optimizer: supports callsite annotations of inlining, fixes #18773

Enable `@inline`/`@noinline` annotations on function callsites.
From #40754.

Now `@inline` and `@noinline` can be applied to a code block and then
the compiler will try to (not) inline calls within the block:
```julia
@inline f(...) # The compiler will try to inline `f`

@inline f(...) + g(...) # The compiler will try to inline `f`, `g` and `+`

@inline f(args...) = ... # Of course annotations on a definition is still allowed
```

Here are couple of notes on how those callsite annotations will work:
- callsite annotation always has the precedence over the annotation
  applied to the definition of the called function, whichever we use
  `@inline`/`@noinline`:
  ```julia
  @inline function explicit_inline(args...)
      # body
  end

  let
      @noinline explicit_inline(args...) # this call will not be inlined
  end
  ```
- when callsite annotations are nested, the innermost annotations has
  the precedence
  ```julia
  @noinline let a0, b0 = ...
      a = @inline f(a0)  # the compiler will try to inline this call
      b = notinlined(b0) # the compiler will NOT try to inline this call
      return a, b
  end
  ```
They're both tested and included in documentations.

Author: aviatesk

base/compiler/abstractinterpretation.jl

@@ -590,7 +590,7 @@ function maybe_get_const_prop_profitable(interp::AbstractInterpreter, result::Me
         return nothing
     end
     mi = mi::MethodInstance
-    if !force && !const_prop_methodinstance_heuristic(interp, method, mi)
+    if !force && !const_prop_methodinstance_heuristic(interp, match, mi)
         add_remark!(interp, sv, "[constprop] Disabled by method instance heuristic")
         return nothing
     end
@@ -692,7 +692,8 @@ end
 # This is a heuristic to avoid trying to const prop through complicated functions
 # where we would spend a lot of time, but are probably unlikely to get an improved
 # result anyway.
-function const_prop_methodinstance_heuristic(interp::AbstractInterpreter, method::Method, mi::MethodInstance)
+function const_prop_methodinstance_heuristic(interp::AbstractInterpreter, match::MethodMatch, mi::MethodInstance)
+    method = match.method
     if method.is_for_opaque_closure
         # Not inlining an opaque closure can be very expensive, so be generous
         # with the const-prop-ability. It is quite possible that we can't infer
@@ -710,7 +711,8 @@ function const_prop_methodinstance_heuristic(interp::AbstractInterpreter, method
     if isdefined(code, :inferred) && !cache_inlineable
         cache_inf = code.inferred
         if !(cache_inf === nothing)
-            cache_inlineable = inlining_policy(interp)(cache_inf) !== nothing
+            # TODO maybe we want to respect callsite `@inline`/`@noinline` annotations here ?
+            cache_inlineable = inlining_policy(interp)(cache_inf, nothing, match) !== nothing
         end
     end
     if !cache_inlineable
@@ -1889,7 +1891,8 @@ function typeinf_local(interp::AbstractInterpreter, frame::InferenceState)
                     if isa(fname, SlotNumber)
                         changes = StateUpdate(fname, VarState(Any, false), changes, false)
                     end
-                elseif hd === :inbounds || hd === :meta || hd === :loopinfo || hd === :code_coverage_effect
+                elseif hd === :code_coverage_effect ||
+                       (hd !== :boundscheck && hd !== nothing && is_meta_expr_head(hd)) # :boundscheck can be narrowed to Bool
                     # these do not generate code
                 else
                     t = abstract_eval_statement(interp, stmt, changes, frame)

base/compiler/optimize.jl

@@ -28,14 +28,20 @@ struct InliningState{S <: Union{EdgeTracker, Nothing}, T, P}
     policy::P
 end
 
-function default_inlining_policy(@nospecialize(src))
+function default_inlining_policy(@nospecialize(src), stmt_flag::Union{Nothing,UInt8}, match::Union{MethodMatch,InferenceResult})
     if isa(src, CodeInfo) || isa(src, Vector{UInt8})
         src_inferred = ccall(:jl_ir_flag_inferred, Bool, (Any,), src)
-        src_inlineable = ccall(:jl_ir_flag_inlineable, Bool, (Any,), src)
+        src_inlineable = is_stmt_inline(stmt_flag) || ccall(:jl_ir_flag_inlineable, Bool, (Any,), src)
         return src_inferred && src_inlineable ? src : nothing
-    end
-    if isa(src, OptimizationState) && isdefined(src, :ir)
-        return src.src.inlineable ? src.ir : nothing
+    elseif isa(src, OptimizationState) && isdefined(src, :ir)
+        return (is_stmt_inline(stmt_flag) || src.src.inlineable) ? src.ir : nothing
+    elseif src === nothing && is_stmt_inline(stmt_flag) && isa(match, MethodMatch)
+        # when the source isn't available at this moment, try to re-infer and inline it
+        # HACK in order to avoid cycles here, we disable inlining and makes sure the following inference never comes here
+        # TODO sort out `AbstractInterpreter` interface to handle this well, and also inference should try to keep the source if the statement will be inlined
+        interp = NativeInterpreter(; opt_params = OptimizationParams(; inlining = false))
+        src, rt = typeinf_code(interp, match.method, match.spec_types, match.sparams, true)
+        return src
     end
     return nothing
 end
@@ -129,6 +135,10 @@ const SLOT_USEDUNDEF    = 32 # slot has uses that might raise UndefVarError
 # This statement was marked as @inbounds by the user. If replaced by inlining,
 # any contained boundschecks may be removed
 const IR_FLAG_INBOUNDS       = 0x01
+# This statement was marked as @inline by the user
+const IR_FLAG_INLINE         = 0x01 << 1
+# This statement was marked as @noinline by the user
+const IR_FLAG_NOINLINE       = 0x01 << 2
 # This statement may be removed if its result is unused. In particular it must
 # thus be both pure and effect free.
 const IR_FLAG_EFFECT_FREE    = 0x01 << 4
@@ -174,6 +184,11 @@ function isinlineable(m::Method, me::OptimizationState, params::OptimizationPara
     return inlineable
 end
 
+is_stmt_inline(stmt_flag::UInt8)   = stmt_flag & IR_FLAG_INLINE != 0
+is_stmt_inline(::Nothing)          = false
+is_stmt_noinline(stmt_flag::UInt8) = stmt_flag & IR_FLAG_NOINLINE != 0
+is_stmt_noinline(::Nothing)        = false # not used for now
+
 # These affect control flow within the function (so may not be removed
 # if there is no usage within the function), but don't affect the purity
 # of the function as a whole.
@@ -360,6 +375,7 @@ function convert_to_ircode(ci::CodeInfo, code::Vector{Any}, coverage::Bool, sv::
     renumber_ir_elements!(code, changemap, labelmap)
 
     inbounds_depth = 0 # Number of stacked inbounds
+    inline_flags = BitVector()
     meta = Any[]
     flags = fill(0x00, length(code))
     for i = 1:length(code)
@@ -374,16 +390,38 @@ function convert_to_ircode(ci::CodeInfo, code::Vector{Any}, coverage::Bool, sv::
                 inbounds_depth -= 1
             end
             stmt = nothing
+        elseif isexpr(stmt, :inline)
+            if stmt.args[1]::Bool
+                push!(inline_flags, true)
+            else
+                pop!(inline_flags)
+            end
+            stmt = nothing
+        elseif isexpr(stmt, :noinline)
+            if stmt.args[1]::Bool
+                push!(inline_flags, false)
+            else
+                pop!(inline_flags)
+            end
+            stmt = nothing
         else
             stmt = normalize(stmt, meta)
         end
         code[i] = stmt
-        if !(stmt === nothing)
+        if stmt !== nothing
             if inbounds_depth > 0
                 flags[i] |= IR_FLAG_INBOUNDS
             end
+            if !isempty(inline_flags)
+                if last(inline_flags)
+                    flags[i] |= IR_FLAG_INLINE
+                else
+                    flags[i] |= IR_FLAG_NOINLINE
+                end
+            end
         end
     end
+    @assert isempty(inline_flags) "malformed meta flags"
     strip_trailing_junk!(ci, code, stmtinfo, flags)
     cfg = compute_basic_blocks(code)
     types = Any[]

base/compiler/ssair/inlining.jl

@@ -604,6 +604,7 @@ function rewrite_apply_exprargs!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::
         argexprs::Vector{Any}, atypes::Vector{Any}, arginfos::Vector{Any},
         arg_start::Int, istate::InliningState)
 
+    flag = ir.stmts[idx][:flag]
     new_argexprs = Any[argexprs[arg_start]]
     new_atypes = Any[atypes[arg_start]]
     # loop over original arguments and flatten any known iterators
@@ -659,8 +660,9 @@ function rewrite_apply_exprargs!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::
                 info = call.info
                 handled = false
                 if isa(info, ConstCallInfo)
-                    if maybe_handle_const_call!(ir, state1.id, new_stmt, info, new_sig,
-                        call.rt, istate, false, todo)
+                    if !is_stmt_noinline(flag) && maybe_handle_const_call!(
+                        ir, state1.id, new_stmt, info, new_sig,call.rt, istate, flag, false, todo)
+
                         handled = true
                     else
                         info = info.call
@@ -671,7 +673,7 @@ function rewrite_apply_exprargs!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::
                         MethodMatchInfo[info] : info.matches
                     # See if we can inline this call to `iterate`
                     analyze_single_call!(ir, todo, state1.id, new_stmt,
-                        new_sig, call.rt, info, istate)
+                        new_sig, call.rt, info, istate, flag)
                 end
                 if i != length(thisarginfo.each)
                     valT = getfield_tfunc(call.rt, Const(1))
@@ -719,16 +721,16 @@ function compileable_specialization(et::Union{EdgeTracker, Nothing}, (; linfo)::
     return mi
 end
 
-function resolve_todo(todo::InliningTodo, state::InliningState)
-    spec = todo.spec::DelayedInliningSpec
+function resolve_todo(todo::InliningTodo, state::InliningState, flag::UInt8)
+    (; match) = todo.spec::DelayedInliningSpec
 
     #XXX: update_valid_age!(min_valid[1], max_valid[1], sv)
     isconst, src = false, nothing
-    if isa(spec.match, InferenceResult)
-        let inferred_src = spec.match.src
+    if isa(match, InferenceResult)
+        let inferred_src = match.src
             if isa(inferred_src, Const)
                 if !is_inlineable_constant(inferred_src.val)
-                    return compileable_specialization(state.et, spec.match)
+                    return compileable_specialization(state.et, match)
                 end
                 isconst, src = true, quoted(inferred_src.val)
             else
@@ -756,12 +758,10 @@ function resolve_todo(todo::InliningTodo, state::InliningState)
         return ConstantCase(src)
     end
 
-    if src !== nothing
-        src = state.policy(src)
-    end
+    src = state.policy(src, flag, match)
 
     if src === nothing
-        return compileable_specialization(et, spec.match)
+        return compileable_specialization(et, match)
     end
 
     if isa(src, IRCode)
@@ -772,9 +772,9 @@ function resolve_todo(todo::InliningTodo, state::InliningState)
     return InliningTodo(todo.mi, src)
 end
 
-function resolve_todo(todo::UnionSplit, state::InliningState)
+function resolve_todo(todo::UnionSplit, state::InliningState, flag::UInt8)
     UnionSplit(todo.fully_covered, todo.atype,
-        Pair{Any,Any}[sig=>resolve_todo(item, state) for (sig, item) in todo.cases])
+        Pair{Any,Any}[sig=>resolve_todo(item, state, flag) for (sig, item) in todo.cases])
 end
 
 function validate_sparams(sparams::SimpleVector)
@@ -785,7 +785,7 @@ function validate_sparams(sparams::SimpleVector)
 end
 
 function analyze_method!(match::MethodMatch, atypes::Vector{Any},
-                         state::InliningState, @nospecialize(stmttyp))
+                         state::InliningState, @nospecialize(stmttyp), flag::UInt8)
     method = match.method
     methsig = method.sig
 
@@ -805,7 +805,7 @@ function analyze_method!(match::MethodMatch, atypes::Vector{Any},
 
     et = state.et
 
-    if !state.params.inlining
+    if !state.params.inlining || is_stmt_noinline(flag)
         return compileable_specialization(et, match)
     end
 
@@ -819,7 +819,7 @@ function analyze_method!(match::MethodMatch, atypes::Vector{Any},
     # If we don't have caches here, delay resolving this MethodInstance
     # until the batch inlining step (or an external post-processing pass)
     state.mi_cache === nothing && return todo
-    return resolve_todo(todo, state)
+    return resolve_todo(todo, state, flag)
 end
 
 function InliningTodo(mi::MethodInstance, ir::IRCode)
@@ -1044,7 +1044,7 @@ is_builtin(s::Signature) =
     s.ft ⊑ Builtin
 
 function inline_invoke!(ir::IRCode, idx::Int, sig::Signature, (; match, result)::InvokeCallInfo,
-        state::InliningState, todo::Vector{Pair{Int, Any}})
+        state::InliningState, todo::Vector{Pair{Int, Any}}, flag::UInt8)
     stmt = ir.stmts[idx][:inst]
     calltype = ir.stmts[idx][:type]
 
@@ -1058,17 +1058,17 @@ function inline_invoke!(ir::IRCode, idx::Int, sig::Signature, (; match, result):
     atypes = atypes[4:end]
     pushfirst!(atypes, atype0)
 
-    if isa(result, InferenceResult)
+    if isa(result, InferenceResult) && !is_stmt_noinline(flag)
         (; mi) = item = InliningTodo(result, atypes, calltype)
         validate_sparams(mi.sparam_vals) || return nothing
         if argtypes_to_type(atypes) <: mi.def.sig
-            state.mi_cache !== nothing && (item = resolve_todo(item, state))
+            state.mi_cache !== nothing && (item = resolve_todo(item, state, flag))
             handle_single_case!(ir, stmt, idx, item, true, todo)
             return nothing
         end
     end
 
-    result = analyze_method!(match, atypes, state, calltype)
+    result = analyze_method!(match, atypes, state, calltype, flag)
     handle_single_case!(ir, stmt, idx, result, true, todo)
     return nothing
 end
@@ -1163,7 +1163,7 @@ end
 
 function analyze_single_call!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::Int, @nospecialize(stmt),
         sig::Signature, @nospecialize(calltype), infos::Vector{MethodMatchInfo},
-        state::InliningState)
+        state::InliningState, flag::UInt8)
     cases = Pair{Any, Any}[]
     signature_union = Union{}
     only_method = nothing  # keep track of whether there is one matching method
@@ -1197,7 +1197,7 @@ function analyze_single_call!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::Int
                 fully_covered = false
                 continue
             end
-            case = analyze_method!(match, sig.atypes, state, calltype)
+            case = analyze_method!(match, sig.atypes, state, calltype, flag)
             if case === nothing
                 fully_covered = false
                 continue
@@ -1224,7 +1224,7 @@ function analyze_single_call!(ir::IRCode, todo::Vector{Pair{Int, Any}}, idx::Int
             match = meth[1]
         end
         fully_covered = true
-        case = analyze_method!(match, sig.atypes, state, calltype)
+        case = analyze_method!(match, sig.atypes, state, calltype, flag)
         case === nothing && return
         push!(cases, Pair{Any,Any}(match.spec_types, case))
     end
@@ -1246,7 +1246,7 @@ end
 
 function maybe_handle_const_call!(ir::IRCode, idx::Int, stmt::Expr,
         info::ConstCallInfo, sig::Signature, @nospecialize(calltype),
-        state::InliningState,
+        state::InliningState, flag::UInt8,
         isinvoke::Bool, todo::Vector{Pair{Int, Any}})
     # when multiple matches are found, bail out and later inliner will union-split this signature
     # TODO effectively use multiple constant analysis results here
@@ -1258,7 +1258,7 @@ function maybe_handle_const_call!(ir::IRCode, idx::Int, stmt::Expr,
     validate_sparams(mi.sparam_vals) || return true
     mthd_sig = mi.def.sig
     mistypes = mi.specTypes
-    state.mi_cache !== nothing && (item = resolve_todo(item, state))
+    state.mi_cache !== nothing && (item = resolve_todo(item, state, flag))
     if sig.atype <: mthd_sig
         handle_single_case!(ir, stmt, idx, item, isinvoke, todo)
         return true
@@ -1296,6 +1296,8 @@ function assemble_inline_todo!(ir::IRCode, state::InliningState)
             info = info.info
         end
 
+        flag = ir.stmts[idx][:flag]
+
         # Inference determined this couldn't be analyzed. Don't question it.
         if info === false
             continue
@@ -1305,23 +1307,24 @@ function assemble_inline_todo!(ir::IRCode, state::InliningState)
         # it'll have performed a specialized analysis for just this case. Use its
         # result.
         if isa(info, ConstCallInfo)
-            if maybe_handle_const_call!(ir, idx, stmt, info, sig, calltype, state, sig.f === Core.invoke, todo)
+            if !is_stmt_noinline(flag) && maybe_handle_const_call!(
+                ir, idx, stmt, info, sig, calltype, state, flag, sig.f === Core.invoke, todo)
                 continue
             else
                 info = info.call
             end
         end
 
         if isa(info, OpaqueClosureCallInfo)
-            result = analyze_method!(info.match, sig.atypes, state, calltype)
+            result = analyze_method!(info.match, sig.atypes, state, calltype, flag)
             handle_single_case!(ir, stmt, idx, result, false, todo)
             continue
         end
 
         # Handle invoke
         if sig.f === Core.invoke
             if isa(info, InvokeCallInfo)
-                inline_invoke!(ir, idx, sig, info, state, todo)
+                inline_invoke!(ir, idx, sig, info, state, todo, flag)
             end
             continue
         end
@@ -1335,7 +1338,7 @@ function assemble_inline_todo!(ir::IRCode, state::InliningState)
             continue
         end
 
-        analyze_single_call!(ir, todo, idx, stmt, sig, calltype, infos, state)
+        analyze_single_call!(ir, todo, idx, stmt, sig, calltype, infos, state, flag)
     end
     todo
 end

base/compiler/typeinfer.jl

@@ -343,7 +343,7 @@ function maybe_compress_codeinfo(interp::AbstractInterpreter, linfo::MethodInsta
             nslots = length(ci.slotflags)
             resize!(ci.slottypes::Vector{Any}, nslots)
             resize!(ci.slotnames, nslots)
-            return ccall(:jl_compress_ir, Any, (Any, Any), def, ci)
+            return ccall(:jl_compress_ir, Vector{UInt8}, (Any, Any), def, ci)
         else
             return ci
         end

base/compiler/utilities.jl

@@ -59,7 +59,8 @@ end
 
 # Meta expression head, these generally can't be deleted even when they are
 # in a dead branch but can be ignored when analyzing uses/liveness.
-is_meta_expr_head(head::Symbol) = (head === :inbounds || head === :boundscheck || head === :meta || head === :loopinfo)
+is_meta_expr_head(head::Symbol) = (head === :inbounds || head === :boundscheck || head === :meta ||
+    head === :loopinfo || head === :inline || head === :noinline)
 
 sym_isless(a::Symbol, b::Symbol) = ccall(:strcmp, Int32, (Ptr{UInt8}, Ptr{UInt8}), a, b) < 0
 
@@ -188,7 +189,7 @@ function specialize_method(method::Method, @nospecialize(atypes), sparams::Simpl
     if preexisting
         # check cached specializations
         # for an existing result stored there
-        return ccall(:jl_specializations_lookup, Any, (Any, Any), method, atypes)
+        return ccall(:jl_specializations_lookup, Any, (Any, Any), method, atypes)::Union{Nothing,MethodInstance}
     end
     return ccall(:jl_specializations_get_linfo, Ref{MethodInstance}, (Any, Any, Any), method, atypes, sparams)
 end