Skip to content

Mark drop calls in landing pads cold instead of noinline #92419

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Jan 1, 2022
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion compiler/rustc_codegen_gcc/src/builder.rs
Original file line number Diff line number Diff line change
Expand Up @@ -1404,7 +1404,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
self.cx
}

fn do_not_inline(&mut self, _llret: RValue<'gcc>) {
fn apply_attrs_to_cleanup_callsite(&mut self, _llret: RValue<'gcc>) {
unimplemented!();
}

Expand Down
11 changes: 9 additions & 2 deletions compiler/rustc_codegen_llvm/src/builder.rs
Original file line number Diff line number Diff line change
Expand Up @@ -1201,8 +1201,15 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
unsafe { llvm::LLVMBuildZExt(self.llbuilder, val, dest_ty, UNNAMED) }
}

fn do_not_inline(&mut self, llret: &'ll Value) {
llvm::Attribute::NoInline.apply_callsite(llvm::AttributePlace::Function, llret);
fn apply_attrs_to_cleanup_callsite(&mut self, llret: &'ll Value) {
// Cleanup is always the cold path.
llvm::Attribute::Cold.apply_callsite(llvm::AttributePlace::Function, llret);

// In LLVM versions with deferred inlining (currently, system LLVM < 14),
// inlining drop glue can lead to exponential size blowup, see #41696 and #92110.
if !llvm_util::is_rust_llvm() && llvm_util::get_version() < (14, 0, 0) {
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Checking for non-system llvm feels a bit hacky, not sure if I should just wait for the LLVM 14 release.

(this code is also assuming that https://reviews.llvm.org/D115497 won't be reverted upstream before then)

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This seems sufficiently reasonable thing to do to me if this does in fact yield any significant benefits. I wouldn't want to land changes that potentially significantly affect performance at some time in the future as it would make those other changes non-representative on their own merits.

llvm::Attribute::NoInline.apply_callsite(llvm::AttributePlace::Function, llret);
}
}
}

Expand Down
2 changes: 2 additions & 0 deletions compiler/rustc_codegen_llvm/src/llvm/ffi.rs
Original file line number Diff line number Diff line change
Expand Up @@ -1902,6 +1902,8 @@ extern "C" {
pub fn LLVMRustVersionMinor() -> u32;
pub fn LLVMRustVersionPatch() -> u32;

pub fn LLVMRustIsRustLLVM() -> bool;

pub fn LLVMRustAddModuleFlag(M: &Module, name: *const c_char, value: u32);

pub fn LLVMRustMetadataAsValue<'a>(C: &'a Context, MD: &'a Metadata) -> &'a Value;
Expand Down
6 changes: 6 additions & 0 deletions compiler/rustc_codegen_llvm/src/llvm_util.rs
Original file line number Diff line number Diff line change
Expand Up @@ -217,6 +217,12 @@ pub fn get_version() -> (u32, u32, u32) {
}
}

/// Returns `true` if this LLVM is Rust's bundled LLVM (and not system LLVM).
pub fn is_rust_llvm() -> bool {
// Can be called without initializing LLVM
unsafe { llvm::LLVMRustIsRustLLVM() }
}

pub fn print_passes() {
// Can be called without initializing LLVM
unsafe {
Expand Down
6 changes: 1 addition & 5 deletions compiler/rustc_codegen_ssa/src/mir/block.rs
Original file line number Diff line number Diff line change
Expand Up @@ -160,11 +160,7 @@ impl<'a, 'tcx> TerminatorCodegenHelper<'tcx> {
let llret = bx.call(fn_ty, fn_ptr, &llargs, self.funclet(fx));
bx.apply_attrs_callsite(&fn_abi, llret);
if fx.mir[self.bb].is_cleanup {
// Cleanup is always the cold path. Don't inline
// drop glue. Also, when there is a deeply-nested
// struct, there are "symmetry" issues that cause
// exponential inlining - see issue #41696.
bx.do_not_inline(llret);
bx.apply_attrs_to_cleanup_callsite(llret);
}

if let Some((ret_dest, target)) = destination {
Expand Down
2 changes: 1 addition & 1 deletion compiler/rustc_codegen_ssa/src/traits/builder.rs
Original file line number Diff line number Diff line change
Expand Up @@ -311,5 +311,5 @@ pub trait BuilderMethods<'a, 'tcx>:
) -> Self::Value;
fn zext(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value;

fn do_not_inline(&mut self, llret: Self::Value);
fn apply_attrs_to_cleanup_callsite(&mut self, llret: Self::Value);
}
8 changes: 8 additions & 0 deletions compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -716,6 +716,14 @@ extern "C" uint32_t LLVMRustVersionMinor() { return LLVM_VERSION_MINOR; }

extern "C" uint32_t LLVMRustVersionMajor() { return LLVM_VERSION_MAJOR; }

extern "C" bool LLVMRustIsRustLLVM() {
#ifdef LLVM_RUSTLLVM
return true;
#else
return false;
#endif
}

extern "C" void LLVMRustAddModuleFlag(LLVMModuleRef M, const char *Name,
uint32_t Value) {
unwrap(M)->addModuleFlag(Module::Warning, Name, Value);
Expand Down
15 changes: 15 additions & 0 deletions src/test/codegen/unwind-landingpad-cold.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,15 @@
// no-system-llvm: needs #92110
// compile-flags: -Cno-prepopulate-passes
#![crate_type = "lib"]

// This test checks that drop calls in unwind landing pads
// get the `cold` attribute.

// CHECK-LABEL: @check_cold
// CHECK: call void {{.+}}drop_in_place{{.+}} [[ATTRIBUTES:#[0-9]+]]
// CHECK: attributes [[ATTRIBUTES]] = { cold }
#[no_mangle]
pub fn check_cold(f: fn(), x: Box<u32>) {
// this may unwind
f();
}
37 changes: 37 additions & 0 deletions src/test/codegen/unwind-landingpad-inline.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,37 @@
// no-system-llvm: needs #92110 + patch for Rust alloc/dealloc functions
// compile-flags: -Copt-level=3
#![crate_type = "lib"]

// This test checks that we can inline drop_in_place in
// unwind landing pads.

// Without inlining, the box pointers escape via the call to drop_in_place,
// and LLVM will not optimize out the pointer comparison.
// With inlining, everything should be optimized out.
// See https://github.com/rust-lang/rust/issues/46515
// CHECK-LABEL: @check_no_escape_in_landingpad
// CHECK: start:
// CHECK-NEXT: ret void
#[no_mangle]
pub fn check_no_escape_in_landingpad(f: fn()) {
let x = &*Box::new(0);
let y = &*Box::new(0);

if x as *const _ == y as *const _ {
f();
}
}

// Without inlining, the compiler can't tell that
// dropping an empty string (in a landing pad) does nothing.
// With inlining, the landing pad should be optimized out.
// See https://github.com/rust-lang/rust/issues/87055
// CHECK-LABEL: @check_eliminate_noop_drop
// CHECK: start:
// CHECK-NEXT: call void %g()
// CHECK-NEXT: ret void
#[no_mangle]
pub fn check_eliminate_noop_drop(g: fn()) {
let _var = String::new();
g();
}