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.

base/compiler/abstractinterpretation.jl

@@ -595,7 +595,7 @@ 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
@@ -699,7 +699,8 @@ 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
@@ -717,7 +718,8 @@ 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, 0x00, match) !== nothing
         end
     end
     if !cache_inlineable
@@ -1896,7 +1898,8 @@ 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

@@ -21,21 +21,30 @@ function push!(et::EdgeTracker, ci::CodeInstance)
     push!(et, ci.def)
 end
 
-struct InliningState{S <: Union{EdgeTracker, Nothing}, T, P}
+struct InliningState{S <: Union{EdgeTracker, Nothing}, T, I<:AbstractInterpreter}
     params::OptimizationParams
     et::S
     mi_cache::T
-    policy::P
+    interp::I
 end
 
-function default_inlining_policy(@nospecialize(src))
+function inlining_policy(interp::AbstractInterpreter, @nospecialize(src), stmt_flag::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
+        # NOTE we can make inference try to keep the source if the call is going to be inlined,
+        # but then inlining will depend on local state of inference and so the first entry
+        # and the succeeding ones may generate different code; rather we always re-infer
+        # the source to avoid the problem while it's obviously not most efficient
+        # HACK disable inlining for the re-inference to avoid cycles by making sure the following inference never comes here again
+        interp = NativeInterpreter(get_world_counter(interp); opt_params = OptimizationParams(; inlining = false))
+        src, rt = typeinf_code(interp, match.method, match.spec_types, match.sparams, true)
+        return src
     end
     return nothing
 end
@@ -57,7 +66,7 @@ mutable struct OptimizationState
         inlining = InliningState(params,
             EdgeTracker(s_edges, frame.valid_worlds),
             WorldView(code_cache(interp), frame.world),
-            inlining_policy(interp))
+            interp)
         return new(frame.linfo,
                    frame.src, nothing, frame.stmt_info, frame.mod,
                    frame.sptypes, frame.slottypes, false,
@@ -86,7 +95,7 @@ mutable struct OptimizationState
         inlining = InliningState(params,
             nothing,
             WorldView(code_cache(interp), get_world_counter()),
-            inlining_policy(interp))
+            interp)
         return new(linfo,
                    src, nothing, stmt_info, mod,
                    sptypes_from_meth_instance(linfo), slottypes, false,
@@ -128,6 +137,10 @@ 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
@@ -173,6 +186,9 @@ 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_noinline(stmt_flag::UInt8) = stmt_flag & IR_FLAG_NOINLINE != 0
+
 # 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.
@@ -359,6 +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)
@@ -373,16 +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[]