I hope I don't need to explain why this is not ideal:
|
// Fallback: reserve a range of memory large enough to find a |
|
// sufficiently aligned address, then free the entire range and |
|
// immediately allocate the desired subset. Another thread may have won |
|
// the race to map the target range, in which case a retry is needed. |
|
windows.VirtualFree(addr, 0, windows.MEM_RELEASE); |
|
|
|
const overalloc_len = n + alignment_bytes - page_size; |
|
const aligned_len = mem.alignForward(usize, n, page_size); |
|
|
|
while (true) { |
|
const reserved_addr = windows.VirtualAlloc( |
|
null, |
|
overalloc_len, |
|
windows.MEM_RESERVE, |
|
windows.PAGE_NOACCESS, |
|
) catch return null; |
|
const aligned_addr = mem.alignForward(usize, @intFromPtr(reserved_addr), alignment_bytes); |
|
windows.VirtualFree(reserved_addr, 0, windows.MEM_RELEASE); |
|
const ptr = windows.VirtualAlloc( |
|
@ptrFromInt(aligned_addr), |
|
aligned_len, |
|
windows.MEM_COMMIT | windows.MEM_RESERVE, |
|
windows.PAGE_READWRITE, |
|
) catch continue; |
|
return @ptrCast(ptr); |
|
} |
This issue is a subtask of #1840.
Issue close criteria:
As a bonus, measure SmpAllocator performance on Windows before and after, as well as providing a comparison with kernel32 HeapAlloc and related functions.
Related:
I hope I don't need to explain why this is not ideal:
zig/lib/std/heap/PageAllocator.zig
Lines 44 to 69 in d4c8507
This issue is a subtask of #1840.
Issue close criteria:
VirtualAllocfrom std libVirtualFreefrom std libNtAllocateVirtualMemoryinsteadMEM_RESERVE_PLACEHOLDERto guarantee no raceAs a bonus, measure SmpAllocator performance on Windows before and after, as well as providing a comparison with kernel32 HeapAlloc and related functions.
Related:
next_mmap_addr_hinton Windows #17377