carbon-language/carbon-lang
1
0
Fork 0
mirror of https://github.com/carbon-language/carbon-lang.git synced 2026-10-04 22:02:52 +01:00
Code Issues Packages Projects Releases 10 Wiki Activity

Add support for running LLVM optimizer. (#6225)

Browse Source 
Adds a flag `--optimize=<mode>` that specifies what to optimize for:

* `--optimize=none` turns off the optimizer as much as possible, but
still respects always_inline.
* `--optimize=debug` aims to be the equivalent of `-Og` / `-O1`, and
provides optimizations that don't affect the ability to debug the
program. This is the default.
* `--optimize=size` optimizes for the size of the produced program, and
aims to be the equivalent of `-Oz`.
* `--optimize=speed` optimizes for the execution time of the produced
program, and aims to be the equivalent of `-O3`.

Following the approach taken by Clang, the optimization level feeds into
both the configuration of the LLVM pass pipeline and the attributes
added to function definitions generated by the frontend.

Optimization is performed in a new phase, `optimize`, which runs between
`lower` and `codegen`.

---------

Co-authored-by: Dana Jansens <danakj@orodu.net>
Co-authored-by: Chandler Carruth <chandlerc@gmail.com>
This commit is contained in:
Richard Smith
2025-11-05 00:15:14 +00:00
committed by GitHub
co-authored by Dana Jansens Chandler Carruth
parent db150ffc5c
commit aa69a484eb
202 changed files with 3376 additions and 1253 deletions
Download Patch File Download Diff File Expand all files Collapse all files
toolchain/driver/compile_subcommand.cpp
+216 -19
View File
@@ -16,6 +16,9 @@
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/ScopeExit.h"
#include "llvm/MC/TargetRegistry.h"
#include "llvm/Passes/OptimizationLevel.h"
#include "llvm/Passes/PassBuilder.h"
#include "llvm/Passes/StandardInstrumentations.h"
#include "toolchain/base/clang_invocation.h"
#include "toolchain/base/timings.h"
#include "toolchain/check/check.h"
@@ -60,6 +63,7 @@ compile to machine code.
arg_b.OneOfValue("parse", Phase::Parse),
arg_b.OneOfValue("check", Phase::Check),
arg_b.OneOfValue("lower", Phase::Lower),
arg_b.OneOfValue("optimize", Phase::Optimize),
arg_b.OneOfValue("codegen", Phase::CodeGen).Default(true),
},
&phase);
@@ -109,6 +113,33 @@ object output can be forced by enabling `--force-obj-output`.
},
[&](auto& arg_b) { arg_b.Set(&output_filename); });
b.AddOneOfOption(
{
.name = "optimize",
.help = R"""(
Selects the amount of optimization to perform.
)""",
},
[&](auto& arg_b) {
arg_b.SetOneOf(
{
// We intentionally don't expose O2 and Os. The difference
// between these levels tends to reflect what achieves the
// best speed for a specific application, as they all
// largely optimize for speed as the primary factor.
//
// Instead of controlling this with more nuanced flags, we
// plan to support profile and in-source hints to the
// optimizer to adjust its strategy in the specific places
// where the default doesn't have the desired results.
arg_b.OneOfValue("none", Lower::OptimizationLevel::None),
arg_b.OneOfValue("debug", Lower::OptimizationLevel::Debug),
arg_b.OneOfValue("speed", Lower::OptimizationLevel::Speed),
arg_b.OneOfValue("size", Lower::OptimizationLevel::Size),
},
&opt_level);
});
// Include the common code generation options at this point to render it
// after the more common options above, but before the more unusual options
// below.
@@ -368,6 +399,8 @@ static auto PhaseToString(CompileOptions::Phase phase) -> std::string {
return "check";
case CompileOptions::Phase::Lower:
return "lower";
case CompileOptions::Phase::Optimize:
return "optimize";
case CompileOptions::Phase::CodeGen:
return "codegen";
}
@@ -408,6 +441,7 @@ auto CompileSubcommand::ValidateOptions(
}
[[fallthrough]];
case Phase::Lower:
case Phase::Optimize:
case Phase::CodeGen:
// Everything can be dumped in these phases.
break;
@@ -422,11 +456,12 @@ class MultiUnitCache;
// Ties together information for a file being compiled.
class CompilationUnit {
public:
// `driver_env`, `options`, and `consumer` must be non-null.
// `driver_env`, `options`, `consumer`, and `target` must be non-null.
explicit CompilationUnit(SemIR::CheckIRId check_ir_id, int total_ir_count,
DriverEnv* driver_env, const CompileOptions* options,
Diagnostics::Consumer* consumer,
llvm::StringRef input_filename);
llvm::StringRef input_filename,
const llvm::Target* target);
// Sets the multi-unit cache and initializes dependent member state.
auto SetMultiUnitCache(MultiUnitCache* cache) -> void;
@@ -446,6 +481,14 @@ class CompilationUnit {
// Lower SemIR to LLVM IR.
auto RunLower() -> void;
// Runs the optimization pipeline.
auto RunOptimize() -> void;
// Runs post-lowering-to-LLVM-IR logic. This is always called if we do any
// lowering work, after we've finished building the IR in RunLower() and,
// optionally, RunOptimize().
auto PostLower() -> void;
auto RunCodeGen() -> void;
// Runs post-compile logic. This is always called, and called after all other
@@ -484,6 +527,9 @@ class CompilationUnit {
// Returns true if the current file should be included in debug dumps.
auto IncludeInDumps() -> bool;
// Builds the LLVM target machine.
auto MakeTargetMachine() -> void;
// The index of the unit amongst all units.
SemIR::CheckIRId check_ir_id_;
// The number of units in total.
@@ -491,6 +537,7 @@ class CompilationUnit {
DriverEnv* driver_env_;
const CompileOptions* options_;
const llvm::Target* target_;
SharedValueStores value_stores_;
@@ -527,6 +574,7 @@ class CompilationUnit {
std::unique_ptr<clang::ASTUnit> clang_ast_unit_;
std::unique_ptr<llvm::LLVMContext> llvm_context_;
std::unique_ptr<llvm::Module> module_;
std::unique_ptr<llvm::TargetMachine> target_machine_;
};
// Caches lists that are shared cross-unit. Accessors do lazy caching because
@@ -611,11 +659,13 @@ CompilationUnit::CompilationUnit(SemIR::CheckIRId check_ir_id,
int total_ir_count, DriverEnv* driver_env,
const CompileOptions* options,
Diagnostics::Consumer* consumer,
llvm::StringRef input_filename)
llvm::StringRef input_filename,
const llvm::Target* target)
: check_ir_id_(check_ir_id),
total_ir_count_(total_ir_count),
driver_env_(driver_env),
options_(options),
target_(target),
input_filename_(input_filename),
vlog_stream_(driver_env_->vlog_stream) {
if (vlog_stream_ != nullptr || options_->stream_errors) {
@@ -743,15 +793,149 @@ auto CompilationUnit::RunLower() -> void {
options_->run_llvm_verifier ? driver_env_->error_stream : nullptr;
options.want_debug_info = options_->include_debug_info;
options.vlog_stream = vlog_stream_;
if (options_->dump_llvm_ir && IncludeInDumps()) {
options.dump_stream = driver_env_->output_stream;
}
options.opt_level = options_->opt_level;
module_ = Lower::LowerToLLVM(*llvm_context_, driver_env_->fs,
cache_->tree_and_subtrees_getters(), *sem_ir_,
total_ir_count_, options);
});
}
auto CompilationUnit::MakeTargetMachine() -> void {
CARBON_CHECK(module_, "Must call RunLower first");
CARBON_CHECK(!target_machine_, "Should not call this multiple times");
// Set the target on the module.
// TODO: We should do this earlier. Lower should be passed the target triple
// so it can create the module with this already set.
llvm::Triple target_triple(options_->codegen_options.target);
module_->setTargetTriple(target_triple);
// TODO: Provide flags to control these.
constexpr llvm::StringLiteral CPU = "generic";
constexpr llvm::StringLiteral Features = "";
llvm::TargetOptions target_opts;
target_machine_.reset(target_->createTargetMachine(
target_triple, CPU, Features, target_opts, llvm::Reloc::PIC_));
}
// Get the LLVM optimization level corresponding to a Carbon optimization level.
static auto GetLLVMOptimizationLevel(Lower::OptimizationLevel opt_level)
-> llvm::OptimizationLevel {
switch (opt_level) {
case Lower::OptimizationLevel::None:
return llvm::OptimizationLevel::O0;
case Lower::OptimizationLevel::Debug:
return llvm::OptimizationLevel::O1;
case Lower::OptimizationLevel::Size:
return llvm::OptimizationLevel::Oz;
case Lower::OptimizationLevel::Speed:
return llvm::OptimizationLevel::O3;
}
}
// Get the `-O` flag corresponding to an optimization level.
static auto GetClangOptimizationFlag(Lower::OptimizationLevel opt_level)
-> llvm::StringLiteral {
switch (opt_level) {
case Lower::OptimizationLevel::None:
return "-O0";
case Lower::OptimizationLevel::Debug:
return "-O1";
case Lower::OptimizationLevel::Size:
return "-Oz";
case Lower::OptimizationLevel::Speed:
return "-O3";
}
}
auto CompilationUnit::RunOptimize() -> void {
CARBON_CHECK(module_, "Must call RunLower first");
// TODO: A lot of the work done here duplicates work done by Clang setting up
// its pass manager. Moreover, we probably want to pick up Clang's
// customizations and make use of its flags for controlling LLVM passes. We
// should consider whether we would be better off running Clang's pass
// pipeline rather than building one of our own, or factoring out enough of
// Clang's pipeline builder that we can reuse and further customize it.
MakeTargetMachine();
// TODO: There's no way to set these automatically from an
// llvm::OptimizationLevel. Add such a mechanism to LLVM and use it from
// here. For now we reconstruct what Clang does by default.
llvm::PipelineTuningOptions pto;
bool opt_for_speed = options_->opt_level == Lower::OptimizationLevel::Speed;
bool opt_for_size_or_speed =
opt_for_speed || options_->opt_level == Lower::OptimizationLevel::Size;
// Loop unrolling is enabled by `--optimize=size` but isn't actually performed
// because we add `optsize` attributes to the function definitions we emit.
pto.LoopUnrolling = opt_for_size_or_speed;
pto.LoopInterleaving = opt_for_size_or_speed;
pto.LoopVectorization = opt_for_speed;
pto.SLPVectorization = opt_for_size_or_speed;
llvm::LoopAnalysisManager lam;
llvm::FunctionAnalysisManager fam;
llvm::CGSCCAnalysisManager cgam;
llvm::ModuleAnalysisManager mam;
llvm::PassInstrumentationCallbacks pic;
// Register standard pass instrumentations. This adds support for things like
// `-print-after-all`.
llvm::StandardInstrumentations si(module_->getContext(),
/*DebugLogging=*/false);
si.registerCallbacks(pic);
llvm::PassBuilder builder(target_machine_.get(), pto,
/*PGOOpt=*/std::nullopt, &pic);
// TODO: Add an AssignmentTrackingPass for at least `--optimize=debug`.
// Set up target library information and add an analysis pass to supply it.
std::unique_ptr<llvm::TargetLibraryInfoImpl> tlii(llvm::driver::createTLII(
module_->getTargetTriple(), llvm::driver::VectorLibrary::NoLibrary));
fam.registerPass([&] { return llvm::TargetLibraryAnalysis(*tlii); });
builder.registerModuleAnalyses(mam);
builder.registerCGSCCAnalyses(cgam);
builder.registerFunctionAnalyses(fam);
builder.registerLoopAnalyses(lam);
builder.crossRegisterProxies(lam, fam, cgam, mam);
llvm::ModulePassManager pass_manager = builder.buildPerModuleDefaultPipeline(
GetLLVMOptimizationLevel(options_->opt_level));
if (vlog_stream_) {
CARBON_VLOG("*** Running pass pipeline: ");
pass_manager.printPipeline(
*vlog_stream_, [&pic](llvm::StringRef class_name) {
auto pass_name = pic.getPassNameForClassName(class_name);
return pass_name.empty() ? class_name : pass_name;
});
CARBON_VLOG(" ***\n");
}
LogCall("ModulePassManager::run", "optimize",
[&] { pass_manager.run(*module_, mam); });
if (vlog_stream_) {
CARBON_VLOG("*** Optimized llvm::Module ***\n");
module_->print(*vlog_stream_, /*AAW=*/nullptr,
/*ShouldPreserveUseListOrder=*/false,
/*IsForDebug=*/true);
}
}
auto CompilationUnit::PostLower() -> void {
CARBON_CHECK(module_, "Must call RunLower first");
if (options_->dump_llvm_ir && IncludeInDumps()) {
module_->print(*driver_env_->output_stream, /*AAW=*/nullptr,
/*ShouldPreserveUseListOrder=*/true);
}
}
auto CompilationUnit::RunCodeGen() -> void {
CARBON_CHECK(module_, "Must call RunLower first");
LogCall("CodeGen", "codegen", [&] { success_ = RunCodeGenHelper(); });
@@ -778,14 +962,13 @@ auto CompilationUnit::PostCompile() -> void {
}
auto CompilationUnit::RunCodeGenHelper() -> bool {
std::optional<CodeGen> codegen =
CodeGen::Make(module_.get(), options_->codegen_options.target, consumer_);
if (!codegen) {
return false;
}
CARBON_CHECK(module_, "Must call RunLower first");
CARBON_CHECK(target_machine_, "Must call MakeTargetMachine first");
CodeGen codegen(module_.get(), target_machine_.get(), consumer_);
if (vlog_stream_) {
CARBON_VLOG("*** Assembly ***\n");
codegen->EmitAssembly(*vlog_stream_);
codegen.EmitAssembly(*vlog_stream_);
}
if (options_->output_filename == "-") {
@@ -793,11 +976,11 @@ auto CompilationUnit::RunCodeGenHelper() -> bool {
// textual assembly output are all somewhat linked flags. We should add
// some validation that they are used correctly.
if (options_->force_obj_output) {
if (!codegen->EmitObject(*driver_env_->output_stream)) {
if (!codegen.EmitObject(*driver_env_->output_stream)) {
return false;
}
} else {
if (!codegen->EmitAssembly(*driver_env_->output_stream)) {
if (!codegen.EmitAssembly(*driver_env_->output_stream)) {
return false;
}
}
@@ -841,11 +1024,11 @@ auto CompilationUnit::RunCodeGenHelper() -> bool {
return false;
}
if (options_->asm_output) {
if (!codegen->EmitAssembly(output_file)) {
if (!codegen.EmitAssembly(output_file)) {
return false;
}
} else {
if (!codegen->EmitObject(output_file)) {
if (!codegen.EmitObject(output_file)) {
return false;
}
}
@@ -910,6 +1093,10 @@ auto CompileSubcommand::Run(DriverEnv& driver_env) -> DriverResult {
// defaults to be more stable.
// TODO: Decide if we want this.
"-fPIE",
// Propagate our optimization level to Clang as a default. This can be
// overridden by Clang arguments, but doing so will only have an effect
// if those arguments affect Clang's IR, not its pass pipeline.
GetClangOptimizationFlag(options_.opt_level).str(),
};
if (driver_env.fuzzing && !options_.clang_args.empty()) {
// Parsing specific Clang arguments can reach deep into
@@ -925,6 +1112,9 @@ auto CompileSubcommand::Run(DriverEnv& driver_env) -> DriverResult {
if (!clang_invocation) {
return {.success = false};
}
// We will run our own pass pipeline over the IR in the `Optimize` phase, so
// disable Clang's pipeline to avoid optimizing C++ code twice.
clang_invocation->getCodeGenOpts().DisableLLVMPasses = true;
}
// Find the files comprising the prelude if we are importing it.
@@ -952,7 +1142,7 @@ auto CompileSubcommand::Run(DriverEnv& driver_env) -> DriverResult {
++unit_index;
return std::make_unique<CompilationUnit>(
SemIR::CheckIRId(unit_index), total_unit_count, &driver_env, &options_,
&driver_env.consumer, filename);
&driver_env.consumer, filename, target);
};
llvm::append_range(units, llvm::map_range(prelude, unit_builder));
llvm::append_range(units,
@@ -1079,11 +1269,18 @@ auto CompileSubcommand::Run(DriverEnv& driver_env) -> DriverResult {
return make_result();
}
// Lower.
// Lower and optimize.
for (const auto& unit : units) {
unit->RunLower();
if (options_.phase != CompileOptions::Phase::Lower) {
unit->RunOptimize();
}
unit->PostLower();
}
if (options_.phase == CompileOptions::Phase::Lower) {
if (options_.phase == CompileOptions::Phase::Lower ||
options_.phase == CompileOptions::Phase::Optimize) {
return make_result();
}
CARBON_CHECK(options_.phase == CompileOptions::Phase::CodeGen,