AOSP Source Releases Are So Late Developers Disassemble Libraries to Report Bugs

Description

On Android 17 builds from September 2026, wcscspn() and wcspbrk() stop at the first character of the string that is >= U+0100 whenever the reject set has two or more characters that are all < U+0100. The character is reported as a match even though it is not in the set. wcsspn(), wcschr(), wcsrchr(), wmemchr() and wcsstr() are correct.

Any code that scans wide strings for ASCII delimiters, for example newline or wildcard characters, breaks as soon as the text contains an emoji, CJK, Cyrillic, Greek or Latin Extended character.

Affected / unaffected

The public android17-release sources (android-17.0.0_r1, June 2026) build wcscspn/wcspbrk from external/llvm-libc, which is a plain nested loop over wchar_t and is correct. The code on the September builds is different (see "Mechanism"), so the regression was introduced between the June and September releases.

Reproducer

Build with the NDK (aarch64-linux-android30-clang), push to /data/local/tmp, run with adb shell:

#include <stdio.h>
#include <wchar.h>

int main(void)
{
    wchar_t s[] = {'a', 0x1F680, 0x100, 'b', 'x', '\n', 0};
    printf("wcscspn(s, L\"x\\n\") = %zu (expected 4)\n", wcscspn(s, L"x\n"));
    printf("wcspbrk(s, L\"x\\n\") = index %ld (expected 4)\n", (long)(wcspbrk(s, L"x\n") - s));
    printf("wcsspn(s, L\"a\\U0001F680\") = %zu (expected 2)\n", wcsspn(s, L"a\U0001F680"));
    return 0;
}

Output on CP3A.260905.009:

wcscspn(s, L"x\n") = 1 (expected 4)
wcspbrk(s, L"x\n") = index 1 (expected 4)
wcsspn(s, L"a\U0001F680") = 2 (expected 2)

A sweep of every code point from U+0001 to U+10FFFF (skipping surrogates and the two set members) with wcscspn(L"<cp>\n", L"x\n") returns 0 instead of 1 for 1,111,808 code points: every one from U+0100 to U+10FFFF. U+0001 to U+00FF are correct.

Which set shapes trigger it (all measured on the Pixel 8a):

Mechanism (from the shipped libc.so)

Disassembly of wcscspn in /apex/com.android.runtime/lib64/bionic/libc.so on CP3A.260905.009 shows three paths selected by the set:

1. One character: a plain compare loop. Correct.
2. Two or more characters: a 256-byte bool table on the stack is filled from the set while the set characters are in 1..0xFF. If a set character >= 0x100 is seen, the function falls back to calling wcschr() for each haystack character. Correct.
3. Otherwise the haystack is scanned against the table:

a2564:  ldrb  w8, [x9, x8]        ; table[c]
a2568:  tbnz  w8, #0, a2578       ; in set -> stop (correct)
a256c:  ldr   w8, [x22, #0x4]!    ; next character
a2570:  cmp   w8, #0x100
a2574:  b.lo  a2564               ; c < 0x100 -> keep scanning
a2578:  ...                       ; c >= 0x100 -> STOP, reported as a match

A character >= 0x100 cannot be in a set whose members are all < 0x100, so for wcscspn the scan must continue past it. The code stops instead. wcsspn has the same structure, and there stopping is the correct answer (the character is not in the keep set), which is why only the complement functions are wrong. wcspbrk is implemented on top of this scan and inherits the bug.

The existing bionic test TEST(wchar, wcspbrk) in tests/wchar_test.cpp only uses ASCII input, so it does not catch this. A test such as EXPECT_EQ(3u, wcscspn(L"aĀb\n", L"x\n")) would.

Real-world impact

Visual Pinball on Android (libwinevbs, a Unix build of Wine's vbscript engine) stopped loading any table whose VBScript source contains a comment with an emoji after the September update: the lexer uses wcspbrk(ptr, L"\n\r") to skip comments, got the position of the first emoji instead of the line end, and reported "Invalid character". FileSystem.MoveFolder likewise treated every path containing a character >= U+0100 as a wildcard pattern via wcspbrk(src, L"*?><").

Worked around downstream by replacing the calls with plain loops: vpinball/libwinevbs#41 (shipped in vpinball/vpinball@e7837322f4).

Full disassembly of the affected functions and the probe sources are available on request.

I am using a supported NDK

• I have checked and the NDK I'm using is currently supported

Affected versions

r29