Skip to content

Commit

Permalink
Merge pull request #13513 from ziglang/faster-wasm-gpa
Browse files Browse the repository at this point in the history 
WebAssembly-only fast allocator
  • Loading branch information
@andrewrk
andrewrk authored Nov 30, 2022
2 parents e35f297 + 7f063b2 commit 71038c4
Show file tree
Hide file tree
Showing 4 changed files with 679 additions and 340 deletions.
357 changes: 17 additions & 340 deletions lib/std/heap.zig
View file
Original file line number Diff line number Diff line change
@@ -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;
Expand All @@ -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 {
Expand Down Expand Up @@ -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,
Expand Down Expand Up @@ -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);
Expand Down Expand Up @@ -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;
}
}
Loading

0 comments on commit 71038c4

Please sign in to comment.