codegen: mark svec fields GC as release stores

Without a release store, it seems LLVM considers it a data race to have
read the initial state on another thread. Marking this as a release
store seems sufficient to prevent that optimization. It is also more
consistent with how we initialize and write to most other structs,
particularly since #55767.

Fixes #59547 (more)

Author: vtjnash

src/cgutils.cpp

@@ -2228,6 +2228,9 @@ static jl_cgval_t typed_load(jl_codectx_t &ctx, Value *ptr, Value *idx_0based, j
     }
     Value *instr = nullptr;
     if (!isboxed && jl_is_genericmemoryref_type(jltype)) {
+        //We don't specify the stronger expected memory ordering here because of fears it may interfere with vectorization and other optimizations
+        //if (Order == AtomicOrdering::NotAtomic)
+        //    Order = AtomicOrdering::Monotonic;
         // load these FCA as individual fields, so LLVM does not need to split them later
         Value *fld0 = ctx.builder.CreateStructGEP(elty, ptr, 0);
         LoadInst *load0 = ctx.builder.CreateAlignedLoad(elty->getStructElementType(0), fld0, Align(alignment), false);
@@ -2403,11 +2406,26 @@ static jl_cgval_t typed_store(jl_codectx_t &ctx,
             instr = load;
         }
         if (r) {
-            StoreInst *store = ctx.builder.CreateAlignedStore(r, ptr, Align(alignment));
-            store->setOrdering(Order == AtomicOrdering::NotAtomic && isboxed ? AtomicOrdering::Release : Order);
             jl_aliasinfo_t ai = jl_aliasinfo_t::fromTBAA(ctx, tbaa);
             ai.noalias = MDNode::concatenate(aliasscope, ai.noalias);
-            ai.decorateInst(store);
+            if (false && !isboxed && Order == AtomicOrdering::NotAtomic && jl_is_genericmemoryref_type(jltype)) {
+                // if enabled, store these FCA as individual fields, so LLVM does not need to split them later and they can use release ordering
+                assert(r->getType() == ctx.types().T_jlgenericmemory);
+                Value *f1 = ctx.builder.CreateExtractValue(r, 0);
+                Value *f2 = ctx.builder.CreateExtractValue(r, 1);
+                static_assert(offsetof(jl_genericmemoryref_t, ptr_or_offset) == 0, "wrong field order");
+                StoreInst *store = ctx.builder.CreateAlignedStore(f1, ctx.builder.CreateStructGEP(ctx.types().T_jlgenericmemory, ptr, 0), Align(alignment));
+                store->setOrdering(AtomicOrdering::Release);
+                ai.decorateInst(store);
+                store = ctx.builder.CreateAlignedStore(f2, ctx.builder.CreateStructGEP(ctx.types().T_jlgenericmemory, ptr, 1), Align(alignment));
+                store->setOrdering(AtomicOrdering::Release);
+                ai.decorateInst(store);
+            }
+            else {
+                StoreInst *store = ctx.builder.CreateAlignedStore(r, ptr, Align(alignment));
+                store->setOrdering(Order == AtomicOrdering::NotAtomic && isboxed ? AtomicOrdering::Release : Order);
+                ai.decorateInst(store);
+            }
         }
         else {
             assert(Order == AtomicOrdering::NotAtomic && !isboxed && rhs.typ == jltype);
@@ -4435,10 +4453,11 @@ static jl_cgval_t emit_new_struct(jl_codectx_t &ctx, jl_value_t *ty, size_t narg
         for (size_t i = nargs; i < nf; i++) {
             if (!jl_field_isptr(sty, i) && jl_is_uniontype(jl_field_type(sty, i))) {
                 jl_aliasinfo_t ai = jl_aliasinfo_t::fromTBAA(ctx, strctinfo.tbaa);
-                ai.decorateInst(ctx.builder.CreateAlignedStore(
+                auto *store = ctx.builder.CreateAlignedStore(
                         ConstantInt::get(getInt8Ty(ctx.builder.getContext()), 0),
                         emit_ptrgep(ctx, strct, jl_field_offset(sty, i) + jl_field_size(sty, i) - 1),
-                        Align(1)));
+                        Align(1));
+                ai.decorateInst(store);
             }
         }
         // TODO: verify that nargs <= nf (currently handled by front-end)

src/codegen.cpp

@@ -4123,7 +4123,8 @@ static bool emit_builtin_call(jl_codectx_t &ctx, jl_cgval_t *ret, jl_value_t *f,
         for (size_t i = 0; i < nargs; i++) {
             Value *elem = boxed(ctx, argv[i + 1]);
             Value *elem_ptr = emit_ptrgep(ctx, svec_derived, ctx.types().sizeof_ptr * (i + 1));
-            ctx.builder.CreateAlignedStore(elem, elem_ptr, Align(ctx.types().sizeof_ptr));
+            auto *store = ctx.builder.CreateAlignedStore(elem, elem_ptr, Align(ctx.types().sizeof_ptr));
+            store->setOrdering(AtomicOrdering::Release);
             emit_write_barrier(ctx, svec, elem);
         }
         *ret = mark_julia_type(ctx, svec, true, jl_simplevector_type);