This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

conditional lim


Could you please review my patch for predicated lim?

Let me note some details about it:

1)      Phi statements are still moved only if they have 1 or 2
arguments. However, phi statements could be move under conditions (as
itâs done for the other statements).  Probably, phi statement motion
with 3 + arguments could be implemented in the next patch after
predicated lim.

2)      Patch has limitations/features like (it was ok to me to
implement it such way, maybe Iâm not correct. ):

a)      Loop1


              If (a)



                       Stmt - Invariant for Loop1



       In this case Stmt will be moved only out of Loop2, because of if (a).

b)      Or




         If (cond1)

              If (cond2)

                  If (cond3)



Stmt will be moved out only if cond1 is always executed in Loop1.

c)       It took me a long time to write all of these code, so there
might be other peculiarities which I forgot to mention. :)

                       Letâs discuss these ones as you will review my patch.

3)      Patch consists of 9 files:

a)      gcc/testsuite/gcc.dg/tree-ssa/loop-7.c,
gcc/testsuite/gcc.dg/tree-ssa/recip-3.c â changed tests:

-          gcc/testsuite/gcc.dg/tree-ssa/loop-7.c  changed as
predicated lim moves 2 more statements out of the loop;

-          gcc/testsuite/gcc.dg/tree-ssa/recip-3.c â with conditional
lim recip optimization in this test doesnât work (the corresponding
value is below threshold as I could see in the code for recip, 1<3).
So to have recip working in this test I changed test a little bit.

b)      gcc/tree-ssa-loop-im.c â the patched lim per se

c)       gcc/testsuite/gcc.dg/tree-ssa/ssa-lim-13.c,



the tests for conditional lim.

4)      Patch testing:

a)      make âk check (no difference in results for me for the clean
build and the patched one,

-          Revision: 222849,

-          uname -a

               Linux Istanbul 3.16.0-23-generic #31-Ubuntu SMP Tue Oct
21 18:00:35 UTC 2014 i686 i686 i686 GNU/Linux

b)      Bootstrap.

 It goes well now, however to fix it I have made a temporary hack in
the lim code. And with this fix patch definitely shouldnât be

I did so, as I would like to discuss this issue first.

The problem is: I got stage2-stage3 comparison failure on the single
file (tree-vect-data-refs.o). After some investigation I understood
that tree-vect-data-refs.o differs being compiled with and without
â-gâ option (yes, more exactly on stage 2 this is  â-g âO2 âgtoggleâ,
and for stage 3 this is â-g âO2â. But to simplify things I can
reproduce this difference on the same build (even not bootstrapped),
with option â âgâ and without it).

Of course, the file compiled with âg option will contain debug
information and will differ from the corresponding file without debug
information. I mean there is the difference reported by
contrib/compare-debug (I mean we talk about stripped files).

Then I compared assemblers and lim dump files and made a conclusion
the difference is:

There is statement _484=something, which is conditionally moved out of
loop X. In non debug cases no usages of _484 are left inside the loop
X. In debug case, there is the single use in debug statement. So for
debug case my patch generates phi statement for _484 to make it
available inside the loop X. For non debug case, no such phi statement
generated as there is no uses of _484.

There is one more statement with lhs=_493, with exactly this pattern
of use. But this is not important for our discussion.

So, in my opinion itâs ok to generate additional phi node for debug
case. But Iâm not a compiler expert and maybe there is some
requirement that debug and non-debug versions should differ only by
debug statements, I donât know.

My temporary hack fix is just removing of such debug statements (no
debug statements, no problems J).

The alternatives I see are:

-          Move debug statements outside the loop (not good in my opinion);

-          Somehow reset values in debug statements (not good in my
opinion, if I could provide correct information for them);

-          Generate phi for debug statements and fix bootstrap (say,
letâs have some list of files, which we have special check for. I mean
for my case, the difference is not in stage2 and stage3, itâs in debug
and non-debug). I like this variant. However, as I donât see such list
of special files in the whole gcc, I think I might be missing

What do you think?



Attachment: patch_May8_3
Description: Binary data

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]