Fix the SVOffset values for loads and stores produced by

memcpy/memset expansion. It was a bug for the SVOffset value
to be used in the actual address calculations.

llvm-svn: 50359
This commit is contained in:
Dan Gohman 2008-04-28 17:15:20 +00:00
parent 1b7238e6e4
commit 0285c1e9bb
6 changed files with 43 additions and 39 deletions

View File

@ -326,17 +326,17 @@ public:
SDOperand getMemcpy(SDOperand Chain, SDOperand Dst, SDOperand Src,
SDOperand Size, unsigned Align,
bool AlwaysInline,
const Value *DstSV, uint64_t DstOff,
const Value *SrcSV, uint64_t SrcOff);
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff);
SDOperand getMemmove(SDOperand Chain, SDOperand Dst, SDOperand Src,
SDOperand Size, unsigned Align,
const Value *DstSV, uint64_t DstOff,
const Value *SrcSV, uint64_t SrcOff);
const Value *DstSV, uint64_t DstOSVff,
const Value *SrcSV, uint64_t SrcSVOff);
SDOperand getMemset(SDOperand Chain, SDOperand Dst, SDOperand Src,
SDOperand Size, unsigned Align,
const Value *DstSV, uint64_t DstOff);
const Value *DstSV, uint64_t DstSVOff);
/// getSetCC - Helper function to make it easier to build SetCC's if you just
/// have an ISD::CondCode instead of an SDOperand.

View File

@ -2590,8 +2590,8 @@ static SDOperand getMemcpyLoadsAndStores(SelectionDAG &DAG,
SDOperand Src, uint64_t Size,
unsigned Align,
bool AlwaysInline,
const Value *DstSV, uint64_t DstOff,
const Value *SrcSV, uint64_t SrcOff) {
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff){
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
// Expand memcpy to a series of store ops if the size operand falls below
@ -2610,6 +2610,7 @@ static SDOperand getMemcpyLoadsAndStores(SelectionDAG &DAG,
GlobalAddressSDNode *G = NULL;
std::string Str;
bool CopyFromStr = false;
uint64_t SrcOff = 0, DstOff = 0;
if (Src.getOpcode() == ISD::GlobalAddress)
G = cast<GlobalAddressSDNode>(Src);
@ -2640,15 +2641,15 @@ static SDOperand getMemcpyLoadsAndStores(SelectionDAG &DAG,
Store =
DAG.getStore(Chain, Value,
getMemBasePlusOffset(Dst, DstOff, DAG),
DstSV, DstOff);
DstSV, DstSVOff + DstOff);
} else {
Value = DAG.getLoad(VT, Chain,
getMemBasePlusOffset(Src, SrcOff, DAG),
SrcSV, SrcOff, false, Align);
SrcSV, SrcSVOff + SrcOff, false, Align);
Store =
DAG.getStore(Chain, Value,
getMemBasePlusOffset(Dst, DstOff, DAG),
DstSV, DstOff, false, Align);
DstSV, DstSVOff + DstOff, false, Align);
}
OutChains.push_back(Store);
SrcOff += VTSize;
@ -2663,7 +2664,7 @@ static SDOperand getMemsetStores(SelectionDAG &DAG,
SDOperand Chain, SDOperand Dst,
SDOperand Src, uint64_t Size,
unsigned Align,
const Value *DstSV, uint64_t DstOff) {
const Value *DstSV, uint64_t DstSVOff) {
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
// Expand memset to a series of load/store ops if the size operand
@ -2674,6 +2675,7 @@ static SDOperand getMemsetStores(SelectionDAG &DAG,
return SDOperand();
SmallVector<SDOperand, 8> OutChains;
uint64_t DstOff = 0;
unsigned NumMemOps = MemOps.size();
for (unsigned i = 0; i < NumMemOps; i++) {
@ -2682,7 +2684,7 @@ static SDOperand getMemsetStores(SelectionDAG &DAG,
SDOperand Value = getMemsetValue(Src, VT, DAG);
SDOperand Store = DAG.getStore(Chain, Value,
getMemBasePlusOffset(Dst, DstOff, DAG),
DstSV, DstOff);
DstSV, DstSVOff + DstOff);
OutChains.push_back(Store);
DstOff += VTSize;
}
@ -2694,8 +2696,8 @@ static SDOperand getMemsetStores(SelectionDAG &DAG,
SDOperand SelectionDAG::getMemcpy(SDOperand Chain, SDOperand Dst,
SDOperand Src, SDOperand Size,
unsigned Align, bool AlwaysInline,
const Value *DstSV, uint64_t DstOff,
const Value *SrcSV, uint64_t SrcOff) {
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff) {
// Check to see if we should lower the memcpy to loads and stores first.
// For cases within the target-specified limits, this is the best choice.
@ -2707,7 +2709,7 @@ SDOperand SelectionDAG::getMemcpy(SDOperand Chain, SDOperand Dst,
SDOperand Result =
getMemcpyLoadsAndStores(*this, Chain, Dst, Src, ConstantSize->getValue(),
Align, false, DstSV, DstOff, SrcSV, SrcOff);
Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff);
if (Result.Val)
return Result;
}
@ -2717,7 +2719,7 @@ SDOperand SelectionDAG::getMemcpy(SDOperand Chain, SDOperand Dst,
SDOperand Result =
TLI.EmitTargetCodeForMemcpy(*this, Chain, Dst, Src, Size, Align,
AlwaysInline,
DstSV, DstOff, SrcSV, SrcOff);
DstSV, DstSVOff, SrcSV, SrcSVOff);
if (Result.Val)
return Result;
@ -2727,7 +2729,7 @@ SDOperand SelectionDAG::getMemcpy(SDOperand Chain, SDOperand Dst,
assert(ConstantSize && "AlwaysInline requires a constant size!");
return getMemcpyLoadsAndStores(*this, Chain, Dst, Src,
ConstantSize->getValue(), Align, true,
DstSV, DstOff, SrcSV, SrcOff);
DstSV, DstSVOff, SrcSV, SrcSVOff);
}
// Emit a library call.
@ -2748,8 +2750,8 @@ SDOperand SelectionDAG::getMemcpy(SDOperand Chain, SDOperand Dst,
SDOperand SelectionDAG::getMemmove(SDOperand Chain, SDOperand Dst,
SDOperand Src, SDOperand Size,
unsigned Align,
const Value *DstSV, uint64_t DstOff,
const Value *SrcSV, uint64_t SrcOff) {
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff) {
// TODO: Optimize small memmove cases with simple loads and stores,
// ensuring that all loads precede all stores. This can cause severe
@ -2759,7 +2761,7 @@ SDOperand SelectionDAG::getMemmove(SDOperand Chain, SDOperand Dst,
// code. If the target chooses to do this, this is the next best.
SDOperand Result =
TLI.EmitTargetCodeForMemmove(*this, Chain, Dst, Src, Size, Align,
DstSV, DstOff, SrcSV, SrcOff);
DstSV, DstSVOff, SrcSV, SrcSVOff);
if (Result.Val)
return Result;
@ -2781,7 +2783,7 @@ SDOperand SelectionDAG::getMemmove(SDOperand Chain, SDOperand Dst,
SDOperand SelectionDAG::getMemset(SDOperand Chain, SDOperand Dst,
SDOperand Src, SDOperand Size,
unsigned Align,
const Value *DstSV, uint64_t DstOff) {
const Value *DstSV, uint64_t DstSVOff) {
// Check to see if we should lower the memset to stores first.
// For cases within the target-specified limits, this is the best choice.
@ -2793,7 +2795,7 @@ SDOperand SelectionDAG::getMemset(SDOperand Chain, SDOperand Dst,
SDOperand Result =
getMemsetStores(*this, Chain, Dst, Src, ConstantSize->getValue(), Align,
DstSV, DstOff);
DstSV, DstSVOff);
if (Result.Val)
return Result;
}
@ -2802,7 +2804,7 @@ SDOperand SelectionDAG::getMemset(SDOperand Chain, SDOperand Dst,
// code. If the target chooses to do this, this is the next best.
SDOperand Result =
TLI.EmitTargetCodeForMemset(*this, Chain, Dst, Src, Size, Align,
DstSV, DstOff);
DstSV, DstSVOff);
if (Result.Val)
return Result;

View File

@ -1247,8 +1247,8 @@ ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
SDOperand Dst, SDOperand Src,
SDOperand Size, unsigned Align,
bool AlwaysInline,
const Value *DstSV, uint64_t DstOff,
const Value *SrcSV, uint64_t SrcOff){
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff){
// Do repeated 4-byte loads and stores. To be improved.
// This requires 4-byte alignment.
if ((Align & 3) != 0)
@ -1271,6 +1271,7 @@ ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
const unsigned MAX_LOADS_IN_LDM = 6;
SDOperand TFOps[MAX_LOADS_IN_LDM];
SDOperand Loads[MAX_LOADS_IN_LDM];
uint64_t SrcOff = 0, DstOff = 0;
// Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
// same number of stores. The loads and stores will get combined into
@ -1281,7 +1282,7 @@ ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
Loads[i] = DAG.getLoad(VT, Chain,
DAG.getNode(ISD::ADD, MVT::i32, Src,
DAG.getConstant(SrcOff, MVT::i32)),
SrcSV, SrcOff);
SrcSV, SrcSVOff + SrcOff);
TFOps[i] = Loads[i].getValue(1);
SrcOff += VTSize;
}
@ -1292,7 +1293,7 @@ ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
TFOps[i] = DAG.getStore(Chain, Loads[i],
DAG.getNode(ISD::ADD, MVT::i32, Dst,
DAG.getConstant(DstOff, MVT::i32)),
DstSV, DstOff);
DstSV, DstSVOff + DstOff);
DstOff += VTSize;
}
Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, &TFOps[0], i);
@ -1318,7 +1319,7 @@ ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
Loads[i] = DAG.getLoad(VT, Chain,
DAG.getNode(ISD::ADD, MVT::i32, Src,
DAG.getConstant(SrcOff, MVT::i32)),
SrcSV, SrcOff);
SrcSV, SrcSVOff + SrcOff);
TFOps[i] = Loads[i].getValue(1);
++i;
SrcOff += VTSize;
@ -1340,7 +1341,7 @@ ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
TFOps[i] = DAG.getStore(Chain, Loads[i],
DAG.getNode(ISD::ADD, MVT::i32, Dst,
DAG.getConstant(DstOff, MVT::i32)),
DstSV, DstOff);
DstSV, DstSVOff + DstOff);
++i;
DstOff += VTSize;
BytesLeft -= VTSize;

View File

@ -149,8 +149,8 @@ namespace llvm {
SDOperand Dst, SDOperand Src,
SDOperand Size, unsigned Align,
bool AlwaysInline,
const Value *DstSV, uint64_t DstOff,
const Value *SrcSV, uint64_t SrcOff);
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff);
};
}

View File

@ -4763,7 +4763,7 @@ X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
SDOperand Chain,
SDOperand Dst, SDOperand Src,
SDOperand Size, unsigned Align,
const Value *DstSV, uint64_t DstOff) {
const Value *DstSV, uint64_t DstSVOff) {
ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
/// If not DWORD aligned or size is more than the threshold, call the library.
@ -4890,7 +4890,7 @@ X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG,
DAG.getConstant(Offset, AddrVT)),
Src,
DAG.getConstant(BytesLeft, SizeVT),
Align, DstSV, 0);
Align, DstSV, DstSVOff + Offset);
}
// TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
@ -4903,8 +4903,8 @@ X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
SDOperand Dst, SDOperand Src,
SDOperand Size, unsigned Align,
bool AlwaysInline,
const Value *DstSV, uint64_t DstOff,
const Value *SrcSV, uint64_t SrcOff){
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff){
// This requires the copy size to be a constant, preferrably
// within a subtarget-specific limit.
@ -4964,7 +4964,8 @@ X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG,
DAG.getConstant(Offset, SrcVT)),
DAG.getConstant(BytesLeft, SizeVT),
Align, AlwaysInline,
DstSV, 0, SrcSV, 0));
DstSV, DstSVOff + Offset,
SrcSV, SrcSVOff + Offset));
}
return DAG.getNode(ISD::TokenFactor, MVT::Other, &Results[0], Results.size());

View File

@ -538,14 +538,14 @@ namespace llvm {
SDOperand Chain,
SDOperand Dst, SDOperand Src,
SDOperand Size, unsigned Align,
const Value *DstSV, uint64_t DstOff);
const Value *DstSV, uint64_t DstSVOff);
SDOperand EmitTargetCodeForMemcpy(SelectionDAG &DAG,
SDOperand Chain,
SDOperand Dst, SDOperand Src,
SDOperand Size, unsigned Align,
bool AlwaysInline,
const Value *DstSV, uint64_t DstOff,
const Value *SrcSV, uint64_t SrcOff);
const Value *DstSV, uint64_t DstSVOff,
const Value *SrcSV, uint64_t SrcSVOff);
};
}