Skip to content
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

Patch2 #7

Merged
merged 15 commits into from
Jul 9, 2021
Merged

Patch2 #7

merged 15 commits into from
Jul 9, 2021

Conversation

N5N3
Copy link
Owner

@N5N3 N5N3 commented Jul 9, 2021

No description provided.

@N5N3 N5N3 merged commit 36adcbd into master Jul 9, 2021
N5N3 pushed a commit that referenced this pull request Mar 28, 2022
Follows up JuliaLang#44708 -- in that PR I missed the most obvious optimization
opportunity, i.e. we can safely eliminate `isdefined` checks when all
fields are defined at allocation site.
This change allows us to eliminate capturing closure constructions when
the body and callsite of capture closure is available within a optimized
frame, e.g.:
```julia
function abmult(r::Int, x0)
    if r < 0
        r = -r
    end
    f = x -> x * r
    return @inline f(x0)
end
```
```diff
diff --git a/_master.jl b/_pr.jl
index ea06d865b75..c38f221090f 100644
--- a/_master.jl
+++ b/_pr.jl
@@ -1,24 +1,19 @@
 julia> @code_typed abmult(-3, 3)
 CodeInfo(
-1 ── %1  = Core.Box::Type{Core.Box}
-│    %2  = %new(%1, r@_2)::Core.Box
-│    %3  = Core.isdefined(%2, :contents)::Bool
-└───       goto #3 if not %3
+1 ──       goto #3 if not true
 2 ──       goto #4
 3 ──       $(Expr(:throw_undef_if_not, :r, false))::Any
-4 ┄─ %7  = (r@_2 < 0)::Any
-└───       goto #9 if not %7
-5 ── %9  = Core.isdefined(%2, :contents)::Bool
-└───       goto #7 if not %9
+4 ┄─ %4  = (r@_2 < 0)::Any
+└───       goto #9 if not %4
+5 ──       goto #7 if not true
 6 ──       goto #8
 7 ──       $(Expr(:throw_undef_if_not, :r, false))::Any
-8 ┄─ %13 = -r@_2::Any
-9 ┄─ %14 = φ (#4 => r@_2, #8 => %13)::Any
-│    %15 = Core.isdefined(%2, :contents)::Bool
-└───       goto #11 if not %15
+8 ┄─ %9  = -r@_2::Any
+9 ┄─ %10 = φ (#4 => r@_2, #8 => %9)::Any
+└───       goto #11 if not true
 10 ─       goto #12
 11 ─       $(Expr(:throw_undef_if_not, :r, false))::Any
-12 ┄ %19 = (x0 * %14)::Any
+12 ┄ %14 = (x0 * %10)::Any
 └───       goto #13
-13 ─       return %19
+13 ─       return %14
 ) => Any
```
N5N3 pushed a commit that referenced this pull request Apr 3, 2022
Currently the optimizer handles abstract callsite only when there is a
single dispatch candidate (in most cases), and so inlining and static-dispatch
are prohibited when the callsite is union-split (in other word, union-split
happens only when all the dispatch candidates are concrete).

However, there are certain patterns of code (most notably our Julia-level compiler code)
that inherently need to deal with abstract callsite.
The following example is taken from `Core.Compiler` utility:
```julia
julia> @inline isType(@nospecialize t) = isa(t, DataType) && t.name === Type.body.name
isType (generic function with 1 method)

julia> code_typed((Any,)) do x # abstract, but no union-split, successful inlining
           isType(x)
       end |> only
CodeInfo(
1 ─ %1 = (x isa Main.DataType)::Bool
└──      goto #3 if not %1
2 ─ %3 = π (x, DataType)
│   %4 = Base.getfield(%3, :name)::Core.TypeName
│   %5 = Base.getfield(Type{T}, :name)::Core.TypeName
│   %6 = (%4 === %5)::Bool
└──      goto #4
3 ─      goto #4
4 ┄ %9 = φ (#2 => %6, #3 => false)::Bool
└──      return %9
) => Bool

julia> code_typed((Union{Type,Nothing},)) do x # abstract, union-split, unsuccessful inlining
           isType(x)
       end |> only
CodeInfo(
1 ─ %1 = (isa)(x, Nothing)::Bool
└──      goto #3 if not %1
2 ─      goto #4
3 ─ %4 = Main.isType(x)::Bool
└──      goto #4
4 ┄ %6 = φ (#2 => false, #3 => %4)::Bool
└──      return %6
) => Bool
```
(note that this is a limitation of the inlining algorithm, and so any
user-provided hints like callsite inlining annotation doesn't help here)

This commit enables inlining and static dispatch for abstract union-split callsite.
The core idea here is that we can simulate our dispatch semantics by
generating `isa` checks in order of the specialities of dispatch candidates:
```julia
julia> code_typed((Union{Type,Nothing},)) do x # union-split, unsuccessful inlining
                  isType(x)
              end |> only
CodeInfo(
1 ─ %1  = (isa)(x, Nothing)::Bool
└──       goto #3 if not %1
2 ─       goto #9
3 ─ %4  = (isa)(x, Type)::Bool
└──       goto #8 if not %4
4 ─ %6  = π (x, Type)
│   %7  = (%6 isa Main.DataType)::Bool
└──       goto #6 if not %7
5 ─ %9  = π (%6, DataType)
│   %10 = Base.getfield(%9, :name)::Core.TypeName
│   %11 = Base.getfield(Type{T}, :name)::Core.TypeName
│   %12 = (%10 === %11)::Bool
└──       goto #7
6 ─       goto #7
7 ┄ %15 = φ (#5 => %12, #6 => false)::Bool
└──       goto #9
8 ─       Core.throw(ErrorException("fatal error in type inference (type bound)"))::Union{}
└──       unreachable
9 ┄ %19 = φ (#2 => false, #7 => %15)::Bool
└──       return %19
) => Bool
```

Inlining/static-dispatch of abstract union-split callsite will improve
the performance in such situations (and so this commit will improve the
latency of our JIT compilation). Especially, this commit helps us avoid
excessive specializations of `Core.Compiler` code by statically-resolving
`@nospecialize`d callsites, and as the result, the # of precompiled
statements is now reduced from  `2005` ([`master`](f782430)) to `1912` (this commit).

And also, as a side effect, the implementation of our inlining algorithm
gets much simplified now since we no longer need the previous special
handlings for abstract callsites.

One possible drawback would be increased code size.
This change seems to certainly increase the size of sysimage,
but I think these numbers are in an acceptable range:
> [`master`](f782430)
```
❯ du -shk usr/lib/julia/*
17604	usr/lib/julia/corecompiler.ji
194072	usr/lib/julia/sys-o.a
169424	usr/lib/julia/sys.dylib
23784	usr/lib/julia/sys.dylib.dSYM
103772	usr/lib/julia/sys.ji
```

> this commit
```
❯ du -shk usr/lib/julia/*
17512	usr/lib/julia/corecompiler.ji
195588	usr/lib/julia/sys-o.a
170908	usr/lib/julia/sys.dylib
23776	usr/lib/julia/sys.dylib.dSYM
105360	usr/lib/julia/sys.ji
```
N5N3 pushed a commit that referenced this pull request Jun 27, 2022
…Lang#45790)

Currently the `@nospecialize`-d `push!(::Vector{Any}, ...)` can only
take a single item and we will end up with runtime dispatch when we try
to call it with multiple items:
```julia
julia> code_typed(push!, (Vector{Any}, Any))
1-element Vector{Any}:
 CodeInfo(
1 ─      $(Expr(:foreigncall, :(:jl_array_grow_end), Nothing, svec(Any, UInt64), 0, :(:ccall), Core.Argument(2), 0x0000000000000001, 0x0000000000000001))::Nothing
│   %2 = Base.arraylen(a)::Int64
│        Base.arrayset(true, a, item, %2)::Vector{Any}
└──      return a
) => Vector{Any}

julia> code_typed(push!, (Vector{Any}, Any, Any))
1-element Vector{Any}:
 CodeInfo(
1 ─ %1 = Base.append!(a, iter)::Vector{Any}
└──      return %1
) => Vector{Any}
```

This commit adds a new specialization that it can take arbitrary-length
items. Our compiler should still be able to optimize the single-input 
case as before via the dispatch mechanism.
```julia
julia> code_typed(push!, (Vector{Any}, Any))
1-element Vector{Any}:
 CodeInfo(
1 ─      $(Expr(:foreigncall, :(:jl_array_grow_end), Nothing, svec(Any, UInt64), 0, :(:ccall), Core.Argument(2), 0x0000000000000001, 0x0000000000000001))::Nothing
│   %2 = Base.arraylen(a)::Int64
│        Base.arrayset(true, a, item, %2)::Vector{Any}
└──      return a
) => Vector{Any}

julia> code_typed(push!, (Vector{Any}, Any, Any))
1-element Vector{Any}:
 CodeInfo(
1 ─ %1  = Base.arraylen(a)::Int64
│         $(Expr(:foreigncall, :(:jl_array_grow_end), Nothing, svec(Any, UInt64), 0, :(:ccall), Core.Argument(2), 0x0000000000000002, 0x0000000000000002))::Nothing
└──       goto #7 if not true
2 ┄ %4  = φ (#1 => 1, #6 => %14)::Int64
│   %5  = φ (#1 => 1, #6 => %15)::Int64
│   %6  = Base.getfield(x, %4, true)::Any
│   %7  = Base.add_int(%1, %4)::Int64
│         Base.arrayset(true, a, %6, %7)::Vector{Any}
│   %9  = (%5 === 2)::Bool
└──       goto #4 if not %9
3 ─       goto #5
4 ─ %12 = Base.add_int(%5, 1)::Int64
└──       goto #5
5 ┄ %14 = φ (#4 => %12)::Int64
│   %15 = φ (#4 => %12)::Int64
│   %16 = φ (#3 => true, #4 => false)::Bool
│   %17 = Base.not_int(%16)::Bool
└──       goto #7 if not %17
6 ─       goto #2
7 ┄       return a
) => Vector{Any}
```

This commit also adds the equivalent implementations for `pushfirst!`.
N5N3 pushed a commit that referenced this pull request Jun 29, 2022
When calling `jl_error()` or `jl_errorf()`, we must check to see if we
are so early in the bringup process that it is dangerous to attempt to
construct a backtrace because the data structures used to provide line
information are not properly setup.

This can be easily triggered by running:

```
julia -C invalid
```

On an `i686-linux-gnu` build, this will hit the "Invalid CPU Name"
branch in `jitlayers.cpp`, which calls `jl_errorf()`.  This in turn
calls `jl_throw()`, which will eventually call `jl_DI_for_fptr` as part
of the backtrace printing process, which fails as the object maps are
not fully initialized.  See the below `gdb` stacktrace for details:

```
$ gdb -batch -ex 'r' -ex 'bt' --args ./julia -C invalid
...
fatal: error thrown and no exception handler available.
ErrorException("Invalid CPU name "invalid".")

Thread 1 "julia" received signal SIGSEGV, Segmentation fault.
0xf75bd665 in std::_Rb_tree<unsigned int, std::pair<unsigned int const, JITDebugInfoRegistry::ObjectInfo>, std::_Select1st<std::pair<unsigned int const, JITDebugInfoRegistry::ObjectInfo> >, std::greater<unsigned int>, std::allocator<std::pair<unsigned int const, JITDebugInfoRegistry::ObjectInfo> > >::lower_bound (__k=<optimized out>, this=0x248) at /usr/local/i686-linux-gnu/include/c++/9.1.0/bits/stl_tree.h:1277
1277    /usr/local/i686-linux-gnu/include/c++/9.1.0/bits/stl_tree.h: No such file or directory.
 #0  0xf75bd665 in std::_Rb_tree<unsigned int, std::pair<unsigned int const, JITDebugInfoRegistry::ObjectInfo>, std::_Select1st<std::pair<unsigned int const, JITDebugInfoRegistry::ObjectInfo> >, std::greater<unsigned int>, std::allocator<std::pair<unsigned int const, JITDebugInfoRegistry::ObjectInfo> > >::lower_bound (__k=<optimized out>, this=0x248) at /usr/local/i686-linux-gnu/include/c++/9.1.0/bits/stl_tree.h:1277
 #1  std::map<unsigned int, JITDebugInfoRegistry::ObjectInfo, std::greater<unsigned int>, std::allocator<std::pair<unsigned int const, JITDebugInfoRegistry::ObjectInfo> > >::lower_bound (__x=<optimized out>, this=0x248) at /usr/local/i686-linux-gnu/include/c++/9.1.0/bits/stl_map.h:1258
 #2  jl_DI_for_fptr (fptr=4155049385, symsize=symsize@entry=0xffffcfa8, slide=slide@entry=0xffffcfa0, Section=Section@entry=0xffffcfb8, context=context@entry=0xffffcf94) at /cache/build/default-amdci5-4/julialang/julia-master/src/debuginfo.cpp:1181
 #3  0xf75c056a in jl_getFunctionInfo_impl (frames_out=0xffffd03c, pointer=4155049385, skipC=0, noInline=0) at /cache/build/default-amdci5-4/julialang/julia-master/src/debuginfo.cpp:1210
 #4  0xf7a6ca98 in jl_print_native_codeloc (ip=4155049385) at /cache/build/default-amdci5-4/julialang/julia-master/src/stackwalk.c:636
 #5  0xf7a6cd54 in jl_print_bt_entry_codeloc (bt_entry=0xf0798018) at /cache/build/default-amdci5-4/julialang/julia-master/src/stackwalk.c:657
 #6  jlbacktrace () at /cache/build/default-amdci5-4/julialang/julia-master/src/stackwalk.c:1090
 #7  0xf7a3cd2b in ijl_no_exc_handler (e=0xf0794010) at /cache/build/default-amdci5-4/julialang/julia-master/src/task.c:605
 #8  0xf7a3d10a in throw_internal (ct=ct@entry=0xf070c010, exception=<optimized out>, exception@entry=0xf0794010) at /cache/build/default-amdci5-4/julialang/julia-master/src/task.c:638
 #9  0xf7a3d330 in ijl_throw (e=0xf0794010) at /cache/build/default-amdci5-4/julialang/julia-master/src/task.c:654
 #10 0xf7a905aa in ijl_errorf (fmt=fmt@entry=0xf7647cd4 "Invalid CPU name \"%s\".") at /cache/build/default-amdci5-4/julialang/julia-master/src/rtutils.c:77
 #11 0xf75a4b22 in (anonymous namespace)::createTargetMachine () at /cache/build/default-amdci5-4/julialang/julia-master/src/jitlayers.cpp:823
 #12 JuliaOJIT::JuliaOJIT (this=<optimized out>) at /cache/build/default-amdci5-4/julialang/julia-master/src/jitlayers.cpp:1044
 #13 0xf7531793 in jl_init_llvm () at /cache/build/default-amdci5-4/julialang/julia-master/src/codegen.cpp:8585
 #14 0xf75318a8 in jl_init_codegen_impl () at /cache/build/default-amdci5-4/julialang/julia-master/src/codegen.cpp:8648
 #15 0xf7a51a52 in jl_restore_system_image_from_stream (f=<optimized out>) at /cache/build/default-amdci5-4/julialang/julia-master/src/staticdata.c:2131
 #16 0xf7a55c03 in ijl_restore_system_image_data (buf=0xe859c1c0 <jl_system_image_data> "8'\031\003", len=125161105) at /cache/build/default-amdci5-4/julialang/julia-master/src/staticdata.c:2184
 #17 0xf7a55cf9 in jl_load_sysimg_so () at /cache/build/default-amdci5-4/julialang/julia-master/src/staticdata.c:424
 #18 ijl_restore_system_image (fname=0x80a0900 "/build/bk_download/julia-d78fdad601/lib/julia/sys.so") at /cache/build/default-amdci5-4/julialang/julia-master/src/staticdata.c:2157
 #19 0xf7a3bdfc in _finish_julia_init (rel=rel@entry=JL_IMAGE_JULIA_HOME, ct=<optimized out>, ptls=<optimized out>) at /cache/build/default-amdci5-4/julialang/julia-master/src/init.c:741
 #20 0xf7a3c8ac in julia_init (rel=<optimized out>) at /cache/build/default-amdci5-4/julialang/julia-master/src/init.c:728
 #21 0xf7a7f61d in jl_repl_entrypoint (argc=<optimized out>, argv=0xffffddf4) at /cache/build/default-amdci5-4/julialang/julia-master/src/jlapi.c:705
 #22 0x080490a7 in main (argc=3, argv=0xffffddf4) at /cache/build/default-amdci5-4/julialang/julia-master/cli/loader_exe.c:59
```

To prevent this, we simply avoid calling `jl_errorf` this early in the
process, punting the problem to a later PR that can update guard
conditions within `jl_error*`.
N5N3 pushed a commit that referenced this pull request Jul 18, 2023
This makes it easier to correlate LLVM IR with the originating source
code by including both argument name and argument type in the LLVM
argument variable.

<details>
<summary>Example 1</summary>

```julia
julia> function f(a, b, c, d, g...)
           e = a + b + c + d
           f = does_not_exist(e) + e
           f
       end
f (generic function with 1 method)

julia> @code_llvm f(0,0,0,0,0)
```
```llvm
;  @ REPL[1]:1 within `f`
define nonnull {}* @julia_f_141(i64 signext %"a::Int64", i64 signext %"b::Int64", i64 signext %"c::Int64", i64 signext %"d::Int64", i64 signext %"g[0]::Int64") #0 {
top:
  %0 = alloca [2 x {}*], align 8
  %gcframe3 = alloca [4 x {}*], align 16
  %gcframe3.sub = getelementptr inbounds [4 x {}*], [4 x {}*]* %gcframe3, i64 0, i64 0
  %1 = bitcast [4 x {}*]* %gcframe3 to i8*
  call void @llvm.memset.p0i8.i64(i8* align 16 %1, i8 0, i64 32, i1 true)
  %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #7
  %tls_ppgcstack = getelementptr i8, i8* %thread_ptr, i64 -8
  %2 = bitcast i8* %tls_ppgcstack to {}****
  %tls_pgcstack = load {}***, {}**** %2, align 8
;  @ REPL[1]:3 within `f`
  %3 = bitcast [4 x {}*]* %gcframe3 to i64*
  store i64 8, i64* %3, align 16
  %4 = getelementptr inbounds [4 x {}*], [4 x {}*]* %gcframe3, i64 0, i64 1
  %5 = bitcast {}** %4 to {}***
  %6 = load {}**, {}*** %tls_pgcstack, align 8
  store {}** %6, {}*** %5, align 8
  %7 = bitcast {}*** %tls_pgcstack to {}***
  store {}** %gcframe3.sub, {}*** %7, align 8
  %Main.does_not_exist.cached = load atomic {}*, {}** @0 unordered, align 8
  %iscached.not = icmp eq {}* %Main.does_not_exist.cached, null
  br i1 %iscached.not, label %notfound, label %found

notfound:                                         ; preds = %top
  %Main.does_not_exist.found = call {}* @ijl_get_binding_or_error({}* nonnull inttoptr (i64 139831437630272 to {}*), {}* nonnull inttoptr (i64 139831600565400 to {}*))
  store atomic {}* %Main.does_not_exist.found, {}** @0 release, align 8
  br label %found

found:                                            ; preds = %notfound, %top
  %Main.does_not_exist = phi {}* [ %Main.does_not_exist.cached, %top ], [ %Main.does_not_exist.found, %notfound ]
  %8 = bitcast {}* %Main.does_not_exist to {}**
  %does_not_exist.checked = load atomic {}*, {}** %8 unordered, align 8
  %.not = icmp eq {}* %does_not_exist.checked, null
  br i1 %.not, label %err, label %ok

err:                                              ; preds = %found
  call void @ijl_undefined_var_error({}* inttoptr (i64 139831600565400 to {}*))
  unreachable

ok:                                               ; preds = %found
  %.sub = getelementptr inbounds [2 x {}*], [2 x {}*]* %0, i64 0, i64 0
;  @ REPL[1]:2 within `f`
; ┌ @ operators.jl:587 within `+` @ int.jl:87
   %9 = add i64 %"b::Int64", %"a::Int64"
   %10 = add i64 %9, %"c::Int64"
; │ @ operators.jl:587 within `+`
; │┌ @ operators.jl:544 within `afoldl`
; ││┌ @ int.jl:87 within `+`
     %11 = add i64 %10, %"d::Int64"
     %12 = getelementptr inbounds [4 x {}*], [4 x {}*]* %gcframe3, i64 0, i64 3
     store {}* %does_not_exist.checked, {}** %12, align 8
; └└└
;  @ REPL[1]:3 within `f`
  %13 = call nonnull {}* @ijl_box_int64(i64 signext %11)
  %14 = getelementptr inbounds [4 x {}*], [4 x {}*]* %gcframe3, i64 0, i64 2
  store {}* %13, {}** %14, align 16
  store {}* %13, {}** %.sub, align 8
  %15 = call nonnull {}* @ijl_apply_generic({}* nonnull %does_not_exist.checked, {}** nonnull %.sub, i32 1)
  store {}* %15, {}** %12, align 8
  %16 = call nonnull {}* @ijl_box_int64(i64 signext %11)
  store {}* %16, {}** %14, align 16
  store {}* %15, {}** %.sub, align 8
  %17 = getelementptr inbounds [2 x {}*], [2 x {}*]* %0, i64 0, i64 1
  store {}* %16, {}** %17, align 8
  %18 = call nonnull {}* @ijl_apply_generic({}* inttoptr (i64 139831370516384 to {}*), {}** nonnull %.sub, i32 2)
  %19 = load {}*, {}** %4, align 8
  %20 = bitcast {}*** %tls_pgcstack to {}**
  store {}* %19, {}** %20, align 8
;  @ REPL[1]:4 within `f`
  ret {}* %18
}
```
</details>

<details>
<summary>Example 2</summary>

```julia
julia> function g(a, b, c, d; kwarg=0)
           a + b + c + d + kwarg
       end
g (generic function with 1 method)

julia> @code_llvm g(0,0,0,0,kwarg=0)
```
```llvm
;  @ REPL[3]:1 within `g`
define i64 @julia_g_160([1 x i64]* nocapture noundef nonnull readonly align 8 dereferenceable(8) %"#1::NamedTuple", i64 signext %"a::Int64", i64 signext %"b::Int64", i64 signext %"c::Int64", i64 signext %"d::Int64") #0 {
top:
  %0 = getelementptr inbounds [1 x i64], [1 x i64]* %"#1::NamedTuple", i64 0, i64 0
; ┌ @ REPL[3]:2 within `#g#1`
; │┌ @ operators.jl:587 within `+` @ int.jl:87
    %1 = add i64 %"b::Int64", %"a::Int64"
    %2 = add i64 %1, %"c::Int64"
; ││ @ operators.jl:587 within `+`
; ││┌ @ operators.jl:544 within `afoldl`
; │││┌ @ int.jl:87 within `+`
      %3 = add i64 %2, %"d::Int64"
; │││└
; │││ @ operators.jl:545 within `afoldl`
; │││┌ @ int.jl:87 within `+`
      %unbox = load i64, i64* %0, align 8
      %4 = add i64 %3, %unbox
; └└└└
  ret i64 %4
}
```
</details>
N5N3 pushed a commit that referenced this pull request Aug 8, 2023
Pass the types to the allocator functions.

-------

Before this PR, we were missing the types for allocations in two cases:

1. allocations from codegen
2. allocations in `gc_managed_realloc_`

The second one is easy: those are always used for buffers, right?

For the first one: we extend the allocation functions called from
codegen, to take the type as a parameter, and set the tag there.

I kept the old interfaces around, since I think that they cannot be
removed due to supporting legacy code?

------

An example of the generated code:
```julia
  %ptls_field6 = getelementptr inbounds {}**, {}*** %4, i64 2
  %13 = bitcast {}*** %ptls_field6 to i8**
  %ptls_load78 = load i8*, i8** %13, align 8
  %box = call noalias nonnull dereferenceable(32) {}* @ijl_gc_pool_alloc_typed(i8* %ptls_load78, i32 1184, i32 32, i64 4366152144) #7
```

Fixes JuliaLang#43688.
Fixes JuliaLang#45268.

Co-authored-by: Valentin Churavy <vchuravy@users.noreply.github.com>
mbauman pushed a commit that referenced this pull request Dec 11, 2023
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
```
N5N3 pushed a commit that referenced this pull request Oct 23, 2024
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).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

9 participants