[CodeGen] Rework MVT representation of capabilities and add type inference #726
Conversation
|
NB: I have not yet tested CheriBSD with this. |
| // Complement of In. | ||
| auto CompIn = [&In](MVT T) -> bool { return !In.count(T); }; | ||
| namespace { | ||
| enum class TypeCat { |
There was a problem hiding this comment.
I will likely rewrite this to be less overkill with the templates. At one point I thought it was going to be a lot more complicated, but as it turns out the function really just needs an MVT and an MVT->bool.
| ArrayRef<ValueTypeByHwMode> PtrTys({ValueTypeByHwMode(MVT::iFATPTRAny), | ||
| ValueTypeByHwMode(MVT::iPTR)}); | ||
| return NodeToApply->UpdateNodeType(ResNo, PtrTys, TP); | ||
| // TODO: Make this less ugly |
There was a problem hiding this comment.
I intend to fix this before it gets merged
There was a problem hiding this comment.
This looks like a huge improvement. I have never looked at those tablegen internals, so if it works that sounds good to me.
One or two of the clang format suggestions are legitimate since the shorter name now makes the line fit but most of them are quite odd suggestions...
…rence This more closely aligns how capabilites are handled with iPTR and iN types, introducing a new cPTR wildcard and fixed-width cN types that it maps to, with iPTRAny being iPTR and cPTR. As a result we no longer need to duplicate patterns for each fixed-width type in certain cases, and the few upstream MIPS tests we've long marked as XFAIL'ed now pass. As part of this, a few of the APIs around MVTs have been renamed from fat pointer terminology to use capability instead. Where relatively straightforward to do these have then been adopted throughout, but some are pervasive and have deprecated aliases under the old names, whose uses should be incrementally reduced.
|
Clean builds of cheribsd-riscv64-purecap have cheribsdtest-* all passing on CHERI-QEMU and CHERI-PURECAP-QEMU kernels. |
|
FYI llvm/llvm-project#81688 for upstreaming the intersect redesign. |
This more closely aligns how capabilites are handled with iPTR and iN
types, introducing a new cPTR wildcard and fixed-width cN types that it
maps to, with iPTRAny being iPTR and cPTR. As a result we no longer need
to duplicate patterns for each fixed-width type in certain cases, and
the few upstream MIPS tests we've long marked as XFAIL'ed now pass.
As part of this, a few of the APIs around MVTs have been renamed from
fat pointer terminology to use capability instead. Where relatively
straightforward to do these have then been adopted throughout, but some
are pervasive and have deprecated aliases under the old names, whose
uses should be incrementally reduced.