Use all 32bits of shape_id_t on all platforms

Followup: ruby#13341 / [Feature #21353]

Even thought `shape_id_t` has been make 32bits, we were still limited
to use only the lower 16 bits because they had to fit alongside `attr_index_t`
inside a `uintptr_t` in inline caches.

By enlarging inline caches we can unlock the full 32bits on all
platforms, allowing to use these extra bits for tagging.

Author: byroot

shape.h

@@ -3,34 +3,22 @@
 
 #include "internal/gc.h"
 
-#if (SIZEOF_UINT64_T <= SIZEOF_VALUE)
-
 #define SIZEOF_SHAPE_T 4
-typedef uint32_t attr_index_t;
-typedef uint32_t shape_id_t;
-# define SHAPE_ID_NUM_BITS 32
-
-#else
-
-#define SIZEOF_SHAPE_T 2
 typedef uint16_t attr_index_t;
-typedef uint16_t shape_id_t;
-# define SHAPE_ID_NUM_BITS 16
-
-#endif
+typedef uint32_t shape_id_t;
+#define SHAPE_ID_NUM_BITS 32
 
 typedef uint32_t redblack_id_t;
 
 #define SHAPE_MAX_FIELDS (attr_index_t)(-1)
 
-# define SHAPE_FLAG_MASK (((VALUE)-1) >> SHAPE_ID_NUM_BITS)
-
-# define SHAPE_FLAG_SHIFT ((SIZEOF_VALUE * 8) - SHAPE_ID_NUM_BITS)
+#define SHAPE_FLAG_MASK (((VALUE)-1) >> SHAPE_ID_NUM_BITS)
+#define SHAPE_FLAG_SHIFT ((SIZEOF_VALUE * 8) - SHAPE_ID_NUM_BITS)
 
-# define SHAPE_MAX_VARIATIONS 8
+#define SHAPE_MAX_VARIATIONS 8
 
-# define INVALID_SHAPE_ID (((uintptr_t)1 << SHAPE_ID_NUM_BITS) - 1)
-#define ATTR_INDEX_NOT_SET (attr_index_t)-1
+#define INVALID_SHAPE_ID ((shape_id_t)-1)
+#define ATTR_INDEX_NOT_SET ((attr_index_t)-1)
 
 #define ROOT_SHAPE_ID               0x0
 #define SPECIAL_CONST_SHAPE_ID      0x1
@@ -44,13 +32,13 @@ typedef struct redblack_node redblack_node_t;
 struct rb_shape {
     VALUE edges; // id_table from ID (ivar) to next shape
     ID edge_name; // ID (ivar) for transition from parent to rb_shape
+    redblack_node_t *ancestor_index;
+    shape_id_t parent_id;
     attr_index_t next_field_index; // Fields are either ivars or internal properties like `object_id`
     attr_index_t capacity; // Total capacity of the object with this shape
     uint8_t type;
     uint8_t heap_index;
     uint8_t flags;
-    shape_id_t parent_id;
-    redblack_node_t *ancestor_index;
 };
 
 typedef struct rb_shape rb_shape_t;

vm_callinfo.h

@@ -289,7 +289,7 @@ struct rb_callcache {
 
     union {
         struct {
-          uintptr_t value; // Shape ID in upper bits, index in lower bits
+          uint64_t value; // Shape ID in upper bits, index in lower bits
         } attr;
         const enum method_missing_reason method_missing_reason; /* used by method_missing */
         VALUE v;
@@ -415,25 +415,43 @@ vm_cc_call(const struct rb_callcache *cc)
     return cc->call_;
 }
 
+static inline uint64_t
+vm_uint64_t_atomic_read(const uint64_t *value)
+{
+#if (SIZEOF_UINT64_T == SIZEOF_VALUE)
+    // Atomically read pointer sized value.
+    return *value;
+#elif defined(HAVE_GCC_ATOMIC_BUILTINS)
+    return __atomic_load_n(value, __ATOMIC_RELAXED);
+#elif defined(_WIN32)
+    uint64_t val = *value;
+    return InterlockedCompareExchange64(value, val, val);
+#elif defined(__sun) && defined(HAVE_ATOMIC_H) && (defined(_LP64) || defined(_I32LPx))
+    uint64_t val = *value;
+    return atomic_cas_64(value, val, val);
+#else
+    return *value;
+#endif
+}
+
 static inline void
-vm_unpack_shape_and_index(uintptr_t cache_value, shape_id_t *shape_id, attr_index_t *index)
+vm_unpack_shape_and_index(const uint64_t *cache_value_ptr, shape_id_t *shape_id, attr_index_t *index)
 {
+    const uint64_t cache_value = vm_uint64_t_atomic_read(cache_value_ptr);
     *shape_id = (shape_id_t)(cache_value >> SHAPE_FLAG_SHIFT);
     *index = (attr_index_t)(cache_value & SHAPE_FLAG_MASK) - 1;
 }
 
 static inline void
 vm_cc_atomic_shape_and_index(const struct rb_callcache *cc, shape_id_t *shape_id, attr_index_t *index)
 {
-    // Atomically read uintptr_t
-    vm_unpack_shape_and_index(cc->aux_.attr.value, shape_id, index);
+    vm_unpack_shape_and_index(&cc->aux_.attr.value, shape_id, index);
 }
 
 static inline void
 vm_ic_atomic_shape_and_index(const struct iseq_inline_iv_cache_entry *ic, shape_id_t *shape_id, attr_index_t *index)
 {
-    // Atomically read uintptr_t
-    vm_unpack_shape_and_index(ic->value, shape_id, index);
+    vm_unpack_shape_and_index(&ic->value, shape_id, index);
 }
 
 static inline unsigned int
@@ -496,16 +514,32 @@ vm_cc_ivar_p(const struct rb_callcache *cc)
     return (cc->flags & VM_CALLCACHE_IVAR) != 0;
 }
 
+static inline void
+vm_uint64_t_atomic_write(uint64_t *address, uint64_t value)
+{
+#if (SIZEOF_UINT64_T == SIZEOF_VALUE)
+    *address = value;
+#elif defined(HAVE_GCC_ATOMIC_BUILTINS)
+    __atomic_store_n(address, value, __ATOMIC_RELAXED);
+#elif defined(_WIN32)
+    InterlockedExchange64(address, value);
+#elif defined(__sun) && defined(HAVE_ATOMIC_H) && (defined(_LP64) || defined(_I32LPx))
+    atomic_swap_64(address, value);
+#else
+    *address = value;
+#endif
+}
+
 static inline void
 vm_ic_attr_index_set(const rb_iseq_t *iseq, const struct iseq_inline_iv_cache_entry *ic, attr_index_t index, shape_id_t dest_shape_id)
 {
-    *(uintptr_t *)&ic->value = vm_pack_shape_and_index(dest_shape_id, index);
+    vm_uint64_t_atomic_write((uint64_t *)&ic->value, vm_pack_shape_and_index(dest_shape_id, index));
 }
 
 static inline void
 vm_ic_attr_index_initialize(const struct iseq_inline_iv_cache_entry *ic, shape_id_t shape_id)
 {
-    *(uintptr_t *)&ic->value = vm_pack_shape_and_index(shape_id, ATTR_INDEX_NOT_SET);
+    vm_uint64_t_atomic_write((uint64_t *)&ic->value, vm_pack_shape_and_index(shape_id, ATTR_INDEX_NOT_SET));
 }
 
 static inline void

vm_core.h

@@ -288,7 +288,7 @@ struct iseq_inline_constant_cache {
 };
 
 struct iseq_inline_iv_cache_entry {
-    uintptr_t value; // attr_index in lower bits, dest_shape_id in upper bits
+    uint64_t value; // attr_index in lower bits, dest_shape_id in upper bits
     ID iv_set_name;
 };
 

vm_insnhelper.c

@@ -4774,7 +4774,7 @@ vm_call_method_each_type(rb_execution_context_t *ec, rb_control_frame_t *cfp, st
                     .call_ = cc->call_,
                     .aux_  = {
                         .attr = {
-                            .value = INVALID_SHAPE_ID << SHAPE_FLAG_SHIFT,
+                            .value = vm_pack_shape_and_index(INVALID_SHAPE_ID, ATTR_INDEX_NOT_SET),
                         }
                     },
             });

yjit/src/codegen.rs

@@ -2904,7 +2904,7 @@ fn gen_get_ivar(
 
     let ivar_index = unsafe {
         let shape_id = comptime_receiver.shape_id_of();
-        let mut ivar_index: u32 = 0;
+        let mut ivar_index: u16 = 0;
         if rb_shape_get_iv_index(shape_id, ivar_name, &mut ivar_index) {
             Some(ivar_index as usize)
         } else {
@@ -3106,7 +3106,7 @@ fn gen_set_ivar(
     let shape_too_complex = comptime_receiver.shape_too_complex();
     let ivar_index = if !shape_too_complex {
         let shape_id = comptime_receiver.shape_id_of();
-        let mut ivar_index: u32 = 0;
+        let mut ivar_index: u16 = 0;
         if unsafe { rb_shape_get_iv_index(shape_id, ivar_name, &mut ivar_index) } {
             Some(ivar_index as usize)
         } else {
@@ -3395,7 +3395,7 @@ fn gen_definedivar(
 
     let shape_id = comptime_receiver.shape_id_of();
     let ivar_exists = unsafe {
-        let mut ivar_index: u32 = 0;
+        let mut ivar_index: u16 = 0;
         rb_shape_get_iv_index(shape_id, ivar_name, &mut ivar_index)
     };