[CIR] Make use of !invariant.group metadata for const allocas #1159
Conversation
Lancern
force-pushed
the
const-lowering
branch
from
909a116
to
2ae7176
Compare
I think we should emit the metadata under -O1+ to prevent LLVM from any possible optimization under -O0. And I realized that LowerToLLVM code does not have access to the code generation option which keeps the optimization flags. Sent #1171 for this and I'm marking this PR as draft until that PR lands. |
Lancern
force-pushed
the
const-lowering
branch
from
2ae7176
to
3f1749d
Compare
|
Rebased onto the latest main. |
Lancern
force-pushed
the
const-lowering
branch
from
3f1749d
to
fe2f762
Compare
This patch updates the LLVM lowering part of load and stores to const allocas and makes use of the !invariant.group metadata in the result LLVM IR. The HoistAlloca pass is also updated. The const flag on a hoisted alloca is removed for now since their uses are not always invariants. Will update in later patches to teach their invariants.
Lancern
force-pushed
the
const-lowering
branch
from
fe2f762
to
b249b48
Compare
| alloca->moveBefore(insertPoint); | ||
| if (alloca.getConstant()) { | ||
| // For now, we remove the const flag on nested allocas for constant | ||
| // variables when hoisting them. |
There was a problem hiding this comment.
In theory the alloca is still const, this is a workaround for the invariant side of things that are going to be generated later, in theory we could be propagating another attribute that has the specific semantic and delete it instead. I'm not sure it's worth adding that right now though. Can you explain that in the comment instead?
There was a problem hiding this comment.
in theory we could be propagating another attribute that has the specific semantic and delete it instead
I'm wondering recently if we could leverage the still-missing CIR lifetime markers as pointed out in #1129 . The rough idea is the constness property would only hold in the program points between the lifetime start and end markers, and then we don't have to remove the const flag here. What do you think?
Anyway, this would not block this PR. Will update the comment here later.
This PR updates the LLVM lowering part of load and stores to const allocas and makes use of the !invariant.group metadata in the result LLVM IR.
The HoistAlloca pass is also updated. The const flag on a hoisted alloca is removed for now since their uses are not always invariants. Will update in later PRs to teach their invariants.