[Bug c/104187] New: Call site specific attribute to control inliner

[Bug c/104187] New: Call site specific attribute to control inliner

[david.bolvansky at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

            Bug ID: 104187
           Summary: Call site specific attribute to control inliner
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: david.bolvansky at gmail dot com
  Target Milestone: ---

It could be useful to have more control over inlining. Use cases:



int foo();

void bar();

int g; 

void test()
{
 g = __builtin_always_inline(foo()); // force inlining of foo() here
 __builtin_noinline(bar()); // never inline bar to this function
}

[Bug ipa/104187] Call site specific attribute to control inliner

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |marxin at gcc dot gnu.org
           Keywords|                            |missed-optimization
           Severity|normal                      |enhancement
          Component|c                           |ipa

[Bug ipa/104187] Call site specific attribute to control inliner

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

Martin Liška <marxin at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2022-01-24
     Ever confirmed|0                           |1

[Bug ipa/104187] Call site specific attribute to control inliner

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  For C++ an attribute on the call stmt might work even better.

Note we're low on bits to represent this on the GENERIC CALL_EXPR and
the GIMPLE gcall.

[Bug ipa/104187] Call site specific attribute to control inliner

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #1)
> Confirmed.  For C++ an attribute on the call stmt might work even better.

I think even C, it might be better too. [[]] syntax is there for GNU C now too.
But those only apply to the full statement and not just a subset of one ...

[Bug ipa/104187] Call site specific attribute to control inliner

[hubicka at kam dot mff.cuni.cz]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

--- Comment #3 from hubicka at kam dot mff.cuni.cz ---
> > Confirmed.  For C++ an attribute on the call stmt might work even better.
> 
> I think even C, it might be better too. [[]] syntax is there for GNU C now too.
> But those only apply to the full statement and not just a subset of one ...
I agree it is quite desirable thing to have

Implementaiton on the inliner side is quite easy: all we need to do is
to detect the attribute and put it into the inline summary and then
handle it same way as always_inline.

However what is the way to add the attribute to front-ends and get it
attached to the gimple call statement?

[Bug ipa/104187] Call site specific attribute to control inliner

[joseph at codesourcery dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

--- Comment #4 from joseph at codesourcery dot com <joseph at codesourcery dot com> ---
Right now, all the C front end does with statement attributes is parses 
them and then passes them to c_warn_unused_attributes; it doesn't have any 
other handling for such attributes.

[Bug ipa/104187] Call site specific attribute to control inliner

[david.bolvansky at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

--- Comment #5 from Dávid Bolvanský <david.bolvansky at gmail dot com> ---
So you prefer eg.

g = a[i] - [[gnu::always_inline]] foo(x, y) + 2 * bar();

over

g = a[i] - __builtin_always_inline(foo(x, y)) + 2 * bar();

?

What is your proposed syntax?

[Bug ipa/104187] Call site specific attribute to control inliner

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

--- Comment #6 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
For Gimple representation __builtin_always_inline(foo(x, y))
does not make very good sense since gimple is not nested and foo may return
void so there may not be any link between call and the builtin.

I see that llvm has something like this
https://reviews.llvm.org/D51200
(or at least it was proposed), but I think in our setting it would be more
natural to have flag in the gimple call statement or some generic way to attach
attributes to statements like we attach histograms.

[Bug ipa/104187] Call site specific attribute to control inliner

[aaron at aaronballman dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

Aaron Ballman <aaron at aaronballman dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |aaron at aaronballman dot com

--- Comment #7 from Aaron Ballman <aaron at aaronballman dot com> ---
(In reply to Dávid Bolvanský from comment #5)
> So you prefer eg.
> 
> g = a[i] - [[gnu::always_inline]] foo(x, y) + 2 * bar();

To be clear, this syntax isn't valid in either C or C++. There are
statement-level attributes, there are not expression-level attributes.

[Bug ipa/104187] Call site specific attribute to control inliner

[david.bolvansky at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104187

--- Comment #8 from Dávid Bolvanský <david.bolvansky at gmail dot com> ---
So this works in Clang now

int foo(int x, int y) { // any compiler will happily inline this function
    return x / y;
}

int test(int x, int y) {
    int r = 0;
    [[clang::noinline]] r += foo(x, y); // for some reason we don't want any
inlining here
    return r;
}


foo(int, int): # @foo(int, int)
  mov eax, edi
  cdq
  idiv esi
  ret
test(int, int): # @test(int, int)
  jmp foo(int, int) # TAILCALL

foo(int, int): # @foo(int, int)
  mov eax, edi
  cdq
  idiv esi
  ret
test(int, int): # @test(int, int)
  jmp foo(int, int) # TAILCALL