> >        foo(&x);
>> >> >        y += x;  // this load is now optimized out
>> >> >    }
>> >> >    return y;
>> >> > }
>> >> >
>> >> > The load of x is now optimized out since it is undefined behaviour if bar()
>> >> > casts the const away when x is declared to be const.
>> >> >
>> >> > Now what strikes me as odd however is that declaring the function access
>> >> > attribute to read_only does not hint the compiler to optimize out the load
>> >> > of x even though read_only is defined as being stronger than const ("The
>> >> > mode implies a stronger guarantee than the const qualifier which, when cast
>> >> > away from a pointer, does not prevent the pointed-to object from being
>> >> > modified."), so in the following code:
>> >> >
>> >> > __attribute__ ((access (read_only, 1))) void foo(const int *);
>> >> >
>> >> > int bar(void)
>> >> > {
>> >> >    int x = 0;
>> >> >    int y = 0;
>> >> >
>> >> >    for (int i = 0; i < 10; i++) {
>> >> >        foo(&x);
>> >> >        y += x;  // this load not optimized out even though we have set the
>> >> > access to read_only
>> >> >    }
>> >> >    return y;
>> >> > }
>> >> >
>> >> > The load of x should really be optimized out but isn't. So is this an
>> >> > oversight in gcc or is the access attribute completely ignored by the
>> >> > optimizer for some good reason?
>> >>
>> >> It’s ignored because it is not thoroughly specified.  There’s an alternate representation the language Frontend can rewrite the attribute to to take advantage in optimization if it’s semantics matches.
>> >>
>> >> Richard
>> >
>> > Ok, didn't really understand the bit about the language Frontend but I guess that you are talking about internal GCC things here and thus there is nothing that I as a GCC user can do to inform the optimizer that a variable is read_only as a hint for external functions. And if so could it be "thoroughly specified" to enable this type of optimization or is this just "the way it is" ?
>>
>> Yes, there's currently nothing the user can do.  Looking at the access
>> attribute specification it could be used
>> to initialize the middle-end used 'fn spec' specification - for
>> example the Fortran Frontend uses that to ferry
>> the guarantees by the 'INTENT' argument specification.
>>
>> Richard.
>
> Ok, so patches to utilize the access attribute to inform the optimizer might be accepted?

No, patches transforming the access attribute into appropriate 'fn
spec' might be accepted.
See attr-fnspec.h for how that works.

Richard.

> /HH
>>
>>
>> >
>> > /HH
>> >>
>> >>
>> >>
>> >> > If there is no good reason for this then changing this to hint the
>> >> > optimizer should enable some nice optimizations of external functions where
>> >> > const in the declaration is not cast away.
>> >> >
>> >> > Regards,
>> >> >  Henrik Holst