[permalink] [raw]

Subject: Re: [PATCH 0/6] bitops: let optimize out non-atomic bitops on compile-time constants

From: Geert Uytterhoeven <[email protected]>
Date: Wed, 8 Jun 2022 11:55:08 +0200

> Hi Olek,
>
> On Tue, Jun 7, 2022 at 5:51 PM Alexander Lobakin
> <[email protected]> wrote:
> > From: Geert Uytterhoeven <[email protected]>
> > > On Mon, Jun 6, 2022 at 1:50 PM Alexander Lobakin
> > > <[email protected]> wrote:
> > > > While I was working on converting some structure fields from a fixed
> > > > type to a bitmap, I started observing code size increase not only in
> > > > places where the code works with the converted structure fields, but
> > > > also where the converted vars were on the stack. That said, the
> > > > following code:
> > > >
> > > > DECLARE_BITMAP(foo, BITS_PER_LONG) = { }; // -> unsigned long foo[1];
> > > > unsigned long bar = BIT(BAR_BIT);
> > > > unsigned long baz = 0;
> > > >
> > > > __set_bit(FOO_BIT, foo);
> > > > baz |= BIT(BAZ_BIT);
> > > >
> > > > BUILD_BUG_ON(!__builtin_constant_p(test_bit(FOO_BIT, foo));
> > > > BUILD_BUG_ON(!__builtin_constant_p(bar & BAR_BIT));
> > > > BUILD_BUG_ON(!__builtin_constant_p(baz & BAZ_BIT));
> > > >
> > > > triggers the first assertion on x86_64, which means that the
> > > > compiler is unable to evaluate it to a compile-time initializer
> > > > when the architecture-specific bitop is used even if it's obvious.
> > > > I found that this is due to that many architecture-specific
> > > > non-atomic bitop implementations use inline asm or other hacks which
> > > > are faster or more robust when working with "real" variables (i.e.
> > > > fields from the structures etc.), but the compilers have no clue how
> > > > to optimize them out when called on compile-time constants.
> > > >
> > > > So, in order to let the compiler optimize out such cases, expand the
> > > > test_bit() and __*_bit() definitions with a compile-time condition
> > > > check, so that they will pick the generic C non-atomic bitop
> > > > implementations when all of the arguments passed are compile-time
> > > > constants, which means that the result will be a compile-time
> > > > constant as well and the compiler will produce more efficient and
> > > > simple code in 100% cases (no changes when there's at least one
> > > > non-compile-time-constant argument).
> > > > The condition itself:
> > > >
> > > > if (
> > > > __builtin_constant_p(nr) && /* <- bit position is constant */
> > > > __builtin_constant_p(!!addr) && /* <- compiler knows bitmap addr is
> > > > always either NULL or not */
> > > > addr && /* <- bitmap addr is not NULL */
> > > > __builtin_constant_p(*addr) /* <- compiler knows the value of
> > > > the target bitmap */
> > > > )
> > > > /* then pick the generic C variant
> > > > else
> > > > /* old code path, arch-specific
> > > >
> > > > I also tried __is_constexpr() as suggested by Andy, but it was
> > > > always returning 0 ('not a constant') for the 2,3 and 4th
> > > > conditions.
> > > >
> > > > The savings on x86_64 with LLVM are insane (.text):
> > > >
> > > > $ scripts/bloat-o-meter -c vmlinux.{base,test}
> > > > add/remove: 72/75 grow/shrink: 182/518 up/down: 53925/-137810 (-83885)
> > > >
> > > > $ scripts/bloat-o-meter -c vmlinux.{base,mod}
> > > > add/remove: 7/1 grow/shrink: 1/19 up/down: 1135/-4082 (-2947)
> > > >
> > > > $ scripts/bloat-o-meter -c vmlinux.{base,all}
> > > > add/remove: 79/76 grow/shrink: 184/537 up/down: 55076/-141892 (-86816)
> > >
> > > Thank you!
> > >
> > > I gave it a try on m68k, and am a bit disappointed seeing an increase
> > > in code size:
> > >
> > > add/remove: 49/13 grow/shrink: 279/138 up/down: 6434/-3342 (3092)
> >
> > Ufff, that sucks =\
> > Could you please try to compile the following code snippet (with the
> > series applied)?
> >
> > unsigned long map;
> >
> > bitmap_zero(&map, BITS_PER_LONG);
> > __set_bit(1, &map);
> > BUILD_BUG_ON(!__builtin_constant_p(map));
> >
> > If it fails during the vmlinux linkage, it will mean that on your
> > architecture/setup the compiler is unable to optimize the generic
> > implementations to compile-time constants and I'll need to debug
> > this more (probably via some compiler explorer).
>
> Builds and links fine.

Ok, so that means the main task is still achieved.

>
> > You could also check the vmlinux size after applying each patch
> > to see which one does this if you feel like it :)
>
> The (incremental) impact of the various patches is shown below:
>
> bitops: unify non-atomic bitops prototypes across architectures
> add/remove: 4/11 grow/shrink: 123/160 up/down: 1700/-2786 (-1086)
>
> bitops: let optimize out non-atomic bitops on compile-time constants
> add/remove: 50/7 grow/shrink: 280/101 up/down: 6798/-2620 (4178)
>
> I.e. the total impact is -1086 + 4178 = +3092
>
> Looking at the impact of the last change on a single file, with rather
> small functions to make it easier to analyze, the results are:
>
> bloat-o-meter net/core/sock.o{.orig,}
> add/remove: 3/1 grow/shrink: 20/3 up/down: 286/-68 (218)
> Function old new delta
> sock_flag - 38 +38
> sock_set_flag - 22 +22
> sock_reset_flag - 22 +22
> sock_recv_errqueue 264 284 +20
> sock_alloc_send_pskb 406 424 +18
> __sock_set_timestamps 104 122 +18
> sock_setsockopt 2412 2428 +16
> sock_pfree 52 66 +14
> sock_wfree 236 248 +12
> sk_wait_data 222 234 +12
> sk_destruct 70 82 +12
> sk_wake_async 40 50 +10
> sk_set_memalloc 74 84 +10
> __sock_queue_rcv_skb 254 264 +10
> __sk_backlog_rcv 92 102 +10
> sock_getsockopt 1734 1742 +8
> sock_no_linger 36 42 +6
> sk_clone_lock 478 484 +6
> sk_clear_memalloc 98 104 +6
> __sk_receive_skb 194 200 +6
> sock_init_data 344 348 +4
> __sock_cmsg_send 196 200 +4
> sk_common_release 152 154 +2
> sock_set_keepalive 62 60 -2
> sock_enable_timestamp 80 72 -8
> sock_valbool_flag 34 12 -22
> bset_mem_set_bit 36 - -36
> Total: Before=18862, After=19080, chg +1.16%
>
> Several small inline functions are no longer inlined.
> And e.g. __sk_backlog_rcv() increases as it now calls sock_flag()
> out-of-line, and needs to save more on the stack:
>
> __sk_backlog_rcv:
> + move.l %a3,-(%sp) |,
> move.l %d2,-(%sp) |,
> - move.l 8(%sp),%a0 | sk, sk
> -| arch/m68k/include/asm/bitops.h:163: return (addr[nr >> 5] & (1UL
> << (nr & 31))) != 0;
> - move.l 76(%a0),%d0 | MEM[(const long unsigned int
> *)sk_6(D) + 76B], _14
> + move.l 12(%sp),%a3 | sk, sk
> | net/core/sock.c:330: BUG_ON(!sock_flag(sk, SOCK_MEMALLOC));
> - btst #14,%d0 |, _14
> - jne .L193 |
> + pea 14.w |
> + move.l %a3,-(%sp) | sk,
> + jsr sock_flag |
> + addq.l #8,%sp |,
> + tst.b %d0 | tmp50
> + jne .L192 |
> pea 330.w |
> pea .LC0 |
> pea .LC3 |
> jsr _printk |
> trap #7
>
> Note that the above is with atari_defconfig, which has
> CONFIG_CC_OPTIMIZE_FOR_SIZE=y.
>
> Switching to CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE=y results in a kernel
> that is ca. 25% larger, and the net impact of this series is:
>
> add/remove: 24/27 grow/shrink: 227/233 up/down: 7494/-8080 (-586)
>
> i.e. a reduction in size...

Much thanks for such an analysis!

I see that previously it was uninlining bset_mem_set_bit(), which is
m68k implementation of __set_bit(), and now it uninlined the "outer"
sock_*_flag().
I'd name it sort of normal, but the intention was to not change the
existing code when we're unable to optimize it out... And those sock
helpers are usually from that side, since @sk is never anything
constant. It's unexpected that the compiler is unable to resolve the
expression to the original bitop call on `-Os`. Dunno if it's
okay <O> __builtin_constant_p() still gives the desired results, so
I guess it is, but at the same time that function layout rebalance
can potentially hurt performance. Or improve it ._.

BTW, after updating LLVM 13 -> 14 I'm now getting -3Kb (was -86).
GCC 12 still gives -30.

>
> Gr{oetje,eeting}s,
>
> Geert
>
> --
> Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]
>
> In personal conversations with technical people, I call myself a hacker. But
> when I'm talking to journalists I just say "programmer" or something like that.
> -- Linus Torvalds

Thanks,
Olek