mirror of
https://github.com/ziglang/zig.git
synced 2024-11-15 16:45:27 +00:00
Merge pull request #13513 from ziglang/faster-wasm-gpa
WebAssembly-only fast allocator
This commit is contained in:
commit
71038c42f5
357
lib/std/heap.zig
357
lib/std/heap.zig
@ -1,13 +1,13 @@
|
||||
const std = @import("std.zig");
|
||||
const builtin = @import("builtin");
|
||||
const root = @import("root");
|
||||
const debug = std.debug;
|
||||
const assert = debug.assert;
|
||||
const assert = std.debug.assert;
|
||||
const testing = std.testing;
|
||||
const mem = std.mem;
|
||||
const os = std.os;
|
||||
const c = std.c;
|
||||
const maxInt = std.math.maxInt;
|
||||
const Allocator = std.mem.Allocator;
|
||||
|
||||
pub const LoggingAllocator = @import("heap/logging_allocator.zig").LoggingAllocator;
|
||||
pub const loggingAllocator = @import("heap/logging_allocator.zig").loggingAllocator;
|
||||
@ -16,8 +16,12 @@ pub const LogToWriterAllocator = @import("heap/log_to_writer_allocator.zig").Log
|
||||
pub const logToWriterAllocator = @import("heap/log_to_writer_allocator.zig").logToWriterAllocator;
|
||||
pub const ArenaAllocator = @import("heap/arena_allocator.zig").ArenaAllocator;
|
||||
pub const GeneralPurposeAllocator = @import("heap/general_purpose_allocator.zig").GeneralPurposeAllocator;
|
||||
pub const WasmAllocator = @import("heap/WasmAllocator.zig");
|
||||
pub const WasmPageAllocator = @import("heap/WasmPageAllocator.zig");
|
||||
pub const PageAllocator = @import("heap/PageAllocator.zig");
|
||||
|
||||
const Allocator = mem.Allocator;
|
||||
/// TODO Utilize this on Windows.
|
||||
pub var next_mmap_addr_hint: ?[*]align(mem.page_size) u8 = null;
|
||||
|
||||
const CAllocator = struct {
|
||||
comptime {
|
||||
@ -227,303 +231,6 @@ pub fn alignPageAllocLen(full_len: usize, len: usize) usize {
|
||||
return aligned_len;
|
||||
}
|
||||
|
||||
/// TODO Utilize this on Windows.
|
||||
pub var next_mmap_addr_hint: ?[*]align(mem.page_size) u8 = null;
|
||||
|
||||
const PageAllocator = struct {
|
||||
const vtable = Allocator.VTable{
|
||||
.alloc = alloc,
|
||||
.resize = resize,
|
||||
.free = free,
|
||||
};
|
||||
|
||||
fn alloc(_: *anyopaque, n: usize, log2_align: u8, ra: usize) ?[*]u8 {
|
||||
_ = ra;
|
||||
_ = log2_align;
|
||||
assert(n > 0);
|
||||
if (n > maxInt(usize) - (mem.page_size - 1)) return null;
|
||||
const aligned_len = mem.alignForward(n, mem.page_size);
|
||||
|
||||
if (builtin.os.tag == .windows) {
|
||||
const w = os.windows;
|
||||
const addr = w.VirtualAlloc(
|
||||
null,
|
||||
aligned_len,
|
||||
w.MEM_COMMIT | w.MEM_RESERVE,
|
||||
w.PAGE_READWRITE,
|
||||
) catch return null;
|
||||
return @ptrCast([*]align(mem.page_size) u8, @alignCast(mem.page_size, addr));
|
||||
}
|
||||
|
||||
const hint = @atomicLoad(@TypeOf(next_mmap_addr_hint), &next_mmap_addr_hint, .Unordered);
|
||||
const slice = os.mmap(
|
||||
hint,
|
||||
aligned_len,
|
||||
os.PROT.READ | os.PROT.WRITE,
|
||||
os.MAP.PRIVATE | os.MAP.ANONYMOUS,
|
||||
-1,
|
||||
0,
|
||||
) catch return null;
|
||||
assert(mem.isAligned(@ptrToInt(slice.ptr), mem.page_size));
|
||||
const new_hint = @alignCast(mem.page_size, slice.ptr + aligned_len);
|
||||
_ = @cmpxchgStrong(@TypeOf(next_mmap_addr_hint), &next_mmap_addr_hint, hint, new_hint, .Monotonic, .Monotonic);
|
||||
return slice.ptr;
|
||||
}
|
||||
|
||||
fn resize(
|
||||
_: *anyopaque,
|
||||
buf_unaligned: []u8,
|
||||
log2_buf_align: u8,
|
||||
new_size: usize,
|
||||
return_address: usize,
|
||||
) bool {
|
||||
_ = log2_buf_align;
|
||||
_ = return_address;
|
||||
const new_size_aligned = mem.alignForward(new_size, mem.page_size);
|
||||
|
||||
if (builtin.os.tag == .windows) {
|
||||
const w = os.windows;
|
||||
if (new_size <= buf_unaligned.len) {
|
||||
const base_addr = @ptrToInt(buf_unaligned.ptr);
|
||||
const old_addr_end = base_addr + buf_unaligned.len;
|
||||
const new_addr_end = mem.alignForward(base_addr + new_size, mem.page_size);
|
||||
if (old_addr_end > new_addr_end) {
|
||||
// For shrinking that is not releasing, we will only
|
||||
// decommit the pages not needed anymore.
|
||||
w.VirtualFree(
|
||||
@intToPtr(*anyopaque, new_addr_end),
|
||||
old_addr_end - new_addr_end,
|
||||
w.MEM_DECOMMIT,
|
||||
);
|
||||
}
|
||||
return true;
|
||||
}
|
||||
const old_size_aligned = mem.alignForward(buf_unaligned.len, mem.page_size);
|
||||
if (new_size_aligned <= old_size_aligned) {
|
||||
return true;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
const buf_aligned_len = mem.alignForward(buf_unaligned.len, mem.page_size);
|
||||
if (new_size_aligned == buf_aligned_len)
|
||||
return true;
|
||||
|
||||
if (new_size_aligned < buf_aligned_len) {
|
||||
const ptr = @alignCast(mem.page_size, buf_unaligned.ptr + new_size_aligned);
|
||||
// TODO: if the next_mmap_addr_hint is within the unmapped range, update it
|
||||
os.munmap(ptr[0 .. buf_aligned_len - new_size_aligned]);
|
||||
return true;
|
||||
}
|
||||
|
||||
// TODO: call mremap
|
||||
// TODO: if the next_mmap_addr_hint is within the remapped range, update it
|
||||
return false;
|
||||
}
|
||||
|
||||
fn free(_: *anyopaque, slice: []u8, log2_buf_align: u8, return_address: usize) void {
|
||||
_ = log2_buf_align;
|
||||
_ = return_address;
|
||||
|
||||
if (builtin.os.tag == .windows) {
|
||||
os.windows.VirtualFree(slice.ptr, 0, os.windows.MEM_RELEASE);
|
||||
} else {
|
||||
const buf_aligned_len = mem.alignForward(slice.len, mem.page_size);
|
||||
const ptr = @alignCast(mem.page_size, slice.ptr);
|
||||
os.munmap(ptr[0..buf_aligned_len]);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
const WasmPageAllocator = struct {
|
||||
comptime {
|
||||
if (!builtin.target.isWasm()) {
|
||||
@compileError("WasmPageAllocator is only available for wasm32 arch");
|
||||
}
|
||||
}
|
||||
|
||||
const vtable = Allocator.VTable{
|
||||
.alloc = alloc,
|
||||
.resize = resize,
|
||||
.free = free,
|
||||
};
|
||||
|
||||
const PageStatus = enum(u1) {
|
||||
used = 0,
|
||||
free = 1,
|
||||
|
||||
pub const none_free: u8 = 0;
|
||||
};
|
||||
|
||||
const FreeBlock = struct {
|
||||
data: []u128,
|
||||
|
||||
const Io = std.packed_int_array.PackedIntIo(u1, .Little);
|
||||
|
||||
fn totalPages(self: FreeBlock) usize {
|
||||
return self.data.len * 128;
|
||||
}
|
||||
|
||||
fn isInitialized(self: FreeBlock) bool {
|
||||
return self.data.len > 0;
|
||||
}
|
||||
|
||||
fn getBit(self: FreeBlock, idx: usize) PageStatus {
|
||||
const bit_offset = 0;
|
||||
return @intToEnum(PageStatus, Io.get(mem.sliceAsBytes(self.data), idx, bit_offset));
|
||||
}
|
||||
|
||||
fn setBits(self: FreeBlock, start_idx: usize, len: usize, val: PageStatus) void {
|
||||
const bit_offset = 0;
|
||||
var i: usize = 0;
|
||||
while (i < len) : (i += 1) {
|
||||
Io.set(mem.sliceAsBytes(self.data), start_idx + i, bit_offset, @enumToInt(val));
|
||||
}
|
||||
}
|
||||
|
||||
// Use '0xFFFFFFFF' as a _missing_ sentinel
|
||||
// This saves ~50 bytes compared to returning a nullable
|
||||
|
||||
// We can guarantee that conventional memory never gets this big,
|
||||
// and wasm32 would not be able to address this memory (32 GB > usize).
|
||||
|
||||
// Revisit if this is settled: https://github.com/ziglang/zig/issues/3806
|
||||
const not_found = std.math.maxInt(usize);
|
||||
|
||||
fn useRecycled(self: FreeBlock, num_pages: usize, log2_align: u8) usize {
|
||||
@setCold(true);
|
||||
for (self.data) |segment, i| {
|
||||
const spills_into_next = @bitCast(i128, segment) < 0;
|
||||
const has_enough_bits = @popCount(segment) >= num_pages;
|
||||
|
||||
if (!spills_into_next and !has_enough_bits) continue;
|
||||
|
||||
var j: usize = i * 128;
|
||||
while (j < (i + 1) * 128) : (j += 1) {
|
||||
var count: usize = 0;
|
||||
while (j + count < self.totalPages() and self.getBit(j + count) == .free) {
|
||||
count += 1;
|
||||
const addr = j * mem.page_size;
|
||||
if (count >= num_pages and mem.isAlignedLog2(addr, log2_align)) {
|
||||
self.setBits(j, num_pages, .used);
|
||||
return j;
|
||||
}
|
||||
}
|
||||
j += count;
|
||||
}
|
||||
}
|
||||
return not_found;
|
||||
}
|
||||
|
||||
fn recycle(self: FreeBlock, start_idx: usize, len: usize) void {
|
||||
self.setBits(start_idx, len, .free);
|
||||
}
|
||||
};
|
||||
|
||||
var _conventional_data = [_]u128{0} ** 16;
|
||||
// Marking `conventional` as const saves ~40 bytes
|
||||
const conventional = FreeBlock{ .data = &_conventional_data };
|
||||
var extended = FreeBlock{ .data = &[_]u128{} };
|
||||
|
||||
fn extendedOffset() usize {
|
||||
return conventional.totalPages();
|
||||
}
|
||||
|
||||
fn nPages(memsize: usize) usize {
|
||||
return mem.alignForward(memsize, mem.page_size) / mem.page_size;
|
||||
}
|
||||
|
||||
fn alloc(_: *anyopaque, len: usize, log2_align: u8, ra: usize) ?[*]u8 {
|
||||
_ = ra;
|
||||
if (len > maxInt(usize) - (mem.page_size - 1)) return null;
|
||||
const page_count = nPages(len);
|
||||
const page_idx = allocPages(page_count, log2_align) catch return null;
|
||||
return @intToPtr([*]u8, page_idx * mem.page_size);
|
||||
}
|
||||
|
||||
fn allocPages(page_count: usize, log2_align: u8) !usize {
|
||||
{
|
||||
const idx = conventional.useRecycled(page_count, log2_align);
|
||||
if (idx != FreeBlock.not_found) {
|
||||
return idx;
|
||||
}
|
||||
}
|
||||
|
||||
const idx = extended.useRecycled(page_count, log2_align);
|
||||
if (idx != FreeBlock.not_found) {
|
||||
return idx + extendedOffset();
|
||||
}
|
||||
|
||||
const next_page_idx = @wasmMemorySize(0);
|
||||
const next_page_addr = next_page_idx * mem.page_size;
|
||||
const aligned_addr = mem.alignForwardLog2(next_page_addr, log2_align);
|
||||
const drop_page_count = @divExact(aligned_addr - next_page_addr, mem.page_size);
|
||||
const result = @wasmMemoryGrow(0, @intCast(u32, drop_page_count + page_count));
|
||||
if (result <= 0)
|
||||
return error.OutOfMemory;
|
||||
assert(result == next_page_idx);
|
||||
const aligned_page_idx = next_page_idx + drop_page_count;
|
||||
if (drop_page_count > 0) {
|
||||
freePages(next_page_idx, aligned_page_idx);
|
||||
}
|
||||
return @intCast(usize, aligned_page_idx);
|
||||
}
|
||||
|
||||
fn freePages(start: usize, end: usize) void {
|
||||
if (start < extendedOffset()) {
|
||||
conventional.recycle(start, @min(extendedOffset(), end) - start);
|
||||
}
|
||||
if (end > extendedOffset()) {
|
||||
var new_end = end;
|
||||
if (!extended.isInitialized()) {
|
||||
// Steal the last page from the memory currently being recycled
|
||||
// TODO: would it be better if we use the first page instead?
|
||||
new_end -= 1;
|
||||
|
||||
extended.data = @intToPtr([*]u128, new_end * mem.page_size)[0 .. mem.page_size / @sizeOf(u128)];
|
||||
// Since this is the first page being freed and we consume it, assume *nothing* is free.
|
||||
mem.set(u128, extended.data, PageStatus.none_free);
|
||||
}
|
||||
const clamped_start = @max(extendedOffset(), start);
|
||||
extended.recycle(clamped_start - extendedOffset(), new_end - clamped_start);
|
||||
}
|
||||
}
|
||||
|
||||
fn resize(
|
||||
_: *anyopaque,
|
||||
buf: []u8,
|
||||
log2_buf_align: u8,
|
||||
new_len: usize,
|
||||
return_address: usize,
|
||||
) bool {
|
||||
_ = log2_buf_align;
|
||||
_ = return_address;
|
||||
const aligned_len = mem.alignForward(buf.len, mem.page_size);
|
||||
if (new_len > aligned_len) return false;
|
||||
const current_n = nPages(aligned_len);
|
||||
const new_n = nPages(new_len);
|
||||
if (new_n != current_n) {
|
||||
const base = nPages(@ptrToInt(buf.ptr));
|
||||
freePages(base + new_n, base + current_n);
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
fn free(
|
||||
_: *anyopaque,
|
||||
buf: []u8,
|
||||
log2_buf_align: u8,
|
||||
return_address: usize,
|
||||
) void {
|
||||
_ = log2_buf_align;
|
||||
_ = return_address;
|
||||
const aligned_len = mem.alignForward(buf.len, mem.page_size);
|
||||
const current_n = nPages(aligned_len);
|
||||
const base = nPages(@ptrToInt(buf.ptr));
|
||||
freePages(base, base + current_n);
|
||||
}
|
||||
};
|
||||
|
||||
pub const HeapAllocator = switch (builtin.os.tag) {
|
||||
.windows => struct {
|
||||
heap_handle: ?HeapHandle,
|
||||
@ -859,43 +566,6 @@ test "raw_c_allocator" {
|
||||
}
|
||||
}
|
||||
|
||||
test "WasmPageAllocator internals" {
|
||||
if (comptime builtin.target.isWasm()) {
|
||||
const conventional_memsize = WasmPageAllocator.conventional.totalPages() * mem.page_size;
|
||||
const initial = try page_allocator.alloc(u8, mem.page_size);
|
||||
try testing.expect(@ptrToInt(initial.ptr) < conventional_memsize); // If this isn't conventional, the rest of these tests don't make sense. Also we have a serious memory leak in the test suite.
|
||||
|
||||
var inplace = try page_allocator.realloc(initial, 1);
|
||||
try testing.expectEqual(initial.ptr, inplace.ptr);
|
||||
inplace = try page_allocator.realloc(inplace, 4);
|
||||
try testing.expectEqual(initial.ptr, inplace.ptr);
|
||||
page_allocator.free(inplace);
|
||||
|
||||
const reuse = try page_allocator.alloc(u8, 1);
|
||||
try testing.expectEqual(initial.ptr, reuse.ptr);
|
||||
page_allocator.free(reuse);
|
||||
|
||||
// This segment may span conventional and extended which has really complex rules so we're just ignoring it for now.
|
||||
const padding = try page_allocator.alloc(u8, conventional_memsize);
|
||||
page_allocator.free(padding);
|
||||
|
||||
const extended = try page_allocator.alloc(u8, conventional_memsize);
|
||||
try testing.expect(@ptrToInt(extended.ptr) >= conventional_memsize);
|
||||
|
||||
const use_small = try page_allocator.alloc(u8, 1);
|
||||
try testing.expectEqual(initial.ptr, use_small.ptr);
|
||||
page_allocator.free(use_small);
|
||||
|
||||
inplace = try page_allocator.realloc(extended, 1);
|
||||
try testing.expectEqual(extended.ptr, inplace.ptr);
|
||||
page_allocator.free(inplace);
|
||||
|
||||
const reuse_extended = try page_allocator.alloc(u8, conventional_memsize);
|
||||
try testing.expectEqual(extended.ptr, reuse_extended.ptr);
|
||||
page_allocator.free(reuse_extended);
|
||||
}
|
||||
}
|
||||
|
||||
test "PageAllocator" {
|
||||
const allocator = page_allocator;
|
||||
try testAllocator(allocator);
|
||||
@ -1163,7 +833,14 @@ pub fn testAllocatorAlignedShrink(base_allocator: mem.Allocator) !void {
|
||||
try testing.expect(slice[60] == 0x34);
|
||||
}
|
||||
|
||||
test "heap" {
|
||||
_ = @import("heap/logging_allocator.zig");
|
||||
_ = @import("heap/log_to_writer_allocator.zig");
|
||||
test {
|
||||
_ = LoggingAllocator;
|
||||
_ = LogToWriterAllocator;
|
||||
_ = ScopedLoggingAllocator;
|
||||
_ = ArenaAllocator;
|
||||
_ = GeneralPurposeAllocator;
|
||||
if (comptime builtin.target.isWasm()) {
|
||||
_ = WasmAllocator;
|
||||
_ = WasmPageAllocator;
|
||||
}
|
||||
}
|
||||
|
110
lib/std/heap/PageAllocator.zig
Normal file
110
lib/std/heap/PageAllocator.zig
Normal file
@ -0,0 +1,110 @@
|
||||
const std = @import("../std.zig");
|
||||
const builtin = @import("builtin");
|
||||
const Allocator = std.mem.Allocator;
|
||||
const mem = std.mem;
|
||||
const os = std.os;
|
||||
const maxInt = std.math.maxInt;
|
||||
const assert = std.debug.assert;
|
||||
|
||||
pub const vtable = Allocator.VTable{
|
||||
.alloc = alloc,
|
||||
.resize = resize,
|
||||
.free = free,
|
||||
};
|
||||
|
||||
fn alloc(_: *anyopaque, n: usize, log2_align: u8, ra: usize) ?[*]u8 {
|
||||
_ = ra;
|
||||
_ = log2_align;
|
||||
assert(n > 0);
|
||||
if (n > maxInt(usize) - (mem.page_size - 1)) return null;
|
||||
const aligned_len = mem.alignForward(n, mem.page_size);
|
||||
|
||||
if (builtin.os.tag == .windows) {
|
||||
const w = os.windows;
|
||||
const addr = w.VirtualAlloc(
|
||||
null,
|
||||
aligned_len,
|
||||
w.MEM_COMMIT | w.MEM_RESERVE,
|
||||
w.PAGE_READWRITE,
|
||||
) catch return null;
|
||||
return @ptrCast([*]align(mem.page_size) u8, @alignCast(mem.page_size, addr));
|
||||
}
|
||||
|
||||
const hint = @atomicLoad(@TypeOf(std.heap.next_mmap_addr_hint), &std.heap.next_mmap_addr_hint, .Unordered);
|
||||
const slice = os.mmap(
|
||||
hint,
|
||||
aligned_len,
|
||||
os.PROT.READ | os.PROT.WRITE,
|
||||
os.MAP.PRIVATE | os.MAP.ANONYMOUS,
|
||||
-1,
|
||||
0,
|
||||
) catch return null;
|
||||
assert(mem.isAligned(@ptrToInt(slice.ptr), mem.page_size));
|
||||
const new_hint = @alignCast(mem.page_size, slice.ptr + aligned_len);
|
||||
_ = @cmpxchgStrong(@TypeOf(std.heap.next_mmap_addr_hint), &std.heap.next_mmap_addr_hint, hint, new_hint, .Monotonic, .Monotonic);
|
||||
return slice.ptr;
|
||||
}
|
||||
|
||||
fn resize(
|
||||
_: *anyopaque,
|
||||
buf_unaligned: []u8,
|
||||
log2_buf_align: u8,
|
||||
new_size: usize,
|
||||
return_address: usize,
|
||||
) bool {
|
||||
_ = log2_buf_align;
|
||||
_ = return_address;
|
||||
const new_size_aligned = mem.alignForward(new_size, mem.page_size);
|
||||
|
||||
if (builtin.os.tag == .windows) {
|
||||
const w = os.windows;
|
||||
if (new_size <= buf_unaligned.len) {
|
||||
const base_addr = @ptrToInt(buf_unaligned.ptr);
|
||||
const old_addr_end = base_addr + buf_unaligned.len;
|
||||
const new_addr_end = mem.alignForward(base_addr + new_size, mem.page_size);
|
||||
if (old_addr_end > new_addr_end) {
|
||||
// For shrinking that is not releasing, we will only
|
||||
// decommit the pages not needed anymore.
|
||||
w.VirtualFree(
|
||||
@intToPtr(*anyopaque, new_addr_end),
|
||||
old_addr_end - new_addr_end,
|
||||
w.MEM_DECOMMIT,
|
||||
);
|
||||
}
|
||||
return true;
|
||||
}
|
||||
const old_size_aligned = mem.alignForward(buf_unaligned.len, mem.page_size);
|
||||
if (new_size_aligned <= old_size_aligned) {
|
||||
return true;
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
const buf_aligned_len = mem.alignForward(buf_unaligned.len, mem.page_size);
|
||||
if (new_size_aligned == buf_aligned_len)
|
||||
return true;
|
||||
|
||||
if (new_size_aligned < buf_aligned_len) {
|
||||
const ptr = @alignCast(mem.page_size, buf_unaligned.ptr + new_size_aligned);
|
||||
// TODO: if the next_mmap_addr_hint is within the unmapped range, update it
|
||||
os.munmap(ptr[0 .. buf_aligned_len - new_size_aligned]);
|
||||
return true;
|
||||
}
|
||||
|
||||
// TODO: call mremap
|
||||
// TODO: if the next_mmap_addr_hint is within the remapped range, update it
|
||||
return false;
|
||||
}
|
||||
|
||||
fn free(_: *anyopaque, slice: []u8, log2_buf_align: u8, return_address: usize) void {
|
||||
_ = log2_buf_align;
|
||||
_ = return_address;
|
||||
|
||||
if (builtin.os.tag == .windows) {
|
||||
os.windows.VirtualFree(slice.ptr, 0, os.windows.MEM_RELEASE);
|
||||
} else {
|
||||
const buf_aligned_len = mem.alignForward(slice.len, mem.page_size);
|
||||
const ptr = @alignCast(mem.page_size, slice.ptr);
|
||||
os.munmap(ptr[0..buf_aligned_len]);
|
||||
}
|
||||
}
|
317
lib/std/heap/WasmAllocator.zig
Normal file
317
lib/std/heap/WasmAllocator.zig
Normal file
@ -0,0 +1,317 @@
|
||||
//! This is intended to be merged into GeneralPurposeAllocator at some point.
|
||||
|
||||
const std = @import("../std.zig");
|
||||
const builtin = @import("builtin");
|
||||
const Allocator = std.mem.Allocator;
|
||||
const mem = std.mem;
|
||||
const assert = std.debug.assert;
|
||||
const wasm = std.wasm;
|
||||
const math = std.math;
|
||||
|
||||
comptime {
|
||||
if (!builtin.target.isWasm()) {
|
||||
@compileError("WasmPageAllocator is only available for wasm32 arch");
|
||||
}
|
||||
}
|
||||
|
||||
pub const vtable = Allocator.VTable{
|
||||
.alloc = alloc,
|
||||
.resize = resize,
|
||||
.free = free,
|
||||
};
|
||||
|
||||
pub const Error = Allocator.Error;
|
||||
|
||||
const max_usize = math.maxInt(usize);
|
||||
const ushift = math.Log2Int(usize);
|
||||
const bigpage_size = 64 * 1024;
|
||||
const pages_per_bigpage = bigpage_size / wasm.page_size;
|
||||
const bigpage_count = max_usize / bigpage_size;
|
||||
|
||||
/// Because of storing free list pointers, the minimum size class is 3.
|
||||
const min_class = math.log2(math.ceilPowerOfTwoAssert(usize, 1 + @sizeOf(usize)));
|
||||
const size_class_count = math.log2(bigpage_size) - min_class;
|
||||
/// 0 - 1 bigpage
|
||||
/// 1 - 2 bigpages
|
||||
/// 2 - 4 bigpages
|
||||
/// etc.
|
||||
const big_size_class_count = math.log2(bigpage_count);
|
||||
|
||||
var next_addrs = [1]usize{0} ** size_class_count;
|
||||
/// For each size class, points to the freed pointer.
|
||||
var frees = [1]usize{0} ** size_class_count;
|
||||
/// For each big size class, points to the freed pointer.
|
||||
var big_frees = [1]usize{0} ** big_size_class_count;
|
||||
|
||||
fn alloc(ctx: *anyopaque, len: usize, log2_align: u8, return_address: usize) ?[*]u8 {
|
||||
_ = ctx;
|
||||
_ = return_address;
|
||||
// Make room for the freelist next pointer.
|
||||
const alignment = @as(usize, 1) << @intCast(Allocator.Log2Align, log2_align);
|
||||
const actual_len = @max(len +| @sizeOf(usize), alignment);
|
||||
const slot_size = math.ceilPowerOfTwo(usize, actual_len) catch return null;
|
||||
const class = math.log2(slot_size) - min_class;
|
||||
if (class < size_class_count) {
|
||||
const addr = a: {
|
||||
const top_free_ptr = frees[class];
|
||||
if (top_free_ptr != 0) {
|
||||
const node = @intToPtr(*usize, top_free_ptr + (slot_size - @sizeOf(usize)));
|
||||
frees[class] = node.*;
|
||||
break :a top_free_ptr;
|
||||
}
|
||||
|
||||
const next_addr = next_addrs[class];
|
||||
if (next_addr % wasm.page_size == 0) {
|
||||
const addr = allocBigPages(1);
|
||||
if (addr == 0) return null;
|
||||
//std.debug.print("allocated fresh slot_size={d} class={d} addr=0x{x}\n", .{
|
||||
// slot_size, class, addr,
|
||||
//});
|
||||
next_addrs[class] = addr + slot_size;
|
||||
break :a addr;
|
||||
} else {
|
||||
next_addrs[class] = next_addr + slot_size;
|
||||
break :a next_addr;
|
||||
}
|
||||
};
|
||||
return @intToPtr([*]u8, addr);
|
||||
}
|
||||
const bigpages_needed = bigPagesNeeded(actual_len);
|
||||
const addr = allocBigPages(bigpages_needed);
|
||||
return @intToPtr([*]u8, addr);
|
||||
}
|
||||
|
||||
fn resize(
|
||||
ctx: *anyopaque,
|
||||
buf: []u8,
|
||||
log2_buf_align: u8,
|
||||
new_len: usize,
|
||||
return_address: usize,
|
||||
) bool {
|
||||
_ = ctx;
|
||||
_ = return_address;
|
||||
// We don't want to move anything from one size class to another, but we
|
||||
// can recover bytes in between powers of two.
|
||||
const buf_align = @as(usize, 1) << @intCast(Allocator.Log2Align, log2_buf_align);
|
||||
const old_actual_len = @max(buf.len + @sizeOf(usize), buf_align);
|
||||
const new_actual_len = @max(new_len +| @sizeOf(usize), buf_align);
|
||||
const old_small_slot_size = math.ceilPowerOfTwoAssert(usize, old_actual_len);
|
||||
const old_small_class = math.log2(old_small_slot_size) - min_class;
|
||||
if (old_small_class < size_class_count) {
|
||||
const new_small_slot_size = math.ceilPowerOfTwo(usize, new_actual_len) catch return false;
|
||||
return old_small_slot_size == new_small_slot_size;
|
||||
} else {
|
||||
const old_bigpages_needed = bigPagesNeeded(old_actual_len);
|
||||
const old_big_slot_pages = math.ceilPowerOfTwoAssert(usize, old_bigpages_needed);
|
||||
const new_bigpages_needed = bigPagesNeeded(new_actual_len);
|
||||
const new_big_slot_pages = math.ceilPowerOfTwo(usize, new_bigpages_needed) catch return false;
|
||||
return old_big_slot_pages == new_big_slot_pages;
|
||||
}
|
||||
}
|
||||
|
||||
fn free(
|
||||
ctx: *anyopaque,
|
||||
buf: []u8,
|
||||
log2_buf_align: u8,
|
||||
return_address: usize,
|
||||
) void {
|
||||
_ = ctx;
|
||||
_ = return_address;
|
||||
const buf_align = @as(usize, 1) << @intCast(Allocator.Log2Align, log2_buf_align);
|
||||
const actual_len = @max(buf.len + @sizeOf(usize), buf_align);
|
||||
const slot_size = math.ceilPowerOfTwoAssert(usize, actual_len);
|
||||
const class = math.log2(slot_size) - min_class;
|
||||
const addr = @ptrToInt(buf.ptr);
|
||||
if (class < size_class_count) {
|
||||
const node = @intToPtr(*usize, addr + (slot_size - @sizeOf(usize)));
|
||||
node.* = frees[class];
|
||||
frees[class] = addr;
|
||||
} else {
|
||||
const bigpages_needed = bigPagesNeeded(actual_len);
|
||||
const pow2_pages = math.ceilPowerOfTwoAssert(usize, bigpages_needed);
|
||||
const big_slot_size_bytes = pow2_pages * bigpage_size;
|
||||
const node = @intToPtr(*usize, addr + (big_slot_size_bytes - @sizeOf(usize)));
|
||||
const big_class = math.log2(pow2_pages);
|
||||
node.* = big_frees[big_class];
|
||||
big_frees[big_class] = addr;
|
||||
}
|
||||
}
|
||||
|
||||
inline fn bigPagesNeeded(byte_count: usize) usize {
|
||||
return (byte_count + (bigpage_size + (@sizeOf(usize) - 1))) / bigpage_size;
|
||||
}
|
||||
|
||||
fn allocBigPages(n: usize) usize {
|
||||
const pow2_pages = math.ceilPowerOfTwoAssert(usize, n);
|
||||
const slot_size_bytes = pow2_pages * bigpage_size;
|
||||
const class = math.log2(pow2_pages);
|
||||
|
||||
const top_free_ptr = big_frees[class];
|
||||
if (top_free_ptr != 0) {
|
||||
const node = @intToPtr(*usize, top_free_ptr + (slot_size_bytes - @sizeOf(usize)));
|
||||
big_frees[class] = node.*;
|
||||
return top_free_ptr;
|
||||
}
|
||||
|
||||
const page_index = @wasmMemoryGrow(0, pow2_pages * pages_per_bigpage);
|
||||
if (page_index <= 0) return 0;
|
||||
const addr = @intCast(u32, page_index) * wasm.page_size;
|
||||
return addr;
|
||||
}
|
||||
|
||||
const test_ally = Allocator{
|
||||
.ptr = undefined,
|
||||
.vtable = &vtable,
|
||||
};
|
||||
|
||||
test "small allocations - free in same order" {
|
||||
var list: [513]*u64 = undefined;
|
||||
|
||||
var i: usize = 0;
|
||||
while (i < 513) : (i += 1) {
|
||||
const ptr = try test_ally.create(u64);
|
||||
list[i] = ptr;
|
||||
}
|
||||
|
||||
for (list) |ptr| {
|
||||
test_ally.destroy(ptr);
|
||||
}
|
||||
}
|
||||
|
||||
test "small allocations - free in reverse order" {
|
||||
var list: [513]*u64 = undefined;
|
||||
|
||||
var i: usize = 0;
|
||||
while (i < 513) : (i += 1) {
|
||||
const ptr = try test_ally.create(u64);
|
||||
list[i] = ptr;
|
||||
}
|
||||
|
||||
i = list.len;
|
||||
while (i > 0) {
|
||||
i -= 1;
|
||||
const ptr = list[i];
|
||||
test_ally.destroy(ptr);
|
||||
}
|
||||
}
|
||||
|
||||
test "large allocations" {
|
||||
const ptr1 = try test_ally.alloc(u64, 42768);
|
||||
const ptr2 = try test_ally.alloc(u64, 52768);
|
||||
test_ally.free(ptr1);
|
||||
const ptr3 = try test_ally.alloc(u64, 62768);
|
||||
test_ally.free(ptr3);
|
||||
test_ally.free(ptr2);
|
||||
}
|
||||
|
||||
test "very large allocation" {
|
||||
try std.testing.expectError(error.OutOfMemory, test_ally.alloc(u8, math.maxInt(usize)));
|
||||
}
|
||||
|
||||
test "realloc" {
|
||||
var slice = try test_ally.alignedAlloc(u8, @alignOf(u32), 1);
|
||||
defer test_ally.free(slice);
|
||||
slice[0] = 0x12;
|
||||
|
||||
// This reallocation should keep its pointer address.
|
||||
const old_slice = slice;
|
||||
slice = try test_ally.realloc(slice, 2);
|
||||
try std.testing.expect(old_slice.ptr == slice.ptr);
|
||||
try std.testing.expect(slice[0] == 0x12);
|
||||
slice[1] = 0x34;
|
||||
|
||||
// This requires upgrading to a larger size class
|
||||
slice = try test_ally.realloc(slice, 17);
|
||||
try std.testing.expect(slice[0] == 0x12);
|
||||
try std.testing.expect(slice[1] == 0x34);
|
||||
}
|
||||
|
||||
test "shrink" {
|
||||
var slice = try test_ally.alloc(u8, 20);
|
||||
defer test_ally.free(slice);
|
||||
|
||||
mem.set(u8, slice, 0x11);
|
||||
|
||||
try std.testing.expect(test_ally.resize(slice, 17));
|
||||
slice = slice[0..17];
|
||||
|
||||
for (slice) |b| {
|
||||
try std.testing.expect(b == 0x11);
|
||||
}
|
||||
|
||||
try std.testing.expect(test_ally.resize(slice, 16));
|
||||
slice = slice[0..16];
|
||||
|
||||
for (slice) |b| {
|
||||
try std.testing.expect(b == 0x11);
|
||||
}
|
||||
}
|
||||
|
||||
test "large object - grow" {
|
||||
var slice1 = try test_ally.alloc(u8, bigpage_size * 2 - 20);
|
||||
defer test_ally.free(slice1);
|
||||
|
||||
const old = slice1;
|
||||
slice1 = try test_ally.realloc(slice1, bigpage_size * 2 - 10);
|
||||
try std.testing.expect(slice1.ptr == old.ptr);
|
||||
|
||||
slice1 = try test_ally.realloc(slice1, bigpage_size * 2);
|
||||
slice1 = try test_ally.realloc(slice1, bigpage_size * 2 + 1);
|
||||
}
|
||||
|
||||
test "realloc small object to large object" {
|
||||
var slice = try test_ally.alloc(u8, 70);
|
||||
defer test_ally.free(slice);
|
||||
slice[0] = 0x12;
|
||||
slice[60] = 0x34;
|
||||
|
||||
// This requires upgrading to a large object
|
||||
const large_object_size = bigpage_size * 2 + 50;
|
||||
slice = try test_ally.realloc(slice, large_object_size);
|
||||
try std.testing.expect(slice[0] == 0x12);
|
||||
try std.testing.expect(slice[60] == 0x34);
|
||||
}
|
||||
|
||||
test "shrink large object to large object" {
|
||||
var slice = try test_ally.alloc(u8, bigpage_size * 2 + 50);
|
||||
defer test_ally.free(slice);
|
||||
slice[0] = 0x12;
|
||||
slice[60] = 0x34;
|
||||
|
||||
try std.testing.expect(test_ally.resize(slice, bigpage_size * 2 + 1));
|
||||
slice = slice[0 .. bigpage_size * 2 + 1];
|
||||
try std.testing.expect(slice[0] == 0x12);
|
||||
try std.testing.expect(slice[60] == 0x34);
|
||||
|
||||
try std.testing.expect(test_ally.resize(slice, bigpage_size * 2 + 1));
|
||||
try std.testing.expect(slice[0] == 0x12);
|
||||
try std.testing.expect(slice[60] == 0x34);
|
||||
|
||||
slice = try test_ally.realloc(slice, bigpage_size * 2);
|
||||
try std.testing.expect(slice[0] == 0x12);
|
||||
try std.testing.expect(slice[60] == 0x34);
|
||||
}
|
||||
|
||||
test "realloc large object to small object" {
|
||||
var slice = try test_ally.alloc(u8, bigpage_size * 2 + 50);
|
||||
defer test_ally.free(slice);
|
||||
slice[0] = 0x12;
|
||||
slice[16] = 0x34;
|
||||
|
||||
slice = try test_ally.realloc(slice, 19);
|
||||
try std.testing.expect(slice[0] == 0x12);
|
||||
try std.testing.expect(slice[16] == 0x34);
|
||||
}
|
||||
|
||||
test "objects of size 1024 and 2048" {
|
||||
const slice = try test_ally.alloc(u8, 1025);
|
||||
const slice2 = try test_ally.alloc(u8, 3000);
|
||||
|
||||
test_ally.free(slice);
|
||||
test_ally.free(slice2);
|
||||
}
|
||||
|
||||
test "standard allocator tests" {
|
||||
try std.heap.testAllocator(test_ally);
|
||||
try std.heap.testAllocatorAligned(test_ally);
|
||||
}
|
235
lib/std/heap/WasmPageAllocator.zig
Normal file
235
lib/std/heap/WasmPageAllocator.zig
Normal file
@ -0,0 +1,235 @@
|
||||
const WasmPageAllocator = @This();
|
||||
const std = @import("../std.zig");
|
||||
const builtin = @import("builtin");
|
||||
const Allocator = std.mem.Allocator;
|
||||
const mem = std.mem;
|
||||
const maxInt = std.math.maxInt;
|
||||
const assert = std.debug.assert;
|
||||
|
||||
comptime {
|
||||
if (!builtin.target.isWasm()) {
|
||||
@compileError("WasmPageAllocator is only available for wasm32 arch");
|
||||
}
|
||||
}
|
||||
|
||||
pub const vtable = Allocator.VTable{
|
||||
.alloc = alloc,
|
||||
.resize = resize,
|
||||
.free = free,
|
||||
};
|
||||
|
||||
const PageStatus = enum(u1) {
|
||||
used = 0,
|
||||
free = 1,
|
||||
|
||||
pub const none_free: u8 = 0;
|
||||
};
|
||||
|
||||
const FreeBlock = struct {
|
||||
data: []u128,
|
||||
|
||||
const Io = std.packed_int_array.PackedIntIo(u1, .Little);
|
||||
|
||||
fn totalPages(self: FreeBlock) usize {
|
||||
return self.data.len * 128;
|
||||
}
|
||||
|
||||
fn isInitialized(self: FreeBlock) bool {
|
||||
return self.data.len > 0;
|
||||
}
|
||||
|
||||
fn getBit(self: FreeBlock, idx: usize) PageStatus {
|
||||
const bit_offset = 0;
|
||||
return @intToEnum(PageStatus, Io.get(mem.sliceAsBytes(self.data), idx, bit_offset));
|
||||
}
|
||||
|
||||
fn setBits(self: FreeBlock, start_idx: usize, len: usize, val: PageStatus) void {
|
||||
const bit_offset = 0;
|
||||
var i: usize = 0;
|
||||
while (i < len) : (i += 1) {
|
||||
Io.set(mem.sliceAsBytes(self.data), start_idx + i, bit_offset, @enumToInt(val));
|
||||
}
|
||||
}
|
||||
|
||||
// Use '0xFFFFFFFF' as a _missing_ sentinel
|
||||
// This saves ~50 bytes compared to returning a nullable
|
||||
|
||||
// We can guarantee that conventional memory never gets this big,
|
||||
// and wasm32 would not be able to address this memory (32 GB > usize).
|
||||
|
||||
// Revisit if this is settled: https://github.com/ziglang/zig/issues/3806
|
||||
const not_found = maxInt(usize);
|
||||
|
||||
fn useRecycled(self: FreeBlock, num_pages: usize, log2_align: u8) usize {
|
||||
@setCold(true);
|
||||
for (self.data) |segment, i| {
|
||||
const spills_into_next = @bitCast(i128, segment) < 0;
|
||||
const has_enough_bits = @popCount(segment) >= num_pages;
|
||||
|
||||
if (!spills_into_next and !has_enough_bits) continue;
|
||||
|
||||
var j: usize = i * 128;
|
||||
while (j < (i + 1) * 128) : (j += 1) {
|
||||
var count: usize = 0;
|
||||
while (j + count < self.totalPages() and self.getBit(j + count) == .free) {
|
||||
count += 1;
|
||||
const addr = j * mem.page_size;
|
||||
if (count >= num_pages and mem.isAlignedLog2(addr, log2_align)) {
|
||||
self.setBits(j, num_pages, .used);
|
||||
return j;
|
||||
}
|
||||
}
|
||||
j += count;
|
||||
}
|
||||
}
|
||||
return not_found;
|
||||
}
|
||||
|
||||
fn recycle(self: FreeBlock, start_idx: usize, len: usize) void {
|
||||
self.setBits(start_idx, len, .free);
|
||||
}
|
||||
};
|
||||
|
||||
var _conventional_data = [_]u128{0} ** 16;
|
||||
// Marking `conventional` as const saves ~40 bytes
|
||||
const conventional = FreeBlock{ .data = &_conventional_data };
|
||||
var extended = FreeBlock{ .data = &[_]u128{} };
|
||||
|
||||
fn extendedOffset() usize {
|
||||
return conventional.totalPages();
|
||||
}
|
||||
|
||||
fn nPages(memsize: usize) usize {
|
||||
return mem.alignForward(memsize, mem.page_size) / mem.page_size;
|
||||
}
|
||||
|
||||
fn alloc(ctx: *anyopaque, len: usize, log2_align: u8, ra: usize) ?[*]u8 {
|
||||
_ = ctx;
|
||||
_ = ra;
|
||||
if (len > maxInt(usize) - (mem.page_size - 1)) return null;
|
||||
const page_count = nPages(len);
|
||||
const page_idx = allocPages(page_count, log2_align) catch return null;
|
||||
return @intToPtr([*]u8, page_idx * mem.page_size);
|
||||
}
|
||||
|
||||
fn allocPages(page_count: usize, log2_align: u8) !usize {
|
||||
{
|
||||
const idx = conventional.useRecycled(page_count, log2_align);
|
||||
if (idx != FreeBlock.not_found) {
|
||||
return idx;
|
||||
}
|
||||
}
|
||||
|
||||
const idx = extended.useRecycled(page_count, log2_align);
|
||||
if (idx != FreeBlock.not_found) {
|
||||
return idx + extendedOffset();
|
||||
}
|
||||
|
||||
const next_page_idx = @wasmMemorySize(0);
|
||||
const next_page_addr = next_page_idx * mem.page_size;
|
||||
const aligned_addr = mem.alignForwardLog2(next_page_addr, log2_align);
|
||||
const drop_page_count = @divExact(aligned_addr - next_page_addr, mem.page_size);
|
||||
const result = @wasmMemoryGrow(0, @intCast(u32, drop_page_count + page_count));
|
||||
if (result <= 0)
|
||||
return error.OutOfMemory;
|
||||
assert(result == next_page_idx);
|
||||
const aligned_page_idx = next_page_idx + drop_page_count;
|
||||
if (drop_page_count > 0) {
|
||||
freePages(next_page_idx, aligned_page_idx);
|
||||
}
|
||||
return @intCast(usize, aligned_page_idx);
|
||||
}
|
||||
|
||||
fn freePages(start: usize, end: usize) void {
|
||||
if (start < extendedOffset()) {
|
||||
conventional.recycle(start, @min(extendedOffset(), end) - start);
|
||||
}
|
||||
if (end > extendedOffset()) {
|
||||
var new_end = end;
|
||||
if (!extended.isInitialized()) {
|
||||
// Steal the last page from the memory currently being recycled
|
||||
// TODO: would it be better if we use the first page instead?
|
||||
new_end -= 1;
|
||||
|
||||
extended.data = @intToPtr([*]u128, new_end * mem.page_size)[0 .. mem.page_size / @sizeOf(u128)];
|
||||
// Since this is the first page being freed and we consume it, assume *nothing* is free.
|
||||
mem.set(u128, extended.data, PageStatus.none_free);
|
||||
}
|
||||
const clamped_start = @max(extendedOffset(), start);
|
||||
extended.recycle(clamped_start - extendedOffset(), new_end - clamped_start);
|
||||
}
|
||||
}
|
||||
|
||||
fn resize(
|
||||
ctx: *anyopaque,
|
||||
buf: []u8,
|
||||
log2_buf_align: u8,
|
||||
new_len: usize,
|
||||
return_address: usize,
|
||||
) bool {
|
||||
_ = ctx;
|
||||
_ = log2_buf_align;
|
||||
_ = return_address;
|
||||
const aligned_len = mem.alignForward(buf.len, mem.page_size);
|
||||
if (new_len > aligned_len) return false;
|
||||
const current_n = nPages(aligned_len);
|
||||
const new_n = nPages(new_len);
|
||||
if (new_n != current_n) {
|
||||
const base = nPages(@ptrToInt(buf.ptr));
|
||||
freePages(base + new_n, base + current_n);
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
fn free(
|
||||
ctx: *anyopaque,
|
||||
buf: []u8,
|
||||
log2_buf_align: u8,
|
||||
return_address: usize,
|
||||
) void {
|
||||
_ = ctx;
|
||||
_ = log2_buf_align;
|
||||
_ = return_address;
|
||||
const aligned_len = mem.alignForward(buf.len, mem.page_size);
|
||||
const current_n = nPages(aligned_len);
|
||||
const base = nPages(@ptrToInt(buf.ptr));
|
||||
freePages(base, base + current_n);
|
||||
}
|
||||
|
||||
test "internals" {
|
||||
const page_allocator = std.heap.page_allocator;
|
||||
const testing = std.testing;
|
||||
|
||||
const conventional_memsize = WasmPageAllocator.conventional.totalPages() * mem.page_size;
|
||||
const initial = try page_allocator.alloc(u8, mem.page_size);
|
||||
try testing.expect(@ptrToInt(initial.ptr) < conventional_memsize); // If this isn't conventional, the rest of these tests don't make sense. Also we have a serious memory leak in the test suite.
|
||||
|
||||
var inplace = try page_allocator.realloc(initial, 1);
|
||||
try testing.expectEqual(initial.ptr, inplace.ptr);
|
||||
inplace = try page_allocator.realloc(inplace, 4);
|
||||
try testing.expectEqual(initial.ptr, inplace.ptr);
|
||||
page_allocator.free(inplace);
|
||||
|
||||
const reuse = try page_allocator.alloc(u8, 1);
|
||||
try testing.expectEqual(initial.ptr, reuse.ptr);
|
||||
page_allocator.free(reuse);
|
||||
|
||||
// This segment may span conventional and extended which has really complex rules so we're just ignoring it for now.
|
||||
const padding = try page_allocator.alloc(u8, conventional_memsize);
|
||||
page_allocator.free(padding);
|
||||
|
||||
const ext = try page_allocator.alloc(u8, conventional_memsize);
|
||||
try testing.expect(@ptrToInt(ext.ptr) >= conventional_memsize);
|
||||
|
||||
const use_small = try page_allocator.alloc(u8, 1);
|
||||
try testing.expectEqual(initial.ptr, use_small.ptr);
|
||||
page_allocator.free(use_small);
|
||||
|
||||
inplace = try page_allocator.realloc(ext, 1);
|
||||
try testing.expectEqual(ext.ptr, inplace.ptr);
|
||||
page_allocator.free(inplace);
|
||||
|
||||
const reuse_extended = try page_allocator.alloc(u8, conventional_memsize);
|
||||
try testing.expectEqual(ext.ptr, reuse_extended.ptr);
|
||||
page_allocator.free(reuse_extended);
|
||||
}
|
Loading…
Reference in New Issue
Block a user