[v2,2/2] Use modulo instead of mul+sub in __secs_to_tm

Submitted by Mattias Andrée on Feb. 28, 2021, 7:22 p.m.

Details

Message ID 20210228192210.1665554-2-maandree@kth.se
State New
Series "Remove unnecessary if in __secs_to_tm"
Headers show

Commit Message

Mattias Andrée Feb. 28, 2021, 7:22 p.m.
On x86 modulo is free when doing division, so this removes
a multiplication and at the cost of replacing a conditional
move with a conditional jump, but it still appears to be
faster.
(Similar architectures: nds32le)

ARM doesn't have modulo, instead an multiply-and-subtract
operation is done after the division, so the diffence
here is either none at all, or a move and a multiply-and-add
being replaced with a multiply-and-subtract.
(Similar architectures: or1k)

RISC-V on the other hand has a separate modulo
instruction and will perform a separate modulo instead of
an assignment, a multiplication, and an addition with
this change. GCC does change how the modulo operation is
realised depending on the optimisation level. I don't know
how this affects the performance, however a simple test on
x86 suggests that doing a modulo operations is actually
faster than assign–multiply–add.
---
 src/time/__secs_to_tm.c | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

Patch hide | download patch | download mbox

diff --git a/src/time/__secs_to_tm.c b/src/time/__secs_to_tm.c
index 62219df5..348e51ec 100644
--- a/src/time/__secs_to_tm.c
+++ b/src/time/__secs_to_tm.c
@@ -39,16 +39,28 @@  int __secs_to_tm(long long t, struct tm *tm)
 		qc_cycles--;
 	}
 
+#if 1
+	c_cycles = remdays / DAYS_PER_100Y;
+	remdays %= DAYS_PER_100Y;
+	if (c_cycles == 4) {
+		remdays += DAYS_PER_100Y;
+		c_cycles--;
+	}
+#else
 	c_cycles = remdays / DAYS_PER_100Y;
 	if (c_cycles == 4) c_cycles--;
 	remdays -= c_cycles * DAYS_PER_100Y;
+#endif
 
 	q_cycles = remdays / DAYS_PER_4Y;
-	remdays -= q_cycles * DAYS_PER_4Y;
+	remdays %= DAYS_PER_4Y;
 
 	remyears = remdays / 365;
-	if (remyears == 4) remyears--;
-	remdays -= remyears * 365;
+	remdays %= 365;
+	if (remyears == 4) {
+		remdays += 365;
+		remyears--;
+	}
 
 	leap = !remyears && (q_cycles || !c_cycles);
 	yday = remdays + 31 + 28 + leap;

Comments

Szabolcs Nagy Feb. 28, 2021, 7:37 p.m.
* Mattias Andrée <maandree@kth.se> [2021-02-28 20:22:10 +0100]:
> On x86 modulo is free when doing division, so this removes

there should be no division.

div by const is transformed to mul and shift at -O1 and
that's what we should be using instead of manual hacks.

https://godbolt.org/z/Wsxq5h

> a multiplication and at the cost of replacing a conditional
> move with a conditional jump, but it still appears to be
> faster.
> (Similar architectures: nds32le)
> 
> ARM doesn't have modulo, instead an multiply-and-subtract
> operation is done after the division, so the diffence
> here is either none at all, or a move and a multiply-and-add
> being replaced with a multiply-and-subtract.
> (Similar architectures: or1k)
> 
> RISC-V on the other hand has a separate modulo
> instruction and will perform a separate modulo instead of
> an assignment, a multiplication, and an addition with
> this change. GCC does change how the modulo operation is
> realised depending on the optimisation level. I don't know
> how this affects the performance, however a simple test on
> x86 suggests that doing a modulo operations is actually
> faster than assign–multiply–add.

did you benchmark with CFLAGS=-O2 or -Os ?

> ---
>  src/time/__secs_to_tm.c | 18 +++++++++++++++---
>  1 file changed, 15 insertions(+), 3 deletions(-)
> 
> diff --git a/src/time/__secs_to_tm.c b/src/time/__secs_to_tm.c
> index 62219df5..348e51ec 100644
> --- a/src/time/__secs_to_tm.c
> +++ b/src/time/__secs_to_tm.c
> @@ -39,16 +39,28 @@ int __secs_to_tm(long long t, struct tm *tm)
>  		qc_cycles--;
>  	}
>  
> +#if 1
> +	c_cycles = remdays / DAYS_PER_100Y;
> +	remdays %= DAYS_PER_100Y;
> +	if (c_cycles == 4) {
> +		remdays += DAYS_PER_100Y;
> +		c_cycles--;
> +	}
> +#else
>  	c_cycles = remdays / DAYS_PER_100Y;
>  	if (c_cycles == 4) c_cycles--;
>  	remdays -= c_cycles * DAYS_PER_100Y;
> +#endif
>  
>  	q_cycles = remdays / DAYS_PER_4Y;
> -	remdays -= q_cycles * DAYS_PER_4Y;
> +	remdays %= DAYS_PER_4Y;
>  
>  	remyears = remdays / 365;
> -	if (remyears == 4) remyears--;
> -	remdays -= remyears * 365;
> +	remdays %= 365;
> +	if (remyears == 4) {
> +		remdays += 365;
> +		remyears--;
> +	}
>  
>  	leap = !remyears && (q_cycles || !c_cycles);
>  	yday = remdays + 31 + 28 + leap;
> -- 
> 2.30.1
Mattias Andrée Feb. 28, 2021, 7:52 p.m.
On Sun, 28 Feb 2021 20:37:33 +0100
Szabolcs Nagy <nsz@port70.net> wrote:

> * Mattias Andrée <maandree@kth.se> [2021-02-28 20:22:10 +0100]:
> > On x86 modulo is free when doing division, so this removes  
> 
> there should be no division.
> 
> div by const is transformed to mul and shift at -O1 and
> that's what we should be using instead of manual hacks.
> 
> https://godbolt.org/z/Wsxq5h

For -Os, the currently used optimisation, it does division.
But for other optimisations, it makes no difference as the
compiler will do a multiply–subtract either way.

> 
> > a multiplication and at the cost of replacing a conditional
> > move with a conditional jump, but it still appears to be
> > faster.
> > (Similar architectures: nds32le)
> > 
> > ARM doesn't have modulo, instead an multiply-and-subtract
> > operation is done after the division, so the diffence
> > here is either none at all, or a move and a multiply-and-add
> > being replaced with a multiply-and-subtract.
> > (Similar architectures: or1k)
> > 
> > RISC-V on the other hand has a separate modulo
> > instruction and will perform a separate modulo instead of
> > an assignment, a multiplication, and an addition with
> > this change. GCC does change how the modulo operation is
> > realised depending on the optimisation level. I don't know
> > how this affects the performance, however a simple test on
> > x86 suggests that doing a modulo operations is actually
> > faster than assign–multiply–add.  
> 
> did you benchmark with CFLAGS=-O2 or -Os ?

I guess it must have been -O0 or -Os, but what I did was
I made a trivial program and checked that assembly output,
to see which method was faster. The important part here
was that the compiler didn't change the division operation,
so adding optimisation might have bad the test pointless.
As I wrote, for RISC-V the compiler did exactly what was
written, no matter the optimisation level, that is, for
RISC-V I tried, -O0, -O1, -O2, -O3, and -Os.

> 
> > ---
> >  src/time/__secs_to_tm.c | 18 +++++++++++++++---
> >  1 file changed, 15 insertions(+), 3 deletions(-)
> > 
> > diff --git a/src/time/__secs_to_tm.c b/src/time/__secs_to_tm.c
> > index 62219df5..348e51ec 100644
> > --- a/src/time/__secs_to_tm.c
> > +++ b/src/time/__secs_to_tm.c
> > @@ -39,16 +39,28 @@ int __secs_to_tm(long long t, struct tm *tm)
> >  		qc_cycles--;
> >  	}
> >  
> > +#if 1
> > +	c_cycles = remdays / DAYS_PER_100Y;
> > +	remdays %= DAYS_PER_100Y;
> > +	if (c_cycles == 4) {
> > +		remdays += DAYS_PER_100Y;
> > +		c_cycles--;
> > +	}
> > +#else
> >  	c_cycles = remdays / DAYS_PER_100Y;
> >  	if (c_cycles == 4) c_cycles--;
> >  	remdays -= c_cycles * DAYS_PER_100Y;
> > +#endif
> >  
> >  	q_cycles = remdays / DAYS_PER_4Y;
> > -	remdays -= q_cycles * DAYS_PER_4Y;
> > +	remdays %= DAYS_PER_4Y;
> >  
> >  	remyears = remdays / 365;
> > -	if (remyears == 4) remyears--;
> > -	remdays -= remyears * 365;
> > +	remdays %= 365;
> > +	if (remyears == 4) {
> > +		remdays += 365;
> > +		remyears--;
> > +	}
> >  
> >  	leap = !remyears && (q_cycles || !c_cycles);
> >  	yday = remdays + 31 + 28 + leap;
> > -- 
> > 2.30.1
Jon Chesterfield Feb. 28, 2021, 7:58 p.m.
On Sun, 28 Feb 2021, 19:37 Szabolcs Nagy, <nsz@port70.net> wrote:

> * Mattias Andrée <maandree@kth.se> [2021-02-28 20:22:10 +0100]:
> > On x86 modulo is free when doing division, so this removes
>
> there should be no division.
>
> div by const is transformed to mul and shift at -O1 and
> that's what we should be using instead of manual hacks.


Right. Divide by constant is cheap because compilers have a bunch of
transforms to get rid of the divide in favour of one of more cheaper
instructions.

Note that module coming for free with division doesn't make it cheap.
Integer division is far more expensive that integer multiply on ~ every
architecture. Several architectures implement division in software. It's
not cheap on x86, despite the dedicated instruction.

Cheers
Markus Wichmann March 1, 2021, 7:26 p.m.
On Sun, Feb 28, 2021 at 07:58:27PM +0000, Jon Chesterfield wrote:
> Note that module coming for free with division doesn't make it cheap.
> Integer division is far more expensive that integer multiply on ~ every
> architecture. Several architectures implement division in software. It's
> not cheap on x86, despite the dedicated instruction.
>
> Cheers

And then there's PowerPC, which does have a divide instruction but no
modulo. The manual explicitly states that if you need a modulo, you are
supposed to divide, multiply, and subtract.

Ciao,
Markus