Backports for 1.12.0-rc2 (#59110)

Author: KristofferC

Compiler/src/abstractinterpretation.jl

@@ -369,7 +369,7 @@ function find_union_split_method_matches(interp::AbstractInterpreter, argtypes::
         end
         valid_worlds = intersect(valid_worlds, thismatches.valid_worlds)
         thisfullmatch = any(match::MethodMatch->match.fully_covers, thismatches)
-        mt = Core.GlobalMethods
+        mt = Core.methodtable
         thisinfo = MethodMatchInfo(thismatches, mt, sig_n, thisfullmatch)
         push!(infos, thisinfo)
         for idx = 1:length(thismatches)
@@ -390,7 +390,7 @@ function find_simple_method_matches(interp::AbstractInterpreter, @nospecialize(a
         return FailedMethodMatch("Too many methods matched")
     end
     fullmatch = any(match::MethodMatch->match.fully_covers, matches)
-    mt = Core.GlobalMethods
+    mt = Core.methodtable
     info = MethodMatchInfo(matches, mt, atype, fullmatch)
     applicable = MethodMatchTarget[MethodMatchTarget(matches[idx], info.edges, idx) for idx = 1:length(matches)]
     return MethodMatches(applicable, info, matches.valid_worlds)
@@ -3214,7 +3214,7 @@ function abstract_eval_copyast(interp::AbstractInterpreter, e::Expr, sstate::Sta
     return RTEffects(rt, Any, effects)
 end
 
-function abstract_eval_isdefined_expr(interp::AbstractInterpreter, e::Expr, sstate::StatementState,
+function abstract_eval_isdefined_expr(::AbstractInterpreter, e::Expr, sstate::StatementState,
                                       sv::AbsIntState)
     sym = e.args[1]
     if isa(sym, SlotNumber) && sstate.vtypes !== nothing
@@ -3444,7 +3444,59 @@ function refine_partial_type(@nospecialize t)
     return t
 end
 
+abstract_eval_nonlinearized_foreigncall_name(interp::AbstractInterpreter, e, sstate::StatementState, sv::IRInterpretationState) = nothing
+
+function abstract_eval_nonlinearized_foreigncall_name(interp::AbstractInterpreter, e, sstate::StatementState, sv::AbsIntState)
+    if isexpr(e, :call)
+        n = length(e.args)
+        argtypes = Vector{Any}(undef, n)
+        callresult = Future{CallMeta}()
+        i::Int = 1
+        nextstate::UInt8 = 0x0
+        local ai, res
+        function evalargs(interp, sv)
+            if nextstate === 0x1
+                @goto state1
+            elseif nextstate === 0x2
+                @goto state2
+            end
+            while i <= n
+                ai = abstract_eval_nonlinearized_foreigncall_name(interp, e.args[i], sstate, sv)
+                if !isready(ai)
+                    nextstate = 0x1
+                    return false
+                    @label state1
+                end
+                argtypes[i] = ai[].rt
+                i += 1
+            end
+            res = abstract_call(interp, ArgInfo(e.args, argtypes), sstate, sv)
+            if !isready(res)
+                nextstate = 0x2
+                return false
+                @label state2
+            end
+            callresult[] = res[]
+            return true
+        end
+        evalargs(interp, sv) || push!(sv.tasks, evalargs)
+        return callresult
+    else
+        return Future(abstract_eval_basic_statement(interp, e, sstate, sv))
+    end
+end
+
 function abstract_eval_foreigncall(interp::AbstractInterpreter, e::Expr, sstate::StatementState, sv::AbsIntState)
+    callee = e.args[1]
+    if isexpr(callee, :call) && length(callee.args) > 1 && callee.args[1] == GlobalRef(Core, :tuple)
+        # NOTE these expressions are not properly linearized
+        abstract_eval_nonlinearized_foreigncall_name(interp, callee.args[2], sstate, sv)
+        if length(callee.args) > 2
+            abstract_eval_nonlinearized_foreigncall_name(interp, callee.args[3], sstate, sv)
+        end
+    else
+        abstract_eval_value(interp, callee, sstate, sv)
+    end
     mi = frame_instance(sv)
     t = sp_type_rewrap(e.args[2], mi, true)
     for i = 3:length(e.args)

Compiler/src/stmtinfo.jl

@@ -49,14 +49,14 @@ function _add_edges_impl(edges::Vector{Any}, info::MethodMatchInfo, mi_edge::Boo
     if !fully_covering(info)
         exists = false
         for i in 2:length(edges)
-            if edges[i] === Core.GlobalMethods && edges[i-1] == info.atype
+            if edges[i] === Core.methodtable && edges[i-1] == info.atype
                 exists = true
                 break
             end
         end
         if !exists
             push!(edges, info.atype)
-            push!(edges, Core.GlobalMethods)
+            push!(edges, Core.methodtable)
         end
     end
     nmatches = length(info.results)

Compiler/src/tfuncs.jl

@@ -3170,7 +3170,7 @@ function _hasmethod_tfunc(interp::AbstractInterpreter, argtypes::Vector{Any}, sv
     if match === nothing
         rt = Const(false)
         vresults = MethodLookupResult(Any[], valid_worlds, true)
-        mt = Core.GlobalMethods
+        mt = Core.methodtable
         vinfo = MethodMatchInfo(vresults, mt, types, false) # XXX: this should actually be an info with invoke-type edge
     else
         rt = Const(true)

Compiler/src/typelattice.jl

@@ -40,6 +40,9 @@ struct Conditional
                          isdefined::Bool=false)
         assert_nested_slotwrapper(thentype)
         assert_nested_slotwrapper(elsetype)
+        if thentype isa LimitedAccuracy || elsetype isa LimitedAccuracy
+            return Bool
+        end
         return new(slot, thentype, elsetype, isdefined)
     end
 end
@@ -83,6 +86,9 @@ struct MustAlias
         assert_nested_slotwrapper(fldtyp)
         # @assert !isalreadyconst(vartyp) "vartyp is already const"
         # @assert !isalreadyconst(fldtyp) "fldtyp is already const"
+        if vartyp isa LimitedAccuracy || fldtyp isa LimitedAccuracy
+            return fldtyp
+        end
         return new(slot, vartyp, fldidx, fldtyp)
     end
 end
@@ -104,6 +110,9 @@ struct InterMustAlias
         assert_nested_slotwrapper(fldtyp)
         # @assert !isalreadyconst(vartyp) "vartyp is already const"
         # @assert !isalreadyconst(fldtyp) "fldtyp is already const"
+        if vartyp isa LimitedAccuracy || fldtyp isa LimitedAccuracy
+            return fldtyp
+        end
         return new(slot, vartyp, fldidx, fldtyp)
     end
 end

NEWS.md

@@ -33,13 +33,6 @@ New language features
 Language changes
 ----------------
 
-* Julia now defaults to 1 "interactive" thread, in addition to the 1 default "worker" thread. i.e. `-t1,1`.
-  This means in default configuration the main task and repl (when in interactive mode), which both run on
-  thread 1, now run within the `interactive` threadpool. The libuv IO loop also runs on thread 1,
-  helping efficient utilization of the worker threadpool used by `Threads.@spawn`. Asking for specifically 1 thread
-  (`-t1`/`JULIA_NUM_THREADS=1`) or passing `0` will disable the interactive thread i.e. `-t1,0` or `JULIA_NUM_THREADS=1,0`
-  , or `-tauto,0` etc. Asking for more than 1 thread will enable the interactive thread so
-  `-t2` will set the equivalent of `-t2,1` ([#57087]).
 * When a method is replaced with an exactly equivalent one, the old method is not deleted. Instead, the
   new method takes priority and becomes more specific than the old method. Thus if the new method is deleted
   later, the old method will resume operating. This can be useful in mocking frameworks (as in SparseArrays,
@@ -88,6 +81,14 @@ Command-line option changes
 Multi-threading changes
 -----------------------
 
+* Julia now defaults to 1 "interactive" thread, in addition to the 1 default "worker" thread. i.e. `-t1,1`.
+  This means in default configuration the main task and repl (when in interactive mode), which both run on
+  thread 1, now run within the `interactive` threadpool. The libuv IO loop also runs on thread 1,
+  helping efficient utilization of the worker threadpool used by `Threads.@spawn`. Asking for specifically 1 thread
+  (`-t1`/`JULIA_NUM_THREADS=1`) or `0` interactive threads will disable the interactive thread i.e. `-t1,0` or `JULIA_NUM_THREADS=1,0`
+  , or `-tauto,0` etc. Asking for more than 1 thread will enable the interactive thread so
+  `-t2` will set the equivalent of `-t2,1`. As a reminder, buffers
+  [should not be managed based on `threadid()`](https://docs.julialang.org/en/v1/manual/multi-threading/#Using-@threads-without-data-races) ([#57087]).
 * New types are defined to handle the pattern of code that must run once per process, called
   a `OncePerProcess{T}` type, which allows defining a function that should be run exactly once
   the first time it is called, and then always return the same result value of type `T`

base/error.jl

@@ -233,14 +233,12 @@ macro assert(ex, msgs...)
         msg = msg # pass-through
     elseif !isempty(msgs) && (isa(msg, Expr) || isa(msg, Symbol))
         # message is an expression needing evaluating
-        # N.B. To reduce the risk of invalidation caused by the complex callstack involved
-        # with `string`, use `inferencebarrier` here to hide this `string` from the compiler.
-        msg = :(Main.Base.inferencebarrier(Main.Base.string)($(esc(msg))))
+        msg = :($_assert_tostring($(esc(msg))))
     elseif isdefined(Main, :Base) && isdefined(Main.Base, :string) && applicable(Main.Base.string, msg)
         msg = Main.Base.string(msg)
     else
         # string() might not be defined during bootstrap
-        msg = :(_assert_tostring($(Expr(:quote,msg))))
+        msg = :($_assert_tostring($(Expr(:quote,msg))))
     end
     return :($(esc(ex)) ? $(nothing) : throw(AssertionError($msg)))
 end

base/initdefs.jl

@@ -494,10 +494,13 @@ end
 ## hook for disabling threaded libraries ##
 
 library_threading_enabled::Bool = true
-const disable_library_threading_hooks = []
+
+# Base.OncePerProcess ensures that any registered hooks do not outlive the session.
+# (even if they are registered during the sysimage build process by top-level code)
+const disable_library_threading_hooks = Base.OncePerProcess(Vector{Any})
 
 function at_disable_library_threading(f)
-    push!(disable_library_threading_hooks, f)
+    push!(disable_library_threading_hooks(), f)
     if !library_threading_enabled
         disable_library_threading()
     end
@@ -506,8 +509,8 @@ end
 
 function disable_library_threading()
     global library_threading_enabled = false
-    while !isempty(disable_library_threading_hooks)
-        f = pop!(disable_library_threading_hooks)
+    while !isempty(disable_library_threading_hooks())
+        f = pop!(disable_library_threading_hooks())
         try
             f()
         catch err

base/int.jl

@@ -845,14 +845,166 @@ widemul(x::Bool,y::Number) = x * y
 widemul(x::Number,y::Bool) = x * y
 
 
-# Int128 multiply and divide
-*(x::T, y::T) where {T<:Union{Int128,UInt128}}  = mul_int(x, y)
+## wide multiplication, Int128 multiply and divide ##
+
+if Core.sizeof(Int) == 4
+    function widemul(u::Int64, v::Int64)
+        local u0::UInt64, v0::UInt64, w0::UInt64
+        local u1::Int64, v1::Int64, w1::UInt64, w2::Int64, t::UInt64
+
+        u0 = u & 0xffffffff; u1 = u >> 32
+        v0 = v & 0xffffffff; v1 = v >> 32
+        w0 = u0 * v0
+        t = reinterpret(UInt64, u1) * v0 + (w0 >>> 32)
+        w2 = reinterpret(Int64, t) >> 32
+        w1 = u0 * reinterpret(UInt64, v1) + (t & 0xffffffff)
+        hi = u1 * v1 + w2 + (reinterpret(Int64, w1) >> 32)
+        lo = w0 & 0xffffffff + (w1 << 32)
+        return Int128(hi) << 64 + Int128(lo)
+    end
+
+    function widemul(u::UInt64, v::UInt64)
+        local u0::UInt64, v0::UInt64, w0::UInt64
+        local u1::UInt64, v1::UInt64, w1::UInt64, w2::UInt64, t::UInt64
+
+        u0 = u & 0xffffffff; u1 = u >>> 32
+        v0 = v & 0xffffffff; v1 = v >>> 32
+        w0 = u0 * v0
+        t = u1 * v0 + (w0 >>> 32)
+        w2 = t >>> 32
+        w1 = u0 * v1 + (t & 0xffffffff)
+        hi = u1 * v1 + w2 + (w1 >>> 32)
+        lo = w0 & 0xffffffff + (w1 << 32)
+        return UInt128(hi) << 64 + UInt128(lo)
+    end
+
+    function *(u::Int128, v::Int128)
+        u0 = u % UInt64; u1 = Int64(u >> 64)
+        v0 = v % UInt64; v1 = Int64(v >> 64)
+        lolo = widemul(u0, v0)
+        lohi = widemul(reinterpret(Int64, u0), v1)
+        hilo = widemul(u1, reinterpret(Int64, v0))
+        t = reinterpret(UInt128, hilo) + (lolo >>> 64)
+        w1 = reinterpret(UInt128, lohi) + (t & 0xffffffffffffffff)
+        return Int128(lolo & 0xffffffffffffffff) + reinterpret(Int128, w1) << 64
+    end
+
+    function *(u::UInt128, v::UInt128)
+        u0 = u % UInt64; u1 = UInt64(u>>>64)
+        v0 = v % UInt64; v1 = UInt64(v>>>64)
+        lolo = widemul(u0, v0)
+        lohi = widemul(u0, v1)
+        hilo = widemul(u1, v0)
+        t = hilo + (lolo >>> 64)
+        w1 = lohi + (t & 0xffffffffffffffff)
+        return (lolo & 0xffffffffffffffff) + UInt128(w1) << 64
+    end
+
+    function _setbit(x::UInt128, i)
+        # faster version of `return x | (UInt128(1) << i)`
+        j = i >> 5
+        y = UInt128(one(UInt32) << (i & 0x1f))
+        if j == 0
+            return x | y
+        elseif j == 1
+            return x | (y << 32)
+        elseif j == 2
+            return x | (y << 64)
+        elseif j == 3
+            return x | (y << 96)
+        end
+        return x
+    end
 
-div(x::Int128,  y::Int128)  = checked_sdiv_int(x, y)
-div(x::UInt128, y::UInt128) = checked_udiv_int(x, y)
+    function divrem(x::UInt128, y::UInt128)
+        iszero(y) && throw(DivideError())
+        if (x >> 64) % UInt64 == 0
+            if (y >> 64) % UInt64 == 0
+                # fast path: upper 64 bits are zero, so we can fallback to UInt64 division
+                q64, x64 = divrem(x % UInt64, y % UInt64)
+                return UInt128(q64), UInt128(x64)
+            else
+                # this implies y>x, so
+                return zero(UInt128), x
+            end
+        end
+        n = leading_zeros(y) - leading_zeros(x)
+        q = zero(UInt128)
+        ys = y << n
+        while n >= 0
+            # ys == y * 2^n
+            if ys <= x
+                x -= ys
+                q = _setbit(q, n)
+                if (x >> 64) % UInt64 == 0
+                    # exit early, similar to above fast path
+                    if (y >> 64) % UInt64 == 0
+                        q64, x64 = divrem(x % UInt64, y % UInt64)
+                        q |= q64
+                        x = UInt128(x64)
+                    end
+                    return q, x
+                end
+            end
+            ys >>>= 1
+            n -= 1
+        end
+        return q, x
+    end
 
-rem(x::Int128,  y::Int128)  = checked_srem_int(x, y)
-rem(x::UInt128, y::UInt128) = checked_urem_int(x, y)
+    function div(x::Int128, y::Int128)
+        (x == typemin(Int128)) & (y == -1) && throw(DivideError())
+        return Int128(div(BigInt(x), BigInt(y)))::Int128
+    end
+    div(x::UInt128, y::UInt128) = divrem(x, y)[1]
+
+    function rem(x::Int128, y::Int128)
+        return Int128(rem(BigInt(x), BigInt(y)))::Int128
+    end
+
+    function rem(x::UInt128, y::UInt128)
+        iszero(y) && throw(DivideError())
+        if (x >> 64) % UInt64 == 0
+            if (y >> 64) % UInt64 == 0
+                # fast path: upper 64 bits are zero, so we can fallback to UInt64 division
+                return UInt128(rem(x % UInt64, y % UInt64))
+            else
+                # this implies y>x, so
+                return x
+            end
+        end
+        n = leading_zeros(y) - leading_zeros(x)
+        ys = y << n
+        while n >= 0
+            # ys == y * 2^n
+            if ys <= x
+                x -= ys
+                if (x >> 64) % UInt64 == 0
+                    # exit early, similar to above fast path
+                    if (y >> 64) % UInt64 == 0
+                        x = UInt128(rem(x % UInt64, y % UInt64))
+                    end
+                    return x
+                end
+            end
+            ys >>>= 1
+            n -= 1
+        end
+        return x
+    end
+
+    function mod(x::Int128, y::Int128)
+        return Int128(mod(BigInt(x), BigInt(y)))::Int128
+    end
+else
+    *(x::T, y::T) where {T<:Union{Int128,UInt128}}  = mul_int(x, y)
+
+    div(x::Int128,  y::Int128)  = checked_sdiv_int(x, y)
+    div(x::UInt128, y::UInt128) = checked_udiv_int(x, y)
+
+    rem(x::Int128,  y::Int128)  = checked_srem_int(x, y)
+    rem(x::UInt128, y::UInt128) = checked_urem_int(x, y)
+end
 
 # issue #15489: since integer ops are unchecked, they shouldn't check promotion
 for op in (:+, :-, :*, :&, :|, :xor)