[opt] load promotion in mandatory combine

[zoecarver]

This patch adds support for load promotion in mandatory combine. This is from #28536 and will be needed (or similar) to support all the optimizations done by closure cleanup in mandatory inlining.

For example:

store x to as
y = load as
use y

Turns into (minus copies):

store x to as
use x

I plan to also add copy/destroy and store elimination in future patches which should help reduce the dead noise left behind.

This is one of the biggest single patches from #28536 the rest will be smaller (in size and scope). A lot of the logic here could also apply to mem2reg and #30308.

[zoecarver]

Implemented in #30464.

[zoecarver]

@swift-ci please benchmark.

[swift-ci]

Performance: -O

Regression	OLD	NEW	DELTA	RATIO
FlattenListLoop	2899	3307	+14.1%	0.88x (?)
FlattenListFlatMap	5656	6234	+10.2%	0.91x (?)
PrefixArrayLazy	13	14	+7.7%	0.93x (?)

Code size: -O

Performance: -Osize

Code size: -Osize

Performance: -Onone

Code size: -swiftlibs

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

[gottesmm]

I don't think it makes sense to do this at -Onone. It would make more sense to do a small bit of analysis to recognize the specific case. We need to be really careful at -Onone with compile time. @atrick your thoughts?

[zoecarver]

That's a good point. Two alternatives: a) only scan single blocks (which would greatly reduce the complexity here) and b) only do this for parital_applys at the same time as the other optimizations. If we want to support everything from closure cleanup we would have to do (at least) the second one.

[zoecarver]

@swift-ci please smoke test compiler performance.

[zoecarver]

@swift-ci please smoke test compiler performance.

[zoecarver]

@swift-ci please smoke test compiler performance

[swift-ci]

Summary for master smoketest

No regressions above thresholds

• Debug
  • debug brief
  • debug detailed
• Release
  • release brief
  • release detailed

Debug

debug brief

Regressed (0)

name	old	new	delta	delta_pct

Improved (0)

name	old	new	delta	delta_pct

Unchanged (delta < 1.0% or delta < 100.0ms) (3)

name	old	new	delta	delta_pct
Frontend.NumInstructionsExecuted	1,019,638,163,967	1,020,095,119,703	456,955,736	0.04%
LLVM.NumLLVMBytesOutput	54,130,888	54,130,896	8	0.0%
time.swift-driver.wall	80.4s	81.1s	695.8ms	0.87%

debug detailed

Regressed (0)

name	old	new	delta	delta_pct

Improved (0)

name	old	new	delta	delta_pct

Unchanged (delta < 1.0% or delta < 100.0ms) (17)

name	old	new	delta	delta_pct
AST.NumLoadedModules	5,721	5,721	0	0.0%
AST.NumTotalClangImportedEntities	54,209	53,733	-476	-0.88%
IRModule.NumIRBasicBlocks	350,303	350,303	0	0.0%
IRModule.NumIRFunctions	99,881	99,881	0	0.0%
IRModule.NumIRGlobals	111,354	111,354	0	0.0%
IRModule.NumIRInsts	2,590,079	2,590,023	-56	-0.0%
IRModule.NumIRValueSymbols	192,568	192,568	0	0.0%
LLVM.NumLLVMBytesOutput	54,130,888	54,130,896	8	0.0%
SILModule.NumSILGenFunctions	49,539	49,539	0	0.0%
SILModule.NumSILOptFunctions	67,926	67,926	0	0.0%
Sema.NumConformancesDeserialized	190,801	188,982	-1,819	-0.95%
Sema.NumConstraintScopes	1,006,545	1,006,339	-206	-0.02%
Sema.NumDeclsDeserialized	1,504,955	1,489,847	-15,108	-1.0%
Sema.NumGenericSignatureBuilders	26,151	26,076	-75	-0.29%
Sema.NumLazyIterableDeclContexts	177,206	176,465	-741	-0.42%
Sema.NumTypesDeserialized	476,592	474,487	-2,105	-0.44%
Sema.NumTypesValidated	31,198	31,198	0	0.0%

Release

release brief

Regressed (0)

name	old	new	delta	delta_pct

Improved (0)

name	old	new	delta	delta_pct

Unchanged (delta < 1.0% or delta < 100.0ms) (3)

name	old	new	delta	delta_pct
Frontend.NumInstructionsExecuted	1,395,320,547,280	1,396,105,749,659	785,202,379	0.06%
LLVM.NumLLVMBytesOutput	55,688,200	55,688,020	-180	-0.0%
time.swift-driver.wall	141.2s	141.6s	382.9ms	0.27%

release detailed

Regressed (0)

name	old	new	delta	delta_pct

Improved (0)

name	old	new	delta	delta_pct

Unchanged (delta < 1.0% or delta < 100.0ms) (17)

name	old	new	delta	delta_pct
AST.NumLoadedModules	552	552	0	0.0%
AST.NumTotalClangImportedEntities	14,604	14,604	0	0.0%
IRModule.NumIRBasicBlocks	182,221	182,221	0	0.0%
IRModule.NumIRFunctions	76,522	76,522	0	0.0%
IRModule.NumIRGlobals	89,717	89,717	0	0.0%
IRModule.NumIRInsts	1,465,156	1,465,156	0	0.0%
IRModule.NumIRValueSymbols	155,325	155,325	0	0.0%
LLVM.NumLLVMBytesOutput	55,688,200	55,688,020	-180	-0.0%
SILModule.NumSILGenFunctions	29,229	29,229	0	0.0%
SILModule.NumSILOptFunctions	24,856	24,856	0	0.0%
Sema.NumConformancesDeserialized	78,928	78,928	0	0.0%
Sema.NumConstraintScopes	989,933	989,933	0	0.0%
Sema.NumDeclsDeserialized	198,327	198,327	0	0.0%
Sema.NumGenericSignatureBuilders	5,235	5,235	0	0.0%
Sema.NumLazyIterableDeclContexts	24,195	24,195	0	0.0%
Sema.NumTypesDeserialized	107,360	107,360	0	0.0%
Sema.NumTypesValidated	18,705	18,705	0	0.0%

[zoecarver]

Some todos:

• use dominance analysis instead of dominance info.
• move into another pass, not mandatory combine.

[shahmishal]

Please update the base branch to main by Oct 5th otherwise the pull request will be closed automatically.

• How to change the base branch: (Link)
• More detail about the branch update: (Link)