Profiling and Performance

This guide covers profiling the Rust compiler to understand and improve its performance, including compilation speed, memory usage, and runtime efficiency.

Why Profile the Compiler?

Compilation Speed

Reduce time spent compiling Rust code

Memory Usage

Minimize compiler memory consumption

Query Optimization

Optimize the query system for better caching

Code Quality

Improve generated code performance

Profiling Setup

Build for profiling

Configure config.toml for profiling:

[rust]
# Enable profiler runtime
profiler = true

# Debug info for profiling
debuginfo-level = 1
debuginfo-level-rustc = 1

# Optimize for profiling
optimize = true

# Keep frame pointers for better stack traces
frame-pointers = true

Build the compiler

# Build optimized compiler with profiling support
./x.py build --stage 1

# Or build specific components
./x.py build compiler/rustc_driver

Install profiling tools

Linux
macOS
Windows

# perf
sudo apt-get install linux-tools-common linux-tools-generic

# flamegraph
cargo install flamegraph

# measureme
cargo install measureme

# samply
cargo install samply

# Instruments (Xcode)
xcode-select --install

# cargo-instruments
cargo install cargo-instruments

# measureme
cargo install measureme

# samply
cargo install samply

# measureme
cargo install measureme

# Windows Performance Analyzer
# Download from Microsoft

# samply
cargo install samply

Self-Profiling with measureme

Rustc has built-in self-profiling using the measureme framework:

Basic Self-Profiling

Enable Self-Profiling
Profile Events
Analyze Results

# Profile a compilation
rustc -Z self-profile file.rs

# This creates file-pid.mm_profdata
# Profile with custom name
rustc -Z self-profile=my_profile file.rs

# Profile specific events
rustc -Z self-profile-events=default,args file.rs

# Available event types:
# - default: basic profiling
# - query-cache-hits: query cache statistics  
# - query-blocked: query blocking information
# - incr-result-hashing: incremental hash computation
# - args: function arguments

# Install analysis tools
cargo install measureme

# Summarize results
summarize file-pid.mm_profdata

# Generate flamegraph
flamegraph file-pid.mm_profdata

# Generate Chrome trace
crox file-pid.mm_profdata
# Open chrome://tracing and load file.json

Analyzing Self-Profile Data

# Get summary statistics
summarize file-pid.mm_profdata

# Output shows:
# - Total time
# - Time per query
# - Invocation counts
# - Cache hit rates

Time-Passes Profiling

Track time spent in each compiler pass:

Basic Time Tracking
With LLVM Timing
Pretty Printing

# Time all passes
rustc -Z time-passes file.rs

# Output shows time for each pass:
# - parsing
# - expansion  
# - type checking
# - borrowck
# - codegen

# Include LLVM pass timing
rustc -Z time-passes -Z time-llvm-passes file.rs

# Shows detailed LLVM optimization times

# Pretty print timing info
rustc -Z time-passes file.rs 2>&1 | grep "time:"

# Save to file for analysis
rustc -Z time-passes file.rs 2>&1 | tee timing.txt

CPU Profiling

Using perf (Linux)

Record profile

# Record CPU profile
perf record -F 99 -g ./build/stage1/bin/rustc file.rs

# Record with call graph
perf record -F 99 --call-graph dwarf ./build/stage1/bin/rustc file.rs

Analyze results

perf report
Flamegraph

# Interactive report
perf report

# Show call graph
perf report --stdio

# Filter by function
perf report --stdio | grep function_name

# Generate flamegraph from perf data
perf script | stackcollapse-perf.pl | flamegraph.pl > flame.svg

# Or use cargo-flamegraph
cargo flamegraph --bin rustc -- file.rs

Using Instruments (macOS)

# Profile with Time Profiler
cargo instruments --template time --bin rustc -- file.rs

# Opens Instruments.app with results

Using samply (Cross-Platform)

Basic Profiling
With Arguments

# Profile with samply
samply record ./build/stage1/bin/rustc file.rs

# Opens profiler.firefox.com with results

# Profile with rustc arguments
samply record ./build/stage1/bin/rustc -C opt-level=3 file.rs

Memory Profiling

Heap Profiling with Valgrind

Massif
DHAT

# Profile heap usage over time
valgrind --tool=massif ./build/stage1/bin/rustc file.rs

# Analyze with ms_print
ms_print massif.out.12345

# Visualize with massif-visualizer
massif-visualizer massif.out.12345

# Detailed heap analyzer
valgrind --tool=dhat ./build/stage1/bin/rustc file.rs

# Shows:
# - Allocation patterns
# - Peak memory usage
# - Lifetime of allocations

Memory Statistics

# Show size of types in compilation
rustc -Z print-type-sizes file.rs

# Helps identify large types causing memory issues

Query Performance Analysis

Query Profiling

The compiler’s query system is a major performance factor:

Query Statistics
Dependency Graph
Query Blocking

# Profile query execution
rustc -Z self-profile-events=default,query-cache-hits file.rs

# Analyze cache hit rates
summarize file-pid.mm_profdata --query-stats

# Dump query dependency graph
rustc -Z dump-dep-graph file.rs

# Creates dep-graph.txt showing query dependencies

# Track query blocking
rustc -Z self-profile-events=query-blocked file.rs

# Shows which queries block on others

Incremental Compilation

# Enable incremental compilation stats
RUSTC_LOG=rustc_incremental=debug rustc file.rs

# Shows:
# - Cache hits/misses
# - Reused query results
# - Invalidated queries

LLVM Performance

LLVM Pass Timing

Time LLVM Passes
Optimization Remarks
LLVM Stats

# Time all LLVM passes
rustc -Z time-llvm-passes file.rs

# With detailed output
rustc -C llvm-args=-time-passes file.rs

# Get optimization remarks
rustc -C llvm-args="-pass-remarks=inline" file.rs

# All optimization remarks
rustc -C llvm-args="-pass-remarks=.*" file.rs 2>&1 | tee remarks.txt

# Enable LLVM statistics
rustc -C llvm-args=-stats file.rs

# Shows:
# - Number of functions
# - Instructions processed
# - Pass execution counts

LLVM Profiling

# Step 1: Build with instrumentation
rustc -C profile-generate=./pgo-data file.rs

# Step 2: Run the program
./file

# Step 3: Process profile data
llvm-profdata merge -o merged.profdata ./pgo-data

# Step 4: Build with profile
rustc -C profile-use=merged.profdata file.rs

Benchmarking

Compiler Benchmarks

Run rustc-perf

# Clone rustc-perf
git clone https://github.com/rust-lang/rustc-perf
cd rustc-perf

# Build collector
cargo build --release -p collector

# Run benchmarks
./target/release/collector bench_local ./build/stage1/bin/rustc

Analyze results

# View results
./target/release/collector analyze

# Compare with baseline
./target/release/collector compare baseline.json new.json

Custom Benchmarks

# Install hyperfine
cargo install hyperfine

# Benchmark compilation
hyperfine './build/stage1/bin/rustc file.rs'

# Compare different versions
hyperfine \
  --warmup 3 \
  './old-rustc file.rs' \
  './new-rustc file.rs'

Performance Analysis Workflow

Establish baseline

# Profile before changes
rustc -Z self-profile=baseline file.rs
summarize baseline-pid.mm_profdata > baseline.txt

Make changes

Implement your performance improvement

Profile again

# Profile after changes
rustc -Z self-profile=improved file.rs
summarize improved-pid.mm_profdata > improved.txt

Compare results

# Compare profiles
mmview diff baseline-pid.mm_profdata improved-pid.mm_profdata

# Look for:
# - Reduced query execution time
# - Better cache hit rates
# - Fewer invocations

Validate improvement

# Run benchmarks
./target/release/collector bench_local ./build/stage1/bin/rustc

# Check against perf suite
./x.py test src/tools/rustc-perf

Common Performance Issues

Slow query execution

Symptoms: High time in specific queriesDiagnosis:

rustc -Z self-profile file.rs
summarize file-pid.mm_profdata --query-stats

Solutions:

Add caching
Optimize hot paths
Reduce query invocations

Poor cache hit rate

Symptoms: Low query cache hitsDiagnosis:

rustc -Z self-profile-events=query-cache-hits file.rs

Solutions:

Improve query key design
Enable incremental compilation
Reduce query invalidation

High memory usage

Symptoms: Excessive memory consumptionDiagnosis:

valgrind --tool=massif ./build/stage1/bin/rustc file.rs

Solutions:

Reduce type sizes
Optimize data structures
Free unused memory

Slow LLVM codegen

Symptoms: Long time in codegen phaseDiagnosis:

rustc -Z time-llvm-passes file.rs

Solutions:

Reduce codegen units
Optimize MIR
Disable expensive LLVM passes

Optimization Tips

Query Optimization

Cache expensive computations
Minimize query parameters
Use shallow queries when possible
Avoid unnecessary query dependencies

Memory Optimization

Use arena allocation
Intern strings and types
Free temporary data
Use compact data structures

Compilation Speed

Enable incremental compilation
Use query caching effectively
Parallelize independent work
Reduce redundant computation

Code Generation

Optimize MIR before codegen
Use appropriate optimization levels
Enable LTO for release builds
Profile-guided optimization

Performance Testing in CI

Setting Up Performance Tests

name: Performance Test

on: [pull_request]

jobs:
  perf:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v2
      
      - name: Build
        run: ./x.py build --stage 1
      
      - name: Profile
        run: |
          rustc -Z self-profile test.rs
          summarize test-*.mm_profdata > results.txt
      
      - name: Check for regressions
        run: ./check-perf-regression.sh

Resources

Profiling Guide

Rustc dev guide profiling chapter

measureme

Self-profiling framework

rustc-perf

Compiler performance benchmarking

Performance Book

Rust performance book

Next Steps

Debugging

Learn debugging techniques

Compiler Development

Return to compiler development

Library Development

Optimize library code

Contributing Guide

Development

Documentation Index

​Why Profile the Compiler?

Compilation Speed

Memory Usage

Query Optimization

Code Quality

​Profiling Setup

​Self-Profiling with measureme

​Basic Self-Profiling

​Analyzing Self-Profile Data

​Time-Passes Profiling

​CPU Profiling

​Using perf (Linux)

​Using Instruments (macOS)

​Using samply (Cross-Platform)

​Memory Profiling

​Heap Profiling with Valgrind

​Memory Statistics

​Query Performance Analysis

​Query Profiling

​Incremental Compilation

​LLVM Performance

​LLVM Pass Timing

​LLVM Profiling

​Benchmarking

​Compiler Benchmarks

​Custom Benchmarks

​Performance Analysis Workflow

​Common Performance Issues

​Optimization Tips

Query Optimization

Memory Optimization

Compilation Speed

Code Generation

​Performance Testing in CI

​Setting Up Performance Tests

​Resources

Profiling Guide

measureme

rustc-perf

Performance Book

​Next Steps

Debugging

Compiler Development

Library Development

Build docs developers (and LLMs) love

Why Profile the Compiler?

Profiling Setup

Self-Profiling with measureme

Basic Self-Profiling

Analyzing Self-Profile Data

Time-Passes Profiling

CPU Profiling

Using perf (Linux)

Using Instruments (macOS)

Using samply (Cross-Platform)

Memory Profiling

Heap Profiling with Valgrind

Memory Statistics

Query Performance Analysis

Query Profiling

Incremental Compilation

LLVM Performance

LLVM Pass Timing

LLVM Profiling

Benchmarking

Compiler Benchmarks

Custom Benchmarks

Performance Analysis Workflow

Common Performance Issues

Optimization Tips

Performance Testing in CI

Setting Up Performance Tests

Resources

Next Steps