Implement LLVM x86 AVX2 intrinsics #3492
Conversation
eduardosm
force-pushed
the
intrinsics-x86-avx2
branch
from
964ade8 to
80316e1
Compare
|
My hope was that we'd implement them in libstd as fallback bodies instead (#3397), but I guess now that we have a shim impl and tests, doing that would be simpler as we have some help against messing up the fallback bodies |
… on `rustc_target::abi::Size`
eduardosm
force-pushed
the
intrinsics-x86-avx2
branch
from
80316e1 to
af9fc5d
Compare
|
@bors r+ |
|
☀️ Test successful - checks-actions |
| let dest = this.project_index(&dest, i)?; | ||
|
|
||
| // Converting to a host "i128" works since the input is always signed. | ||
| let res = op.to_int(dest.layout.size)?.unsigned_abs(); |
There was a problem hiding this comment.
This may be worth an assertion.
Alternatively, a moire clear implementation might avoid using host integers entirely and just use binop_with_overflow.
There was a problem hiding this comment.
I'm not sure what to check with an assertion. How would you implement this with binop_with_overflow?
There was a problem hiding this comment.
The assertion would check that the element type is signed.
How would you implement this with binop_with_overflow?
Sorry, I meant wrapping_binary_op.
Basically run the equivalent of if op < 0 { -op } else { op }. It's one BinOp::lt and one... okay you need wrapping_unary_op as well, UnOp::Neg. You can use ImmTy::from_int to create a 0 of the right type.
| Ok(()) | ||
| } | ||
|
|
||
| fn pack_generic<'tcx>( |
| let (mask, mask_len) = this.operand_to_simd(mask)?; | ||
| let (dest, dest_len) = this.mplace_to_simd(dest)?; | ||
|
|
||
| // There are cases like dest: i32x4, offsets: i64x2 |
There was a problem hiding this comment.
What happens if the offsets are longer than the dest?
There was a problem hiding this comment.
Extra elements are ignored.
There was a problem hiding this comment.
And that's how the intrinsic is intended to work? Wow.
Would be good to have a comment explaining that.
No description provided.