Fix ExistentialSpecializer to correctly cleanup storage.

[atrick]

Handle calling conventions and cleanups in all the places (hopefully).

• when ExistentialSpecializer copies the specialized concrete arg into the
  original existential value
• when ExistentialSpecializer generates a think
• when SILCombine substitutes concrete values in place of the opened
  existential.

One particularly nasty problem is the existential boxes need to be
destroyed. It is not ok to simply destroy their value. The "leaks"
tool does not catch this problem.

Ownership SIL will make this all much more robust.

Fixes rdar://problem/50595630 Multiple leaks detected - Swift Perf

[atrick]

@swift-ci test.

[atrick]

@swift-ci test source compatibility

[atrick]

@swift-ci benchmark

[swift-ci]

Performance: -O

TEST	OLD	NEW	DELTA	RATIO
Regression
SortLargeExistentials	6600	16300	+147.0%	0.40x
StackPromo	12800	14100	+10.2%	0.91x
Improvement
BucketSort	2531	150	-94.1%	16.87x
ProtocolDispatch2	207	76	-63.3%	2.72x
ObserverForwarderStruct	660	575	-12.9%	1.15x (?)
ObserverUnappliedMethod	1710	1510	-11.7%	1.13x (?)
DataAccessBytesMedium	47	43	-8.5%	1.09x (?)

Code size: -O

TEST	OLD	NEW	DELTA	RATIO
Regression
BucketSort.o	11050	11618	+5.1%	0.95x
NIOChannelPipeline.o	4218	4357	+3.3%	0.97x
StackPromo.o	1892	1939	+2.5%	0.98x
Improvement
ProtocolDispatch2.o	1940	1740	-10.3%	1.11x
ObserverForwarderStruct.o	3728	3648	-2.1%	1.02x

Performance: -Osize

TEST	OLD	NEW	DELTA	RATIO
Regression
SortLargeExistentials	6700	16200	+141.8%	0.41x
ProtocolDispatch2	206	326	+58.3%	0.63x
DropLastCountableRangeLazy	4	5	+25.0%	0.80x (?)
DropLastAnyCollection	28	33	+17.9%	0.85x (?)
StackPromo	13000	14500	+11.5%	0.90x
PrefixAnyCollection	81	88	+8.6%	0.92x (?)
PrefixWhileAnySequence	1194	1288	+7.9%	0.93x (?)
Improvement
BucketSort	2523	145	-94.3%	17.40x
SuffixCountableRangeLazy	5	4	-20.0%	1.25x (?)
ObserverForwarderStruct	710	585	-17.6%	1.21x (?)
DropFirstAnyCollection	103	86	-16.5%	1.20x
ObserverUnappliedMethod	1740	1520	-12.6%	1.14x (?)
ReversedDictionary2	239	217	-9.2%	1.10x (?)
DataAccessBytesMedium	58	54	-6.9%	1.07x (?)

Code size: -Osize

TEST	OLD	NEW	DELTA	RATIO
Regression
NIOChannelPipeline.o	3651	3963	+8.5%	0.92x
ProtocolDispatch2.o	1968	2084	+5.9%	0.94x
BucketSort.o	11159	11431	+2.4%	0.98x
Improvement
ObserverForwarderStruct.o	3908	3708	-5.1%	1.05x
ObserverUnappliedMethod.o	5563	5354	-3.8%	1.04x

How to read the data

The tables contain differences in performance which are larger than 8% and differences in code size which are larger than 1%.

If you see any unexpected regressions, you should consider fixing the
regressions before you merge the PR.

Noise: Sometimes the performance results (not code size!) contain false
alarms. Unexpected regressions which are marked with '(?)' are probably noise.
If you see regressions which you cannot explain you can try to run the
benchmarks again. If regressions still show up, please consult with the
performance team (@eeckstein).

Hardware Overview

  Model Name: Mac mini
  Model Identifier: Macmini8,1
  Processor Name: Intel Core i7
  Processor Speed: 3.2 GHz
  Number of Processors: 1
  Total Number of Cores: 6
  L2 Cache (per Core): 256 KB
  L3 Cache: 12 MB
  Memory: 64 GB

[atrick]

I cannot reproduce the SortLargeExistentials regression locally. It only showed up in Osize which usually indicates something wasn't inlined.

@rajbarik will you have a chance to review these changes?

[rajbarik]

@atrick Sorry, I got sidetracked with something and did not see the email until now. I will have a look at it today.