- User Since
- Jun 1 2015, 8:58 AM (363 w, 5 d)
Thu, May 19
While this bug is reproducible on main, there's no way to properly fix the codegen without D125787 as we lack the ability to ensure RVV alignment beyond 8, and we don't need to ensure an alignment of 8 as we get that for free. So I thought it better to base this patch on D125787.
fix up a comment: RVV padding isn't only required when *accessing* RVV objects;
it affects the calculation of *any* objects when we have a non-zero RVV stack
rebase on D125964
Wed, May 18
Why was this reverted?
LGTM - that last comment isn't a blocker, btw.
Tue, May 17
add rv32 lines and zve64x
Mon, May 16
add zve64x run lines: update comments
Thu, May 12
LGTM but yeah I think [RISCV] seems more canonical to me.
From what I gather about the bug report, RISCV doesn't suffer from it as we always dyn_cast to Instruction. Is that right? Just wondering if we should add a similar test - I don't think we have constant expr coverage in sink-splat-operands.ll.
LGTM with a nit
I think @simoll should be included in VP reviews.
Wed, May 11
I think this can be exhibited even more simply. If we have one RVV object and call a function, that called function expects its stack to be 16-byte aligned. We're not ensuring that.
resuscitate old test
This is what I was thinking of. Short of speculatively materializing an instruction only to throw it away, I don't see a better way.
LGTM but I don't think you need to say [RISCV] in the title.
Also LGTM aside from nits and possible restructuring. Definitely nicer than the other patch.
100% agreed, thank you. FWIW I think there's scope to do something similar for our ISD nodes.
Tue, May 10
LGTM, I had kind of assumed we were already enabling this. So thanks! This should help us embedding liveness bugs deep into our code - I've seen targets struggle to right these when enabling subreg liveness.
Looks good to me but I'd defer to @craig.topper in case I've missed something.
Mon, May 9
pre-emptively include D125133
LGTM but yeah I prefer @gkm's suggestion.
Fri, May 6
remove need for TmpStatus (and BBInfo.Change, really)
rebase on new test which now fails
It's uncommon to add nodes and patterns without test coverage. I realise that'd involve adding custom lowering code, but can't that be done in this patch too?
fix and enable bitcasts for scalable vectors
Thu, May 5