Merge pull request #22019 from JuliaLang/jn/inferrable-functions

improve inferability of base

Author: vtjnash

base/abstractarray.jl

@@ -202,13 +202,11 @@ julia> strides(A)
 (1, 3, 12)
 ```
 """
-strides(A::AbstractArray) = _strides((1,), A)
-_strides(out::Tuple{Int}, A::AbstractArray{<:Any,0}) = ()
-_strides(out::NTuple{N,Int}, A::AbstractArray{<:Any,N}) where {N} = out
-function _strides(out::NTuple{M,Int}, A::AbstractArray) where M
-    @_inline_meta
-    _strides((out..., out[M]*size(A, M)), A)
-end
+strides(A::AbstractArray) = size_to_strides(1, size(A)...)
+@inline size_to_strides(s, d, sz...) = (s, size_to_strides(s * d, sz...)...)
+size_to_strides(s, d) = (s,)
+size_to_strides(s) = ()
+
 
 function isassigned(a::AbstractArray, i::Int...)
     try
@@ -1160,30 +1158,32 @@ cat_similar(A::AbstractArray, T, shape) = similar(A, T, shape)
 
 cat_shape(dims, shape::Tuple) = shape
 @inline cat_shape(dims, shape::Tuple, nshape::Tuple, shapes::Tuple...) =
-    cat_shape(dims, _cshp(dims, (), shape, nshape), shapes...)
+    cat_shape(dims, _cshp(1, dims, shape, nshape), shapes...)
 
-_cshp(::Tuple{}, out, ::Tuple{}, ::Tuple{}) = out
-_cshp(::Tuple{}, out, ::Tuple{}, nshape) = (out..., nshape...)
-_cshp(dims, out, ::Tuple{}, ::Tuple{}) = (out..., map(b -> 1, dims)...)
-@inline _cshp(dims, out, shape, ::Tuple{}) =
-    _cshp(tail(dims), (out..., shape[1] + dims[1]), tail(shape), ())
-@inline _cshp(dims, out, ::Tuple{}, nshape) =
-    _cshp(tail(dims), (out..., nshape[1]), (), tail(nshape))
-@inline function _cshp(::Tuple{}, out, shape, ::Tuple{})
-    _cs(length(out) + 1, false, shape[1], 1)
-    _cshp((), (out..., 1), tail(shape), ())
+_cshp(ndim::Int, ::Tuple{}, ::Tuple{}, ::Tuple{}) = ()
+_cshp(ndim::Int, ::Tuple{}, ::Tuple{}, nshape) = nshape
+_cshp(ndim::Int, dims, ::Tuple{}, ::Tuple{}) = ntuple(b -> 1, Val{length(dims)})
+@inline _cshp(ndim::Int, dims, shape, ::Tuple{}) =
+    (shape[1] + dims[1], _cshp(ndim + 1, tail(dims), tail(shape), ())...)
+@inline _cshp(ndim::Int, dims, ::Tuple{}, nshape) =
+    (nshape[1], _cshp(ndim + 1, tail(dims), (), tail(nshape))...)
+@inline function _cshp(ndim::Int, ::Tuple{}, shape, ::Tuple{})
+    _cs(ndim, shape[1], 1)
+    (1, _cshp(ndim + 1, (), tail(shape), ())...)
 end
-@inline function _cshp(::Tuple{}, out, shape, nshape)
-    next = _cs(length(out) + 1, false, shape[1], nshape[1])
-    _cshp((), (out..., next), tail(shape), tail(nshape))
+@inline function _cshp(ndim::Int, ::Tuple{}, shape, nshape)
+    next = _cs(ndim, shape[1], nshape[1])
+    (next, _cshp(ndim + 1, (), tail(shape), tail(nshape))...)
 end
-@inline function _cshp(dims, out, shape, nshape)
-    next = _cs(length(out) + 1, dims[1], shape[1], nshape[1])
-    _cshp(tail(dims), (out..., next), tail(shape), tail(nshape))
+@inline function _cshp(ndim::Int, dims, shape, nshape)
+    a = shape[1]
+    b = nshape[1]
+    next = dims[1] ? a + b : _cs(ndim, a, b)
+    (next, _cshp(ndim + 1, tail(dims), tail(shape), tail(nshape))...)
 end
 
-_cs(d, concat, a, b) = concat ? (a + b) : (a == b ? a : throw(DimensionMismatch(string(
-    "mismatch in dimension ", d, " (expected ", a, " got ", b, ")"))))
+_cs(d, a, b) = (a == b ? a : throw(DimensionMismatch(
+    "mismatch in dimension $d (expected $a got $b)")))
 
 dims2cat{n}(::Type{Val{n}}) = ntuple(i -> (i == n), Val{n})
 dims2cat(dims) = ntuple(i -> (i in dims), maximum(dims))
@@ -1668,15 +1668,15 @@ end
 function _sub2ind!(Iout, inds, Iinds, I)
     @_noinline_meta
     for i in Iinds
-        # Iout[i] = sub2ind(inds, map(Ij->Ij[i], I)...)
+        # Iout[i] = sub2ind(inds, map(Ij -> Ij[i], I)...)
         Iout[i] = sub2ind_vec(inds, i, I)
     end
     Iout
 end
 
-sub2ind_vec(inds, i, I) = (@_inline_meta; _sub2ind_vec(inds, (), i, I...))
-_sub2ind_vec(inds, out, i, I1, I...) = (@_inline_meta; _sub2ind_vec(inds, (out..., I1[i]), i, I...))
-_sub2ind_vec(inds, out, i) = (@_inline_meta; sub2ind(inds, out...))
+sub2ind_vec(inds, i, I) = (@_inline_meta; sub2ind(inds, _sub2ind_vec(i, I...)...))
+_sub2ind_vec(i, I1, I...) = (@_inline_meta; (I1[i], _sub2ind_vec(i, I...)...))
+_sub2ind_vec(i) = ()
 
 function ind2sub(inds::Union{DimsInteger{N},Indices{N}}, ind::AbstractVector{<:Integer}) where N
     M = length(ind)

base/array.jl

@@ -73,12 +73,7 @@ end
 size(a::Array, d) = arraysize(a, d)
 size(a::Vector) = (arraysize(a,1),)
 size(a::Matrix) = (arraysize(a,1), arraysize(a,2))
-size(a::Array) = (@_inline_meta; _size((), a))
-_size(out::NTuple{N}, A::Array{_,N}) where {_,N} = out
-function _size(out::NTuple{M}, A::Array{_,N}) where _ where M where N
-    @_inline_meta
-    _size((out..., size(A,M+1)), A)
-end
+size(a::Array{<:Any,N}) where {N} = (@_inline_meta; ntuple(M -> size(a, M), Val{N}))
 
 asize_from(a::Array, n) = n > ndims(a) ? () : (arraysize(a,n), asize_from(a, n+1)...)
 

base/broadcast.jl

@@ -45,23 +45,22 @@ promote_containertype(::Type{T}, ::Type{T}) where {T} = T
 ## Calculate the broadcast indices of the arguments, or error if incompatible
 # array inputs
 broadcast_indices() = ()
-broadcast_indices(A) = broadcast_indices(containertype(A), A)
-broadcast_indices(::ScalarType, A) = ()
-broadcast_indices(::Type{Tuple}, A) = (OneTo(length(A)),)
-broadcast_indices(::Type{Array}, A::Ref) = ()
-broadcast_indices(::Type{Array}, A) = indices(A)
-@inline broadcast_indices(A, B...) = broadcast_shape((), broadcast_indices(A), map(broadcast_indices, B)...)
+broadcast_indices(A) = _broadcast_indices(containertype(A), A)
+@inline broadcast_indices(A, B...) = broadcast_shape(broadcast_indices(A), broadcast_indices(B...))
+_broadcast_indices(::Type, A) = ()
+_broadcast_indices(::Type{Tuple}, A) = (OneTo(length(A)),)
+_broadcast_indices(::Type{Array}, A::Ref) = ()
+_broadcast_indices(::Type{Array}, A) = indices(A)
 
 # shape (i.e., tuple-of-indices) inputs
 broadcast_shape(shape::Tuple) = shape
-@inline broadcast_shape(shape::Tuple, shape1::Tuple, shapes::Tuple...) = broadcast_shape(_bcs((), shape, shape1), shapes...)
+@inline broadcast_shape(shape::Tuple, shape1::Tuple, shapes::Tuple...) = broadcast_shape(_bcs(shape, shape1), shapes...)
 # _bcs consolidates two shapes into a single output shape
-_bcs(out, ::Tuple{}, ::Tuple{}) = out
-@inline _bcs(out, ::Tuple{}, newshape) = _bcs((out..., newshape[1]), (), tail(newshape))
-@inline _bcs(out, shape, ::Tuple{}) = _bcs((out..., shape[1]), tail(shape), ())
-@inline function _bcs(out, shape, newshape)
-    newout = _bcs1(shape[1], newshape[1])
-    _bcs((out..., newout), tail(shape), tail(newshape))
+_bcs(::Tuple{}, ::Tuple{}) = ()
+@inline _bcs(::Tuple{}, newshape::Tuple) = (newshape[1], _bcs((), tail(newshape))...)
+@inline _bcs(shape::Tuple, ::Tuple{}) = (shape[1], _bcs(tail(shape), ())...)
+@inline function _bcs(shape::Tuple, newshape::Tuple)
+    return (_bcs1(shape[1], newshape[1]), _bcs(tail(shape), tail(newshape))...)
 end
 # _bcs1 handles the logic for a single dimension
 _bcs1(a::Integer, b::Integer) = a == 1 ? b : (b == 1 ? a : (a == b ? a : throw(DimensionMismatch("arrays could not be broadcast to a common size"))))

base/inference.jl

@@ -15,6 +15,7 @@ struct InferenceParams
     inlining::Bool
 
     # parameters limiting potentially-infinite types (configurable)
+    MAX_METHODS::Int
     MAX_TUPLETYPE_LEN::Int
     MAX_TUPLE_DEPTH::Int
     MAX_TUPLE_SPLAT::Int
@@ -24,12 +25,13 @@ struct InferenceParams
     # reasonable defaults
     function InferenceParams(world::UInt;
                     inlining::Bool = inlining_enabled(),
+                    max_methods::Int = 4,
                     tupletype_len::Int = 15,
                     tuple_depth::Int = 4,
                     tuple_splat::Int = 16,
                     union_splitting::Int = 4,
                     apply_union_enum::Int = 8)
-        return new(world, inlining, tupletype_len,
+        return new(world, inlining, max_methods, tupletype_len,
             tuple_depth, tuple_splat, union_splitting, apply_union_enum)
     end
 end
@@ -1280,7 +1282,7 @@ function abstract_call_gf_by_type(f::ANY, atype::ANY, sv::InferenceState)
     end
     min_valid = UInt[typemin(UInt)]
     max_valid = UInt[typemax(UInt)]
-    applicable = _methods_by_ftype(argtype, 4, sv.params.world, min_valid, max_valid)
+    applicable = _methods_by_ftype(argtype, sv.params.MAX_METHODS, sv.params.world, min_valid, max_valid)
     rettype = Bottom
     if applicable === false
         # this means too many methods matched
@@ -1431,7 +1433,7 @@ function precise_container_type(arg::ANY, typ::ANY, vtypes::VarTable, sv::Infere
     if isa(typ, Const)
         val = typ.val
         if isa(val, SimpleVector) || isa(val, Tuple)
-            return Any[ abstract_eval_constant(x) for x in val ]
+            return Any[ Const(val[i]) for i in 1:length(val) ] # avoid making a tuple Generator here!
         end
     end
 
@@ -1499,44 +1501,64 @@ function abstract_iteration(itertype::ANY, vtypes::VarTable, sv::InferenceState)
     return Vararg{valtype}
 end
 
+function tuple_tail_elem(init::ANY, ct)
+    return Vararg{widenconst(foldl((a, b) -> tmerge(a, unwrapva(b)), init, ct))}
+end
+
 # do apply(af, fargs...), where af is a function value
-function abstract_apply(af::ANY, fargs::Vector{Any}, aargtypes::Vector{Any}, vtypes::VarTable, sv::InferenceState)
+function abstract_apply(aft::ANY, fargs::Vector{Any}, aargtypes::Vector{Any}, vtypes::VarTable, sv::InferenceState)
+    if !isa(aft, Const) && !isconstType(aft)
+        if !(isleaftype(aft) || aft <: Type) || (aft <: Builtin) || (aft <: IntrinsicFunction)
+            return Any
+        end
+        # non-constant function, but type is known
+    end
     res = Union{}
     nargs = length(fargs)
     assert(nargs == length(aargtypes))
-    splitunions = countunionsplit(aargtypes) <= sv.params.MAX_APPLY_UNION_ENUM
-    ctypes = Any[Any[]]
+    splitunions = 1 < countunionsplit(aargtypes) <= sv.params.MAX_APPLY_UNION_ENUM
+    ctypes = Any[Any[aft]]
     for i = 1:nargs
         if aargtypes[i] === Any
             # bail out completely and infer as f(::Any...)
-            # instead could keep what we got so far and just append a Vararg{Any} (by just
-            # using the normal logic from below), but that makes the time of the subarray
-            # test explode
-            ctypes = Any[Any[Vararg{Any}]]
+            # instead could infer the precise types for the types up to this point and just append a Vararg{Any}
+            # (by just using the normal logic from below), but that makes the time of the subarray test explode
+            push!(ctypes[1], Vararg{Any})
             break
         end
-        ctypes´ = []
-        for ti in (splitunions ? uniontypes(aargtypes[i]) : Any[aargtypes[i]])
-            cti = precise_container_type(fargs[i], ti, vtypes, sv)
-            for ct in ctypes
-                if !isempty(ct) && isvarargtype(ct[end])
-                    tail = foldl((a,b)->tmerge(a,unwrapva(b)), unwrapva(ct[end]), cti)
-                    push!(ctypes´, push!(ct[1:end-1], Vararg{widenconst(tail)}))
-                else
-                    push!(ctypes´, append_any(ct, cti))
+    end
+    if length(ctypes[1]) == 1
+        for i = 1:nargs
+            ctypes´ = []
+            for ti in (splitunions ? uniontypes(aargtypes[i]) : Any[aargtypes[i]])
+                cti = precise_container_type(fargs[i], ti, vtypes, sv)
+                for ct in ctypes
+                    if !isempty(ct) && isvarargtype(ct[end])
+                        tail = tuple_tail_elem(unwrapva(ct[end]), cti)
+                        push!(ctypes´, push!(ct[1:(end - 1)], tail))
+                    else
+                        push!(ctypes´, append_any(ct, cti))
+                    end
                 end
             end
+            ctypes = ctypes´
         end
-        ctypes = ctypes´
     end
     for ct in ctypes
         if length(ct) > sv.params.MAX_TUPLETYPE_LEN
-            tail = foldl((a,b)->tmerge(a,unwrapva(b)), Bottom, ct[sv.params.MAX_TUPLETYPE_LEN:end])
+            tail = tuple_tail_elem(Bottom, ct[sv.params.MAX_TUPLETYPE_LEN:end])
             resize!(ct, sv.params.MAX_TUPLETYPE_LEN)
-            ct[end] = Vararg{widenconst(tail)}
+            ct[end] = tail
+        end
+        if isa(aft, Const)
+            rt = abstract_call(aft.val, (), ct, vtypes, sv)
+        elseif isconstType(aft)
+            rt = abstract_call(aft.parameters[1], (), ct, vtypes, sv)
+        else
+            astype = argtypes_to_type(ct)
+            rt = abstract_call_gf_by_type(nothing, astype, sv)
         end
-        at = append_any(Any[Const(af)], ct)
-        res = tmerge(res, abstract_call(af, (), at, vtypes, sv))
+        res = tmerge(res, rt)
         if res === Any
             break
         end
@@ -1651,20 +1673,7 @@ typename_static(t::ANY) = isType(t) ? _typename(t.parameters[1]) : Any
 function abstract_call(f::ANY, fargs::Union{Tuple{},Vector{Any}}, argtypes::Vector{Any}, vtypes::VarTable, sv::InferenceState)
     if f === _apply
         length(fargs) > 1 || return Any
-        aft = argtypes[2]
-        if isa(aft, Const)
-            af = aft.val
-        else
-            if isType(aft) && isleaftype(aft.parameters[1])
-                af = aft.parameters[1]
-            elseif isleaftype(aft) && isdefined(aft, :instance)
-                af = aft.instance
-            else
-                # TODO jb/functions: take advantage of case where non-constant `af`'s type is known
-                return Any
-            end
-        end
-        return abstract_apply(af, fargs[3:end], argtypes[3:end], vtypes, sv)
+        return abstract_apply(argtypes[2], fargs[3:end], argtypes[3:end], vtypes, sv)
     end
 
     la = length(argtypes)
@@ -2508,12 +2517,14 @@ function typeinf_edge(method::Method, atypes::ANY, sparams::SimpleVector, caller
     frame = resolve_call_cycle!(code, caller)
     if frame === nothing
         code.inInference = true
-        frame = InferenceState(code, true, true, caller.params) # always optimize and cache edge targets
+        frame = InferenceState(code, #=optimize=#true, #=cached=#true, caller.params) # always optimize and cache edge targets
         if frame === nothing
             code.inInference = false
             return Any, nothing
         end
-        frame.parent = caller
+        if caller.cached # don't involve uncached functions in cycle resolution
+            frame.parent = caller
+        end
         typeinf(frame)
         return frame.bestguess, frame.inferred ? frame.linfo : nothing
     end
@@ -2849,6 +2860,7 @@ end
 #### finalize and record the result of running type inference ####
 
 function isinlineable(m::Method, src::CodeInfo)
+    # compute the cost (size) of inlining this code
     inlineable = false
     cost = 1000
     if m.module === _topmod(m.module)
@@ -2941,7 +2953,25 @@ function optimize(me::InferenceState)
     end
 
     # determine and cache inlineability
-    if !me.src.inlineable && !force_noinline && isdefined(me.linfo, :def)
+    if !force_noinline
+        # don't keep ASTs for functions specialized on a Union argument
+        # TODO: this helps avoid a type-system bug mis-computing sparams during intersection
+        sig = unwrap_unionall(me.linfo.specTypes)
+        if isa(sig, DataType) && sig.name === Tuple.name
+            for P in sig.parameters
+                P = unwrap_unionall(P)
+                if isa(P, Union)
+                    force_noinline = true
+                    break
+                end
+            end
+        else
+            force_noinline = true
+        end
+    end
+    if force_noinline
+        me.src.inlineable = false
+    elseif !me.src.inlineable && isdefined(me.linfo, :def)
         me.src.inlineable = isinlineable(me.linfo.def, me.src)
     end
     me.src.inferred = true

base/int.jl

@@ -362,22 +362,24 @@ end
 # @doc isn't available when running in Core at this point.
 # Tuple syntax for documention two function signatures at the same time
 # doesn't work either at this point.
-isdefined(Main, :Base) && for fname in (:mod, :rem)
-    @eval @doc """
-        rem(x::Integer, T::Type{<:Integer}) -> T
-        mod(x::Integer, T::Type{<:Integer}) -> T
-        %(x::Integer, T::Type{<:Integer}) -> T
-
-    Find `y::T` such that `x` ≡ `y` (mod n), where n is the number of integers representable
-    in `T`, and `y` is an integer in `[typemin(T),typemax(T)]`.
-    If `T` can represent any integer (e.g. `T == BigInt`), then this operation corresponds to
-    a conversion to `T`.
-
-    ```jldoctest
-    julia> 129 % Int8
-    -127
-    ```
-    """ -> $fname(x::Integer, T::Type{<:Integer})
+if module_name(current_module()) === :Base
+    for fname in (:mod, :rem)
+        @eval @doc ("""
+            rem(x::Integer, T::Type{<:Integer}) -> T
+            mod(x::Integer, T::Type{<:Integer}) -> T
+            %(x::Integer, T::Type{<:Integer}) -> T
+
+        Find `y::T` such that `x` ≡ `y` (mod n), where n is the number of integers representable
+        in `T`, and `y` is an integer in `[typemin(T),typemax(T)]`.
+        If `T` can represent any integer (e.g. `T == BigInt`), then this operation corresponds to
+        a conversion to `T`.
+
+        ```jldoctest
+        julia> 129 % Int8
+        -127
+        ```
+        """ -> $fname(x::Integer, T::Type{<:Integer}))
+    end
 end
 
 rem(x::T, ::Type{T}) where {T<:Integer} = x

base/multidimensional.jl

@@ -138,9 +138,10 @@ module IteratorsMD
     eachindex(::IndexCartesian, A::AbstractArray) = CartesianRange(indices(A))
 
     @inline eachindex(::IndexCartesian, A::AbstractArray, B::AbstractArray...) =
-        CartesianRange(maxsize((), A, B...))
-    maxsize(sz) = sz
-    @inline maxsize(sz, A, B...) = maxsize(maxt(sz, size(A)), B...)
+        CartesianRange(maxsize(A, B...))
+    maxsize() = ()
+    @inline maxsize(A) = size(A)
+    @inline maxsize(A, B...) = maxt(size(A), maxsize(B...))
     @inline maxt(a::Tuple{}, b::Tuple{}) = ()
     @inline maxt(a::Tuple{}, b::Tuple)   = b
     @inline maxt(a::Tuple,   b::Tuple{}) = a