Files
clang-p2996/llvm/lib/Target/BPF/BPFISelDAGToDAG.cpp
Jiong Wang ec51851026 bpf: fix wrong truncation elimination when there is back-edge/loop
Currently, BPF backend is doing truncation elimination. If one truncation
is performed on a value defined by narrow loads, then it could be redundant
given BPF loads zero extend the destination register implicitly.

When the definition of the truncated value is a merging value (PHI node)
that could come from different code paths, then checks need to be done on
all possible code paths.

Above described optimization was introduced as r306685, however it doesn't
work when there is back-edge, for example when loop is used inside BPF
code.

For example for the following code, a zero-extended value should be stored
into b[i], but the "and reg, 0xffff" is wrongly eliminated which then
generates corrupted data.

void cal1(unsigned short *a, unsigned long *b, unsigned int k)
{
  unsigned short e;

  e = *a;
  for (unsigned int i = 0; i < k; i++) {
    b[i] = e;
    e = ~e;
  }
}

The reason is r306685 was trying to do the PHI node checks inside isel
DAG2DAG phase, and the checks are done on MachineInstr. This is actually
wrong, because MachineInstr is being built during isel phase and the
associated information is not completed yet. A quick search shows none
target other than BPF is access MachineInstr info during isel phase.

For an PHI node, when you reached it during isel phase, it may have all
predecessors linked, but not successors. It seems successors are linked to
PHI node only when doing SelectionDAGISel::FinishBasicBlock and this
happens later than PreprocessISelDAG hook.

Previously, BPF program doesn't allow loop, there is probably the reason
why this bug was not exposed.

This patch therefore fixes the bug by the following approach:
 - The existing truncation elimination code and the associated
   "load_to_vreg_" records are removed.
 - Instead, implement truncation elimination using MachineSSA pass, this
   is where all information are built, and keep the pass together with other
   similar peephole optimizations inside BPFMIPeephole.cpp. Redundant move
   elimination logic is updated accordingly.
 - Unit testcase included + no compilation errors for kernel BPF selftest.

Patch Review
===
Patch was sent to and reviewed by BPF community at:

  https://lore.kernel.org/bpf

Reported-by: David Beckett <david.beckett@netronome.com>
Reviewed-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
llvm-svn: 375007
2019-10-16 15:27:59 +00:00

503 lines
17 KiB
C++

//===-- BPFISelDAGToDAG.cpp - A dag to dag inst selector for BPF ----------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file defines a DAG pattern matching instruction selector for BPF,
// converting from a legalized dag to a BPF dag.
//
//===----------------------------------------------------------------------===//
#include "BPF.h"
#include "BPFRegisterInfo.h"
#include "BPFSubtarget.h"
#include "BPFTargetMachine.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/Endian.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"
using namespace llvm;
#define DEBUG_TYPE "bpf-isel"
// Instruction Selector Implementation
namespace {
class BPFDAGToDAGISel : public SelectionDAGISel {
/// Subtarget - Keep a pointer to the BPFSubtarget around so that we can
/// make the right decision when generating code for different subtargets.
const BPFSubtarget *Subtarget;
public:
explicit BPFDAGToDAGISel(BPFTargetMachine &TM)
: SelectionDAGISel(TM), Subtarget(nullptr) {}
StringRef getPassName() const override {
return "BPF DAG->DAG Pattern Instruction Selection";
}
bool runOnMachineFunction(MachineFunction &MF) override {
// Reset the subtarget each time through.
Subtarget = &MF.getSubtarget<BPFSubtarget>();
return SelectionDAGISel::runOnMachineFunction(MF);
}
void PreprocessISelDAG() override;
bool SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintCode,
std::vector<SDValue> &OutOps) override;
private:
// Include the pieces autogenerated from the target description.
#include "BPFGenDAGISel.inc"
void Select(SDNode *N) override;
// Complex Pattern for address selection.
bool SelectAddr(SDValue Addr, SDValue &Base, SDValue &Offset);
bool SelectFIAddr(SDValue Addr, SDValue &Base, SDValue &Offset);
// Node preprocessing cases
void PreprocessLoad(SDNode *Node, SelectionDAG::allnodes_iterator &I);
void PreprocessCopyToReg(SDNode *Node);
void PreprocessTrunc(SDNode *Node, SelectionDAG::allnodes_iterator &I);
// Find constants from a constant structure
typedef std::vector<unsigned char> val_vec_type;
bool fillGenericConstant(const DataLayout &DL, const Constant *CV,
val_vec_type &Vals, uint64_t Offset);
bool fillConstantDataArray(const DataLayout &DL, const ConstantDataArray *CDA,
val_vec_type &Vals, int Offset);
bool fillConstantArray(const DataLayout &DL, const ConstantArray *CA,
val_vec_type &Vals, int Offset);
bool fillConstantStruct(const DataLayout &DL, const ConstantStruct *CS,
val_vec_type &Vals, int Offset);
bool getConstantFieldValue(const GlobalAddressSDNode *Node, uint64_t Offset,
uint64_t Size, unsigned char *ByteSeq);
// Mapping from ConstantStruct global value to corresponding byte-list values
std::map<const void *, val_vec_type> cs_vals_;
};
} // namespace
// ComplexPattern used on BPF Load/Store instructions
bool BPFDAGToDAGISel::SelectAddr(SDValue Addr, SDValue &Base, SDValue &Offset) {
// if Address is FI, get the TargetFrameIndex.
SDLoc DL(Addr);
if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i64);
Offset = CurDAG->getTargetConstant(0, DL, MVT::i64);
return true;
}
if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
Addr.getOpcode() == ISD::TargetGlobalAddress)
return false;
// Addresses of the form Addr+const or Addr|const
if (CurDAG->isBaseWithConstantOffset(Addr)) {
ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1));
if (isInt<16>(CN->getSExtValue())) {
// If the first operand is a FI, get the TargetFI Node
if (FrameIndexSDNode *FIN =
dyn_cast<FrameIndexSDNode>(Addr.getOperand(0)))
Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i64);
else
Base = Addr.getOperand(0);
Offset = CurDAG->getTargetConstant(CN->getSExtValue(), DL, MVT::i64);
return true;
}
}
Base = Addr;
Offset = CurDAG->getTargetConstant(0, DL, MVT::i64);
return true;
}
// ComplexPattern used on BPF FI instruction
bool BPFDAGToDAGISel::SelectFIAddr(SDValue Addr, SDValue &Base,
SDValue &Offset) {
SDLoc DL(Addr);
if (!CurDAG->isBaseWithConstantOffset(Addr))
return false;
// Addresses of the form Addr+const or Addr|const
ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1));
if (isInt<16>(CN->getSExtValue())) {
// If the first operand is a FI, get the TargetFI Node
if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr.getOperand(0)))
Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i64);
else
return false;
Offset = CurDAG->getTargetConstant(CN->getSExtValue(), DL, MVT::i64);
return true;
}
return false;
}
bool BPFDAGToDAGISel::SelectInlineAsmMemoryOperand(
const SDValue &Op, unsigned ConstraintCode, std::vector<SDValue> &OutOps) {
SDValue Op0, Op1;
switch (ConstraintCode) {
default:
return true;
case InlineAsm::Constraint_m: // memory
if (!SelectAddr(Op, Op0, Op1))
return true;
break;
}
SDLoc DL(Op);
SDValue AluOp = CurDAG->getTargetConstant(ISD::ADD, DL, MVT::i32);;
OutOps.push_back(Op0);
OutOps.push_back(Op1);
OutOps.push_back(AluOp);
return false;
}
void BPFDAGToDAGISel::Select(SDNode *Node) {
unsigned Opcode = Node->getOpcode();
// If we have a custom node, we already have selected!
if (Node->isMachineOpcode()) {
LLVM_DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << '\n');
return;
}
// tablegen selection should be handled here.
switch (Opcode) {
default:
break;
case ISD::SDIV: {
DebugLoc Empty;
const DebugLoc &DL = Node->getDebugLoc();
if (DL != Empty)
errs() << "Error at line " << DL.getLine() << ": ";
else
errs() << "Error: ";
errs() << "Unsupport signed division for DAG: ";
Node->print(errs(), CurDAG);
errs() << "Please convert to unsigned div/mod.\n";
break;
}
case ISD::INTRINSIC_W_CHAIN: {
unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
switch (IntNo) {
case Intrinsic::bpf_load_byte:
case Intrinsic::bpf_load_half:
case Intrinsic::bpf_load_word: {
SDLoc DL(Node);
SDValue Chain = Node->getOperand(0);
SDValue N1 = Node->getOperand(1);
SDValue Skb = Node->getOperand(2);
SDValue N3 = Node->getOperand(3);
SDValue R6Reg = CurDAG->getRegister(BPF::R6, MVT::i64);
Chain = CurDAG->getCopyToReg(Chain, DL, R6Reg, Skb, SDValue());
Node = CurDAG->UpdateNodeOperands(Node, Chain, N1, R6Reg, N3);
break;
}
}
break;
}
case ISD::FrameIndex: {
int FI = cast<FrameIndexSDNode>(Node)->getIndex();
EVT VT = Node->getValueType(0);
SDValue TFI = CurDAG->getTargetFrameIndex(FI, VT);
unsigned Opc = BPF::MOV_rr;
if (Node->hasOneUse()) {
CurDAG->SelectNodeTo(Node, Opc, VT, TFI);
return;
}
ReplaceNode(Node, CurDAG->getMachineNode(Opc, SDLoc(Node), VT, TFI));
return;
}
}
// Select the default instruction
SelectCode(Node);
}
void BPFDAGToDAGISel::PreprocessLoad(SDNode *Node,
SelectionDAG::allnodes_iterator &I) {
union {
uint8_t c[8];
uint16_t s;
uint32_t i;
uint64_t d;
} new_val; // hold up the constant values replacing loads.
bool to_replace = false;
SDLoc DL(Node);
const LoadSDNode *LD = cast<LoadSDNode>(Node);
uint64_t size = LD->getMemOperand()->getSize();
if (!size || size > 8 || (size & (size - 1)))
return;
SDNode *LDAddrNode = LD->getOperand(1).getNode();
// Match LDAddr against either global_addr or (global_addr + offset)
unsigned opcode = LDAddrNode->getOpcode();
if (opcode == ISD::ADD) {
SDValue OP1 = LDAddrNode->getOperand(0);
SDValue OP2 = LDAddrNode->getOperand(1);
// We want to find the pattern global_addr + offset
SDNode *OP1N = OP1.getNode();
if (OP1N->getOpcode() <= ISD::BUILTIN_OP_END || OP1N->getNumOperands() == 0)
return;
LLVM_DEBUG(dbgs() << "Check candidate load: "; LD->dump(); dbgs() << '\n');
const GlobalAddressSDNode *GADN =
dyn_cast<GlobalAddressSDNode>(OP1N->getOperand(0).getNode());
const ConstantSDNode *CDN = dyn_cast<ConstantSDNode>(OP2.getNode());
if (GADN && CDN)
to_replace =
getConstantFieldValue(GADN, CDN->getZExtValue(), size, new_val.c);
} else if (LDAddrNode->getOpcode() > ISD::BUILTIN_OP_END &&
LDAddrNode->getNumOperands() > 0) {
LLVM_DEBUG(dbgs() << "Check candidate load: "; LD->dump(); dbgs() << '\n');
SDValue OP1 = LDAddrNode->getOperand(0);
if (const GlobalAddressSDNode *GADN =
dyn_cast<GlobalAddressSDNode>(OP1.getNode()))
to_replace = getConstantFieldValue(GADN, 0, size, new_val.c);
}
if (!to_replace)
return;
// replacing the old with a new value
uint64_t val;
if (size == 1)
val = new_val.c[0];
else if (size == 2)
val = new_val.s;
else if (size == 4)
val = new_val.i;
else {
val = new_val.d;
}
LLVM_DEBUG(dbgs() << "Replacing load of size " << size << " with constant "
<< val << '\n');
SDValue NVal = CurDAG->getConstant(val, DL, MVT::i64);
// After replacement, the current node is dead, we need to
// go backward one step to make iterator still work
I--;
SDValue From[] = {SDValue(Node, 0), SDValue(Node, 1)};
SDValue To[] = {NVal, NVal};
CurDAG->ReplaceAllUsesOfValuesWith(From, To, 2);
I++;
// It is safe to delete node now
CurDAG->DeleteNode(Node);
}
void BPFDAGToDAGISel::PreprocessISelDAG() {
// Iterate through all nodes, interested in the following case:
//
// . loads from ConstantStruct or ConstantArray of constructs
// which can be turns into constant itself, with this we can
// avoid reading from read-only section at runtime.
//
// . Removing redundant AND for intrinsic narrow loads.
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
E = CurDAG->allnodes_end();
I != E;) {
SDNode *Node = &*I++;
unsigned Opcode = Node->getOpcode();
if (Opcode == ISD::LOAD)
PreprocessLoad(Node, I);
else if (Opcode == ISD::AND)
PreprocessTrunc(Node, I);
}
}
bool BPFDAGToDAGISel::getConstantFieldValue(const GlobalAddressSDNode *Node,
uint64_t Offset, uint64_t Size,
unsigned char *ByteSeq) {
const GlobalVariable *V = dyn_cast<GlobalVariable>(Node->getGlobal());
if (!V || !V->hasInitializer())
return false;
const Constant *Init = V->getInitializer();
const DataLayout &DL = CurDAG->getDataLayout();
val_vec_type TmpVal;
auto it = cs_vals_.find(static_cast<const void *>(Init));
if (it != cs_vals_.end()) {
TmpVal = it->second;
} else {
uint64_t total_size = 0;
if (const ConstantStruct *CS = dyn_cast<ConstantStruct>(Init))
total_size =
DL.getStructLayout(cast<StructType>(CS->getType()))->getSizeInBytes();
else if (const ConstantArray *CA = dyn_cast<ConstantArray>(Init))
total_size = DL.getTypeAllocSize(CA->getType()->getElementType()) *
CA->getNumOperands();
else
return false;
val_vec_type Vals(total_size, 0);
if (fillGenericConstant(DL, Init, Vals, 0) == false)
return false;
cs_vals_[static_cast<const void *>(Init)] = Vals;
TmpVal = std::move(Vals);
}
// test whether host endianness matches target
union {
uint8_t c[2];
uint16_t s;
} test_buf;
uint16_t test_val = 0x2345;
if (DL.isLittleEndian())
support::endian::write16le(test_buf.c, test_val);
else
support::endian::write16be(test_buf.c, test_val);
bool endian_match = test_buf.s == test_val;
for (uint64_t i = Offset, j = 0; i < Offset + Size; i++, j++)
ByteSeq[j] = endian_match ? TmpVal[i] : TmpVal[Offset + Size - 1 - j];
return true;
}
bool BPFDAGToDAGISel::fillGenericConstant(const DataLayout &DL,
const Constant *CV,
val_vec_type &Vals, uint64_t Offset) {
uint64_t Size = DL.getTypeAllocSize(CV->getType());
if (isa<ConstantAggregateZero>(CV) || isa<UndefValue>(CV))
return true; // already done
if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV)) {
uint64_t val = CI->getZExtValue();
LLVM_DEBUG(dbgs() << "Byte array at offset " << Offset << " with value "
<< val << '\n');
if (Size > 8 || (Size & (Size - 1)))
return false;
// Store based on target endian
for (uint64_t i = 0; i < Size; ++i) {
Vals[Offset + i] = DL.isLittleEndian()
? ((val >> (i * 8)) & 0xFF)
: ((val >> ((Size - i - 1) * 8)) & 0xFF);
}
return true;
}
if (const ConstantDataArray *CDA = dyn_cast<ConstantDataArray>(CV))
return fillConstantDataArray(DL, CDA, Vals, Offset);
if (const ConstantArray *CA = dyn_cast<ConstantArray>(CV))
return fillConstantArray(DL, CA, Vals, Offset);
if (const ConstantStruct *CVS = dyn_cast<ConstantStruct>(CV))
return fillConstantStruct(DL, CVS, Vals, Offset);
return false;
}
bool BPFDAGToDAGISel::fillConstantDataArray(const DataLayout &DL,
const ConstantDataArray *CDA,
val_vec_type &Vals, int Offset) {
for (unsigned i = 0, e = CDA->getNumElements(); i != e; ++i) {
if (fillGenericConstant(DL, CDA->getElementAsConstant(i), Vals, Offset) ==
false)
return false;
Offset += DL.getTypeAllocSize(CDA->getElementAsConstant(i)->getType());
}
return true;
}
bool BPFDAGToDAGISel::fillConstantArray(const DataLayout &DL,
const ConstantArray *CA,
val_vec_type &Vals, int Offset) {
for (unsigned i = 0, e = CA->getNumOperands(); i != e; ++i) {
if (fillGenericConstant(DL, CA->getOperand(i), Vals, Offset) == false)
return false;
Offset += DL.getTypeAllocSize(CA->getOperand(i)->getType());
}
return true;
}
bool BPFDAGToDAGISel::fillConstantStruct(const DataLayout &DL,
const ConstantStruct *CS,
val_vec_type &Vals, int Offset) {
const StructLayout *Layout = DL.getStructLayout(CS->getType());
for (unsigned i = 0, e = CS->getNumOperands(); i != e; ++i) {
const Constant *Field = CS->getOperand(i);
uint64_t SizeSoFar = Layout->getElementOffset(i);
if (fillGenericConstant(DL, Field, Vals, Offset + SizeSoFar) == false)
return false;
}
return true;
}
void BPFDAGToDAGISel::PreprocessTrunc(SDNode *Node,
SelectionDAG::allnodes_iterator &I) {
ConstantSDNode *MaskN = dyn_cast<ConstantSDNode>(Node->getOperand(1));
if (!MaskN)
return;
// The Reg operand should be a virtual register, which is defined
// outside the current basic block. DAG combiner has done a pretty
// good job in removing truncating inside a single basic block except
// when the Reg operand comes from bpf_load_[byte | half | word] for
// which the generic optimizer doesn't understand their results are
// zero extended.
SDValue BaseV = Node->getOperand(0);
if (BaseV.getOpcode() != ISD::INTRINSIC_W_CHAIN)
return;
unsigned IntNo = cast<ConstantSDNode>(BaseV->getOperand(1))->getZExtValue();
uint64_t MaskV = MaskN->getZExtValue();
if (!((IntNo == Intrinsic::bpf_load_byte && MaskV == 0xFF) ||
(IntNo == Intrinsic::bpf_load_half && MaskV == 0xFFFF) ||
(IntNo == Intrinsic::bpf_load_word && MaskV == 0xFFFFFFFF)))
return;
LLVM_DEBUG(dbgs() << "Remove the redundant AND operation in: ";
Node->dump(); dbgs() << '\n');
I--;
CurDAG->ReplaceAllUsesWith(SDValue(Node, 0), BaseV);
I++;
CurDAG->DeleteNode(Node);
return;
}
FunctionPass *llvm::createBPFISelDag(BPFTargetMachine &TM) {
return new BPFDAGToDAGISel(TM);
}