On 24/10/2019 00.56, Andrew Morton wrote:
> On Wed, 23 Oct 2019 16:13:08 +0300 Jani Nikula <jani.nikula@xxxxxxxxx> wrote:
>
>> +
>> +static inline const char *yesno(bool v)
>> +{
>> + return v ? "yes" : "no";
>> +}
>> +
>> +static inline const char *onoff(bool v)
>> +{
>> + return v ? "on" : "off";
>> +}
>> +
>> +static inline const char *enableddisabled(bool v)
>> +{
>> + return v ? "enabled" : "disabled";
>> +}
>> +
>> +static inline const char *plural(long v)
>> +{
>> + return v == 1 ? "" : "s";
>> +}
>> +
>> +#endif /* __STRING_CHOICE_H__ */
>
> These aren't very good function names. Better to create a kernel-style
> namespace such as "choice_" and then add the expected underscores:
>
> choice_yes_no()
> choice_enabled_disabled()
> choice_plural()
I think I prefer the short names (no choice_ prefix), it's rather
obvious what they do. I also asked for underscores, especially for the
enableddisabled case, but Jani didn't want to change existing callers.
But I'll keep out of the naming discussion from now on.
> Also, I worry that making these functions inline means that each .o
> file will contain its own copy of the strings
They will, in .rodata.str1.1
("yes", "no", "enabled",
> etc) if the .c file calls the relevant helper. I'm not sure if the
> linker is smart enough (yet) to fix this up.
AFAIK, that's an optimization the linker has done forever - the whole
reason the SHF_MERGE | SHF_STRINGS (the MS in readelf -S output) flags
exist (and AFAICT they have been part of the ELF spec since forever) is
so the linker can do that trick. So no, do not make them ool.
> And doing this will cause additional savings: calling a single-arg
> out-of-line function generates less .text than calling yesno(). When I
> did this:
>
> --- a/include/linux/string-choice.h~string-choice-add-yesno-onoff-enableddisabled-plural-helpers-fix
> +++ a/include/linux/string-choice.h
> @@ -8,10 +8,7 @@
>
> #include <linux/types.h>
>
> -static inline const char *yesno(bool v)
> -{
> - return v ? "yes" : "no";
> -}
> +const char *yesno(bool v);
>
> static inline const char *onoff(bool v)
> {
>
> The text segment of drivers/net/ethernet/chelsio/cxgb4/cxgb4_debugfs.o
> (78 callsites) shrunk by 118 bytes.
>
Interesting, and not at all what I see. Mind dumping rss_config_show
before/after? Even better, cxgb4_debugfs.s before/after. Here's what I get:
static inline yesno:
cbb: 49 c7 c4 00 00 00 00 mov $0x0,%r12 cbe:
R_X86_64_32S .rodata.str1.1+0x197
cc2: 44 89 ea mov %r13d,%edx
cc5: 48 c7 c6 00 00 00 00 mov $0x0,%rsi cc8:
R_X86_64_32S .rodata.str1.1+0x1b7
ccc: 48 c7 c5 00 00 00 00 mov $0x0,%rbp ccf:
R_X86_64_32S .rodata.str1.1+0x19b
Load "yes" into %12 and "no" into %rbp (or vice versa).
cd3: 4d 89 e7 mov %r12,%r15
cd6: e8 00 00 00 00 callq cdb
<rss_config_show+0x3b> cd7: R_X86_64_PC32 seq_printf-0x4
cdb: 45 85 ed test %r13d,%r13d
cde: 4c 89 e2 mov %r12,%rdx
ce1: 48 89 df mov %rbx,%rdi
ce4: 48 0f 49 d5 cmovns %rbp,%rdx
ce8: 48 c7 c6 00 00 00 00 mov $0x0,%rsi ceb:
R_X86_64_32S .rodata.str1.1+0x1cb
cef: e8 00 00 00 00 callq cf4
<rss_config_show+0x54> cf0: R_X86_64_PC32 seq_printf-0x4
cf4: 41 f7 c5 00 00 00 40 test $0x40000000,%r13d
cfb: 4c 89 e2 mov %r12,%rdx
cfe: 48 89 df mov %rbx,%rdi
d01: 48 0f 44 d5 cmove %rbp,%rdx
d05: 48 c7 c6 00 00 00 00 mov $0x0,%rsi d08:
R_X86_64_32S .rodata.str1.1+0x1e1
d0c: e8 00 00 00 00 callq d11
<rss_config_show+0x71> d0d: R_X86_64_PC32 seq_printf-0x4
Test a bit, move "yes" into %rdx, conditionally move "no" into %rdx
instead, call seq_printf.
d11: 41 f7 c5 00 00 00 20 test $0x20000000,%r13d
d18: 4c 89 e2 mov %r12,%rdx
d1b: 48 89 df mov %rbx,%rdi
d1e: 48 0f 44 d5 cmove %rbp,%rdx
d22: 48 c7 c6 00 00 00 00 mov $0x0,%rsi d25:
R_X86_64_32S .rodata.str1.1+0x1f7
d29: e8 00 00 00 00 callq d2e
<rss_config_show+0x8e> d2a: R_X86_64_PC32 seq_printf-0x4
etc. That's a marginal (i.e., after the preamble storing "yes" and "no"
in callee-saved registers) cost of six instructions/29 bytes per
seq_printf, three of which are to implement the yesno() call.
extern const char *yesno():
64e7: 48 c7 c6 00 00 00 00 mov $0x0,%rsi 64ea:
R_X86_64_32S .rodata.str1.1+0x8e4
64ee: 89 ea mov %ebp,%edx
64f0: 41 89 ed mov %ebp,%r13d
64f3: e8 00 00 00 00 callq 64f8
<rss_config_show+0x28> 64f4: R_X86_64_PC32 seq_printf-0x4
64f8: 89 ef mov %ebp,%edi
64fa: c1 ef 1f shr $0x1f,%edi
64fd: e8 00 00 00 00 callq 6502
<rss_config_show+0x32> 64fe: R_X86_64_PC32 yesno-0x4
Three instructions to prepare the argument to yesno and call it.
6502: 48 c7 c6 00 00 00 00 mov $0x0,%rsi 6505:
R_X86_64_32S .rodata.str1.1+0x8f8
6509: 48 89 c2 mov %rax,%rdx
One more to put the return from yesno in the right register.
650c: 48 89 df mov %rbx,%rdi
650f: e8 00 00 00 00 callq 6514
<rss_config_show+0x44> 6510: R_X86_64_PC32 seq_printf-0x4
So not a lot, but still one more instruction, for a total of 31 bytes.
bloat-o-meter says
$ scripts/bloat-o-meter
drivers/net/ethernet/chelsio/cxgb4/cxgb4_debugfs.o.{1,2}
add/remove: 0/0 grow/shrink: 3/0 up/down: 301/0 (301)
Function old new delta
rss_config_show 2343 2482 +139
rss_vf_config_show 263 363 +100
rss_pf_config_show 342 404 +62
which is more then 2*78, but I haven't looked at the code patterns in
the other functions.
Did you use size(1) to compare when you say "text segment"? That would
include .rodata (and, more importantly, .rodata.strX.Y) in its text
column. Maybe your compiler doesn't do string literal merging (since the
linker does it anyway), so your .rodata.str1.1 might contain several
copies of "yes" and "no", but they shouldn't really be counted.
Rasmus
