use a local cache for activity reg

Author: vchuravy

src/analyses/activity.jl

@@ -1,14 +1,3 @@
-@enum ActivityState begin
-    AnyState = 0
-    ActiveState = 1
-    DupState = 2
-    MixedState = 3
-end
-
-@inline function Base.:|(a1::ActivityState, a2::ActivityState)
-    ActivityState(Int(a1) | Int(a2))
-end
-
 @inline element(::Val{T}) where {T} = T
 
 @inline ptreltype(::Type{Ptr{T}}) where {T} = T
@@ -393,6 +382,14 @@ Base.@nospecializeinfer @inline function active_reg_inner(
     return ty
 end
 
+function active_reg_cached(ctx::EnzymeContext, @nospecialize(ST::Type); justActive=false, UnionSret = false, AbstractIsMixed = false)
+    key = (ST, justActive, UnionSret, AbstractIsMixed)
+    get!(ctx.activity_cache, key) do
+        set = Base.IdSet{Type}()
+        active_reg_inner(ST, set, ctx.world, justActive, UnionSret, AbstractIsMixed)
+    end
+end
+
 Base.@nospecializeinfer @inline function active_reg(@nospecialize(ST::Type), world::UInt; justActive=false, UnionSret = false, AbstractIsMixed = false)
     set = Base.IdSet{Type}()
     return active_reg_inner(ST, set, world, justActive, UnionSret, AbstractIsMixed)

src/compiler.jl

@@ -41,6 +41,25 @@ import LLVM: Target, TargetMachine
 import SparseArrays
 using Printf
 
+@enum ActivityState begin
+    AnyState = 0
+    ActiveState = 1
+    DupState = 2
+    MixedState = 3
+end
+
+@inline function Base.:|(a1::ActivityState, a2::ActivityState)
+    ActivityState(Int(a1) | Int(a2))
+end
+
+mutable struct EnzymeContext
+    world::UInt
+    activity_cache::Dict{Tuple{Type,Bool,Bool,Bool},ActivityState}
+    function EnzymeContext(world)
+        new(world, Dict{Tuple{Type,Bool,Bool,Bool},ActivityState}())
+    end
+end
+
 using Preferences
 
 bitcode_replacement() = parse(Bool, @load_preference("bitcode_replacement", "true"))
@@ -3109,7 +3128,7 @@ function create_abi_wrapper(
         # 3 is index of shadow
         if existed[3] != 0 &&
            sret_union &&
-           active_reg(pactualRetType, world; justActive=true, UnionSret=true) == ActiveState
+           active_reg_cached(interp.context, pactualRetType; justActive=true, UnionSret=true) == ActiveState
             rewrite_union_returns_as_ref(enzymefn, data[3], world, width)
         end
         returnNum = 0
@@ -4773,7 +4792,7 @@ end
     if params.err_if_func_written
         FT = TT.parameters[1]
         Ty = eltype(FT)
-        reg = active_reg(Ty, job.world)
+        reg = active_reg_cached(interp.context, Ty)
         if reg == DupState || reg == MixedState
             swiftself = has_swiftself(primalf)
             todo = LLVM.Value[parameters(primalf)[1+swiftself]]
@@ -4797,7 +4816,7 @@ end
                     if !mayWriteToMemory(user)
                         slegal, foundv, byref = abs_typeof(user)
                         if slegal
-                            reg2 = active_reg(foundv, job.world)
+                            reg2 = active_reg_cached(interp.context, foundv)
                             if reg2 == ActiveState || reg2 == AnyState
                                 continue
                             end
@@ -4825,7 +4844,7 @@ end
                         if operands(user)[2] == cur
                             slegal, foundv, byref = abs_typeof(operands(user)[1])
                             if slegal
-                                reg2 = active_reg(foundv, job.world)
+                                reg2 = active_reg_cached(interp.context, foundv)
                                 if reg2 == AnyState
                                     continue
                                 end
@@ -4859,7 +4878,7 @@ end
                             if is_readonly(called)
                                 slegal, foundv, byref = abs_typeof(user)
                                 if slegal
-                                    reg2 = active_reg(foundv, job.world)
+                                    reg2 = active_reg_cached(interp.context, foundv)
                                     if reg2 == ActiveState || reg2 == AnyState
                                         continue
                                     end
@@ -4877,7 +4896,7 @@ end
                                 end
                                 slegal, foundv, byref = abs_typeof(user)
                                 if slegal
-                                    reg2 = active_reg(foundv, job.world)
+                                    reg2 = active_reg_cached(interp.context, foundv)
                                     if reg2 == ActiveState || reg2 == AnyState
                                         continue
                                     end

src/compiler/interpreter.jl

@@ -136,6 +136,8 @@ struct EnzymeInterpreter{T} <: AbstractInterpreter
     within_autodiff_rewrite::Bool
 
     handler::T
+
+    context::Enzyme.Compiler.EnzymeContext
 end
 
 const SigCache = Dict{Tuple, Dict{UInt, Base.IdSet{Type}}}()
@@ -247,7 +249,8 @@ function EnzymeInterpreter(
         inactive_rules::Bool,
         broadcast_rewrite::Bool,
         within_autodiff_rewrite::Bool,
-        handler
+        handler,
+        Enzyme.Compiler.EnzymeContext(world)
     )
 end
 
@@ -278,7 +281,9 @@ function EnzymeInterpreter(interp::EnzymeInterpreter;
     inactive_rules = interp.inactive_rules,
     broadcast_rewrite = interp.broadcast_rewrite,
     within_autodiff_rewrite = interp.within_autodiff_rewrite,
-    handler = interp.handler)
+    handler = interp.handler,
+    context = interp.context,)
+    @assert context.world == world
     return EnzymeInterpreter(
         cache_or_token,
         mt,
@@ -291,7 +296,8 @@ function EnzymeInterpreter(interp::EnzymeInterpreter;
         inactive_rules,
         broadcast_rewrite,
         within_autodiff_rewrite,
-        handler
+        handler,
+        context
     )
 end