effects: fix effects of atomic pointer operations (#57806)

The `_unsetindex!(::GenericMemoryRef)` function was being concretely
evaluated (meaning the code was actually being run during inference) by
the `REPLInterpreter`.

The root cause is that the Base compiler system forgot to mark
`atomic_pointerset` as "effect-free".

This is simply a bug in the effect system. It didn't cause problems
during normal Base compilation because `atomic_pointerset` was correctly
marked as "consistent" (concrete evaluation requires both "consistent"
and "effect-free"). However, this was still a pretty risky situation.

The reason this only caused problems with REPL completion is that the
`REPLInterpreter` intentionally ignores the "consistent" requirement
under certain conditions to achieve better completion accuracy. This is
usually fine, but it relies on "effect-free" being correctly marked. So,
when there's a critical bug like this in the effect system, these kinds
of dangerous issues can occur.

As discussed with Jameson earlier, the effects of atomic pointer
operations are not precisely defined.
This commit includes the minimal changes necessary to fix
#57780, but a more extensive audit is planned for later.

- closes #57780

---------

Co-authored-by: Jameson Nash <[email protected]>

Author: aviatesk

Compiler/src/tfuncs.jl

@@ -2407,11 +2407,8 @@ const _ARGMEM_BUILTINS = Any[
 ]
 
 const _INCONSISTENT_INTRINSICS = Any[
-    Intrinsics.pointerref,      # this one is volatile
-    Intrinsics.sqrt_llvm_fast,  # this one may differ at runtime (by a few ulps)
-    Intrinsics.have_fma,        # this one depends on the runtime environment
-    Intrinsics.cglobal,         # cglobal lookup answer changes at runtime
-    # ... and list fastmath intrinsics:
+    # all is_pure_intrinsic_infer plus
+    # ... all the unsound fastmath functions which should have been in is_pure_intrinsic_infer
     # join(string.("Intrinsics.", sort(filter(endswith("_fast")∘string, names(Core.Intrinsics)))), ",\n")
     Intrinsics.add_float_fast,
     Intrinsics.div_float_fast,
@@ -2936,36 +2933,48 @@ function intrinsic_nothrow(f::IntrinsicFunction, argtypes::Vector{Any})
     return intrinsic_exct(SimpleInferenceLattice.instance, f, argtypes) === Union{}
 end
 
-# whether `f` is pure for inference
-function is_pure_intrinsic_infer(f::IntrinsicFunction)
-    return !(f === Intrinsics.pointerref || # this one is volatile
-             f === Intrinsics.pointerset || # this one is never effect-free
-             f === Intrinsics.llvmcall ||   # this one is never effect-free
-             f === Intrinsics.sqrt_llvm_fast ||  # this one may differ at runtime (by a few ulps)
-             f === Intrinsics.have_fma ||  # this one depends on the runtime environment
-             f === Intrinsics.cglobal)  # cglobal lookup answer changes at runtime
+function _is_effect_free_infer(f::IntrinsicFunction)
+     return !(f === Intrinsics.pointerset ||
+              f === Intrinsics.atomic_pointerref ||
+              f === Intrinsics.atomic_pointerset ||
+              f === Intrinsics.atomic_pointerswap ||
+              # f === Intrinsics.atomic_pointermodify ||
+              f === Intrinsics.atomic_pointerreplace ||
+              f === Intrinsics.atomic_fence)
 end
 
-# whether `f` is effect free if nothrow
-function intrinsic_effect_free_if_nothrow(@nospecialize f)
-    return f === Intrinsics.pointerref ||
-           f === Intrinsics.have_fma ||
-           is_pure_intrinsic_infer(f)
+# whether `f` is pure for inference
+function is_pure_intrinsic_infer(f::IntrinsicFunction, is_effect_free::Union{Nothing,Bool}=nothing)
+    if is_effect_free === nothing
+        is_effect_free = _is_effect_free_infer(f)
+    end
+    return is_effect_free && !(
+            f === Intrinsics.llvmcall ||              # can do arbitrary things
+            f === Intrinsics.atomic_pointermodify ||  # can do arbitrary things
+            f === Intrinsics.pointerref ||            # this one is volatile
+            f === Intrinsics.sqrt_llvm_fast ||        # this one may differ at runtime (by a few ulps)
+            f === Intrinsics.have_fma ||              # this one depends on the runtime environment
+            f === Intrinsics.cglobal)                 # cglobal lookup answer changes at runtime
 end
 
 function intrinsic_effects(f::IntrinsicFunction, argtypes::Vector{Any})
     if f === Intrinsics.llvmcall
         # llvmcall can do arbitrary things
         return Effects()
+    elseif f === atomic_pointermodify
+        # atomic_pointermodify has memory effects, plus any effects from the ModifyOpInfo
+        return Effects()
     end
-    if contains_is(_INCONSISTENT_INTRINSICS, f)
-        consistent = ALWAYS_FALSE
-    else
+    is_effect_free = _is_effect_free_infer(f)
+    effect_free = is_effect_free ? ALWAYS_TRUE : ALWAYS_FALSE
+    if ((is_pure_intrinsic_infer(f, is_effect_free) && !contains_is(_INCONSISTENT_INTRINSICS, f)) ||
+        f === Intrinsics.pointerset || f === Intrinsics.atomic_pointerset || f === Intrinsics.atomic_fence)
         consistent = ALWAYS_TRUE
+    else
+        consistent = ALWAYS_FALSE
     end
-    effect_free = !(f === Intrinsics.pointerset) ? ALWAYS_TRUE : ALWAYS_FALSE
     nothrow = intrinsic_nothrow(f, argtypes)
-    inaccessiblememonly = ALWAYS_TRUE
+    inaccessiblememonly = is_effect_free && !(f === Intrinsics.pointerref) ? ALWAYS_TRUE : ALWAYS_FALSE
     return Effects(EFFECTS_TOTAL; consistent, effect_free, nothrow, inaccessiblememonly)
 end
 

Compiler/test/effects.jl

@@ -1427,3 +1427,37 @@ end |> !Compiler.is_nothrow
 @test Base.infer_effects((UInt64,)) do x
     return Core.Intrinsics.uitofp(Int64, x)
 end |> !Compiler.is_nothrow
+
+# effects modeling for pointer-related intrinsics
+let effects = Base.infer_effects(Core.Intrinsics.pointerref, Tuple{Vararg{Any}})
+    @test !Compiler.is_consistent(effects)
+    @test Compiler.is_effect_free(effects)
+    @test !Compiler.is_inaccessiblememonly(effects)
+end
+let effects = Base.infer_effects(Core.Intrinsics.pointerset, Tuple{Vararg{Any}})
+    @test Compiler.is_consistent(effects)
+    @test !Compiler.is_effect_free(effects)
+end
+# effects modeling for atomic intrinsics
+# these functions especially need to be marked !effect_free since they imply synchronization
+for atomicfunc = Any[
+        Core.Intrinsics.atomic_pointerref,
+        Core.Intrinsics.atomic_pointerset,
+        Core.Intrinsics.atomic_pointerswap,
+        Core.Intrinsics.atomic_pointerreplace,
+        Core.Intrinsics.atomic_fence]
+    @test !Compiler.is_effect_free(Base.infer_effects(atomicfunc, Tuple{Vararg{Any}}))
+end
+
+# effects modeling for intrinsics that can do arbitrary things
+let effects = Base.infer_effects(Core.Intrinsics.llvmcall, Tuple{Vararg{Any}})
+    @test effects == Compiler.Effects()
+end
+let effects = Base.infer_effects(Core.Intrinsics.atomic_pointermodify, Tuple{Vararg{Any}})
+    @test effects == Compiler.Effects()
+end
+
+# JuliaLang/julia#57780
+let effects = Base.infer_effects(Base._unsetindex!, (MemoryRef{String},))
+    @test !Compiler.is_effect_free(effects)
+end

stdlib/REPL/test/replcompletions.jl

@@ -2484,3 +2484,9 @@ let (c, r, res) = test_complete_context("global xxx::Number = Base.", Main)
     @test res
     @test "pi" ∈ c
 end
+
+# JuliaLang/julia#57780
+const issue57780 = ["a", "b", "c"]
+const issue57780_orig = copy(issue57780)
+test_complete_context("empty!(issue57780).", Main)
+@test issue57780 == issue57780_orig