On Sat, Oct 21, 2023 at 1:12 AM Andi Kleen <ak@xxxxxxxxxxxxxxx> wrote:
>
> A common cause of binary code bloat is excessive inlining. Traditional
> tools (like nm --size-sort -t d) don't address that directly because
> they only see the final functions, but don't know about inlines.
>
> This patch adds inline-account that makes it easy to track that down
> by accounting code bytes to all functions visible in the debug information,
> as well as code lines.
>
> Here are some examples:
>
> Show all inlines that increase code size by >1K in the core scheduler:
>
> $ inline-account.py --min-bytes=1000 kernel/sched/core.o
> Total code bytes seen 75690
>
> Code bytes by functions:
> Function Total Avg Num
> rq_pin_lock 1401 (0.02%) 35 39
> __sched_setscheduler 1277 (0.02%) 41 31
> perf_fetch_caller_regs 1012 (0.01%) 17 58
>
> Code bytes by nearby source line blocks:
> prefix /home/ak/lsrc/git/linux/
> Line Total
> kernel/sched/sched.h:1610 1387 (0.02%)
> include/trace/events/sched.h:16 1172 (0.02%)
> include/trace/events/sched.h:222 1058 (0.01%)
>
> This indicates that rq_pin_lock should likely be not inline,
> and perhaps perf_fetch_caller_regs not either.
>
> Note that not all large inlines are necessary bloat. If there is only
> a single call site it isn't bloat (the tool currently cannot distinguish
> that case). For example it is commonly seen with syscall definitions
> that use single large inlines with only a single caller. In the example
> above I think it's the case with __sched_setscheduler.
>
> Show the >1K inlines in lib/maple_tree.o, which for some reason
> comes in at a incredible 73k of code size:
>
> $ inline-account.py --min-bytes 1000 lib/maple_tree.o
> Total code bytes seen 73578
>
> Code bytes by functions:
> Function Total Avg Num
> mas_mab_cp 5537 (0.08%) 37 149
> mas_pop_node 3798 (0.05%) 28 131
> ma_slots 2368 (0.03%) 14 162
> ma_pivots 2353 (0.03%) 10 222
> mas_destroy_rebalance 2056 (0.03%) 42 48
> mas_start 1661 (0.02%) 13 125
> mas_set_parent 1454 (0.02%) 20 72
> mas_set_alloc_req 1410 (0.02%) 17 80
> mte_node_type 1360 (0.02%) 5 228
> mas_data_end 1189 (0.02%) 16 74
> mte_to_node 1085 (0.01%) 3 276
> mas_split 1053 (0.01%) 65 16
> mas_topiary_replace 1033 (0.01%) 38 27
> mas_root_expand 1001 (0.01%) 35 28
>
> Code bytes by nearby source line blocks:
> prefix /home/ak/lsrc/git/linux/
> Line Total
> lib/maple_tree.c:210 1360 (0.02%)
> include/trace/events/maple_tree.h:80 1283 (0.02%)
> lib/maple_tree.c:649 1193 (0.02%)
> lib/maple_tree.c:288 1097 (0.01%)
>
> It's clear there is a lot of potential for shrinking here, as a quick
> experiment shows:
>
> $ size lib/maple_tree.o
> text data bss dec hex filename
> 72257 5312 8 77577 12f09 lib/maple_tree.o
> $ sed -i -e s/__always_inline// -e 's/ inline/ /' lib/maple_tree.c
> $ make -s lib/maple_tree.o
> $ size lib/maple_tree.o
> text data bss dec hex filename
> 47774 4720 8 52502 cd16 lib/maple_tree.o
>
> 34% reduction just from trusting the compiler. Most of it seems
> to come from abuse of __always_inline. I suppose a large scale
> tree purge of that would give some decent binary size results.
>
> $ inline-account.py --show=5 kernel/workqueue.o
> Total code bytes seen 40403
>
> Code bytes by functions:
> Function Total Avg Num
> bitmap_copy 1477 (0.04%) 26 56
> show_pwq 912 (0.02%) 76 12
> workqueue_init_early 846 (0.02%) 29 29
> __flush_workqueue 753 (0.02%) 31 24
> alloc_and_link_pwqs 558 (0.01%) 69 8
>
> Code bytes by nearby source line blocks:
> prefix /home/ak/lsrc/git/linux/
> Line Total
> include/linux/bitmap.h:268 1336 (0.03%)
> include/trace/events/workqueue.h:23 1038 (0.03%)
> include/trace/events/workqueue.h:108 732 (0.02%)
> include/trace/events/workqueue.h:59 694 (0.02%)
> include/trace/events/workqueue.h:82 670 (0.02%)
> $
>
> This is an interesting case because bitmap_copy is just
>
> static inline void bitmap_copy(unsigned long *dst, const unsigned long *src,
> unsigned int nbits)
> {
> unsigned int len = BITS_TO_LONGS(nbits) * sizeof(unsigned long);
>
> if (small_const_nbits(nbits))
> *dst = *src;
> else
> memcpy(dst, src, len);
> }
>
> memcpy (which is a macro) must sometimes generate a lot of
> code. The small_const_nbits case definitely should be inlined though
> because it's likely even smaller than a call. Would need
> more investigation.
>
> The other large inlines are trace points. Perhaps there is something
> there that could be done to shrink that a bit.
>
> Finally we can do a global accounting (currently with multiple runs):
>
> (ignore the percentage numbers since they are just for the local file)
>
> $ find -name '*.o' | xargs -n1 inline-account.py > a
> $ sort -n -r -k 2 a | head -30
> ZSTD_count 81799 (0.19%) 32 2514
> ZSTD_count 52233 (0.25%) 33 1544
> kmalloc 43324 (0.00%) 12 3334
> pv_queued_spin_unlock 42027 (0.00%) 9 4580
> constant_test_bit 41667 (0.00%) 5 8005
> arch/x86/include/asm/paravirt.h:591 41044 (0.00%)
> arch/x86/include/asm/bitops.h:207 40153 (0.00%)
> __refcount_add 37968 (0.00%) 24 1532
> page_fixed_fake_head 36368 (0.00%) 19 1832
> include/linux/slab.h:599 35654 (0.00%)
> arch/x86/include/asm/jump_label.h:27 35156 (0.00%)
> spin_lock 32170 (0.00%) 10 3007
> __refcount_sub_and_test 32068 (0.00%) 17 1842
> include/linux/spinlock.h:351 31102 (0.00%)
> arch_static_branch 30874 (0.00%) 4 7022
> get_current 30714 (0.00%) 9 3351
> arch/x86/include/asm/current.h:41 29912 (0.00%)
> trace_trigger_soft_disabled 29814 (0.00%) 21 1368
> perf_fetch_caller_regs 27504 (0.00%) 16 1634
> ZSTD_storeSeq 26060 (0.06%) 30 862
> hid_map_usage 25582 (0.00%) 88 288
> ZSTD_compressBlock_lazy_generic 24953 (0.12%) 46 535
> ZSTD_compressBlock_lazy_generic 24953 (0.06%) 46 535
> paravirt_ret0 24152 (0.00%) 24152 1
> spin_unlock_irqrestore 23253 (0.00%) 10 2281
> include/linux/spinlock.h:406 22526 (0.00%)
> ZSTD_RowFindBestMatch 21527 (0.10%) 23 922
> ZSTD_RowFindBestMatch 21527 (0.05%) 23 922
> __list_add 21209 (0.00%) 11 1851
> include/linux/refcount.h:283 20642 (0.00%)
>
> - So my kernel is spending around ~30K just for getting task_structs in
> current.
> - I'm sure ZSTD is great, but is it >200K in duplicated code worth great?
> - page_fixed_fake_head probably shouldn't be inlined
> - There might be some potential in out lining reference counts
> (although that one might be truly performance critical)
> - There's maybe some potential in shrinking trace point bloat?
>
> ... and more similar insights.
>
> Also of course there can be critical inlines that really need
> to be inline in many sites for best performance. But that's rarely the case
> if they are big because it's unlikely the small call overhead is making
> a significant difference for a large chunk of code.
>
> In any case the tool is useful, so I think it deserves its place
> in scripts/
>
> Signed-off-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> ---
> scripts/inline-account.py | 173 ++++++++++++++++++++++++++++++++++++++
> 1 file changed, 173 insertions(+)
> create mode 100755 scripts/inline-account.py
>
> diff --git a/scripts/inline-account.py b/scripts/inline-account.py
> new file mode 100755
> index 000000000000..b6cfe195efe6
> --- /dev/null
> +++ b/scripts/inline-account.py
> @@ -0,0 +1,173 @@
> +#!/usr/bin/env python3
SPDX?
> +# account code bytes per source code / functions from objdump -Sl output
> +# useful to find inline bloat
> +# Author: Andi Kleen
A blank line please between the comment header
and the start of the code.
For example, see scripts/clang-tools/run-clang-tools.py
as a Python script written in a good coding style.
> +import os, sys, re, argparse, multiprocessing
> +from collections import Counter
> +from functools import reduce
> +
> +p = argparse.ArgumentParser(
> + description="""
> +Account code bytes per source code / functions from objdump.
> +Useful to find inline bloat.
> +
> +The line numbers are the beginning of a block, so the actual code can be later.
> +Line numbers can be a also little off due to objdump bugs
> +also some misaccounting can happen due to inexact gcc debug information.
> +The number output for functions may account a single large function multiple
> +times. program/object files need to be built with -g.
> +
> +This is somewhat slow due to objdump -S being slow. It helps to have
> +plenty of cores.""")
> +p.add_argument('--min-bytes', type=int, help='minimum bytes to report', default=100)
> +p.add_argument('--threads', '-t', type=int, default=multiprocessing.cpu_count(),
> + help='Number of objdump processes to run')
> +p.add_argument('--verbose', '-v', action='store_true', help="Print more")
> +p.add_argument('--show', type=int, help='Number of results to show')
> +p.add_argument('file', help='object file/program as input')
> +args = p.parse_args()
Put the code lines in a function,
except import directives and
if __name__ == '__main__':
main()
> +
> +def get_syms(fn):
> + f = os.popen("nm --print-size " + fn)
> + syms = []
> + pc = None
> + for l in f:
> + n = l.split()
> + if len(n) > 2 and n[2].upper() == "T":
> + pc = int(n[0], 16)
> + syms.append(pc)
> + ln = int(n[1], 16)
> + f.close()
with os.popen("nm --print-size " + fn) as f:
then, remove the f.close()
> + if not pc:
> + sys.exit(fn + " has no symbols")
> + syms.append(pc + ln)
I guess you intended 'pc + ln' as the
tail of the .text
Not really because the output from 'nm --print-size'
is not sorted by address.
You need to pass the -n flag to
make 'nm' sort the output by address.
> + return syms
> +
> +class Account:
> + pass
class without any method?
Please make add_account() and account_range()
class methods for readability.
Maybe account_range() to __init().
If you want to use map(),
al = list(map(lambda r: Account(r), ranges))
> +
> +def add_account(a, b):
> + a.funcbytes += b.funcbytes
> + a.linebytes += b.linebytes
> + a.funccount += b.funccount
> + a.nolinebytes += b.nolinebytes
> + a.nofuncbytes += b.nofuncbytes
> + a.total += b.total
> + return a
> +
> +# dont add sys.exit here, causes deadlocks
> +def account_range(r):
> + a = Account()
> + a.funcbytes = Counter()
> + a.linebytes = Counter()
> + a.funccount = Counter()
> + a.nolinebytes = 0
> + a.nofuncbytes = 0
> + a.total = 0
> +
> + line = None
> + func = None
> + codefunc = None
> +
> + cmd = ("objdump -Sl %s --start-address=%#x --stop-address=%#x" %
> + (args.file, r[0], r[1]))
> + f = os.popen(cmd)
Ditto. with os.popen(cmd) as f:
> + for l in f:
> + # 250: e8 00 00 00 00 callq 255 <proc_skip_spaces+0x5>
> + m = re.match(r'\s*([0-9a-fA-F]+):\s+(.*)', l)
> + if m:
> + #print "iscode", func, l,
Do not leave debugging code.
> + bytes = len(re.findall(r'[0-9a-f][0-9a-f] ', m.group(2)))
> + if not func:
> + a.nofuncbytes += bytes
> + continue
> + if not line:
> + a.nolinebytes += bytes
> + continue
> + a.total += bytes
> + a.funcbytes[func] += bytes
> + a.linebytes[(file, line)] += bytes
> + codefunc = func
> + continue
> +
> + # sysctl_init():
> + m = re.match(r'([a-zA-Z_][a-zA-Z0-9_]*)\(\):$', l)
> + if m:
> + if codefunc and m.group(1) != codefunc:
> + a.funccount[codefunc] += 1
> + codefunc = None
> + func = m.group(1)
> + continue
> +
> + # /sysctl.c:1666
> + m = re.match(r'^([^:]+):(\d+)$', l)
> + if m:
> + file, line = m.group(1), int(m.group(2))
> + continue
> + f.close()
> +
> + if codefunc:
> + a.funccount[codefunc] += 1
> + return a
> +
> +# objdump -S is slow, so we parallelize
> +
> +# split symbol table into chunks for parallelization
> +# we split on functions boundaries to avoid mis-accounting
> +# assumes functions have roughly similar length
> +syms = sorted(get_syms(args.file))
This sorted() is too late.
Sorting must be done by 'nm -n'.
> +chunk = max(int(min((len(syms) - 1) / args.threads, len(syms) - 1)), 1)
I do not understand.
Why min()?
If you check args.threads >= 1,
(len(syms) - 1) / args.threads is smaller, isn't it?
> +boundaries = [syms[x] for x in range(0, len(syms) - 1, chunk)] + [syms[-1]]
> +ranges = [(boundaries[x], boundaries[x+1]) for x in range(0, len(boundaries) - 1)
> + if boundaries[x+1] > boundaries[x]]
> +assert ranges[0][0] == syms[0]
> +assert ranges[-1][1] == syms[-1]
> +
> +# map-reduce
> +if args.threads == 1:
> + al = list(map(account_range, ranges))
> +else:
> + al = multiprocessing.Pool(args.threads).map(account_range, ranges)
> +a = reduce(add_account, al)
> +
> +print("Total code bytes seen", a.total)
> +if args.verbose:
> + print("Bytes with no function %d (%.2f%%)" % (a.nofuncbytes, 100.0*(float(a.nofuncbytes)/a.total)))
> + print("Bytes with no lines %d (%.2f%%)" % (a.nolinebytes, 100.0*(float(a.nolinebytes)/a.total)))
> +
> +def sort_map(m):
> + return sorted(list(m.keys()), key=lambda x: m[x], reverse=True)
> +
> +print("\nCode bytes by functions:")
> +print("%-50s %-5s %-5s %-5s %-5s" % ("Function", "Total", "", "Avg", "Num"))
> +for i, j in enumerate(sort_map(a.funcbytes)):
> + if a.funcbytes[j] < args.min_bytes:
> + break
> + if args.show and i >= args.show:
> + break
> + print("%-50s %-5d (%.2f%%) %-5d %-5d" % (
> + j,
> + a.funcbytes[j],
> + a.funcbytes[j] / float(a.total),
> + a.funcbytes[j] / a.funccount[j],
> + a.funccount[j]))
> +
> +for j in list(a.linebytes.keys()):
> + if a.linebytes[j] < args.min_bytes:
> + del a.linebytes[j]
> +
> +# os.path.commonprefix fails with >50k entries
> +# just use the first 10
> +prefix = os.path.commonprefix([x[0] for x in list(a.linebytes.keys())[:10]])
This does not seem to be a good idea.
I see a lot of absolute paths left behind,
and also empty lines.
> +
> +print("\nCode bytes by nearby source line blocks:")
> +print("prefix", prefix)
> +
> +print("%-50s %-5s" % ("Line", "Total"))
> +for i, j in enumerate(sort_map(a.linebytes)):
> + if args.show and i >= args.show:
> + break
> + print("%-50s %-5d (%.2f%%)" % (
> + "%s:%d" % (j[0].replace(prefix, ""), j[1]),
If j[0] and prefix are the same,
the file part disappears.
> + a.linebytes[j],
> + a.linebytes[j] / float(a.total)))
> --
> 2.41.0
>
--
Best Regards
Masahiro Yamada
