On 31 March 2014 19:08, Thomas Gleixner tglx@linutronix.de wrote:
And those cleanups make the compiler generate worse code at least on x86_64:
text data bss dec hex filename 7475 554 0 8029 1f5d kernel/hrtimer.o 7706 554 0 8260 2044 kernel/hrtimer.o
So just removing parameters and return values because you can get the same information from a datastructure is not necessarily a good thing.
So here is the patch by patch analysis (With x86_64_defconfig):
Initial details: text data bss dec hex filename 7989 408 0 8397 20cd ../bx86/kernel/hrtimer.o
73a6cd8 hrtimer: replace 'tab' with 'space' after comma ',' 04223a8 hrtimer: Coalesce format fragments in printk() 52fac3f hrtimer: call hrtimer_set_expires_range() from hrtimer_set_expires_range_ns() 535a552 hrtimer: use base->index instead of basenum in switch_hrtimer_base()
7989 408 0 8397 20cd ../bx86/kernel/hrtimer.o
f3a2cdd hrtimer: no need to rewrite '1' to hrtimer_hres_enabled
7974 408 0 8382 20be ../bx86/kernel/hrtimer.o
This one actually made a smaller :)
479d66f hrtimer: don't rewrite same value to expires_next in hrtimer_force_reprogram()
7990 408 0 8398 20ce ../bx86/kernel/hrtimer.o
0e134df hrtimer: use base->hres_active directly instead of hrtimer_hres_active()
8006 408 0 8414 20de ../bx86/kernel/hrtimer.o
So, reading from a per-cpu variable is efficient as compared to base->hres_active. In that case this patch can be dropped :)
0030ddd hrtimer: remove dummy definition of hrtimer_force_reprogram()
8006 408 0 8414 20de ../bx86/kernel/hrtimer.o
4a0c909 hrtimer: don't check state of base->hres_active in hrtimer_switch_to_hres()
8022 408 0 8430 20ee ../bx86/kernel/hrtimer.o
This one actually removed some code, how come it is bigger :(
213bc01 hrtimer: remove clock_was_set_delayed() from hrtimer.h
8006 408 0 8414 20de ../bx86/kernel/hrtimer.o
Though this one made it shorter, we probably need to drop it as clock_was_set_delayed() is now used in kernel/time/timekeeping.c
3defaf8 hrtimer: don't emulate notifier call to initialize timer base
8071 408 0 8479 211f ../bx86/kernel/hrtimer.o
I thought this will surely generate a smaller text segment :(
e438576 hrtimer: use ffs() to iterate over valid bits from active_bases
8151 408 0 8559 216f ../bx86/kernel/hrtimer.o
1ab3d15 timer: simplify CPU_UP_PREPARE notifier code path c5111c9 timer: don't emulate notifier call to initialize timer base a576aa9 hrtimer: move unlock_hrtimer_base() upwards 9ea8619 hrtimer: reorder code in __remove_hrtimer() 2619470 hrtimer: Create hrtimer_get_monoexpires()
8135 408 0 8543 215f ../bx86/kernel/hrtimer.o
b64d746 hrtimer: remove 'base' parameter from remove_timer() and __remove_timer()
8183 408 0 8591 218f ../bx86/kernel/hrtimer.o
ced918e2 hrtimer: remove 'base' parameter from switch_hrtimer_base()
8135 408 0 8543 215f ../bx86/kernel/hrtimer.o
f046106b hrtimer: remove 'base' parameter from enqueue_hrtimer()
8119 408 0 8527 214f ../bx86/kernel/hrtimer.o
b102337 hrtimer: remove 'base' parameter from hrtimer_{enqueue_}reprogram()
8135 408 0 8543 215f ../bx86/kernel/hrtimer.o
So, these are the four patches that have removed parameters from routines. and it started with 8135 and ended up with 8135 :)
d43b5d1 hrtimer: make switch_hrtimer_base() return void
8199 408 0 8607 219f ../bx86/kernel/hrtimer.o
cd08cd9 hrtimer: make lock_hrtimer_base() return void
8199 408 0 8607 219f ../bx86/kernel/hrtimer.o
74a3a18 hrtimer: make enqueue_hrtimer() return void
8167 408 0 8575 217f ../bx86/kernel/hrtimer.o
b5cb057 hrtimer: don't check if timer is queued in __remove_hrtimer()
8135 408 0 8543 215f ../bx86/kernel/hrtimer.o
b56d3fc hrtimer: rewrite switch_hrtimer_base() to remove extra indentation level fccd30d hrtimer: rewrite remove_hrtimer() to remove extra indentation level d4a6c58 hrtimer: replace base by new_base to get resolution: __hrtimer_start_range_ns() 24bf5c0 hrtimer: create base_on_this_cpu() 3f690eb hrtimer: use base->hres_active directly instead of hrtimer_hres_active()
8135 408 0 8543 215f ../bx86/kernel/hrtimer.o
And finally we ended up at 8135 :)