Enhance ComputeMaskedBits to know that aligned frameindexes
have their low bits set to zero. This allows us to optimize
out explicit stack alignment code like in stack-align.ll:test4 when
it is redundant.
Doing this causes the code generator to start turning FI+cst into
FI|cst all over the place, which is general goodness (that is the
canonical form) except that various pieces of the code generator
don't handle OR aggressively. Fix this by introducing a new
SelectionDAG::isBaseWithConstantOffset predicate, and using it
in places that are looking for ADD(X,CST). The ARM backend in
particular was missing a lot of addressing mode folding opportunities
around OR.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@125470 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 5dd8434..a506cff 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -427,16 +427,19 @@
// Match simple R + imm12 operands.
// Base only.
- if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
+ if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
+ !CurDAG->isBaseWithConstantOffset(N)) {
if (N.getOpcode() == ISD::FrameIndex) {
- // Match frame index...
+ // Match frame index.
int FI = cast<FrameIndexSDNode>(N)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
- } else if (N.getOpcode() == ARMISD::Wrapper &&
- !(Subtarget->useMovt() &&
- N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
+ }
+
+ if (N.getOpcode() == ARMISD::Wrapper &&
+ !(Subtarget->useMovt() &&
+ N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
Base = N.getOperand(0);
} else
Base = N;
@@ -494,11 +497,13 @@
}
}
- if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB)
+ if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
+ // ISD::OR that is equivalent to an ISD::ADD.
+ !CurDAG->isBaseWithConstantOffset(N))
return false;
// Leave simple R +/- imm12 operands for LDRi12
- if (N.getOpcode() == ISD::ADD) {
+ if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::OR) {
int RHSC;
if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
-0x1000+1, 0x1000, RHSC)) // 12 bits.
@@ -510,7 +515,7 @@
return false;
// Otherwise this is R +/- [possibly shifted] R.
- ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
+ ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::SUB ? ARM_AM::sub:ARM_AM::add;
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
unsigned ShAmt = 0;
@@ -535,7 +540,7 @@
}
// Try matching (R shl C) + (R).
- if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift &&
+ if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift &&
!(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
if (ShOpcVal != ARM_AM::no_shift) {
@@ -597,7 +602,9 @@
}
}
- if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
+ if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
+ // ISD::OR that is equivalent to an ADD.
+ !CurDAG->isBaseWithConstantOffset(N)) {
Base = N;
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
@@ -615,7 +622,7 @@
}
// Match simple R +/- imm12 operands.
- if (N.getOpcode() == ISD::ADD) {
+ if (N.getOpcode() != ISD::SUB) {
int RHSC;
if (isScaledConstantInRange(N.getOperand(1), /*Scale=*/1,
-0x1000+1, 0x1000, RHSC)) { // 12 bits.
@@ -649,7 +656,7 @@
}
// Otherwise this is R +/- [possibly shifted] R.
- ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
+ ARM_AM::AddrOpc AddSub = N.getOpcode() != ISD::SUB ? ARM_AM::add:ARM_AM::sub;
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
unsigned ShAmt = 0;
@@ -674,7 +681,7 @@
}
// Try matching (R shl C) + (R).
- if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift &&
+ if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift &&
!(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
if (ShOpcVal != ARM_AM::no_shift) {
@@ -756,7 +763,7 @@
return true;
}
- if (N.getOpcode() != ISD::ADD) {
+ if (!CurDAG->isBaseWithConstantOffset(N)) {
Base = N;
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
@@ -781,7 +788,7 @@
ARM_AM::AddrOpc AddSub = ARM_AM::add;
if (RHSC < 0) {
AddSub = ARM_AM::sub;
- RHSC = - RHSC;
+ RHSC = -RHSC;
}
Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
return true;
@@ -815,7 +822,7 @@
bool ARMDAGToDAGISel::SelectAddrMode5(SDValue N,
SDValue &Base, SDValue &Offset) {
- if (N.getOpcode() != ISD::ADD) {
+ if (!CurDAG->isBaseWithConstantOffset(N)) {
Base = N;
if (N.getOpcode() == ISD::FrameIndex) {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
@@ -843,7 +850,7 @@
ARM_AM::AddrOpc AddSub = ARM_AM::add;
if (RHSC < 0) {
AddSub = ARM_AM::sub;
- RHSC = - RHSC;
+ RHSC = -RHSC;
}
Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
MVT::i32);
@@ -899,8 +906,7 @@
bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue N,
SDValue &Base, SDValue &Offset){
- // FIXME dl should come from the parent load or store, not the address
- if (N.getOpcode() != ISD::ADD) {
+ if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N)) {
ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
if (!NC || !NC->isNullValue())
return false;
@@ -927,7 +933,7 @@
return false; // We want to select tLDRpci instead.
}
- if (N.getOpcode() != ISD::ADD)
+ if (!CurDAG->isBaseWithConstantOffset(N))
return false;
// Thumb does not have [sp, r] address mode.
@@ -983,7 +989,7 @@
return false; // We want to select tLDRpci instead.
}
- if (N.getOpcode() != ISD::ADD) {
+ if (!CurDAG->isBaseWithConstantOffset(N)) {
if (N.getOpcode() == ARMISD::Wrapper &&
!(Subtarget->useMovt() &&
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
@@ -1053,7 +1059,7 @@
return true;
}
- if (N.getOpcode() != ISD::ADD)
+ if (!CurDAG->isBaseWithConstantOffset(N))
return false;
RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
@@ -1108,14 +1114,17 @@
// Match simple R + imm12 operands.
// Base only.
- if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
+ if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
+ !CurDAG->isBaseWithConstantOffset(N)) {
if (N.getOpcode() == ISD::FrameIndex) {
- // Match frame index...
+ // Match frame index.
int FI = cast<FrameIndexSDNode>(N)->getIndex();
Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
OffImm = CurDAG->getTargetConstant(0, MVT::i32);
return true;
- } else if (N.getOpcode() == ARMISD::Wrapper &&
+ }
+
+ if (N.getOpcode() == ARMISD::Wrapper &&
!(Subtarget->useMovt() &&
N.getOperand(0).getOpcode() == ISD::TargetGlobalAddress)) {
Base = N.getOperand(0);
@@ -1156,21 +1165,23 @@
bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N,
SDValue &Base, SDValue &OffImm) {
// Match simple R - imm8 operands.
- if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
- if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
- int RHSC = (int)RHS->getSExtValue();
- if (N.getOpcode() == ISD::SUB)
- RHSC = -RHSC;
+ if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB &&
+ !CurDAG->isBaseWithConstantOffset(N))
+ return false;
+
+ if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
+ int RHSC = (int)RHS->getSExtValue();
+ if (N.getOpcode() == ISD::SUB)
+ RHSC = -RHSC;
- if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
- Base = N.getOperand(0);
- if (Base.getOpcode() == ISD::FrameIndex) {
- int FI = cast<FrameIndexSDNode>(Base)->getIndex();
- Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
- }
- OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
- return true;
+ if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
+ Base = N.getOperand(0);
+ if (Base.getOpcode() == ISD::FrameIndex) {
+ int FI = cast<FrameIndexSDNode>(Base)->getIndex();
+ Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
}
+ OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
+ return true;
}
}
@@ -1198,7 +1209,7 @@
SDValue &Base,
SDValue &OffReg, SDValue &ShImm) {
// (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
- if (N.getOpcode() != ISD::ADD)
+ if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(N))
return false;
// Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp
index 59ddf3b..1845b3c 100644
--- a/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -685,25 +685,6 @@
return false;
}
-/// isLogicallyAddWithConstant - Return true if this node is semantically an
-/// add of a value with a constantint.
-static bool isLogicallyAddWithConstant(SDValue V, SelectionDAG *CurDAG) {
- // Check for (add x, Cst)
- if (V->getOpcode() == ISD::ADD)
- return isa<ConstantSDNode>(V->getOperand(1));
-
- // Check for (or x, Cst), where Cst & x == 0.
- if (V->getOpcode() != ISD::OR ||
- !isa<ConstantSDNode>(V->getOperand(1)))
- return false;
-
- // Handle "X | C" as "X + C" iff X is known to have C bits clear.
- ConstantSDNode *CN = cast<ConstantSDNode>(V->getOperand(1));
-
- // Check to see if the LHS & C is zero.
- return CurDAG->MaskedValueIsZero(V->getOperand(0), CN->getAPIntValue());
-}
-
bool X86DAGToDAGISel::MatchAddressRecursively(SDValue N, X86ISelAddressMode &AM,
unsigned Depth) {
bool is64Bit = Subtarget->is64Bit();
@@ -789,7 +770,7 @@
// Okay, we know that we have a scale by now. However, if the scaled
// value is an add of something and a constant, we can fold the
// constant into the disp field here.
- if (isLogicallyAddWithConstant(ShVal, CurDAG)) {
+ if (CurDAG->isBaseWithConstantOffset(ShVal)) {
AM.IndexReg = ShVal.getNode()->getOperand(0);
ConstantSDNode *AddVal =
cast<ConstantSDNode>(ShVal.getNode()->getOperand(1));
@@ -964,7 +945,7 @@
case ISD::OR:
// Handle "X | C" as "X + C" iff X is known to have C bits clear.
- if (isLogicallyAddWithConstant(N, CurDAG)) {
+ if (CurDAG->isBaseWithConstantOffset(N)) {
X86ISelAddressMode Backup = AM;
ConstantSDNode *CN = cast<ConstantSDNode>(N.getOperand(1));
uint64_t Offset = CN->getSExtValue();
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 596f010..d5b4a1e 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -4201,8 +4201,7 @@
if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
FI = FINode->getIndex();
Offset = 0;
- } else if (Ptr.getOpcode() == ISD::ADD &&
- isa<ConstantSDNode>(Ptr.getOperand(1)) &&
+ } else if (DAG.isBaseWithConstantOffset(Ptr) &&
isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
Offset = Ptr.getConstantOperandVal(1);