asgardius/virtualx-engine
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge pull request #87814 from bruvzg/memalign

Browse source 
[Core] Improve `CowData` and `Memory` metadata alignment.
This commit is contained in:
Rémi Verschelde 2024-02-05 17:10:52 +01:00 committed by GitHub
commit f8f2c8c85a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 99 additions and 52 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

24
core/os/memory.cpp
View file

@ -72,23 +72,23 @@ void *Memory::alloc_static(size_t p_bytes, bool p_pad_align) {
bool prepad = p_pad_align; bool prepad = p_pad_align;
#endif #endif
void *mem = malloc(p_bytes + (prepad ? PAD_ALIGN : 0)); void *mem = malloc(p_bytes + (prepad ? DATA_OFFSET : 0));
ERR_FAIL_NULL_V(mem, nullptr); ERR_FAIL_NULL_V(mem, nullptr);
alloc_count.increment(); alloc_count.increment();
if (prepad) { if (prepad) {
uint64_t *s = (uint64_t *)mem;
*s = p_bytes;
uint8_t *s8 = (uint8_t *)mem; uint8_t *s8 = (uint8_t *)mem;
uint64_t *s = (uint64_t *)(s8 + SIZE_OFFSET);
*s = p_bytes;
#ifdef DEBUG_ENABLED #ifdef DEBUG_ENABLED
uint64_t new_mem_usage = mem_usage.add(p_bytes); uint64_t new_mem_usage = mem_usage.add(p_bytes);
max_usage.exchange_if_greater(new_mem_usage); max_usage.exchange_if_greater(new_mem_usage);
#endif #endif
return s8 + PAD_ALIGN; return s8 + DATA_OFFSET;
} else { } else {
return mem; return mem;
} }
@ -108,8 +108,8 @@ void *Memory::realloc_static(void *p_memory, size_t p_bytes, bool p_pad_align) {
#endif #endif
if (prepad) { if (prepad) {
mem -= PAD_ALIGN; mem -= DATA_OFFSET;
uint64_t *s = (uint64_t *)mem; uint64_t *s = (uint64_t *)(mem + SIZE_OFFSET);
#ifdef DEBUG_ENABLED #ifdef DEBUG_ENABLED
if (p_bytes > *s) { if (p_bytes > *s) {
@ -126,14 +126,14 @@ void *Memory::realloc_static(void *p_memory, size_t p_bytes, bool p_pad_align) {
} else { } else {
*s = p_bytes; *s = p_bytes;
mem = (uint8_t *)realloc(mem, p_bytes + PAD_ALIGN); mem = (uint8_t *)realloc(mem, p_bytes + DATA_OFFSET);
ERR_FAIL_NULL_V(mem, nullptr); ERR_FAIL_NULL_V(mem, nullptr);
s = (uint64_t *)mem; s = (uint64_t *)(mem + SIZE_OFFSET);
*s = p_bytes; *s = p_bytes;
return mem + PAD_ALIGN; return mem + DATA_OFFSET;
} }
} else { } else {
mem = (uint8_t *)realloc(mem, p_bytes); mem = (uint8_t *)realloc(mem, p_bytes);
@ -158,10 +158,10 @@ void Memory::free_static(void *p_ptr, bool p_pad_align) {
alloc_count.decrement(); alloc_count.decrement();
if (prepad) { if (prepad) {
mem -= PAD_ALIGN; mem -= DATA_OFFSET;
#ifdef DEBUG_ENABLED #ifdef DEBUG_ENABLED
uint64_t *s = (uint64_t *)mem; uint64_t *s = (uint64_t *)(mem + SIZE_OFFSET);
mem_usage.sub(*s); mem_usage.sub(*s);
#endif #endif

35
core/os/memory.h
View file

@ -38,10 +38,6 @@
#include <new> #include <new>
#include <type_traits> #include <type_traits>
#ifndef PAD_ALIGN
#define PAD_ALIGN 16 //must always be greater than this at much
#endif
class Memory { class Memory {
#ifdef DEBUG_ENABLED #ifdef DEBUG_ENABLED
static SafeNumeric<uint64_t> mem_usage; static SafeNumeric<uint64_t> mem_usage;
@ -51,6 +47,17 @@ class Memory {
static SafeNumeric<uint64_t> alloc_count; static SafeNumeric<uint64_t> alloc_count;
public: public:
// Alignment: ↓ max_align_t ↓ uint64_t ↓ max_align_t
// ┌─────────────────┬──┬────────────────┬──┬───────────...
// │ uint64_t │░░│ uint64_t │░░│ T[]
// │ alloc size │░░│ element count │░░│ data
// └─────────────────┴──┴────────────────┴──┴───────────...
// Offset: ↑ SIZE_OFFSET ↑ ELEMENT_OFFSET ↑ DATA_OFFSET
static constexpr size_t SIZE_OFFSET = 0;
static constexpr size_t ELEMENT_OFFSET = ((SIZE_OFFSET + sizeof(uint64_t)) % alignof(uint64_t) == 0) ? (SIZE_OFFSET + sizeof(uint64_t)) : ((SIZE_OFFSET + sizeof(uint64_t)) + alignof(uint64_t) - ((SIZE_OFFSET + sizeof(uint64_t)) % alignof(uint64_t)));
static constexpr size_t DATA_OFFSET = ((ELEMENT_OFFSET + sizeof(uint64_t)) % alignof(max_align_t) == 0) ? (ELEMENT_OFFSET + sizeof(uint64_t)) : ((ELEMENT_OFFSET + sizeof(uint64_t)) + alignof(max_align_t) - ((ELEMENT_OFFSET + sizeof(uint64_t)) % alignof(max_align_t)));
static void *alloc_static(size_t p_bytes, bool p_pad_align = false); static void *alloc_static(size_t p_bytes, bool p_pad_align = false);
static void *realloc_static(void *p_memory, size_t p_bytes, bool p_pad_align = false); static void *realloc_static(void *p_memory, size_t p_bytes, bool p_pad_align = false);
static void free_static(void *p_ptr, bool p_pad_align = false); static void free_static(void *p_ptr, bool p_pad_align = false);
@ -133,6 +140,10 @@ void memdelete_allocator(T *p_class) {
#define memnew_arr(m_class, m_count) memnew_arr_template<m_class>(m_count) #define memnew_arr(m_class, m_count) memnew_arr_template<m_class>(m_count)
_FORCE_INLINE_ uint64_t *_get_element_count_ptr(uint8_t *p_ptr) {
return (uint64_t *)(p_ptr - Memory::DATA_OFFSET + Memory::ELEMENT_OFFSET);
}
template <typename T> template <typename T>
T *memnew_arr_template(size_t p_elements) { T *memnew_arr_template(size_t p_elements) {
if (p_elements == 0) { if (p_elements == 0) {
@ -142,10 +153,12 @@ T *memnew_arr_template(size_t p_elements) {
same strategy used by std::vector, and the Vector class, so it should be safe.*/ same strategy used by std::vector, and the Vector class, so it should be safe.*/
size_t len = sizeof(T) * p_elements; size_t len = sizeof(T) * p_elements;
uint64_t *mem = (uint64_t *)Memory::alloc_static(len, true); uint8_t *mem = (uint8_t *)Memory::alloc_static(len, true);
T *failptr = nullptr; //get rid of a warning T *failptr = nullptr; //get rid of a warning
ERR_FAIL_NULL_V(mem, failptr); ERR_FAIL_NULL_V(mem, failptr);
*(mem - 1) = p_elements;
uint64_t *_elem_count_ptr = _get_element_count_ptr(mem);
*(_elem_count_ptr) = p_elements;
if constexpr (!std::is_trivially_constructible_v<T>) { if constexpr (!std::is_trivially_constructible_v<T>) {
T *elems = (T *)mem; T *elems = (T *)mem;
@ -166,16 +179,18 @@ T *memnew_arr_template(size_t p_elements) {
template <typename T> template <typename T>
size_t memarr_len(const T *p_class) { size_t memarr_len(const T *p_class) {
uint64_t *ptr = (uint64_t *)p_class; uint8_t *ptr = (uint8_t *)p_class;
return *(ptr - 1); uint64_t *_elem_count_ptr = _get_element_count_ptr(ptr);
return *(_elem_count_ptr);
} }
template <typename T> template <typename T>
void memdelete_arr(T *p_class) { void memdelete_arr(T *p_class) {
uint64_t *ptr = (uint64_t *)p_class; uint8_t *ptr = (uint8_t *)p_class;
if constexpr (!std::is_trivially_destructible_v<T>) { if constexpr (!std::is_trivially_destructible_v<T>) {
uint64_t elem_count = *(ptr - 1); uint64_t *_elem_count_ptr = _get_element_count_ptr(ptr);
uint64_t elem_count = *(_elem_count_ptr);
for (uint64_t i = 0; i < elem_count; i++) { for (uint64_t i = 0; i < elem_count; i++) {
p_class[i].~T(); p_class[i].~T();

92
core/templates/cowdata.h
View file

@ -46,7 +46,7 @@ class CharString;
template <class T, class V> template <class T, class V>
class VMap; class VMap;
SAFE_NUMERIC_TYPE_PUN_GUARANTEES(uint64_t) static_assert(std::is_trivially_destructible_v<std::atomic<uint64_t>>);
// Silence a false positive warning (see GH-52119). // Silence a false positive warning (see GH-52119).
#if defined(__GNUC__) && !defined(__clang__) #if defined(__GNUC__) && !defined(__clang__)
@ -89,18 +89,39 @@ private:
return ++x; return ++x;
} }
static constexpr USize ALLOC_PAD = sizeof(USize) * 2; // For size and atomic refcount. // Alignment: ↓ max_align_t ↓ USize ↓ max_align_t
// ┌────────────────────┬──┬─────────────┬──┬───────────...
// │ SafeNumeric<USize> │░░│ USize │░░│ T[]
// │ ref. count │░░│ data size │░░│ data
// └────────────────────┴──┴─────────────┴──┴───────────...
// Offset: ↑ REF_COUNT_OFFSET ↑ SIZE_OFFSET ↑ DATA_OFFSET
static constexpr size_t REF_COUNT_OFFSET = 0;
static constexpr size_t SIZE_OFFSET = ((REF_COUNT_OFFSET + sizeof(SafeNumeric<USize>)) % alignof(USize) == 0) ? (REF_COUNT_OFFSET + sizeof(SafeNumeric<USize>)) : ((REF_COUNT_OFFSET + sizeof(SafeNumeric<USize>)) + alignof(USize) - ((REF_COUNT_OFFSET + sizeof(SafeNumeric<USize>)) % alignof(USize)));
static constexpr size_t DATA_OFFSET = ((SIZE_OFFSET + sizeof(USize)) % alignof(max_align_t) == 0) ? (SIZE_OFFSET + sizeof(USize)) : ((SIZE_OFFSET + sizeof(USize)) + alignof(max_align_t) - ((SIZE_OFFSET + sizeof(USize)) % alignof(max_align_t)));
mutable T *_ptr = nullptr; mutable T *_ptr = nullptr;
// internal helpers // internal helpers
static _FORCE_INLINE_ SafeNumeric<USize> *_get_refcount_ptr(uint8_t *p_ptr) {
return (SafeNumeric<USize> *)(p_ptr + REF_COUNT_OFFSET);
}
static _FORCE_INLINE_ USize *_get_size_ptr(uint8_t *p_ptr) {
return (USize *)(p_ptr + SIZE_OFFSET);
}
static _FORCE_INLINE_ T *_get_data_ptr(uint8_t *p_ptr) {
return (T *)(p_ptr + DATA_OFFSET);
}
_FORCE_INLINE_ SafeNumeric<USize> *_get_refcount() const { _FORCE_INLINE_ SafeNumeric<USize> *_get_refcount() const {
if (!_ptr) { if (!_ptr) {
return nullptr; return nullptr;
} }
return reinterpret_cast<SafeNumeric<USize> *>(_ptr) - 2; return (SafeNumeric<USize> *)((uint8_t *)_ptr - DATA_OFFSET + REF_COUNT_OFFSET);
} }
_FORCE_INLINE_ USize *_get_size() const { _FORCE_INLINE_ USize *_get_size() const {
@ -108,7 +129,7 @@ private:
return nullptr; return nullptr;
} }
return reinterpret_cast<USize *>(_ptr) - 1; return (USize *)((uint8_t *)_ptr - DATA_OFFSET + SIZE_OFFSET);
} }
_FORCE_INLINE_ USize _get_alloc_size(USize p_elements) const { _FORCE_INLINE_ USize _get_alloc_size(USize p_elements) const {
@ -244,7 +265,7 @@ void CowData<T>::_unref(void *p_data) {
} }
// free mem // free mem
Memory::free_static(((uint8_t *)p_data) - ALLOC_PAD, false); Memory::free_static(((uint8_t *)p_data) - DATA_OFFSET, false);
} }
template <class T> template <class T>
@ -260,26 +281,27 @@ typename CowData<T>::USize CowData<T>::_copy_on_write() {
/* in use by more than me */ /* in use by more than me */
USize current_size = *_get_size(); USize current_size = *_get_size();
USize *mem_new = (USize *)Memory::alloc_static(_get_alloc_size(current_size) + ALLOC_PAD, false); uint8_t *mem_new = (uint8_t *)Memory::alloc_static(_get_alloc_size(current_size) + DATA_OFFSET, false);
mem_new += 2; ERR_FAIL_NULL_V(mem_new, 0);
new (mem_new - 2) SafeNumeric<USize>(1); //refcount SafeNumeric<USize> *_refc_ptr = _get_refcount_ptr(mem_new);
*(mem_new - 1) = current_size; //size USize *_size_ptr = _get_size_ptr(mem_new);
T *_data_ptr = _get_data_ptr(mem_new);
T *_data = (T *)(mem_new); new (_refc_ptr) SafeNumeric<USize>(1); //refcount
*(_size_ptr) = current_size; //size
// initialize new elements // initialize new elements
if constexpr (std::is_trivially_copyable_v<T>) { if constexpr (std::is_trivially_copyable_v<T>) {
memcpy(mem_new, _ptr, current_size * sizeof(T)); memcpy((uint8_t *)_data_ptr, _ptr, current_size * sizeof(T));
} else { } else {
for (USize i = 0; i < current_size; i++) { for (USize i = 0; i < current_size; i++) {
memnew_placement(&_data[i], T(_ptr[i])); memnew_placement(&_data_ptr[i], T(_ptr[i]));
} }
} }
_unref(_ptr); _unref(_ptr);
_ptr = _data; _ptr = _data_ptr;
rc = 1; rc = 1;
} }
@ -315,21 +337,28 @@ Error CowData<T>::resize(Size p_size) {
if (alloc_size != current_alloc_size) { if (alloc_size != current_alloc_size) {
if (current_size == 0) { if (current_size == 0) {
// alloc from scratch // alloc from scratch
USize *ptr = (USize *)Memory::alloc_static(alloc_size + ALLOC_PAD, false); uint8_t *mem_new = (uint8_t *)Memory::alloc_static(alloc_size + DATA_OFFSET, false);
ptr += 2; ERR_FAIL_NULL_V(mem_new, ERR_OUT_OF_MEMORY);
ERR_FAIL_NULL_V(ptr, ERR_OUT_OF_MEMORY);
*(ptr - 1) = 0; //size, currently none
new (ptr - 2) SafeNumeric<USize>(1); //refcount
_ptr = (T *)ptr; SafeNumeric<USize> *_refc_ptr = _get_refcount_ptr(mem_new);
USize *_size_ptr = _get_size_ptr(mem_new);
T *_data_ptr = _get_data_ptr(mem_new);
new (_refc_ptr) SafeNumeric<USize>(1); //refcount
*(_size_ptr) = 0; //size, currently none
_ptr = _data_ptr;
} else { } else {
USize *_ptrnew = (USize *)Memory::realloc_static(((uint8_t *)_ptr) - ALLOC_PAD, alloc_size + ALLOC_PAD, false); uint8_t *mem_new = (uint8_t *)Memory::realloc_static(((uint8_t *)_ptr) - DATA_OFFSET, alloc_size + DATA_OFFSET, false);
ERR_FAIL_NULL_V(_ptrnew, ERR_OUT_OF_MEMORY); ERR_FAIL_NULL_V(mem_new, ERR_OUT_OF_MEMORY);
_ptrnew += 2;
new (_ptrnew - 2) SafeNumeric<USize>(rc); //refcount
_ptr = (T *)(_ptrnew); SafeNumeric<USize> *_refc_ptr = _get_refcount_ptr(mem_new);
T *_data_ptr = _get_data_ptr(mem_new);
new (_refc_ptr) SafeNumeric<USize>(rc); //refcount
_ptr = _data_ptr;
} }
} }
@ -355,12 +384,15 @@ Error CowData<T>::resize(Size p_size) {
} }
if (alloc_size != current_alloc_size) { if (alloc_size != current_alloc_size) {
USize *_ptrnew = (USize *)Memory::realloc_static(((uint8_t *)_ptr) - ALLOC_PAD, alloc_size + ALLOC_PAD, false); uint8_t *mem_new = (uint8_t *)Memory::realloc_static(((uint8_t *)_ptr) - DATA_OFFSET, alloc_size + DATA_OFFSET, false);
ERR_FAIL_NULL_V(_ptrnew, ERR_OUT_OF_MEMORY); ERR_FAIL_NULL_V(mem_new, ERR_OUT_OF_MEMORY);
_ptrnew += 2;
new (_ptrnew - 2) SafeNumeric<USize>(rc); //refcount
_ptr = (T *)(_ptrnew); SafeNumeric<USize> *_refc_ptr = _get_refcount_ptr(mem_new);
T *_data_ptr = _get_data_ptr(mem_new);
new (_refc_ptr) SafeNumeric<USize>(rc); //refcount
_ptr = _data_ptr;
} }
*_get_size() = p_size; *_get_size() = p_size;