Backports for 1.12.0-rc3 (or 1.12.0) (#59577)

Author: KristofferC

Compiler/src/ssair/passes.jl

@@ -872,6 +872,49 @@ function perform_lifting!(compact::IncrementalCompact,
     return Pair{Any, PhiNest}(stmt_val, PhiNest(visited_philikes, lifted_philikes, lifted_leaves, reverse_mapping, walker_callback))
 end
 
+function lift_apply_args!(compact::IncrementalCompact, idx::Int, stmt::Expr, 𝕃ₒ::AbstractLattice)
+    # Handle _apply_iterate calls: convert arguments to use `Core.svec`. The behavior of Core.svec (with boxing) better matches the ABI of codegen.
+    compact[idx] = nothing
+    for i in 4:length(stmt.args)  # Skip iterate function, f, and first iterator
+        arg = stmt.args[i]
+        arg_type = argextype(arg, compact)
+        svec_args = nothing
+        if isa(arg_type, DataType) && arg_type.name === Tuple.name
+            if isa(arg, SSAValue)
+                arg_stmt = compact[arg][:stmt]
+                if is_known_call(arg_stmt, Core.tuple, compact)
+                    svec_args = copy(arg_stmt.args)
+                end
+            end
+            if svec_args === nothing
+                # Fallback path: generate getfield calls for tuple elements
+                tuple_length = length(arg_type.parameters)
+                if tuple_length > 0 && !isvarargtype(arg_type.parameters[tuple_length])
+                    svec_args = Vector{Any}(undef, tuple_length + 1)
+                    for j in 1:tuple_length
+                        getfield_call = Expr(:call, GlobalRef(Core, :getfield), arg, j)
+                        getfield_type = arg_type.parameters[j]
+                        inst = compact[SSAValue(idx)]
+                        getfield_ssa = insert_node!(compact, SSAValue(idx), NewInstruction(getfield_call, getfield_type, NoCallInfo(), inst[:line], inst[:flag]))
+                        svec_args[j + 1] = getfield_ssa
+                    end
+                end
+            end
+        end
+        # Create Core.svec call if we have arguments
+        if svec_args !== nothing
+            svec_args[1] = GlobalRef(Core, :svec)
+            new_svec_call = Expr(:call)
+            new_svec_call.args = svec_args
+            inst = compact[SSAValue(idx)]
+            new_svec_ssa = insert_node!(compact, SSAValue(idx), NewInstruction(new_svec_call, SimpleVector, NoCallInfo(), inst[:line], inst[:flag]))
+            stmt.args[i] = new_svec_ssa
+        end
+    end
+    compact[idx] = stmt
+    nothing
+end
+
 function lift_svec_ref!(compact::IncrementalCompact, idx::Int, stmt::Expr)
     length(stmt.args) != 3 && return
 
@@ -1375,6 +1418,9 @@ function sroa_pass!(ir::IRCode, inlining::Union{Nothing,InliningState}=nothing)
                 compact[SSAValue(idx)] = (compact[enter_ssa][:stmt]::EnterNode).scope
             elseif isexpr(stmt, :new)
                 refine_new_effects!(𝕃ₒ, compact, idx, stmt)
+            elseif is_known_call(stmt, Core._apply_iterate, compact)
+                length(stmt.args) >= 4 || continue
+                lift_apply_args!(compact, idx, stmt, 𝕃ₒ)
             end
             continue
         end

Compiler/src/tfuncs.jl

@@ -585,6 +585,15 @@ end
 add_tfunc(nfields, 1, 1, nfields_tfunc, 1)
 add_tfunc(Core._expr, 1, INT_INF, @nospecs((𝕃::AbstractLattice, args...)->Expr), 100)
 add_tfunc(svec, 0, INT_INF, @nospecs((𝕃::AbstractLattice, args...)->SimpleVector), 20)
+
+@nospecs function _svec_len_tfunc(𝕃::AbstractLattice, s)
+    if isa(s, Const) && isa(s.val, SimpleVector)
+        return Const(length(s.val))
+    end
+    return Int
+end
+add_tfunc(Core._svec_len, 1, 1, _svec_len_tfunc, 1)
+
 @nospecs function _svec_ref_tfunc(𝕃::AbstractLattice, s, i)
     if isa(s, Const) && isa(i, Const)
         s, i = s.val, i.val
@@ -1986,15 +1995,8 @@ function tuple_tfunc(𝕃::AbstractLattice, argtypes::Vector{Any})
         # UnionAll context is missing around this.
         pop!(argtypes)
     end
-    all_are_const = true
-    for i in 1:length(argtypes)
-        if !isa(argtypes[i], Const)
-            all_are_const = false
-            break
-        end
-    end
-    if all_are_const
-        return Const(ntuple(i::Int->argtypes[i].val, length(argtypes)))
+    if is_all_const_arg(argtypes, 1) # repeated from builtin_tfunction for the benefit of callers that use this tfunc directly
+        return Const(tuple(collect_const_args(argtypes, 1)...))
     end
     params = Vector{Any}(undef, length(argtypes))
     anyinfo = false
@@ -2359,14 +2361,17 @@ function _builtin_nothrow(𝕃::AbstractLattice, @nospecialize(f::Builtin), argt
     elseif f === Core.compilerbarrier
         na == 2 || return false
         return compilerbarrier_nothrow(argtypes[1], nothing)
+    elseif f === Core._svec_len
+        na == 1 || return false
+        return _svec_len_tfunc(𝕃, argtypes[1]) isa Const
     elseif f === Core._svec_ref
         na == 2 || return false
         return _svec_ref_tfunc(𝕃, argtypes[1], argtypes[2]) isa Const
     end
     return false
 end
 
-# known to be always effect-free (in particular nothrow)
+# known to be always effect-free (in particular also nothrow)
 const _PURE_BUILTINS = Any[
     tuple,
     svec,
@@ -2395,6 +2400,8 @@ const _CONSISTENT_BUILTINS = Any[
     donotdelete,
     memoryrefnew,
     memoryrefoffset,
+    Core._svec_len,
+    Core._svec_ref,
 ]
 
 # known to be effect-free (but not necessarily nothrow)
@@ -2419,6 +2426,7 @@ const _EFFECT_FREE_BUILTINS = [
     Core.throw_methoderror,
     getglobal,
     compilerbarrier,
+    Core._svec_len,
     Core._svec_ref,
 ]
 
@@ -2453,6 +2461,7 @@ const _ARGMEM_BUILTINS = Any[
     replacefield!,
     setfield!,
     swapfield!,
+    Core._svec_len,
     Core._svec_ref,
 ]
 
@@ -2637,7 +2646,7 @@ function builtin_effects(𝕃::AbstractLattice, @nospecialize(f::Builtin), argty
     else
         if contains_is(_CONSISTENT_BUILTINS, f)
             consistent = ALWAYS_TRUE
-        elseif f === memoryrefget || f === memoryrefset! || f === memoryref_isassigned || f === Core._svec_ref
+        elseif f === memoryrefget || f === memoryrefset! || f === memoryref_isassigned || f === Core._svec_len || f === Core._svec_ref
             consistent = CONSISTENT_IF_INACCESSIBLEMEMONLY
         elseif f === Core._typevar || f === Core.memorynew
             consistent = CONSISTENT_IF_NOTRETURNED
@@ -2746,11 +2755,12 @@ end
 function builtin_tfunction(interp::AbstractInterpreter, @nospecialize(f), argtypes::Vector{Any},
                            sv::Union{AbsIntState, Nothing})
     𝕃ᵢ = typeinf_lattice(interp)
-    if isa(f, IntrinsicFunction)
-        if is_pure_intrinsic_infer(f) && all(@nospecialize(a) -> isa(a, Const), argtypes)
-            argvals = anymap(@nospecialize(a) -> (a::Const).val, argtypes)
+    # Early constant evaluation for foldable builtins with all const args
+    if isa(f, IntrinsicFunction) ? is_pure_intrinsic_infer(f) : (f in _PURE_BUILTINS || (f in _CONSISTENT_BUILTINS && f in _EFFECT_FREE_BUILTINS))
+        if is_all_const_arg(argtypes, 1)
+            argvals = collect_const_args(argtypes, 1)
             try
-                # unroll a few cases which have specialized codegen
+                # unroll a few common cases for better codegen
                 if length(argvals) == 1
                     return Const(f(argvals[1]))
                 elseif length(argvals) == 2
@@ -2764,6 +2774,8 @@ function builtin_tfunction(interp::AbstractInterpreter, @nospecialize(f), argtyp
                 return Bottom
             end
         end
+    end
+    if isa(f, IntrinsicFunction)
         iidx = Int(reinterpret(Int32, f)) + 1
         if iidx < 0 || iidx > length(T_IFUNC)
             # unknown intrinsic
@@ -2790,6 +2802,7 @@ function builtin_tfunction(interp::AbstractInterpreter, @nospecialize(f), argtyp
         end
         tf = T_FFUNC_VAL[fidx]
     end
+
     if hasvarargtype(argtypes)
         if length(argtypes) - 1 > tf[2]
             # definitely too many arguments

Compiler/test/codegen.jl

@@ -133,14 +133,14 @@ if !is_debug_build && opt_level > 0
     # Array
     test_loads_no_call(strip_debug_calls(get_llvm(sizeof, Tuple{Vector{Int}})), [Iptr])
     # As long as the eltype is known we don't need to load the elsize, but do need to check isvector
-    @test_skip test_loads_no_call(strip_debug_calls(get_llvm(sizeof, Tuple{Array{Any}})), ["atomic $Iptr", "ptr", "ptr", Iptr, Iptr, "ptr",  Iptr])
+    @test_skip test_loads_no_call(strip_debug_calls(get_llvm(sizeof, Tuple{Array{Any}})), ["atomic volatile $Iptr", "ptr", "ptr", Iptr, Iptr, "ptr",  Iptr])
     # Memory
     test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory{Int}})), [Iptr])
     # As long as the eltype is known we don't need to load the elsize
     test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory{Any}})), [Iptr])
     # Check that we load the elsize and isunion from the typeof layout
-    test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory})), [Iptr, "atomic $Iptr", "ptr", "i32", "i16"])
-    test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory})), [Iptr, "atomic $Iptr", "ptr", "i32", "i16"])
+    test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory})), [Iptr, "atomic volatile $Iptr", "ptr", "i32", "i16"])
+    test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory})), [Iptr, "atomic volatile $Iptr", "ptr", "i32", "i16"])
     # Primitive Type size should be folded to a constant
     test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Ptr})), String[])
 

Compiler/test/effects.jl

@@ -1466,7 +1466,7 @@ end
 let effects = Base.infer_effects((Core.SimpleVector,Int); optimize=false) do svec, i
         Core._svec_ref(svec, i)
     end
-    @test !Compiler.is_consistent(effects)
+    @test Compiler.is_consistent(effects)
     @test Compiler.is_effect_free(effects)
     @test !Compiler.is_nothrow(effects)
     @test Compiler.is_terminates(effects)

base/Base.jl

@@ -174,6 +174,7 @@ using .Filesystem
 include("cmd.jl")
 include("process.jl")
 include("terminfo.jl")
+include("Terminals.jl") # Moved from REPL to reduce invalidations
 include("secretbuffer.jl")
 
 # core math functions

stdlib/REPL/src/Terminals.jl renamed to base/Terminals.jl

@@ -934,11 +934,7 @@ setindex!(A::MemoryRef{Any}, @nospecialize(x)) = (memoryrefset!(A, x, :not_atomi
 
 getindex(v::SimpleVector, i::Int) = (@_foldable_meta; Core._svec_ref(v, i))
 function length(v::SimpleVector)
-    @_total_meta
-    t = @_gc_preserve_begin v
-    len = unsafe_load(Ptr{Int}(pointer_from_objref(v)))
-    @_gc_preserve_end t
-    return len
+    Core._svec_len(v)
 end
 firstindex(v::SimpleVector) = 1
 lastindex(v::SimpleVector) = length(v)

base/essentials.jl

@@ -563,12 +563,15 @@ function _sort!(v::UnwrappableSubArray, a::SubArrayOptimization, o::Ordering, kw
     @getkw lo hi
     # @assert v.stride1 == 1
     parent = v.parent
-    if parent isa Array && !(parent isa Vector) && hi - lo < 100
+    if parent isa Array && !(parent isa Vector) && hi - lo < 100 || !iszero(v.offset1)
         # vec(::Array{T, ≠1}) allocates and is therefore somewhat expensive.
         # We don't want that for small inputs.
+
+        # Additionally, if offset1 is non-zero, then this optimization is incompatible with
+        # algorithms that track absolute first and last indices (e.g. ScratchQuickSort)
         _sort!(v, a.next, o, kw)
     else
-        _sort!(vec(parent), a.next, o, (;kw..., lo = lo + v.offset1, hi = hi + v.offset1))
+        _sort!(vec(parent), a.next, o, kw)
     end
 end
 

base/sort.jl

@@ -378,7 +378,7 @@ end
 
 function isopen(x::Union{LibuvStream, LibuvServer})
     if x.status == StatusUninit || x.status == StatusInit || x.handle === C_NULL
-        throw(ArgumentError("$x is not initialized"))
+        throw(ArgumentError("stream not initialized"))
     end
     return x.status != StatusClosed
 end

base/stream.jl

@@ -34,6 +34,14 @@ hardcoded_precompile_statements = """
 precompile(Base.unsafe_string, (Ptr{UInt8},))
 precompile(Base.unsafe_string, (Ptr{Int8},))
 
+# used by REPL
+precompile(Tuple{typeof(Base.getproperty), Base.Terminals.TTYTerminal, Symbol})
+precompile(Tuple{typeof(Base.reseteof), Base.Terminals.TTYTerminal})
+precompile(Tuple{typeof(Base.Terminals.enable_bracketed_paste), Base.Terminals.TTYTerminal})
+precompile(Tuple{typeof(Base.Terminals.width), Base.Terminals.TTYTerminal})
+precompile(Tuple{typeof(Base.Terminals.height), Base.Terminals.TTYTerminal})
+precompile(Tuple{typeof(Base.write), Base.Terminals.TTYTerminal, Array{UInt8, 1}})
+
 # loading.jl - without these each precompile worker would precompile these because they're hit before pkgimages are loaded
 precompile(Base.__require, (Module, Symbol))
 precompile(Base.__require, (Base.PkgId,))