diff --git a/src/irgen.jl b/src/irgen.jl
index 2dbf5ddd..ac253a75 100644
--- a/src/irgen.jl
+++ b/src/irgen.jl
@@ -343,28 +343,24 @@ function inline_unreachable_control_flow!(@nospecialize(job::CompilerJob), mod::
     return changed
 end
 
-# demote LLVM atomic loads and stores to plain ones
+# demote `unordered` LLVM atomic loads and stores to plain ones
 #
-# Julia marks accesses to heap references `unordered` so that a read racing with the GC, or
-# with another thread's write, cannot observe a torn pointer, and stores the type tag of a
-# freshly allocated object with `release` ordering so that no other thread can observe the
-# object before its header. There is no device GC and no such race for GPUCompiler to
-# protect against, so these orderings carry no meaning here, but not every back-end can
-# express them: SPIR-V's OpAtomicLoad/OpAtomicStore only take scalar integer or
-# floating-point operands, so the Khronos translator turns an atomic access of a pointer
-# into an invalid pointer-typed atomic that consumers reject (Intel's compiler fails with an
-# undefined `__spirv_AtomicLoad(long**, int, int)`), and AIR has no atomic load or store
-# instructions at all: Apple's back-end aborts on them (`XPC_ERROR_CONNECTION_INTERRUPTED`
-# from the driver; the macOS 26 AGX compiler reports `unable to legalize instruction:
-# store release (p0)` for a type-tag store through the generic pointer the device allocator
-# returns). Run after optimization, where dropping the ordering cannot enable new
-# transformations. Device-side atomics proper go through target intrinsics, not these
-# instructions, so every remaining one is such Julia bookkeeping and gets demoted.
-function demote_atomics!(mod::LLVM.Module)
+# Julia marks accesses to heap references `unordered`, so that a read racing with the GC or
+# with another thread's write cannot observe a torn pointer. There is no device GC, and
+# aligned accesses don't tear, so the ordering carries no meaning here. Not every back-end can
+# express it, though: SPIR-V's OpAtomicLoad/OpAtomicStore only take scalar integer or
+# floating-point operands, so the Khronos translator turns an atomic access of a pointer into
+# an invalid pointer-typed atomic, and Metal only has atomics on device and threadgroup memory,
+# while these accesses are often of objects passed by reference (e.g. an immutable struct
+# with a `DataType` field, or with a reference to a mutable object). Rejecting allocations of
+# objects with references (see `check_allocation!`) does not remove them, as they don't
+# involve an allocation. Other orderings are left alone, because they are user atomics (e.g.
+# UnsafeAtomics' `load`/`store!`) that the back-end must see.
+function demote_unordered_atomics!(mod::LLVM.Module)
     changed = false
     for f in functions(mod), bb in blocks(f), inst in instructions(bb)
         (inst isa LLVM.LoadInst || inst isa LLVM.StoreInst) || continue
-        is_atomic(inst) || continue
+        is_atomic(inst) && ordering(inst) == LLVM.API.LLVMAtomicOrderingUnordered || continue
         ordering!(inst, LLVM.API.LLVMAtomicOrderingNotAtomic)
         changed = true
     end
diff --git a/src/metal.jl b/src/metal.jl
index 65b3da95..3180b02f 100644
--- a/src/metal.jl
+++ b/src/metal.jl
@@ -597,6 +597,10 @@ function finish_ir!(@nospecialize(job::CompilerJob{MetalCompilerTarget}), mod::L
                                   entry::LLVM.Function)
     entry_fn = LLVM.name(entry)
 
+    # Julia's `unordered` heap-reference accesses are pointer-sized, and often of thread
+    # memory, neither of which AIR has atomics for (see `demote_unordered_atomics!`)
+    demote_unordered_atomics!(mod)
+
     # convert the kernel state argument to a reference
     if job.config.kernel && kernel_state_type(job) !== Nothing
         entry = kernel_state_to_reference!(job, mod, entry)
@@ -659,9 +663,6 @@ function lower_air!(@nospecialize(job::CompilerJob{MetalCompilerTarget}), mod::L
     # Metal.malloc uses.
     rewrite_generic_null_selects!(mod)
 
-    # AIR does not support LLVM atomic load/store instructions (see `demote_atomics!`)
-    demote_atomics!(mod)
-
     # the macOS 27 back-end rejects bare LLVM fences (Metal.jl#968)
     lower_fences!(job, mod)
 
diff --git a/src/spirv.jl b/src/spirv.jl
index 3a4515ba..a1486d69 100644
--- a/src/spirv.jl
+++ b/src/spirv.jl
@@ -116,8 +116,8 @@ function finish_ir!(job::CompilerJob{SPIRVCompilerTarget}, mod::LLVM.Module,
     lower_unreachable_control_flow!(job, mod)
 
     # SPIR-V cannot express atomic loads and stores of pointers, which is what Julia's
-    # heap-reference accesses and type-tag stores are; the orderings serve no purpose on device
-    demote_atomics!(mod)
+    # `unordered` heap-reference accesses are (see `demote_unordered_atomics!`)
+    demote_unordered_atomics!(mod)
 
     # the SPIR-V back-ends lower `llvm.minimum`/`llvm.maximum` to NaN-ignoring `fmin`/`fmax`
     lower_minimum_maximum!(mod)
diff --git a/test/metal.jl b/test/metal.jl
index 67c7b11f..d5d5af90 100644
--- a/test/metal.jl
+++ b/test/metal.jl
@@ -1806,7 +1806,6 @@ end
     end
 end
 
-
 @testset "unsupported allocations" begin
     # without a garbage collector, device code can only allocate objects that do not
     # reference other objects
@@ -1902,77 +1901,28 @@ end
     end
 end
 
-@testset "atomic demotion" begin
-    # Demote every LLVM atomic load/store, whatever its ordering or value type.
-    Context() do ctx
-        ir = """
-        define void @f(i8** %p, i8** %q, i64* %r) {
-        entry:
-          %a = load atomic i8*, i8** %p unordered, align 8
-          store atomic i8* %a, i8** %q unordered, align 8
-          %t = load atomic i8*, i8** %p acquire, align 8
-          store atomic i8* %t, i8** %q release, align 8
-          %b = load atomic i64, i64* %r monotonic, align 8
-          store atomic i64 %b, i64* %r release, align 8
-          store i8* %a, i8** %p, align 8
-          ret void
-        }
-        """
-        mod = parse(LLVM.Module, ir)
-        insts() = [i for f in functions(mod) for bb in blocks(f) for i in instructions(bb)]
-        memops() = filter(i -> i isa LLVM.LoadInst || i isa LLVM.StoreInst, insts())
-
-        @test count(is_atomic, memops()) == 6
-        @test GPUCompiler.demote_atomics!(mod)
-        @test count(is_atomic, memops()) == 0
-        @test !occursin("atomic", string(mod))
-        @test (verify(mod); true)
-        # idempotent
-        @test !GPUCompiler.demote_atomics!(mod)
-    end
-
-    # end-to-end: Julia's own `:unordered` accesses (as codegen emits for heap-reference
-    # fields) must reach the AIR as plain loads and stores
-    function kernel(p::Core.LLVMPtr{Int,1}, q::Core.LLVMPtr{Int,1})
-        x = Core.Intrinsics.atomic_pointerref(reinterpret(Ptr{Int}, p), :unordered)
-        Core.Intrinsics.atomic_pointerset(reinterpret(Ptr{Int}, q), x, :unordered)
+@testset "LLVM atomics" begin
+    # atomics in user code (e.g. UnsafeAtomics' `load`/`store!`) must reach the back-end,
+    # while Julia's `unordered` heap-reference accesses, which AIR cannot express when they
+    # are of pointers or outside device and threadgroup memory, become plain ones
+    function kernel(p::Core.LLVMPtr{Int32,1}, q::Core.LLVMPtr{Int32,1})
+        r = reinterpret(Ptr{Ptr{Int32}}, p)
+        y = Core.Intrinsics.atomic_pointerref(r, :unordered)
+        Core.Intrinsics.atomic_pointerset(reinterpret(Ptr{Ptr{Int32}}, q), y, :unordered)
+        x = Core.Intrinsics.atomic_pointerref(reinterpret(Ptr{Int32}, p), :acquire)
+        Core.Intrinsics.atomic_pointerset(reinterpret(Ptr{Int32}, q), x, :release)
         return
     end
-    source = methodinstance(typeof(kernel), Tuple{Core.LLVMPtr{Int,1}, Core.LLVMPtr{Int,1}},
-                            Base.get_world_counter())
-    target = MetalCompilerTarget(; macos=v"12.2", metal=v"3.0", air=v"3.0")
-    config = CompilerConfig(target, Metal.CompilerParams(); kernel=true)
-    job = CompilerJob(source, config)
-
-    # precondition: the accesses survive optimization as unordered atomics
-    ir = sprint(io->GPUCompiler.code_llvm(io, job; dump_module=true))
-    @test occursin(r"load atomic .* unordered", ir)
-    @test occursin(r"store atomic .* unordered", ir)
 
-    air = sprint(io->GPUCompiler.code_native(io, job; dump_module=true))
-    @test !occursin(r"(load|store) atomic", air)
-    @test occursin(r"load i64", air)
-    @test occursin(r"store i64", air)
-
-    # likewise for the `release` store of a pointer that codegen emits for an allocated
-    # object's type tag. Only Julia 1.12+ lowers `Ptr` values to LLVM pointers, which is
-    # what makes this the case Apple's back-end cannot legalize.
-    @static if VERSION >= v"1.12"
-    function tagged(p::Core.LLVMPtr{Int,1}, q::Core.LLVMPtr{Int,1})
-        x = Core.Intrinsics.atomic_pointerref(reinterpret(Ptr{Ptr{Int}}, p), :acquire)
-        Core.Intrinsics.atomic_pointerset(reinterpret(Ptr{Ptr{Int}}, q), x, :release)
-        return
-    end
-    source = methodinstance(typeof(tagged), Tuple{Core.LLVMPtr{Int,1}, Core.LLVMPtr{Int,1}},
-                            Base.get_world_counter())
-    job = CompilerJob(source, config)
-
-    ir = sprint(io->GPUCompiler.code_llvm(io, job; dump_module=true))
-    @test occursin(r"load atomic ptr.* acquire", ir)
-    @test occursin(r"store atomic ptr.* release", ir)
-
-    air = sprint(io->GPUCompiler.code_native(io, job; dump_module=true))
-    @test !occursin(r"(load|store) atomic", air)
+    @test @filecheck begin
+        @check_not "unordered"
+        @check "load atomic i32"
+        @check_same "acquire"
+        @check "store atomic i32"
+        @check_same "release"
+        @check_not "unordered"
+        Metal.code_native(kernel, Tuple{Core.LLVMPtr{Int32,1}, Core.LLVMPtr{Int32,1}};
+                          kernel=true, dump_module=true)
     end
 end
 
diff --git a/test/spirv.jl b/test/spirv.jl
index 28d6a997..5c89775a 100644
--- a/test/spirv.jl
+++ b/test/spirv.jl
@@ -417,27 +417,29 @@ end
     end
 end
 
-
-@testset "atomic demotion" begin
-    # Julia's `unordered` heap-reference accesses and `release` type-tag stores cannot be
-    # expressed in SPIR-V when they involve pointers (OpAtomicLoad/OpAtomicStore take scalars
-    # only): the translator would emit an invalid pointer-typed atomic. They carry no meaning
-    # without a device GC, so `demote_atomics!` turns them into plain accesses.
+@testset "LLVM atomics" begin
+    # atomics in user code (e.g. UnsafeAtomics' `load`/`store!`, Atomix' `get`/`set!`) must
+    # reach the back-end, while Julia's `unordered` heap-reference accesses, which SPIR-V
+    # cannot express when they are of pointers, become plain ones
     mod = @eval module $(gensym())
-        function kernel(p::Ptr{Ptr{Int}}, q::Ptr{Ptr{Int}})
-            x = Core.Intrinsics.atomic_pointerref(p, :unordered)
-            Core.Intrinsics.atomic_pointerset(q, x, :unordered)
-            y = Core.Intrinsics.atomic_pointerref(p, :acquire)
-            Core.Intrinsics.atomic_pointerset(q, y, :release)
+        function kernel(p::Ptr{Int32}, q::Ptr{Int32})
+            y = Core.Intrinsics.atomic_pointerref(reinterpret(Ptr{Ptr{Int32}}, p), :unordered)
+            Core.Intrinsics.atomic_pointerset(reinterpret(Ptr{Ptr{Int32}}, q), y, :unordered)
+            x = Core.Intrinsics.atomic_pointerref(p, :acquire)
+            Core.Intrinsics.atomic_pointerset(q, x, :release)
             return
         end
     end
-    tt = Tuple{Ptr{Ptr{Int}}, Ptr{Ptr{Int}}}
+    tt = Tuple{Ptr{Int32}, Ptr{Int32}}
 
     @test @filecheck begin
         @check_label "define spir_kernel void @_Z6kernel"
-        @check_not "load atomic"
-        @check_not "store atomic"
+        @check_not "unordered"
+        @check "load atomic i32"
+        @check_same "acquire"
+        @check "store atomic i32"
+        @check_same "release"
+        @check_not "unordered"
         @check "ret void"
         SPIRV.code_llvm(mod.kernel, tt; backend, kernel=true)
     end
@@ -446,8 +448,8 @@ end
     @test @filecheck begin
         @check "OpEntryPoint Kernel %[[KERNEL:[^ ]+]]"
         @check "%[[KERNEL]] = OpFunction %void None"
-        @check_not "OpAtomicLoad"
-        @check_not "OpAtomicStore"
+        @check "OpAtomicLoad"
+        @check "OpAtomicStore"
         SPIRV.code_native(mod.kernel, tt; backend, kernel=true)
     end
 end