[Bug tree-optimization/57579] Problem with vectorization

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57579

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |INVALID

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Federico Carminati from comment #0)
> Good evening,
>    all my apologises if this is a stupid question, however I have a terribly
> simple loop
> 
> 
> include <math.h>
> 
> typedef struct
> {
>    double x,y,z;
>    double dummy;
> } P;
> 
> 
> void foo(const P * __restrict__ points, double * __restrict__ d)
> {
>    for(int i=0;i<100;++i) {
>       d[i]=points[i].x*points[i].x*points[i].x*points[i].x+
> 	 points[i].y*points[i].y*points[i].y+
> 	 points[i].z*points[i].z*points[i].z;
>    }
> }
>  
> if I compile with
> 
> /opt/gcc-4.8.1/bin/g++ -c -std=c++0x -O3 -msse4.1 -Wall -Wstrict-aliasing=2
> -ftree-vectorizer-verbose=2
> 
> I obtain the diagnostic at the bottom of the page. Is there a place where I
> can find an explanation for the g++ vectorizer diagnostic messages? I
> frankly do not understand what it is talking about and google does not seem
> to be willing to help me this time. Any help in deciphering these messages
> would be greatly appreciated. Thanks and sorry for the bother
> 
> Analyzing loop at testvec1.cxx:12
> 
> testvec1.cxx:12: note: Data access with gaps requires scalar epilogue loop
> testvec1.cxx:12: note: vector alignment may not be reachable
> testvec1.cxx:12: note: virtual phi. skip.
> testvec1.cxx:12: note: not ssa-name.
> testvec1.cxx:12: note: use not simple.
> testvec1.cxx:12: note: not ssa-name.
> testvec1.cxx:12: note: use not simple.
> testvec1.cxx:12: note: no array mode for V2DF[4]
> testvec1.cxx:12: note: not ssa-name.
> testvec1.cxx:12: note: use not simple.
> testvec1.cxx:12: note: not ssa-name.
> testvec1.cxx:12: note: use not simple.
> testvec1.cxx:12: note: no array mode for V2DF[4]
> testvec1.cxx:12: note: not ssa-name.
> testvec1.cxx:12: note: use not simple.
> testvec1.cxx:12: note: not ssa-name.
> testvec1.cxx:12: note: use not simple.
> testvec1.cxx:12: note: no array mode for V2DF[4]
> 
> Vectorizing loop at testvec1.cxx:12

^^^^

the loop is vectorized.

> testvec1.cxx:12: note: virtual phi. skip.
> testvec1.cxx:12: note: no array mode for V2DF[4]
> testvec1.cxx:12: note: no array mode for V2DF[4]
> testvec1.cxx:12: note: no array mode for V2DF[4]
> testvec1.cxx:10: note: vectorized 1 loops in function.
> 
> testvec1.cxx:10: note: not vectorized: not enough data-refs in basic block.
> 
> testvec1.cxx:13: note: not vectorized: no vectype for stmt: vect_var_.10_29
> = MEM[(const struct P *)vect_ppoints.6_31];
>  scalar_type: const vector(2) double
> testvec1.cxx:13: note: Failed to SLP the basic block.
> testvec1.cxx:13: note: not vectorized: failed to find SLP opportunities in
> basic block.
> 
> testvec1.cxx:13: note: Build SLP failed: grouped loads have gaps _59 =
> _60->x;
> 
> testvec1.cxx:13: note: Failed to SLP the basic block.
> testvec1.cxx:13: note: not vectorized: failed to find SLP opportunities in
> basic block.
> 
> testvec1.cxx:10: note: not vectorized: not enough data-refs in basic block.

the rest of the "complaints" are from basic-block vectorization trying to
vectorize the loop body again.  Which of course fails, it is already
vectorized.