[Bug middle-end/71614] ppc __gcc_qmul is mis-optimized not to be commutative

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71614

--- Comment #5 from Andrew Pinski  ---
The other thing to note is:
https://gcc.gnu.org/ml/gcc-patches/2014-01/msg00157.html

[Bug middle-end/71614] ppc __gcc_qmul is mis-optimized not to be commutative

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71614

--- Comment #4 from Andrew Pinski  ---
The fix would then put a barrier to force GCC not to do contracting.

Something like:
  v = a*d;
  w = b*c;
  asm("":"+f"(v));
  asm("":"+f"(w));
  tau += v + w; /* Add in other second-order terms.  */

[Bug middle-end/71614] ppc __gcc_qmul is mis-optimized not to be commutative

[timshen at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71614

--- Comment #3 from Tim Shen  ---
(In reply to Andrew Pinski from comment #2)
> Two things, try -fno-strict-aliasing first.
> Since your main violates C aliasing rules.
> Second thing to try is -ffp-contract=off as I am suspecting:
>   v = a*d;
>   w = b*c;
>   tau += v + w;   /* Add in other second-order terms.  */
>   u = t + tau;
> 
> is using an fmadd here.

Passing -fno-strict-aliasing doesn't change the -O2 result; -ffp-contract=off
does give the consistent result as -O1.

What result is __gcc_qmul supposed to give? As a user I'd expect the
commutativity here.

[Bug middle-end/71614] ppc __gcc_qmul is mis-optimized not to be commutative

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71614

--- Comment #2 from Andrew Pinski  ---
Two things, try -fno-strict-aliasing first.
Since your main violates C aliasing rules.
Second thing to try is -ffp-contract=off as I am suspecting:
  v = a*d;
  w = b*c;
  tau += v + w; /* Add in other second-order terms.  */
  u = t + tau;

is using an fmadd here.

[Bug middle-end/71614] ppc __gcc_qmul is mis-optimized not to be commutative

[timshen at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71614

--- Comment #1 from Tim Shen  ---
> Suppose the attached file is a.c

s/a\.c/b.c