RE: Is VRP is too conservative to identify boolean value 0 and 1?
> -Original Message-
> From: Richard Guenther [mailto:richard.guent...@gmail.com]
> Sent: Friday, September 02, 2011 5:07 PM
> To: Jiangning Liu
> Cc: gcc@gcc.gnu.org
> Subject: Re: Is VRP is too conservative to identify boolean value 0 and
> 1?
>
> On Fri, Sep 2, 2011 at 7:58 AM, Jiangning Liu
> wrote:
> > Hi,
> >
> > For the following small case,
> >
> > int f(int i, int j)
> > {
> > if (i==1 && j==2)
> > return i;
> > else
> > return j;
> > }
> >
> > with -O2 option, GCC has vrp2 dump like below,
> >
> > ==
> >
> > Value ranges after VRP:
> >
> > i_1: VARYING
> > i_2(D): VARYING
> > D.1249_3: [0, +INF]
> > j_4(D): VARYING
> > D.1250_5: [0, +INF]
> > D.1251_6: [0, +INF]
> > j_10: [2, 2] EQUIVALENCES: { j_4(D) } (1 elements)
> >
> >
> > Removing basic block 3
> > f (int i, int j)
> > {
> > _Bool D.1251;
> > _Bool D.1250;
> > _Bool D.1249;
> >
> > :
> > D.1249_3 = i_2(D) == 1;
> > D.1250_5 = j_4(D) == 2;
> > D.1251_6 = D.1250_5 & D.1249_3;
> > if (D.1251_6 != 0)
> > goto ;
> > else
> > goto ;
> >
> > :
> >
> > :
> > # i_1 = PHI <1(3), j_4(D)(2)>
> > return i_1;
> >
> > }
> >
> >
> >
> > Variable D.1249_3, D.1250_5 and D.1251_6 should be boolean values, so
> the
> > their value ranges should be
> >
> > D.1249_3: [0, 1]
> > D.1250_5: [0, 1]
> > D.1251_6: [0, 1]
> >
> > So why current VRP can't find out this value range?
>
> It does - it just prints it as [0, +INF], they are bools with
> TYPE_MAX_VALUE
> == 1 after all.
Richard,
May I use REG_EXPR(rtx of D.1249_3) in xxx.md file to detect whether
D.1249_3 is a bool or not?
Some comments in GCC says REG_EXPR may be lost in back-end. True?
If we do have REG_EXPR info for some cases in back-end, is it guaranteed to
be correct? May I implementing back-end peephole optimization depending on
REG_EXPR?
Thanks,
-Jiangning
>
> Richard.
>
> >
> > I'm asking this question because the optimizations in back-end need
> this
> > info to do advanced optimization.
> >
> > Thanks,
> > -Jiangning
> >
> >
> >
RE: Is VRP is too conservative to identify boolean value 0 and 1?
Andrew, I realize I needn't back-end solution for my case at all, and in middle end I can directly use the _Bool type info! Appreciate your reply! Thanks, -Jiangning > -Original Message- > From: Andrew Pinski [mailto:pins...@gmail.com] > Sent: Friday, September 02, 2011 2:27 PM > To: Jiangning Liu > Cc: gcc@gcc.gnu.org > Subject: Re: Is VRP is too conservative to identify boolean value 0 and > 1? > > On Thu, Sep 1, 2011 at 10:58 PM, Jiangning Liu > wrote: > > D.1249_3: [0, 1] > > D.1250_5: [0, 1] > > D.1251_6: [0, 1] > > Those are equivalent to [0, MAX] as _Bool only has two different > values, 0 and 1 (MAX). Can you explain more about the optimization > which you are working on that needs the ranges as (int)[0,1] rather > than (_Bool)[0,MAX] ? > > Thanks, > Andrew Pinski
Re: Is VRP is too conservative to identify boolean value 0 and 1?
On Fri, Sep 2, 2011 at 7:58 AM, Jiangning Liu wrote:
> Hi,
>
> For the following small case,
>
> int f(int i, int j)
> {
> if (i==1 && j==2)
> return i;
> else
> return j;
> }
>
> with -O2 option, GCC has vrp2 dump like below,
>
> ==
>
> Value ranges after VRP:
>
> i_1: VARYING
> i_2(D): VARYING
> D.1249_3: [0, +INF]
> j_4(D): VARYING
> D.1250_5: [0, +INF]
> D.1251_6: [0, +INF]
> j_10: [2, 2] EQUIVALENCES: { j_4(D) } (1 elements)
>
>
> Removing basic block 3
> f (int i, int j)
> {
> _Bool D.1251;
> _Bool D.1250;
> _Bool D.1249;
>
> :
> D.1249_3 = i_2(D) == 1;
> D.1250_5 = j_4(D) == 2;
> D.1251_6 = D.1250_5 & D.1249_3;
> if (D.1251_6 != 0)
> goto ;
> else
> goto ;
>
> :
>
> :
> # i_1 = PHI <1(3), j_4(D)(2)>
> return i_1;
>
> }
>
>
>
> Variable D.1249_3, D.1250_5 and D.1251_6 should be boolean values, so the
> their value ranges should be
>
> D.1249_3: [0, 1]
> D.1250_5: [0, 1]
> D.1251_6: [0, 1]
>
> So why current VRP can't find out this value range?
It does - it just prints it as [0, +INF], they are bools with TYPE_MAX_VALUE
== 1 after all.
Richard.
>
> I'm asking this question because the optimizations in back-end need this
> info to do advanced optimization.
>
> Thanks,
> -Jiangning
>
>
>
Re: Is VRP is too conservative to identify boolean value 0 and 1?
On Thu, Sep 1, 2011 at 10:58 PM, Jiangning Liu wrote: > D.1249_3: [0, 1] > D.1250_5: [0, 1] > D.1251_6: [0, 1] Those are equivalent to [0, MAX] as _Bool only has two different values, 0 and 1 (MAX). Can you explain more about the optimization which you are working on that needs the ranges as (int)[0,1] rather than (_Bool)[0,MAX] ? Thanks, Andrew Pinski
Is VRP is too conservative to identify boolean value 0 and 1?
Hi,
For the following small case,
int f(int i, int j)
{
if (i==1 && j==2)
return i;
else
return j;
}
with -O2 option, GCC has vrp2 dump like below,
==
Value ranges after VRP:
i_1: VARYING
i_2(D): VARYING
D.1249_3: [0, +INF]
j_4(D): VARYING
D.1250_5: [0, +INF]
D.1251_6: [0, +INF]
j_10: [2, 2] EQUIVALENCES: { j_4(D) } (1 elements)
Removing basic block 3
f (int i, int j)
{
_Bool D.1251;
_Bool D.1250;
_Bool D.1249;
:
D.1249_3 = i_2(D) == 1;
D.1250_5 = j_4(D) == 2;
D.1251_6 = D.1250_5 & D.1249_3;
if (D.1251_6 != 0)
goto ;
else
goto ;
:
:
# i_1 = PHI <1(3), j_4(D)(2)>
return i_1;
}
Variable D.1249_3, D.1250_5 and D.1251_6 should be boolean values, so the
their value ranges should be
D.1249_3: [0, 1]
D.1250_5: [0, 1]
D.1251_6: [0, 1]
So why current VRP can't find out this value range?
I'm asking this question because the optimizations in back-end need this
info to do advanced optimization.
Thanks,
-Jiangning
