llvm / llvm/llvm-project

Inconsistent big-endian layout for packed boolean vectors

Open
#224,659 1 comment 0 reactions 0 assignees View on GitHub
clang:frontend
Dominant language
LLVM
Stars
40.5k
Forks
18.7k
PR merge metrics
PR metrics pending

Description

On big-endian targets, static initialization of packed boolean vectors with size less than 8 uses the low bits of a byte, while loads and stores use the high bits.

```c
typedef _Bool bool2 __attribute__((ext_vector_type(2)));

bool2 global_v = {1, 1};

void set(bool2 *p) {
*p = (bool2){1, 1};
}

_Bool first(void) {
return global_v[0];
}
```

```sh
clang --target=armv7-none-eabi -march=armv7-m -mbig-endian \
-O1 -S repro.c -o -
```

**Actual:** `global_v` is emitted as `0x03`, but `set()` stores `0xC0`.
`first()` reads bit 7, returning `0` initially and `1` after `set(&global_v)`.

**Expected:** Initialization, loads, and stores must agree; `first()` should
return `1` in both cases. The little-endian variant behaves correctly.

[Compiler Explorer example](https://godbolt.org/z/Y5ae6hjn9).

Since packed boolean vectors with size >=8 have their bits allocated to the lower bits of the bytes, I believe that the memory layout used by the static initialization is the correct one.

Contributor guide

Open the contributing guide

Research direction

Start by compiling the issue's repro.c with the provided clang armv7 big-endian command and compare the emitted global initializer with set() and first(). The fix is complete when packed boolean vectors smaller than 8 use the same bit layout for static initialization, loads, and stores, with first() returning 1 before and after set().

Written by the indexing model from the issue text.

Assessment

Tech stack
c
Domain
compilers
Issue type
Bug
Difficulty
3/5
Estimated time
1-2 days
Activity status
Active
Clarity
Clearly specified
Newbie friendliness
58/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.