[patch] Add more loop simplification passes to LTO

[Duncan P. N. Exon Smith]

On 2014 Apr 10, at 13:25, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:

> On Apr 9, 2014, at 10:10, Rafael Espíndola <rafael.espindola at gmail.com> wrote:
> 
>> Some coworkers were trying LTO and noticed that not as many loops
>> where being optimized when a function was inlined in LTO instead of
>> during at regular build when the function was made inline.
>> 
>> In the end, the problem was really simple, we are not running indvars
>> or -loop-deletion in lto. This patch then seems fairly analogous to
>> r202051.
>> 
>> I haven't run the test-suite to benchmark this (I don't have in my
>> macbook air), but do you guys think this patch is reasonable? If so I
>> can benchmark and commit it when I get home (April 20th) if no one
>> beats me to it .
>> 
> 
> Nice find!
> 
> I plan to benchmark this over the weekend... if someone gets to it
> first, let me know.

Gerolf volunteered offline to check results on ARM64, but in the
meantime I ran LNT --large with 5 sample points on x86-64 using
r206149 and -O3 -flto:

Performance Regressions - Execution Time:
Δ       Previous Current  σ       Benchmark
  3.79%  2.1767   2.2591  0.0052  MultiSource/Benchmarks/TSVC/InductionVariable-flt/InductionVariable-flt
  3.74%  3.5570   3.6901  0.0084  SingleSource/Benchmarks/CoyoteBench/almabench
  1.81% 16.0277  16.3184  0.0850  MultiSource/Applications/lua/lua

Performance Improvements - Execution Time:
Δ       Previous Current  σ       Benchmark
-19.91%  0.4034   0.3231  0.0013  SingleSource/Benchmarks/Dhrystone/dry
-14.11%  0.8462   0.7268  0.0007  MultiSource/Benchmarks/Ptrdist/ks/ks
 -7.51%  4.9761   4.6026  0.0060  MultiSource/Applications/hexxagon/hexxagon
 -4.25%  1.6009   1.5328  0.0011  MultiSource/Benchmarks/TSVC/ControlLoops-flt/ControlLoops-flt
 -3.49%  2.5562   2.4670  0.0019  MultiSource/Benchmarks/TSVC/Reductions-dbl/Reductions-dbl
 -3.14%  2.1693   2.1011  0.0061  MultiSource/Benchmarks/TSVC/ControlLoops-dbl/ControlLoops-dbl
 -1.88%  4.8174   4.7267  0.0051  MultiSource/Benchmarks/TSVC/Reductions-flt/Reductions-flt
 -1.31%  0.8478   0.8367  0.0028  MultiSource/Benchmarks/TSVC/Equivalencing-flt/Equivalencing-flt
 -1.29%  2.9246   2.8869  0.0052  MultiSource/Benchmarks/TSVC/ControlFlow-dbl/ControlFlow-dbl
 -1.28%  2.9317   2.8941  0.0044  MultiSource/Benchmarks/sim/sim
 -1.07%  1.2226   1.2095  0.0032  MultiSource/Applications/viterbi/viterbi

Results look good to me.  If no one has any objections, I'll
commit Rafael's patch when I get a chance.