Auto merge of #150605 - RalfJung:fallback-intrinsic-skip, r=mati865

skip codegen for intrinsics with big fallback bodies if backend does not need them

This hopefully fixes the perf regression from #148478. I only added the intrinsics with big fallback bodies to the list; it doesn't seem worth the effort of going through the entire list.

Fixes #149945
Cc @scottmcm @bjorn3

Author: bors

compiler/rustc_codegen_llvm/src/lib.rs

@@ -43,7 +43,7 @@ use rustc_middle::ty::TyCtxt;
 use rustc_middle::util::Providers;
 use rustc_session::Session;
 use rustc_session::config::{OptLevel, OutputFilenames, PrintKind, PrintRequest};
-use rustc_span::Symbol;
+use rustc_span::{Symbol, sym};
 use rustc_target::spec::{RelocModel, TlsModel};
 
 use crate::llvm::ToLlvmBool;
@@ -344,6 +344,10 @@ impl CodegenBackend for LlvmCodegenBackend {
         target_config(sess)
     }
 
+    fn replaced_intrinsics(&self) -> Vec<Symbol> {
+        vec![sym::unchecked_funnel_shl, sym::unchecked_funnel_shr, sym::carrying_mul_add]
+    }
+
     fn codegen_crate<'tcx>(&self, tcx: TyCtxt<'tcx>) -> Box<dyn Any> {
         Box::new(rustc_codegen_ssa::base::codegen_crate(
             LlvmCodegenBackend(()),

compiler/rustc_codegen_ssa/src/traits/backend.rs

@@ -74,6 +74,12 @@ pub trait CodegenBackend {
 
     fn print_version(&self) {}
 
+    /// Returns a list of all intrinsics that this backend definitely
+    /// replaces, which means their fallback bodies do not need to be monomorphized.
+    fn replaced_intrinsics(&self) -> Vec<Symbol> {
+        vec![]
+    }
+
     /// Value printed by `--print=backend-has-zstd`.
     ///
     /// Used by compiletest to determine whether tests involving zstd compression

compiler/rustc_interface/src/interface.rs

@@ -474,6 +474,7 @@ pub fn run_compiler<R: Send>(config: Config, f: impl FnOnce(&Compiler) -> R + Se
             );
 
             codegen_backend.init(&sess);
+            sess.replaced_intrinsics = FxHashSet::from_iter(codegen_backend.replaced_intrinsics());
 
             let cfg = parse_cfg(sess.dcx(), config.crate_cfg);
             let mut cfg = config::build_configuration(&sess, cfg);

compiler/rustc_monomorphize/src/collector.rs

@@ -1002,11 +1002,12 @@ fn visit_instance_use<'tcx>(
             if tcx.should_codegen_locally(panic_instance) {
                 output.push(create_fn_mono_item(tcx, panic_instance, source));
             }
-        } else if !intrinsic.must_be_overridden {
+        } else if !intrinsic.must_be_overridden
+            && !tcx.sess.replaced_intrinsics.contains(&intrinsic.name)
+        {
             // Codegen the fallback body of intrinsics with fallback bodies.
-            // We explicitly skip this otherwise to ensure we get a linker error
-            // if anyone tries to call this intrinsic and the codegen backend did not
-            // override the implementation.
+            // We have to skip this otherwise as there's no body to codegen.
+            // We also skip intrinsics the backend handles, to reduce monomorphizations.
             let instance = ty::Instance::new_raw(instance.def_id(), instance.args);
             if tcx.should_codegen_locally(instance) {
                 output.push(create_fn_mono_item(tcx, instance, source));

compiler/rustc_session/src/session.rs

@@ -8,7 +8,7 @@ use std::{env, io};
 use rand::{RngCore, rng};
 use rustc_data_structures::base_n::{CASE_INSENSITIVE, ToBaseN};
 use rustc_data_structures::flock;
-use rustc_data_structures::fx::{FxHashMap, FxIndexSet};
+use rustc_data_structures::fx::{FxHashMap, FxHashSet, FxIndexSet};
 use rustc_data_structures::profiling::{SelfProfiler, SelfProfilerRef};
 use rustc_data_structures::sync::{DynSend, DynSync, Lock, MappedReadGuard, ReadGuard, RwLock};
 use rustc_errors::annotate_snippet_emitter_writer::AnnotateSnippetEmitter;
@@ -154,6 +154,10 @@ pub struct Session {
     /// preserved with a flag like `-C save-temps`, since these files may be
     /// hard linked.
     pub invocation_temp: Option<String>,
+
+    /// The names of intrinsics that the current codegen backend replaces
+    /// with its own implementations.
+    pub replaced_intrinsics: FxHashSet<Symbol>,
 }
 
 #[derive(Clone, Copy)]
@@ -1092,6 +1096,7 @@ pub fn build_session(
         target_filesearch,
         host_filesearch,
         invocation_temp,
+        replaced_intrinsics: FxHashSet::default(), // filled by `run_compiler`
     };
 
     validate_commandline_args_with_session_available(&sess);

tests/codegen-llvm/intrinsics/carrying_mul_add.rs

@@ -11,10 +11,9 @@
 
 use std::intrinsics::{carrying_mul_add, fallback};
 
-// The fallbacks are emitted even when they're never used, but optimize out.
+// The fallbacks should not be emitted.
 
-// RAW: wide_mul_u128
-// OPT-NOT: wide_mul_u128
+// NOT: wide_mul_u128
 
 // CHECK-LABEL: @cma_u8
 #[no_mangle]