Add functions to allow "external" memory registration #6

rdaum · 2023-05-03T17:30:47Z

Goal: allow a mechanism to make the GC aware of non-Julia process memory usage.

Background in RAI-7988

Adds calls to register allocations and frees of process memory
unaccounted for by live Julia objects, but relevant for GC
collection heuristics. (jl_gc_allocd_external and jl_gc_freed_external)
Adds a separate call to allow the Julia program to adjust the
default_collect_interval and makes the latter a global static
instead of #defined constant. (jl_gc_set_default_collect_interval)
Increases the default for default_collect_interval (to avoid excessive
full sweeps once external memory is reported)

…ated memory. Goal: allow a mechanism to make the GC aware of non-Julia process memory usage. Background in RAI-7988 * Adds calls to register allocations and frees of process memory unaccounted for by live Julia objects, but relevant for GC collection heuristics. * Adds a separate call to allow the Julia program to adjust the `default_collect_interval`. * Increases the default collection interval (to avoid excessive full sweeps once external memory is reported)

kpamnany

Okay, let's get this done.

@d-netto: leaving aside the question of the racy update of allocd, can you confirm that we should split the add to allocd across threads? If we just update allocd for the current thread, i.e. just use jl_gc_count_allocd(), will it work fine?

@rdaum: what problems did you see with just using jl_gc_count_allocd()?

Also, let's drop the default_collect_interval related changes.

rdaum · 2023-05-03T20:00:36Z

src/gc.c

+    // thread goes over max_collect_interval.
+    int64_t per_tls_delta = sz_delta / jl_n_threads;
+    int64_t remainder = sz_delta % jl_n_threads;
+


I believe I get the intent, but not sure it's possible to grab and compare all of jl_all_tls_states like this? jl_atomic_load_relaxed works on a single atomic value, and there'd be no way to atomically load or compare the values from all threads at once, would there? And I suspect combine_thread_gc_counts below is either incorrect or is behind a locking synchronization point?

I'm trying to recall for sure if per-thread allocd being off balance was a concern or not and what led to me writing this. It's possible just sticking it in the current thread and letting combine_thread_gc_counts work it out is good enough.

Sorry, my comments are only relevant for Julia 1.9 and up. jl_all_tls_states is not an _Atomic in 1.8.2, so you can ignore all that.

To address your question anyway -- in 1.9, it is possible to add "foreign" threads through the Julia C library interface. When that happens, a new jl_ptls_t * array is created and jl_all_tls_states is atomically switched to point at the new array -- the old array is lazily freed. So my suggestions were basically about checking to see if jl_all_tls_states changed underneath us while we were doing our updates and if it did, to redo them.

@d-netto agreed with the need to partition the allocd across threads, FWIW, maybe he can clarify on that question.

Also, there are plans afoot to make changes to the GC's heuristics -- default_collect_interval may go away altogether. So if we're not immediately using the calls to adjust it, let's drop them from this patch to keep it minimal.

rdaum · 2023-05-03T20:00:36Z

src/gc.c

+    // thread goes over max_collect_interval.
+    int64_t per_tls_delta = sz_delta / jl_n_threads;
+    int64_t remainder = sz_delta % jl_n_threads;
+


I believe I get the intent, but not sure it's possible to grab and compare all of jl_all_tls_states like this? jl_atomic_load_relaxed works on a single atomic value, and there'd be no way to atomically load or compare the values from all threads at once, would there? And I suspect combine_thread_gc_counts below is either incorrect or is behind a locking synchronization point?

I'm trying to recall for sure if per-thread allocd being off balance was a concern or not and what led to me writing this. It's possible just sticking it in the current thread and letting combine_thread_gc_counts work it out is good enough.

d-netto

I think the sequence jl_atomic_store_relaxed(&ptls->gc_num.allocd, jl_atomic_load_relaxed(&ptls->gc_num.allocd) + per_tls_delta) might not be atomic (though the individual stores and loads are, of course).

I was wondering whether an allocation by thread A, which interleaves with thread B's increments to thread A's ptls->gc_num.allocd could lead to an incorrect reported allocd due to a race.

kpamnany · 2023-05-04T17:14:45Z

I think the sequence jl_atomic_store_relaxed(&ptls->gc_num.allocd, jl_atomic_load_relaxed(&ptls->gc_num.allocd) + per_tls_delta) might not be atomic

True. jl_atomic_fetch_add is a costlier operation though, which, I think, is why this non-atomic version is being used throughout the GC code, e.g. here.

I think the thinking is that it is not critical that this number be exact?

d-netto · 2023-05-04T17:22:13Z

I think the thinking is that it is not critical that this number be exact?

This race condition could lead us to lose the per_tls_delta incremented bytes, which seems concerning:

Thread A loads ptls->gc_num.allocd in jl_gc_big_alloc_inner.
Thread B loads ptls->gc_num.allocd in jl_gc_allocd_external.
Thread B stores ptls->gc_num.allocd + per_tls_delta in jl_gc_allocd_external.
Thread A stores ptls->gc_num.allocd + sz in jl_gc_big_alloc_inner.

kpamnany · 2023-05-04T20:08:59Z

Okay, so after discussion with Diogo, the only way to safely distribute an external allocation/free request across threads is to convert all the GC metrics updates into fetch-add atomics, which will likely be noticeably costly. We may have to pay that price, but we can try something else first.

In https://github.com/RelationalAI/raicode/pull/13506, I've added a different patch that simply makes jl_gc_count_allocd() external and adds a jl_gc_count_freed(). These will attribute the memory allocated/freed to the calling thread, as a result of which there will most likely be an immediate collection.

To prevent these extra collections from becoming excessive, I think @rdaum had (in his earlier work on this) tweaked the GC heuristics. We'd prefer to avoid doing that since all GC heuristics are going to be undergoing a revamp soon. An alternative solution might be for the pager to call these functions only when the amount of allocated/freed memory exceeds some threshold, rather than to call these functions on every allocation/free.

rdaum · 2023-05-08T12:03:45Z

I was wondering whether an allocation by thread A, which interleaves with thread B's increments to thread A's ptls->gc_num.allocd could lead to an incorrect reported allocd due to a race.

I think there's definitely potential for a race here.

I'm also not convinced that the work I did here to distribute the load across threads is actually necessary. I would have to revisit the experiment with an actual running workload and see. I suspect just incrementing or decrementing current thread count is likely fine. There were many moving levers back when I was doing my benchmark runs, and this was only one that I played with, and so I'd want to try pulling it out again and seeing where we end up.

kpamnany · 2023-05-26T15:06:43Z

Given that we can't really split an allocation notification across threads without bigger changes, this approach won't work. Also, the periodic GC.gc call has shown good results, so we'll likely go with that.

@benchmark

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 #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 ```

@something

`@something` eagerly unwraps any `Some` given to it, while keeping the variable between its arguments the same. This can be an issue if a previously unpacked value is used as input to `@something`, leading to a type instability on more than two arguments (e.g. because of a fallback to `Some(nothing)`). By using different variables for each argument, type inference has an easier time handling these cases that are isolated to single branches anyway. This also adds some comments to the macro, since it's non-obvious what it does. Benchmarking the specific case I encountered this in led to a ~2x performance improvement on multiple machines. 1.10-beta3/master: ``` [sukera@tower 01]$ jl1100 -q --project=. -L 01.jl -e 'bench()' v"1.10.0-beta3" BenchmarkTools.Trial: 10000 samples with 1 evaluation. Range (min … max): 38.670 μs … 70.350 μs ┊ GC (min … max): 0.00% … 0.00% Time (median): 43.340 μs ┊ GC (median): 0.00% Time (mean ± σ): 43.395 μs ± 1.518 μs ┊ GC (mean ± σ): 0.00% ± 0.00% ▆█▂ ▁▁ ▂▂▂▂▂▂▂▂▂▁▂▂▂▃▃▃▂▂▃▃▃▂▂▂▂▂▄▇███▆██▄▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂ ▃ 38.7 μs Histogram: frequency by time 48 μs < Memory estimate: 0 bytes, allocs estimate: 0. ``` This PR: ``` [sukera@tower 01]$ julia -q --project=. -L 01.jl -e 'bench()' v"1.11.0-DEV.970" BenchmarkTools.Trial: 10000 samples with 1 evaluation. Range (min … max): 22.820 μs … 44.980 μs ┊ GC (min … max): 0.00% … 0.00% Time (median): 24.300 μs ┊ GC (median): 0.00% Time (mean ± σ): 24.370 μs ± 832.239 ns ┊ GC (mean ± σ): 0.00% ± 0.00% ▂▅▇██▇▆▅▁ ▂▂▂▂▂▂▂▂▃▃▄▅▇███████████▅▄▃▃▂▂▂▂▂▂▂▂▂▂▁▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▁▁▂▂ ▃ 22.8 μs Histogram: frequency by time 27.7 μs < Memory estimate: 0 bytes, allocs estimate: 0. ``` <details> <summary>Benchmarking code (spoilers for Advent Of Code 2023 Day 01, Part 01). Running this requires the input of that Advent Of Code day.</summary> ```julia using BenchmarkTools using InteractiveUtils isdigit(d::UInt8) = UInt8('0') <= d <= UInt8('9') someDigit(c::UInt8) = isdigit(c) ? Some(c - UInt8('0')) : nothing function part1(data) total = 0 may_a = nothing may_b = nothing for c in data digitRes = someDigit(c) may_a = @something may_a digitRes Some(nothing) may_b = @something digitRes may_b Some(nothing) if c == UInt8('\n') digit_a = may_a::UInt8 digit_b = may_b::UInt8 total += digit_a*0xa + digit_b may_a = nothing may_b = nothing end end return total end function bench() data = read("input.txt") display(VERSION) println() display(@benchmark part1($data)) nothing end ``` </details> <details> <summary>`@code_warntype` before</summary> ```julia julia> @code_warntype part1(data) MethodInstance for part1(::Vector{UInt8}) from part1(data) @ Main ~/Documents/projects/AOC/2023/01/01.jl:7 Arguments #self#::Core.Const(part1) data::Vector{UInt8} Locals @_3::Union{Nothing, Tuple{UInt8, Int64}} may_b::Union{Nothing, UInt8} may_a::Union{Nothing, UInt8} total::Int64 c::UInt8 digit_b::UInt8 digit_a::UInt8 val@_10::Any val@_11::Any digitRes::Union{Nothing, Some{UInt8}} @_13::Union{Some{Nothing}, Some{UInt8}, UInt8} @_14::Union{Some{Nothing}, Some{UInt8}} @_15::Some{Nothing} @_16::Union{Some{Nothing}, Some{UInt8}, UInt8} @_17::Union{Some{Nothing}, UInt8} @_18::Some{Nothing} Body::Int64 1 ── (total = 0) │ (may_a = Main.nothing) │ (may_b = Main.nothing) │ %4 = data::Vector{UInt8} │ (@_3 = Base.iterate(%4)) │ %6 = (@_3 === nothing)::Bool │ %7 = Base.not_int(%6)::Bool └─── goto #24 if not %7 2 ┄─ Core.NewvarNode(:(digit_b)) │ Core.NewvarNode(:(digit_a)) │ Core.NewvarNode(:(val@_10)) │ %12 = @_3::Tuple{UInt8, Int64} │ (c = Core.getfield(%12, 1)) │ %14 = Core.getfield(%12, 2)::Int64 │ (digitRes = Main.someDigit(c)) │ (val@_11 = may_a) │ %17 = (val@_11::Union{Nothing, UInt8} !== Base.nothing)::Bool └─── goto #4 if not %17 3 ── (@_13 = val@_11::UInt8) └─── goto #11 4 ── (val@_11 = digitRes) │ %22 = (val@_11::Union{Nothing, Some{UInt8}} !== Base.nothing)::Bool └─── goto #6 if not %22 5 ── (@_14 = val@_11::Some{UInt8}) └─── goto #10 6 ── (val@_11 = Main.Some(Main.nothing)) │ %27 = (val@_11::Core.Const(Some(nothing)) !== Base.nothing)::Core.Const(true) └─── goto #8 if not %27 7 ── (@_15 = val@_11::Core.Const(Some(nothing))) └─── goto #9 8 ── Core.Const(:(@_15 = Base.nothing)) 9 ┄─ (@_14 = @_15) 10 ┄ (@_13 = @_14) 11 ┄ %34 = @_13::Union{Some{Nothing}, Some{UInt8}, UInt8} │ (may_a = Base.something(%34)) │ (val@_10 = digitRes) │ %37 = (val@_10::Union{Nothing, Some{UInt8}} !== Base.nothing)::Bool └─── goto #13 if not %37 12 ─ (@_16 = val@_10::Some{UInt8}) └─── goto #20 13 ─ (val@_10 = may_b) │ %42 = (val@_10::Union{Nothing, UInt8} !== Base.nothing)::Bool └─── goto #15 if not %42 14 ─ (@_17 = val@_10::UInt8) └─── goto #19 15 ─ (val@_10 = Main.Some(Main.nothing)) │ %47 = (val@_10::Core.Const(Some(nothing)) !== Base.nothing)::Core.Const(true) └─── goto #17 if not %47 16 ─ (@_18 = val@_10::Core.Const(Some(nothing))) └─── goto #18 17 ─ Core.Const(:(@_18 = Base.nothing)) 18 ┄ (@_17 = @_18) 19 ┄ (@_16 = @_17) 20 ┄ %54 = @_16::Union{Some{Nothing}, Some{UInt8}, UInt8} │ (may_b = Base.something(%54)) │ %56 = c::UInt8 │ %57 = Main.UInt8('\n')::Core.Const(0x0a) │ %58 = (%56 == %57)::Bool └─── goto #22 if not %58 21 ─ (digit_a = Core.typeassert(may_a, Main.UInt8)) │ (digit_b = Core.typeassert(may_b, Main.UInt8)) │ %62 = total::Int64 │ %63 = (digit_a * 0x0a)::UInt8 │ %64 = (%63 + digit_b)::UInt8 │ (total = %62 + %64) │ (may_a = Main.nothing) └─── (may_b = Main.nothing) 22 ┄ (@_3 = Base.iterate(%4, %14)) │ %69 = (@_3 === nothing)::Bool │ %70 = Base.not_int(%69)::Bool └─── goto #24 if not %70 23 ─ goto #2 24 ┄ return total ``` </details> <details> <summary>`@code_native debuginfo=:none` Before </summary> ```julia julia> @code_native debuginfo=:none part1(data) .text .file "part1" .globl julia_part1_418 # -- Begin function julia_part1_418 .p2align 4, 0x90 .type julia_part1_418,@function julia_part1_418: # @julia_part1_418 # %bb.0: # %top push rbp mov rbp, rsp push r15 push r14 push r13 push r12 push rbx sub rsp, 40 mov rax, qword ptr [rdi + 8] test rax, rax je .LBB0_1 # %bb.2: # %L17 mov rcx, qword ptr [rdi] dec rax mov r10b, 1 xor r14d, r14d # implicit-def: $r12b # implicit-def: $r13b # implicit-def: $r9b # implicit-def: $sil mov qword ptr [rbp - 64], rax # 8-byte Spill mov al, 1 mov dword ptr [rbp - 48], eax # 4-byte Spill # implicit-def: $al # kill: killed $al xor eax, eax mov qword ptr [rbp - 56], rax # 8-byte Spill mov qword ptr [rbp - 72], rcx # 8-byte Spill # implicit-def: $cl jmp .LBB0_3 .p2align 4, 0x90 .LBB0_8: # in Loop: Header=BB0_3 Depth=1 mov dword ptr [rbp - 48], 0 # 4-byte Folded Spill .LBB0_24: # %post_union_move # in Loop: Header=BB0_3 Depth=1 movzx r13d, byte ptr [rbp - 41] # 1-byte Folded Reload mov r12d, r8d cmp qword ptr [rbp - 64], r14 # 8-byte Folded Reload je .LBB0_13 .LBB0_25: # %guard_exit113 # in Loop: Header=BB0_3 Depth=1 inc r14 mov r10d, ebx .LBB0_3: # %L19 # =>This Inner Loop Header: Depth=1 mov rax, qword ptr [rbp - 72] # 8-byte Reload xor ebx, ebx xor edi, edi movzx r15d, r9b movzx ecx, cl movzx esi, sil mov r11b, 1 # implicit-def: $r9b movzx edx, byte ptr [rax + r14] lea eax, [rdx - 58] lea r8d, [rdx - 48] cmp al, -10 setae bl setb dil test r10b, 1 cmovne r15d, edi mov edi, 0 cmovne ecx, ebx mov bl, 1 cmovne esi, edi test r15b, 1 jne .LBB0_7 # %bb.4: # %L76 # in Loop: Header=BB0_3 Depth=1 mov r11b, 2 test cl, 1 jne .LBB0_5 # %bb.6: # %L78 # in Loop: Header=BB0_3 Depth=1 mov ebx, r10d mov r9d, r15d mov byte ptr [rbp - 41], r13b # 1-byte Spill test sil, 1 je .LBB0_26 .LBB0_7: # %L82 # in Loop: Header=BB0_3 Depth=1 cmp al, -11 jbe .LBB0_9 jmp .LBB0_8 .p2align 4, 0x90 .LBB0_5: # in Loop: Header=BB0_3 Depth=1 mov ecx, r8d mov sil, 1 xor ebx, ebx mov byte ptr [rbp - 41], r8b # 1-byte Spill xor r9d, r9d xor ecx, ecx cmp al, -11 ja .LBB0_8 .LBB0_9: # %L90 # in Loop: Header=BB0_3 Depth=1 test byte ptr [rbp - 48], 1 # 1-byte Folded Reload jne .LBB0_23 # %bb.10: # %L115 # in Loop: Header=BB0_3 Depth=1 cmp dl, 10 jne .LBB0_11 # %bb.14: # %L122 # in Loop: Header=BB0_3 Depth=1 test r15b, 1 jne .LBB0_15 # %bb.12: # %L130.thread # in Loop: Header=BB0_3 Depth=1 movzx eax, byte ptr [rbp - 41] # 1-byte Folded Reload mov bl, 1 add eax, eax lea eax, [rax + 4*rax] add al, r12b movzx eax, al add qword ptr [rbp - 56], rax # 8-byte Folded Spill mov al, 1 mov dword ptr [rbp - 48], eax # 4-byte Spill cmp qword ptr [rbp - 64], r14 # 8-byte Folded Reload jne .LBB0_25 jmp .LBB0_13 .p2align 4, 0x90 .LBB0_23: # %L115.thread # in Loop: Header=BB0_3 Depth=1 mov al, 1 # implicit-def: $r8b mov dword ptr [rbp - 48], eax # 4-byte Spill cmp dl, 10 jne .LBB0_24 jmp .LBB0_21 .LBB0_11: # in Loop: Header=BB0_3 Depth=1 mov r8d, r12d jmp .LBB0_24 .LBB0_1: xor eax, eax mov qword ptr [rbp - 56], rax # 8-byte Spill .LBB0_13: # %L159 mov rax, qword ptr [rbp - 56] # 8-byte Reload add rsp, 40 pop rbx pop r12 pop r13 pop r14 pop r15 pop rbp ret .LBB0_21: # %L122.thread test r15b, 1 jne .LBB0_15 # %bb.22: # %post_box_union58 movabs rdi, offset .L_j_str1 movabs rax, offset ijl_type_error movabs rsi, 140008511215408 movabs rdx, 140008667209736 call rax .LBB0_15: # %fail cmp r11b, 1 je .LBB0_19 # %bb.16: # %fail movzx eax, r11b cmp eax, 2 jne .LBB0_17 # %bb.20: # %box_union54 movzx eax, byte ptr [rbp - 41] # 1-byte Folded Reload movabs rcx, offset jl_boxed_uint8_cache mov rdx, qword ptr [rcx + 8*rax] jmp .LBB0_18 .LBB0_26: # %L80 movabs rax, offset ijl_throw movabs rdi, 140008495049392 call rax .LBB0_19: # %box_union movabs rdx, 140008667209736 jmp .LBB0_18 .LBB0_17: xor edx, edx .LBB0_18: # %post_box_union movabs rdi, offset .L_j_str1 movabs rax, offset ijl_type_error movabs rsi, 140008511215408 call rax .Lfunc_end0: .size julia_part1_418, .Lfunc_end0-julia_part1_418 # -- End function .type .L_j_str1,@object # @_j_str1 .section .rodata.str1.1,"aMS",@progbits,1 .L_j_str1: .asciz "typeassert" .size .L_j_str1, 11 .section ".note.GNU-stack","",@progbits ``` </details> <details> <summary>`@code_warntype` After</summary> ```julia [sukera@tower 01]$ julia -q --project=. -L 01.jl julia> data = read("input.txt"); julia> @code_warntype part1(data) MethodInstance for part1(::Vector{UInt8}) from part1(data) @ Main ~/Documents/projects/AOC/2023/01/01.jl:7 Arguments #self#::Core.Const(part1) data::Vector{UInt8} Locals @_3::Union{Nothing, Tuple{UInt8, Int64}} may_b::Union{Nothing, UInt8} may_a::Union{Nothing, UInt8} total::Int64 val@_7::Union{} val@_8::Union{} c::UInt8 digit_b::UInt8 digit_a::UInt8 #JuliaLang#215::Some{Nothing} #JuliaLang#216::Union{Nothing, UInt8} #JuliaLang#217::Union{Nothing, Some{UInt8}} #JuliaLang#212::Some{Nothing} #JuliaLang#213::Union{Nothing, Some{UInt8}} #JuliaLang#214::Union{Nothing, UInt8} digitRes::Union{Nothing, Some{UInt8}} @_19::Union{Nothing, UInt8} @_20::Union{Nothing, UInt8} @_21::Nothing @_22::Union{Nothing, UInt8} @_23::Union{Nothing, UInt8} @_24::Nothing Body::Int64 1 ── (total = 0) │ (may_a = Main.nothing) │ (may_b = Main.nothing) │ %4 = data::Vector{UInt8} │ (@_3 = Base.iterate(%4)) │ %6 = @_3::Union{Nothing, Tuple{UInt8, Int64}} │ %7 = (%6 === nothing)::Bool │ %8 = Base.not_int(%7)::Bool └─── goto #24 if not %8 2 ┄─ Core.NewvarNode(:(val@_7)) │ Core.NewvarNode(:(val@_8)) │ Core.NewvarNode(:(digit_b)) │ Core.NewvarNode(:(digit_a)) │ Core.NewvarNode(:(#JuliaLang#215)) │ Core.NewvarNode(:(#JuliaLang#216)) │ Core.NewvarNode(:(#JuliaLang#217)) │ Core.NewvarNode(:(#JuliaLang#212)) │ Core.NewvarNode(:(#JuliaLang#213)) │ %19 = @_3::Tuple{UInt8, Int64} │ (c = Core.getfield(%19, 1)) │ %21 = Core.getfield(%19, 2)::Int64 │ %22 = c::UInt8 │ (digitRes = Main.someDigit(%22)) │ %24 = may_a::Union{Nothing, UInt8} │ (#JuliaLang#214 = %24) │ %26 = Base.:!::Core.Const(!) │ %27 = #JuliaLang#214::Union{Nothing, UInt8} │ %28 = Base.isnothing(%27)::Bool │ %29 = (%26)(%28)::Bool └─── goto #4 if not %29 3 ── %31 = #JuliaLang#214::UInt8 │ (@_19 = Base.something(%31)) └─── goto #11 4 ── %34 = digitRes::Union{Nothing, Some{UInt8}} │ (#JuliaLang#213 = %34) │ %36 = Base.:!::Core.Const(!) │ %37 = #JuliaLang#213::Union{Nothing, Some{UInt8}} │ %38 = Base.isnothing(%37)::Bool │ %39 = (%36)(%38)::Bool └─── goto #6 if not %39 5 ── %41 = #JuliaLang#213::Some{UInt8} │ (@_20 = Base.something(%41)) └─── goto #10 6 ── %44 = Main.Some::Core.Const(Some) │ %45 = Main.nothing::Core.Const(nothing) │ (#JuliaLang#212 = (%44)(%45)) │ %47 = Base.:!::Core.Const(!) │ %48 = #JuliaLang#212::Core.Const(Some(nothing)) │ %49 = Base.isnothing(%48)::Core.Const(false) │ %50 = (%47)(%49)::Core.Const(true) └─── goto #8 if not %50 7 ── %52 = #JuliaLang#212::Core.Const(Some(nothing)) │ (@_21 = Base.something(%52)) └─── goto #9 8 ── Core.Const(nothing) │ Core.Const(:(val@_8 = Base.something(Base.nothing))) │ Core.Const(nothing) │ Core.Const(:(val@_8)) └─── Core.Const(:(@_21 = %58)) 9 ┄─ %60 = @_21::Core.Const(nothing) └─── (@_20 = %60) 10 ┄ %62 = @_20::Union{Nothing, UInt8} └─── (@_19 = %62) 11 ┄ %64 = @_19::Union{Nothing, UInt8} │ (may_a = %64) │ %66 = digitRes::Union{Nothing, Some{UInt8}} │ (#JuliaLang#217 = %66) │ %68 = Base.:!::Core.Const(!) │ %69 = #JuliaLang#217::Union{Nothing, Some{UInt8}} │ %70 = Base.isnothing(%69)::Bool │ %71 = (%68)(%70)::Bool └─── goto #13 if not %71 12 ─ %73 = #JuliaLang#217::Some{UInt8} │ (@_22 = Base.something(%73)) └─── goto #20 13 ─ %76 = may_b::Union{Nothing, UInt8} │ (#JuliaLang#216 = %76) │ %78 = Base.:!::Core.Const(!) │ %79 = #JuliaLang#216::Union{Nothing, UInt8} │ %80 = Base.isnothing(%79)::Bool │ %81 = (%78)(%80)::Bool └─── goto #15 if not %81 14 ─ %83 = #JuliaLang#216::UInt8 │ (@_23 = Base.something(%83)) └─── goto #19 15 ─ %86 = Main.Some::Core.Const(Some) │ %87 = Main.nothing::Core.Const(nothing) │ (#JuliaLang#215 = (%86)(%87)) │ %89 = Base.:!::Core.Const(!) │ %90 = #JuliaLang#215::Core.Const(Some(nothing)) │ %91 = Base.isnothing(%90)::Core.Const(false) │ %92 = (%89)(%91)::Core.Const(true) └─── goto #17 if not %92 16 ─ %94 = #JuliaLang#215::Core.Const(Some(nothing)) │ (@_24 = Base.something(%94)) └─── goto #18 17 ─ Core.Const(nothing) │ Core.Const(:(val@_7 = Base.something(Base.nothing))) │ Core.Const(nothing) │ Core.Const(:(val@_7)) └─── Core.Const(:(@_24 = %100)) 18 ┄ %102 = @_24::Core.Const(nothing) └─── (@_23 = %102) 19 ┄ %104 = @_23::Union{Nothing, UInt8} └─── (@_22 = %104) 20 ┄ %106 = @_22::Union{Nothing, UInt8} │ (may_b = %106) │ %108 = Main.:(==)::Core.Const(==) │ %109 = c::UInt8 │ %110 = Main.UInt8('\n')::Core.Const(0x0a) │ %111 = (%108)(%109, %110)::Bool └─── goto #22 if not %111 21 ─ %113 = may_a::Union{Nothing, UInt8} │ (digit_a = Core.typeassert(%113, Main.UInt8)) │ %115 = may_b::Union{Nothing, UInt8} │ (digit_b = Core.typeassert(%115, Main.UInt8)) │ %117 = Main.:+::Core.Const(+) │ %118 = total::Int64 │ %119 = Main.:+::Core.Const(+) │ %120 = Main.:*::Core.Const(*) │ %121 = digit_a::UInt8 │ %122 = (%120)(%121, 0x0a)::UInt8 │ %123 = digit_b::UInt8 │ %124 = (%119)(%122, %123)::UInt8 │ (total = (%117)(%118, %124)) │ (may_a = Main.nothing) └─── (may_b = Main.nothing) 22 ┄ (@_3 = Base.iterate(%4, %21)) │ %129 = @_3::Union{Nothing, Tuple{UInt8, Int64}} │ %130 = (%129 === nothing)::Bool │ %131 = Base.not_int(%130)::Bool └─── goto #24 if not %131 23 ─ goto #2 24 ┄ %134 = total::Int64 └─── return %134 ``` </details> <details> <summary>`@code_native debuginfo=:none` After </summary> ```julia julia> @code_native debuginfo=:none part1(data) .text .file "part1" .globl julia_part1_1203 # -- Begin function julia_part1_1203 .p2align 4, 0x90 .type julia_part1_1203,@function julia_part1_1203: # @julia_part1_1203 ; Function Signature: part1(Array{UInt8, 1}) # %bb.0: # %top #DEBUG_VALUE: part1:data <- [DW_OP_deref] $rdi push rbp mov rbp, rsp push r15 push r14 push r13 push r12 push rbx sub rsp, 40 vxorps xmm0, xmm0, xmm0 #APP mov rax, qword ptr fs:[0] #NO_APP lea rdx, [rbp - 64] vmovaps xmmword ptr [rbp - 64], xmm0 mov qword ptr [rbp - 48], 0 mov rcx, qword ptr [rax - 8] mov qword ptr [rbp - 64], 4 mov rax, qword ptr [rcx] mov qword ptr [rbp - 72], rcx # 8-byte Spill mov qword ptr [rbp - 56], rax mov qword ptr [rcx], rdx #DEBUG_VALUE: part1:data <- [DW_OP_deref] 0 mov r15, qword ptr [rdi + 16] test r15, r15 je .LBB0_1 # %bb.2: # %L34 mov r14, qword ptr [rdi] dec r15 mov r11b, 1 mov r13b, 1 # implicit-def: $r12b # implicit-def: $r10b xor eax, eax jmp .LBB0_3 .p2align 4, 0x90 .LBB0_4: # in Loop: Header=BB0_3 Depth=1 xor r11d, r11d mov ebx, edi mov r10d, r8d .LBB0_9: # %L114 # in Loop: Header=BB0_3 Depth=1 mov r12d, esi test r15, r15 je .LBB0_12 .LBB0_10: # %guard_exit126 # in Loop: Header=BB0_3 Depth=1 inc r14 dec r15 mov r13d, ebx .LBB0_3: # %L36 # =>This Inner Loop Header: Depth=1 movzx edx, byte ptr [r14] test r13b, 1 movzx edi, r13b mov ebx, 1 mov ecx, 0 cmove ebx, edi cmovne edi, ecx movzx ecx, r10b lea esi, [rdx - 48] lea r9d, [rdx - 58] movzx r8d, sil cmove r8d, ecx cmp r9b, -11 ja .LBB0_4 # %bb.5: # %L89 # in Loop: Header=BB0_3 Depth=1 test r11b, 1 jne .LBB0_8 # %bb.6: # %L102 # in Loop: Header=BB0_3 Depth=1 cmp dl, 10 jne .LBB0_7 # %bb.13: # %L106 # in Loop: Header=BB0_3 Depth=1 test r13b, 1 jne .LBB0_14 # %bb.11: # %L114.thread # in Loop: Header=BB0_3 Depth=1 add ecx, ecx mov bl, 1 mov r11b, 1 lea ecx, [rcx + 4*rcx] add cl, r12b movzx ecx, cl add rax, rcx test r15, r15 jne .LBB0_10 jmp .LBB0_12 .p2align 4, 0x90 .LBB0_8: # %L102.thread # in Loop: Header=BB0_3 Depth=1 mov r11b, 1 # implicit-def: $sil cmp dl, 10 jne .LBB0_9 jmp .LBB0_15 .LBB0_7: # in Loop: Header=BB0_3 Depth=1 mov esi, r12d jmp .LBB0_9 .LBB0_1: xor eax, eax .LBB0_12: # %L154 mov rcx, qword ptr [rbp - 56] mov rdx, qword ptr [rbp - 72] # 8-byte Reload mov qword ptr [rdx], rcx add rsp, 40 pop rbx pop r12 pop r13 pop r14 pop r15 pop rbp ret .LBB0_15: # %L106.thread test r13b, 1 jne .LBB0_14 # %bb.16: # %post_box_union47 movabs rax, offset jl_nothing movabs rcx, offset jl_small_typeof movabs rdi, offset ".L_j_str_typeassert#1" mov rdx, qword ptr [rax] mov rsi, qword ptr [rcx + 336] movabs rax, offset ijl_type_error mov qword ptr [rbp - 48], rsi call rax .LBB0_14: # %post_box_union movabs rax, offset jl_nothing movabs rcx, offset jl_small_typeof movabs rdi, offset ".L_j_str_typeassert#1" mov rdx, qword ptr [rax] mov rsi, qword ptr [rcx + 336] movabs rax, offset ijl_type_error mov qword ptr [rbp - 48], rsi call rax .Lfunc_end0: .size julia_part1_1203, .Lfunc_end0-julia_part1_1203 # -- End function .type ".L_j_str_typeassert#1",@object # @"_j_str_typeassert#1" .section .rodata.str1.1,"aMS",@progbits,1 ".L_j_str_typeassert#1": .asciz "typeassert" .size ".L_j_str_typeassert#1", 11 .section ".note.GNU-stack","",@progbits ``` </details> Co-authored-by: Sukera <[email protected]>

…ce. (JuliaLang#54113) The former also handles vectors of pointers, which can occur after vectorization: ``` #5 0x00007f5bfe94de5e in llvm::cast<llvm::PointerType, llvm::Type> (Val=<optimized out>) at llvm/Support/Casting.h:578 578 assert(isa<To>(Val) && "cast<Ty>() argument of incompatible type!"); (rr) up #6 GCInvariantVerifier::visitAddrSpaceCastInst (this=this@entry=0x7ffd022fbf56, I=...) at julia/src/llvm-gc-invariant-verifier.cpp:66 66 unsigned ToAS = cast<PointerType>(I.getDestTy())->getAddressSpace(); (rr) call I.dump() %23 = addrspacecast <4 x ptr addrspace(10)> %wide.load to <4 x ptr addrspace(11)>, !dbg !43 ``` Fixes aborts seen in JuliaLang#53070

@noinline

E.g. this allows `finalizer` inlining in the following case: ```julia mutable struct ForeignBuffer{T} const ptr::Ptr{T} end const foreign_buffer_finalized = Ref(false) function foreign_alloc(::Type{T}, length) where T ptr = Libc.malloc(sizeof(T) * length) ptr = Base.unsafe_convert(Ptr{T}, ptr) obj = ForeignBuffer{T}(ptr) return finalizer(obj) do obj Base.@assume_effects :notaskstate :nothrow foreign_buffer_finalized[] = true Libc.free(obj.ptr) end end function f_EA_finalizer(N::Int) workspace = foreign_alloc(Float64, N) GC.@preserve workspace begin (;ptr) = workspace Base.@assume_effects :nothrow @noinline println(devnull, "ptr = ", ptr) end end ``` ```julia julia> @code_typed f_EA_finalizer(42) CodeInfo( 1 ── %1 = Base.mul_int(8, N)::Int64 │ %2 = Core.lshr_int(%1, 63)::Int64 │ %3 = Core.trunc_int(Core.UInt8, %2)::UInt8 │ %4 = Core.eq_int(%3, 0x01)::Bool └─── goto #3 if not %4 2 ── invoke Core.throw_inexacterror(:convert::Symbol, UInt64::Type, %1::Int64)::Union{} └─── unreachable 3 ── goto #4 4 ── %9 = Core.bitcast(Core.UInt64, %1)::UInt64 └─── goto #5 5 ── goto #6 6 ── goto #7 7 ── goto #8 8 ── %14 = $(Expr(:foreigncall, :(:malloc), Ptr{Nothing}, svec(UInt64), 0, :(:ccall), :(%9), :(%9)))::Ptr{Nothing} └─── goto #9 9 ── %16 = Base.bitcast(Ptr{Float64}, %14)::Ptr{Float64} │ %17 = %new(ForeignBuffer{Float64}, %16)::ForeignBuffer{Float64} └─── goto #10 10 ─ %19 = $(Expr(:gc_preserve_begin, :(%17))) │ %20 = Base.getfield(%17, :ptr)::Ptr{Float64} │ invoke Main.println(Main.devnull::Base.DevNull, "ptr = "::String, %20::Ptr{Float64})::Nothing │ $(Expr(:gc_preserve_end, :(%19))) │ %23 = Main.foreign_buffer_finalized::Base.RefValue{Bool} │ Base.setfield!(%23, :x, true)::Bool │ %25 = Base.getfield(%17, :ptr)::Ptr{Float64} │ %26 = Base.bitcast(Ptr{Nothing}, %25)::Ptr{Nothing} │ $(Expr(:foreigncall, :(:free), Nothing, svec(Ptr{Nothing}), 0, :(:ccall), :(%26), :(%25)))::Nothing └─── return nothing ) => Nothing ``` However, this is still a WIP. Before merging, I want to improve EA's precision a bit and at least fix the test case that is currently marked as `broken`. I also need to check its impact on compiler performance. Additionally, I believe this feature is not yet practical. In particular, there is still significant room for improvement in the following areas: - EA's interprocedural capabilities: currently EA is performed ad-hoc for limited frames because of latency reasons, which significantly reduces its precision in the presence of interprocedural calls. - Relaxing the `:nothrow` check for finalizer inlining: the current algorithm requires `:nothrow`-ness on all paths from the allocation of the mutable struct to its last use, which is not practical for real-world cases. Even when `:nothrow` cannot be guaranteed, auxiliary optimizations such as inserting a `finalize` call after the last use might still be possible (JuliaLang#55990).

rdaum changed the base branch from master to 1.8.2-RAI May 3, 2023 17:31

rdaum requested review from d-netto and kpamnany May 3, 2023 17:32

rdaum changed the base branch from 1.8.2-RAI to v1.8.2+RAI May 3, 2023 17:33

rdaum force-pushed the rd-external-gc-alloc-hooks branch from 099f906 to 1408468 Compare May 3, 2023 17:35

kpamnany reviewed May 3, 2023

View reviewed changes

rdaum commented May 3, 2023

View reviewed changes

d-netto reviewed May 4, 2023

View reviewed changes

kpamnany closed this May 26, 2023

kpamnany deleted the rd-external-gc-alloc-hooks branch May 26, 2023 15:07

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add functions to allow "external" memory registration #6

Add functions to allow "external" memory registration #6

rdaum commented May 3, 2023

kpamnany left a comment •

edited

Loading

rdaum May 3, 2023

kpamnany May 3, 2023

rdaum May 3, 2023

d-netto left a comment

kpamnany commented May 4, 2023

d-netto commented May 4, 2023

kpamnany commented May 4, 2023

rdaum commented May 8, 2023

kpamnany commented May 26, 2023

Add functions to allow "external" memory registration #6

Add functions to allow "external" memory registration #6

Conversation

rdaum commented May 3, 2023

kpamnany left a comment • edited Loading

Choose a reason for hiding this comment

rdaum May 3, 2023

Choose a reason for hiding this comment

kpamnany May 3, 2023

Choose a reason for hiding this comment

rdaum May 3, 2023

Choose a reason for hiding this comment

d-netto left a comment

Choose a reason for hiding this comment

kpamnany commented May 4, 2023

d-netto commented May 4, 2023

kpamnany commented May 4, 2023

rdaum commented May 8, 2023

kpamnany commented May 26, 2023

kpamnany left a comment •

edited

Loading