Use special DAG-to-DAG preprocessing to allow mem-mem instructions to be selected.

Yay for ASCII graphics!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@84808 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Anton Korobeynikov 2009-10-22 00:16:00 +00:00
parent 443b2153e3
commit f32df4ce3e
3 changed files with 149 additions and 21 deletions

View File

@ -65,6 +65,9 @@ namespace {
return "MSP430 DAG->DAG Pattern Instruction Selection";
}
bool IsLegalAndProfitableToFold(SDNode *N, SDNode *U,
SDNode *Root) const;
virtual bool
SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
std::vector<SDValue> &OutOps);
@ -73,6 +76,7 @@ namespace {
#include "MSP430GenDAGISel.inc"
private:
DenseMap<SDNode*, SDNode*> RMWStores;
void PreprocessForRMW();
SDNode *Select(SDValue Op);
bool SelectAddr(SDValue Op, SDValue Addr, SDValue &Base, SDValue &Disp);
@ -139,7 +143,6 @@ bool MSP430DAGToDAGISel::SelectAddr(SDValue Op, SDValue Addr,
return true;
}
bool MSP430DAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
std::vector<SDValue> &OutOps) {
@ -157,6 +160,42 @@ SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
return false;
}
bool MSP430DAGToDAGISel::IsLegalAndProfitableToFold(SDNode *N, SDNode *U,
SDNode *Root) const {
if (OptLevel == CodeGenOpt::None) return false;
/// RMW preprocessing creates the following code:
/// [Load1]
/// ^ ^
/// | \
/// | \
/// [Load2] |
/// ^ ^ |
/// | | |
/// | \-|
/// | |
/// | [Op]
/// | ^
/// | |
/// \ /
/// \ /
/// [Store]
///
/// The path Store => Load2 => Load1 is via chain. Note that in general it is
/// not allowed to fold Load1 into Op (and Store) since it will creates a
/// cycle. However, this is perfectly legal for the loads moved below the
/// TokenFactor by PreprocessForRMW. Query the map Store => Load1 (created
/// during preprocessing) to determine whether it's legal to introduce such
/// "cycle" for a moment.
DenseMap<SDNode*, SDNode*>::iterator I = RMWStores.find(Root);
if (I != RMWStores.end() && I->second == N)
return true;
// Proceed to 'generic' cycle finder code
return SelectionDAGISel::IsLegalAndProfitableToFold(N, U, Root);
}
/// MoveBelowTokenFactor - Replace TokenFactor operand with load's chain operand
/// and move load below the TokenFactor. Replace store's chain operand with
/// load's chain result.
@ -176,11 +215,43 @@ static void MoveBelowTokenFactor(SelectionDAG *CurDAG, SDValue Load,
Store.getOperand(2), Store.getOperand(3));
}
/// isRMWLoad - Return true if N is a load that's part of RMW sub-DAG.
/// The chain produced by the load must only be used by the store's chain
/// operand, otherwise this may produce a cycle in the DAG.
static bool isRMWLoad(SDValue N, SDValue Chain, SDValue Address,
SDValue &Load) {
/// MoveBelowTokenFactor2 - Replace TokenFactor operand with load's chain operand
/// and move load below the TokenFactor. Replace store's chain operand with
/// load's chain result. This a version which sinks two loads below token factor.
/// Look into PreprocessForRMW comments for explanation of transform.
static void MoveBelowTokenFactor2(SelectionDAG *CurDAG,
SDValue Load1, SDValue Load2,
SDValue Store, SDValue TF) {
SmallVector<SDValue, 4> Ops;
for (unsigned i = 0, e = TF.getNode()->getNumOperands(); i != e; ++i) {
SDNode* N = TF.getOperand(i).getNode();
if (Load2.getNode() == N)
Ops.push_back(Load2.getOperand(0));
else if (Load1.getNode() != N)
Ops.push_back(TF.getOperand(i));
}
SDValue NewTF = SDValue(CurDAG->MorphNodeTo(TF.getNode(),
TF.getOpcode(),
TF.getNode()->getVTList(),
&Ops[0], Ops.size()), TF.getResNo());
SDValue NewLoad2 = CurDAG->UpdateNodeOperands(Load2, NewTF,
Load2.getOperand(1),
Load2.getOperand(2));
SDValue NewLoad1 = CurDAG->UpdateNodeOperands(Load1, NewLoad2.getValue(1),
Load1.getOperand(1),
Load1.getOperand(2));
CurDAG->UpdateNodeOperands(Store,
NewLoad1.getValue(1),
Store.getOperand(1),
Store.getOperand(2), Store.getOperand(3));
}
/// isAllowedToSink - return true if N a load which can be moved below token
/// factor. Basically, the load should be non-volatile and has single use.
static bool isLoadAllowedToSink(SDValue N, SDValue Chain) {
if (N.getOpcode() == ISD::BIT_CONVERT)
N = N.getOperand(0);
@ -194,10 +265,19 @@ static bool isRMWLoad(SDValue N, SDValue Chain, SDValue Address,
if (ExtType != ISD::NON_EXTLOAD && ExtType != ISD::EXTLOAD)
return false;
if (N.hasOneUse() &&
LD->hasNUsesOfValue(1, 1) &&
N.getOperand(1) == Address &&
LD->isOperandOf(Chain.getNode())) {
return (N.hasOneUse() &&
LD->hasNUsesOfValue(1, 1) &&
LD->isOperandOf(Chain.getNode()));
}
/// isRMWLoad - Return true if N is a load that's part of RMW sub-DAG.
/// The chain produced by the load must only be used by the store's chain
/// operand, otherwise this may produce a cycle in the DAG.
static bool isRMWLoad(SDValue N, SDValue Chain, SDValue Address,
SDValue &Load) {
if (isLoadAllowedToSink(N, Chain) &&
N.getOperand(1) == Address) {
Load = N;
return true;
}
@ -244,6 +324,34 @@ static bool isRMWLoad(SDValue N, SDValue Chain, SDValue Address,
/// \ /
/// \ /
/// [Store]
///
/// We also recognize the case where second operand of Op is load as well and
/// move it below token factor as well creating DAG as follows:
///
/// [Load chain]
/// ^
/// |
/// [TokenFactor]
/// ^
/// |
/// [Load1]
/// ^ ^
/// | \
/// | \
/// [Load2] |
/// ^ ^ |
/// | | |
/// | \-|
/// | |
/// | [Op]
/// | ^
/// | |
/// \ /
/// \ /
/// [Store]
///
/// This allows selection of mem-mem instructions. Yay!
void MSP430DAGToDAGISel::PreprocessForRMW() {
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
E = CurDAG->allnodes_end(); I != E; ++I) {
@ -261,8 +369,8 @@ void MSP430DAGToDAGISel::PreprocessForRMW() {
!N1.hasOneUse())
continue;
bool RModW = false;
SDValue Load;
unsigned RModW = 0;
SDValue Load1, Load2;
unsigned Opcode = N1.getNode()->getOpcode();
switch (Opcode) {
case ISD::ADD:
@ -273,23 +381,44 @@ void MSP430DAGToDAGISel::PreprocessForRMW() {
case ISD::ADDE: {
SDValue N10 = N1.getOperand(0);
SDValue N11 = N1.getOperand(1);
RModW = isRMWLoad(N10, Chain, N2, Load);
if (!RModW)
RModW = isRMWLoad(N11, Chain, N2, Load);
if (isRMWLoad(N10, Chain, N2, Load1)) {
if (isLoadAllowedToSink(N11, Chain)) {
Load2 = N11;
RModW = 2;
} else
RModW = 1;
} else if (isRMWLoad(N11, Chain, N2, Load1)) {
if (isLoadAllowedToSink(N10, Chain)) {
Load2 = N10;
RModW = 2;
} else
RModW = 1;
}
break;
}
case ISD::SUB:
case ISD::SUBC:
case ISD::SUBE: {
SDValue N10 = N1.getOperand(0);
RModW = isRMWLoad(N10, Chain, N2, Load);
SDValue N11 = N1.getOperand(1);
if (isRMWLoad(N10, Chain, N2, Load1)) {
if (isLoadAllowedToSink(N11, Chain)) {
Load2 = N11;
RModW = 2;
} else
RModW = 1;
}
break;
}
}
if (RModW) {
MoveBelowTokenFactor(CurDAG, Load, SDValue(I, 0), Chain);
++NumLoadMoved;
NumLoadMoved += RModW;
if (RModW == 1)
MoveBelowTokenFactor(CurDAG, Load1, SDValue(I, 0), Chain);
else if (RModW == 2) {
MoveBelowTokenFactor2(CurDAG, Load1, Load2, SDValue(I, 0), Chain);
SDNode* Store = I;
RMWStores[Store] = Load2.getNode();
}
}
}
@ -318,6 +447,7 @@ void MSP430DAGToDAGISel::InstructionSelect() {
DEBUG(errs() << "===== Instruction selection ends:\n");
CurDAG->RemoveDeadNodes();
RMWStores.clear();
}
SDNode *MSP430DAGToDAGISel::Select(SDValue Op) {

View File

@ -1,5 +1,4 @@
; RUN: llc -march=msp430 < %s | FileCheck %s
; XFAIL: *
target datalayout = "e-p:16:8:8-i8:8:8-i16:8:8-i32:8:8"
target triple = "msp430-generic-generic"
@foo = common global i16 0, align 2

View File

@ -1,5 +1,4 @@
; RUN: llc -march=msp430 < %s | FileCheck %s
; XFAIL: *
target datalayout = "e-p:16:8:8-i8:8:8-i16:8:8-i32:8:8"
target triple = "msp430-generic-generic"