bug w/ inline of inline?

bug w/ inline of inline?

[Eliot Dresselhaus]

In this example

    static inline int f (int x) { return x + 1; }

    static inline int g (int x, inline int f (int x))
    { return 1 + f (x); }

    int h (int x)
    { return g (x, f); }

is h supposed to optimize to return x + 2 or supposed to actually call
f.  Here's what I get with current gcc mainline cvs -O2

    f:
	    leal	1(%rdi), %eax
	    ret

    h:
	    call	f
	    incl	%eax
	    ret

Is this a bug?  If not, how to I get gcc to emit "return x + 2"?

Thanks,

Eliot

Re: bug w/ inline of inline?

[Zack Weinberg]

Eliot Dresselhaus <edressel@cisco.com> writes:

> In this example
>
>     static inline int f (int x) { return x + 1; }
>
>     static inline int g (int x, inline int f (int x))
>     { return 1 + f (x); }
>
>     int h (int x)
>     { return g (x, f); }
>
> is h supposed to optimize to return x + 2 or supposed to actually call
> f.

GCC is allowed to do either.  It would be good if it produced 
"return x + 2;"  Please file a missed-optimization bug report.

Note that declaring the parameter inline is incorrect C.

zw

Re: bug w/ inline of inline?

[Andrew Pinski]

On Oct 17, 2004, at 8:18 PM, Zack Weinberg wrote:

> Eliot Dresselhaus <edressel@cisco.com> writes:
>
>> In this example
>>
>>     static inline int f (int x) { return x + 1; }
>>
>>     static inline int g (int x, inline int f (int x))
>>     { return 1 + f (x); }
>>
>>     int h (int x)
>>     { return g (x, f); }
>>
>> is h supposed to optimize to return x + 2 or supposed to actually call
>> f.
>
> GCC is allowed to do either.  It would be good if it produced
> "return x + 2;"  Please file a missed-optimization bug report.

We have one already, PR 9079: <http://gcc.gnu.org/PR9079>.

Thanks,
Andrew Pinski
a gcc bug master

Re: bug w/ inline of inline?

[Richard Henderson]

On Sun, Oct 17, 2004 at 10:43:34PM -0400, Richard Kenner wrote:
> I saw this sort of thing a while ago when I was playing with similar
> artificial tests.  The reason the inlining isn't done is because of
> heuristics in cgraph.  It's looking at the number of times the function
> is used and such and the artificial cases don't tend to trigger the
> inlining.

Not even close.

The function isn't inlined because it's called indirectly, and we don't
re-do inlining after we cse the constant function address into the call.



r~

Re: bug w/ inline of inline?

[Richard Kenner]

    GCC is allowed to do either.  It would be good if it produced 
    "return x + 2;"  Please file a missed-optimization bug report.

I saw this sort of thing a while ago when I was playing with similar
artificial tests.  The reason the inlining isn't done is because of
heuristics in cgraph.  It's looking at the number of times the function
is used and such and the artificial cases don't tend to trigger the
inlining.