Unjustified optimization due to restricted struct members?

Unjustified optimization due to restricted struct members?

[Ties Klappe]

[-- Attachment #1: Type: text/plain, Size: 983 bytes --]

When reading section 6.7.3.1 of the C standard (quoted below) about
the *restrict
*type qualifier, the first section talks about *ordinary identifiers*.
These are defined in section 6.2.3, and exclude members of structures.

Let D be a declaration of an ordinary identifier that provides a means of
> designating an object P as a restrict-qualified pointer to type T.


I would assume that this means that in the code excerpt below the function
*h* cannot be optimized by substituting the load of *b.p *for *10*, as the
standard does not specify what it means for a struct member to be restrict
qualified. However, the code is still optimized by gcc (but not Clang), as
can be seen here: https://godbolt.org/z/hEnKKoaae

struct bar {
int* restrict p;
int* restrict q;
};

int h(struct bar b) {
*b.p = 10;
*b.q = 11;
return *b.p;
}

Was this a deliberate choice, or does it simply follow from how restrict is
supported in gcc (and could this be considered a bug w.r.t. the standard)?

Re: Unjustified optimization due to restricted struct members?

[Richard Biener]

On Thu, Nov 30, 2023 at 12:07 PM Ties Klappe via Gcc <gcc@gcc.gnu.org> wrote:
>
> When reading section 6.7.3.1 of the C standard (quoted below) about
> the *restrict
> *type qualifier, the first section talks about *ordinary identifiers*.
> These are defined in section 6.2.3, and exclude members of structures.
>
> Let D be a declaration of an ordinary identifier that provides a means of
> > designating an object P as a restrict-qualified pointer to type T.
>
>
> I would assume that this means that in the code excerpt below the function
> *h* cannot be optimized by substituting the load of *b.p *for *10*, as the
> standard does not specify what it means for a struct member to be restrict
> qualified. However, the code is still optimized by gcc (but not Clang), as
> can be seen here: https://godbolt.org/z/hEnKKoaae
>
> struct bar {
> int* restrict p;
> int* restrict q;
> };
>
> int h(struct bar b) {
> *b.p = 10;
> *b.q = 11;
> return *b.p;
> }
>
> Was this a deliberate choice, or does it simply follow from how restrict is
> supported in gcc (and could this be considered a bug w.r.t. the standard)?

Hmm, this was a deliberate choice (it also works for global 'b'), I didn't think
the standard would exclude that.  Note GCCs C++ standard library makes
use of restrict qualified pointers as structure members for example.

Richard.

Re: Unjustified optimization due to restricted struct members?

[Ties Klappe]

[-- Attachment #1: Type: text/plain, Size: 2003 bytes --]

Thank you Richard.

Similar to the struct example, I was also wondering about why the following
code does *not* get optimized (e.g. https://godbolt.org/z/9eGrjjK81):

int f(int* restrict a[restrict 2]) {
*(a[0]) = 10;
*(a[1]) = 11;
return *(a[0]);
}

Do you happen to know why a reload via a[0] is required? I would have
expected to see the same optimization as is performed for the struct
example.

Kind regards,
Ties

Op do 30 nov 2023 om 13:16 schreef Richard Biener <
richard.guenther@gmail.com>:

> On Thu, Nov 30, 2023 at 12:07 PM Ties Klappe via Gcc <gcc@gcc.gnu.org>
> wrote:
> >
> > When reading section 6.7.3.1 of the C standard (quoted below) about
> > the *restrict
> > *type qualifier, the first section talks about *ordinary identifiers*.
> > These are defined in section 6.2.3, and exclude members of structures.
> >
> > Let D be a declaration of an ordinary identifier that provides a means of
> > > designating an object P as a restrict-qualified pointer to type T.
> >
> >
> > I would assume that this means that in the code excerpt below the
> function
> > *h* cannot be optimized by substituting the load of *b.p *for *10*, as
> the
> > standard does not specify what it means for a struct member to be
> restrict
> > qualified. However, the code is still optimized by gcc (but not Clang),
> as
> > can be seen here: https://godbolt.org/z/hEnKKoaae
> >
> > struct bar {
> > int* restrict p;
> > int* restrict q;
> > };
> >
> > int h(struct bar b) {
> > *b.p = 10;
> > *b.q = 11;
> > return *b.p;
> > }
> >
> > Was this a deliberate choice, or does it simply follow from how restrict
> is
> > supported in gcc (and could this be considered a bug w.r.t. the
> standard)?
>
> Hmm, this was a deliberate choice (it also works for global 'b'), I didn't
> think
> the standard would exclude that.  Note GCCs C++ standard library makes
> use of restrict qualified pointers as structure members for example.
>
> Richard.
>

Re: Unjustified optimization due to restricted struct members?

[Richard Biener]

On Thu, Nov 30, 2023 at 1:50 PM Ties Klappe <tg.klappe@gmail.com> wrote:
>
> Thank you Richard.
>
> Similar to the struct example, I was also wondering about why the following code does not get optimized (e.g. https://godbolt.org/z/9eGrjjK81):
>
> int f(int* restrict a[restrict 2]) {
> *(a[0]) = 10;
> *(a[1]) = 11;
> return *(a[0]);
> }
>
> Do you happen to know why a reload via a[0] is required? I would have expected to see the same optimization as is performed for the struct example.

It's not implemented.  I think there's even a bugreport about this,
basically restrict
gets only "one layer deep".

Richard.

> Kind regards,
> Ties
>
> Op do 30 nov 2023 om 13:16 schreef Richard Biener <richard.guenther@gmail.com>:
>>
>> On Thu, Nov 30, 2023 at 12:07 PM Ties Klappe via Gcc <gcc@gcc.gnu.org> wrote:
>> >
>> > When reading section 6.7.3.1 of the C standard (quoted below) about
>> > the *restrict
>> > *type qualifier, the first section talks about *ordinary identifiers*.
>> > These are defined in section 6.2.3, and exclude members of structures.
>> >
>> > Let D be a declaration of an ordinary identifier that provides a means of
>> > > designating an object P as a restrict-qualified pointer to type T.
>> >
>> >
>> > I would assume that this means that in the code excerpt below the function
>> > *h* cannot be optimized by substituting the load of *b.p *for *10*, as the
>> > standard does not specify what it means for a struct member to be restrict
>> > qualified. However, the code is still optimized by gcc (but not Clang), as
>> > can be seen here: https://godbolt.org/z/hEnKKoaae
>> >
>> > struct bar {
>> > int* restrict p;
>> > int* restrict q;
>> > };
>> >
>> > int h(struct bar b) {
>> > *b.p = 10;
>> > *b.q = 11;
>> > return *b.p;
>> > }
>> >
>> > Was this a deliberate choice, or does it simply follow from how restrict is
>> > supported in gcc (and could this be considered a bug w.r.t. the standard)?
>>
>> Hmm, this was a deliberate choice (it also works for global 'b'), I didn't think
>> the standard would exclude that.  Note GCCs C++ standard library makes
>> use of restrict qualified pointers as structure members for example.
>>
>> Richard.

Re: Unjustified optimization due to restricted struct members?

[Joseph Myers]

On Thu, 30 Nov 2023, Ties Klappe via Gcc wrote:

> When reading section 6.7.3.1 of the C standard (quoted below) about
> the *restrict
> *type qualifier, the first section talks about *ordinary identifiers*.
> These are defined in section 6.2.3, and exclude members of structures.
> 
> Let D be a declaration of an ordinary identifier that provides a means of
> > designating an object P as a restrict-qualified pointer to type T.
> 
> 
> I would assume that this means that in the code excerpt below the function
> *h* cannot be optimized by substituting the load of *b.p *for *10*, as the
> standard does not specify what it means for a struct member to be restrict
> qualified. However, the code is still optimized by gcc (but not Clang), as
> can be seen here: https://godbolt.org/z/hEnKKoaae
> 
> struct bar {
> int* restrict p;
> int* restrict q;
> };
> 
> int h(struct bar b) {

In this code, 'b' is the ordinary identifier; "struct bar b" is the 
declaration D.  The declaration does not itself need to have pointer type.

-- 
Joseph S. Myers
joseph@codesourcery.com