2016-02-10 06:50:34 +08:00
|
|
|
//===- WholeProgramDevirt.cpp - Whole program virtual call optimization ---===//
|
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This pass implements whole program optimization of virtual calls in cases
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
// where we know (via !type metadata) that the list of callees is fixed. This
|
2016-02-10 06:50:34 +08:00
|
|
|
// includes the following:
|
|
|
|
// - Single implementation devirtualization: if a virtual call has a single
|
|
|
|
// possible callee, replace all calls with a direct call to that callee.
|
|
|
|
// - Virtual constant propagation: if the virtual function's return type is an
|
|
|
|
// integer <=64 bits and all possible callees are readnone, for each class and
|
|
|
|
// each list of constant arguments: evaluate the function, store the return
|
|
|
|
// value alongside the virtual table, and rewrite each virtual call as a load
|
|
|
|
// from the virtual table.
|
|
|
|
// - Uniform return value optimization: if the conditions for virtual constant
|
|
|
|
// propagation hold and each function returns the same constant value, replace
|
|
|
|
// each virtual call with that constant.
|
|
|
|
// - Unique return value optimization for i1 return values: if the conditions
|
|
|
|
// for virtual constant propagation hold and a single vtable's function
|
|
|
|
// returns 0, or a single vtable's function returns 1, replace each virtual
|
|
|
|
// call with a comparison of the vptr against that vtable's address.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#include "llvm/Transforms/IPO/WholeProgramDevirt.h"
|
2016-04-18 17:17:29 +08:00
|
|
|
#include "llvm/ADT/ArrayRef.h"
|
2016-02-10 06:50:34 +08:00
|
|
|
#include "llvm/ADT/DenseSet.h"
|
|
|
|
#include "llvm/ADT/MapVector.h"
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
#include "llvm/Analysis/TypeMetadataUtils.h"
|
2016-02-10 06:50:34 +08:00
|
|
|
#include "llvm/IR/CallSite.h"
|
|
|
|
#include "llvm/IR/Constants.h"
|
|
|
|
#include "llvm/IR/DataLayout.h"
|
|
|
|
#include "llvm/IR/IRBuilder.h"
|
|
|
|
#include "llvm/IR/Instructions.h"
|
|
|
|
#include "llvm/IR/Intrinsics.h"
|
|
|
|
#include "llvm/IR/Module.h"
|
|
|
|
#include "llvm/Pass.h"
|
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2016-04-18 17:17:29 +08:00
|
|
|
#include "llvm/Transforms/IPO.h"
|
2016-02-10 06:50:34 +08:00
|
|
|
#include "llvm/Transforms/Utils/Evaluator.h"
|
|
|
|
#include "llvm/Transforms/Utils/Local.h"
|
|
|
|
|
|
|
|
#include <set>
|
|
|
|
|
|
|
|
using namespace llvm;
|
|
|
|
using namespace wholeprogramdevirt;
|
|
|
|
|
|
|
|
#define DEBUG_TYPE "wholeprogramdevirt"
|
|
|
|
|
|
|
|
// Find the minimum offset that we may store a value of size Size bits at. If
|
|
|
|
// IsAfter is set, look for an offset before the object, otherwise look for an
|
|
|
|
// offset after the object.
|
|
|
|
uint64_t
|
|
|
|
wholeprogramdevirt::findLowestOffset(ArrayRef<VirtualCallTarget> Targets,
|
|
|
|
bool IsAfter, uint64_t Size) {
|
|
|
|
// Find a minimum offset taking into account only vtable sizes.
|
|
|
|
uint64_t MinByte = 0;
|
|
|
|
for (const VirtualCallTarget &Target : Targets) {
|
|
|
|
if (IsAfter)
|
|
|
|
MinByte = std::max(MinByte, Target.minAfterBytes());
|
|
|
|
else
|
|
|
|
MinByte = std::max(MinByte, Target.minBeforeBytes());
|
|
|
|
}
|
|
|
|
|
|
|
|
// Build a vector of arrays of bytes covering, for each target, a slice of the
|
|
|
|
// used region (see AccumBitVector::BytesUsed in
|
|
|
|
// llvm/Transforms/IPO/WholeProgramDevirt.h) starting at MinByte. Effectively,
|
|
|
|
// this aligns the used regions to start at MinByte.
|
|
|
|
//
|
|
|
|
// In this example, A, B and C are vtables, # is a byte already allocated for
|
|
|
|
// a virtual function pointer, AAAA... (etc.) are the used regions for the
|
|
|
|
// vtables and Offset(X) is the value computed for the Offset variable below
|
|
|
|
// for X.
|
|
|
|
//
|
|
|
|
// Offset(A)
|
|
|
|
// | |
|
|
|
|
// |MinByte
|
|
|
|
// A: ################AAAAAAAA|AAAAAAAA
|
|
|
|
// B: ########BBBBBBBBBBBBBBBB|BBBB
|
|
|
|
// C: ########################|CCCCCCCCCCCCCCCC
|
|
|
|
// | Offset(B) |
|
|
|
|
//
|
|
|
|
// This code produces the slices of A, B and C that appear after the divider
|
|
|
|
// at MinByte.
|
|
|
|
std::vector<ArrayRef<uint8_t>> Used;
|
|
|
|
for (const VirtualCallTarget &Target : Targets) {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
ArrayRef<uint8_t> VTUsed = IsAfter ? Target.TM->Bits->After.BytesUsed
|
|
|
|
: Target.TM->Bits->Before.BytesUsed;
|
2016-02-10 06:50:34 +08:00
|
|
|
uint64_t Offset = IsAfter ? MinByte - Target.minAfterBytes()
|
|
|
|
: MinByte - Target.minBeforeBytes();
|
|
|
|
|
|
|
|
// Disregard used regions that are smaller than Offset. These are
|
|
|
|
// effectively all-free regions that do not need to be checked.
|
|
|
|
if (VTUsed.size() > Offset)
|
|
|
|
Used.push_back(VTUsed.slice(Offset));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (Size == 1) {
|
|
|
|
// Find a free bit in each member of Used.
|
|
|
|
for (unsigned I = 0;; ++I) {
|
|
|
|
uint8_t BitsUsed = 0;
|
|
|
|
for (auto &&B : Used)
|
|
|
|
if (I < B.size())
|
|
|
|
BitsUsed |= B[I];
|
|
|
|
if (BitsUsed != 0xff)
|
|
|
|
return (MinByte + I) * 8 +
|
|
|
|
countTrailingZeros(uint8_t(~BitsUsed), ZB_Undefined);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// Find a free (Size/8) byte region in each member of Used.
|
|
|
|
// FIXME: see if alignment helps.
|
|
|
|
for (unsigned I = 0;; ++I) {
|
|
|
|
for (auto &&B : Used) {
|
|
|
|
unsigned Byte = 0;
|
|
|
|
while ((I + Byte) < B.size() && Byte < (Size / 8)) {
|
|
|
|
if (B[I + Byte])
|
|
|
|
goto NextI;
|
|
|
|
++Byte;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return (MinByte + I) * 8;
|
|
|
|
NextI:;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void wholeprogramdevirt::setBeforeReturnValues(
|
|
|
|
MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocBefore,
|
|
|
|
unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
|
|
|
|
if (BitWidth == 1)
|
|
|
|
OffsetByte = -(AllocBefore / 8 + 1);
|
|
|
|
else
|
|
|
|
OffsetByte = -((AllocBefore + 7) / 8 + (BitWidth + 7) / 8);
|
|
|
|
OffsetBit = AllocBefore % 8;
|
|
|
|
|
|
|
|
for (VirtualCallTarget &Target : Targets) {
|
|
|
|
if (BitWidth == 1)
|
|
|
|
Target.setBeforeBit(AllocBefore);
|
|
|
|
else
|
|
|
|
Target.setBeforeBytes(AllocBefore, (BitWidth + 7) / 8);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void wholeprogramdevirt::setAfterReturnValues(
|
|
|
|
MutableArrayRef<VirtualCallTarget> Targets, uint64_t AllocAfter,
|
|
|
|
unsigned BitWidth, int64_t &OffsetByte, uint64_t &OffsetBit) {
|
|
|
|
if (BitWidth == 1)
|
|
|
|
OffsetByte = AllocAfter / 8;
|
|
|
|
else
|
|
|
|
OffsetByte = (AllocAfter + 7) / 8;
|
|
|
|
OffsetBit = AllocAfter % 8;
|
|
|
|
|
|
|
|
for (VirtualCallTarget &Target : Targets) {
|
|
|
|
if (BitWidth == 1)
|
|
|
|
Target.setAfterBit(AllocAfter);
|
|
|
|
else
|
|
|
|
Target.setAfterBytes(AllocAfter, (BitWidth + 7) / 8);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
VirtualCallTarget::VirtualCallTarget(Function *Fn, const TypeMemberInfo *TM)
|
|
|
|
: Fn(Fn), TM(TM),
|
2016-02-10 06:50:34 +08:00
|
|
|
IsBigEndian(Fn->getParent()->getDataLayout().isBigEndian()) {}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
// A slot in a set of virtual tables. The TypeID identifies the set of virtual
|
2016-02-10 06:50:34 +08:00
|
|
|
// tables, and the ByteOffset is the offset in bytes from the address point to
|
|
|
|
// the virtual function pointer.
|
|
|
|
struct VTableSlot {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
Metadata *TypeID;
|
2016-02-10 06:50:34 +08:00
|
|
|
uint64_t ByteOffset;
|
|
|
|
};
|
|
|
|
|
|
|
|
}
|
|
|
|
|
2016-02-10 07:01:38 +08:00
|
|
|
namespace llvm {
|
|
|
|
|
2016-02-10 06:50:34 +08:00
|
|
|
template <> struct DenseMapInfo<VTableSlot> {
|
|
|
|
static VTableSlot getEmptyKey() {
|
|
|
|
return {DenseMapInfo<Metadata *>::getEmptyKey(),
|
|
|
|
DenseMapInfo<uint64_t>::getEmptyKey()};
|
|
|
|
}
|
|
|
|
static VTableSlot getTombstoneKey() {
|
|
|
|
return {DenseMapInfo<Metadata *>::getTombstoneKey(),
|
|
|
|
DenseMapInfo<uint64_t>::getTombstoneKey()};
|
|
|
|
}
|
|
|
|
static unsigned getHashValue(const VTableSlot &I) {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
return DenseMapInfo<Metadata *>::getHashValue(I.TypeID) ^
|
2016-02-10 06:50:34 +08:00
|
|
|
DenseMapInfo<uint64_t>::getHashValue(I.ByteOffset);
|
|
|
|
}
|
|
|
|
static bool isEqual(const VTableSlot &LHS,
|
|
|
|
const VTableSlot &RHS) {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
return LHS.TypeID == RHS.TypeID && LHS.ByteOffset == RHS.ByteOffset;
|
2016-02-10 06:50:34 +08:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2016-02-10 07:01:38 +08:00
|
|
|
}
|
|
|
|
|
2016-02-10 06:50:34 +08:00
|
|
|
namespace {
|
|
|
|
|
|
|
|
// A virtual call site. VTable is the loaded virtual table pointer, and CS is
|
|
|
|
// the indirect virtual call.
|
|
|
|
struct VirtualCallSite {
|
|
|
|
Value *VTable;
|
|
|
|
CallSite CS;
|
|
|
|
|
|
|
|
void replaceAndErase(Value *New) {
|
|
|
|
CS->replaceAllUsesWith(New);
|
|
|
|
if (auto II = dyn_cast<InvokeInst>(CS.getInstruction())) {
|
|
|
|
BranchInst::Create(II->getNormalDest(), CS.getInstruction());
|
|
|
|
II->getUnwindDest()->removePredecessor(II->getParent());
|
|
|
|
}
|
|
|
|
CS->eraseFromParent();
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
struct DevirtModule {
|
|
|
|
Module &M;
|
|
|
|
IntegerType *Int8Ty;
|
|
|
|
PointerType *Int8PtrTy;
|
|
|
|
IntegerType *Int32Ty;
|
|
|
|
|
|
|
|
MapVector<VTableSlot, std::vector<VirtualCallSite>> CallSlots;
|
|
|
|
|
|
|
|
DevirtModule(Module &M)
|
|
|
|
: M(M), Int8Ty(Type::getInt8Ty(M.getContext())),
|
|
|
|
Int8PtrTy(Type::getInt8PtrTy(M.getContext())),
|
|
|
|
Int32Ty(Type::getInt32Ty(M.getContext())) {}
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
void buildTypeIdentifierMap(
|
|
|
|
std::vector<VTableBits> &Bits,
|
|
|
|
DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap);
|
|
|
|
bool
|
|
|
|
tryFindVirtualCallTargets(std::vector<VirtualCallTarget> &TargetsForSlot,
|
|
|
|
const std::set<TypeMemberInfo> &TypeMemberInfos,
|
|
|
|
uint64_t ByteOffset);
|
2016-02-10 06:50:34 +08:00
|
|
|
bool trySingleImplDevirt(ArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
MutableArrayRef<VirtualCallSite> CallSites);
|
|
|
|
bool tryEvaluateFunctionsWithArgs(
|
|
|
|
MutableArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
ArrayRef<ConstantInt *> Args);
|
|
|
|
bool tryUniformRetValOpt(IntegerType *RetType,
|
|
|
|
ArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
MutableArrayRef<VirtualCallSite> CallSites);
|
|
|
|
bool tryUniqueRetValOpt(unsigned BitWidth,
|
|
|
|
ArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
MutableArrayRef<VirtualCallSite> CallSites);
|
|
|
|
bool tryVirtualConstProp(MutableArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
ArrayRef<VirtualCallSite> CallSites);
|
|
|
|
|
|
|
|
void rebuildGlobal(VTableBits &B);
|
|
|
|
|
|
|
|
bool run();
|
|
|
|
};
|
|
|
|
|
|
|
|
struct WholeProgramDevirt : public ModulePass {
|
|
|
|
static char ID;
|
|
|
|
WholeProgramDevirt() : ModulePass(ID) {
|
|
|
|
initializeWholeProgramDevirtPass(*PassRegistry::getPassRegistry());
|
|
|
|
}
|
2016-04-23 06:06:11 +08:00
|
|
|
bool runOnModule(Module &M) {
|
|
|
|
if (skipModule(M))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
return DevirtModule(M).run();
|
|
|
|
}
|
2016-02-10 06:50:34 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
} // anonymous namespace
|
|
|
|
|
|
|
|
INITIALIZE_PASS(WholeProgramDevirt, "wholeprogramdevirt",
|
|
|
|
"Whole program devirtualization", false, false)
|
|
|
|
char WholeProgramDevirt::ID = 0;
|
|
|
|
|
|
|
|
ModulePass *llvm::createWholeProgramDevirtPass() {
|
|
|
|
return new WholeProgramDevirt;
|
|
|
|
}
|
|
|
|
|
2016-06-17 08:11:01 +08:00
|
|
|
PreservedAnalyses WholeProgramDevirtPass::run(Module &M,
|
|
|
|
ModuleAnalysisManager &) {
|
2016-06-15 05:44:19 +08:00
|
|
|
if (!DevirtModule(M).run())
|
|
|
|
return PreservedAnalyses::all();
|
|
|
|
return PreservedAnalyses::none();
|
|
|
|
}
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
void DevirtModule::buildTypeIdentifierMap(
|
2016-02-10 06:50:34 +08:00
|
|
|
std::vector<VTableBits> &Bits,
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
DenseMap<Metadata *, std::set<TypeMemberInfo>> &TypeIdMap) {
|
2016-02-10 06:50:34 +08:00
|
|
|
DenseMap<GlobalVariable *, VTableBits *> GVToBits;
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
Bits.reserve(M.getGlobalList().size());
|
|
|
|
SmallVector<MDNode *, 2> Types;
|
|
|
|
for (GlobalVariable &GV : M.globals()) {
|
|
|
|
Types.clear();
|
|
|
|
GV.getMetadata(LLVMContext::MD_type, Types);
|
|
|
|
if (Types.empty())
|
2016-02-10 06:50:34 +08:00
|
|
|
continue;
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
VTableBits *&BitsPtr = GVToBits[&GV];
|
2016-02-10 06:50:34 +08:00
|
|
|
if (!BitsPtr) {
|
|
|
|
Bits.emplace_back();
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
Bits.back().GV = &GV;
|
|
|
|
Bits.back().ObjectSize =
|
|
|
|
M.getDataLayout().getTypeAllocSize(GV.getInitializer()->getType());
|
2016-02-10 06:50:34 +08:00
|
|
|
BitsPtr = &Bits.back();
|
|
|
|
}
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
|
|
|
|
for (MDNode *Type : Types) {
|
|
|
|
auto TypeID = Type->getOperand(1).get();
|
|
|
|
|
|
|
|
uint64_t Offset =
|
|
|
|
cast<ConstantInt>(
|
|
|
|
cast<ConstantAsMetadata>(Type->getOperand(0))->getValue())
|
|
|
|
->getZExtValue();
|
|
|
|
|
|
|
|
TypeIdMap[TypeID].insert({BitsPtr, Offset});
|
|
|
|
}
|
2016-02-10 06:50:34 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DevirtModule::tryFindVirtualCallTargets(
|
|
|
|
std::vector<VirtualCallTarget> &TargetsForSlot,
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
const std::set<TypeMemberInfo> &TypeMemberInfos, uint64_t ByteOffset) {
|
|
|
|
for (const TypeMemberInfo &TM : TypeMemberInfos) {
|
|
|
|
if (!TM.Bits->GV->isConstant())
|
2016-02-10 06:50:34 +08:00
|
|
|
return false;
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
auto Init = dyn_cast<ConstantArray>(TM.Bits->GV->getInitializer());
|
2016-02-10 06:50:34 +08:00
|
|
|
if (!Init)
|
|
|
|
return false;
|
|
|
|
ArrayType *VTableTy = Init->getType();
|
|
|
|
|
|
|
|
uint64_t ElemSize =
|
|
|
|
M.getDataLayout().getTypeAllocSize(VTableTy->getElementType());
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
uint64_t GlobalSlotOffset = TM.Offset + ByteOffset;
|
2016-02-10 06:50:34 +08:00
|
|
|
if (GlobalSlotOffset % ElemSize != 0)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
unsigned Op = GlobalSlotOffset / ElemSize;
|
|
|
|
if (Op >= Init->getNumOperands())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
auto Fn = dyn_cast<Function>(Init->getOperand(Op)->stripPointerCasts());
|
|
|
|
if (!Fn)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// We can disregard __cxa_pure_virtual as a possible call target, as
|
|
|
|
// calls to pure virtuals are UB.
|
|
|
|
if (Fn->getName() == "__cxa_pure_virtual")
|
|
|
|
continue;
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
TargetsForSlot.push_back({Fn, &TM});
|
2016-02-10 06:50:34 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
// Give up if we couldn't find any targets.
|
|
|
|
return !TargetsForSlot.empty();
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DevirtModule::trySingleImplDevirt(
|
|
|
|
ArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
MutableArrayRef<VirtualCallSite> CallSites) {
|
|
|
|
// See if the program contains a single implementation of this virtual
|
|
|
|
// function.
|
|
|
|
Function *TheFn = TargetsForSlot[0].Fn;
|
|
|
|
for (auto &&Target : TargetsForSlot)
|
|
|
|
if (TheFn != Target.Fn)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// If so, update each call site to call that implementation directly.
|
|
|
|
for (auto &&VCallSite : CallSites) {
|
|
|
|
VCallSite.CS.setCalledFunction(ConstantExpr::getBitCast(
|
|
|
|
TheFn, VCallSite.CS.getCalledValue()->getType()));
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DevirtModule::tryEvaluateFunctionsWithArgs(
|
|
|
|
MutableArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
ArrayRef<ConstantInt *> Args) {
|
|
|
|
// Evaluate each function and store the result in each target's RetVal
|
|
|
|
// field.
|
|
|
|
for (VirtualCallTarget &Target : TargetsForSlot) {
|
|
|
|
if (Target.Fn->arg_size() != Args.size() + 1)
|
|
|
|
return false;
|
|
|
|
for (unsigned I = 0; I != Args.size(); ++I)
|
|
|
|
if (Target.Fn->getFunctionType()->getParamType(I + 1) !=
|
|
|
|
Args[I]->getType())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
Evaluator Eval(M.getDataLayout(), nullptr);
|
|
|
|
SmallVector<Constant *, 2> EvalArgs;
|
|
|
|
EvalArgs.push_back(
|
|
|
|
Constant::getNullValue(Target.Fn->getFunctionType()->getParamType(0)));
|
|
|
|
EvalArgs.insert(EvalArgs.end(), Args.begin(), Args.end());
|
|
|
|
Constant *RetVal;
|
|
|
|
if (!Eval.EvaluateFunction(Target.Fn, RetVal, EvalArgs) ||
|
|
|
|
!isa<ConstantInt>(RetVal))
|
|
|
|
return false;
|
|
|
|
Target.RetVal = cast<ConstantInt>(RetVal)->getZExtValue();
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DevirtModule::tryUniformRetValOpt(
|
|
|
|
IntegerType *RetType, ArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
MutableArrayRef<VirtualCallSite> CallSites) {
|
|
|
|
// Uniform return value optimization. If all functions return the same
|
|
|
|
// constant, replace all calls with that constant.
|
|
|
|
uint64_t TheRetVal = TargetsForSlot[0].RetVal;
|
|
|
|
for (const VirtualCallTarget &Target : TargetsForSlot)
|
|
|
|
if (Target.RetVal != TheRetVal)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
auto TheRetValConst = ConstantInt::get(RetType, TheRetVal);
|
|
|
|
for (auto Call : CallSites)
|
|
|
|
Call.replaceAndErase(TheRetValConst);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DevirtModule::tryUniqueRetValOpt(
|
|
|
|
unsigned BitWidth, ArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
MutableArrayRef<VirtualCallSite> CallSites) {
|
|
|
|
// IsOne controls whether we look for a 0 or a 1.
|
|
|
|
auto tryUniqueRetValOptFor = [&](bool IsOne) {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
const TypeMemberInfo *UniqueMember = 0;
|
2016-02-10 06:50:34 +08:00
|
|
|
for (const VirtualCallTarget &Target : TargetsForSlot) {
|
2016-03-08 11:50:36 +08:00
|
|
|
if (Target.RetVal == (IsOne ? 1 : 0)) {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
if (UniqueMember)
|
2016-02-10 06:50:34 +08:00
|
|
|
return false;
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
UniqueMember = Target.TM;
|
2016-02-10 06:50:34 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
// We should have found a unique member or bailed out by now. We already
|
2016-02-10 06:50:34 +08:00
|
|
|
// checked for a uniform return value in tryUniformRetValOpt.
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
assert(UniqueMember);
|
2016-02-10 06:50:34 +08:00
|
|
|
|
|
|
|
// Replace each call with the comparison.
|
|
|
|
for (auto &&Call : CallSites) {
|
|
|
|
IRBuilder<> B(Call.CS.getInstruction());
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
Value *OneAddr = B.CreateBitCast(UniqueMember->Bits->GV, Int8PtrTy);
|
|
|
|
OneAddr = B.CreateConstGEP1_64(OneAddr, UniqueMember->Offset);
|
2016-02-10 06:50:34 +08:00
|
|
|
Value *Cmp = B.CreateICmp(IsOne ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE,
|
|
|
|
Call.VTable, OneAddr);
|
|
|
|
Call.replaceAndErase(Cmp);
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
};
|
|
|
|
|
|
|
|
if (BitWidth == 1) {
|
|
|
|
if (tryUniqueRetValOptFor(true))
|
|
|
|
return true;
|
|
|
|
if (tryUniqueRetValOptFor(false))
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DevirtModule::tryVirtualConstProp(
|
|
|
|
MutableArrayRef<VirtualCallTarget> TargetsForSlot,
|
|
|
|
ArrayRef<VirtualCallSite> CallSites) {
|
|
|
|
// This only works if the function returns an integer.
|
|
|
|
auto RetType = dyn_cast<IntegerType>(TargetsForSlot[0].Fn->getReturnType());
|
|
|
|
if (!RetType)
|
|
|
|
return false;
|
|
|
|
unsigned BitWidth = RetType->getBitWidth();
|
|
|
|
if (BitWidth > 64)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// Make sure that each function does not access memory, takes at least one
|
|
|
|
// argument, does not use its first argument (which we assume is 'this'),
|
|
|
|
// and has the same return type.
|
|
|
|
for (VirtualCallTarget &Target : TargetsForSlot) {
|
|
|
|
if (!Target.Fn->doesNotAccessMemory() || Target.Fn->arg_empty() ||
|
|
|
|
!Target.Fn->arg_begin()->use_empty() ||
|
|
|
|
Target.Fn->getReturnType() != RetType)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Group call sites by the list of constant arguments they pass.
|
|
|
|
// The comparator ensures deterministic ordering.
|
|
|
|
struct ByAPIntValue {
|
|
|
|
bool operator()(const std::vector<ConstantInt *> &A,
|
|
|
|
const std::vector<ConstantInt *> &B) const {
|
|
|
|
return std::lexicographical_compare(
|
|
|
|
A.begin(), A.end(), B.begin(), B.end(),
|
|
|
|
[](ConstantInt *AI, ConstantInt *BI) {
|
|
|
|
return AI->getValue().ult(BI->getValue());
|
|
|
|
});
|
|
|
|
}
|
|
|
|
};
|
|
|
|
std::map<std::vector<ConstantInt *>, std::vector<VirtualCallSite>,
|
|
|
|
ByAPIntValue>
|
|
|
|
VCallSitesByConstantArg;
|
|
|
|
for (auto &&VCallSite : CallSites) {
|
|
|
|
std::vector<ConstantInt *> Args;
|
|
|
|
if (VCallSite.CS.getType() != RetType)
|
|
|
|
continue;
|
|
|
|
for (auto &&Arg :
|
|
|
|
make_range(VCallSite.CS.arg_begin() + 1, VCallSite.CS.arg_end())) {
|
|
|
|
if (!isa<ConstantInt>(Arg))
|
|
|
|
break;
|
|
|
|
Args.push_back(cast<ConstantInt>(&Arg));
|
|
|
|
}
|
|
|
|
if (Args.size() + 1 != VCallSite.CS.arg_size())
|
|
|
|
continue;
|
|
|
|
|
|
|
|
VCallSitesByConstantArg[Args].push_back(VCallSite);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (auto &&CSByConstantArg : VCallSitesByConstantArg) {
|
|
|
|
if (!tryEvaluateFunctionsWithArgs(TargetsForSlot, CSByConstantArg.first))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (tryUniformRetValOpt(RetType, TargetsForSlot, CSByConstantArg.second))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (tryUniqueRetValOpt(BitWidth, TargetsForSlot, CSByConstantArg.second))
|
|
|
|
continue;
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
// Find an allocation offset in bits in all vtables associated with the
|
|
|
|
// type.
|
2016-02-10 06:50:34 +08:00
|
|
|
uint64_t AllocBefore =
|
|
|
|
findLowestOffset(TargetsForSlot, /*IsAfter=*/false, BitWidth);
|
|
|
|
uint64_t AllocAfter =
|
|
|
|
findLowestOffset(TargetsForSlot, /*IsAfter=*/true, BitWidth);
|
|
|
|
|
|
|
|
// Calculate the total amount of padding needed to store a value at both
|
|
|
|
// ends of the object.
|
|
|
|
uint64_t TotalPaddingBefore = 0, TotalPaddingAfter = 0;
|
|
|
|
for (auto &&Target : TargetsForSlot) {
|
|
|
|
TotalPaddingBefore += std::max<int64_t>(
|
|
|
|
(AllocBefore + 7) / 8 - Target.allocatedBeforeBytes() - 1, 0);
|
|
|
|
TotalPaddingAfter += std::max<int64_t>(
|
|
|
|
(AllocAfter + 7) / 8 - Target.allocatedAfterBytes() - 1, 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
// If the amount of padding is too large, give up.
|
|
|
|
// FIXME: do something smarter here.
|
|
|
|
if (std::min(TotalPaddingBefore, TotalPaddingAfter) > 128)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Calculate the offset to the value as a (possibly negative) byte offset
|
|
|
|
// and (if applicable) a bit offset, and store the values in the targets.
|
|
|
|
int64_t OffsetByte;
|
|
|
|
uint64_t OffsetBit;
|
|
|
|
if (TotalPaddingBefore <= TotalPaddingAfter)
|
|
|
|
setBeforeReturnValues(TargetsForSlot, AllocBefore, BitWidth, OffsetByte,
|
|
|
|
OffsetBit);
|
|
|
|
else
|
|
|
|
setAfterReturnValues(TargetsForSlot, AllocAfter, BitWidth, OffsetByte,
|
|
|
|
OffsetBit);
|
|
|
|
|
|
|
|
// Rewrite each call to a load from OffsetByte/OffsetBit.
|
|
|
|
for (auto Call : CSByConstantArg.second) {
|
|
|
|
IRBuilder<> B(Call.CS.getInstruction());
|
|
|
|
Value *Addr = B.CreateConstGEP1_64(Call.VTable, OffsetByte);
|
|
|
|
if (BitWidth == 1) {
|
|
|
|
Value *Bits = B.CreateLoad(Addr);
|
2016-03-31 05:30:00 +08:00
|
|
|
Value *Bit = ConstantInt::get(Int8Ty, 1ULL << OffsetBit);
|
2016-02-10 06:50:34 +08:00
|
|
|
Value *BitsAndBit = B.CreateAnd(Bits, Bit);
|
|
|
|
auto IsBitSet = B.CreateICmpNE(BitsAndBit, ConstantInt::get(Int8Ty, 0));
|
|
|
|
Call.replaceAndErase(IsBitSet);
|
|
|
|
} else {
|
|
|
|
Value *ValAddr = B.CreateBitCast(Addr, RetType->getPointerTo());
|
|
|
|
Value *Val = B.CreateLoad(RetType, ValAddr);
|
|
|
|
Call.replaceAndErase(Val);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
void DevirtModule::rebuildGlobal(VTableBits &B) {
|
|
|
|
if (B.Before.Bytes.empty() && B.After.Bytes.empty())
|
|
|
|
return;
|
|
|
|
|
|
|
|
// Align each byte array to pointer width.
|
|
|
|
unsigned PointerSize = M.getDataLayout().getPointerSize();
|
|
|
|
B.Before.Bytes.resize(alignTo(B.Before.Bytes.size(), PointerSize));
|
|
|
|
B.After.Bytes.resize(alignTo(B.After.Bytes.size(), PointerSize));
|
|
|
|
|
|
|
|
// Before was stored in reverse order; flip it now.
|
|
|
|
for (size_t I = 0, Size = B.Before.Bytes.size(); I != Size / 2; ++I)
|
|
|
|
std::swap(B.Before.Bytes[I], B.Before.Bytes[Size - 1 - I]);
|
|
|
|
|
|
|
|
// Build an anonymous global containing the before bytes, followed by the
|
|
|
|
// original initializer, followed by the after bytes.
|
|
|
|
auto NewInit = ConstantStruct::getAnon(
|
|
|
|
{ConstantDataArray::get(M.getContext(), B.Before.Bytes),
|
|
|
|
B.GV->getInitializer(),
|
|
|
|
ConstantDataArray::get(M.getContext(), B.After.Bytes)});
|
|
|
|
auto NewGV =
|
|
|
|
new GlobalVariable(M, NewInit->getType(), B.GV->isConstant(),
|
|
|
|
GlobalVariable::PrivateLinkage, NewInit, "", B.GV);
|
|
|
|
NewGV->setSection(B.GV->getSection());
|
|
|
|
NewGV->setComdat(B.GV->getComdat());
|
|
|
|
|
|
|
|
// Build an alias named after the original global, pointing at the second
|
|
|
|
// element (the original initializer).
|
|
|
|
auto Alias = GlobalAlias::create(
|
|
|
|
B.GV->getInitializer()->getType(), 0, B.GV->getLinkage(), "",
|
|
|
|
ConstantExpr::getGetElementPtr(
|
|
|
|
NewInit->getType(), NewGV,
|
|
|
|
ArrayRef<Constant *>{ConstantInt::get(Int32Ty, 0),
|
|
|
|
ConstantInt::get(Int32Ty, 1)}),
|
|
|
|
&M);
|
|
|
|
Alias->setVisibility(B.GV->getVisibility());
|
|
|
|
Alias->takeName(B.GV);
|
|
|
|
|
|
|
|
B.GV->replaceAllUsesWith(Alias);
|
|
|
|
B.GV->eraseFromParent();
|
|
|
|
}
|
|
|
|
|
|
|
|
bool DevirtModule::run() {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
Function *TypeTestFunc =
|
|
|
|
M.getFunction(Intrinsic::getName(Intrinsic::type_test));
|
|
|
|
if (!TypeTestFunc || TypeTestFunc->use_empty())
|
2016-02-10 06:50:34 +08:00
|
|
|
return false;
|
|
|
|
|
|
|
|
Function *AssumeFunc = M.getFunction(Intrinsic::getName(Intrinsic::assume));
|
|
|
|
if (!AssumeFunc || AssumeFunc->use_empty())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// Find all virtual calls via a virtual table pointer %p under an assumption
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
// of the form llvm.assume(llvm.type.test(%p, %md)). This indicates that %p
|
|
|
|
// points to a member of the type identifier %md. Group calls by (type ID,
|
|
|
|
// offset) pair (effectively the identity of the virtual function) and store
|
|
|
|
// to CallSlots.
|
2016-02-10 06:50:34 +08:00
|
|
|
DenseSet<Value *> SeenPtrs;
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
for (auto I = TypeTestFunc->use_begin(), E = TypeTestFunc->use_end();
|
2016-02-10 06:50:34 +08:00
|
|
|
I != E;) {
|
|
|
|
auto CI = dyn_cast<CallInst>(I->getUser());
|
|
|
|
++I;
|
|
|
|
if (!CI)
|
|
|
|
continue;
|
|
|
|
|
2016-05-11 02:07:21 +08:00
|
|
|
// Search for virtual calls based on %p and add them to DevirtCalls.
|
|
|
|
SmallVector<DevirtCallSite, 1> DevirtCalls;
|
2016-02-10 06:50:34 +08:00
|
|
|
SmallVector<CallInst *, 1> Assumes;
|
2016-05-11 02:07:21 +08:00
|
|
|
findDevirtualizableCalls(DevirtCalls, Assumes, CI);
|
2016-02-10 06:50:34 +08:00
|
|
|
|
2016-05-11 02:07:21 +08:00
|
|
|
// If we found any, add them to CallSlots. Only do this if we haven't seen
|
|
|
|
// the vtable pointer before, as it may have been CSE'd with pointers from
|
|
|
|
// other call sites, and we don't want to process call sites multiple times.
|
2016-02-10 06:50:34 +08:00
|
|
|
if (!Assumes.empty()) {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
Metadata *TypeId =
|
2016-02-10 06:50:34 +08:00
|
|
|
cast<MetadataAsValue>(CI->getArgOperand(1))->getMetadata();
|
|
|
|
Value *Ptr = CI->getArgOperand(0)->stripPointerCasts();
|
2016-05-11 02:07:21 +08:00
|
|
|
if (SeenPtrs.insert(Ptr).second) {
|
|
|
|
for (DevirtCallSite Call : DevirtCalls) {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
CallSlots[{TypeId, Call.Offset}].push_back(
|
2016-05-11 02:07:21 +08:00
|
|
|
{CI->getArgOperand(0), Call.CS});
|
|
|
|
}
|
|
|
|
}
|
2016-02-10 06:50:34 +08:00
|
|
|
}
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
// We no longer need the assumes or the type test.
|
2016-02-10 06:50:34 +08:00
|
|
|
for (auto Assume : Assumes)
|
|
|
|
Assume->eraseFromParent();
|
|
|
|
// We can't use RecursivelyDeleteTriviallyDeadInstructions here because we
|
|
|
|
// may use the vtable argument later.
|
|
|
|
if (CI->use_empty())
|
|
|
|
CI->eraseFromParent();
|
|
|
|
}
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
// Rebuild type metadata into a map for easy lookup.
|
2016-02-10 06:50:34 +08:00
|
|
|
std::vector<VTableBits> Bits;
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
DenseMap<Metadata *, std::set<TypeMemberInfo>> TypeIdMap;
|
|
|
|
buildTypeIdentifierMap(Bits, TypeIdMap);
|
|
|
|
if (TypeIdMap.empty())
|
2016-02-10 06:50:34 +08:00
|
|
|
return true;
|
|
|
|
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
// For each (type, offset) pair:
|
2016-02-10 06:50:34 +08:00
|
|
|
bool DidVirtualConstProp = false;
|
|
|
|
for (auto &S : CallSlots) {
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
// Search each of the members of the type identifier for the virtual
|
|
|
|
// function implementation at offset S.first.ByteOffset, and add to
|
|
|
|
// TargetsForSlot.
|
2016-02-10 06:50:34 +08:00
|
|
|
std::vector<VirtualCallTarget> TargetsForSlot;
|
IR: New representation for CFI and virtual call optimization pass metadata.
The bitset metadata currently used in LLVM has a few problems:
1. It has the wrong name. The name "bitset" refers to an implementation
detail of one use of the metadata (i.e. its original use case, CFI).
This makes it harder to understand, as the name makes no sense in the
context of virtual call optimization.
2. It is represented using a global named metadata node, rather than
being directly associated with a global. This makes it harder to
manipulate the metadata when rebuilding global variables, summarise it
as part of ThinLTO and drop unused metadata when associated globals are
dropped. For this reason, CFI does not currently work correctly when
both CFI and vcall opt are enabled, as vcall opt needs to rebuild vtable
globals, and fails to associate metadata with the rebuilt globals. As I
understand it, the same problem could also affect ASan, which rebuilds
globals with a red zone.
This patch solves both of those problems in the following way:
1. Rename the metadata to "type metadata". This new name reflects how
the metadata is currently being used (i.e. to represent type information
for CFI and vtable opt). The new name is reflected in the name for the
associated intrinsic (llvm.type.test) and pass (LowerTypeTests).
2. Attach metadata directly to the globals that it pertains to, rather
than using the "llvm.bitsets" global metadata node as we are doing now.
This is done using the newly introduced capability to attach
metadata to global variables (r271348 and r271358).
See also: http://lists.llvm.org/pipermail/llvm-dev/2016-June/100462.html
Differential Revision: http://reviews.llvm.org/D21053
llvm-svn: 273729
2016-06-25 05:21:32 +08:00
|
|
|
if (!tryFindVirtualCallTargets(TargetsForSlot, TypeIdMap[S.first.TypeID],
|
2016-02-10 06:50:34 +08:00
|
|
|
S.first.ByteOffset))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (trySingleImplDevirt(TargetsForSlot, S.second))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
DidVirtualConstProp |= tryVirtualConstProp(TargetsForSlot, S.second);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Rebuild each global we touched as part of virtual constant propagation to
|
|
|
|
// include the before and after bytes.
|
|
|
|
if (DidVirtualConstProp)
|
|
|
|
for (VTableBits &B : Bits)
|
|
|
|
rebuildGlobal(B);
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|