effects: allow override of :nonoverlayed effect bit

aviatesk · aviatesk · commit 077817575450 · 2023-08-29T12:54:26.000+09:00
Certain external `AbstractInterpreters`, such as GPUCompiler.jl, have long sought the ability to allow concrete evaluation for specific overlay-ed methods to achieve optimal inference accuracy. This is currently not permitted, although it should be safe when an overlay-ed method has the same semantics as the original method, and its result can be safely replaced with the result of the original method. Refer to JuliaGPU/GPUCompiler.jl#384 for more examples. To address this issue, this commit introduces the capability to override the `:nonoverlayed` effect bit using `@assume_effects`. With the enhancements in PR #51078, this override behaves similarly to other effect bits. Consequently, external `AbstractInterpreters` can utilize this feature to permit concrete evaluation for annotated overlay-ed methods, e.g. ```julia @overlay OVERLAY_MT Base.@assume_effects :nonoverlayed f(x) = [...] ``` However, it now seems awkward to annotate a method with `Base.@assume_effects :nonoverlayed` when it is actually marked with `@overlay`. A more intuitive terminology, like `native_executable`, might be more appropriate for renaming the `:nonoverlayed` effect bit.
diff --git a/base/boot.jl b/base/boot.jl
@@ -464,7 +464,8 @@ macro _foldable_meta()
         #=:terminates_locally=#false,
         #=:notaskstate=#false,
         #=:inaccessiblememonly=#false,
-        #=:noub=#true))
+        #=:noub=#true,
+        #=:nonoverlayed=#false))
 end
 
 const NTuple{N,T} = Tuple{Vararg{T,N}}
diff --git a/base/c.jl b/base/c.jl
@@ -715,6 +715,6 @@ macro ccall(expr)
     return ccall_macro_lower(:ccall, ccall_macro_parse(expr)...)
 end
 
-macro ccall_effects(effects::UInt8, expr)
+macro ccall_effects(effects::UInt32, expr)
     return ccall_macro_lower((:ccall, effects), ccall_macro_parse(expr)...)
 end
diff --git a/base/compiler/abstractinterpretation.jl b/base/compiler/abstractinterpretation.jl
@@ -2557,7 +2557,7 @@ function abstract_eval_foreigncall(interp::AbstractInterpreter, e::Expr, vtypes:
         abstract_eval_value(interp, x, vtypes, sv)
     end
     cconv = e.args[5]
-    if isa(cconv, QuoteNode) && (v = cconv.value; isa(v, Tuple{Symbol, UInt8}))
+    if isa(cconv, QuoteNode) && (v = cconv.value; isa(v, Tuple{Symbol, UInt32}))
         override = decode_effects_override(v[2])
         effects = Effects(effects;
             consistent          = override.consistent ? ALWAYS_TRUE : effects.consistent,
diff --git a/base/compiler/effects.jl b/base/compiler/effects.jl
@@ -258,29 +258,32 @@ struct EffectsOverride
     notaskstate::Bool
     inaccessiblememonly::Bool
     noub::Bool
+    nonoverlayed::Bool
 end
 
 function encode_effects_override(eo::EffectsOverride)
-    e = 0x00
-    eo.consistent          && (e |= (0x01 << 0))
-    eo.effect_free         && (e |= (0x01 << 1))
-    eo.nothrow             && (e |= (0x01 << 2))
-    eo.terminates_globally && (e |= (0x01 << 3))
-    eo.terminates_locally  && (e |= (0x01 << 4))
-    eo.notaskstate         && (e |= (0x01 << 5))
-    eo.inaccessiblememonly && (e |= (0x01 << 6))
-    eo.noub                && (e |= (0x01 << 7))
+    e = zero(UInt32)
+    eo.consistent          && (e |= (1 << 0) % UInt32)
+    eo.effect_free         && (e |= (1 << 1) % UInt32)
+    eo.nothrow             && (e |= (1 << 2) % UInt32)
+    eo.terminates_globally && (e |= (1 << 3) % UInt32)
+    eo.terminates_locally  && (e |= (1 << 4) % UInt32)
+    eo.notaskstate         && (e |= (1 << 5) % UInt32)
+    eo.inaccessiblememonly && (e |= (1 << 6) % UInt32)
+    eo.noub                && (e |= (1 << 7) % UInt32)
+    eo.nonoverlayed        && (e |= (1 << 8) % UInt32)
     return e
 end
 
-function decode_effects_override(e::UInt8)
+function decode_effects_override(e::UInt32)
     return EffectsOverride(
-        (e & (0x01 << 0)) != 0x00,
-        (e & (0x01 << 1)) != 0x00,
-        (e & (0x01 << 2)) != 0x00,
-        (e & (0x01 << 3)) != 0x00,
-        (e & (0x01 << 4)) != 0x00,
-        (e & (0x01 << 5)) != 0x00,
-        (e & (0x01 << 6)) != 0x00,
-        (e & (0x01 << 7)) != 0x00)
+        !iszero(e & (1 << 0)),
+        !iszero(e & (1 << 1)),
+        !iszero(e & (1 << 2)),
+        !iszero(e & (1 << 3)),
+        !iszero(e & (1 << 4)),
+        !iszero(e & (1 << 5)),
+        !iszero(e & (1 << 6)),
+        !iszero(e & (1 << 7)),
+        !iszero(e & (1 << 8)))
 end
diff --git a/base/compiler/typeinfer.jl b/base/compiler/typeinfer.jl
@@ -500,6 +500,9 @@ function adjust_effects(sv::InferenceState)
         if is_effect_overridden(override, :noub)
             ipo_effects = Effects(ipo_effects; noub=true)
         end
+        if is_effect_overridden(override, :nonoverlayed)
+            ipo_effects = Effects(ipo_effects; nonoverlayed=true)
+        end
     end
 
     return ipo_effects
diff --git a/base/essentials.jl b/base/essentials.jl
@@ -209,7 +209,8 @@ macro _total_meta()
         #=:terminates_locally=#false,
         #=:notaskstate=#true,
         #=:inaccessiblememonly=#true,
-        #=:noub=#true))
+        #=:noub=#true,
+        #=:nonoverlayed=#false))
 end
 # can be used in place of `@assume_effects :foldable` (supposed to be used for bootstrapping)
 macro _foldable_meta()
@@ -221,7 +222,8 @@ macro _foldable_meta()
         #=:terminates_locally=#false,
         #=:notaskstate=#false,
         #=:inaccessiblememonly=#true,
-        #=:noub=#true))
+        #=:noub=#true,
+        #=:nonoverlayed=#false))
 end
 # can be used in place of `@assume_effects :nothrow` (supposed to be used for bootstrapping)
 macro _nothrow_meta()
@@ -233,7 +235,8 @@ macro _nothrow_meta()
         #=:terminates_locally=#false,
         #=:notaskstate=#false,
         #=:inaccessiblememonly=#false,
-        #=:noub=#false))
+        #=:noub=#false,
+        #=:nonoverlayed=#false))
 end
 # can be used in place of `@assume_effects :terminates_locally` (supposed to be used for bootstrapping)
 macro _terminates_locally_meta()
@@ -245,7 +248,8 @@ macro _terminates_locally_meta()
         #=:terminates_locally=#true,
         #=:notaskstate=#false,
         #=:inaccessiblememonly=#false,
-        #=:noub=#false))
+        #=:noub=#false,
+        #=:nonoverlayed=#false))
 end
 # can be used in place of `@assume_effects :effect_free :terminates_locally` (supposed to be used for bootstrapping)
 macro _effect_free_terminates_locally_meta()
@@ -257,7 +261,8 @@ macro _effect_free_terminates_locally_meta()
         #=:terminates_locally=#true,
         #=:notaskstate=#false,
         #=:inaccessiblememonly=#false,
-        #=:noub=#false))
+        #=:noub=#false,
+        #=:nonoverlayed=#false))
 end
 
 # another version of inlining that propagates an inbounds context
diff --git a/base/expr.jl b/base/expr.jl
@@ -714,8 +714,9 @@ macro assume_effects(args...)
         ex = nothing
         idx = length(args)
     end
-    (consistent, effect_free, nothrow, terminates_globally, terminates_locally, notaskstate, inaccessiblememonly, noub) =
-        (false, false, false, false, false, false, false, false, false)
+    (consistent, effect_free, nothrow, terminates_globally, terminates_locally,
+        notaskstate, inaccessiblememonly, noub, nonoverlayed) =
+        (false, false, false, false, false, false, false, false, false, false)
     for org_setting in args[1:idx]
         (setting, val) = compute_assumed_setting(org_setting)
         if setting === :consistent
@@ -734,6 +735,8 @@ macro assume_effects(args...)
             inaccessiblememonly = val
         elseif setting === :noub
             noub = val
+        elseif setting === :nonoverlayed
+            nonoverlayed = val
         elseif setting === :foldable
             consistent = effect_free = terminates_globally = noub = val
         elseif setting === :removable
@@ -746,15 +749,18 @@ macro assume_effects(args...)
     end
     if is_function_def(inner)
         return esc(pushmeta!(ex, :purity,
-            consistent, effect_free, nothrow, terminates_globally, terminates_locally, notaskstate, inaccessiblememonly, noub))
+            consistent, effect_free, nothrow, terminates_globally, terminates_locally,
+            notaskstate, inaccessiblememonly, noub, nonoverlayed))
     elseif isexpr(ex, :macrocall) && ex.args[1] === Symbol("@ccall")
         ex.args[1] = GlobalRef(Base, Symbol("@ccall_effects"))
         insert!(ex.args, 3, Core.Compiler.encode_effects_override(Core.Compiler.EffectsOverride(
-            consistent, effect_free, nothrow, terminates_globally, terminates_locally, notaskstate, inaccessiblememonly, noub)))
+            consistent, effect_free, nothrow, terminates_globally, terminates_locally,
+            notaskstate, inaccessiblememonly, noub, nonoverlayed)))
         return esc(ex)
     else # anonymous function case
         return Expr(:meta, Expr(:purity,
-            consistent, effect_free, nothrow, terminates_globally, terminates_locally, notaskstate, inaccessiblememonly, noub))
+            consistent, effect_free, nothrow, terminates_globally, terminates_locally,
+            notaskstate, inaccessiblememonly, noub, nonoverlayed))
     end
 end
 
diff --git a/base/strings/string.jl b/base/strings/string.jl
@@ -87,7 +87,11 @@ end
 
 # This is @assume_effects :effect_free :nothrow :terminates_globally @ccall jl_alloc_string(n::Csize_t)::Ref{String},
 # but the macro is not available at this time in bootstrap, so we write it manually.
-@eval _string_n(n::Integer) = $(Expr(:foreigncall, QuoteNode(:jl_alloc_string), Ref{String}, Expr(:call, Expr(:core, :svec), :Csize_t), 1, QuoteNode((:ccall,0xe)), :(convert(Csize_t, n))))
+@eval function _string_n(n::Integer)
+    return $(Expr(:foreigncall, QuoteNode(:jl_alloc_string), Ref{String},
+        Expr(:call, Expr(:core, :svec), :Csize_t), 1, QuoteNode((:ccall, 0x0000000e)),
+        :(convert(Csize_t, n))))
+end
 
 """
     String(s::AbstractString)
diff --git a/src/jltypes.c b/src/jltypes.c
@@ -3005,7 +3005,7 @@ void jl_init_types(void) JL_GC_DISABLED
                             jl_bool_type,
                             jl_uint8_type,
                             jl_uint8_type,
-                            jl_uint8_type,
+                            jl_uint32_type,
                             jl_uint16_type),
                         jl_emptysvec,
                         0, 1, 22);
@@ -3074,7 +3074,7 @@ void jl_init_types(void) JL_GC_DISABLED
                             jl_bool_type,
                             jl_uint8_type,
                             jl_uint8_type,
-                            jl_uint8_type),
+                            jl_uint32_type),
                         jl_emptysvec,
                         0, 1, 10);
     //const static uint32_t method_constfields[1] = { 0x03fc065f }; // (1<<0)|(1<<1)|(1<<2)|(1<<3)|(1<<4)|(1<<6)|(1<<9)|(1<<10)|(1<<18)|(1<<19)|(1<<20)|(1<<21)|(1<<22)|(1<<23)|(1<<24)|(1<<25);
diff --git a/src/julia.h b/src/julia.h
@@ -269,8 +269,9 @@ typedef union __jl_purity_overrides_t {
         uint8_t ipo_notaskstate         : 1;
         uint8_t ipo_inaccessiblememonly : 1;
         uint8_t ipo_noub                : 1;
+        uint8_t ipo_nonoverlayed        : 1;
     } overrides;
-    uint8_t bits;
+    uint32_t bits;
 } _jl_purity_overrides_t;
 
 // This type describes a single function body
@@ -428,22 +429,24 @@ typedef struct _jl_code_instance_t {
     // see also encode_effects() and decode_effects() in `base/compiler/effects.jl`,
     uint32_t ipo_purity_bits;
     // ipo_purity_flags:
-    //     uint8_t ipo_consistent          : 2;
+    //     uint8_t ipo_consistent          : 3;
     //     uint8_t ipo_effect_free         : 2;
-    //     uint8_t ipo_nothrow             : 2;
-    //     uint8_t ipo_terminates          : 2;
-    //     uint8_t ipo_nonoverlayed        : 1;
+    //     uint8_t ipo_nothrow             : 1;
+    //     uint8_t ipo_terminates          : 1;
     //     uint8_t ipo_notaskstate         : 2;
     //     uint8_t ipo_inaccessiblememonly : 2;
+    //     uint8_t ipo_nonoverlayed        : 1;
+    //     uint8_t ipo_noinbounds          : 1;
     _Atomic(uint32_t) purity_bits;
     // purity_flags:
-    //     uint8_t consistent          : 2;
-    //     uint8_t effect_free         : 2;
-    //     uint8_t nothrow             : 2;
-    //     uint8_t terminates          : 2;
-    //     uint8_t nonoverlayed        : 1;
-    //     uint8_t notaskstate         : 2;
-    //     uint8_t inaccessiblememonly : 2;
+    //     uint8_t ipo_consistent          : 3;
+    //     uint8_t ipo_effect_free         : 2;
+    //     uint8_t ipo_nothrow             : 1;
+    //     uint8_t ipo_terminates          : 1;
+    //     uint8_t ipo_notaskstate         : 2;
+    //     uint8_t ipo_inaccessiblememonly : 2;
+    //     uint8_t ipo_nonoverlayed        : 1;
+    //     uint8_t ipo_noinbounds          : 1;
     jl_value_t *argescapes; // escape information of call arguments
 
     // compilation state cache
diff --git a/src/method.c b/src/method.c
@@ -192,7 +192,7 @@ static jl_value_t *resolve_globals(jl_value_t *expr, jl_module_t *module, jl_sve
                         jl_error("In ccall calling convention, expected two argument tuple or symbol.");
                     }
                     JL_TYPECHK(ccall method definition, symbol, jl_get_nth_field(cc, 0));
-                    JL_TYPECHK(ccall method definition, uint8, jl_get_nth_field(cc, 1));
+                    JL_TYPECHK(ccall method definition, uint32, jl_get_nth_field(cc, 1));
                 }
                 jl_exprargset(e, 0, resolve_globals(jl_exprarg(e, 0), module, sparam_vals, binding_effects, 1));
                 i++;
@@ -328,7 +328,7 @@ static void jl_code_info_set_ir(jl_code_info_t *li, jl_expr_t *ir)
                 else if (ma == (jl_value_t*)jl_no_constprop_sym)
                     li->constprop = 2;
                 else if (jl_is_expr(ma) && ((jl_expr_t*)ma)->head == jl_purity_sym) {
-                    if (jl_expr_nargs(ma) == 8) {
+                    if (jl_expr_nargs(ma) == 9) {
                         li->purity.overrides.ipo_consistent = jl_unbox_bool(jl_exprarg(ma, 0));
                         li->purity.overrides.ipo_effect_free = jl_unbox_bool(jl_exprarg(ma, 1));
                         li->purity.overrides.ipo_nothrow = jl_unbox_bool(jl_exprarg(ma, 2));
@@ -337,6 +337,7 @@ static void jl_code_info_set_ir(jl_code_info_t *li, jl_expr_t *ir)
                         li->purity.overrides.ipo_notaskstate = jl_unbox_bool(jl_exprarg(ma, 5));
                         li->purity.overrides.ipo_inaccessiblememonly = jl_unbox_bool(jl_exprarg(ma, 6));
                         li->purity.overrides.ipo_noub = jl_unbox_bool(jl_exprarg(ma, 7));
+                        li->purity.overrides.ipo_nonoverlayed = jl_unbox_bool(jl_exprarg(ma, 8));
                     }
                 }
                 else
diff --git a/stdlib/Serialization/src/Serialization.jl b/stdlib/Serialization/src/Serialization.jl
@@ -80,7 +80,7 @@ const TAGS = Any[
 const NTAGS = length(TAGS)
 @assert NTAGS == 255
 
-const ser_version = 24 # do not make changes without bumping the version #!
+const ser_version = 25 # do not make changes without bumping the version #!
 
 format_version(::AbstractSerializer) = ser_version
 format_version(s::Serializer) = s.version
@@ -1028,7 +1028,8 @@ function deserialize(s::AbstractSerializer, ::Type{Method})
     isva = deserialize(s)::Bool
     is_for_opaque_closure = false
     nospecializeinfer = false
-    constprop = purity = 0x00
+    constprop = 0x00
+    purity = format_version(s) >= 25 ? zero(UInt32) : 0x00
     template_or_is_opaque = deserialize(s)
     if isa(template_or_is_opaque, Bool)
         is_for_opaque_closure = template_or_is_opaque
@@ -1039,7 +1040,7 @@ function deserialize(s::AbstractSerializer, ::Type{Method})
             constprop = deserialize(s)::UInt8
         end
         if format_version(s) >= 17
-            purity = deserialize(s)::UInt8
+            purity = deserialize(s)::(format_version(s) >= 25 ? UInt32 : UInt8)
         end
         template = deserialize(s)
     else
@@ -1211,7 +1212,7 @@ function deserialize(s::AbstractSerializer, ::Type{CodeInfo})
         ci.constprop = deserialize(s)::UInt8
     end
     if format_version(s) >= 17
-        ci.purity = deserialize(s)::UInt8
+        ci.purity = deserialize(s)::(format_version(s) >= 25 ? UInt32 : UInt8)
     end
     if format_version(s) >= 22
         ci.inlining_cost = deserialize(s)::UInt16
diff --git a/test/compiler/AbstractInterpreter.jl b/test/compiler/AbstractInterpreter.jl

Original file line number	Diff line number	Diff line change
`@@ -192,7 +192,7 @@ static jl_value_t resolve_globals(jl_value_t expr, jl_module_t *module, jl_sve`
`192`	`192`	`jl_error("In ccall calling convention, expected two argument tuple or symbol.");`
`193`	`193`	`}`
`194`	`194`	`JL_TYPECHK(ccall method definition, symbol, jl_get_nth_field(cc, 0));`
`195`		`- JL_TYPECHK(ccall method definition, uint8, jl_get_nth_field(cc, 1));`
	`195`	`+ JL_TYPECHK(ccall method definition, uint32, jl_get_nth_field(cc, 1));`
`196`	`196`	`}`
`197`	`197`	`jl_exprargset(e, 0, resolve_globals(jl_exprarg(e, 0), module, sparam_vals, binding_effects, 1));`
`198`	`198`	`i++;`
`@@ -328,7 +328,7 @@ static void jl_code_info_set_ir(jl_code_info_t li, jl_expr_t ir)`
`328`	`328`	`else if (ma == (jl_value_t*)jl_no_constprop_sym)`
`329`	`329`	`li->constprop = 2;`
`330`	`330`	`else if (jl_is_expr(ma) && ((jl_expr_t*)ma)->head == jl_purity_sym) {`
`331`		`- if (jl_expr_nargs(ma) == 8) {`
	`331`	`+ if (jl_expr_nargs(ma) == 9) {`
`332`	`332`	`li->purity.overrides.ipo_consistent = jl_unbox_bool(jl_exprarg(ma, 0));`
`333`	`333`	`li->purity.overrides.ipo_effect_free = jl_unbox_bool(jl_exprarg(ma, 1));`
`334`	`334`	`li->purity.overrides.ipo_nothrow = jl_unbox_bool(jl_exprarg(ma, 2));`
`@@ -337,6 +337,7 @@ static void jl_code_info_set_ir(jl_code_info_t li, jl_expr_t ir)`
`337`	`337`	`li->purity.overrides.ipo_notaskstate = jl_unbox_bool(jl_exprarg(ma, 5));`
`338`	`338`	`li->purity.overrides.ipo_inaccessiblememonly = jl_unbox_bool(jl_exprarg(ma, 6));`
`339`	`339`	`li->purity.overrides.ipo_noub = jl_unbox_bool(jl_exprarg(ma, 7));`
	`340`	`+ li->purity.overrides.ipo_nonoverlayed = jl_unbox_bool(jl_exprarg(ma, 8));`
`340`	`341`	`}`
`341`	`342`	`}`
`342`	`343`	`else`