[PATCH] Remove inefficient branchless conditional negate optimization

[Richard Biener]

On Thu, Feb 26, 2015 at 11:20 PM, Jeff Law <law@redhat.com> wrote:
> On 02/26/15 10:30, Wilco Dijkstra wrote:
>>
>> Several GCC versions ago a conditional negate optimization was introduced
>> as a workaround for
>> PR45685. However the branchless expansion for conditional negate is
>> extremely inefficient on most
>> targets (5 sequentially dependent instructions rather than 2 on AArch64).
>> Since the underlying issue
>> has been resolved (the example in PR45685 no longer generates a branch on
>> x64), remove the
>> workaround so that conditional negates are treated in exactly the same way
>> as conditional invert,
>> add, subtract, and, orr, xor etc. Simple example:
>>
>> int f(int x) { if (x > 3) x = -x; return x; }
>
> You need to bootstrap and regression test the change before it can be
> approved.

As Jeff added a testcase for the PHI opt transform to happen I'm sure
testing would shown this as fallout.

> You should turn this little example into a testcase.  It's fine with me if
> this new test is ARM specific.
>
>
> You should also find a way to change the test gcc.dg/tree-ssa/pr45685.c in
> such a way that it ensures there aren't any undesirable branches.

I'd be also interested in results of vectorizing a loop with a
conditional negate.
I can very well imagine reverting this patch causing code quality regressions
there.

> I've got enough history to know this is fixing a regression of sorts for the
> ARM platform.  So once the issues above are addressed it can go forward even
> without a BZ noting the regression.

But I'd say this is stage1 material at this point.

Richard.

> jeff
>