Hi!
On Tue, May 11, 2021 at 06:08:06AM +0000, Christophe Leroy wrote:
Commit 328e7e487a46 ("powerpc: force inlining of csum_partial() to
avoid multiple csum_partial() with GCC10") inlined csum_partial().
Now that csum_partial() is inlined, GCC outlines csum_add() when
called by csum_partial().
Could you build this with -fdump-tree-einline-all and send me the
results? Or open a GCC PR yourself :-)
Something seems to have decided this asm is more expensive than it is.
That isn't always avoidable -- the compiler cannot look inside asms --
but it seems it could be improved here.
Do you have (or can make) a self-contained testcase?
The sum with 0 is useless, should have been skipped.
That isn't something the compiler can do anything about (not sure if you
were suggesting that); it has to be done in the user code (and it tries
to already, see below).
And there is even one completely unused instance of csum_add().
That is strange, that should never happen.
./arch/powerpc/include/asm/checksum.h: In function '__ip6_tnl_rcv':
./arch/powerpc/include/asm/checksum.h:94:22: warning: inlining failed in call to 'csum_add': call is unlikely and code size would grow [-Winline]
94 | static inline __wsum csum_add(__wsum csum, __wsum addend)
| ^~~~~~~~
./arch/powerpc/include/asm/checksum.h:172:31: note: called from here
172 | sum = csum_add(sum, (__force __wsum)*(const u32 *)buff);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
At least we say what happened. Progress! :-)
In the non-inlined version, the first sum with 0 was performed.
Here it is skipped.
That is because of how __builtin_constant_p works, most likely. As we
discussed elsewhere it is evaluated before all forms of loop unrolling.
The patch looks perfect of course :-)
Reviewed-by: Segher Boessenkool <redacted>
Segher
Hi,
Le 11/05/2021 à 12:51, Segher Boessenkool a écrit :
Hi!
On Tue, May 11, 2021 at 06:08:06AM +0000, Christophe Leroy wrote:
quoted
Commit 328e7e487a46 ("powerpc: force inlining of csum_partial() to
avoid multiple csum_partial() with GCC10") inlined csum_partial().
Now that csum_partial() is inlined, GCC outlines csum_add() when
called by csum_partial().
Could you build this with -fdump-tree-einline-all and send me the
results? Or open a GCC PR yourself :-)
Ok, I'll forward it to you in a minute.
Something seems to have decided this asm is more expensive than it is.
That isn't always avoidable -- the compiler cannot look inside asms --
but it seems it could be improved here.
Do you have (or can make) a self-contained testcase?
I have not tried, and I fear it might be difficult, because on a kernel build with dozens of calls
to csum_add(), only ip6_tunnel.o exhibits such an issue.
quoted
The sum with 0 is useless, should have been skipped.
That isn't something the compiler can do anything about (not sure if you
were suggesting that); it has to be done in the user code (and it tries
to already, see below).
I was not suggesting that, only that when properly inlined the sum with 0 is skipped (because we put
the necessary stuff in csum_add() of course).
quoted
And there is even one completely unused instance of csum_add().
That is strange, that should never happen.
It seems that several .o include unused versions of csum_add. After the final link, one remains (in
addition to the used one) in vmlinux.
quoted
./arch/powerpc/include/asm/checksum.h: In function '__ip6_tnl_rcv':
./arch/powerpc/include/asm/checksum.h:94:22: warning: inlining failed in call to 'csum_add': call is unlikely and code size would grow [-Winline]
94 | static inline __wsum csum_add(__wsum csum, __wsum addend)
| ^~~~~~~~
./arch/powerpc/include/asm/checksum.h:172:31: note: called from here
172 | sum = csum_add(sum, (__force __wsum)*(const u32 *)buff);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
At least we say what happened. Progress! :-)
Lol. I've seen this warning for long, that's not something new I guess.
quoted
In the non-inlined version, the first sum with 0 was performed.
Here it is skipped.
That is because of how __builtin_constant_p works, most likely. As we
discussed elsewhere it is evaluated before all forms of loop unrolling.
But we are not talking about loop unrolling here, are we ?
It seems that the reason here is that __builtin_constant_p() is evaluated long after GCC decided to
not inline that call to csum_add().
Christophe
On Wed, May 12, 2021 at 02:56:56PM +0200, Christophe Leroy wrote:
Le 11/05/2021 à 12:51, Segher Boessenkool a écrit :
quoted
Something seems to have decided this asm is more expensive than it is.
That isn't always avoidable -- the compiler cannot look inside asms --
but it seems it could be improved here.
Do you have (or can make) a self-contained testcase?
I have not tried, and I fear it might be difficult, because on a kernel
build with dozens of calls to csum_add(), only ip6_tunnel.o exhibits such
an issue.
Yeah. Sometimes you can force some of the decisions, but that usually
requires knowing too many GCC internals :-/
quoted
quoted
And there is even one completely unused instance of csum_add().
That is strange, that should never happen.
It seems that several .o include unused versions of csum_add. After the
final link, one remains (in addition to the used one) in vmlinux.
But it is a static function, so it should not end up in any object file
where it isn't used.
quoted
quoted
In the non-inlined version, the first sum with 0 was performed.
Here it is skipped.
That is because of how __builtin_constant_p works, most likely. As we
discussed elsewhere it is evaluated before all forms of loop unrolling.
But we are not talking about loop unrolling here, are we ?
Oh, right you are, but that doesn't change much. The
_builtin_constant_p(len) is evaluated long before the compiler sees len
is a constant here.
It seems that the reason here is that __builtin_constant_p() is evaluated
long after GCC decided to not inline that call to csum_add().
Yes, it seems we do not currently do even trivial inlining except very
early in the compiler.
Thanks,
Segher
Le 12/05/2021 à 16:31, Segher Boessenkool a écrit :
On Wed, May 12, 2021 at 02:56:56PM +0200, Christophe Leroy wrote:
quoted
Le 11/05/2021 à 12:51, Segher Boessenkool a écrit :
quoted
Something seems to have decided this asm is more expensive than it is.
That isn't always avoidable -- the compiler cannot look inside asms --
but it seems it could be improved here.
Do you have (or can make) a self-contained testcase?
I have not tried, and I fear it might be difficult, because on a kernel
build with dozens of calls to csum_add(), only ip6_tunnel.o exhibits such
an issue.
Yeah. Sometimes you can force some of the decisions, but that usually
requires knowing too many GCC internals :-/
quoted
quoted
quoted
And there is even one completely unused instance of csum_add().
That is strange, that should never happen.
It seems that several .o include unused versions of csum_add. After the
final link, one remains (in addition to the used one) in vmlinux.
But it is a static function, so it should not end up in any object file
where it isn't used.
Well .... did I dream ?
Now I only find one extra .o with unused csum_add() : That's net/ipv6/exthdrs.o
It matches the one found in vmlinux.
Are you interested in -fdump-tree-einline-all for that one as well ?
Christophe
On Wed, May 12, 2021 at 04:43:33PM +0200, Christophe Leroy wrote:
Le 12/05/2021 à 16:31, Segher Boessenkool a écrit :
quoted
On Wed, May 12, 2021 at 02:56:56PM +0200, Christophe Leroy wrote:
quoted
Le 11/05/2021 à 12:51, Segher Boessenkool a écrit :
quoted
Something seems to have decided this asm is more expensive than it is.
That isn't always avoidable -- the compiler cannot look inside asms --
but it seems it could be improved here.
Do you have (or can make) a self-contained testcase?
I have not tried, and I fear it might be difficult, because on a kernel
build with dozens of calls to csum_add(), only ip6_tunnel.o exhibits such
an issue.
Yeah. Sometimes you can force some of the decisions, but that usually
requires knowing too many GCC internals :-/
quoted
quoted
quoted
And there is even one completely unused instance of csum_add().
That is strange, that should never happen.
It seems that several .o include unused versions of csum_add. After the
final link, one remains (in addition to the used one) in vmlinux.
But it is a static function, so it should not end up in any object file
where it isn't used.
Well .... did I dream ?
Now I only find one extra .o with unused csum_add() : That's
net/ipv6/exthdrs.o
It matches the one found in vmlinux.
Are you interested in -fdump-tree-einline-all for that one as well ?
From: Michael Ellerman <hidden> Date: 2021-06-18 04:23:01
On Tue, 11 May 2021 06:08:06 +0000 (UTC), Christophe Leroy wrote:
Commit 328e7e487a46 ("powerpc: force inlining of csum_partial() to
avoid multiple csum_partial() with GCC10") inlined csum_partial().
Now that csum_partial() is inlined, GCC outlines csum_add() when
called by csum_partial().
c064fb28 <csum_add>:
c064fb28: 7c 63 20 14 addc r3,r3,r4
c064fb2c: 7c 63 01 94 addze r3,r3
c064fb30: 4e 80 00 20 blr
[...]