add array element mutex offset in print and gc

The layout and printing logic need to correctly offset and align the
inset fields to account for the per-element mutex of an atomic array
with large elements.

Fix #58993

Author: vtjnash

base/runtime_internals.jl

@@ -528,6 +528,10 @@ struct DataTypeLayout
     # fielddesc_type : 2;
     # arrayelem_isboxed : 1;
     # arrayelem_isunion : 1;
+    # arrayelem_isatomic : 1;
+    # arrayelem_islocked : 1;
+    # isbitsegal : 1;
+    # padding : 8;
 end
 
 """

src/codegen.cpp

@@ -3862,8 +3862,8 @@ static bool emit_f_opmemory(jl_codectx_t &ctx, jl_cgval_t *ret, jl_value_t *f,
     const jl_datatype_layout_t *layout = ((jl_datatype_t*)mty_dt)->layout;
     bool isboxed = layout->flags.arrayelem_isboxed;
     bool isunion = layout->flags.arrayelem_isunion;
-    bool isatomic = kind == (jl_value_t*)jl_atomic_sym;
-    bool needlock = isatomic && layout->size > MAX_ATOMIC_SIZE;
+    bool isatomic = layout->flags.arrayelem_isatomic || layout->flags.arrayelem_islocked;
+    bool needlock = layout->flags.arrayelem_islocked;
     size_t elsz = layout->size;
     size_t al = layout->alignment;
     if (al > JL_HEAP_ALIGNMENT)
@@ -4231,7 +4231,7 @@ static bool emit_builtin_call(jl_codectx_t &ctx, jl_cgval_t *ret, jl_value_t *f,
             size_t al = layout->alignment;
             if (al > JL_HEAP_ALIGNMENT)
                 al = JL_HEAP_ALIGNMENT;
-            bool needlock = isatomic && !isboxed && elsz > MAX_ATOMIC_SIZE;
+            bool needlock = layout->flags.arrayelem_islocked;
             AtomicOrdering Order = (needlock || order <= jl_memory_order_notatomic)
                                     ? (isboxed ? AtomicOrdering::Unordered : AtomicOrdering::NotAtomic)
                                     : get_llvm_atomic_order(order);
@@ -4317,7 +4317,8 @@ static bool emit_builtin_call(jl_codectx_t &ctx, jl_cgval_t *ret, jl_value_t *f,
                 *ret = jl_cgval_t(); // unreachable
                 return true;
             }
-            bool isatomic = kind == (jl_value_t*)jl_atomic_sym;
+            const jl_datatype_layout_t *layout = ((jl_datatype_t*)mty_dt)->layout;
+            bool isatomic = layout->flags.arrayelem_isatomic || layout->flags.arrayelem_islocked;
             if (!isatomic && order != jl_memory_order_notatomic && order != jl_memory_order_unspecified) {
                 emit_atomic_error(ctx, "memoryref_isassigned: non-atomic memory cannot be accessed atomically");
                 *ret = jl_cgval_t(); // unreachable
@@ -4333,13 +4334,12 @@ static bool emit_builtin_call(jl_codectx_t &ctx, jl_cgval_t *ret, jl_value_t *f,
             }
             jl_value_t *boundscheck = argv[3].constant;
             emit_typecheck(ctx, argv[3], (jl_value_t*)jl_bool_type, fname);
-            const jl_datatype_layout_t *layout = ((jl_datatype_t*)mty_dt)->layout;
             Value *mem = emit_memoryref_mem(ctx, ref, layout);
             Value *mlen = emit_genericmemorylen(ctx, mem, ref.typ);
             Value *oob = bounds_check_enabled(ctx, boundscheck) ? ctx.builder.CreateIsNull(mlen) : nullptr;
             bool isboxed = layout->flags.arrayelem_isboxed;
             if (isboxed || layout->first_ptr >= 0) {
-                bool needlock = isatomic && !isboxed && layout->size > MAX_ATOMIC_SIZE;
+                bool needlock = layout->flags.arrayelem_islocked;
                 AtomicOrdering Order = (needlock || order <= jl_memory_order_notatomic)
                                         ? (isboxed ? AtomicOrdering::Unordered : AtomicOrdering::NotAtomic)
                                         : get_llvm_atomic_order(order);
@@ -4359,13 +4359,12 @@ static bool emit_builtin_call(jl_codectx_t &ctx, jl_cgval_t *ret, jl_value_t *f,
                     ctx.builder.SetInsertPoint(passBB);
                 }
                 Value *elem = emit_memoryref_ptr(ctx, ref, layout);
-                if (needlock) {
+                if (!isboxed)
+                    elem = emit_ptrgep(ctx, elem, layout->first_ptr * sizeof(void*));
+                else if (needlock)
                     // n.b. no actual lock acquire needed, as the check itself only needs to load a single pointer and check for null
                     // elem += sizeof(lock);
                     elem = emit_ptrgep(ctx, elem, LLT_ALIGN(sizeof(jl_mutex_t), JL_SMALL_BYTE_ALIGNMENT));
-                }
-                if (!isboxed)
-                    elem = emit_ptrgep(ctx, elem, layout->first_ptr * sizeof(void*));
                 // emit this using the same type as BUILTIN(memoryrefget)
                 // so that LLVM may be able to load-load forward them and fold the result
                 auto tbaa = isboxed ? ctx.tbaa().tbaa_ptrarraybuf : ctx.tbaa().tbaa_arraybuf;

src/datatype.c

@@ -239,8 +239,10 @@ static jl_datatype_layout_t *jl_get_layout(uint32_t sz,
     flddesc->flags.haspadding = haspadding;
     flddesc->flags.isbitsegal = isbitsegal;
     flddesc->flags.fielddesc_type = fielddesc_type;
-    flddesc->flags.arrayelem_isboxed = arrayelem == 1;
-    flddesc->flags.arrayelem_isunion = arrayelem == 2;
+    flddesc->flags.arrayelem_isboxed = (arrayelem & 1) != 0;
+    flddesc->flags.arrayelem_isunion = (arrayelem & 2) != 0;
+    flddesc->flags.arrayelem_isatomic = (arrayelem & 4) != 0;
+    flddesc->flags.arrayelem_islocked = (arrayelem & 8) != 0;
     flddesc->flags.padding = 0;
     flddesc->npointers = npointers;
     flddesc->first_ptr = first_ptr;
@@ -537,6 +539,7 @@ void jl_get_genericmemory_layout(jl_datatype_t *st)
     uint32_t *pointers = &first_ptr;
     int needlock = 0;
 
+    const jl_datatype_layout_t *el_layout = NULL;
     if (isunboxed) {
         elsz = LLT_ALIGN(elsz, al);
         if (kind == (jl_value_t*)jl_atomic_sym) {
@@ -551,12 +554,12 @@ void jl_get_genericmemory_layout(jl_datatype_t *st)
         else {
             assert(jl_is_datatype(eltype));
             zi = ((jl_datatype_t*)eltype)->zeroinit;
-            const jl_datatype_layout_t *layout = ((jl_datatype_t*)eltype)->layout;
-            if (layout->first_ptr >= 0) {
-                first_ptr = layout->first_ptr;
-                npointers = layout->npointers;
-                if (layout->flags.fielddesc_type == 2) {
-                    pointers = (uint32_t*)jl_dt_layout_ptrs(layout);
+            el_layout = ((jl_datatype_t*)eltype)->layout;
+            if (el_layout->first_ptr >= 0) {
+                first_ptr = el_layout->first_ptr;
+                npointers = el_layout->npointers;
+                if (el_layout->flags.fielddesc_type == 2 && !needlock) {
+                    pointers = (uint32_t*)jl_dt_layout_ptrs(el_layout);
                 }
                 else {
                     pointers = (uint32_t*)alloca(npointers * sizeof(uint32_t));
@@ -568,10 +571,22 @@ void jl_get_genericmemory_layout(jl_datatype_t *st)
         }
         if (needlock) {
             assert(al <= JL_SMALL_BYTE_ALIGNMENT);
-            size_t offset = LLT_ALIGN(sizeof(jl_mutex_t), JL_SMALL_BYTE_ALIGNMENT);
-            elsz += offset;
+            size_t lock_offset = LLT_ALIGN(sizeof(jl_mutex_t), JL_SMALL_BYTE_ALIGNMENT);
+            elsz += lock_offset;
+            if (al < sizeof(void*)) {
+              al = sizeof(void*);
+              elsz = LLT_ALIGN(elsz, al);
+            }
             haspadding = 1;
             zi = 1;
+            // Adjust pointer offsets to account for the lock at the beginning
+            if (first_ptr >= 0) {
+                uint32_t lock_offset_words = lock_offset / sizeof(void*);
+                first_ptr += lock_offset_words;
+                for (int j = 0; j < npointers; j++) {
+                    pointers[j] += lock_offset_words;
+                }
+            }
         }
     }
     else {
@@ -580,13 +595,17 @@ void jl_get_genericmemory_layout(jl_datatype_t *st)
         zi = 1;
     }
 
-    int arrayelem;
+    // arrayelem is a bitfield: 1=isboxed, 2=isunion, 4=isatomic, 8=islocked
+    int arrayelem = 0;
     if (!isunboxed)
-        arrayelem = 1;
-    else if (isunion)
-        arrayelem = 2;
-    else
-        arrayelem = 0;
+        arrayelem |= 1;  // arrayelem_isboxed
+    if (isunion)
+        arrayelem |= 2;  // arrayelem_isunion
+    if (kind == (jl_value_t*)jl_atomic_sym) {
+        arrayelem |= 4;  // arrayelem_isatomic
+        if (needlock)
+            arrayelem |= 8;  // arrayelem_islocked
+    }
     assert(!st->layout);
     st->layout = jl_get_layout(elsz, nfields, npointers, al, haspadding, isbitsegal, arrayelem, NULL, pointers);
     st->zeroinit = zi;
@@ -647,17 +666,17 @@ void jl_compute_field_offsets(jl_datatype_t *st)
         // if we have no fields, we can trivially skip the rest
         if (st == jl_symbol_type || st == jl_string_type) {
             // opaque layout - heap-allocated blob
-            static const jl_datatype_layout_t opaque_byte_layout = {0, 0, 1, -1, 1, { .haspadding = 0, .fielddesc_type=0, .isbitsegal=1, .arrayelem_isboxed=0, .arrayelem_isunion=0 }};
+            static const jl_datatype_layout_t opaque_byte_layout = {0, 0, 1, -1, 1, { .isbitsegal=1 }};
             st->layout = &opaque_byte_layout;
             return;
         }
         else if (st == jl_simplevector_type || st == jl_module_type) {
-            static const jl_datatype_layout_t opaque_ptr_layout = {0, 0, 1, -1, sizeof(void*), { .haspadding = 0, .fielddesc_type=0, .isbitsegal=1, .arrayelem_isboxed=0, .arrayelem_isunion=0 }};
+            static const jl_datatype_layout_t opaque_ptr_layout = {0, 0, 1, -1, sizeof(void*), { .isbitsegal=1 }};
             st->layout = &opaque_ptr_layout;
             return;
         }
         else {
-            static const jl_datatype_layout_t singleton_layout = {0, 0, 0, -1, 1, { .haspadding = 0, .fielddesc_type=0, .isbitsegal=1, .arrayelem_isboxed=0, .arrayelem_isunion=0 }};
+            static const jl_datatype_layout_t singleton_layout = {0, 0, 0, -1, 1, { .isbitsegal=1 }};
             st->layout = &singleton_layout;
         }
     }
@@ -1001,6 +1020,8 @@ JL_DLLEXPORT jl_datatype_t * jl_new_foreign_type(jl_sym_t *name,
     layout->flags.padding = 0;
     layout->flags.arrayelem_isboxed = 0;
     layout->flags.arrayelem_isunion = 0;
+    layout->flags.arrayelem_isatomic = 0;
+    layout->flags.arrayelem_islocked = 0;
     jl_fielddescdyn_t * desc =
       (jl_fielddescdyn_t *) ((char *)layout + sizeof(*layout));
     desc->markfunc = markfunc;

src/genericmemory.c

@@ -262,8 +262,8 @@ JL_DLLEXPORT void jl_genericmemory_copyto(jl_genericmemory_t *dest, char* destda
 
 JL_DLLEXPORT jl_value_t *jl_genericmemoryref(jl_genericmemory_t *mem, size_t i)
 {
-    int isatomic = (jl_tparam0(jl_typetagof(mem)) == (jl_value_t*)jl_atomic_sym);
     const jl_datatype_layout_t *layout = ((jl_datatype_t*)jl_typetagof(mem))->layout;
+    int isatomic = layout->flags.arrayelem_isatomic || layout->flags.arrayelem_islocked;
     jl_genericmemoryref_t m;
     m.mem = mem;
     m.ptr_or_offset = (layout->flags.arrayelem_isunion || layout->size == 0) ? (void*)i : (void*)((char*)mem->ptr + layout->size * i);
@@ -342,8 +342,8 @@ static jl_value_t *jl_ptrmemrefget(jl_genericmemoryref_t m JL_PROPAGATES_ROOT, i
 
 JL_DLLEXPORT jl_value_t *jl_memoryrefget(jl_genericmemoryref_t m, int isatomic)
 {
-    assert(isatomic == (jl_tparam0(jl_typetagof(m.mem)) == (jl_value_t*)jl_atomic_sym));
     const jl_datatype_layout_t *layout = ((jl_datatype_t*)jl_typetagof(m.mem))->layout;
+    assert(isatomic == (layout->flags.arrayelem_isatomic || layout->flags.arrayelem_islocked));
     if (layout->flags.arrayelem_isboxed)
         return jl_ptrmemrefget(m, isatomic);
     jl_value_t *eltype = jl_tparam1(jl_typetagof(m.mem));
@@ -365,7 +365,7 @@ JL_DLLEXPORT jl_value_t *jl_memoryrefget(jl_genericmemoryref_t m, int isatomic)
     assert(data - (char*)m.mem->ptr < layout->size * m.mem->length);
     jl_value_t *r;
     size_t fsz = jl_datatype_size(eltype);
-    int needlock = isatomic && fsz > MAX_ATOMIC_SIZE;
+    int needlock = layout->flags.arrayelem_islocked;
     if (isatomic && !needlock) {
         r = jl_atomic_new_bits(eltype, data);
     }
@@ -393,7 +393,7 @@ static int _jl_memoryref_isassigned(jl_genericmemoryref_t m, int isatomic)
     if (layout->flags.arrayelem_isboxed) {
     }
     else if (layout->first_ptr >= 0) {
-        int needlock = isatomic && layout->size > MAX_ATOMIC_SIZE;
+        int needlock = layout->flags.arrayelem_islocked;
         if (needlock)
             elem = elem + LLT_ALIGN(sizeof(jl_mutex_t), JL_SMALL_BYTE_ALIGNMENT) / sizeof(jl_value_t*);
         elem = &elem[layout->first_ptr];
@@ -411,15 +411,15 @@ JL_DLLEXPORT jl_value_t *jl_memoryref_isassigned(jl_genericmemoryref_t m, int is
 
 JL_DLLEXPORT void jl_memoryrefset(jl_genericmemoryref_t m JL_ROOTING_ARGUMENT, jl_value_t *rhs JL_ROOTED_ARGUMENT JL_MAYBE_UNROOTED, int isatomic)
 {
-    assert(isatomic == (jl_tparam0(jl_typetagof(m.mem)) == (jl_value_t*)jl_atomic_sym));
+    const jl_datatype_layout_t *layout = ((jl_datatype_t*)jl_typetagof(m.mem))->layout;
+    assert(isatomic == (layout->flags.arrayelem_isatomic || layout->flags.arrayelem_islocked));
     jl_value_t *eltype = jl_tparam1(jl_typetagof(m.mem));
     if (eltype != (jl_value_t*)jl_any_type && !jl_typeis(rhs, eltype)) {
         JL_GC_PUSH1(&rhs);
         if (!jl_isa(rhs, eltype))
             jl_type_error("memoryrefset!", eltype, rhs);
         JL_GC_POP();
     }
-    const jl_datatype_layout_t *layout = ((jl_datatype_t*)jl_typetagof(m.mem))->layout;
     if (layout->flags.arrayelem_isboxed) {
         assert((char*)m.ptr_or_offset - (char*)m.mem->ptr < sizeof(jl_value_t*) * m.mem->length);
         if (isatomic)
@@ -449,7 +449,7 @@ JL_DLLEXPORT void jl_memoryrefset(jl_genericmemoryref_t m JL_ROOTING_ARGUMENT, j
     }
     if (layout->size != 0) {
         assert(data - (char*)m.mem->ptr < layout->size * m.mem->length);
-        int needlock = isatomic && layout->size > MAX_ATOMIC_SIZE;
+        int needlock = layout->flags.arrayelem_islocked;
         size_t fsz = jl_datatype_size((jl_datatype_t*)jl_typeof(rhs)); // need to shrink-wrap the final copy
         if (isatomic && !needlock) {
             jl_atomic_store_bits(data, rhs, fsz);

src/julia.h

@@ -574,10 +574,12 @@ typedef struct {
         // metadata bit only for GenericMemory eltype layout
         uint16_t arrayelem_isboxed : 1;
         uint16_t arrayelem_isunion : 1;
+        uint16_t arrayelem_isatomic : 1;
+        uint16_t arrayelem_islocked : 1;
         // If set, this type's egality can be determined entirely by comparing
         // the non-padding bits of this datatype.
         uint16_t isbitsegal : 1;
-        uint16_t padding : 10;
+        uint16_t padding : 8;
     } flags;
     // union {
     //     jl_fielddesc8_t field8[nfields];
@@ -1667,6 +1669,8 @@ static inline int jl_field_isconst(jl_datatype_t *st, int i) JL_NOTSAFEPOINT
 #define jl_is_addrspacecore(v) jl_typetagis(v,jl_addrspacecore_type)
 #define jl_is_abioverride(v) jl_typetagis(v,jl_abioverride_type)
 #define jl_genericmemory_isbitsunion(a) (((jl_datatype_t*)jl_typetagof(a))->layout->flags.arrayelem_isunion)
+#define jl_genericmemory_isatomic(a) (((jl_datatype_t*)jl_typetagof(a))->layout->flags.arrayelem_isatomic)
+#define jl_genericmemory_islocked(a) (((jl_datatype_t*)jl_typetagof(a))->layout->flags.arrayelem_islocked)
 #define jl_is_array_any(v)    jl_typetagis(v,jl_array_any_type)
 
 JL_DLLEXPORT int jl_subtype(jl_value_t *a, jl_value_t *b);

src/rtutils.c

@@ -1290,10 +1290,20 @@ static size_t jl_static_show_x_(JL_STREAM *out, jl_value_t *v, jl_datatype_t *vt
             for (size_t j = 0; j < tlen; j++) {
                 if (layout->flags.arrayelem_isboxed) {
                     jl_value_t **ptr = ((jl_value_t**)m->ptr) + j;
+                    if (layout->flags.arrayelem_islocked) {
+                        // Skip the lock at the beginning for locked arrays
+                        size_t lock_size = sizeof(jl_mutex_t);
+                        ptr = (jl_value_t**)((char*)ptr + lock_size);
+                    }
                     n += jl_static_show_x(out, *ptr, depth, ctx);
                 }
                 else {
                     char *ptr = ((char*)m->ptr) + j * layout->size;
+                    if (layout->flags.arrayelem_islocked) {
+                        // Skip the lock at the beginning for locked arrays
+                        size_t lock_size = sizeof(jl_mutex_t);
+                        ptr += lock_size;
+                    }
                     n += jl_static_show_x_(out, (jl_value_t*)ptr,
                             (jl_datatype_t*)(typetagdata ? jl_nth_union_component(el_type, typetagdata[j]) : el_type),
                             depth, ctx);