Auto merge of #77972 - Mark-Simulacrum:side-effect-loop, r=nagisa

Prevent miscompilation in trivial loop {}

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.

Performance results indicate this is not a major regression, if at all, so it should be safe to land.

cc #28728
This commit is contained in:
bors 2020-10-16 14:06:37 +00:00
commit a78a62fc99
5 changed files with 42 additions and 9 deletions

View File

@ -334,8 +334,8 @@ impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
self.call(expect, &[cond, self.const_bool(expected)], None)
}
fn sideeffect(&mut self) {
if self.tcx.sess.opts.debugging_opts.insert_sideeffect {
fn sideeffect(&mut self, unconditional: bool) {
if unconditional || self.tcx.sess.opts.debugging_opts.insert_sideeffect {
let fnname = self.get_intrinsic(&("llvm.sideeffect"));
self.call(fnname, &[], None);
}
@ -390,7 +390,7 @@ fn codegen_msvc_try(
) {
let llfn = get_rust_try_fn(bx, &mut |mut bx| {
bx.set_personality_fn(bx.eh_personality());
bx.sideeffect();
bx.sideeffect(false);
let mut normal = bx.build_sibling_block("normal");
let mut catchswitch = bx.build_sibling_block("catchswitch");
@ -553,7 +553,7 @@ fn codegen_gnu_try(
// call %catch_func(%data, %ptr)
// ret 1
bx.sideeffect();
bx.sideeffect(false);
let mut then = bx.build_sibling_block("then");
let mut catch = bx.build_sibling_block("catch");
@ -615,7 +615,7 @@ fn codegen_emcc_try(
// call %catch_func(%data, %catch_data)
// ret 1
bx.sideeffect();
bx.sideeffect(false);
let mut then = bx.build_sibling_block("then");
let mut catch = bx.build_sibling_block("catch");

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);
}

View File

@ -153,7 +153,7 @@ pub fn codegen_mir<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
bx.set_personality_fn(cx.eh_personality());
}
bx.sideeffect();
bx.sideeffect(false);
let cleanup_kinds = analyze::cleanup_kinds(&mir);
// Allocate a `Block` for every basic block, except

View File

@ -20,7 +20,9 @@ pub trait IntrinsicCallMethods<'tcx>: BackendTypes {
fn abort(&mut self);
fn assume(&mut self, val: Self::Value);
fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value;
fn sideeffect(&mut self);
/// Normally, sideeffect is only emitted if -Zinsert-sideeffect is passed;
/// in some cases though we want to emit it regardless.
fn sideeffect(&mut self, unconditional: bool);
/// Trait method used to inject `va_start` on the "spoofed" `VaListImpl` in
/// Rust defined C-variadic functions.
fn va_start(&mut self, val: Self::Value) -> Self::Value;

15
src/test/codegen/loop.rs Normal file
View File

@ -0,0 +1,15 @@
// compile-flags: -C opt-level=3
#![crate_type = "lib"]
// CHECK-LABEL: @check_loop
#[no_mangle]
pub fn check_loop() -> u8 {
// CHECK-NOT: unreachable
call_looper()
}
#[no_mangle]
fn call_looper() -> ! {
loop {}
}