inference: remove throw block deoptimization completely (JuliaLang#…

…49260) Co-authored-by: Cody Tapscott <[email protected]> Co-authored-by: Oscar Smith <[email protected]>
lazarusA · Aug 17, 2024 · cf951f7 · cf951f7
1 parent 9bcf3ef
commit cf951f7
Show file tree

Hide file tree

Showing 10 changed files with 33 additions and 96 deletions.
diff --git a/base/compiler/abstractinterpretation.jl b/base/compiler/abstractinterpretation.jl
@@ -43,15 +43,6 @@ function abstract_call_gf_by_type(interp::AbstractInterpreter, @nospecialize(f),
                                   sv::AbsIntState, max_methods::Int)
     𝕃ₚ, 𝕃ᵢ = ipo_lattice(interp), typeinf_lattice(interp)
     ⊑ₚ, ⊔ₚ, ⊔ᵢ  = partialorder(𝕃ₚ), join(𝕃ₚ), join(𝕃ᵢ)
-    if !should_infer_this_call(interp, sv)
-        add_remark!(interp, sv, "Skipped call in throw block")
-        # At this point we are guaranteed to end up throwing on this path,
-        # which is all that's required for :consistent-cy. Of course, we don't
-        # know anything else about this statement.
-        effects = Effects(; consistent=ALWAYS_TRUE)
-        return CallMeta(Any, Any, effects, NoCallInfo())
-    end
-
     argtypes = arginfo.argtypes
     matches = find_method_matches(interp, argtypes, atype; max_methods)
     if isa(matches, FailedMethodMatch)

diff --git a/base/compiler/compiler.jl b/base/compiler/compiler.jl
@@ -184,8 +184,7 @@ baremodule BuildSettings
 using Core: ARGS, include
 using Core.Compiler: >, getindex, length
 
-MAX_METHODS::Int = 3
-UNOPTIMIZE_THROW_BLOCKS::Bool = true
+global MAX_METHODS::Int = 3
 
 if length(ARGS) > 2 && ARGS[2] === "--buildsettings"
     include(BuildSettings, ARGS[3])

diff --git a/base/compiler/inferencestate.jl b/base/compiler/inferencestate.jl
@@ -348,7 +348,6 @@ mutable struct InferenceState
         restrict_abstract_call_sites = isa(def, Module)
 
         # some more setups
-        InferenceParams(interp).unoptimize_throw_blocks && mark_throw_blocks!(src, handler_info)
         !iszero(cache_mode & CACHE_MODE_LOCAL) && push!(get_inference_cache(interp), result)
 
         this = new(
@@ -1102,30 +1101,6 @@ bail_out_apply(::AbstractInterpreter, state::InferenceLoopState, ::InferenceStat
 bail_out_apply(::AbstractInterpreter, state::InferenceLoopState, ::IRInterpretationState) =
     state.rt === Any
 
-function should_infer_this_call(interp::AbstractInterpreter, sv::InferenceState)
-    if InferenceParams(interp).unoptimize_throw_blocks
-        # Disable inference of calls in throw blocks, since we're unlikely to
-        # need their types. There is one exception however: If up until now, the
-        # function has not seen any side effects, we would like to make sure there
-        # aren't any in the throw block either to enable other optimizations.
-        if is_stmt_throw_block(get_curr_ssaflag(sv))
-            should_infer_for_effects(sv) || return false
-        end
-    end
-    return true
-end
-function should_infer_for_effects(sv::InferenceState)
-    def = sv.linfo.def
-    def isa Method || return false # toplevel frame will not be [semi-]concrete-evaluated
-    effects = sv.ipo_effects
-    override = decode_effects_override(def.purity)
-    effects.consistent === ALWAYS_FALSE && !is_effect_overridden(override, :consistent) && return false
-    effects.effect_free === ALWAYS_FALSE && !is_effect_overridden(override, :effect_free) && return false
-    !effects.terminates && !is_effect_overridden(override, :terminates_globally) && return false
-    return true
-end
-should_infer_this_call(::AbstractInterpreter, ::IRInterpretationState) = true
-
 add_remark!(::AbstractInterpreter, ::InferenceState, remark) = return
 add_remark!(::AbstractInterpreter, ::IRInterpretationState, remark) = return
 

diff --git a/base/compiler/optimize.jl b/base/compiler/optimize.jl
@@ -23,31 +23,29 @@ const IR_FLAG_INBOUNDS    = one(UInt32) << 0
 const IR_FLAG_INLINE      = one(UInt32) << 1
 # This statement is marked as @noinline by user
 const IR_FLAG_NOINLINE    = one(UInt32) << 2
-# This statement is on a code path that eventually `throw`s.
-const IR_FLAG_THROW_BLOCK = one(UInt32) << 3
 # An optimization pass has updated this statement in a way that may
 # have exposed information that inference did not see. Re-running
 # inference on this statement may be profitable.
-const IR_FLAG_REFINED     = one(UInt32) << 4
+const IR_FLAG_REFINED     = one(UInt32) << 3
 # This statement is proven :consistent
-const IR_FLAG_CONSISTENT  = one(UInt32) << 5
+const IR_FLAG_CONSISTENT  = one(UInt32) << 4
 # This statement is proven :effect_free
-const IR_FLAG_EFFECT_FREE = one(UInt32) << 6
+const IR_FLAG_EFFECT_FREE = one(UInt32) << 5
 # This statement is proven :nothrow
-const IR_FLAG_NOTHROW     = one(UInt32) << 7
+const IR_FLAG_NOTHROW     = one(UInt32) << 6
 # This statement is proven :terminates
-const IR_FLAG_TERMINATES  = one(UInt32) << 8
+const IR_FLAG_TERMINATES  = one(UInt32) << 7
 # This statement is proven :noub
-const IR_FLAG_NOUB        = one(UInt32) << 9
+const IR_FLAG_NOUB        = one(UInt32) << 8
 # TODO: Both of these should eventually go away once
 # This statement is :effect_free == EFFECT_FREE_IF_INACCESSIBLEMEMONLY
-const IR_FLAG_EFIIMO      = one(UInt32) << 10
+const IR_FLAG_EFIIMO      = one(UInt32) << 9
 # This statement is :inaccessiblememonly == INACCESSIBLEMEM_OR_ARGMEMONLY
-const IR_FLAG_INACCESSIBLEMEM_OR_ARGMEM = one(UInt32) << 11
+const IR_FLAG_INACCESSIBLEMEM_OR_ARGMEM = one(UInt32) << 10
 # This statement has no users and may be deleted if flags get refined to IR_FLAGS_REMOVABLE
-const IR_FLAG_UNUSED      = one(UInt32) << 12
+const IR_FLAG_UNUSED      = one(UInt32) << 11
 
-const NUM_IR_FLAGS = 13 # sync with julia.h
+const NUM_IR_FLAGS = 12 # sync with julia.h
 
 const IR_FLAGS_EFFECTS =
     IR_FLAG_CONSISTENT | IR_FLAG_EFFECT_FREE | IR_FLAG_NOTHROW | IR_FLAG_TERMINATES | IR_FLAG_NOUB
@@ -249,9 +247,8 @@ end
 
 _topmod(sv::OptimizationState) = _topmod(sv.mod)
 
-is_stmt_inline(stmt_flag::UInt32)      = has_flag(stmt_flag, IR_FLAG_INLINE)
-is_stmt_noinline(stmt_flag::UInt32)    = has_flag(stmt_flag, IR_FLAG_NOINLINE)
-is_stmt_throw_block(stmt_flag::UInt32) = has_flag(stmt_flag, IR_FLAG_THROW_BLOCK)
+is_stmt_inline(stmt_flag::UInt32) = has_flag(stmt_flag, IR_FLAG_INLINE)
+is_stmt_noinline(stmt_flag::UInt32) = has_flag(stmt_flag, IR_FLAG_NOINLINE)
 
 function new_expr_effect_flags(𝕃ₒ::AbstractLattice, args::Vector{Any}, src::Union{IRCode,IncrementalCompact}, pattern_match=nothing)
     Targ = args[1]
@@ -1272,7 +1269,7 @@ plus_saturate(x::Int, y::Int) = max(x, y, x+y)
 isknowntype(@nospecialize T) = (T === Union{}) || isa(T, Const) || isconcretetype(widenconst(T))
 
 function statement_cost(ex::Expr, line::Int, src::Union{CodeInfo, IRCode}, sptypes::Vector{VarState},
-                        params::OptimizationParams, error_path::Bool = false)
+                        params::OptimizationParams)
     #=const=# UNKNOWN_CALL_COST = 20
     head = ex.head
     if is_meta_expr_head(head)
@@ -1333,10 +1330,10 @@ function statement_cost(ex::Expr, line::Int, src::Union{CodeInfo, IRCode}, sptyp
                 return 0
             elseif (f === Core.memoryrefget || f === Core.memoryref_isassigned) && length(ex.args) >= 3
                 atyp = argextype(ex.args[2], src, sptypes)
-                return isknowntype(atyp) ? 1 : error_path ? params.inline_error_path_cost : params.inline_nonleaf_penalty
+                return isknowntype(atyp) ? 1 : params.inline_nonleaf_penalty
             elseif f === Core.memoryrefset! && length(ex.args) >= 3
                 atyp = argextype(ex.args[2], src, sptypes)
-                return isknowntype(atyp) ? 5 : error_path ? params.inline_error_path_cost : params.inline_nonleaf_penalty
+                return isknowntype(atyp) ? 5 : params.inline_nonleaf_penalty
             elseif f === typeassert && isconstType(widenconst(argextype(ex.args[3], src, sptypes)))
                 return 1
             end
@@ -1352,7 +1349,7 @@ function statement_cost(ex::Expr, line::Int, src::Union{CodeInfo, IRCode}, sptyp
         if extyp === Union{}
             return 0
         end
-        return error_path ? params.inline_error_path_cost : params.inline_nonleaf_penalty
+        return params.inline_nonleaf_penalty
     elseif head === :foreigncall
         foreigncall = ex.args[1]
         if foreigncall isa QuoteNode && foreigncall.value === :jl_string_ptr
@@ -1375,7 +1372,7 @@ function statement_cost(ex::Expr, line::Int, src::Union{CodeInfo, IRCode}, sptyp
         end
         a = ex.args[2]
         if a isa Expr
-            cost = plus_saturate(cost, statement_cost(a, -1, src, sptypes, params, error_path))
+            cost = plus_saturate(cost, statement_cost(a, -1, src, sptypes, params))
         end
         return cost
     elseif head === :copyast
@@ -1389,8 +1386,7 @@ function statement_or_branch_cost(@nospecialize(stmt), line::Int, src::Union{Cod
     thiscost = 0
     dst(tgt) = isa(src, IRCode) ? first(src.cfg.blocks[tgt].stmts) : tgt
     if stmt isa Expr
-        thiscost = statement_cost(stmt, line, src, sptypes, params,
-                                  is_stmt_throw_block(isa(src, IRCode) ? src.stmts.flag[line] : src.ssaflags[line]))::Int
+        thiscost = statement_cost(stmt, line, src, sptypes, params)::Int
     elseif stmt isa GotoNode
         # loops are generally always expensive
         # but assume that forward jumps are already counted for from

diff --git a/base/compiler/types.jl b/base/compiler/types.jl
@@ -156,11 +156,6 @@ Parameters that control abstract interpretation-based type inference operation.
   information available. [`Base.@constprop :aggressive`](@ref Base.@constprop) can have a
   more fine-grained control on this configuration with per-method annotation basis.
 ---
-- `inf_params.unoptimize_throw_blocks::Bool = true`\\
-  If `true`, skips inferring calls that are in a block that is known to `throw`.
-  It may improve the compiler latency without sacrificing the runtime performance
-  in common situations.
----
 - `inf_params.assume_bindings_static::Bool = false`\\
   If `true`, assumes that no new bindings will be added, i.e. a non-existing binding at
   inference time can be assumed to always not exist at runtime (and thus e.g. any access to
@@ -176,7 +171,6 @@ struct InferenceParams
     tuple_complexity_limit_depth::Int
     ipo_constant_propagation::Bool
     aggressive_constant_propagation::Bool
-    unoptimize_throw_blocks::Bool
     assume_bindings_static::Bool
     ignore_recursion_hardlimit::Bool
 
@@ -188,7 +182,6 @@ struct InferenceParams
         tuple_complexity_limit_depth::Int,
         ipo_constant_propagation::Bool,
         aggressive_constant_propagation::Bool,
-        unoptimize_throw_blocks::Bool,
         assume_bindings_static::Bool,
         ignore_recursion_hardlimit::Bool)
         return new(
@@ -199,7 +192,6 @@ struct InferenceParams
             tuple_complexity_limit_depth,
             ipo_constant_propagation,
             aggressive_constant_propagation,
-            unoptimize_throw_blocks,
             assume_bindings_static,
             ignore_recursion_hardlimit)
     end
@@ -213,7 +205,6 @@ function InferenceParams(
         #=tuple_complexity_limit_depth::Int=# 3,
         #=ipo_constant_propagation::Bool=# true,
         #=aggressive_constant_propagation::Bool=# false,
-        #=unoptimize_throw_blocks::Bool=# BuildSettings.UNOPTIMIZE_THROW_BLOCKS,
         #=assume_bindings_static::Bool=# false,
         #=ignore_recursion_hardlimit::Bool=# false);
     max_methods::Int = params.max_methods,
@@ -223,7 +214,6 @@ function InferenceParams(
     tuple_complexity_limit_depth::Int = params.tuple_complexity_limit_depth,
     ipo_constant_propagation::Bool = params.ipo_constant_propagation,
     aggressive_constant_propagation::Bool = params.aggressive_constant_propagation,
-    unoptimize_throw_blocks::Bool = params.unoptimize_throw_blocks,
     assume_bindings_static::Bool = params.assume_bindings_static,
     ignore_recursion_hardlimit::Bool = params.ignore_recursion_hardlimit)
     return InferenceParams(
@@ -234,7 +224,6 @@ function InferenceParams(
         tuple_complexity_limit_depth,
         ipo_constant_propagation,
         aggressive_constant_propagation,
-        unoptimize_throw_blocks,
         assume_bindings_static,
         ignore_recursion_hardlimit)
 end
@@ -259,10 +248,6 @@ Parameters that control optimizer operation.
   tuple return types (in hopes of splitting it up). `opt_params.inline_tupleret_bonus` will
   be added to `opt_params.inline_cost_threshold` when making inlining decision.
 ---
-- `opt_params.inline_error_path_cost::Int = 20`\\
-  Specifies the penalty cost for an un-optimized dynamic call in a block that is known to
-  `throw`. See also [`(inf_params::InferenceParams).unoptimize_throw_blocks`](@ref InferenceParams).
----
 - `opt_params.max_tuple_splat::Int = 32`\\
   When attempting to inline `Core._apply_iterate`, abort the optimization if the tuple
   contains more than this many elements.
@@ -289,7 +274,6 @@ struct OptimizationParams
     inline_cost_threshold::Int
     inline_nonleaf_penalty::Int
     inline_tupleret_bonus::Int
-    inline_error_path_cost::Int
     max_tuple_splat::Int
     compilesig_invokes::Bool
     assume_fatal_throw::Bool
@@ -300,7 +284,6 @@ struct OptimizationParams
         inline_cost_threshold::Int,
         inline_nonleaf_penalty::Int,
         inline_tupleret_bonus::Int,
-        inline_error_path_cost::Int,
         max_tuple_splat::Int,
         compilesig_invokes::Bool,
         assume_fatal_throw::Bool,
@@ -310,7 +293,6 @@ struct OptimizationParams
             inline_cost_threshold,
             inline_nonleaf_penalty,
             inline_tupleret_bonus,
-            inline_error_path_cost,
             max_tuple_splat,
             compilesig_invokes,
             assume_fatal_throw,
@@ -323,7 +305,6 @@ function OptimizationParams(
         #=inline_cost_threshold::Int=# 100,
         #=inline_nonleaf_penalty::Int=# 1000,
         #=inline_tupleret_bonus::Int=# 250,
-        #=inline_error_path_cost::Int=# 20,
         #=max_tuple_splat::Int=# 32,
         #=compilesig_invokes::Bool=# true,
         #=assume_fatal_throw::Bool=# false,
@@ -332,7 +313,6 @@ function OptimizationParams(
     inline_cost_threshold::Int = params.inline_cost_threshold,
     inline_nonleaf_penalty::Int = params.inline_nonleaf_penalty,
     inline_tupleret_bonus::Int = params.inline_tupleret_bonus,
-    inline_error_path_cost::Int = params.inline_error_path_cost,
     max_tuple_splat::Int = params.max_tuple_splat,
     compilesig_invokes::Bool = params.compilesig_invokes,
     assume_fatal_throw::Bool = params.assume_fatal_throw,
@@ -342,7 +322,6 @@ function OptimizationParams(
         inline_cost_threshold,
         inline_nonleaf_penalty,
         inline_tupleret_bonus,
-        inline_error_path_cost,
         max_tuple_splat,
         compilesig_invokes,
         assume_fatal_throw,

diff --git a/src/julia.h b/src/julia.h
@@ -276,7 +276,7 @@ typedef union __jl_purity_overrides_t {
 } _jl_purity_overrides_t;
 
 #define NUM_EFFECTS_OVERRIDES 10
-#define NUM_IR_FLAGS 13
+#define NUM_IR_FLAGS 12
 
 // This type describes a single function body
 typedef struct _jl_code_info_t {
@@ -288,16 +288,15 @@ typedef struct _jl_code_info_t {
         // 1 << 0 = inbounds region
         // 1 << 1 = callsite inline region
         // 1 << 2 = callsite noinline region
-        // 1 << 3 = throw block
-        // 1 << 4 = refined statement
-        // 1 << 5 = :consistent
-        // 1 << 6 = :effect_free
-        // 1 << 7 = :nothrow
-        // 1 << 8 = :terminates
-        // 1 << 9 = :noub
-        // 1 << 10 = :effect_free_if_inaccessiblememonly
-        // 1 << 11 = :inaccessiblemem_or_argmemonly
-        // 1 << 12-19 = callsite effects overrides
+        // 1 << 3 = refined statement
+        // 1 << 4 = :consistent
+        // 1 << 5 = :effect_free
+        // 1 << 6 = :nothrow
+        // 1 << 7 = :terminates
+        // 1 << 8 = :noub
+        // 1 << 9 = :effect_free_if_inaccessiblememonly
+        // 1 << 10 = :inaccessiblemem_or_argmemonly
+        // 1 << 11-19 = callsite effects overrides
     // miscellaneous data:
     jl_array_t *slotnames; // names of local variables
     jl_array_t *slotflags;  // local var bit flags

diff --git a/stdlib/REPL/src/REPLCompletions.jl b/stdlib/REPL/src/REPLCompletions.jl
@@ -556,8 +556,7 @@ struct REPLInterpreter <: CC.AbstractInterpreter
     function REPLInterpreter(limit_aggressive_inference::Bool=false;
                              world::UInt = Base.get_world_counter(),
                              inf_params::CC.InferenceParams = CC.InferenceParams(;
-                                 aggressive_constant_propagation=true,
-                                 unoptimize_throw_blocks=false),
+                                 aggressive_constant_propagation=true),
                              opt_params::CC.OptimizationParams = CC.OptimizationParams(),
                              inf_cache::Vector{CC.InferenceResult} = CC.InferenceResult[])
         return new(limit_aggressive_inference, world, inf_params, opt_params, inf_cache)

diff --git a/test/compiler/AbstractInterpreter.jl b/test/compiler/AbstractInterpreter.jl
@@ -176,7 +176,6 @@ end == Val{6}
 @newinterp Issue48097Interp
 @MethodTable ISSUE_48097_MT
 CC.method_table(interp::Issue48097Interp) = CC.OverlayMethodTable(CC.get_inference_world(interp), ISSUE_48097_MT)
-CC.InferenceParams(::Issue48097Interp) = CC.InferenceParams(; unoptimize_throw_blocks=false)
 function CC.concrete_eval_eligible(interp::Issue48097Interp,
     @nospecialize(f), result::CC.MethodCallResult, arginfo::CC.ArgInfo, sv::CC.AbsIntState)
     ret = @invoke CC.concrete_eval_eligible(interp::CC.AbstractInterpreter,

diff --git a/test/compiler/codegen.jl b/test/compiler/codegen.jl
@@ -697,7 +697,7 @@ mktempdir() do pfx
         libs_deleted += 1
     end
     @test libs_deleted > 0
-    @test readchomp(`$pfx/bin/$(Base.julia_exename()) -e 'print("no codegen!\n")'`) == "no codegen!"
+    @test readchomp(`$pfx/bin/$(Base.julia_exename()) --startup-file=no -e 'print("no codegen!\n")'`) == "no codegen!"
 
     # PR #47343
     libs_emptied = 0

diff --git a/test/dict.jl b/test/dict.jl
@@ -1510,9 +1510,9 @@ end
 for T in (Int, Float64, String, Symbol)
     @testset let T=T
         @test !Core.Compiler.is_consistent(Base.infer_effects(getindex, (Dict{T,Any}, T)))
-        @test_broken Core.Compiler.is_effect_free(Base.infer_effects(getindex, (Dict{T,Any}, T)))
+        @test Core.Compiler.is_effect_free(Base.infer_effects(getindex, (Dict{T,Any}, T)))
         @test !Core.Compiler.is_nothrow(Base.infer_effects(getindex, (Dict{T,Any}, T)))
-        @test_broken Core.Compiler.is_terminates(Base.infer_effects(getindex, (Dict{T,Any}, T)))
+        @test Core.Compiler.is_terminates(Base.infer_effects(getindex, (Dict{T,Any}, T)))
     end
 end