change compiler to be stackless (#55575)

This change ensures the compiler uses very little stack, making it
compatible with running on any arbitrary system stack size and depths
much more reliably. It also could be further modified now to easily add
various forms of pause-able/resumable inference, since there is no
implicit state on the stack--everything is local and explicit now.

Whereas before, less than 900 frames would crash in less than a second:
```
$ time ./julia -e 'f(::Val{N}) where {N} = N <= 0 ? 0 : f(Val(N - 1)); f(Val(1000))'
Warning: detected a stack overflow; program state may be corrupted, so further execution might be unreliable.
Internal error: during type inference of
f(Base.Val{1000})
Encountered stack overflow.
This might be caused by recursion over very long tuples or argument lists.

[23763] signal 6: Abort trap: 6
in expression starting at none:1
__pthread_kill at /usr/lib/system/libsystem_kernel.dylib (unknown line)
Allocations: 1 (Pool: 1; Big: 0); GC: 0
Abort trap: 6

real	0m0.233s
user	0m0.165s
sys	0m0.049s
````

Now: it is effectively unlimited, as long as you are willing to wait for
it:
```
$ time ./julia -e 'f(::Val{N}) where {N} = N <= 0 ? 0 : f(Val(N - 1)); f(Val(50000))'
info: inference of f(Base.Val{50000}) from f(Base.Val{N}) where {N} exceeding 2500 frames (may be slow).
info: inference of f(Base.Val{50000}) from f(Base.Val{N}) where {N} exceeding 5000 frames (may be slow).
info: inference of f(Base.Val{50000}) from f(Base.Val{N}) where {N} exceeding 10000 frames (may be slow).
info: inference of f(Base.Val{50000}) from f(Base.Val{N}) where {N} exceeding 20000 frames (may be slow).
info: inference of f(Base.Val{50000}) from f(Base.Val{N}) where {N} exceeding 40000 frames (may be slow).
real	7m4.988s

$ time ./julia -e 'f(::Val{N}) where {N} = N <= 0 ? 0 : f(Val(N - 1)); f(Val(1000))'
real	0m0.214s
user	0m0.164s
sys	0m0.044s

$ time ./julia -e '@noinline f(::Val{N}) where {N} = N <= 0 ? GC.safepoint() : f(Val(N - 1)); f(Val(5000))'
info: inference of f(Base.Val{5000}) from f(Base.Val{N}) where {N} exceeding 2500 frames (may be slow).
info: inference of f(Base.Val{5000}) from f(Base.Val{N}) where {N} exceeding 5000 frames (may be slow).
real	0m8.609s
user	0m8.358s
sys	0m0.240s
```

Author: vtjnash

base/compiler/abstractinterpretation.jl

@@ -251,6 +251,7 @@ mutable struct InferenceState
     stmt_info::Vector{CallInfo}
 
     #= intermediate states for interprocedural abstract interpretation =#
+    tasks::Vector{WorkThunk}
     pclimitations::IdSet{InferenceState} # causes of precision restrictions (LimitedAccuracy) on currpc ssavalue
     limitations::IdSet{InferenceState} # causes of precision restrictions (LimitedAccuracy) on return
     cycle_backedges::Vector{Tuple{InferenceState, Int}} # call-graph backedges connecting from callee to caller
@@ -328,6 +329,7 @@ mutable struct InferenceState
         limitations = IdSet{InferenceState}()
         cycle_backedges = Vector{Tuple{InferenceState,Int}}()
         callstack = AbsIntState[]
+        tasks = WorkThunk[]
 
         valid_worlds = WorldRange(1, get_world_counter())
         bestguess = Bottom
@@ -351,7 +353,7 @@ mutable struct InferenceState
         this = new(
             mi, world, mod, sptypes, slottypes, src, cfg, method_info,
             currbb, currpc, ip, handler_info, ssavalue_uses, bb_vartables, ssavaluetypes, stmt_edges, stmt_info,
-            pclimitations, limitations, cycle_backedges, callstack, 0, 0, 0,
+            tasks, pclimitations, limitations, cycle_backedges, callstack, 0, 0, 0,
             result, unreachable, valid_worlds, bestguess, exc_bestguess, ipo_effects,
             restrict_abstract_call_sites, cache_mode, insert_coverage,
             interp)
@@ -800,6 +802,7 @@ mutable struct IRInterpretationState
     const ssa_refined::BitSet
     const lazyreachability::LazyCFGReachability
     valid_worlds::WorldRange
+    const tasks::Vector{WorkThunk}
     const edges::Vector{Any}
     callstack #::Vector{AbsIntState}
     frameid::Int
@@ -825,10 +828,11 @@ mutable struct IRInterpretationState
         ssa_refined = BitSet()
         lazyreachability = LazyCFGReachability(ir)
         valid_worlds = WorldRange(min_world, max_world == typemax(UInt) ? get_world_counter() : max_world)
+        tasks = WorkThunk[]
         edges = Any[]
         callstack = AbsIntState[]
         return new(method_info, ir, mi, world, curridx, argtypes_refined, ir.sptypes, tpdum,
-                ssa_refined, lazyreachability, valid_worlds, edges, callstack, 0, 0)
+                ssa_refined, lazyreachability, valid_worlds, tasks, edges, callstack, 0, 0)
     end
 end
 
@@ -870,6 +874,7 @@ function print_callstack(frame::AbsIntState)
         print(frame_instance(sv))
         is_cached(sv) || print("  [uncached]")
         sv.parentid == idx - 1 || print(" [parent=", sv.parentid, "]")
+        isempty(callers_in_cycle(sv)) || print(" [cycle=", sv.cycleid, "]")
         println()
         @assert sv.frameid == idx
     end
@@ -994,7 +999,10 @@ of the same cycle, only if it is part of a cycle with multiple frames.
 function callers_in_cycle(sv::InferenceState)
     callstack = sv.callstack::Vector{AbsIntState}
     cycletop = cycleid = sv.cycleid
-    while cycletop < length(callstack) && (callstack[cycletop + 1]::InferenceState).cycleid == cycleid
+    while cycletop < length(callstack)
+        frame = callstack[cycletop + 1]
+        frame isa InferenceState || break
+        frame.cycleid == cycleid || break
         cycletop += 1
     end
     return AbsIntCycle(callstack, cycletop == cycleid ? 0 : cycleid, cycletop)
@@ -1054,6 +1062,7 @@ function merge_effects!(::AbstractInterpreter, caller::InferenceState, effects::
         effects = Effects(effects; effect_free=ALWAYS_TRUE)
     end
     caller.ipo_effects = merge_effects(caller.ipo_effects, effects)
+    nothing
 end
 merge_effects!(::AbstractInterpreter, ::IRInterpretationState, ::Effects) = return
 
@@ -1116,3 +1125,90 @@ function get_max_methods_for_module(mod::Module)
     max_methods < 0 && return nothing
     return max_methods
 end
+
+"""
+    Future{T}
+
+Delayed return value for a value of type `T`, similar to RefValue{T}, but
+explicitly represents completed as a `Bool` rather than as `isdefined`.
+Set once with `f[] = v` and accessed with `f[]` afterwards.
+
+Can also be constructed with the `completed` flag value and a closure to
+produce `x`, as well as the additional arguments to avoid always capturing the
+same couple of values.
+"""
+struct Future{T}
+    later::Union{Nothing,RefValue{T}}
+    now::Union{Nothing,T}
+    Future{T}() where {T} = new{T}(RefValue{T}(), nothing)
+    Future{T}(x) where {T} = new{T}(nothing, x)
+    Future(x::T) where {T} = new{T}(nothing, x)
+end
+isready(f::Future) = f.later === nothing
+getindex(f::Future{T}) where {T} = (later = f.later; later === nothing ? f.now::T : later[])
+setindex!(f::Future, v) = something(f.later)[] = v
+convert(::Type{Future{T}}, x) where {T} = Future{T}(x) # support return type conversion
+convert(::Type{Future{T}}, x::Future) where {T} = x::Future{T}
+function Future{T}(f, immediate::Bool, interp::AbstractInterpreter, sv::AbsIntState) where {T}
+    if immediate
+        return Future{T}(f(interp, sv))
+    else
+        @assert applicable(f, interp, sv)
+        result = Future{T}()
+        push!(sv.tasks, function (interp, sv)
+            result[] = f(interp, sv)
+            return true
+        end)
+        return result
+    end
+end
+function Future{T}(f, prev::Future{S}, interp::AbstractInterpreter, sv::AbsIntState) where {T, S}
+    later = prev.later
+    if later === nothing
+        return Future{T}(f(prev[], interp, sv))
+    else
+        @assert Core._hasmethod(Tuple{Core.Typeof(f), S, typeof(interp), typeof(sv)})
+        result = Future{T}()
+        push!(sv.tasks, function (interp, sv)
+            result[] = f(later[], interp, sv) # capture just later, instead of all of prev
+            return true
+        end)
+        return result
+    end
+end
+
+
+"""
+    doworkloop(args...)
+
+Run a tasks inside the abstract interpreter, returning false if there are none.
+Tasks will be run in DFS post-order tree order, such that all child tasks will
+be run in the order scheduled, prior to running any subsequent tasks. This
+allows tasks to generate more child tasks, which will be run before anything else.
+Each task will be run repeatedly when returning `false`, until it returns `true`.
+"""
+function doworkloop(interp::AbstractInterpreter, sv::AbsIntState)
+    tasks = sv.tasks
+    prev = length(tasks)
+    prev == 0 && return false
+    task = pop!(tasks)
+    completed = task(interp, sv)
+    tasks = sv.tasks # allow dropping gc root over the previous call
+    completed isa Bool || throw(TypeError(:return, "", Bool, task)) # print the task on failure as part of the error message, instead of just "@ workloop:line"
+    completed || push!(tasks, task)
+    # efficient post-order visitor: items pushed are executed in reverse post order such
+    # that later items are executed before earlier ones, but are fully executed
+    # (including any dependencies scheduled by them) before going on to the next item
+    reverse!(tasks, #=start=#prev)
+    return true
+end
+
+
+#macro workthunk(name::Symbol, body)
+#    name = esc(name)
+#    body = esc(body)
+#    return replace_linenums!(
+#        :(function $name($(esc(interp)), $(esc(sv)))
+#              $body
+#          end), __source__)
+#end

base/compiler/inferencestate.jl

@@ -1432,6 +1432,7 @@ function process_node!(compact::IncrementalCompact, result_idx::Int, inst::Instr
     elseif isa(stmt, OldSSAValue)
         ssa_rename[idx] = ssa_rename[stmt.id]
     elseif isa(stmt, GotoNode) && cfg_transforms_enabled
+        stmt.label < 0 && (println(stmt); println(compact))
         label = bb_rename_succ[stmt.label]
         @assert label > 0
         ssa_rename[idx] = SSAValue(result_idx)

base/compiler/ssair/ir.jl

@@ -51,8 +51,11 @@ end
 
 function abstract_call(interp::AbstractInterpreter, arginfo::ArgInfo, irsv::IRInterpretationState)
     si = StmtInfo(true) # TODO better job here?
-    call = abstract_call(interp, arginfo, si, irsv)
-    irsv.ir.stmts[irsv.curridx][:info] = call.info
+    call = abstract_call(interp, arginfo, si, irsv)::Future
+    Future{Nothing}(call, interp, irsv) do call, interp, irsv
+        irsv.ir.stmts[irsv.curridx][:info] = call.info
+        nothing
+    end
     return call
 end
 
@@ -143,7 +146,19 @@ function reprocess_instruction!(interp::AbstractInterpreter, inst::Instruction,
         head = stmt.head
         if (head === :call || head === :foreigncall || head === :new || head === :splatnew ||
             head === :static_parameter || head === :isdefined || head === :boundscheck)
-            (; rt, effects) = abstract_eval_statement_expr(interp, stmt, nothing, irsv)
+            @assert isempty(irsv.tasks) # TODO: this whole function needs to be converted to a stackless design to be a valid AbsIntState, but this should work here for now
+            result = abstract_eval_statement_expr(interp, stmt, nothing, irsv)
+            reverse!(irsv.tasks)
+            while true
+                if length(irsv.callstack) > irsv.frameid
+                    typeinf(interp, irsv.callstack[irsv.frameid + 1])
+                elseif !doworkloop(interp, irsv)
+                    break
+                end
+            end
+            @assert length(irsv.callstack) == irsv.frameid && isempty(irsv.tasks)
+            result isa Future && (result = result[])
+            (; rt, effects) = result
             add_flag!(inst, flags_for_effects(effects))
         elseif head === :invoke
             rt, (nothrow, noub) = abstract_eval_invoke_inst(interp, inst, irsv)
@@ -293,7 +308,7 @@ function is_all_const_call(@nospecialize(stmt), interp::AbstractInterpreter, irs
     return true
 end
 
-function _ir_abstract_constant_propagation(interp::AbstractInterpreter, irsv::IRInterpretationState;
+function ir_abstract_constant_propagation(interp::AbstractInterpreter, irsv::IRInterpretationState;
         externally_refined::Union{Nothing,BitSet} = nothing)
     (; ir, tpdum, ssa_refined) = irsv
 
@@ -449,18 +464,3 @@ function _ir_abstract_constant_propagation(interp::AbstractInterpreter, irsv::IR
 
     return Pair{Any,Tuple{Bool,Bool}}(maybe_singleton_const(ultimate_rt), (nothrow, noub))
 end
-
-function ir_abstract_constant_propagation(interp::NativeInterpreter, irsv::IRInterpretationState)
-    if __measure_typeinf__[]
-        inf_frame = Timings.InferenceFrameInfo(irsv.mi, irsv.world, VarState[], Any[], length(irsv.ir.argtypes))
-        Timings.enter_new_timer(inf_frame)
-        ret = _ir_abstract_constant_propagation(interp, irsv)
-        append!(inf_frame.slottypes, irsv.ir.argtypes)
-        Timings.exit_current_timer(inf_frame)
-        return ret
-    else
-        return _ir_abstract_constant_propagation(interp, irsv)
-    end
-end
-ir_abstract_constant_propagation(interp::AbstractInterpreter, irsv::IRInterpretationState) =
-    _ir_abstract_constant_propagation(interp, irsv)

base/compiler/ssair/irinterp.jl

@@ -1,9 +1,11 @@
 # This file is a part of Julia. License is MIT: https://julialang.org/license
 
 function maybe_show_ir(ir::IRCode)
-    if isdefined(Core, :Main)
+    if isdefined(Core, :Main) && isdefined(Core.Main, :Base)
         # ensure we use I/O that does not yield, as this gets called during compilation
         invokelatest(Core.Main.Base.show, Core.stdout, "text/plain", ir)
+    else
+        Core.show(ir)
     end
 end
 
@@ -25,6 +27,7 @@ is_toplevel_expr_head(head::Symbol) = head === :global || head === :method || he
 is_value_pos_expr_head(head::Symbol) = head === :static_parameter
 function check_op(ir::IRCode, domtree::DomTree, @nospecialize(op), use_bb::Int, use_idx::Int, printed_use_idx::Int, print::Bool, isforeigncall::Bool, arg_idx::Int, allow_frontend_forms::Bool)
     if isa(op, SSAValue)
+        op.id > 0 || @verify_error "Def ($(op.id)) is invalid in final IR"
         if op.id > length(ir.stmts)
             def_bb = block_for_inst(ir.cfg, ir.new_nodes.info[op.id - length(ir.stmts)].pos)
         else