[Bug tree-optimization/62012] New: Loop is not vectorized after function inlining (SCEV)

[Bug tree-optimization/62012] New: Loop is not vectorized after function inlining (SCEV)

[ysrumyan at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62012

            Bug ID: 62012
           Summary: Loop is not vectorized after function inlining (SCEV)
           Product: gcc
           Version: 4.10.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ysrumyan at gmail dot com

We noticed that for one important benchmark using '-lto' options leads to
performance degradation which is caused by not-vectorizing the hottest loop
after function inlining. I can able to reproduce this deficiency using simple
test-case:
if we passed class by reference to function (it need to be compiled with
-DPARAM macros) loop is vectorized:
g++ -Ofast -m64 -march=core-avx2 -c test.cpp  -fdump-tree-vect-details -fopenmp
-DPARAM; grep 'note: vectorized' test.cpp.114t.vect                             
test.cpp:45:1: note: vectorized 1 loops in function.

But if we compile it without macros we get:
g++ -Ofast -m64 -march=core-avx2 -c test.cpp  -fdump-tree-vect-details
-fopenmp; grep 'note: vectorized' test.cpp.114t.vect
test.cpp:45:1: note: vectorized 0 loops in function.

It looks like SCEV issue.

[Bug tree-optimization/62012] Loop is not vectorized after function inlining (SCEV)

[ysrumyan at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62012

--- Comment #1 from Yuri Rumyantsev <ysrumyan at gmail dot com> ---
Created attachment 33241
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=33241&action=edit
test-case to reproduce

Options to compile are:
-Ofast -m64 -march=core-avx2 -fopenmp
and macros -DPARAM can be used to get vectorizable version of test.

[Bug tree-optimization/62012] Loop is not vectorized after function inlining (SCEV)

[ysrumyan at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62012

--- Comment #2 from Yuri Rumyantsev <ysrumyan at gmail dot com> ---
Any updates?

Thanks.

[Bug tree-optimization/62012] Loop is not vectorized after function inlining (SCEV)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62012

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2014-09-08
     Ever confirmed|0                           |1

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
So it's independend of LTO.

Confirmed.  We have

  _28 = &MEM[(struct Array *)&fa + 256B].a[0] + _3;
  *_28 = u_23;

which SCEV messes up because it ends up with

(instantiate_scev
  (instantiate_below = 4)
  (evolution_loop = 1)
  (chrec = &MEM[(struct Array *)&fa + 256B].a[0])
  (res = &MEM[(struct Array *)&fa + 256B].a[0]))
(instantiate_scev
  (instantiate_below = 4)
  (evolution_loop = 1)
  (chrec = {(long unsigned int) first_6(D) * 4, +, 4}_1)
  (res = {(long unsigned int) first_6(D) * 4, +, 4}_1))
(set_scalar_evolution
  instantiated_below = 4
  (scalar = _13)
  (scalar_evolution = {&MEM[(struct Array *)&fa + 256B].a[(sizetype)
first_6(D)], +, 4}_1))
)
failed: evolution of base is not affine.

Not sure why it thinks that.

Btw, on trunk we now vectorize this just fine probably because of the fix
for PR63148 which avoids moving first_6 * 4 inside the array-ref and we
get

  (scalar_evolution = {&MEM[(struct Array *)&fa + 256B].a[0] + (sizetype)
((long unsigned int) first_6(D) * 4), +, 4}_1))
)
success.

instead.

So - can you re-check please?

[Bug tree-optimization/62012] Loop is not vectorized after function inlining (SCEV)

[ysrumyan at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62012

--- Comment #4 from Yuri Rumyantsev <ysrumyan at gmail dot com> ---
I checked that our benchmark is successfully vectorized with function inlining.
 So this bug must be closed as fixed/resolved.

[Bug tree-optimization/62012] Loop is not vectorized after function inlining (SCEV)

[ysrumyan at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62012

--- Comment #5 from Yuri Rumyantsev <ysrumyan at gmail dot com> ---
You can close this bug as fixed/resolved (see my comment).

Thanks.
Yuri.

2014-09-08 15:29 GMT+04:00 rguenth at gcc dot gnu.org
<gcc-bugzilla@gcc.gnu.org>:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62012
>
> Richard Biener <rguenth at gcc dot gnu.org> changed:
>
>            What    |Removed                     |Added
> ----------------------------------------------------------------------------
>              Status|UNCONFIRMED                 |WAITING
>    Last reconfirmed|                            |2014-09-08
>      Ever confirmed|0                           |1
>
> --- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
> So it's independend of LTO.
>
> Confirmed.  We have
>
>   _28 = &MEM[(struct Array *)&fa + 256B].a[0] + _3;
>   *_28 = u_23;
>
> which SCEV messes up because it ends up with
>
> (instantiate_scev
>   (instantiate_below = 4)
>   (evolution_loop = 1)
>   (chrec = &MEM[(struct Array *)&fa + 256B].a[0])
>   (res = &MEM[(struct Array *)&fa + 256B].a[0]))
> (instantiate_scev
>   (instantiate_below = 4)
>   (evolution_loop = 1)
>   (chrec = {(long unsigned int) first_6(D) * 4, +, 4}_1)
>   (res = {(long unsigned int) first_6(D) * 4, +, 4}_1))
> (set_scalar_evolution
>   instantiated_below = 4
>   (scalar = _13)
>   (scalar_evolution = {&MEM[(struct Array *)&fa + 256B].a[(sizetype)
> first_6(D)], +, 4}_1))
> )
> failed: evolution of base is not affine.
>
> Not sure why it thinks that.
>
> Btw, on trunk we now vectorize this just fine probably because of the fix
> for PR63148 which avoids moving first_6 * 4 inside the array-ref and we
> get
>
>   (scalar_evolution = {&MEM[(struct Array *)&fa + 256B].a[0] + (sizetype)
> ((long unsigned int) first_6(D) * 4), +, 4}_1))
> )
> success.
>
> instead.
>
> So - can you re-check please?
>
> --
> You are receiving this mail because:
> You reported the bug.

[Bug tree-optimization/62012] Loop is not vectorized after function inlining (SCEV)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62012

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |RESOLVED
         Resolution|---                         |FIXED

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed.