[BUG] `DTypePointer[bool]` packs bits inconsistently

[helehex]

Bug description

When using DTypePointer[bool] store()/load() with different widths, you get inconsistent results.

Steps to reproduce

var ptr = DTypePointer[DType.bool].alloc(4)
ptr.store(0, True)
ptr.store(1, True)
ptr.store(2, True)
ptr.store(3, True)
print(ptr.load[width=4](0))

prints: [True, False, False, False]

System information

Host Information
  ================

  Target Triple: x86_64-unknown-linux
  CPU: skylake
  CPU Features: adx, aes, avx, avx2, bmi, bmi2, clflushopt, cmov, crc32, cx16, cx8, f16c, fma, fsgsbase, fxsr, invpcid, lzcnt, mmx, movbe, pclmul, popcnt, prfchw, rdrnd, rdseed, sahf, sgx, sse, sse2, sse3, sse4.1, sse4.2, ssse3, x87, xsave, xsavec, xsaveopt, xsaves

mojo 2024.5.2405 (fc1ed0a3)

modular 0.8.0 (39a426b5)

[MadAlex1997]

Probably Related to 2492, I am pretty sure there are issues at the DType pointer level with the i1 representation of SIMD[DType.bool,1] and the 8 Bit C style bool used in the pointers.

[bgreni]

Seems as though this has been fixed with recent changes?

var ptr = DTypePointer[DType.bool].alloc(4)
SIMD[size=4].store(ptr, SIMD[DType.bool, 4](True, False, True, True))
print(SIMD[size=4].load(ptr)) # prints [True, False, True, True]

[rparolin]

Looks like the patch mentioned above by @bgreni resolved the issue. I'll close this issue but we can reopen it if knock-ons are discovered. Thank you!

[helehex]

looks like when store/load moved to SIMD, the fix wasn't

[helehex]

var ptr = DTypePointer[DType.bool].alloc(4)
SIMD[DType.bool, 1].store(ptr, 0, True)
SIMD[DType.bool, 1].store(ptr, 1, True)
SIMD[DType.bool, 1].store(ptr, 2, True)
SIMD[DType.bool, 1].store(ptr, 3, True)
print(SIMD[DType.bool, 4].load(ptr, 0))

still prints [True, False, False, False]

[helehex]

The inconsistency is that width packs bits, but the offset is in bytes. This means everything gets weird when using different widths on the same pointer.

SIMD[DType.bool, 1].store(ptr, 0, True)
SIMD[DType.bool, 1].store(ptr, 1, True)
# b00000001
# b00000001

vs:

SIMD[DType.bool, 2].store(ptr, 0, True)
# b00000011

[bgreni]

@helehex hmm yeah I see your point, loading the entire thing in/out works better now, but trying to do piecemeal operations still doesn't work.

[MarkOWiesemann]

Is this supposed to work now? I am on mojo 24.4.0 (2cb57382)

[MadAlex1997]

Is this supposed to work now? I am on mojo 24.4.0 (2cb57382)

They fixed it in nightly but it did not make it into 24.4. And I don't know if it got reverted in nightly.

def main():
    var boolpointer = DTypePointer[DType.bool]().alloc(32)
    boolpointer.store[width=64](0,SIMD[DType.bool,1](True))
    print(boolpointer)
    for i in range(32):
        print(boolpointer[i])

True
True
True
True
True
True
True
True
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False
False