bjoernager/rust
bjoernager/rust
1
Fork 0
Code Activity

Prevent miscompilation in trivial loop {}

Browse source 
Ideally, we would want to handle a broader set of cases to fully fix the
underlying bug here. That is currently relatively expensive at compile and
runtime, so we don't do that for now.
This commit is contained in:
Mark Rousskov 2020-10-15 09:28:27 -04:00
parent 7f58716810
commit e2efec8976
5 changed files with 42 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

20
compiler/rustc_codegen_ssa/src/mir/block.rs
View file

@ -163,7 +163,7 @@ impl<'a, 'tcx> TerminatorCodegenHelper<'tcx> {
target <= self.bb
&& target.start_location().is_predecessor_of(self.bb.start_location(), mir)
}) {
bx.sideeffect();
bx.sideeffect(false);
}
}
}
@ -964,7 +964,23 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
}
mir::TerminatorKind::Goto { target } => {
helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
if bb == target {
// This is an unconditional branch back to this same basic
// block. That means we have something like a `loop {}`
// statement. Currently LLVM miscompiles this because it
// assumes forward progress. We want to prevent this in all
// cases, but that has a fairly high cost to compile times
// currently. Instead, try to handle this specific case
// which comes up commonly in practice (e.g., in embedded
// code).
//
// The `true` here means we insert side effects regardless
// of -Zinsert-sideeffect being passed on unconditional
// branching to the same basic block.
bx.sideeffect(true);
} else {
helper.maybe_sideeffect(self.mir, &mut bx, &[target]);
}
helper.funclet_br(self, &mut bx, target);
}