Leading zeros and other bit manipulations

[TheIronBorn]

#188

[gnzlbg]

might want to split these into two lines - rustfmt chokes on these

[gnzlbg]

LGTM, thank you!

[gnzlbg]

For the following architectures you need to:

• fill in an LLVM bug for the respective backends (aarch64 and s390x): ideally you just write LLVM-IR in godbolt for one case that works, say in x86_64, compile it for one of these targets, and show the LLVM error.
• open a bug here referencing the LLVM bug
• add a workaround to codegen for these targets that, instead of using the LLVM intrinsic, uses pure rust code to compute these - these workaround should have a FIXME referencing the bug in these repo

It looks like these are not available in aarch64:

LLVM ERROR: Cannot select: 0x7fb76ee85270: v8i8 = cttz 0x7fb76ef05270, src/codegen/bit_manip.rs:176:39
  0x7fb76ef05270: v8i8,ch = load<(dereferenceable load 8 from %ir.5)> 0x7fb76ef23758, FrameIndex:i64<2>, undef:i64, src/codegen/bit_manip.rs:176:31
    0x7fb76ee5b138: i64 = FrameIndex<2>
    0x7fb76ee90958: i64 = undef
In function: _ZN110_$LT$packed_simd..Simd$LT$$u5b$u8$u3b$$u20$_$u5d$$GT$$u20$as$u20$packed_simd..codegen..bit_manip..BitManip$GT$4cttz17h3aae67eb204ee427E
error: Could not compile `packed_simd`.

And in s390x:

Intrinsic has incorrect return type!
void (<16 x i8>*, <16 x i8>*)* @llvm.ctpop.v16i8
LLVM ERROR: Broken function found, compilation aborted!
error: Could not compile `packed_simd`.

Both appears to be bugs in the respective LLVM backends.

[gnzlbg]

See above.

[TheIronBorn]

Turns out aarch64 has instructions for ctlz and ctpop. We should add those to assert_instr tests.

[gnzlbg]

Please ping me when CI is green :) I try to check on this every now and then, but if you don't hear from me, then just ping me :D

[TheIronBorn]

The remaining CI failures seem to be unrelated.

[gnzlbg]

Thank you, this looks good. I've restarted the failing build bots, but yeah the failures look unrelated. Do you want to rebase this into a single commit ?

[gnzlbg]

Also could you open an issue about optimizing the aarch64 intrinsics here (and maybe include there adding assert_instr tests for those) ?

[TheIronBorn]

It seemed like every s390x intrinsic was failing so I just disabled them entirely

[gnzlbg]

@TheIronBorn You have contributed a lot to the crate, you know it's structure, have good judgement, and I think it would be awesome if you could be a maintainer of packed_simd, are you interested ?