Merge pull request #45946 from JuliaLang/backports-release-1.8

Backports for 1.8-rc2/1.8.0

base/array.jl

@@ -1833,6 +1833,11 @@ function reverseind(a::AbstractVector, i::Integer)
     first(li) + last(li) - i
 end
 
+# This implementation of `midpoint` is performance-optimized but safe
+# only if `lo <= hi`.
+midpoint(lo::T, hi::T) where T<:Integer = lo + ((hi - lo) >>> 0x01)
+midpoint(lo::Integer, hi::Integer) = midpoint(promote(lo, hi)...)
+
 """
     reverse!(v [, start=1 [, stop=length(v) ]]) -> v
 
@@ -1861,17 +1866,18 @@ julia> A
 """
 function reverse!(v::AbstractVector, start::Integer, stop::Integer=lastindex(v))
     s, n = Int(start), Int(stop)
-    liv = LinearIndices(v)
-    if n <= s  # empty case; ok
-    elseif !(first(liv) ≤ s ≤ last(liv))
-        throw(BoundsError(v, s))
-    elseif !(first(liv) ≤ n ≤ last(liv))
-        throw(BoundsError(v, n))
-    end
-    r = n
-    @inbounds for i in s:div(s+n-1, 2)
-        v[i], v[r] = v[r], v[i]
-        r -= 1
+    if n > s # non-empty and non-trivial
+        liv = LinearIndices(v)
+        if !(first(liv) ≤ s ≤ last(liv))
+            throw(BoundsError(v, s))
+        elseif !(first(liv) ≤ n ≤ last(liv))
+            throw(BoundsError(v, n))
+        end
+        r = n
+        @inbounds for i in s:midpoint(s, n-1)
+            v[i], v[r] = v[r], v[i]
+            r -= 1
+        end
     end
     return v
 end

base/broadcast.jl

@@ -1172,7 +1172,7 @@ Base.@propagate_inbounds dotview(B::BitArray, i::BitArray) = BitMaskedBitArray(B
 Base.show(io::IO, B::BitMaskedBitArray) = foreach(arg->show(io, arg), (typeof(B), (B.parent, B.mask)))
 # Override materialize! to prevent the BitMaskedBitArray from escaping to an overrideable method
 @inline materialize!(B::BitMaskedBitArray, bc::Broadcasted{<:Any,<:Any,typeof(identity),Tuple{Bool}}) = fill!(B, bc.args[1])
-@inline materialize!(B::BitMaskedBitArray, bc::Broadcasted{<:Any}) = materialize!(SubArray(B.parent, to_indices(B.parent, (B.mask,))), bc)
+@inline materialize!(B::BitMaskedBitArray, bc::Broadcasted{<:Any}) = materialize!(@inbounds(view(B.parent, B.mask)), bc)
 function Base.fill!(B::BitMaskedBitArray, b::Bool)
     Bc = B.parent.chunks
     Ic = B.mask.chunks

base/reinterpretarray.jl

@@ -152,23 +152,15 @@ strides(a::Union{DenseArray,StridedReshapedArray,StridedReinterpretArray}) = siz
 stride(A::Union{DenseArray,StridedReshapedArray,StridedReinterpretArray}, k::Integer) =
     k ≤ ndims(A) ? strides(A)[k] : length(A)
 
-function strides(a::ReshapedReinterpretArray)
-    ap = parent(a)
-    els, elp = elsize(a), elsize(ap)
-    stp = strides(ap)
-    els == elp && return stp
-    els < elp && return (1, _checked_strides(stp, els, elp)...)
+function strides(a::ReinterpretArray{T,<:Any,S,<:AbstractArray{S},IsReshaped}) where {T,S,IsReshaped}
+    _checkcontiguous(Bool, a) && return size_to_strides(1, size(a))
+    stp = strides(parent(a))
+    els, elp = sizeof(T), sizeof(S)
+    els == elp && return stp # 0dim parent is also handled here.
+    IsReshaped && els < elp && return (1, _checked_strides(stp, els, elp)...)
     stp[1] == 1 || throw(ArgumentError("Parent must be contiguous in the 1st dimension!"))
-    return _checked_strides(tail(stp), els, elp)
-end
-
-function strides(a::NonReshapedReinterpretArray)
-    ap = parent(a)
-    els, elp = elsize(a), elsize(ap)
-    stp = strides(ap)
-    els == elp && return stp
-    stp[1] == 1 || throw(ArgumentError("Parent must be contiguous in the 1st dimension!"))
-    return (1, _checked_strides(tail(stp), els, elp)...)
+    st′ = _checked_strides(tail(stp), els, elp)
+    return IsReshaped ? st′ : (1, st′...)
 end
 
 @inline function _checked_strides(stp::Tuple, els::Integer, elp::Integer)

base/reshapedarray.jl

@@ -294,14 +294,51 @@ unsafe_convert(::Type{Ptr{T}}, V::SubArray{T,N,P,<:Tuple{Vararg{Union{RangeIndex
     unsafe_convert(Ptr{T}, V.parent) + (first_index(V)-1)*sizeof(T)
 
 
-_checkcontiguous(::Type{Bool}, A::AbstractArray) = size_to_strides(1, size(A)...) == strides(A)
-_checkcontiguous(::Type{Bool}, A::Array) = true
+_checkcontiguous(::Type{Bool}, A::AbstractArray) = false
+# `strides(A::DenseArray)` calls `size_to_strides` by default.
+# Thus it's OK to assume all `DenseArray`s are contiguously stored.
+_checkcontiguous(::Type{Bool}, A::DenseArray) = true
 _checkcontiguous(::Type{Bool}, A::ReshapedArray) = _checkcontiguous(Bool, parent(A))
 _checkcontiguous(::Type{Bool}, A::FastContiguousSubArray) = _checkcontiguous(Bool, parent(A))
 
 function strides(a::ReshapedArray)
-    # We can handle non-contiguous parent if it's a StridedVector
-    ndims(parent(a)) == 1 && return size_to_strides(only(strides(parent(a))), size(a)...)
-    _checkcontiguous(Bool, a) || throw(ArgumentError("Parent must be contiguous."))
-    size_to_strides(1, size(a)...)
+    _checkcontiguous(Bool, a) && return size_to_strides(1, size(a)...)
+    apsz::Dims = size(a.parent)
+    apst::Dims = strides(a.parent)
+    msz, mst, n = merge_adjacent_dim(apsz, apst) # Try to perform "lazy" reshape
+    n == ndims(a.parent) && return size_to_strides(mst, size(a)...) # Parent is stridevector like
+    return _reshaped_strides(size(a), 1, msz, mst, n, apsz, apst)
+end
+
+function _reshaped_strides(::Dims{0}, reshaped::Int, msz::Int, ::Int, ::Int, ::Dims, ::Dims)
+    reshaped == msz && return ()
+    throw(ArgumentError("Input is not strided."))
+end
+function _reshaped_strides(sz::Dims, reshaped::Int, msz::Int, mst::Int, n::Int, apsz::Dims, apst::Dims)
+    st = reshaped * mst
+    reshaped = reshaped * sz[1]
+    if length(sz) > 1 && reshaped == msz && sz[2] != 1
+        msz, mst, n = merge_adjacent_dim(apsz, apst, n + 1)
+        reshaped = 1
+    end
+    sts = _reshaped_strides(tail(sz), reshaped, msz, mst, n, apsz, apst)
+    return (st, sts...)
+end
+
+merge_adjacent_dim(::Dims{0}, ::Dims{0}) = 1, 1, 0
+merge_adjacent_dim(apsz::Dims{1}, apst::Dims{1}) = apsz[1], apst[1], 1
+function merge_adjacent_dim(apsz::Dims{N}, apst::Dims{N}, n::Int = 1) where {N}
+    sz, st = apsz[n], apst[n]
+    while n < N
+        szₙ, stₙ = apsz[n+1], apst[n+1]
+        if sz == 1
+            sz, st = szₙ, stₙ
+        elseif stₙ == st * sz || szₙ == 1
+            sz *= szₙ
+        else
+            break
+        end
+        n += 1
+    end
+    return sz, st, n
 end

base/sort.jl

@@ -11,7 +11,7 @@ using .Base: copymutable, LinearIndices, length, (:),
     AbstractMatrix, AbstractUnitRange, isless, identity, eltype, >, <, <=, >=, |, +, -, *, !,
     extrema, sub_with_overflow, add_with_overflow, oneunit, div, getindex, setindex!,
     length, resize!, fill, Missing, require_one_based_indexing, keytype,
-    UnitRange, max, min
+    UnitRange, max, min, midpoint
 
 using .Base: >>>, !==
 
@@ -166,11 +166,6 @@ same thing as `partialsort!` but leaving `v` unmodified.
 partialsort(v::AbstractVector, k::Union{Integer,OrdinalRange}; kws...) =
     partialsort!(copymutable(v), k; kws...)
 
-# This implementation of `midpoint` is performance-optimized but safe
-# only if `lo <= hi`.
-midpoint(lo::T, hi::T) where T<:Integer = lo + ((hi - lo) >>> 0x01)
-midpoint(lo::Integer, hi::Integer) = midpoint(promote(lo, hi)...)
-
 # reference on sorted binary search:
 #   http://www.tbray.org/ongoing/When/200x/2003/03/22/Binary
 
@@ -505,12 +500,12 @@ function sort!(v::AbstractVector, lo::Integer, hi::Integer, ::InsertionSortAlg,
         j = i
         x = v[i]
         while j > lo
-            if lt(o, x, v[j-1])
-                v[j] = v[j-1]
-                j -= 1
-                continue
+            y = v[j-1]
+            if !lt(o, x, y)
+                break
             end
-            break
+            v[j] = y
+            j -= 1
         end
         v[j] = x
     end

src/Makefile

@@ -115,6 +115,11 @@ PUBLIC_HEADER_TARGETS := $(addprefix $(build_includedir)/julia/,$(notdir $(PUBLI
 LLVM_LDFLAGS := $(shell $(LLVM_CONFIG_HOST) --ldflags)
 LLVM_CXXFLAGS := $(shell $(LLVM_CONFIG_HOST) --cxxflags)
 
+# llvm-config --cxxflags does not return -DNDEBUG
+ifeq ($(shell $(LLVM_CONFIG_HOST) --assertion-mode),OFF)
+LLVM_CXXFLAGS += -DNDEBUG
+endif
+
 ifeq ($(JULIACODEGEN),LLVM)
 ifneq ($(USE_SYSTEM_LLVM),0)
 CG_LLVMLINK += $(LLVM_LDFLAGS) $(shell $(LLVM_CONFIG_HOST) --libs --system-libs)

src/jl_uv.c

@@ -60,6 +60,7 @@ void JL_UV_LOCK(void)
     }
     else {
         jl_atomic_fetch_add_relaxed(&jl_uv_n_waiters, 1);
+        jl_fence(); // [^store_buffering_2]
         jl_wake_libuv();
         JL_LOCK(&jl_uv_mutex);
         jl_atomic_fetch_add_relaxed(&jl_uv_n_waiters, -1);

src/jltypes.c

@@ -420,10 +420,8 @@ static int datatype_name_cmp(jl_value_t *a, jl_value_t *b) JL_NOTSAFEPOINT
 
 // sort singletons first, then DataTypes, then UnionAlls,
 // ties broken alphabetically including module name & type parameters
-static int union_sort_cmp(const void *ap, const void *bp) JL_NOTSAFEPOINT
+static int union_sort_cmp(jl_value_t *a, jl_value_t *b) JL_NOTSAFEPOINT
 {
-    jl_value_t *a = *(jl_value_t**)ap;
-    jl_value_t *b = *(jl_value_t**)bp;
     if (a == NULL)
         return b == NULL ? 0 : 1;
     if (b == NULL)
@@ -458,16 +456,33 @@ static int union_sort_cmp(const void *ap, const void *bp) JL_NOTSAFEPOINT
     }
 }
 
+static void isort_union(jl_value_t **a, size_t len) JL_NOTSAFEPOINT
+{
+    size_t i, j;
+    for (i = 1; i < len; i++) {
+        jl_value_t *x = a[i];
+        for (j = i; j > 0; j--) {
+            jl_value_t *y = a[j - 1];
+            if (!(union_sort_cmp(x, y) < 0))
+                break;
+            a[j] = y;
+        }
+        a[j] = x;
+    }
+}
+
 JL_DLLEXPORT jl_value_t *jl_type_union(jl_value_t **ts, size_t n)
 {
-    if (n == 0) return (jl_value_t*)jl_bottom_type;
+    if (n == 0)
+        return (jl_value_t*)jl_bottom_type;
     size_t i;
-    for(i=0; i < n; i++) {
+    for (i = 0; i < n; i++) {
         jl_value_t *pi = ts[i];
         if (!(jl_is_type(pi) || jl_is_typevar(pi)))
             jl_type_error("Union", (jl_value_t*)jl_type_type, pi);
     }
-    if (n == 1) return ts[0];
+    if (n == 1)
+        return ts[0];
 
     size_t nt = count_union_components(ts, n);
     jl_value_t **temp;
@@ -476,9 +491,9 @@ JL_DLLEXPORT jl_value_t *jl_type_union(jl_value_t **ts, size_t n)
     flatten_type_union(ts, n, temp, &count);
     assert(count == nt);
     size_t j;
-    for(i=0; i < nt; i++) {
-        int has_free = temp[i]!=NULL && jl_has_free_typevars(temp[i]);
-        for(j=0; j < nt; j++) {
+    for (i = 0; i < nt; i++) {
+        int has_free = temp[i] != NULL && jl_has_free_typevars(temp[i]);
+        for (j = 0; j < nt; j++) {
             if (j != i && temp[i] && temp[j]) {
                 if (temp[i] == jl_bottom_type ||
                     temp[j] == (jl_value_t*)jl_any_type ||
@@ -490,7 +505,7 @@ JL_DLLEXPORT jl_value_t *jl_type_union(jl_value_t **ts, size_t n)
             }
         }
     }
-    qsort(temp, nt, sizeof(jl_value_t*), union_sort_cmp);
+    isort_union(temp, nt);
     jl_value_t **ptu = &temp[nt];
     *ptu = jl_bottom_type;
     int k;

src/partr.c

@@ -40,7 +40,14 @@ static const int16_t sleeping = 1;
 //   * 2a: `multiq_insert`
 //   * 2b: `jl_atomic_load_relaxed(&ptls->sleep_check_state)` in `jl_wakeup_thread` returns `not_sleeping`
 // i.e., the dequeuer misses the enqueue and enqueuer misses the sleep state transition.
-
+// [^store_buffering_2]: and also
+// * Enqueuer:
+//   * 1a: `jl_atomic_store_relaxed(jl_uv_n_waiters, 1)` in `JL_UV_LOCK`
+//   * 1b: "cheap read" of `handle->pending` in `uv_async_send` (via `JL_UV_LOCK`) loads `0`
+// * Dequeuer:
+//   * 2a: store `2` to `handle->pending` in `uv_async_send` (via `JL_UV_LOCK` in `jl_task_get_next`)
+//   * 2b: `jl_atomic_load_relaxed(jl_uv_n_waiters)` in `jl_task_get_next` returns `0`
+// i.e., the dequeuer misses the `n_waiters` is set and enqueuer misses the `uv_stop` flag (in `signal_async`) transition to cleared
 
 JULIA_DEBUG_SLEEPWAKE(
 uint64_t wakeup_enter;
@@ -462,7 +469,7 @@ static int may_sleep(jl_ptls_t ptls) JL_NOTSAFEPOINT
     // by the thread itself. As a result, if this returns false, it will
     // continue returning false. If it returns true, we know the total
     // modification order of the fences.
-    jl_fence(); // [^store_buffering_1]
+    jl_fence(); // [^store_buffering_1] [^store_buffering_2]
     return jl_atomic_load_relaxed(&ptls->sleep_check_state) == sleeping;
 }
 

stdlib/LinearAlgebra/src/blas.jl

@@ -169,18 +169,19 @@ end
 # Level 1
 # A help function to pick the pointer and inc for 1d like inputs.
 @inline function vec_pointer_stride(x::AbstractArray, stride0check = nothing)
-    isdense(x) && return pointer(x), 1 # simpify runtime check when possibe
-    ndims(x) == 1 || strides(x) == Base.size_to_strides(stride(x, 1), size(x)...) ||
-        throw(ArgumentError("only support vector like inputs"))
-    st = stride(x, 1)
+    Base._checkcontiguous(Bool, x) && return pointer(x), 1 # simpify runtime check when possibe
+    st, ptr = checkedstride(x), pointer(x)
     isnothing(stride0check) || (st == 0 && throw(stride0check))
-    ptr = st > 0 ? pointer(x) : pointer(x, lastindex(x))
+    ptr += min(st, 0) * sizeof(eltype(x)) * (length(x) - 1)
     ptr, st
 end
-isdense(x) = x isa DenseArray
-isdense(x::Base.FastContiguousSubArray) = isdense(parent(x))
-isdense(x::Base.ReshapedArray) = isdense(parent(x))
-isdense(x::Base.ReinterpretArray) = isdense(parent(x))
+function checkedstride(x::AbstractArray)
+    szs::Dims = size(x)
+    sts::Dims = strides(x)
+    _, st, n = Base.merge_adjacent_dim(szs, sts)
+    n === ndims(x) && return st
+    throw(ArgumentError("only support vector like inputs"))
+end
 ## copy
 
 """
@@ -1565,11 +1566,27 @@ for (mfname, elty) in ((:dsymm_,:Float64),
             require_one_based_indexing(A, B, C)
             m, n = size(C)
             j = checksquare(A)
-            if j != (side == 'L' ? m : n)
-                throw(DimensionMismatch(lazy"A has size $(size(A)), C has size ($m,$n)"))
-            end
-            if size(B,2) != n
-                throw(DimensionMismatch(lazy"B has second dimension $(size(B,2)) but needs to match second dimension of C, $n"))
+            M, N = size(B)
+            if side == 'L'
+                if j != m
+                    throw(DimensionMismatch(lazy"A has first dimension $j but needs to match first dimension of C, $m"))
+                end
+                if N != n
+                    throw(DimensionMismatch(lazy"B has second dimension $N but needs to match second dimension of C, $n"))
+                end
+                if j != M
+                    throw(DimensionMismatch(lazy"A has second dimension $j but needs to match first dimension of B, $M"))
+                end
+            else
+                if j != n
+                    throw(DimensionMismatch(lazy"B has second dimension $j but needs to match second dimension of C, $n"))
+                end
+                if N != j
+                    throw(DimensionMismatch(lazy"A has second dimension $N but needs to match first dimension of B, $j"))
+                end
+                if M != m
+                    throw(DimensionMismatch(lazy"A has first dimension $M but needs to match first dimension of C, $m"))
+                end
             end
             chkstride1(A)
             chkstride1(B)
@@ -1639,11 +1656,27 @@ for (mfname, elty) in ((:zhemm_,:ComplexF64),
             require_one_based_indexing(A, B, C)
             m, n = size(C)
             j = checksquare(A)
-            if j != (side == 'L' ? m : n)
-                throw(DimensionMismatch(lazy"A has size $(size(A)), C has size ($m,$n)"))
-            end
-            if size(B,2) != n
-                throw(DimensionMismatch(lazy"B has second dimension $(size(B,2)) but needs to match second dimension of C, $n"))
+            M, N = size(B)
+            if side == 'L'
+                if j != m
+                    throw(DimensionMismatch(lazy"A has first dimension $j but needs to match first dimension of C, $m"))
+                end
+                if N != n
+                    throw(DimensionMismatch(lazy"B has second dimension $N but needs to match second dimension of C, $n"))
+                end
+                if j != M
+                    throw(DimensionMismatch(lazy"A has second dimension $j but needs to match first dimension of B, $M"))
+                end
+            else
+                if j != n
+                    throw(DimensionMismatch(lazy"B has second dimension $j but needs to match second dimension of C, $n"))
+                end
+                if N != j
+                    throw(DimensionMismatch(lazy"A has second dimension $N but needs to match first dimension of B, $j"))
+                end
+                if M != m
+                    throw(DimensionMismatch(lazy"A has first dimension $M but needs to match first dimension of C, $m"))
+                end
             end
             chkstride1(A)
             chkstride1(B)

stdlib/LinearAlgebra/src/bunchkaufman.jl

@@ -81,7 +81,7 @@ BunchKaufman(A::AbstractMatrix{T}, ipiv::AbstractVector{<:Integer}, uplo::Abstra
         BunchKaufman{T,typeof(A),typeof(ipiv)}(A, ipiv, uplo, symmetric, rook, info)
 # backwards-compatible constructors (remove with Julia 2.0)
 @deprecate(BunchKaufman(LD, ipiv, uplo, symmetric, rook, info) where {T,S},
-           BunchKaufman{T,S,typeof(ipiv)}(LD, ipiv, uplo, symmetric, rook, info))
+           BunchKaufman{T,S,typeof(ipiv)}(LD, ipiv, uplo, symmetric, rook, info), false)
 
 # iteration for destructuring into components
 Base.iterate(S::BunchKaufman) = (S.D, Val(:UL))