[PATCH] D18226: Codegen: Tail-duplicate during placement.

[Quentin Colombet via llvm-commits]

qcolombet added a comment.

Hi Kyle,

I have some early results (see below) and I see a bunch of regressions on the llvm test-suite already. I thought you already benchmarked the test-sutie, so I am wondering if I missed something in my run because some of the regressions look bad.
The base line is O3 without your patch and I compare with O3 -mllvm  -outline-optional-branches -mllvm -tail-dup-size=3 with your patch.

Let me know if you need more information to reproduce.

I am waiting for your feedbacks to know if I am doing something wrong because going further.

Cheers,
-Q

Run-Over-Run Changes Detail

Performance Regressions - Execution Time	Δ 	Previous	Current	σ 
SingleSource/Benchmarks/Shootout/fib2 http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.296=3	28.43%	1.7542	2.2530	0.0207
SingleSource/Benchmarks/BenchmarkGame/recursive http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.202=3	8.65%	0.6450	0.7008	0.0021
MultiSource/Applications/lambda-0_1_3/lambda http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.33=3	6.20%	3.3535	3.5614	0.0138
MultiSource/Benchmarks/Trimaran/netbench-url/netbench-url http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.253=3	4.80%	3.8169	4.0002	0.0212
SingleSource/Benchmarks/Misc/lowercase http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.312=3	4.74%	2.5230	2.6425	0.0026
SingleSource/Benchmarks/Shootout/heapsort http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.83=3	2.48%	2.5172	2.5796	0.0025

MultiSource/Benchmarks/Trimaran/enc-rc4/enc-rc4 http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.171=3	1.86%	0.8480	0.8638	0.0034
SingleSource/Benchmarks/CoyoteBench/huffbench http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.160=3	1.42%	12.0151	12.1854	0.0026

[…]

Performance Improvements - Execution Time	Δ 	Previous	Current	σ 
MultiSource/Benchmarks/Ptrdist/anagram/anagram http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.92=3	-11.35%	0.9561	0.8476	0.0101
SingleSource/Benchmarks/Shootout/methcall http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.16=3	-6.15%	4.1681	3.9117	0.0448
SingleSource/Benchmarks/Misc/ReedSolomon http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.270=3	-4.89%	4.4497	4.2320	0.0199
SingleSource/Benchmarks/Misc/salsa20 http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.61=3	-4.85%	6.3418	6.0343	0.0021
MultiSource/Applications/aha/aha http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.233=3	-4.51%	2.0488	1.9564	0.0049
MultiSource/Benchmarks/TSVC/InductionVariable-flt/InductionVariable-flt http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.231=3	-2.59%	2.7844	2.7124	0.0228

SingleSource/Benchmarks/Polybench/linear-algebra/kernels/syrk/syrk http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.177=3	-1.98%	1.3657	1.3387	0.0082

MultiSource/Applications/JM/lencod/lencod http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.286=3	-1.31%	5.0791	5.0127	0.0139

SingleSource/Benchmarks/Shootout/sieve http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.278=3	-1.14%	4.4957	4.4446	0.0080
MultiSource/Benchmarks/Fhourstones-3_1/fhourstones3_1 http://qcolombet-haswell.apple.com/perf/v4/nts/2/graph?test.239=3	-1.09%	1.0943	1.0824	0.0037


http://reviews.llvm.org/D18226