| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
On ARM all divisions have to be performed "manually". For 64-bit
divisions that may take more than a hundred cycles in many cases.
With 32-bit divisions gcc already use the recyprocal of constant
divisors to perform a multiplication, but not with 64-bit divisions.
Since the kernel is increasingly relying upon 64-bit divisions it is
worth optimizing at least those cases where the divisor is a constant.
This is what this patch does using plain C code that gets optimized away
at compile time.
For example, despite the amount of added C code, do_div(x, 10000) now
produces the following assembly code (where x is assigned to r0-r1):
adr r4, .L0
ldmia r4, {r4-r5}
umull r2, r3, r4, r0
mov r2, #0
umlal r3, r2, r5, r0
umlal r3, r2, r4, r1
mov r3, #0
umlal r2, r3, r5, r1
mov r0, r2, lsr #11
orr r0, r0, r3, lsl #21
mov r1, r3, lsr #11
...
.L0:
.word 948328779
.word 879609302
which is the fastest that can be done for any value of x in that case,
many times faster than the __do_div64 code (except for the small x value
space for which the result ends up being zero or a single bit).
The fact that this code is generated inline produces a tiny increase in
.text size, but not significant compared to the needed code around each
__do_div64 call site this code is replacing.
The algorithm used has been validated on a 16-bit scale for all possible
values, and then recodified for 64-bit values. Furthermore I've been
running it with the final BUG_ON() uncommented for over two months now
with no problem.
Note that this new code is compiled with gcc versions 4.0 or later.
Earlier gcc versions proved themselves too problematic and only the
original code is used with them.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
|
