IPO: Add use-list-order verifier

Add a -verify-use-list-order pass, which shuffles use-list order, writes
to bitcode, reads back, and verifies that the (shuffled) order matches.

  - The utility functions live in lib/IR/UseListOrder.cpp.

  - Moved (and renamed) the command-line option to enable writing
    use-lists, so that this pass can return early if the use-list orders
    aren't being serialized.

It's not clear that this pass is the right direction long-term (perhaps
a separate tool instead?), but short-term it's a great way to test the
use-list order prototype.  I've added an XFAIL-ed testcase that I'm
hoping to get working pretty quickly.

This is part of PR5680.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213945 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Duncan P. N. Exon Smith 2014-07-25 14:49:26 +00:00
parent db4f73f6d9
commit 7bf73bd378
11 changed files with 621 additions and 7 deletions

View File

@ -0,0 +1,42 @@
//===- llvm/IR/UseListOrder.h - LLVM Use List Order functions ---*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file has functions to modify the use-list order and to verify that it
// doesn't change after serialization.
//
//===----------------------------------------------------------------------===//
#ifndef LLVM_IR_USELISTORDER_H
#define LLVM_IR_USELISTORDER_H
namespace llvm {
class Module;
/// \brief Whether to preserve use-list ordering.
bool shouldPreserveBitcodeUseListOrder();
/// \brief Shuffle all use-lists in a module.
///
/// Adds \c SeedOffset to the default seed for the random number generator.
void shuffleUseLists(Module &M, unsigned SeedOffset = 0);
/// \brief Verify use-list order after serializing to bitcode.
///
/// \return \c true if there are no errors.
bool verifyBitcodeUseListOrder(const Module &M);
/// \brief Verify use-list order after serializing to assembly.
///
/// \return \c true if there are no errors.
bool verifyAssemblyUseListOrder(const Module &M);
} // end namespace llvm
#endif

View File

@ -268,6 +268,7 @@ void initializeUnifyFunctionExitNodesPass(PassRegistry&);
void initializeUnreachableBlockElimPass(PassRegistry&);
void initializeUnreachableMachineBlockElimPass(PassRegistry&);
void initializeVerifierLegacyPassPass(PassRegistry&);
void initializeVerifyUseListOrderPass(PassRegistry&);
void initializeVirtRegMapPass(PassRegistry&);
void initializeVirtRegRewriterPass(PassRegistry&);
void initializeInstSimplifierPass(PassRegistry&);

View File

@ -161,6 +161,7 @@ namespace {
(void) llvm::createPartiallyInlineLibCallsPass();
(void) llvm::createScalarizerPass();
(void) llvm::createSeparateConstOffsetFromGEPPass();
(void) llvm::createVerifyUseListOrderPass();
(void)new llvm::IntervalPartition();
(void)new llvm::FindUsedTypes();

View File

@ -118,6 +118,11 @@ ModulePass *createInternalizePass(ArrayRef<const char *> ExportList);
/// createInternalizePass - Same as above, but with an empty exportList.
ModulePass *createInternalizePass();
/// \brief Verify that use-list order doesn't change after shuffling.
///
/// \note This is a transformation, since the use-list order changes.
ModulePass *createVerifyUseListOrderPass();
//===----------------------------------------------------------------------===//
/// createDeadArgEliminationPass - This pass removes arguments from functions
/// which are not used by the body of the function.

View File

@ -22,6 +22,7 @@
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/Operator.h"
#include "llvm/IR/UseListOrder.h"
#include "llvm/IR/ValueSymbolTable.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
@ -32,12 +33,6 @@
#include <map>
using namespace llvm;
static cl::opt<bool>
EnablePreserveUseListOrdering("enable-bc-uselist-preserve",
cl::desc("Turn on experimental support for "
"use-list order preservation."),
cl::init(false), cl::Hidden);
/// These are manifest constants used by the bitcode writer. They do not need to
/// be kept in sync with the reader, but need to be consistent within this file.
enum {
@ -1975,7 +1970,7 @@ static void WriteModule(const Module *M, BitstreamWriter &Stream) {
WriteValueSymbolTable(M->getValueSymbolTable(), VE, Stream);
// Emit use-lists.
if (EnablePreserveUseListOrdering)
if (shouldPreserveBitcodeUseListOrder())
WriteModuleUseLists(M, VE, Stream);
// Emit function bodies.

View File

@ -39,6 +39,7 @@ add_llvm_library(LLVMCore
Type.cpp
TypeFinder.cpp
Use.cpp
UseListOrder.cpp
User.cpp
Value.cpp
ValueSymbolTable.cpp

423
lib/IR/UseListOrder.cpp Normal file
View File

@ -0,0 +1,423 @@
//===- UseListOrder.cpp - Implement Use List Order functions --------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// Implement use list order functions to modify use-list order and verify it
// doesn't change after serialization.
//
//===----------------------------------------------------------------------===//
#include "llvm/IR/UseListOrder.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DenseSet.h"
#include "llvm/AsmParser/Parser.h"
#include "llvm/Bitcode/ReaderWriter.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Module.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/FileUtilities.h"
#include "llvm/Support/MemoryBuffer.h"
#include "llvm/Support/SourceMgr.h"
#include <random>
#include <vector>
#define DEBUG_TYPE "use-list-order"
using namespace llvm;
static cl::opt<bool> PreserveBitcodeUseListOrder(
"preserve-bc-use-list-order",
cl::desc("Experimental support to preserve bitcode use-list order."),
cl::init(false), cl::Hidden);
bool llvm::shouldPreserveBitcodeUseListOrder() {
return PreserveBitcodeUseListOrder;
}
static void shuffleValueUseLists(Value *V, std::minstd_rand0 &Gen,
DenseSet<Value *> &Seen) {
if (!Seen.insert(V).second)
return;
if (auto *C = dyn_cast<Constant>(V))
if (!isa<GlobalValue>(C))
for (Value *Op : C->operands())
shuffleValueUseLists(Op, Gen, Seen);
if (V->use_empty() || std::next(V->use_begin()) == V->use_end())
// Nothing to shuffle for 0 or 1 users.
return;
// Generate random numbers between 10 and 99, which will line up nicely in
// debug output. We're not worried about collisons here.
DEBUG(dbgs() << "V = "; V->dump());
std::uniform_int_distribution<short> Dist(10, 99);
SmallDenseMap<const Use *, short, 16> Order;
for (const Use &U : V->uses()) {
auto I = Dist(Gen);
Order[&U] = I;
DEBUG(dbgs() << " - order: " << I << ", U = "; U.getUser()->dump());
}
DEBUG(dbgs() << " => shuffle\n");
V->sortUseList(
[&Order](const Use &L, const Use &R) { return Order[&L] < Order[&R]; });
DEBUG({
for (const Use &U : V->uses())
DEBUG(dbgs() << " - order: " << Order.lookup(&U) << ", U = ";
U.getUser()->dump());
});
}
void llvm::shuffleUseLists(Module &M, unsigned SeedOffset) {
DEBUG(dbgs() << "*** shuffle-use-lists ***\n");
std::minstd_rand0 Gen(std::minstd_rand0::default_seed + SeedOffset);
DenseSet<Value *> Seen;
// Shuffle the use-list of each value that would be serialized to an IR file
// (bitcode or assembly).
auto shuffle = [&](Value *V) { shuffleValueUseLists(V, Gen, Seen); };
// Globals.
for (GlobalVariable &G : M.globals())
shuffle(&G);
for (GlobalAlias &A : M.aliases())
shuffle(&A);
for (Function &F : M)
shuffle(&F);
// Constants used by globals.
for (GlobalVariable &G : M.globals())
if (G.hasInitializer())
shuffle(G.getInitializer());
for (GlobalAlias &A : M.aliases())
shuffle(A.getAliasee());
for (Function &F : M)
if (F.hasPrefixData())
shuffle(F.getPrefixData());
// Function bodies.
for (Function &F : M) {
for (Argument &A : F.args())
shuffle(&A);
for (BasicBlock &BB : F)
shuffle(&BB);
for (BasicBlock &BB : F)
for (Instruction &I : BB)
shuffle(&I);
// Constants used by instructions.
for (BasicBlock &BB : F)
for (Instruction &I : BB)
for (Value *Op : I.operands())
if ((isa<Constant>(Op) && !isa<GlobalValue>(*Op)) ||
isa<InlineAsm>(Op))
shuffle(Op);
}
DEBUG(dbgs() << "\n");
}
namespace {
struct TempFile {
std::string Filename;
FileRemover Remover;
bool init(const std::string &Ext);
bool writeBitcode(const Module &M) const;
bool writeAssembly(const Module &M) const;
std::unique_ptr<Module> readBitcode(LLVMContext &Context) const;
std::unique_ptr<Module> readAssembly(LLVMContext &Context) const;
};
struct ValueMapping {
DenseMap<const Value *, unsigned> IDs;
std::vector<const Value *> Values;
/// \brief Construct a value mapping for module.
///
/// Creates mapping from every value in \c M to an ID. This mapping includes
/// un-referencable values.
///
/// Every \a Value that gets serialized in some way should be represented
/// here. The order needs to be deterministic, but it's unnecessary to match
/// the value-ids in the bitcode writer.
///
/// All constants that are referenced by other values are included in the
/// mapping, but others -- which wouldn't be serialized -- are not.
ValueMapping(const Module &M);
/// \brief Map a value.
///
/// Maps a value. If it's a constant, maps all of its operands first.
void map(const Value *V);
unsigned lookup(const Value *V) const { return IDs.lookup(V); }
};
} // end namespace
bool TempFile::init(const std::string &Ext) {
SmallVector<char, 64> Vector;
DEBUG(dbgs() << " - create-temp-file\n");
if (auto EC = sys::fs::createTemporaryFile("use-list-order", Ext, Vector)) {
DEBUG(dbgs() << "error: " << EC.message() << "\n");
return true;
}
assert(!Vector.empty());
Filename.assign(Vector.data(), Vector.data() + Vector.size());
Remover.setFile(Filename);
DEBUG(dbgs() << " - filename = " << Filename << "\n");
return false;
}
bool TempFile::writeBitcode(const Module &M) const {
DEBUG(dbgs() << " - write bitcode\n");
std::string ErrorInfo;
raw_fd_ostream OS(Filename.c_str(), ErrorInfo, sys::fs::F_None);
if (!ErrorInfo.empty()) {
DEBUG(dbgs() << "error: " << ErrorInfo << "\n");
return true;
}
WriteBitcodeToFile(&M, OS);
return false;
}
bool TempFile::writeAssembly(const Module &M) const {
DEBUG(dbgs() << " - write assembly\n");
std::string ErrorInfo;
raw_fd_ostream OS(Filename.c_str(), ErrorInfo, sys::fs::F_Text);
if (!ErrorInfo.empty()) {
DEBUG(dbgs() << "error: " << ErrorInfo << "\n");
return true;
}
OS << M;
return false;
}
std::unique_ptr<Module> TempFile::readBitcode(LLVMContext &Context) const {
DEBUG(dbgs() << " - read bitcode\n");
ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOr =
MemoryBuffer::getFile(Filename);
if (!BufferOr) {
DEBUG(dbgs() << "error: " << BufferOr.getError().message() << "\n");
return nullptr;
}
std::unique_ptr<MemoryBuffer> Buffer = std::move(BufferOr.get());
ErrorOr<Module *> ModuleOr = parseBitcodeFile(Buffer.release(), Context);
if (!ModuleOr) {
DEBUG(dbgs() << "error: " << ModuleOr.getError().message() << "\n");
return nullptr;
}
return std::unique_ptr<Module>(ModuleOr.get());
}
std::unique_ptr<Module> TempFile::readAssembly(LLVMContext &Context) const {
DEBUG(dbgs() << " - read assembly\n");
SMDiagnostic Err;
std::unique_ptr<Module> M(ParseAssemblyFile(Filename, Err, Context));
if (!M.get())
DEBUG(dbgs() << "error: "; Err.print("verify-use-list-order", dbgs()));
return M;
}
ValueMapping::ValueMapping(const Module &M) {
// Every value should be mapped, including things like void instructions and
// basic blocks that are kept out of the ValueEnumerator.
//
// The current mapping order makes it easier to debug the tables. It happens
// to be similar to the ID mapping when writing ValueEnumerator, but they
// aren't (and needn't be) in sync.
// Globals.
for (const GlobalVariable &G : M.globals())
map(&G);
for (const GlobalAlias &A : M.aliases())
map(&A);
for (const Function &F : M)
map(&F);
// Constants used by globals.
for (const GlobalVariable &G : M.globals())
if (G.hasInitializer())
map(G.getInitializer());
for (const GlobalAlias &A : M.aliases())
map(A.getAliasee());
for (const Function &F : M)
if (F.hasPrefixData())
map(F.getPrefixData());
// Function bodies.
for (const Function &F : M) {
for (const Argument &A : F.args())
map(&A);
for (const BasicBlock &BB : F)
map(&BB);
for (const BasicBlock &BB : F)
for (const Instruction &I : BB)
map(&I);
// Constants used by instructions.
for (const BasicBlock &BB : F)
for (const Instruction &I : BB)
for (const Value *Op : I.operands())
if ((isa<Constant>(Op) && !isa<GlobalValue>(*Op)) ||
isa<InlineAsm>(Op))
map(Op);
}
}
void ValueMapping::map(const Value *V) {
if (IDs.lookup(V))
return;
if (auto *C = dyn_cast<Constant>(V))
if (!isa<GlobalValue>(C))
for (const Value *Op : C->operands())
map(Op);
Values.push_back(V);
IDs[V] = Values.size();
}
#ifndef NDEBUG
static void dumpMapping(const ValueMapping &VM) {
dbgs() << "value-mapping (size = " << VM.Values.size() << "):\n";
for (unsigned I = 0, E = VM.Values.size(); I != E; ++I) {
dbgs() << " - id = " << I << ", value = ";
VM.Values[I]->dump();
}
}
static void debugValue(const ValueMapping &M, unsigned I, StringRef Desc) {
const Value *V = M.Values[I];
dbgs() << " - " << Desc << " value = ";
V->dump();
for (const Use &U : V->uses()) {
dbgs() << " => use: op = " << U.getOperandNo()
<< ", user-id = " << M.IDs.lookup(U.getUser()) << ", user = ";
U.getUser()->dump();
}
}
static void debugUserMismatch(const ValueMapping &L, const ValueMapping &R,
unsigned I) {
dbgs() << " - fail: user mismatch: ID = " << I << "\n";
debugValue(L, I, "LHS");
debugValue(R, I, "RHS");
dbgs() << "\nlhs-";
dumpMapping(L);
dbgs() << "\nrhs-";
dumpMapping(R);
}
static void debugSizeMismatch(const ValueMapping &L, const ValueMapping &R) {
dbgs() << " - fail: map size: " << L.Values.size()
<< " != " << R.Values.size() << "\n";
dbgs() << "\nlhs-";
dumpMapping(L);
dbgs() << "\nrhs-";
dumpMapping(R);
}
#endif
static bool matches(const ValueMapping &LM, const ValueMapping &RM) {
DEBUG(dbgs() << "compare value maps\n");
if (LM.Values.size() != RM.Values.size()) {
DEBUG(debugSizeMismatch(LM, RM));
return false;
}
// This mapping doesn't include dangling constant users, since those don't
// get serialized. However, checking if users are constant and calling
// isConstantUsed() on every one is very expensive. Instead, just check if
// the user is mapped.
auto skipUnmappedUsers =
[&](Value::const_use_iterator &U, Value::const_use_iterator E,
const ValueMapping &M) {
while (U != E && !M.lookup(U->getUser()))
++U;
};
// Iterate through all values, and check that both mappings have the same
// users.
for (unsigned I = 0, E = LM.Values.size(); I != E; ++I) {
const Value *L = LM.Values[I];
const Value *R = RM.Values[I];
auto LU = L->use_begin(), LE = L->use_end();
auto RU = R->use_begin(), RE = R->use_end();
skipUnmappedUsers(LU, LE, LM);
skipUnmappedUsers(RU, RE, RM);
while (LU != LE) {
if (RU == RE) {
DEBUG(debugUserMismatch(LM, RM, I));
return false;
}
if (LM.lookup(LU->getUser()) != RM.lookup(RU->getUser())) {
DEBUG(debugUserMismatch(LM, RM, I));
return false;
}
if (LU->getOperandNo() != RU->getOperandNo()) {
DEBUG(debugUserMismatch(LM, RM, I));
return false;
}
skipUnmappedUsers(++LU, LE, LM);
skipUnmappedUsers(++RU, RE, RM);
}
if (RU != RE) {
DEBUG(debugUserMismatch(LM, RM, I));
return false;
}
}
return true;
}
bool llvm::verifyBitcodeUseListOrder(const Module &M) {
DEBUG(dbgs() << "*** verify-use-list-order: bitcode ***\n");
TempFile F;
if (F.init("bc"))
return false;
if (F.writeBitcode(M))
return false;
LLVMContext Context;
std::unique_ptr<Module> OtherM = F.readBitcode(Context);
if (!OtherM)
return false;
return matches(ValueMapping(M), ValueMapping(*OtherM));
}
bool llvm::verifyAssemblyUseListOrder(const Module &M) {
DEBUG(dbgs() << "*** verify-use-list-order: assembly ***\n");
TempFile F;
if (F.init("ll"))
return false;
if (F.writeAssembly(M))
return false;
LLVMContext Context;
std::unique_ptr<Module> OtherM = F.readAssembly(Context);
if (!OtherM)
return false;
return matches(ValueMapping(M), ValueMapping(*OtherM));
}

View File

@ -20,6 +20,7 @@ add_llvm_library(LLVMipo
PruneEH.cpp
StripDeadPrototypes.cpp
StripSymbols.cpp
VerifyUseListOrder.cpp
)
add_dependencies(LLVMipo intrinsics_gen)

View File

@ -44,6 +44,7 @@ void llvm::initializeIPO(PassRegistry &Registry) {
initializeStripDebugDeclarePass(Registry);
initializeStripDeadDebugInfoPass(Registry);
initializeStripNonDebugSymbolsPass(Registry);
initializeVerifyUseListOrderPass(Registry);
initializeBarrierNoopPass(Registry);
}

View File

@ -0,0 +1,61 @@
//===- VerifyUseListOrder.cpp - Use List Order Verifier ---------*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// Pass to verify use-list order doesn't change after serialization.
//
// Despite it being a verifier, this pass *does* transform the module, since it
// shuffles the use-list of every value.
//
//===----------------------------------------------------------------------===//
#include "llvm/Transforms/IPO.h"
#include "llvm/IR/UseListOrder.h"
#include "llvm/Pass.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
using namespace llvm;
#define DEBUG_TYPE "use-list-order"
namespace {
class VerifyUseListOrder : public ModulePass {
public:
static char ID;
VerifyUseListOrder();
bool runOnModule(Module &M) override;
};
} // end anonymous namespace
char VerifyUseListOrder::ID = 0;
INITIALIZE_PASS(VerifyUseListOrder, "verify-use-list-order",
"Verify Use List Order", false, false)
VerifyUseListOrder::VerifyUseListOrder() : ModulePass(ID) {
initializeVerifyUseListOrderPass(*PassRegistry::getPassRegistry());
}
bool VerifyUseListOrder::runOnModule(Module &M) {
DEBUG(dbgs() << "*** verify-use-list-order ***\n");
if (!shouldPreserveBitcodeUseListOrder()) {
// Can't verify if order isn't preserved.
DEBUG(dbgs() << "warning: cannot verify bitcode; "
"try -preserve-bc-use-list-order\n");
return false;
}
shuffleUseLists(M);
if (!verifyBitcodeUseListOrder(M))
report_fatal_error("bitcode use-list order changed");
return true;
}
ModulePass *llvm::createVerifyUseListOrderPass() {
return new VerifyUseListOrder;
}

View File

@ -0,0 +1,83 @@
; RUN: opt -S < %s -preserve-bc-use-list-order -verify-use-list-order
; XFAIL: *
@a = global [4 x i1] [i1 0, i1 1, i1 0, i1 1]
@b = alias i1* getelementptr ([4 x i1]* @a, i64 0, i64 2)
define i64 @f(i64 %f) {
entry:
%sum = add i64 %f, 0
ret i64 %sum
}
define i64 @g(i64 %g) {
entry:
%sum = add i64 %g, 0
ret i64 %sum
}
define i64 @h(i64 %h) {
entry:
%sum = add i64 %h, 0
ret i64 %sum
}
define i64 @i(i64 %i) {
entry:
%sum = add i64 %i, 1
ret i64 %sum
}
define i64 @j(i64 %j) {
entry:
%sum = add i64 %j, 1
ret i64 %sum
}
define i64 @k(i64 %k) {
entry:
%sum = add i64 %k, 1
ret i64 %sum
}
define i64 @l(i64 %l) {
entry:
%sum = add i64 %l, 1
ret i64 %sum
}
define i1 @loadb() {
entry:
%b = load i1* @b
ret i1 %b
}
define i1 @loada() {
entry:
%a = load i1* getelementptr ([4 x i1]* @a, i64 0, i64 2)
ret i1 %a
}
define i32 @f32(i32 %a, i32 %b, i32 %c, i32 %d) {
entry:
br label %first
second:
%eh = mul i32 %e, %h
%sum = add i32 %eh, %ef
br label %exit
exit:
%product = phi i32 [%ef, %first], [%sum, %second]
ret i32 %product
first:
%e = add i32 %a, 7
%f = add i32 %b, 7
%g = add i32 %c, 8
%h = add i32 %d, 8
%ef = mul i32 %e, %f
%gh = mul i32 %g, %h
%gotosecond = icmp slt i32 %gh, -9
br i1 %gotosecond, label %second, label %exit
}