Hi!
On Thu, Nov 22, 2018 at 02:47:10PM -0500, Vladimir Makarov wrote:
> Today I committed a patch
>
> https://gcc.gnu.org/ml/gcc-patches/2018-11/msg01945.html
>
> But it makes gcc.target/powerpc/pr70669.c to fail. Here is the patch
> to fix the failure. The expected code assumes that variable r should
> get a general reg. I suspect the expectation is wrong. There are three
> possible choices for r: a general reg, a vsx reg, and memory.
>
> The cost of move of vsx to/from mem is 4, the cost of move of general
> reg of this mode to/from mem is 8, and the cost of move of general reg
> to/from vsx is 12.
>
> So approximately if r is assigned memory we have cost 8 * 2 (asm) + 4
> * 2 (plus) = 24 (when we use memory at *q), if r is assigned to vsx the
> cost is 12*2(asm) + 4(r=*q) = 28, and if r is assigned a general reg the
> cost is 12 * 2 (plus) + 8 (r = *q) = 32.
>
> ira-costs should choose mem although it chose a general reg before my
> patch (it is a big topic why it chose a general reg. In brief, it is
> because some inherent drawbacks of the current cost calculation
> algorithm). The patch I recently submitted solves some drawbacks and
> memory is chosen for r.
That sounds all correct. Thanks for fixing this!
> To generate the expected code with r in a general reg, I just decrease
> the number of getting r value into vsx reg by changing r + r onto -r.
>
> Is the following patch ok for trunk?
Yes, the test now again tests what it is meant to test (that you get
m[tf]vsrd from reloads). It's fine for trunk, thanks!
Segher
2018-11-22 Vladimir Makarov <vmaka...@redhat.com>
* gcc.target/powerpc/pr70669.c: Use unary minus instead of
addition.
