bjoernager/rust
bjoernager/rust
1
Fork 0
Code Activity

Auto merge of #105119 - JakobDegen:inline-experiments, r=cjgillot

Browse source 
Disable top down MIR inlining

The current MIR inliner has exponential behavior in some cases: <https://godbolt.org/z/7jnWah4fE>. The cause of this is top-down inlining, where we repeatedly do inlining like `call_a() => { call_b(); call_b(); }`. Each decision on its own seems to make sense, but the result is exponential.

Disabling top-down inlining fundamentally prevents this. Each call site in the original, unoptimized source code is now considered for inlining exactly one time, which means that the total growth in MIR size is limited to number of call sites * inlining threshold.

Top down inlining may be worth re-introducing at some point, but it needs to be accompanied with a principled way to prevent this kind of behavior.
This commit is contained in:
bors 2022-12-06 00:53:01 +00:00
commit 226202d902
10 changed files with 245 additions and 117 deletions
Download patch file Download diff file Expand all files Collapse all files

28
compiler/rustc_mir_transform/src/inline.rs
View file

@ -8,7 +8,6 @@ use rustc_middle::mir::visit::*;
use rustc_middle::mir::*;
use rustc_middle::ty::{self, Instance, InstanceDef, ParamEnv, Ty, TyCtxt};
use rustc_session::config::OptLevel;
use rustc_span::def_id::DefId;
use rustc_span::{hygiene::ExpnKind, ExpnData, LocalExpnId, Span};
use rustc_target::abi::VariantIdx;
use rustc_target::spec::abi::Abi;
@ -87,13 +86,8 @@ fn inline<'tcx>(tcx: TyCtxt<'tcx>, body: &mut Body<'tcx>) -> bool {
let param_env = tcx.param_env_reveal_all_normalized(def_id);
let mut this = Inliner {
tcx,
param_env,
codegen_fn_attrs: tcx.codegen_fn_attrs(def_id),
history: Vec::new(),
changed: false,
};
let mut this =
Inliner { tcx, param_env, codegen_fn_attrs: tcx.codegen_fn_attrs(def_id), changed: false };
let blocks = BasicBlock::new(0)..body.basic_blocks.next_index();
this.process_blocks(body, blocks);
this.changed
@ -104,12 +98,6 @@ struct Inliner<'tcx> {
param_env: ParamEnv<'tcx>,
/// Caller codegen attributes.
codegen_fn_attrs: &'tcx CodegenFnAttrs,
/// Stack of inlined instances.
/// We only check the `DefId` and not the substs because we want to
/// avoid inlining cases of polymorphic recursion.
/// The number of `DefId`s is finite, so checking history is enough
/// to ensure that we do not loop endlessly while inlining.
history: Vec<DefId>,
/// Indicates that the caller body has been modified.
changed: bool,
}
@ -134,12 +122,12 @@ impl<'tcx> Inliner<'tcx> {
debug!("not-inlined {} [{}]", callsite.callee, reason);
continue;
}
Ok(new_blocks) => {
Ok(_) => {
debug!("inlined {}", callsite.callee);
self.changed = true;
self.history.push(callsite.callee.def_id());
self.process_blocks(caller_body, new_blocks);
self.history.pop();
// We could process the blocks returned by `try_inlining` here. However, that
// leads to exponential compile times due to the top-down nature of this kind
// of inlining.
}
}
}
@ -313,10 +301,6 @@ impl<'tcx> Inliner<'tcx> {
return None;
}
if self.history.contains(&callee.def_id()) {
return None;
}
let fn_sig = self.tcx.bound_fn_sig(def_id).subst(self.tcx, substs);
return Some(CallSite {