x64: Add non-SSE4.1 lowerings of pmov{s,z}x*
#6279
Conversation
This commit adds lowerings for a suite of sign/zero extension
instructions which don't require SSE4.1. Like before these lowerings are
based on LLVM's output.
This commit also deletes special casees for `i16x8.extmul_{low,high}_*`
since the output of the special case is the same as the default lowering
of all the component instructions used within as well.
github-actions
bot
added
cranelift
|
|
||
| ;; Same as `uwiden_high`, but interleaving high lanes instead. | ||
| (rule (lower (has_type $I16X8 (uwiden_high val @ (value_type $I8X16)))) | ||
| (x64_punpckhbw val (xmm_zero $I8X16))) |
There was a problem hiding this comment.
Can't remember all the history of PALIGNR + PMOVZX above but this PUNPCKH might be a better lowering even in the SSE 4.1 case (fewer dependencies?). Just a thought, haven't looked too closely at this...
There was a problem hiding this comment.
I believe llvm-mca agrees with you, and LLVM appears to, even when I use the intrinsics for palignr + pmovzx, to still use the pupckh combo with an xor-to-create-zero.
|
Also, just a general thought: how can we check that fuzzing is using all of these new <SSE4.1 lowerings? |
LLVM prefers the `punpckh*`-based lowerings and at least according to `llvm-mca` these are slightly better cycle-wise too.
|
Currently there's no fuzz coverage due to this but at the end of this road I'll be able to disable that meaning we'll get fuzz coverage. Once that's removed though we should get fuzz coverage. For example after #6206 the |
* x64: Add non-SSE4.1 lowerings of `pmov{s,z}x*`
This commit adds lowerings for a suite of sign/zero extension
instructions which don't require SSE4.1. Like before these lowerings are
based on LLVM's output.
This commit also deletes special casees for `i16x8.extmul_{low,high}_*`
since the output of the special case is the same as the default lowering
of all the component instructions used within as well.
* Remove SSE4.1 specialization of `uwiden_high`
LLVM prefers the `punpckh*`-based lowerings and at least according to
`llvm-mca` these are slightly better cycle-wise too.
This commit adds lowerings for a suite of sign/zero extension instructions which don't require SSE4.1. Like before these lowerings are based on LLVM's output.
This commit also deletes special casees for
i16x8.extmul_{low,high}_*since the output of the special case is the same as the default lowering of all the component instructions used within as well.