[9fans] what does #pragma fpround do?

[9fans] what does #pragma fpround do?

[erik quanstrom]

reading 8c (the only compiler that appears to
implement this pragma), the difference appears
to be in the conversion of (double, float) ->
any type of integer.  however,

	#include <u.h>
	#include <libc.h>

	void
	main(void)
	{
		double d;

		d = 1.99999999999;
	#pragma fpround on
		print("%d\n", (int)d);
	#pragma fpround 0
		print("%d\n", (int)d);
	}

generates the same output and same assembly
for both casts.  can anyone explain what this pragma
is supposed to do?

- erik

Re: [9fans] what does #pragma fpround do?

[Russ Cox]

> generates the same output and same assembly
> for both casts.  can anyone explain what this pragma
> is supposed to do?

it changes the rounding mode from the standard
truncate to integer (expensive on a 387) to
round to nearest (incorrect but cheap).

       #pragma fpround on
               print("%d\n", (int)d);
       #pragma fpround 0
               print("%d\n", (int)d);

your examples compiles to the same code in both
cases because the rounding mode is only consulted
during code generation, which happens at the
function's final }.  you'd need to write two different
functions to demonstrate the difference.

russ

Re: [9fans] what does #pragma fpround do?

[erik quanstrom]

> > generates the same output and same assembly
> > for both casts.  can anyone explain what this pragma
> > is supposed to do?
>
> it changes the rounding mode from the standard
> truncate to integer (expensive on a 387) to
> round to nearest (incorrect but cheap).

sure enough.  here's timings in µsec for 1<<20 cycles

			slow	fast	factor
amd64 (2.0ghz)		82717	2873	29
xeon 5000 (1.6ghz)	17595	1963	9
core i7 (2.6ghz)		4734	1185	4

>        #pragma fpround on
>                print("%d\n", (int)d);
>        #pragma fpround 0
>                print("%d\n", (int)d);
>
> your examples compiles to the same code in both
> cases because the rounding mode is only consulted
> during code generation, which happens at the
> function's final }.  you'd need to write two different
> functions to demonstrate the difference.

ah.  that escaped me.  thanks.  it would be good to
document somwhere.  would updating /sys/doc/compiler.ms
be a really bad idea?

- erik