Remove no-op cleanups on MIR post-monomorphization

This speeds up LLVM and improves codegen overall. As an example, for
cargo this cuts ~5% of the LLVM IR lines we generate (measured with
-Cno-prepopulate-passes).

Author: Mark-Simulacrum

Cargo.lock

@@ -3505,6 +3505,7 @@ dependencies = [
  "rustc_macros",
  "rustc_metadata",
  "rustc_middle",
+ "rustc_mir_transform",
  "rustc_query_system",
  "rustc_serialize",
  "rustc_session",

compiler/rustc_codegen_ssa/Cargo.toml

@@ -37,6 +37,7 @@ rustc_span = { path = "../rustc_span" }
 rustc_symbol_mangling = { path = "../rustc_symbol_mangling" }
 rustc_target = { path = "../rustc_target" }
 rustc_trait_selection = { path = "../rustc_trait_selection" }
+rustc_mir_transform = { path = "../rustc_mir_transform" }
 serde_json = "1.0.59"
 smallvec = { version = "1.8.1", features = ["union", "may_dangle"] }
 tempfile = "3.2"

compiler/rustc_codegen_ssa/src/mir/analyze.rs

@@ -279,10 +279,14 @@ impl CleanupKind {
 /// MSVC requires unwinding code to be split to a tree of *funclets*, where each funclet can only
 /// branch to itself or to its parent. Luckily, the code we generates matches this pattern.
 /// Recover that structure in an analyze pass.
-pub(crate) fn cleanup_kinds(mir: &mir::Body<'_>) -> IndexVec<mir::BasicBlock, CleanupKind> {
+pub(crate) fn cleanup_kinds(
+    mir: &mir::Body<'_>,
+    nop_landing_pads: &DenseBitSet<mir::BasicBlock>,
+) -> IndexVec<mir::BasicBlock, CleanupKind> {
     fn discover_masters<'tcx>(
         result: &mut IndexSlice<mir::BasicBlock, CleanupKind>,
         mir: &mir::Body<'tcx>,
+        nop_landing_pads: &DenseBitSet<mir::BasicBlock>,
     ) {
         for (bb, data) in mir.basic_blocks.iter_enumerated() {
             match data.terminator().kind {
@@ -301,7 +305,9 @@ pub(crate) fn cleanup_kinds(mir: &mir::Body<'_>) -> IndexVec<mir::BasicBlock, Cl
                 | TerminatorKind::InlineAsm { unwind, .. }
                 | TerminatorKind::Assert { unwind, .. }
                 | TerminatorKind::Drop { unwind, .. } => {
-                    if let mir::UnwindAction::Cleanup(unwind) = unwind {
+                    if let mir::UnwindAction::Cleanup(unwind) = unwind
+                        && !nop_landing_pads.contains(unwind)
+                    {
                         debug!(
                             "cleanup_kinds: {:?}/{:?} registering {:?} as funclet",
                             bb, data, unwind
@@ -382,7 +388,7 @@ pub(crate) fn cleanup_kinds(mir: &mir::Body<'_>) -> IndexVec<mir::BasicBlock, Cl
 
     let mut result = IndexVec::from_elem(CleanupKind::NotCleanup, &mir.basic_blocks);
 
-    discover_masters(&mut result, mir);
+    discover_masters(&mut result, mir, &nop_landing_pads);
     propagate(&mut result, mir);
     debug!("cleanup_kinds: result={:?}", result);
     result

compiler/rustc_codegen_ssa/src/mir/block.rs

@@ -200,7 +200,13 @@ impl<'a, 'tcx> TerminatorCodegenHelper<'tcx> {
         }
 
         let unwind_block = match unwind {
-            mir::UnwindAction::Cleanup(cleanup) => Some(self.llbb_with_cleanup(fx, cleanup)),
+            mir::UnwindAction::Cleanup(cleanup) => {
+                if !fx.nop_landing_pads.contains(cleanup) {
+                    Some(self.llbb_with_cleanup(fx, cleanup))
+                } else {
+                    None
+                }
+            }
             mir::UnwindAction::Continue => None,
             mir::UnwindAction::Unreachable => None,
             mir::UnwindAction::Terminate(reason) => {
@@ -286,7 +292,13 @@ impl<'a, 'tcx> TerminatorCodegenHelper<'tcx> {
         mergeable_succ: bool,
     ) -> MergingSucc {
         let unwind_target = match unwind {
-            mir::UnwindAction::Cleanup(cleanup) => Some(self.llbb_with_cleanup(fx, cleanup)),
+            mir::UnwindAction::Cleanup(cleanup) => {
+                if !fx.nop_landing_pads.contains(cleanup) {
+                    Some(self.llbb_with_cleanup(fx, cleanup))
+                } else {
+                    None
+                }
+            }
             mir::UnwindAction::Terminate(reason) => Some(fx.terminate_block(reason)),
             mir::UnwindAction::Continue => None,
             mir::UnwindAction::Unreachable => None,

compiler/rustc_codegen_ssa/src/mir/mod.rs

@@ -97,6 +97,8 @@ pub struct FunctionCx<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> {
     /// A cold block is a block that is unlikely to be executed at runtime.
     cold_blocks: IndexVec<mir::BasicBlock, bool>,
 
+    nop_landing_pads: DenseBitSet<mir::BasicBlock>,
+
     /// The location where each MIR arg/var/tmp/ret is stored. This is
     /// usually an `PlaceRef` representing an alloca, but not always:
     /// sometimes we can skip the alloca and just store the value
@@ -176,8 +178,14 @@ pub fn codegen_mir<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 
     let mut mir = tcx.instance_mir(instance.def);
 
-    let fn_abi = cx.fn_abi_of_instance(instance, ty::List::empty());
-    debug!("fn_abi: {:?}", fn_abi);
+    let nop_landing_pads = rustc_mir_transform::remove_noop_landing_pads::find_noop_landing_pads(
+        mir,
+        Some(rustc_mir_transform::remove_noop_landing_pads::ExtraInfo {
+            tcx,
+            instance,
+            typing_env: cx.typing_env(),
+        }),
+    );
 
     if tcx.features().ergonomic_clones() {
         let monomorphized_mir = instance.instantiate_mir_and_normalize_erasing_regions(
@@ -188,19 +196,23 @@ pub fn codegen_mir<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
         mir = tcx.arena.alloc(optimize_use_clone::<Bx>(cx, monomorphized_mir));
     }
 
+    let fn_abi = cx.fn_abi_of_instance(instance, ty::List::empty());
+    debug!("fn_abi: {:?}", fn_abi);
+
     let debug_context = cx.create_function_debug_context(instance, fn_abi, llfn, &mir);
 
     let start_llbb = Bx::append_block(cx, llfn, "start");
     let mut start_bx = Bx::build(cx, start_llbb);
 
-    if mir.basic_blocks.iter().any(|bb| {
-        bb.is_cleanup || matches!(bb.terminator().unwind(), Some(mir::UnwindAction::Terminate(_)))
+    if mir::traversal::mono_reachable(&mir, tcx, instance).any(|(bb, block)| {
+        (block.is_cleanup && !nop_landing_pads.contains(bb))
+            || matches!(block.terminator().unwind(), Some(mir::UnwindAction::Terminate(_)))
     }) {
         start_bx.set_personality_fn(cx.eh_personality());
     }
 
-    let cleanup_kinds =
-        base::wants_new_eh_instructions(tcx.sess).then(|| analyze::cleanup_kinds(&mir));
+    let cleanup_kinds = base::wants_new_eh_instructions(tcx.sess)
+        .then(|| analyze::cleanup_kinds(&mir, &nop_landing_pads));
 
     let cached_llbbs: IndexVec<mir::BasicBlock, CachedLlbb<Bx::BasicBlock>> =
         mir.basic_blocks
@@ -228,6 +240,7 @@ pub fn codegen_mir<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
         debug_context,
         per_local_var_debug_info: None,
         caller_location: None,
+        nop_landing_pads,
     };
 
     // It may seem like we should iterate over `required_consts` to ensure they all successfully
@@ -239,7 +252,36 @@ pub fn codegen_mir<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
         fx.compute_per_local_var_debug_info(&mut start_bx).unzip();
     fx.per_local_var_debug_info = per_local_var_debug_info;
 
-    let traversal_order = traversal::mono_reachable_reverse_postorder(mir, tcx, instance);
+    let mut traversal_order = traversal::mono_reachable_reverse_postorder(mir, tcx, instance);
+
+    // Filter out blocks that won't be codegen'd because of nop_landing_pads optimization.
+    // FIXME: We might want to integrate the nop_landing_pads analysis into mono reachability.
+    {
+        let mut reachable = DenseBitSet::new_empty(mir.basic_blocks.len());
+        let mut to_visit = vec![mir::START_BLOCK];
+        while let Some(next) = to_visit.pop() {
+            if !reachable.insert(next) {
+                continue;
+            }
+
+            let block = &mir.basic_blocks[next];
+            if let Some(mir::UnwindAction::Cleanup(target)) = block.terminator().unwind()
+                && fx.nop_landing_pads.contains(*target)
+            {
+                // This edge will not be followed when we actually codegen, so skip generating it here.
+                //
+                // It's guaranteed that the cleanup block (`target`) occurs only in
+                // UnwindAction::Cleanup(...) -- i.e., we can't incorrectly filter too much here --
+                // because cleanup transitions must happen via UnwindAction::Cleanup.
+                to_visit.extend(block.terminator().successors().filter(|s| s != target));
+            } else {
+                to_visit.extend(block.terminator().successors());
+            }
+        }
+
+        traversal_order.retain(|bb| reachable.contains(*bb));
+    }
+
     let memory_locals = analyze::non_ssa_locals(&fx, &traversal_order);
 
     // Allocate variable and temp allocas

compiler/rustc_mir_transform/src/lib.rs

@@ -158,7 +158,7 @@ declare_passes! {
     mod prettify : ReorderBasicBlocks, ReorderLocals;
     mod promote_consts : PromoteTemps;
     mod ref_prop : ReferencePropagation;
-    mod remove_noop_landing_pads : RemoveNoopLandingPads;
+    pub mod remove_noop_landing_pads : RemoveNoopLandingPads;
     mod remove_place_mention : RemovePlaceMention;
     mod remove_storage_markers : RemoveStorageMarkers;
     mod remove_uninit_drops : RemoveUninitDrops;