The embedded bitcode should always be prepared for LTO/ThinLTO

This commit is contained in:
DianQK 2024-11-17 14:21:23 +08:00
parent 1805b33483
commit 1a99ca8da9
No known key found for this signature in database
13 changed files with 265 additions and 69 deletions

View File

@ -210,7 +210,7 @@ fn produce_final_output_artifacts(
// to get rid of it.
for output_type in crate_output.outputs.keys() {
match *output_type {
OutputType::Bitcode | OutputType::ThinLinkBitcode => {
OutputType::Bitcode | OutputType::ThinLinkBitcode | OutputType::ThinBitcode => {
// Cranelift doesn't have bitcode
// user_wants_bitcode = true;
// // Copy to .bc, but always keep the .0.bc. There is a later

View File

@ -2,6 +2,7 @@ use std::collections::BTreeMap;
use std::ffi::{CStr, CString};
use std::fs::File;
use std::path::Path;
use std::ptr::NonNull;
use std::sync::Arc;
use std::{io, iter, slice};
@ -655,14 +656,14 @@ pub(crate) fn run_pass_manager(
}
unsafe {
write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, stage)?;
write::llvm_optimize(cgcx, dcx, module, None, config, opt_level, opt_stage, stage)?;
}
if cfg!(llvm_enzyme) && enable_ad {
let opt_stage = llvm::OptStage::FatLTO;
let stage = write::AutodiffStage::PostAD;
unsafe {
write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, stage)?;
write::llvm_optimize(cgcx, dcx, module, None, config, opt_level, opt_stage, stage)?;
}
// This is the final IR, so people should be able to inspect the optimized autodiff output.
@ -729,6 +730,11 @@ impl ThinBuffer {
ThinBuffer(buffer)
}
}
pub unsafe fn from_raw_ptr(ptr: *mut llvm::ThinLTOBuffer) -> ThinBuffer {
let mut ptr = NonNull::new(ptr).unwrap();
ThinBuffer(unsafe { ptr.as_mut() })
}
}
impl ThinBufferMethods for ThinBuffer {

View File

@ -1,6 +1,7 @@
use std::ffi::{CStr, CString};
use std::io::{self, Write};
use std::path::{Path, PathBuf};
use std::ptr::null_mut;
use std::sync::Arc;
use std::{fs, slice, str};
@ -15,7 +16,7 @@ use rustc_codegen_ssa::back::write::{
TargetMachineFactoryFn,
};
use rustc_codegen_ssa::traits::*;
use rustc_codegen_ssa::{CompiledModule, ModuleCodegen};
use rustc_codegen_ssa::{CompiledModule, ModuleCodegen, ModuleKind};
use rustc_data_structures::profiling::SelfProfilerRef;
use rustc_data_structures::small_c_str::SmallCStr;
use rustc_errors::{DiagCtxtHandle, FatalError, Level};
@ -551,6 +552,7 @@ pub(crate) unsafe fn llvm_optimize(
cgcx: &CodegenContext<LlvmCodegenBackend>,
dcx: DiagCtxtHandle<'_>,
module: &ModuleCodegen<ModuleLlvm>,
thin_lto_buffer: Option<&mut *mut llvm::ThinLTOBuffer>,
config: &ModuleConfig,
opt_level: config::OptLevel,
opt_stage: llvm::OptStage,
@ -584,7 +586,17 @@ pub(crate) unsafe fn llvm_optimize(
vectorize_loop = config.vectorize_loop;
}
trace!(?unroll_loops, ?vectorize_slp, ?vectorize_loop, ?run_enzyme);
let using_thin_buffers = opt_stage == llvm::OptStage::PreLinkThinLTO || config.bitcode_needed();
if thin_lto_buffer.is_some() {
assert!(
matches!(
opt_stage,
llvm::OptStage::PreLinkNoLTO
| llvm::OptStage::PreLinkFatLTO
| llvm::OptStage::PreLinkThinLTO
),
"the bitcode for LTO can only be obtained at the pre-link stage"
);
}
let pgo_gen_path = get_pgo_gen_path(config);
let pgo_use_path = get_pgo_use_path(config);
let pgo_sample_use_path = get_pgo_sample_use_path(config);
@ -644,7 +656,9 @@ pub(crate) unsafe fn llvm_optimize(
config.no_prepopulate_passes,
config.verify_llvm_ir,
config.lint_llvm_ir,
using_thin_buffers,
thin_lto_buffer,
config.emit_thin_lto,
config.emit_thin_lto_summary,
config.merge_functions,
unroll_loops,
vectorize_slp,
@ -705,9 +719,56 @@ pub(crate) unsafe fn optimize(
// Otherwise we pretend AD is already done and run the normal opt pipeline (=PostAD).
let consider_ad = cfg!(llvm_enzyme) && config.autodiff.contains(&config::AutoDiff::Enable);
let autodiff_stage = if consider_ad { AutodiffStage::PreAD } else { AutodiffStage::PostAD };
return unsafe {
llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, autodiff_stage)
// The embedded bitcode is used to run LTO/ThinLTO.
// The bitcode obtained during the `codegen` phase is no longer suitable for performing LTO.
// It may have undergone LTO due to ThinLocal, so we need to obtain the embedded bitcode at
// this point.
let mut thin_lto_buffer = if (module.kind == ModuleKind::Regular
&& config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full))
|| config.emit_thin_lto_summary
{
Some(null_mut())
} else {
None
};
unsafe {
llvm_optimize(
cgcx,
dcx,
module,
thin_lto_buffer.as_mut(),
config,
opt_level,
opt_stage,
autodiff_stage,
)
}?;
if let Some(thin_lto_buffer) = thin_lto_buffer {
let thin_lto_buffer = unsafe { ThinBuffer::from_raw_ptr(thin_lto_buffer) };
let thin_bc_out = cgcx.output_filenames.temp_path(OutputType::ThinBitcode, module_name);
if let Err(err) = fs::write(&thin_bc_out, thin_lto_buffer.data()) {
dcx.emit_err(WriteBytecode { path: &thin_bc_out, err });
}
let bc_summary_out =
cgcx.output_filenames.temp_path(OutputType::ThinLinkBitcode, module_name);
if config.emit_thin_lto_summary
&& let Some(thin_link_bitcode_filename) = bc_summary_out.file_name()
{
let summary_data = thin_lto_buffer.thin_link_data();
cgcx.prof.artifact_size(
"llvm_bitcode_summary",
thin_link_bitcode_filename.to_string_lossy(),
summary_data.len() as u64,
);
let _timer = cgcx.prof.generic_activity_with_arg(
"LLVM_module_codegen_emit_bitcode_summary",
&*module.name,
);
if let Err(err) = fs::write(&bc_summary_out, summary_data) {
dcx.emit_err(WriteBytecode { path: &bc_summary_out, err });
}
}
}
}
Ok(())
}
@ -760,59 +821,47 @@ pub(crate) unsafe fn codegen(
// otherwise requested.
let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name);
let bc_summary_out =
cgcx.output_filenames.temp_path(OutputType::ThinLinkBitcode, module_name);
let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name);
if config.bitcode_needed() {
let _timer = cgcx
.prof
.generic_activity_with_arg("LLVM_module_codegen_make_bitcode", &*module.name);
let thin = ThinBuffer::new(llmod, config.emit_thin_lto, config.emit_thin_lto_summary);
let data = thin.data();
if let Some(bitcode_filename) = bc_out.file_name() {
cgcx.prof.artifact_size(
"llvm_bitcode",
bitcode_filename.to_string_lossy(),
data.len() as u64,
);
}
if config.emit_thin_lto_summary
&& let Some(thin_link_bitcode_filename) = bc_summary_out.file_name()
{
let summary_data = thin.thin_link_data();
cgcx.prof.artifact_size(
"llvm_bitcode_summary",
thin_link_bitcode_filename.to_string_lossy(),
summary_data.len() as u64,
);
let _timer = cgcx.prof.generic_activity_with_arg(
"LLVM_module_codegen_emit_bitcode_summary",
&*module.name,
);
if let Err(err) = fs::write(&bc_summary_out, summary_data) {
dcx.emit_err(WriteBytecode { path: &bc_summary_out, err });
}
}
if config.emit_bc || config.emit_obj == EmitObj::Bitcode {
let thin = {
let _timer = cgcx.prof.generic_activity_with_arg(
"LLVM_module_codegen_make_bitcode",
&*module.name,
);
ThinBuffer::new(llmod, config.emit_thin_lto, false)
};
let data = thin.data();
let _timer = cgcx
.prof
.generic_activity_with_arg("LLVM_module_codegen_emit_bitcode", &*module.name);
if let Some(bitcode_filename) = bc_out.file_name() {
cgcx.prof.artifact_size(
"llvm_bitcode",
bitcode_filename.to_string_lossy(),
data.len() as u64,
);
}
if let Err(err) = fs::write(&bc_out, data) {
dcx.emit_err(WriteBytecode { path: &bc_out, err });
}
}
if config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full) {
if config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full)
&& module.kind == ModuleKind::Regular
{
let _timer = cgcx
.prof
.generic_activity_with_arg("LLVM_module_codegen_embed_bitcode", &*module.name);
let thin_bc_out =
cgcx.output_filenames.temp_path(OutputType::ThinBitcode, module_name);
assert!(thin_bc_out.exists(), "cannot find {:?} as embedded bitcode", thin_bc_out);
let data = fs::read(&thin_bc_out).unwrap();
debug!("removing embed bitcode file {:?}", thin_bc_out);
ensure_removed(dcx, &thin_bc_out);
unsafe {
embed_bitcode(cgcx, llcx, llmod, &config.bc_cmdline, data);
embed_bitcode(cgcx, llcx, llmod, &config.bc_cmdline, &data);
}
}
}

View File

@ -2421,7 +2421,9 @@ unsafe extern "C" {
NoPrepopulatePasses: bool,
VerifyIR: bool,
LintIR: bool,
UseThinLTOBuffers: bool,
ThinLTOBuffer: Option<&mut *mut ThinLTOBuffer>,
EmitThinLTO: bool,
EmitThinLTOSummary: bool,
MergeFunctions: bool,
UnrollLoops: bool,
SLPVectorize: bool,

View File

@ -626,6 +626,9 @@ fn produce_final_output_artifacts(
// them for making an rlib.
copy_if_one_unit(OutputType::Bitcode, true);
}
OutputType::ThinBitcode => {
copy_if_one_unit(OutputType::ThinBitcode, true);
}
OutputType::ThinLinkBitcode => {
copy_if_one_unit(OutputType::ThinLinkBitcode, false);
}

View File

@ -7,6 +7,7 @@
#include "llvm/Analysis/Lint.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Bitcode/BitcodeWriter.h"
#include "llvm/Bitcode/BitcodeWriterPass.h"
#include "llvm/CodeGen/CommandFlags.h"
#include "llvm/IR/AssemblyAnnotationWriter.h"
#include "llvm/IR/AutoUpgrade.h"
@ -37,6 +38,7 @@
#include "llvm/Transforms/Instrumentation/InstrProfiling.h"
#include "llvm/Transforms/Instrumentation/MemorySanitizer.h"
#include "llvm/Transforms/Instrumentation/ThreadSanitizer.h"
#include "llvm/Transforms/Scalar/AnnotationRemarks.h"
#include "llvm/Transforms/Utils/CanonicalizeAliases.h"
#include "llvm/Transforms/Utils/FunctionImportUtils.h"
#include "llvm/Transforms/Utils/NameAnonGlobals.h"
@ -195,6 +197,19 @@ extern "C" void LLVMRustTimeTraceProfilerFinish(const char *FileName) {
GEN_SUBTARGETS
#undef SUBTARGET
// This struct and various functions are sort of a hack right now, but the
// problem is that we've got in-memory LLVM modules after we generate and
// optimize all codegen-units for one compilation in rustc. To be compatible
// with the LTO support above we need to serialize the modules plus their
// ThinLTO summary into memory.
//
// This structure is basically an owned version of a serialize module, with
// a ThinLTO summary attached.
struct LLVMRustThinLTOBuffer {
std::string data;
std::string thin_link_data;
};
extern "C" bool LLVMRustHasFeature(LLVMTargetMachineRef TM,
const char *Feature) {
TargetMachine *Target = unwrap(TM);
@ -704,7 +719,8 @@ extern "C" LLVMRustResult LLVMRustOptimize(
LLVMModuleRef ModuleRef, LLVMTargetMachineRef TMRef,
LLVMRustPassBuilderOptLevel OptLevelRust, LLVMRustOptStage OptStage,
bool IsLinkerPluginLTO, bool NoPrepopulatePasses, bool VerifyIR,
bool LintIR, bool UseThinLTOBuffers, bool MergeFunctions, bool UnrollLoops,
bool LintIR, LLVMRustThinLTOBuffer **ThinLTOBufferRef, bool EmitThinLTO,
bool EmitThinLTOSummary, bool MergeFunctions, bool UnrollLoops,
bool SLPVectorize, bool LoopVectorize, bool DisableSimplifyLibCalls,
bool EmitLifetimeMarkers, bool RunEnzyme,
LLVMRustSanitizerOptions *SanitizerOptions, const char *PGOGenPath,
@ -952,7 +968,10 @@ extern "C" LLVMRustResult LLVMRustOptimize(
}
ModulePassManager MPM;
bool NeedThinLTOBufferPasses = UseThinLTOBuffers;
bool NeedThinLTOBufferPasses = EmitThinLTO;
auto ThinLTOBuffer = std::make_unique<LLVMRustThinLTOBuffer>();
raw_string_ostream ThinLTODataOS(ThinLTOBuffer->data);
raw_string_ostream ThinLinkDataOS(ThinLTOBuffer->thin_link_data);
if (!NoPrepopulatePasses) {
// The pre-link pipelines don't support O0 and require using
// buildO0DefaultPipeline() instead. At the same time, the LTO pipelines do
@ -976,7 +995,25 @@ extern "C" LLVMRustResult LLVMRustOptimize(
switch (OptStage) {
case LLVMRustOptStage::PreLinkNoLTO:
MPM = PB.buildPerModuleDefaultPipeline(OptLevel);
if (ThinLTOBufferRef) {
// This is similar to LLVM's `buildFatLTODefaultPipeline`, where the
// bitcode for embedding is obtained after performing
// `ThinLTOPreLinkDefaultPipeline`.
MPM.addPass(PB.buildThinLTOPreLinkDefaultPipeline(OptLevel));
if (EmitThinLTO) {
MPM.addPass(ThinLTOBitcodeWriterPass(
ThinLTODataOS, EmitThinLTOSummary ? &ThinLinkDataOS : nullptr));
} else {
MPM.addPass(BitcodeWriterPass(ThinLTODataOS));
}
*ThinLTOBufferRef = ThinLTOBuffer.release();
MPM.addPass(PB.buildModuleOptimizationPipeline(
OptLevel, ThinOrFullLTOPhase::None));
MPM.addPass(
createModuleToFunctionPassAdaptor(AnnotationRemarksPass()));
} else {
MPM = PB.buildPerModuleDefaultPipeline(OptLevel);
}
break;
case LLVMRustOptStage::PreLinkThinLTO:
MPM = PB.buildThinLTOPreLinkDefaultPipeline(OptLevel);
@ -1022,6 +1059,16 @@ extern "C" LLVMRustResult LLVMRustOptimize(
MPM.addPass(CanonicalizeAliasesPass());
MPM.addPass(NameAnonGlobalPass());
}
// For `-Copt-level=0`, ThinLTO, or LTO.
if (ThinLTOBufferRef && *ThinLTOBufferRef == nullptr) {
if (EmitThinLTO) {
MPM.addPass(ThinLTOBitcodeWriterPass(
ThinLTODataOS, EmitThinLTOSummary ? &ThinLinkDataOS : nullptr));
} else {
MPM.addPass(BitcodeWriterPass(ThinLTODataOS));
}
*ThinLTOBufferRef = ThinLTOBuffer.release();
}
// now load "-enzyme" pass:
#ifdef ENZYME
@ -1500,19 +1547,6 @@ extern "C" bool LLVMRustPrepareThinLTOImport(const LLVMRustThinLTOData *Data,
return true;
}
// This struct and various functions are sort of a hack right now, but the
// problem is that we've got in-memory LLVM modules after we generate and
// optimize all codegen-units for one compilation in rustc. To be compatible
// with the LTO support above we need to serialize the modules plus their
// ThinLTO summary into memory.
//
// This structure is basically an owned version of a serialize module, with
// a ThinLTO summary attached.
struct LLVMRustThinLTOBuffer {
std::string data;
std::string thin_link_data;
};
extern "C" LLVMRustThinLTOBuffer *
LLVMRustThinLTOBufferCreate(LLVMModuleRef M, bool is_thin, bool emit_summary) {
auto Ret = std::make_unique<LLVMRustThinLTOBuffer>();

View File

@ -541,6 +541,7 @@ impl FromStr for SplitDwarfKind {
pub enum OutputType {
Bitcode,
ThinLinkBitcode,
ThinBitcode,
Assembly,
LlvmAssembly,
Mir,
@ -571,6 +572,7 @@ impl OutputType {
OutputType::Exe | OutputType::DepInfo | OutputType::Metadata => true,
OutputType::Bitcode
| OutputType::ThinLinkBitcode
| OutputType::ThinBitcode
| OutputType::Assembly
| OutputType::LlvmAssembly
| OutputType::Mir
@ -582,6 +584,7 @@ impl OutputType {
match *self {
OutputType::Bitcode => "llvm-bc",
OutputType::ThinLinkBitcode => "thin-link-bitcode",
OutputType::ThinBitcode => "thin-llvm-bc",
OutputType::Assembly => "asm",
OutputType::LlvmAssembly => "llvm-ir",
OutputType::Mir => "mir",
@ -599,6 +602,7 @@ impl OutputType {
"mir" => OutputType::Mir,
"llvm-bc" => OutputType::Bitcode,
"thin-link-bitcode" => OutputType::ThinLinkBitcode,
"thin-llvm-bc" => OutputType::ThinBitcode,
"obj" => OutputType::Object,
"metadata" => OutputType::Metadata,
"link" => OutputType::Exe,
@ -609,9 +613,10 @@ impl OutputType {
fn shorthands_display() -> String {
format!(
"`{}`, `{}`, `{}`, `{}`, `{}`, `{}`, `{}`, `{}`, `{}`",
"`{}`, `{}`, `{}`, `{}`, `{}`, `{}`, `{}`, `{}`, `{}`, `{}`",
OutputType::Bitcode.shorthand(),
OutputType::ThinLinkBitcode.shorthand(),
OutputType::ThinBitcode.shorthand(),
OutputType::Assembly.shorthand(),
OutputType::LlvmAssembly.shorthand(),
OutputType::Mir.shorthand(),
@ -626,6 +631,7 @@ impl OutputType {
match *self {
OutputType::Bitcode => "bc",
OutputType::ThinLinkBitcode => "indexing.o",
OutputType::ThinBitcode => "thin.bc",
OutputType::Assembly => "s",
OutputType::LlvmAssembly => "ll",
OutputType::Mir => "mir",
@ -644,6 +650,7 @@ impl OutputType {
| OutputType::DepInfo => true,
OutputType::Bitcode
| OutputType::ThinLinkBitcode
| OutputType::ThinBitcode
| OutputType::Object
| OutputType::Metadata
| OutputType::Exe => false,
@ -731,6 +738,7 @@ impl OutputTypes {
self.0.keys().any(|k| match *k {
OutputType::Bitcode
| OutputType::ThinLinkBitcode
| OutputType::ThinBitcode
| OutputType::Assembly
| OutputType::LlvmAssembly
| OutputType::Mir
@ -745,6 +753,7 @@ impl OutputTypes {
self.0.keys().any(|k| match *k {
OutputType::Bitcode
| OutputType::ThinLinkBitcode
| OutputType::ThinBitcode
| OutputType::Assembly
| OutputType::LlvmAssembly
| OutputType::Mir

View File

@ -9,14 +9,14 @@ struct S(i32);
struct SmallStruct(f32, Option<S>, &'static [f32]);
// CHECK: @0 = private unnamed_addr constant
// CHECK: [[const:@.*]] = private unnamed_addr constant
// CHECK-SAME: , align 8
#[no_mangle]
pub fn overaligned_constant() {
// CHECK-LABEL: @overaligned_constant
// CHECK: [[full:%_.*]] = alloca [32 x i8], align 8
// CHECK: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[full]], ptr align 8 @0, i64 32, i1 false)
// CHECK: call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[full]], ptr align 8 [[const]], i64 32, i1 false)
let mut s = S(1);
s.0 = 3;

View File

@ -11,15 +11,15 @@ pub struct PartiallyUninit {
y: MaybeUninit<[u8; 10]>,
}
// CHECK: [[FULLY_UNINIT:@[0-9]+]] = private unnamed_addr constant <{ [10 x i8] }> undef
// CHECK: [[FULLY_UNINIT:@.*]] = private unnamed_addr constant <{ [10 x i8] }> undef
// CHECK: [[PARTIALLY_UNINIT:@[0-9]+]] = private unnamed_addr constant <{ [4 x i8], [12 x i8] }> <{ [4 x i8] c"{{\\EF\\BE\\AD\\DE|\\DE\\AD\\BE\\EF}}", [12 x i8] undef }>, align 4
// CHECK: [[PARTIALLY_UNINIT:@.*]] = private unnamed_addr constant <{ [4 x i8], [12 x i8] }> <{ [4 x i8] c"{{\\EF\\BE\\AD\\DE|\\DE\\AD\\BE\\EF}}", [12 x i8] undef }>, align 4
// This shouldn't contain undef, since it contains more chunks
// than the default value of uninit_const_chunk_threshold.
// CHECK: [[UNINIT_PADDING_HUGE:@[0-9]+]] = private unnamed_addr constant <{ [32768 x i8] }> <{ [32768 x i8] c"{{.+}}" }>, align 4
// CHECK: [[UNINIT_PADDING_HUGE:@.*]] = private unnamed_addr constant <{ [32768 x i8] }> <{ [32768 x i8] c"{{.+}}" }>, align 4
// CHECK: [[FULLY_UNINIT_HUGE:@[0-9]+]] = private unnamed_addr constant <{ [16384 x i8] }> undef
// CHECK: [[FULLY_UNINIT_HUGE:@.*]] = private unnamed_addr constant <{ [16384 x i8] }> undef
// CHECK-LABEL: @fully_uninit
#[no_mangle]

View File

@ -0,0 +1,16 @@
#![crate_name = "interesting"]
#![crate_type = "rlib"]
extern crate opaque;
#[no_mangle]
#[inline(never)]
pub fn function_called_once() {
opaque::foo();
}
// CHECK-LABEL: @function_called_once
// CHECK-SAME: !prof [[function_called_once_id:![0-9]+]] {
// CHECK: "CG Profile"
// CHECK-NOT: "CG Profile"
// CHECK-DAG: [[function_called_once_id]] = !{!"function_entry_count", i64 1}

View File

@ -0,0 +1,5 @@
extern crate interesting;
fn main() {
interesting::function_called_once();
}

View File

@ -0,0 +1,5 @@
#![crate_name = "opaque"]
#![crate_type = "rlib"]
#[inline(never)]
pub fn foo() {}

View File

@ -0,0 +1,67 @@
// This test case verifies that we successfully complete an LTO build with PGO
// using the embedded bitcode.
// It also ensures that the generated IR correctly includes the call results.
//@ needs-profiler-runtime
//@ ignore-cross-compile
use std::path::Path;
use run_make_support::{
has_extension, has_prefix, llvm_filecheck, llvm_profdata, rfs, run, rustc, shallow_find_files,
};
fn run_test(cg_units: usize) {
let path_prof_data_dir = Path::new("prof_data_dir");
if path_prof_data_dir.exists() {
rfs::remove_dir_all(path_prof_data_dir);
}
rfs::create_dir_all(&path_prof_data_dir);
let path_merged_profdata = path_prof_data_dir.join("merged.profdata");
rustc().input("opaque.rs").codegen_units(1).run();
rustc()
.input("interesting.rs")
.profile_generate(&path_prof_data_dir)
.opt()
.crate_type("lib,cdylib")
.codegen_units(cg_units)
.run();
rustc()
.input("main.rs")
.arg("-Clto=thin")
.opt()
.codegen_units(cg_units)
.profile_generate(&path_prof_data_dir)
.opt()
.run();
run("main");
llvm_profdata().merge().output(&path_merged_profdata).input(path_prof_data_dir).run();
rustc()
.input("interesting.rs")
.profile_use(&path_merged_profdata)
.opt()
.crate_type("lib,cdylib")
.codegen_units(cg_units)
.emit("link")
.run();
rustc()
.input("main.rs")
.arg("-Clto=thin")
.opt()
.codegen_units(cg_units)
.profile_use(&path_merged_profdata)
.emit("llvm-ir,link")
.opt()
.run();
let files = shallow_find_files(".", |path| {
has_prefix(path, "main.interesting.interesting") && has_extension(path, "ll")
});
assert_eq!(files.len(), 1);
let llvm_ir = &files[0];
llvm_filecheck().patterns("interesting.rs").stdin_buf(rfs::read(llvm_ir)).run();
}
fn main() {
run_test(1);
run_test(16);
}