cgen: load full 24B header on whole str/slice deref-by-value

`let s: str = *h` (a str/slice loaded by value through *str / *[]T)
fell through the N_UN deref arm to the scalar load, emitting a single
MOVQ that read only the 8B .ptr and left .len/.cap from stale registers,
so len(*p) returned garbage. Both stages emitted byte-identical wrong
code, so the self-compile byte-id gate was blind to it. Add a str/slice
arm that loads the full {ptr,len,cap} via cgslicehdr when the chased
pointee is TY_STR/TY_SLICE.

Surfaced by the codegen miscompile hunt (finding C1b). Pinned by
test/lang/deref_hdr_test.ww, which interposes a different-sized decoy
header so the test reddens when the arm is reverted.
This commit is contained in:
2026-06-27 11:47:54 +09:00
parent b855b3bdbc
commit c67f362bbd
3 changed files with 105 additions and 0 deletions

View File

@@ -4748,6 +4748,19 @@ cgexpr(Cg *c, Node *n, Local *locals)
if (ru && ru->kind == TY_TAGGED
&& !ru->nullable && ru->size > 8)
break;
/* C1b: a whole str/slice loaded BY VALUE
* through *str / *[]T — AX (the operand value) IS
* the 24B {ptr,len,cap} header address. The scalar
* load below pulled ONLY word0 (.ptr); .len/.cap
* were then stored from stale BX/CX, so len(*p) read
* garbage — byte-id-blind on both stages. Reuse the
* same 3-word header load as the slice-FIELD /
* N_INDEX str-element arms (cgslicehdr). */
if (ru && (ru->kind == TY_STR
|| ru->kind == TY_SLICE)) {
cgslicehdr(c, D_AX);
break;
}
}
/* f64/f32 result rides X0 (SSE), not AX — an integer
* MOVQ strands the value off the float ABI and the

View File

@@ -5525,6 +5525,19 @@ fn cgun(c: *cgen, n: *syntax.node) void = {
if (rti != nil && rti.kind == syntax.tykind.TY_TAGGED) {
if (rti.nullable == 0 && rti.size: i32 > 8) { return; };
};
// C1b: a whole str/slice loaded BY VALUE through *str /
// *[]T — AX (the operand value) IS the 24B {ptr,len,cap}
// header address. The scalar load below pulled ONLY word0
// (.ptr); .len/.cap were then stored from stale BX/CX, so
// len(*p) read garbage — byte-id-blind on both stages. Reuse
// the same 3-word header load as the slice-field / cgindex
// str-element arms.
if (rti != nil) {
if (rti.kind == syntax.tykind.TY_STR || rti.kind == syntax.tykind.TY_SLICE) {
cgslicehdr(c, "AX");
return;
};
};
// f64/f32 result rides X0 (SSE), not AX — an integer MOVQ
// strands the value off the float ABI and the caller's
// MOVSD X0 reads stale bits (#96). Mirrors the float

View File

@@ -0,0 +1,79 @@
// deref_hdr_test — runtime contract for a whole 24B str/slice header loaded
// BY VALUE through a pointer (`let s = *h` where h:*str / h:*[]T). C1b (#4):
// the N_UN TK_STAR deref arm fell through to a scalar MOVQ (AX),AX that loaded
// ONLY word0 (.ptr); .len/.cap were then stored from stale BX/CX, so len(*h)
// and cap(*h) read garbage. byte-id proves the two stages agree, NOT that the
// emitted code is correct (this gap was byte-id-identical on both stages).
//
// WHY the decoy: the pre-fix bug leaves .len/.cap in whatever BX/CX happened to
// hold. A naive `let backing="abcde"; let s=*h` PASSES even broken, because the
// literal's header-load left BX/CX = backing's own len/cap moments earlier —
// the exact register coincidence that let this class hide. Each fn interposes a
// DIFFERENT-sized decoy header right before the deref, so any stale BX/CX hold
// the decoy's len/cap (≠ backing's). Verified: with the fix reverted all four
// fns RED, with the fix all GREEN. The s[i] element reads additionally pin .ptr
// (loaded into AX) — covering the whole header on str and slice, in both the
// let-bind and fn-return shapes that propagate the bug.
package deref_hdr_test;
fn ret_str_deref(h: *str) str = { return *h; };
fn ret_slice_deref(h: *[]i32) []i32 = { return *h; };
@test fn str_whole_deref() void = {
// Whole str loaded by value through *str. The decoy ("xy", len 2) is the
// last header materialized before `*h`, so a pre-fix stale BX/CX hold
// 2/2, not backing's 5/5 — the .len/.cap asserts then RED on the bug.
// The s[i] byte reads stride from .ptr, pinning AX (word0) too.
let backing: str = "abcde";
let h: *str = &backing;
let decoy: str = "xy";
assert(decoy.len: i32 == 2);
let s: str = *h;
assert(s.len: i32 == 5);
assert(s.cap: i32 == 5);
assert(s[0]: i32 == 97);
assert(s[1]: i32 == 98);
assert(s[4]: i32 == 101);
};
@test fn slice_whole_deref() void = {
// Whole slice loaded by value through *[]i32 — same 3-word header load.
// Decoy (len 1) clobbers the stale-register path away from backing's 3.
let backing: []i32 = [10i32, 20i32, 30i32];
let h: *[]i32 = &backing;
let decoy: []i32 = [1i32];
assert(decoy.len: i32 == 1);
let s: []i32 = *h;
assert(s.len: i32 == 3);
assert(s.cap: i32 == 3);
assert(s[0]: i32 == 10);
assert(s[1]: i32 == 20);
assert(s[2]: i32 == 30);
};
@test fn str_deref_as_return() void = {
// `return *h` feeds the same arm into the return registers. The decoy
// ("xyz", len 3) before the call seeds BX/CX wrong, so a pre-fix callee
// that returns only .ptr surfaces 3/3 here instead of backing's 4/4.
let backing: str = "abcd";
let h: *str = &backing;
let decoy: str = "xyz";
assert(decoy.len: i32 == 3);
let s: str = ret_str_deref(h);
assert(s.len: i32 == 4);
assert(s.cap: i32 == 4);
assert(s[0]: i32 == 97);
};
@test fn slice_deref_as_return() void = {
let backing: []i32 = [7i32, 8i32];
let h: *[]i32 = &backing;
let decoy: []i32 = [1i32, 2i32, 3i32, 4i32];
assert(decoy.len: i32 == 4);
let s: []i32 = ret_slice_deref(h);
assert(s.len: i32 == 2);
assert(s.cap: i32 == 2);
assert(s[1]: i32 == 8);
};