SSAIR: improve inlining performance with in-place IR-inflation

[aviatesk]

This commit improves the performance of a huge hot-spot within inflate_ir
by using the in-place version of it (inflate_ir!) and avoiding some
unnecessary allocations.
For NativeInterpreter, CodeInfo-IR passed to inflate_ir can come
from two ways:

1. from global cache: uncompressed from compressed format
2. from local cache: inferred CodeInfo as-is managed by InferenceResult

And in the case of 1, an uncompressed CodeInfo is an newly-allocated
object already and thus we can use the in-place version safely. And it
turns out that this helps us avoid many unnecessary allocations.
The original non-destructive inflate_ir remains there for testing or
interactive purpose.

---

@nanosoldier runbenchmarks("inference", vs=":master")

[nanosoldier]

Your benchmark job has completed - no performance regressions were detected. A full report can be found here.

[Keno]

This seems reasonable to me. I usually profile with Cthulhu when I profile inference, which doesn't go through the compression/decompression cycle, so I don't have a good idea of the usual performance impact, but the nanosolider results look favorable, so I'm in favor. My only concern would be ending up with corrupted CodeInfos in some of the generated function cases. I've run into hard to debug bugs of that kind before, but certainly for the compression case, it's a no brainer to do the inflation destructively. Eventually we may want to even skip the decompression step and just inline directly from the compressed representation as we go along, but that's obviously future work.

[aviatesk]

My only concern would be ending up with corrupted CodeInfos in some of the generated function cases.

For the uncompressed case, this PR makes a copy of that CodeInfo and so I think it is also safe. Maybe CodeInstance.inferred manages CodeInfo directly for generated function case? (EDIT: that case happens for external AbstractInterpreters whose may_compress return false) It'd be handled in the same way anyway though.

julia/base/compiler/ssair/inlining.jl

Lines 916 to 926 in f34c577

	function InliningTodo(mi::MethodInstance, src::Union{CodeInfo, Vector{UInt8}}, effects::Effects)
	if !isa(src, CodeInfo)
	src = ccall(:jl_uncompress_ir, Any, (Any, Ptr{Cvoid}, Any), mi.def, C_NULL, src::Vector{UInt8})::CodeInfo
	else
	src = copy(src)
	end
	@timeit "inline IR inflation" begin
	ir = inflate_ir!(src, mi)::IRCode
	return InliningTodo(mi, ResolvedInliningSpec(ir, effects))
	end
	end

---

@nanosoldier runtests(ALL, vs = ":master")

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected. A full report can be found here.

[aviatesk]

@nanosoldier runtests(["AStarSearch", "ArchGDAL", "AugmentedGPLikelihoods", "AutocorrelationShell", "BigO", "BioStructures", "ClimateModels", "CollectiveSpins", "ConceptnetNumberbatch", "CovarianceMatrices", "CrypticCrosswords", "CryptoGroups", "DarkCurves", "ExactOptimalTransport", "FSimZoo", "Faker", "Ferrite", "GrayCoding", "Individual", "InfrastructureSystems", "Ipaper", "JSONSchema", "Karnak", "MDInclude", "MixedModelsPermutations", "NonconvexIpopt", "OhMyREPL", "PoissonRandom", "PolynomialBases", "PowerSimulations", "QuantumAlgebra", "SimpleGraphs", "SimpleTweaks", "SkipLists", "StringDistances", "Surrogates", "ToeplitzMatrices", "VertexFinder", "VoxelRayTracers", "BasicInterpolators", "BayesianQuadrature", "BitSAD", "BundlerIO", "CiteEXchange", "Evolutionary", "GraphMLDatasets", "JuliaCon", "Probably", "QuantumTomography", "ReadVTK", "Retriever", "SBML", "StochasticDelayDiffEq", "ImageTracking", "Tapestree"], vs = ":master")

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected. A full report can be found here.

[aviatesk]

@nanosoldier runtests(["JSONSchema", "ToeplitzMatrices", "Tapestree"], vs = ":master")

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected. A full report can be found here.

[aviatesk]

Just confirmed the JSONSchema test suite runs successfully on my local machine on this branch. Going to merge.