[polly] r223965 - Do not run dead code elimination by default

[Tobias Grosser]

On 10.12.2014 22:12, Tobias Grosser wrote:
> Author: grosser
> Date: Wed Dec 10 15:12:23 2014
> New Revision: 223965
>
> URL: http://llvm.org/viewvc/llvm-project?rev=223965&view=rev
> Log:
> Do not run dead code elimination by default
>
> The dead code elimination is a pass that looks very promising, but needs some
> more compile-time tuning before enabling it by default seems sensible.

Our LNT servers show no run-time regressions for 'clang -O3 -polly', 
meaning the dead code elimination we just disabled did not give any 
benefits yet. Zino reported benefits on their internal test suite, so it 
would be interesting to understand which kind of codes benefit (Maybe D 
where zero initializations are common?)

We see 29 compile-time reductions after dropping dce with the top 4 being:

Stanford/Puzzle

	-73.54%  1.85  0.49

Polybench/linear-algebra/solvers/dynprog/dynprog

	-30.87%  0.94  0.65

Polybench/datamining/correlation/correlation 		
	-21.81%  0.70  0.55

Polybench/linear-algebra/solvers/gramschmidt/gramschmidt
	-12.82%  0.47  0.41
  	
(http://llvm.org/perf/db_default/v4/nts/14590?num_comparison_runs=0&test_filter=&test_min_value_filter=&aggregation_fn=median&MW_confidence_lv=0.01&compare_to=14535&submit=Update

The following link gives a comparision of 'clang -O3' vs. 'clang -O3 
-polly' after this commit:

http://llvm.org/perf/db_default/v4/nts/14590?compare_to=14581

this link is just before:

http://llvm.org/perf/db_default/v4/nts/14535?compare_to=14581

the compile time cost for the most costly benchmarks has decreased, but 
there is still some room for improvement.

Cheers,
Tobias