codegen: mark write barrier field load as volatile (#59559)

This is as an alternative to changing all allocation functions to
mutating all memory and returning an aliasing pointer. This operates
usually late in the pipeline, so either should not make much difference,
but I think it should suffice to mark this volatile to prevent any
de-duplication of this load, and this should also be most conservative
for GC but more liberal for other optimizations.

Fixes #59547

Produced with dubious help by Claude.

Author: vtjnash

Compiler/test/codegen.jl

@@ -134,14 +134,14 @@ if !is_debug_build && opt_level > 0
     # Array
     test_loads_no_call(strip_debug_calls(get_llvm(sizeof, Tuple{Vector{Int}})), [Iptr])
     # As long as the eltype is known we don't need to load the elsize, but do need to check isvector
-    @test_skip test_loads_no_call(strip_debug_calls(get_llvm(sizeof, Tuple{Array{Any}})), ["atomic $Iptr", "ptr", "ptr", Iptr, Iptr, "ptr",  Iptr])
+    @test_skip test_loads_no_call(strip_debug_calls(get_llvm(sizeof, Tuple{Array{Any}})), ["atomic volatile $Iptr", "ptr", "ptr", Iptr, Iptr, "ptr",  Iptr])
     # Memory
     test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory{Int}})), [Iptr])
     # As long as the eltype is known we don't need to load the elsize
     test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory{Any}})), [Iptr])
     # Check that we load the elsize and isunion from the typeof layout
-    test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory})), [Iptr, "atomic $Iptr", "ptr", "i32", "i16"])
-    test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory})), [Iptr, "atomic $Iptr", "ptr", "i32", "i16"])
+    test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory})), [Iptr, "atomic volatile $Iptr", "ptr", "i32", "i16"])
+    test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Memory})), [Iptr, "atomic volatile $Iptr", "ptr", "i32", "i16"])
     # Primitive Type size should be folded to a constant
     test_loads_no_call(strip_debug_calls(get_llvm(core_sizeof, Tuple{Ptr})), String[])
 

src/llvm-late-gc-lowering.cpp

@@ -1913,6 +1913,9 @@ Value *LateLowerGCFrame::EmitLoadTag(IRBuilder<> &builder, Type *T_size, Value *
     auto &M = *builder.GetInsertBlock()->getModule();
     LoadInst *load = builder.CreateAlignedLoad(T_size, addr, M.getDataLayout().getPointerABIAlignment(0), V->getName() + ".tag");
     load->setOrdering(AtomicOrdering::Unordered);
+    // Mark as volatile to prevent optimizers from treating GC tag loads as constants
+    // since GC mark bits can change during runtime (issue #59547)
+    load->setVolatile(true);
     load->setMetadata(LLVMContext::MD_tbaa, tbaa_tag);
     MDBuilder MDB(load->getContext());
     auto *NullInt = ConstantInt::get(T_size, 0);

test/llvmpasses/gc-writebarrier-volatile.ll

@@ -0,0 +1,41 @@
+; This file is a part of Julia. License is MIT: https://julialang.org/license
+
+; RUN: opt --load-pass-plugin=libjulia-codegen%shlibext -passes='function(LateLowerGCFrame,FinalLowerGC,gvn)' -S %s | FileCheck %s
+
+; Test for issue #59547: Ensure write barrier GC tag loads are volatile
+; This test verifies that the LateLowerGCFrame pass marks GC tag loads as volatile
+; to prevent GVN from incorrectly constant-folding them, which would eliminate
+; necessary write barrier checks.
+
+@tag = external addrspace(10) global {}, align 16
+
+declare void @julia.write_barrier({} addrspace(10)*, {} addrspace(10)*)
+declare {}*** @julia.get_pgcstack()
+declare {} addrspace(10)* @julia.gc_alloc_obj({}**, i64, {} addrspace(10)*)
+
+; Test that write barrier expansion produces volatile GC tag loads
+; CHECK-LABEL: @test_writebarrier_volatile_tags
+define {} addrspace(10)* @test_writebarrier_volatile_tags() {
+top:
+  %pgcstack = call {}*** @julia.get_pgcstack()
+  %current_task = bitcast {}*** %pgcstack to {}**
+  %parent = call {} addrspace(10)* @julia.gc_alloc_obj({}** %current_task, i64 8, {} addrspace(10)* @tag)
+  %child = call {} addrspace(10)* @julia.gc_alloc_obj({}** %current_task, i64 8, {} addrspace(10)* @tag)
+  call void @julia.write_barrier({} addrspace(10)* %parent, {} addrspace(10)* %child)
+  ret {} addrspace(10)* %parent
+
+; The critical test: GC tag loads must be volatile to prevent constant folding
+; CHECK: load atomic volatile i64, ptr {{.*}} unordered, align 8, {{.*}}!tbaa
+; CHECK: and i64 {{.*}}, 3
+; CHECK: icmp eq i64 {{.*}}, 3
+; CHECK: br i1 {{.*}}, label %may_trigger_wb, label
+
+; CHECK: may_trigger_wb:
+; CHECK: load atomic volatile i64, ptr {{.*}} unordered, align 8, {{.*}}!tbaa
+; CHECK: and i64 {{.*}}, 1
+; CHECK: icmp eq i64 {{.*}}, 0
+; CHECK: br i1 {{.*}}, label %trigger_wb, label
+
+; CHECK: trigger_wb:
+; CHECK: call void @ijl_gc_queue_root(ptr {{.*}})
+}