Feature #9997: Specifying non volatile registers for increase performence in ppc64

• Status: Closed
• Assignee: lbianc (Leonardo Bianconi)
• Target version: 2.2.0

[lbianc (Leonardo Bianconi)]

To increase performance on ppc64 machines, the registers 14 and 15 were specified for the architecture, just like for x86_64 machines.

I've executed some benchmarks, see below:

bm_loop_for.rb
Before After
real 0m2.524s real 0m2.781s
user 0m2.520s user 0m2.777s
sys 0m0.004s sys 0m0.004s

bm_loop_whileloop.rb
Before After
real 0m1.142s real 0m1.267s
user 0m1.138s user 0m1.263s
sys 0m0.004s sys 0m0.004s

bm_so_array.rb
Before After
real 0m1.556s real 0m1.832s
user 0m1.552s user 0m1.828s
sys 0m0.004s sys 0m0.004s

bm_so_matrix.rb
Before After
real 0m1.070s real 0m1.107s
user 0m1.066s user 0m1.104s
sys 0m0.004s sys 0m0.004s

[lbianc (Leonardo Bianconi)]

Sending the benchmark results again, the previous was wrong, sorry.

bm_loop_for.rb
Before
real 0m2.781s
user 0m2.777s
sys 0m0.004s

After
real 0m2.524s
user 0m2.520s
sys 0m0.004s

bm_loop_whileloop.rb
Before
real 0m1.267s
user 0m1.263s
sys 0m0.004s

After
real 0m1.142s
user 0m1.138s
sys 0m0.004s

bm_so_array.rb
Before
real 0m1.832s
user 0m1.828s
sys 0m0.004s

After
real 0m1.556s
user 0m1.552s
sys 0m0.004s

bm_so_matrix.rb
Before
real 0m1.107s
user 0m1.104s
sys 0m0.004s

After
real 0m1.070s
user 0m1.066s
sys 0m0.004s

[nobu (Nobuyoshi Nakada)]

The first results are inverted?

[lbianc (Leonardo Bianconi)]

• File Ruby_ppc64_benchmark.pdf Ruby_ppc64_benchmark.pdf added

Hi Nobuyoshi, yes, my mistake. The correct results are the ones I added later.
I'm attaching the results for the "Computer Language Benchmarks Game" as well. Check the time enhancement.

[lbianc (Leonardo Bianconi)]

ping... anyone to review this patch?

[hsbt (Hiroshi SHIBATA)]

• Status changed from Open to Closed
• % Done changed from 0 to 100

Applied in changeset r47149.

---

• vm_exec.c: improve performance in ppc64 arch.
  [ruby-core:63437] [Feature #9997]