optimizer: handle GlobalRef based on assume_bindings_static
#54148
Conversation
base/compiler/ssair/ir.jl
Outdated
| @@ -428,22 +428,26 @@ struct IRCode | |||
| cfg::CFG | |||
| new_nodes::NewNodeStream | |||
| meta::Vector{Expr} | |||
| assume_bindings_static::Bool # XXX propagate `interp::AbstractInterpreter` here? | |||
There was a problem hiding this comment.
I don't think adding this field here is reasonable. We should find some other way to propagate this information.
There was a problem hiding this comment.
Yeah.. Would it be too hacky to temporarily keep Expr(:assume_bindings_static, true) in meta::Vector{Expr}?
There was a problem hiding this comment.
I'm not very happy with that solution.
Assuming that some implicit subset of InferenceParams have been copied into the meta vector is very subtle and bakes implicit pipeline ordering into the construction of an IRCode.
For example, it looks like a caller of ssa_inlining_pass! will have their Interpreter's parameters ignored when stmt_effect_flags are computed internally (despite the fact that InliningState contains an interp which has params).
Is the caller supposed to provide a "meta-expanded" IRCode? How are they supposed to know which settings will be ignored?
aviatesk
force-pushed
the
avi/opt-assume-bindings-static
branch
2 times, most recently
from
cf8295d
to
f714901
Compare
|
@nanosoldier |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. |
aviatesk
force-pushed
the
avi/opt-assume-bindings-static
branch
from
f714901
to
94d07a6
Compare
There was a problem hiding this comment.
I really don't like how this means that our optimization passes now pull some information from InferenceParams(interp) and other information from secret/undocumented entries in the ir.meta Vector.
I think this needs a more significant re-work, so that we always have a coherent view of the lattice + inference parameters within any given function/pass.
|
Indeed, this could be a great chance to introduce something similar to LLVM's optimization manager, which would centralize the management of optimization passes. I'll start thinking about the design. |
Following up #53750. External abstract interpreters with `assume_bindings_static` enabled assume nothrow-ness of defined `GlobalRef`s no matter if they are constant during abstract interpretation. However, post-#53750, the optimizer’s verification switched to stricter checks, causing errors in such interpreters. This commit makes `IRCode` hold `assume_bindings_static::Bool` info to ensure that inlining and verification behaviors are synced with abstract interpretation.
aviatesk
force-pushed
the
avi/opt-assume-bindings-static
branch
from
94d07a6
to
40fb628
Compare
Following up #53750.
External abstract interpreters with
assume_bindings_staticenabled assume nothrow-ness of definedGlobalRefs no matter if they are constant during abstract interpretation. However, post-#53750, the optimizer’s verification switched to stricter checks, causing errors in such interpreters. This commit makesIRCodeholdassume_bindings_static::Boolinfo to ensure that inlining and verification behaviors are synced with abstract interpretation.