[stdlib] Fix SIMD[..., u128/s128] construction from UInt
#4339
Conversation
| alias a = rebind[UInt](-1) | ||
| alias a_str = String(a) | ||
| assert_equal(a_str, String(UInt128(a))) |
There was a problem hiding this comment.
We can't do much better than stringify due to #4049.
mojo/stdlib/src/builtin/simd.mojo
Outdated
| self = Self(value.value) | ||
| @parameter | ||
| if bitwidthof[dtype]() > bitwidthof[DType.index](): | ||
| self = Self(rebind[UInt64](value)) |
There was a problem hiding this comment.
Cannot use UInt64(value) here. I think there's a compiler bug(?) where v.cast[index]().cast[int128]() gets simplified to v.cast[int128]() (on the pop level), which lowers to sext yet again. So we use rebind to skip the constructor from index entirely.
There was a problem hiding this comment.
Actually, it's probably not really a compiler bug. Suppose the target type is UInt128. If we go through UInt64, we get roughly the following MLIR:
var v: index = ...
var a = pop.cast[int64](v)
var b = pop.cast[int128](a)Since index is treated as signed (?), the first cast is a no-op, and the second is a sign extension.
There was a problem hiding this comment.
Question: Aren't you assuming 64 bitwidth machines by using UInt64?
There was a problem hiding this comment.
Yea, I'm being lazy here. Will fix it properly when the team confirmes that rebind is an acceptable approach.
There was a problem hiding this comment.
Actually, rebind[T] doesn't work when alias T = Scalar[_unsigned_integral_type_of[DType.index]()], so I opted for a different fix.
soraros
force-pushed
the
index-to-i128
branch
2 times, most recently
from
fb3be68 to
e883e8b
Compare
Signed-off-by: Yiwu Chen <[email protected]>
Signed-off-by: Yiwu Chen <[email protected]>
Close #4338.