Re: Expanding roundeven (Was: Re: About GSOC.)

[Tejas Joshi <tejasjoshi9673@gmail.com>]

Hello.
Looking at machine description of functions like ceil/floor and I
think specifically:

(define_expand "<rounding_insn><mode>2"

Roundeven should be described as define_expand (or within this
alongside ceil and floor?) with expand functions to emit instructions
in i386-builtins.c?
Referring to <www.info.univ-angers.fr/pub/richer/ens/l3info/ao/intel_sse4.pdf>,
operands of these instructions work with xmm registers, memory and
imm8 for rounding mode.

> existing ROUND_NO_EXC definition in GCC.  A new definition will need
> adding alongside ROUND_FLOOR, ROUND_CEIL and ROUND_TRUNC to correspond to
> rounding to nearest with ties to even, evaluating to 0.)

So (ROUND_ROUNDEVEN   0x0) be declared for rounding towards nearest
even for rounding mode argument? But reference says that RC field
should end up as 00B for rounding ties to even? I am also much
confused about machine-type flags used as C conditions. How do we
determine which ones to check and does it depends on exceptions and
conditions described in the reference? Also, do ROUNDEVEN needs to be
defined as (how):

(define_int_attr rounding_insn
[(UNSPEC_FRNDINT_FLOOR "floor")

Thanks,
--Tejas


On Fri, 14 Jun 2019 at 23:02, Martin Jambor <mjambor@suse.cz> wrote:
>
> Hi,
>
> On Fri, Jun 14 2019, Tejas Joshi wrote:
> >> Of course the instruction is not present there, that is the next step in
> >> your project :-)
> >
> > Yes, of course, but what I meant was that instructions for existing
> > round/ceil functions and not roundeven. If inlining for these
> > functions is available, how can I inspect such instructions getting
> > inlined maybe in a *.s file?
>
> Make sure you compile to a target that has the rounding instruction,
> i.e. by using an appropriate -march or -mavx) and also specify
> -ffast-math on the command line.  I have not double checked, but I
> assume the latter is necessary (mainly) because it implies
> -fno-math-errno and most of the math builtin expansion into instructions
> is guarded by check for !flag_errno_math.
>
> So e.g. my test input:
>
>   double
>   callplusone (double d)
>   {
>     return __builtin_round (d) + 1;
>   }
>
> has vroundsd instruction in the output of:
>
>   ~/gcc/trunk/inst/bin/gcc -O2 -S expand_example.c -mavx -ffast-math
>
> >
> > Also, I am trying to find appropriate places to introduce changes for
> > roundeven to be inlined. Attached patch is what I have tried so far to
> > find the places.
>
> So far I only quickly glanced over it but it seems to be going in the
> right direction.  I'll try to answer the rest of your questions as soon
> as I can but I also have to look a bit into the machine descriptions
> myself first.
>
> Martin
>
>
> > ix86_expand_roundeven have dummy code for the sake of time.
> >
> > In i386.md:
> > 1. How should roundeven be defined at certain places where existing
> > floor is defined as:
> > (define_int_attr rounding_insn
> >     [(UNSPEC_FRNDINT_FLOOR "floor")
> >
> > 2. Also, can roundeven be handled like round_floor is handled by:
> > (define_expand "<rounding_insn><mode>2"
> > but definitely with certain changes of flags, options and macros.
> >
> > Thanks,
> > --Tejas