Use subword loads instead of a 4-byte load when the size of a structure (or a
piece of it) that is being passed by value is smaller than a word.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@138007 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/Mips/MipsISelLowering.cpp b/lib/Target/Mips/MipsISelLowering.cpp
index 97b5a14..f2682cf 100644
--- a/lib/Target/Mips/MipsISelLowering.cpp
+++ b/lib/Target/Mips/MipsISelLowering.cpp
@@ -1805,43 +1805,90 @@
SmallVector<SDValue, 8>& MemOpChains, int& LastFI,
MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg,
const CCValAssign &VA, const ISD::ArgFlagsTy& Flags,
- MVT PtrType) {
- unsigned FirstWord = VA.getLocMemOffset() / 4;
- unsigned NumWords = (Flags.getByValSize() + 3) / 4;
- unsigned LastWord = FirstWord + NumWords;
- unsigned CurWord;
+ MVT PtrType, bool isLittle) {
+ unsigned LocMemOffset = VA.getLocMemOffset();
+ unsigned Offset = 0;
+ uint32_t RemainingSize = Flags.getByValSize();
unsigned ByValAlign = Flags.getByValAlign();
- // copy the first 4 words of byval arg to registers A0 - A3
- for (CurWord = FirstWord; CurWord < std::min(LastWord, O32IntRegsSize);
- ++CurWord) {
+ // Copy the first 4 words of byval arg to registers A0 - A3.
+ // FIXME: Use a stricter alignment if it enables better optimization in passes
+ // run later.
+ for (; RemainingSize >= 4 && LocMemOffset < 4 * 4;
+ Offset += 4, RemainingSize -= 4, LocMemOffset += 4) {
SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
- DAG.getConstant((CurWord - FirstWord) * 4,
- MVT::i32));
+ DAG.getConstant(Offset, MVT::i32));
SDValue LoadVal = DAG.getLoad(MVT::i32, dl, Chain, LoadPtr,
MachinePointerInfo(),
false, false, std::min(ByValAlign,
(unsigned )4));
MemOpChains.push_back(LoadVal.getValue(1));
- unsigned DstReg = O32IntRegs[CurWord];
+ unsigned DstReg = O32IntRegs[LocMemOffset / 4];
RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
}
- // copy remaining part of byval arg to stack.
- if (CurWord < LastWord) {
- unsigned SizeInBytes = (LastWord - CurWord) * 4;
- SDValue Src = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
- DAG.getConstant((CurWord - FirstWord) * 4,
- MVT::i32));
- LastFI = MFI->CreateFixedObject(SizeInBytes, CurWord * 4, true);
- SDValue Dst = DAG.getFrameIndex(LastFI, PtrType);
- Chain = DAG.getMemcpy(Chain, dl, Dst, Src,
- DAG.getConstant(SizeInBytes, MVT::i32),
- /*Align*/ByValAlign,
- /*isVolatile=*/false, /*AlwaysInline=*/false,
- MachinePointerInfo(0), MachinePointerInfo(0));
- MemOpChains.push_back(Chain);
+ if (RemainingSize == 0)
+ return;
+
+ // If there still is a register available for argument passing, write the
+ // remaining part of the structure to it using subword loads and shifts.
+ if (LocMemOffset < 4 * 4) {
+ assert(RemainingSize <= 3 && RemainingSize >= 1 &&
+ "There must be one to three bytes remaining.");
+ unsigned LoadSize = (RemainingSize == 3 ? 2 : RemainingSize);
+ SDValue LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
+ DAG.getConstant(Offset, MVT::i32));
+ unsigned Alignment = std::min(ByValAlign, (unsigned )4);
+ SDValue LoadVal = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
+ LoadPtr, MachinePointerInfo(),
+ MVT::getIntegerVT(LoadSize * 8), false,
+ false, Alignment);
+ MemOpChains.push_back(LoadVal.getValue(1));
+
+ // If target is big endian, shift it to the most significant half-word or
+ // byte.
+ if (!isLittle)
+ LoadVal = DAG.getNode(ISD::SHL, dl, MVT::i32, LoadVal,
+ DAG.getConstant(32 - LoadSize * 8, MVT::i32));
+
+ Offset += LoadSize;
+ RemainingSize -= LoadSize;
+
+ // Read second subword if necessary.
+ if (RemainingSize != 0) {
+ assert(RemainingSize == 1 && "There must be one byte remaining.");
+ LoadPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
+ DAG.getConstant(Offset, MVT::i32));
+ unsigned Alignment = std::min(ByValAlign, (unsigned )2);
+ SDValue Subword = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, Chain,
+ LoadPtr, MachinePointerInfo(),
+ MVT::i8, false, false, Alignment);
+ MemOpChains.push_back(Subword.getValue(1));
+ // Insert the loaded byte to LoadVal.
+ // FIXME: Use INS if supported by target.
+ unsigned ShiftAmt = isLittle ? 16 : 8;
+ SDValue Shift = DAG.getNode(ISD::SHL, dl, MVT::i32, Subword,
+ DAG.getConstant(ShiftAmt, MVT::i32));
+ LoadVal = DAG.getNode(ISD::OR, dl, MVT::i32, LoadVal, Shift);
+ }
+
+ unsigned DstReg = O32IntRegs[LocMemOffset / 4];
+ RegsToPass.push_back(std::make_pair(DstReg, LoadVal));
+ return;
}
+
+ // Create a fixed object on stack at offset LocMemOffset and copy
+ // remaining part of byval arg to it using memcpy.
+ SDValue Src = DAG.getNode(ISD::ADD, dl, MVT::i32, Arg,
+ DAG.getConstant(Offset, MVT::i32));
+ LastFI = MFI->CreateFixedObject(RemainingSize, LocMemOffset, true);
+ SDValue Dst = DAG.getFrameIndex(LastFI, PtrType);
+ Chain = DAG.getMemcpy(Chain, dl, Dst, Src,
+ DAG.getConstant(RemainingSize, MVT::i32),
+ std::min(ByValAlign, (unsigned)4),
+ /*isVolatile=*/false, /*AlwaysInline=*/false,
+ MachinePointerInfo(0), MachinePointerInfo(0));
+ MemOpChains.push_back(Chain);
}
/// LowerCall - functions arguments are copied from virtual regs to
@@ -1974,7 +2021,7 @@
assert(Flags.getByValSize() &&
"ByVal args of size 0 should have been ignored by front-end.");
WriteByValArg(Chain, dl, RegsToPass, MemOpChains, LastFI, MFI, DAG, Arg,
- VA, Flags, getPointerTy());
+ VA, Flags, getPointerTy(), Subtarget->isLittle());
continue;
}