Fix V8 UnitGUID x86 ABI stack leak
This commit is contained in:
+11
-1
@@ -203,7 +203,17 @@ pub fn getMapId() u32 {
|
||||
|
||||
/// UnitGUID("player") / UnitGUID("target") → 64-bit GUID.
|
||||
pub fn unitGUID(unit_id: [*:0]const u8) u64 {
|
||||
return hook.call(fn ([*:0]const u8) callconv(hook.cc.fastcall) u64, o.FN_UNIT_GUID, .{unit_id});
|
||||
// UnitGUID is __fastcall with its single argument in ECX.
|
||||
// hook.call + hook.cc.fastcall is currently miscompiled by Zig 0.16 on x86
|
||||
// as a stack argument. Because the callee consumes no stack argument, that
|
||||
// leaked 4 bytes per call; Outline calls UnitGUID("player"/"target") every
|
||||
// frame, so ESP drifted until execution returned into stack data.
|
||||
//
|
||||
// With a single register argument, x86_thiscall is ABI-compatible here:
|
||||
// ECX=unit_id, no stack args, u64 returned in EDX:EAX.
|
||||
const f_native: *const fn ([*:0]const u8) callconv(.{ .x86_thiscall = .{} }) u64 =
|
||||
@ptrFromInt(o.FN_UNIT_GUID);
|
||||
return @call(.never_tail, f_native, .{unit_id});
|
||||
}
|
||||
|
||||
/// Resolve a GUID → object pointer via the object manager hash table.
|
||||
|
||||
Reference in New Issue
Block a user