string-choice: add yesno(), onoff(), enableddisabled(), plural() helpers
Subsystems (6)
PATCH v4 1/1 string-choice: add yesno(), onoff(), enableddisabled(), plural() helpers
The kernel has plenty of ternary operators to choose between constant strings, such as condition ? "yes" : "no", as well as value == 1 ? "" :
"s":
$ git grep '? "yes" : "no"' | wc -l 258 $ git grep '? "on" : "off"' | wc -l 204 $ git grep '? "enabled" : "disabled"' | wc -l 196 $ git grep '? "" : "s"' | wc -l 25
Additionally, there are some occurences of the same in reverse order, split to multiple lines, or otherwise not caught by the simple grep.
Add helpers to return the constant strings. Remove existing equivalent and conflicting functions in i915, cxgb4, and USB core. Further conversion can be done incrementally.
The main goal here is to abstract recurring patterns, and slightly clean up the code base by not open coding the ternary operators.
---
v2: add string-choice.[ch] to not clutter kernel.h and to actually save space on string constants.
v3: back to static inlines based on Rasmus' feedback
v4: Massaged commit message about space savings to make it less fluffy based on Rasmus' feedback.
Example of further cleanup possibilities are at [1], to be done incrementally afterwards.
[1] http://lore.kernel.org/r/[email protected]
--- drivers/gpu/drm/i915/i915_utils.h | 16 +--------- .../ethernet/chelsio/cxgb4/cxgb4_debugfs.c | 12 +------ drivers/usb/core/config.c | 6 +--- drivers/usb/core/generic.c | 6 +--- include/linux/string-choice.h | 31 +++++++++++++++++++ 5 files changed, 35 insertions(+), 36 deletions(-) create mode 100644 include/linux/string-choice.h
drivers/gpu/drm/i915/i915_utils.h +1 -15
@@ -28,6 +28,7 @@#include <linux/list.h> #include <linux/overflow.h> #include <linux/sched.h> +#include <linux/string-choice.h> #include <linux/types.h> #include <linux/workqueue.h>
@@ -395,21 +396,6 @@ wait_remaining_ms_from_jiffies(unsigned long timestamp_jiffies, int to_wait_ms)#define MBps (x) KBps(1000 * (x))#define GBps (x) ((u64)1000 * MBps((x))) -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 void add_taint_for_CI (unsigned int taint) {/*
drivers/net/ethernet/chelsio/cxgb4/cxgb4_debugfs.c +1 -11
@@ -35,6 +35,7 @@#include <linux/seq_file.h> #include <linux/debugfs.h> #include <linux/string_helpers.h> +#include <linux/string-choice.h> #include <linux/sort.h> #include <linux/ctype.h>
@@ -2023,17 +2024,6 @@ static const struct file_operations rss_debugfs_fops = {/* RSS Configuration. */ -/* Small utility function to return the strings "yes" or "no" if the supplied -* argument is non-zero. -*/ -static const char *yesno (int x) -{ -static const char *yes ="yes" ; -static const char *no ="no" ; - -return x ? yes : no; -} -static int rss_config_show (struct seq_file *seq,void *v) {struct adapter *adapter = seq->private;
drivers/usb/core/config.c +1 -5
@@ -10,6 +10,7 @@#include <linux/module.h> #include <linux/slab.h> #include <linux/device.h> +#include <linux/string-choice.h> #include <asm/byteorder.h> #include "usb.h"
@@ -19,11 +20,6 @@#define USB_MAXCONFIG 8/* Arbitrary limit */ -static inline const char *plural (int n) -{ -return (n ==1 ?"" :"s" ); -} -static int find_next_descriptor (unsigned char *buffer,int size,int dt1,int dt2,int *num_skipped) {
drivers/usb/core/generic.c +1 -5
@@ -21,14 +21,10 @@#include <linux/usb.h> #include <linux/usb/hcd.h> +#include <linux/string-choice.h> #include <uapi/linux/usb/audio.h> #include "usb.h" -static inline const char *plural (int n) -{ -return (n ==1 ?"" :"s" ); -} -static int is_rndis (struct usb_interface_descriptor *desc) {return desc->bInterfaceClass ==USB_CLASS_COMM
include/linux/string-choice.h new file +31 -0
@@ -0,0 +1,31 @@ +/* SPDX-License-Identifier: MIT */ +/* +* Copyright © 2019 Intel Corporation +*/ + +#ifndef __STRING_CHOICE_H__ +#define __STRING_CHOICE_H__ + +#include <linux/types.h> + +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__ */
Signature · 1 line
2.20.1
Quoted 7 lines · Jani Nikula
The kernel has plenty of ternary operators to choose between constant strings, such as condition ? "yes" : "no", as well as value == 1 ? "" :
"s":v4: Massaged commit message about space savings to make it less fluffy based on Rasmus' feedback.
Thanks, it looks good to me. FWIW,
Quoted 6 lines · Rasmus Villemoes
Quoted 5 lines · Jani Nikula
The kernel has plenty of ternary operators to choose between constant strings, such as condition ? "yes" : "no", as well as value == 1 ? "" :
"s":
Quoted 6 lines
Quoted 2 lines
v4: Massaged commit message about space savings to make it less fluffy based on Rasmus' feedback.
Thanks, it looks good to me. FWIW,
Acked-by Rasmus Villemoes <[email protected]>
Thanks.
I think the question is, which tree to apply this to, who's going to pick it up? I'm fine with any route.
BR,
Jani.
Signature · 5 lines
Jani Nikula, Intel Open Source Graphics Center _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 16 lines · On Wed, 23 Oct 2019 16:13:08 +0300 Jani Nikula <[email protected]> wrote:
The kernel has plenty of ternary operators to choose between constant strings, such as condition ? "yes" : "no", as well as value == 1 ? "" :
"s":$ git grep '? "yes" : "no"' | wc -l 258 $ git grep '? "on" : "off"' | wc -l 204 $ git grep '? "enabled" : "disabled"' | wc -l 196 $ git grep '? "" : "s"' | wc -l 25Additionally, there are some occurences of the same in reverse order, split to multiple lines, or otherwise not caught by the simple grep.
Quoted 6 lines
Add helpers to return the constant strings. Remove existing equivalent and conflicting functions in i915, cxgb4, and USB core. Further conversion can be done incrementally.
The main goal here is to abstract recurring patterns, and slightly clean up the code base by not open coding the ternary operators.
Fair enough.
Quoted 24 lines
include/linux/string-choice.h new file +21
-0@@ -0,0 +1,31 @@ +/* SPDX-License-Identifier: MIT */ +/* +* Copyright © 2019 Intel Corporation +*/ + +#ifndef __STRING_CHOICE_H__ +#define __STRING_CHOICE_H__ + +#include <linux/types.h> + +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)
include/linux/string-choice.h line 31
Quoted 10 lines
include/linux/string-choice.h new file +10
-0+{ +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()
(Example: note that slabinfo.c already has an "onoff()").
Also, I worry that making these functions inline means that each .o file will contain its own copy of the strings ("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. If not, we will end up with a smaller kernel by uninlining these functions. lib/string-choice.c would suit.
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:
include/linux/string-choice.h renamed from include/linux/string-choice.h~string-choice-add-yesno-onoff-enableddisabled-plural-helpers-fix +1 -4
@@ -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.
Quoted 2 lines · Andrew Morton
And doing this will cause additional savings: calling a single-arg out-of-line function generates less .text than calling
yesno().
I get no change in size at all with any of
extern
static __always_inline
with either of bool or int argument.
gcc 8.3, defconfig with CONFIG_CHELSIO_T4
Quoted 65 lines · Andrew Morton
Quoted 16 lines · On Wed, 23 Oct 2019 16:13:08 +0300 Jani Nikula <[email protected]> wrote:
The kernel has plenty of ternary operators to choose between constant strings, such as condition ? "yes" : "no", as well as value == 1 ? "" :
"s":$ git grep '? "yes" : "no"' | wc -l 258 $ git grep '? "on" : "off"' | wc -l 204 $ git grep '? "enabled" : "disabled"' | wc -l 196 $ git grep '? "" : "s"' | wc -l 25Additionally, there are some occurences of the same in reverse order, split to multiple lines, or otherwise not caught by the simple grep.
Quoted 6 lines
Add helpers to return the constant strings. Remove existing equivalent and conflicting functions in i915, cxgb4, and USB core. Further conversion can be done incrementally.
The main goal here is to abstract recurring patterns, and slightly clean up the code base by not open coding the ternary operators.
Fair enough.
Quoted 24 lines
include/linux/string-choice.h new file +21
-0@@ -0,0 +1,31 @@ +/* SPDX-License-Identifier: MIT */ +/* +* Copyright © 2019 Intel Corporation +*/ + +#ifndef __STRING_CHOICE_H__ +#define __STRING_CHOICE_H__ + +#include <linux/types.h> + +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)include/linux/string-choice.h line 31
Quoted 10 lines
include/linux/string-choice.h new file +10
-0+{ +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:
Quoted 3 lines
choice_yes_no()choice_enabled_disabled()choice_plural()
I was merely using existing function names used in several drivers in the kernel. But I can rename no problem.
Are your suggestions the names we can settle on now, or should I expect to receive more opinions, but only after I send v5?
Quoted 1 line
(Example: note that
slabinfo.calready has an "onoff()").
Under tools/ though? I did mean to address all conflicts in this patch.
Quoted 26 lines
Also, I worry that making these functions inline means that each .o file will contain its own copy of the strings ("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. If not, we will end up with a smaller kernel by uninlining these functions.
lib/string-choice.cwould suit.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:include/linux/string-choice.h renamed from include/linux/string-choice.h~string-choice-add-yesno-onoff-enableddisabled-plural-helpers-fix +1
-4@@ -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) {
Quoted 2 lines
The text segment of
drivers/net/ethernet/chelsio/cxgb4/cxgb4_debugfs.o(78 callsites) shrunk by 118 bytes.
So we've already been back and forth on that particular topic in the history of this patch. v2 had lib/string-choice.c and no inlines [1].
In the end, starting to use functions, inline or not, will let us rework the implementation as we see fit, without touching the callers.
Again, it's no problem to go back to lib/string-choice.c, *once* more, and the effort is trivial, but the ping-pong is getting old.
BR,
Jani.
[1] http://lore.kernel.org/r/[email protected]
Signature · 5 lines
Jani Nikula, Intel Open Source Graphics Center _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 28 lines · Andrew Morton
Quoted 12 lines · On Wed, 23 Oct 2019 16:13:08 +0300 Jani Nikula <[email protected]> wrote:
(quoted diff) +12
-0+ +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)Review
Quoted 10 lines
(quoted diff) +10
-0+{ + 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:
Quoted 3 lines
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.
Quoted 2 lines
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",
Quoted 2 lines
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.
Quoted 19 lines
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:include/linux/string-choice.h renamed from include/linux/string-choice.h~string-choice-add-yesno-onoff-enableddisabled-plural-helpers-fix +1
-4@@ -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) {
Quoted 3 lines
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 +62which 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
Quoted 3 lines · Rasmus Villemoes
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.
Sorry, that's of course nonsense - the strings only appear once in the TU (inside the static inline function), so gcc must treat them all as the same object - as opposed to the case where the implementation was
#define yesno(x) ((x) ? "yes" : "no")
So that can't explain why you saw a smaller text segment using the OOL version.
Rasmus
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Patchwork
CI✗ Fi.CI.BUILD: failure for string-choice: add yesno(), onoff(), enableddisabled(), plural() helpers (rev2)
In reply to Jani Nikula== Series Details ==
== Series Details ==
Series: string-choice: add yesno(), onoff(), enableddisabled(), plural() helpers (rev2)
URL : https://patchwork.freedesktop.org/series/68461/
State : failure
== Summary ==
Applying: string-choice: add yesno(), onoff(), enableddisabled(), plural() helpers
error: sha1 information is lacking or useless (include/linux/string-choice.h).
error: could not build fake ancestor
hint: Use 'git am --show-current-patch' to see the failed patch
Patch failed at 0001 string-choice: add yesno(), onoff(), enableddisabled(), plural() helpers When you have resolved this problem, run "git am --continue".
If you prefer to skip this patch, run "git am --skip" instead.
To restore the original branch and stop patching, run "git am --abort".
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx