Allow more top-down inlining for single-BB callees

This means that things like `<usize as Step>::forward_unchecked` and `<PartialOrd for f32>::le` will inline even if we've already done a bunch of inlining to find the calls to them.

Author: scottmcm

compiler/rustc_mir_transform/src/inline.rs

@@ -26,6 +26,7 @@ use crate::{check_inline, util};
 
 pub(crate) mod cycle;
 
+const HISTORY_DEPTH_LIMIT: usize = 20;
 const TOP_DOWN_DEPTH_LIMIT: usize = 5;
 
 #[derive(Clone, Debug)]
@@ -117,6 +118,13 @@ trait Inliner<'tcx> {
     /// Should inlining happen for a given callee?
     fn should_inline_for_callee(&self, def_id: DefId) -> bool;
 
+    fn check_codegen_attributes_extra(
+        &self,
+        _callee_attrs: &CodegenFnAttrs,
+    ) -> Result<(), &'static str> {
+        Ok(())
+    }
+
     fn check_caller_mir_body(&self, body: &Body<'tcx>) -> bool;
 
     /// Returns inlining decision that is based on the examination of callee MIR body.
@@ -128,10 +136,6 @@ trait Inliner<'tcx> {
         callee_attrs: &CodegenFnAttrs,
     ) -> Result<(), &'static str>;
 
-    // How many callsites in a body are we allowed to inline? We need to limit this in order
-    // to prevent super-linear growth in MIR size.
-    fn inline_limit_for_block(&self) -> Option<usize>;
-
     /// Called when inlining succeeds.
     fn on_inline_success(
         &mut self,
@@ -142,9 +146,6 @@ trait Inliner<'tcx> {
 
     /// Called when inlining failed or was not performed.
     fn on_inline_failure(&self, callsite: &CallSite<'tcx>, reason: &'static str);
-
-    /// Called when the inline limit for a body is reached.
-    fn on_inline_limit_reached(&self) -> bool;
 }
 
 struct ForceInliner<'tcx> {
@@ -224,10 +225,6 @@ impl<'tcx> Inliner<'tcx> for ForceInliner<'tcx> {
         }
     }
 
-    fn inline_limit_for_block(&self) -> Option<usize> {
-        Some(usize::MAX)
-    }
-
     fn on_inline_success(
         &mut self,
         callsite: &CallSite<'tcx>,
@@ -261,10 +258,6 @@ impl<'tcx> Inliner<'tcx> for ForceInliner<'tcx> {
             justification: justification.map(|sym| crate::errors::ForceInlineJustification { sym }),
         });
     }
-
-    fn on_inline_limit_reached(&self) -> bool {
-        false
-    }
 }
 
 struct NormalInliner<'tcx> {
@@ -278,13 +271,23 @@ struct NormalInliner<'tcx> {
     /// 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>,
+    /// How many (multi-call) callsites have we inlined for the top-level call?
+    ///
+    /// We need to limit this in order to prevent super-linear growth in MIR size.
+    top_down_counter: usize,
     /// Indicates that the caller body has been modified.
     changed: bool,
     /// Indicates that the caller is #[inline] and just calls another function,
     /// and thus we can inline less into it as it'll be inlined itself.
     caller_is_inline_forwarder: bool,
 }
 
+impl<'tcx> NormalInliner<'tcx> {
+    fn past_depth_limit(&self) -> bool {
+        self.history.len() > HISTORY_DEPTH_LIMIT || self.top_down_counter > TOP_DOWN_DEPTH_LIMIT
+    }
+}
+
 impl<'tcx> Inliner<'tcx> for NormalInliner<'tcx> {
     fn new(tcx: TyCtxt<'tcx>, def_id: DefId, body: &Body<'tcx>) -> Self {
         let typing_env = body.typing_env(tcx);
@@ -295,6 +298,7 @@ impl<'tcx> Inliner<'tcx> for NormalInliner<'tcx> {
             typing_env,
             def_id,
             history: Vec::new(),
+            top_down_counter: 0,
             changed: false,
             caller_is_inline_forwarder: matches!(
                 codegen_fn_attrs.inline,
@@ -327,6 +331,17 @@ impl<'tcx> Inliner<'tcx> for NormalInliner<'tcx> {
         true
     }
 
+    fn check_codegen_attributes_extra(
+        &self,
+        callee_attrs: &CodegenFnAttrs,
+    ) -> Result<(), &'static str> {
+        if self.past_depth_limit() && matches!(callee_attrs.inline, InlineAttr::None) {
+            Err("Past depth limit so not inspecting unmarked callee")
+        } else {
+            Ok(())
+        }
+    }
+
     fn check_caller_mir_body(&self, body: &Body<'tcx>) -> bool {
         // Avoid inlining into coroutines, since their `optimized_mir` is used for layout computation,
         // which can create a cycle, even when no attempt is made to inline the function in the other
@@ -351,6 +366,10 @@ impl<'tcx> Inliner<'tcx> for NormalInliner<'tcx> {
             return Err("body has errors");
         }
 
+        if self.past_depth_limit() && callee_body.basic_blocks.len() > 1 {
+            return Err("Not inlining multi-block body as we're past a depth limit");
+        }
+
         let mut threshold = if self.caller_is_inline_forwarder {
             tcx.sess.opts.unstable_opts.inline_mir_forwarder_threshold.unwrap_or(30)
         } else if tcx.cross_crate_inlinable(callsite.callee.def_id()) {
@@ -431,14 +450,6 @@ impl<'tcx> Inliner<'tcx> for NormalInliner<'tcx> {
         }
     }
 
-    fn inline_limit_for_block(&self) -> Option<usize> {
-        match self.history.len() {
-            0 => Some(usize::MAX),
-            1..=TOP_DOWN_DEPTH_LIMIT => Some(1),
-            _ => None,
-        }
-    }
-
     fn on_inline_success(
         &mut self,
         callsite: &CallSite<'tcx>,
@@ -447,13 +458,26 @@ impl<'tcx> Inliner<'tcx> for NormalInliner<'tcx> {
     ) {
         self.changed = true;
 
+        let new_calls_count = new_blocks
+            .clone()
+            .filter(|&bb| {
+                matches!(
+                    caller_body.basic_blocks[bb].terminator().kind,
+                    TerminatorKind::Call { .. },
+                )
+            })
+            .count();
+        if new_calls_count > 1 {
+            self.top_down_counter += 1;
+        }
+
         self.history.push(callsite.callee.def_id());
         process_blocks(self, caller_body, new_blocks);
         self.history.pop();
-    }
 
-    fn on_inline_limit_reached(&self) -> bool {
-        true
+        if self.history.is_empty() {
+            self.top_down_counter = 0;
+        }
     }
 
     fn on_inline_failure(&self, _: &CallSite<'tcx>, _: &'static str) {}
@@ -482,8 +506,6 @@ fn process_blocks<'tcx, I: Inliner<'tcx>>(
     caller_body: &mut Body<'tcx>,
     blocks: Range<BasicBlock>,
 ) {
-    let Some(inline_limit) = inliner.inline_limit_for_block() else { return };
-    let mut inlined_count = 0;
     for bb in blocks {
         let bb_data = &caller_body[bb];
         if bb_data.is_cleanup {
@@ -505,13 +527,6 @@ fn process_blocks<'tcx, I: Inliner<'tcx>>(
             Ok(new_blocks) => {
                 debug!("inlined {}", callsite.callee);
                 inliner.on_inline_success(&callsite, caller_body, new_blocks);
-
-                inlined_count += 1;
-                if inlined_count == inline_limit {
-                    if inliner.on_inline_limit_reached() {
-                        return;
-                    }
-                }
             }
         }
     }
@@ -584,6 +599,7 @@ fn try_inlining<'tcx, I: Inliner<'tcx>>(
     let callee_attrs = tcx.codegen_fn_attrs(callsite.callee.def_id());
     check_inline::is_inline_valid_on_fn(tcx, callsite.callee.def_id())?;
     check_codegen_attributes(inliner, callsite, callee_attrs)?;
+    inliner.check_codegen_attributes_extra(callee_attrs)?;
 
     let terminator = caller_body[callsite.block].terminator.as_ref().unwrap();
     let TerminatorKind::Call { args, destination, .. } = &terminator.kind else { bug!() };
@@ -770,6 +786,8 @@ fn check_codegen_attributes<'tcx, I: Inliner<'tcx>>(
         return Err("has DoNotOptimize attribute");
     }
 
+    inliner.check_codegen_attributes_extra(callee_attrs)?;
+
     // Reachability pass defines which functions are eligible for inlining. Generally inlining
     // other functions is incorrect because they could reference symbols that aren't exported.
     let is_generic = callsite.callee.args.non_erasable_generics().next().is_some();

tests/mir-opt/inline/exponential_runtime.rs

@@ -87,10 +87,15 @@ fn main() {
     // CHECK-LABEL: fn main(
     // CHECK-NOT: inlined
     // CHECK: (inlined <() as G>::call)
+    // CHECK-NOT: inlined
     // CHECK: (inlined <() as F>::call)
+    // CHECK-NOT: inlined
     // CHECK: (inlined <() as E>::call)
+    // CHECK-NOT: inlined
     // CHECK: (inlined <() as D>::call)
+    // CHECK-NOT: inlined
     // CHECK: (inlined <() as C>::call)
+    // CHECK-NOT: inlined
     // CHECK: (inlined <() as B>::call)
     // CHECK-NOT: inlined
     <() as G>::call();

tests/mir-opt/inline/inline_diverging.h.Inline.panic-abort.diff

@@ -6,6 +6,7 @@
       let _1: (!, !);
 +     let mut _2: fn() -> ! {sleep};
 +     let mut _7: ();
++     let mut _8: ();
 +     scope 1 (inlined call_twice::<!, fn() -> ! {sleep}>) {
 +         debug f => _2;
 +         let mut _3: &fn() -> ! {sleep};
@@ -17,6 +18,10 @@
 +             scope 3 {
 +                 debug b => _6;
 +             }
++             scope 6 (inlined <fn() -> ! {sleep} as Fn<()>>::call - shim(fn() -> ! {sleep})) {
++                 scope 7 (inlined sleep) {
++                 }
++             }
 +         }
 +         scope 4 (inlined <fn() -> ! {sleep} as Fn<()>>::call - shim(fn() -> ! {sleep})) {
 +             scope 5 (inlined sleep) {

tests/mir-opt/inline/inline_diverging.h.Inline.panic-unwind.diff

@@ -6,6 +6,7 @@
       let _1: (!, !);
 +     let mut _2: fn() -> ! {sleep};
 +     let mut _8: ();
++     let mut _9: ();
 +     scope 1 (inlined call_twice::<!, fn() -> ! {sleep}>) {
 +         debug f => _2;
 +         let mut _3: &fn() -> ! {sleep};
@@ -18,6 +19,10 @@
 +             scope 3 {
 +                 debug b => _6;
 +             }
++             scope 6 (inlined <fn() -> ! {sleep} as Fn<()>>::call - shim(fn() -> ! {sleep})) {
++                 scope 7 (inlined sleep) {
++                 }
++             }
 +         }
 +         scope 4 (inlined <fn() -> ! {sleep} as Fn<()>>::call - shim(fn() -> ! {sleep})) {
 +             scope 5 (inlined sleep) {