[Bug tree-optimization/103797] Clang vectorized LightPixel while GCC does not

[hubicka at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103797

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW

--- Comment #7 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
OK, here is completely fake testcase that does similar operaitons:

#include <math.h>
struct test {float x; float y; float z;} test;
float f;
void
t()
{
  float x = test.x;
  float y = test.y;
  float z = test.z;

  x = x * f;
  y = y * f;
  z = z * f;
  x = sqrt (x);
  y = sqrt (y);
  z = sqrt (z);
  x = x / f;
  y = y / f;
  z = z / f;
  test.x=x;
  test.y=y;
  test.z=z;
}

We seem to fail to vectorize it with:

t.c:20:9: missed:   op not supported by target.                                 
t.c:17:5: missed:   not vectorized: relevant stmt not supported: x_15 = x_24 /
f.0_1;

clang seems to use divps happilly, so I am not sure why it is not supported.
Even more funny is that with -Ofast it is compiled into multiplication by
reciprocal:

t:
.LFB0:
        .cfi_startproc
        movss   f(%rip), %xmm4
        movss   .LC0(%rip), %xmm2
        movss   test(%rip), %xmm0
        movss   test+4(%rip), %xmm3
        divss   %xmm4, %xmm2
        movss   test+8(%rip), %xmm1
        mulss   %xmm4, %xmm0
        mulss   %xmm4, %xmm3
        mulss   %xmm4, %xmm1
        sqrtss  %xmm0, %xmm0
        sqrtss  %xmm3, %xmm3
        sqrtss  %xmm1, %xmm1
        mulss   %xmm2, %xmm0
        mulss   %xmm2, %xmm3
        mulss   %xmm2, %xmm1
        unpcklps        %xmm3, %xmm0
        movlps  %xmm0, test(%rip)
        movss   %xmm1, test+8(%rip)
        ret


and rewriting it that way by hand:

#include <math.h>
struct test {float x; float y; float z;} test;
float f;
void
t()
{
  float x = test.x;
  float y = test.y;
  float z = test.z;
  float m = 1/f;

  x = x * f;
  y = y * f;
  z = z * f;
  x = sqrt (x);
  y = sqrt (y);
  z = sqrt (z);
  x = x * m;
  y = y * m;
  z = z * m;
  test.x=x;
  test.y=y;
  test.z=z;
}

gets the expected result:
t:
.LFB0:
        .cfi_startproc
        movss   f(%rip), %xmm0
        movq    test(%rip), %xmm1
        movaps  %xmm0, %xmm2
        shufps  $0xe0, %xmm2, %xmm2
        mulps   %xmm1, %xmm2
        movss   .LC0(%rip), %xmm1
        divss   %xmm0, %xmm1
        mulss   test+8(%rip), %xmm0
        sqrtps  %xmm2, %xmm2
        sqrtss  %xmm0, %xmm0
        movaps  %xmm1, %xmm3
        shufps  $0xe0, %xmm3, %xmm3
        mulss   %xmm0, %xmm1
        mulps   %xmm3, %xmm2
        movss   %xmm1, test+8(%rip)
        movlps  %xmm2, test(%rip)
        ret
        .cfi_endproc

Having this however I do not see slp analyzing the divide in the original code
at all.