-
-
Notifications
You must be signed in to change notification settings - Fork 5.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Compiler support for optimizing PersistentDict #51993
Conversation
LGTM |
this is really cool. A little bit scary but I prefer it over moving the definition into the compiler wholesale. |
What are the consequences of actions that break interface assumptions? julia> dict = Base.PersistentDict(0=>false)
Base.PersistentDict{Int64, Bool} with 1 entry:
0 => 0
julia> dict.trie.data[1] = Base.HashArrayMappedTries.Leaf{Int64, Bool}(0, true)
Base.HashArrayMappedTries.Leaf{Int64, Bool}(0, true)
julia> dict
Base.PersistentDict{Int64, Bool} with 1 entry:
0 => 1 Obviously a user "shouldn't" do this, but would this now trigger UB? |
Yes, any misuse is undefined behavior. There are some basic sanity checks that avoid applying the optimization if you things are obviously wrong, but we can't really say anything else. |
I've given this some further thought, and I don't think the callback design I have here is all that good. I'm planning to rejigger this with an optional return for the getter, having it return |
#length(stmt.args) == 4 || return | ||
|
||
collection = stmt.args[end-1] | ||
key = stmt.args[end] |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think we need some kind of a mutability check on key
, since PersistentDict allows mutable keys but does not behave very sensibly for them.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can you expand what you mean by doesn't behave sensible? Or open a separate issue?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The main problem for this optimization is:
julia> x = Int64[0];
julia> d = Base.PersistentDict(x => false, [1] => true);
julia> x[1] = 1
julia> d[x]
true
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We could make it an IdDict equivalent.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
That still leaves types like IdDict
out in the cold, since it is an immutable struct, that references mutable contents to compute the hash. There is isidentityfree
that solves that issue though, if desired.
base/compiler/ssair/passes.jl
Outdated
leaves::Vector{Any}, 𝕃ₒ::AbstractLattice) | ||
# For every leaf, the lifted value | ||
lifted_leaves = LiftedLeaves() | ||
maybe_undef = false |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
maybe_undef = false |
base/compiler/ssair/passes.jl
Outdated
lifted_result = lift_leaves_keyvalue(compact, key, leaves, 𝕃ₒ) | ||
lifted_result === nothing && return | ||
lifted_leaves, any_undef = lifted_result |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
lifted_result = lift_leaves_keyvalue(compact, key, leaves, 𝕃ₒ) | |
lifted_result === nothing && return | |
lifted_leaves, any_undef = lifted_result | |
lifted_leaves = lift_leaves_keyvalue(compact, key, leaves, 𝕃ₒ) | |
lifted_leaves === nothing && return |
What is the main reason for this? Is it because it makes debugging difficult when users are developing code to utilize this interface? |
Rebased against the latest master. |
The callbacks are a bit awkward, since it's not clear when to infer them, if you want to optimize them. |
Review addressed, rebased, and updated in line with what I said I would do above. Needs #52193 to be sound, but other than that, I don't see much reason not to try this. The one remaining concern is that because of the |
As discussed in #51352, this gives `EnterNode` the ability to set (and restore on leave or catch edge) jl_current_task->scope. Manual modifications of the task field after the task has started are considered undefined behavior. In addition, we gain a new intrinsic to access current_task->scope and both inference and the optimizer will forward scopes from EnterNodes to this intrinsic (non-interprocedurally). Together with #51993 this is sufficient to fully optimize ScopedValues (non-interprocedurally at least).
As discussed in #51352, this gives `EnterNode` the ability to set (and restore on leave or catch edge) jl_current_task->scope. Manual modifications of the task field after the task has started are considered undefined behavior. In addition, we gain a new intrinsic to access current_task->scope and both inference and the optimizer will forward scopes from EnterNodes to this intrinsic (non-interprocedurally). Together with #51993 this is sufficient to fully optimize ScopedValues (non-interprocedurally at least).
This is part of the work to address #51352 by attempting to allow the compiler to perform SRAO on persistent data structures like `PersistentDict` as if they were regular immutable data structures. These sorts of data structures have very complicated internals (with lots of mutation, memory sharing, etc.), but a relatively simple interface. As such, it is unlikely that our compiler will have sufficient power to optimize this interface by analyzing the implementation. We thus need to come up with some other mechanism that gives the compiler license to perform the requisite optimization. One way would be to just hardcode `PersistentDict` into the compiler, optimizing it like any of the other builtin datatypes. However, this is of course very unsatisfying. At the other end of the spectrum would be something like a generic rewrite rule system (e-graphs anyone?) that would let the PersistentDict implementation declare its interface to the compiler and the compiler would use this for optimization (in a perfect world, the actual rewrite would then be checked using some sort of formal methods). I think that would be interesting, but we're very far from even being able to design something like that (at least in Base - experiments with external AbstractInterpreters in this direction are encouraged). This PR tries to come up with a reasonable middle ground, where the compiler gets some knowledge of the protocol hardcoded without having to know about the implementation details of the data structure. The basic ideas is that `Core` provides some magic generic functions that implementations can extend. Semantically, they are not special. They dispatch as usual, and implementations are expected to work properly even in the absence of any compiler optimizations. However, the compiler is semantically permitted to perform structural optimization using these magic generic functions. In the concrete case, this PR introduces the `KeyValue` interface which consists of two generic functions, `get` and `set`. The core optimization is that the compiler is allowed to rewrite any occurrence of `get(set(x, k, v), k)` into `v` without additional legality checks. In particular, the compiler performs no type checks, conversions, etc. The higher level implementation code is expected to do all that. This approach closely matches the general direction we've been taking in external AbstractInterpreters for embedding additional semantics and optimization opportunities into Julia code (although we generally use methods there, rather than full generic functions), so I think we have some evidence that this sort of approach works reasonably well. Nevertheless, this is certainly an experiment and the interface is explicitly declared unstable. This is fully working and implemented, but the optimization currently bails on anything but the simplest cases. Filling all those cases in is not particularly hard, but should be done along with a more invasive refactoring of SROA, so we should figure out the general direction here first and then we can finish all that up in a follow-up cleanup. Before: ``` julia> using BenchmarkTools julia> function foo() a = Base.PersistentDict(:a => 1) return a[:a] end foo (generic function with 1 method) julia> @benchmark foo() BenchmarkTools.Trial: 10000 samples with 993 evaluations. Range (min … max): 32.940 ns … 28.754 μs ┊ GC (min … max): 0.00% … 99.76% Time (median): 49.647 ns ┊ GC (median): 0.00% Time (mean ± σ): 57.519 ns ± 333.275 ns ┊ GC (mean ± σ): 10.81% ± 2.22% ▃█▅ ▁▃▅▅▃▁ ▁▃▂ ▂ ▁▂▄▃▅▇███▇▃▁▂▁▁▁▁▁▁▁▁▂▂▅██████▅▂▁▁▁▁▁▁▁▁▁▁▂▃▃▇███▇▆███▆▄▃▃▂▂ ▃ 32.9 ns Histogram: frequency by time 68.6 ns < Memory estimate: 128 bytes, allocs estimate: 4. julia> @code_typed foo() CodeInfo( 1 ─ %1 = invoke Vector{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}}(Base.HashArrayMappedTries.undef::UndefInitializer, 1::Int64)::Vector{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} │ %2 = %new(Base.HashArrayMappedTries.HAMT{Symbol, Int64}, %1, 0x00000000)::Base.HashArrayMappedTries.HAMT{Symbol, Int64} │ %3 = %new(Base.HashArrayMappedTries.Leaf{Symbol, Int64}, :a, 1)::Base.HashArrayMappedTries.Leaf{Symbol, Int64} │ %4 = Base.getfield(%2, :data)::Vector{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} │ %5 = $(Expr(:boundscheck, true))::Bool └── goto #5 if not %5 2 ─ %7 = Base.sub_int(1, 1)::Int64 │ %8 = Base.bitcast(UInt64, %7)::UInt64 │ %9 = Base.getfield(%4, :size)::Tuple{Int64} │ %10 = $(Expr(:boundscheck, true))::Bool │ %11 = Base.getfield(%9, 1, %10)::Int64 │ %12 = Base.bitcast(UInt64, %11)::UInt64 │ %13 = Base.ult_int(%8, %12)::Bool └── goto #4 if not %13 3 ─ goto #5 4 ─ %16 = Core.tuple(1)::Tuple{Int64} │ invoke Base.throw_boundserror(%4::Vector{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}}, %16::Tuple{Int64})::Union{} └── unreachable 5 ┄ %19 = Base.getfield(%4, :ref)::MemoryRef{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} │ %20 = Base.memoryref(%19, 1, false)::MemoryRef{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} │ Base.memoryrefset!(%20, %3, :not_atomic, false)::MemoryRef{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} └── goto #6 6 ─ %23 = Base.getfield(%2, :bitmap)::UInt32 │ %24 = Base.or_int(%23, 0x00010000)::UInt32 │ Base.setfield!(%2, :bitmap, %24)::UInt32 └── goto #7 7 ─ %27 = %new(Base.PersistentDict{Symbol, Int64}, %2)::Base.PersistentDict{Symbol, Int64} └── goto #8 8 ─ %29 = invoke Base.getindex(%27::Base.PersistentDict{Symbol, Int64},🅰️ :Symbol)::Int64 └── return %29 ``` After: ``` julia> using BenchmarkTools julia> function foo() a = Base.PersistentDict(:a => 1) return a[:a] end foo (generic function with 1 method) julia> @benchmark foo() BenchmarkTools.Trial: 10000 samples with 1000 evaluations. Range (min … max): 2.459 ns … 11.320 ns ┊ GC (min … max): 0.00% … 0.00% Time (median): 2.460 ns ┊ GC (median): 0.00% Time (mean ± σ): 2.469 ns ± 0.183 ns ┊ GC (mean ± σ): 0.00% ± 0.00% ▂ █ ▁ █ ▂ █▁▁▁▁█▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁█▁▁▁▁█ █ 2.46 ns Histogram: log(frequency) by time 2.47 ns < Memory estimate: 0 bytes, allocs estimate: 0. julia> @code_typed foo() CodeInfo( 1 ─ return 1 ```
Rebased. I think this is basically good to go. I'm gonna do some more experiments with this, but those might go into a follow up PR, depending. |
Working well in my downstream testing. It does still not fully optimize ScopedValue, because we do not yet have the ability to remove all the setup code, which will need some effect tweaks, but that's independent of this, so I'll go ahead and merge this. |
As discussed in #51352, this gives `EnterNode` the ability to set (and restore on leave or catch edge) jl_current_task->scope. Manual modifications of the task field after the task has started are considered undefined behavior. In addition, we gain a new intrinsic to access current_task->scope and both inference and the optimizer will forward scopes from EnterNodes to this intrinsic (non-interprocedurally). Together with #51993 this is sufficient to fully optimize ScopedValues (non-interprocedurally at least).
This is part of the work to address JuliaLang#51352 by attempting to allow the compiler to perform SRAO on persistent data structures like `PersistentDict` as if they were regular immutable data structures. These sorts of data structures have very complicated internals (with lots of mutation, memory sharing, etc.), but a relatively simple interface. As such, it is unlikely that our compiler will have sufficient power to optimize this interface by analyzing the implementation. We thus need to come up with some other mechanism that gives the compiler license to perform the requisite optimization. One way would be to just hardcode `PersistentDict` into the compiler, optimizing it like any of the other builtin datatypes. However, this is of course very unsatisfying. At the other end of the spectrum would be something like a generic rewrite rule system (e-graphs anyone?) that would let the PersistentDict implementation declare its interface to the compiler and the compiler would use this for optimization (in a perfect world, the actual rewrite would then be checked using some sort of formal methods). I think that would be interesting, but we're very far from even being able to design something like that (at least in Base - experiments with external AbstractInterpreters in this direction are encouraged). This PR tries to come up with a reasonable middle ground, where the compiler gets some knowledge of the protocol hardcoded without having to know about the implementation details of the data structure. The basic ideas is that `Core` provides some magic generic functions that implementations can extend. Semantically, they are not special. They dispatch as usual, and implementations are expected to work properly even in the absence of any compiler optimizations. However, the compiler is semantically permitted to perform structural optimization using these magic generic functions. In the concrete case, this PR introduces the `KeyValue` interface which consists of two generic functions, `get` and `set`. The core optimization is that the compiler is allowed to rewrite any occurrence of `get(set(x, k, v), k)` into `v` without additional legality checks. In particular, the compiler performs no type checks, conversions, etc. The higher level implementation code is expected to do all that. This approach closely matches the general direction we've been taking in external AbstractInterpreters for embedding additional semantics and optimization opportunities into Julia code (although we generally use methods there, rather than full generic functions), so I think we have some evidence that this sort of approach works reasonably well. Nevertheless, this is certainly an experiment and the interface is explicitly declared unstable. ## Current Status This is fully working and implemented, but the optimization currently bails on anything but the simplest cases. Filling all those cases in is not particularly hard, but should be done along with a more invasive refactoring of SROA, so we should figure out the general direction here first and then we can finish all that up in a follow-up cleanup. ## Obligatory benchmark Before: ``` julia> using BenchmarkTools julia> function foo() a = Base.PersistentDict(:a => 1) return a[:a] end foo (generic function with 1 method) julia> @benchmark foo() BenchmarkTools.Trial: 10000 samples with 993 evaluations. Range (min … max): 32.940 ns … 28.754 μs ┊ GC (min … max): 0.00% … 99.76% Time (median): 49.647 ns ┊ GC (median): 0.00% Time (mean ± σ): 57.519 ns ± 333.275 ns ┊ GC (mean ± σ): 10.81% ± 2.22% ▃█▅ ▁▃▅▅▃▁ ▁▃▂ ▂ ▁▂▄▃▅▇███▇▃▁▂▁▁▁▁▁▁▁▁▂▂▅██████▅▂▁▁▁▁▁▁▁▁▁▁▂▃▃▇███▇▆███▆▄▃▃▂▂ ▃ 32.9 ns Histogram: frequency by time 68.6 ns < Memory estimate: 128 bytes, allocs estimate: 4. julia> @code_typed foo() CodeInfo( 1 ─ %1 = invoke Vector{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}}(Base.HashArrayMappedTries.undef::UndefInitializer, 1::Int64)::Vector{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} │ %2 = %new(Base.HashArrayMappedTries.HAMT{Symbol, Int64}, %1, 0x00000000)::Base.HashArrayMappedTries.HAMT{Symbol, Int64} │ %3 = %new(Base.HashArrayMappedTries.Leaf{Symbol, Int64}, :a, 1)::Base.HashArrayMappedTries.Leaf{Symbol, Int64} │ %4 = Base.getfield(%2, :data)::Vector{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} │ %5 = $(Expr(:boundscheck, true))::Bool └── goto JuliaLang#5 if not %5 2 ─ %7 = Base.sub_int(1, 1)::Int64 │ %8 = Base.bitcast(UInt64, %7)::UInt64 │ %9 = Base.getfield(%4, :size)::Tuple{Int64} │ %10 = $(Expr(:boundscheck, true))::Bool │ %11 = Base.getfield(%9, 1, %10)::Int64 │ %12 = Base.bitcast(UInt64, %11)::UInt64 │ %13 = Base.ult_int(%8, %12)::Bool └── goto JuliaLang#4 if not %13 3 ─ goto JuliaLang#5 4 ─ %16 = Core.tuple(1)::Tuple{Int64} │ invoke Base.throw_boundserror(%4::Vector{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}}, %16::Tuple{Int64})::Union{} └── unreachable 5 ┄ %19 = Base.getfield(%4, :ref)::MemoryRef{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} │ %20 = Base.memoryref(%19, 1, false)::MemoryRef{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} │ Base.memoryrefset!(%20, %3, :not_atomic, false)::MemoryRef{Union{Base.HashArrayMappedTries.HAMT{Symbol, Int64}, Base.HashArrayMappedTries.Leaf{Symbol, Int64}}} └── goto JuliaLang#6 6 ─ %23 = Base.getfield(%2, :bitmap)::UInt32 │ %24 = Base.or_int(%23, 0x00010000)::UInt32 │ Base.setfield!(%2, :bitmap, %24)::UInt32 └── goto JuliaLang#7 7 ─ %27 = %new(Base.PersistentDict{Symbol, Int64}, %2)::Base.PersistentDict{Symbol, Int64} └── goto JuliaLang#8 8 ─ %29 = invoke Base.getindex(%27::Base.PersistentDict{Symbol, Int64},🅰️ :Symbol)::Int64 └── return %29 ``` After: ``` julia> using BenchmarkTools julia> function foo() a = Base.PersistentDict(:a => 1) return a[:a] end foo (generic function with 1 method) julia> @benchmark foo() BenchmarkTools.Trial: 10000 samples with 1000 evaluations. Range (min … max): 2.459 ns … 11.320 ns ┊ GC (min … max): 0.00% … 0.00% Time (median): 2.460 ns ┊ GC (median): 0.00% Time (mean ± σ): 2.469 ns ± 0.183 ns ┊ GC (mean ± σ): 0.00% ± 0.00% ▂ █ ▁ █ ▂ █▁▁▁▁█▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁█▁▁▁▁█ █ 2.46 ns Histogram: log(frequency) by time 2.47 ns < Memory estimate: 0 bytes, allocs estimate: 0. julia> @code_typed foo() CodeInfo( 1 ─ return 1 ```
As discussed in #51352, this gives `EnterNode` the ability to set (and restore on leave or catch edge) jl_current_task->scope. Manual modifications of the task field after the task has started are considered undefined behavior. In addition, we gain a new intrinsic to access current_task->scope and both inference and the optimizer will forward scopes from EnterNodes to this intrinsic (non-interprocedurally). Together with #51993 this is sufficient to fully optimize ScopedValues (non-interprocedurally at least).
As discussed in #51352, this gives `EnterNode` the ability to set (and restore on leave or catch edge) jl_current_task->scope. Manual modifications of the task field after the task has started are considered undefined behavior. In addition, we gain a new intrinsic to access current_task->scope and both inference and the optimizer will forward scopes from EnterNodes to this intrinsic (non-interprocedurally). Together with #51993 this is sufficient to fully optimize ScopedValues (non-interprocedurally at least).
This is part of the work to address #51352 by attempting to allow the compiler to perform SRAO on persistent data structures like
PersistentDict
as if they were regular immutable data structures. These sorts of data structures have very complicated internals (with lots of mutation, memory sharing, etc.), but a relatively simple interface. As such, it is unlikely that our compiler will have sufficient power to optimize this interface by analyzing the implementation.We thus need to come up with some other mechanism that gives the compiler license to perform the requisite optimization. One way would be to just hardcode
PersistentDict
into the compiler, optimizing it like any of the other builtin datatypes. However, this is of course very unsatisfying. At the other end of the spectrum would be something like a generic rewrite rule system (e-graphs anyone?) that would let the PersistentDict implementation declare its interface to the compiler and the compiler would use this for optimization (in a perfect world, the actual rewrite would then be checked using some sort of formal methods). I think that would be interesting, but we're very far from even being able to design something like that (at least in Base - experiments with external AbstractInterpreters in this direction are encouraged).This PR tries to come up with a reasonable middle ground, where the compiler gets some knowledge of the protocol hardcoded without having to know about the implementation details of the data structure.
The basic ideas is that
Core
provides some magic generic functions that implementations can extend. Semantically, they are not special. They dispatch as usual, and implementations are expected to work properly even in the absence of any compiler optimizations.However, the compiler is semantically permitted to perform structural optimization using these magic generic functions. In the concrete case, this PR introduces the
KeyValue
interface which consists of two generic functions,get
andset
. The core optimization is that the compiler is allowed to rewrite any occurrence ofget(set(x, k, v), k)
intov
without additional legality checks. In particular, the compiler performs no type checks, conversions, etc. The higher level implementation code is expected to do all that.This approach closely matches the general direction we've been taking in external AbstractInterpreters for embedding additional semantics and optimization opportunities into Julia code (although we generally use methods there, rather than full generic functions), so I think we have some evidence that this sort of approach works reasonably well.
Nevertheless, this is certainly an experiment and the interface is explicitly declared unstable.
Current Status
This is fully working and implemented, but the optimization currently bails on anything but the simplest cases. Filling all those cases in is not particularly hard, but should be done along with a more invasive refactoring of SROA, so we should figure out the general direction here first and then we can finish all that up in a follow-up cleanup.
Obligatory benchmark
Before:
After: