[Bug tree-optimization/71264] [5 Regression] ICE in convert_move

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264

Jakub Jelinek  changed:

   What|Removed |Added

 Status|ASSIGNED|RESOLVED
 CC||jakub at gcc dot gnu.org
 Resolution|--- |FIXED
   Target Milestone|5.5 |6.2

--- Comment #25 from Jakub Jelinek  ---
GCC 5 branch has been closed, should be fixed in GCC 6.2 and later.

[Bug tree-optimization/71264] [5 Regression] ICE in convert_move

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264

--- Comment #24 from Richard Biener  ---
Author: rguenth
Date: Fri Jan 20 08:01:27 2017
New Revision: 244683

URL: https://gcc.gnu.org/viewcvs?rev=244683=gcc=rev
Log:
2017-01-20  Richard Biener  

PR tree-optimization/71264
* gcc.dg/vect/pr71264.c: Make sure memory is aligned.

Modified:
trunk/gcc/testsuite/ChangeLog
trunk/gcc/testsuite/gcc.dg/vect/pr71264.c

[Bug tree-optimization/71264] [5 Regression] ICE in convert_move

[ro at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264

--- Comment #23 from Rainer Orth  ---
(In reply to rguent...@suse.de from comment #22)
> On Mon, 24 Oct 2016, ebotcazou at gcc dot gnu.org wrote:
> 
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264
> > 
> > --- Comment #21 from Eric Botcazou  ---
> > > It is supposed to be vectorized on targets w/o V4QImode support.
> > > Not sure if the list of targets that FAIL the testcase would be smaller
> > > if I change the scan to an explicit list of targets (x86)?
> > 
> > Not sure indeed, it passes on PowerPC & PowerPC64 too for example.
> > 
> > > Meanwhile just add more targets to the xfail -- it really is a missed
> > > optimization as I guess for example arm also has V4SImode.
> > 
> > There is no XFAIL currently but we can certainly add one.
> 
> Ah, right ... my tree has local changes:
> 
> Index: gcc/testsuite/gcc.dg/vect/pr71264.c
> ===
> --- gcc/testsuite/gcc.dg/vect/pr71264.c (revision 241395)
> +++ gcc/testsuite/gcc.dg/vect/pr71264.c (working copy)
> @@ -7,6 +7,8 @@ typedef uint8_t footype __attribute__((v
>  void test(uint8_t *ptr, uint8_t *mask)
>  {
>footype mv;
> +  ptr = __builtin_assume_aligned (ptr, __alignof__ (footype));
> +  mask = __builtin_assume_aligned (mask, __alignof__ (footype));
>__builtin_memcpy(, mask, sizeof(mv));
>for (unsigned i = 0; i < 16; i += 4)
>  {
> @@ -17,4 +19,4 @@ void test(uint8_t *ptr, uint8_t *mask)
>  }
>  }
>  
> -/* { dg-final { scan-tree-dump "vectorized 1 loops in function" "vect" } 
> } */
> +/* { dg-final { scan-tree-dump "vectorized 1 loops in function" "vect" { 
> xfail sparc*-*-* } } } */

Shouldn't this patch be applied then?

  Rainer

[Bug tree-optimization/71264] [5 Regression] ICE in convert_move

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264

--- Comment #22 from rguenther at suse dot de  ---
On Mon, 24 Oct 2016, ebotcazou at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264
> 
> --- Comment #21 from Eric Botcazou  ---
> > It is supposed to be vectorized on targets w/o V4QImode support.
> > Not sure if the list of targets that FAIL the testcase would be smaller
> > if I change the scan to an explicit list of targets (x86)?
> 
> Not sure indeed, it passes on PowerPC & PowerPC64 too for example.
> 
> > Meanwhile just add more targets to the xfail -- it really is a missed
> > optimization as I guess for example arm also has V4SImode.
> 
> There is no XFAIL currently but we can certainly add one.

Ah, right ... my tree has local changes:

Index: gcc/testsuite/gcc.dg/vect/pr71264.c
===
--- gcc/testsuite/gcc.dg/vect/pr71264.c (revision 241395)
+++ gcc/testsuite/gcc.dg/vect/pr71264.c (working copy)
@@ -7,6 +7,8 @@ typedef uint8_t footype __attribute__((v
 void test(uint8_t *ptr, uint8_t *mask)
 {
   footype mv;
+  ptr = __builtin_assume_aligned (ptr, __alignof__ (footype));
+  mask = __builtin_assume_aligned (mask, __alignof__ (footype));
   __builtin_memcpy(, mask, sizeof(mv));
   for (unsigned i = 0; i < 16; i += 4)
 {
@@ -17,4 +19,4 @@ void test(uint8_t *ptr, uint8_t *mask)
 }
 }

-/* { dg-final { scan-tree-dump "vectorized 1 loops in function" "vect" } 
} */
+/* { dg-final { scan-tree-dump "vectorized 1 loops in function" "vect" { 
xfail sparc*-*-* } } } */

[Bug tree-optimization/71264] [5 Regression] ICE in convert_move

[ebotcazou at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264

--- Comment #21 from Eric Botcazou  ---
> It is supposed to be vectorized on targets w/o V4QImode support.
> Not sure if the list of targets that FAIL the testcase would be smaller
> if I change the scan to an explicit list of targets (x86)?

Not sure indeed, it passes on PowerPC & PowerPC64 too for example.

> Meanwhile just add more targets to the xfail -- it really is a missed
> optimization as I guess for example arm also has V4SImode.

There is no XFAIL currently but we can certainly add one.

[Bug tree-optimization/71264] [5 Regression] ICE in convert_move

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264

--- Comment #20 from rguenther at suse dot de  ---
On Sat, 22 Oct 2016, ebotcazou at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264
> 
> Eric Botcazou  changed:
> 
>What|Removed |Added
> 
>  CC||ebotcazou at gcc dot gnu.org
> 
> --- Comment #19 from Eric Botcazou  ---
> Richard, can you make the testcase compile-only if it cannot be vectorized in
> the end (except for x86 by chance apparently)?

It is supposed to be vectorized on targets w/o V4QImode support.
Not sure if the list of targets that FAIL the testcase would be smaller
if I change the scan to an explicit list of targets (x86)?

Meanwhile just add more targets to the xfail -- it really is a missed
optimization as I guess for example arm also has V4SImode.

(I tried to fix the vectorizer to deal with vector types in the IL
but failed to quickly handle it)

[Bug tree-optimization/71264] [5 Regression] ICE in convert_move

[ebotcazou at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71264

Eric Botcazou  changed:

   What|Removed |Added

 CC||ebotcazou at gcc dot gnu.org

--- Comment #19 from Eric Botcazou  ---
Richard, can you make the testcase compile-only if it cannot be vectorized in
the end (except for x86 by chance apparently)?