diff options
author | Daniel Sanders <daniel.sanders@imgtec.com> | 2013-09-24 13:33:07 +0000 |
---|---|---|
committer | Daniel Sanders <daniel.sanders@imgtec.com> | 2013-09-24 13:33:07 +0000 |
commit | acfa5a203c01d99aac1bdc1e045c08153bcdbbf6 (patch) | |
tree | cf4cd051e24f258f85e61dc004a1b1ea017550b2 /lib/Target/Mips/MipsSEISelDAGToDAG.cpp | |
parent | ad16ddeb8e07a259d17ef203c9f443f816f6ae7b (diff) | |
download | external_llvm-acfa5a203c01d99aac1bdc1e045c08153bcdbbf6.zip external_llvm-acfa5a203c01d99aac1bdc1e045c08153bcdbbf6.tar.gz external_llvm-acfa5a203c01d99aac1bdc1e045c08153bcdbbf6.tar.bz2 |
[mips][msa] Remove the VSPLAT and VSPLATD nodes in favour of matching BUILD_VECTOR.
Most constant BUILD_VECTOR's are matched using ComplexPatterns which cover
bitcasted as well as normal vectors. However, it doesn't seem to be possible to
match ldi.[bhwd] in a type-agnostic manner (e.g. to support the widest range of
immediates, it should be possible to use ldi.b to load v2i64) using TableGen so
ldi.[bhwd] is matched using custom code in MipsSEISelDAGToDAG.cpp
This made the majority of the constant splat BUILD_VECTOR lowering redundant.
The only transformation remaining for constant splats is when an (up-to) 32-bit
constant splat is possible but the value does not fit into a 10-bit signed
integer. In this case, the BUILD_VECTOR is transformed into a bitcasted
BUILD_VECTOR so that fill.[bhw] can be used to splat the vector from a GPR32
register (which is initialized using the usual lui/addui sequence).
There are no additional tests since this is a re-implementation of previous
functionality. The change is intended to make it easier to implement some of
the upcoming instruction selection patches since they can rely on existing
support for BUILD_VECTOR's in the DAGCombiner.
compare_float.ll changed slightly because a BITCAST is no longer
introduced during legalization.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@191299 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/Target/Mips/MipsSEISelDAGToDAG.cpp')
-rw-r--r-- | lib/Target/Mips/MipsSEISelDAGToDAG.cpp | 217 |
1 files changed, 217 insertions, 0 deletions
diff --git a/lib/Target/Mips/MipsSEISelDAGToDAG.cpp b/lib/Target/Mips/MipsSEISelDAGToDAG.cpp index 10221d5..8a9481c 100644 --- a/lib/Target/Mips/MipsSEISelDAGToDAG.cpp +++ b/lib/Target/Mips/MipsSEISelDAGToDAG.cpp @@ -374,6 +374,147 @@ bool MipsSEDAGToDAGISel::selectIntAddrMM(SDValue Addr, SDValue &Base, selectAddrDefault(Addr, Base, Offset); } +// Select constant vector splats. +// +// Returns true and sets Imm if: +// * MSA is enabled +// * N is a ISD::BUILD_VECTOR representing a constant splat +// * The splat value fits in a signed 32-bit value. +// +// That last requirement isn't strictly a requirement of the instruction set +// but it simplifies the callers by allowing them to assume they don't have to +// handle 64-bit values. The callers will also be placing stricter requirements +// on the immediates so this doesn't prohibit selection of legal immediates. +bool MipsSEDAGToDAGISel::selectVSplat(SDNode *N, APInt &Imm) const { + if (!Subtarget.hasMSA()) + return false; + + BuildVectorSDNode *Node = dyn_cast<BuildVectorSDNode>(N); + + if (Node == NULL) + return false; + + APInt SplatValue, SplatUndef; + unsigned SplatBitSize; + bool HasAnyUndefs; + + if (!Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, + HasAnyUndefs, 8, + !Subtarget.isLittle())) + return false; + + // None of the immediate forms can handle more than 32 bits + if (!SplatValue.isIntN(32)) + return false; + + Imm = SplatValue; + + return true; +} + +// Select constant vector splats. +// +// In addition to the requirements of selectVSplat(), this function returns +// true and sets Imm if: +// * The splat value is the same width as the elements of the vector +// * The splat value fits in an integer with the specified signed-ness and +// width. +// +// This function looks through ISD::BITCAST nodes. +// TODO: This might not be appropriate for big-endian MSA since BITCAST is +// sometimes a shuffle in big-endian mode. +// +// It's worth noting that this function is not used as part of the selection +// of ldi.[bhwd] since it does not permit using the wrong-typed ldi.[bhwd] +// instruction to achieve the desired bit pattern. ldi.[bhwd] is selected in +// MipsSEDAGToDAGISel::selectNode. +bool MipsSEDAGToDAGISel:: +selectVSplatCommon(SDValue N, SDValue &Imm, bool Signed, + unsigned ImmBitSize) const { + APInt ImmValue; + EVT EltTy = N->getValueType(0).getVectorElementType(); + + if (N->getOpcode() == ISD::BITCAST) + N = N->getOperand(0); + + if (selectVSplat (N.getNode(), ImmValue) && + ImmValue.getBitWidth() == EltTy.getSizeInBits()) { + if (( Signed && ImmValue.isSignedIntN(ImmBitSize)) || + (!Signed && ImmValue.isIntN(ImmBitSize))) { + Imm = CurDAG->getTargetConstant(ImmValue, EltTy); + return true; + } + } + + return false; +} + +// Select constant vector splats. +bool MipsSEDAGToDAGISel:: +selectVSplatUimm3(SDValue N, SDValue &Imm) const { + return selectVSplatCommon(N, Imm, false, 3); +} + +// Select constant vector splats. +bool MipsSEDAGToDAGISel:: +selectVSplatUimm4(SDValue N, SDValue &Imm) const { + return selectVSplatCommon(N, Imm, false, 4); +} + +// Select constant vector splats. +bool MipsSEDAGToDAGISel:: +selectVSplatUimm5(SDValue N, SDValue &Imm) const { + return selectVSplatCommon(N, Imm, false, 5); +} + +// Select constant vector splats. +bool MipsSEDAGToDAGISel:: +selectVSplatUimm6(SDValue N, SDValue &Imm) const { + return selectVSplatCommon(N, Imm, false, 6); +} + +// Select constant vector splats. +bool MipsSEDAGToDAGISel:: +selectVSplatUimm8(SDValue N, SDValue &Imm) const { + return selectVSplatCommon(N, Imm, false, 8); +} + +// Select constant vector splats. +bool MipsSEDAGToDAGISel:: +selectVSplatSimm5(SDValue N, SDValue &Imm) const { + return selectVSplatCommon(N, Imm, true, 5); +} + +// Select constant vector splats whose value is a power of 2. +// +// In addition to the requirements of selectVSplat(), this function returns +// true and sets Imm if: +// * The splat value is the same width as the elements of the vector +// * The splat value is a power of two. +// +// This function looks through ISD::BITCAST nodes. +// TODO: This might not be appropriate for big-endian MSA since BITCAST is +// sometimes a shuffle in big-endian mode. +bool MipsSEDAGToDAGISel::selectVSplatUimmPow2(SDValue N, SDValue &Imm) const { + APInt ImmValue; + EVT EltTy = N->getValueType(0).getVectorElementType(); + + if (N->getOpcode() == ISD::BITCAST) + N = N->getOperand(0); + + if (selectVSplat (N.getNode(), ImmValue) && + ImmValue.getBitWidth() == EltTy.getSizeInBits()) { + int32_t Log2 = ImmValue.exactLogBase2(); + + if (Log2 != -1) { + Imm = CurDAG->getTargetConstant(Log2, EltTy); + return true; + } + } + + return false; +} + std::pair<bool, SDNode*> MipsSEDAGToDAGISel::selectNode(SDNode *Node) { unsigned Opcode = Node->getOpcode(); SDLoc DL(Node); @@ -545,6 +686,82 @@ std::pair<bool, SDNode*> MipsSEDAGToDAGISel::selectNode(SDNode *Node) { MVT::Untyped, Ops); return std::make_pair(true, Res); } + + case ISD::BUILD_VECTOR: { + // Select appropriate ldi.[bhwd] instructions for constant splats of + // 128-bit when MSA is enabled. Fixup any register class mismatches that + // occur as a result. + // + // This allows the compiler to use a wider range of immediates than would + // otherwise be allowed. If, for example, v4i32 could only use ldi.h then + // it would not be possible to load { 0x01010101, 0x01010101, 0x01010101, + // 0x01010101 } without using a constant pool. This would be sub-optimal + // when // 'ldi.b wd, 1' is capable of producing that bit-pattern in the + // same set/ of registers. Similarly, ldi.h isn't capable of producing { + // 0x00000000, 0x00000001, 0x00000000, 0x00000001 } but 'ldi.d wd, 1' can. + + BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Node); + APInt SplatValue, SplatUndef; + unsigned SplatBitSize; + bool HasAnyUndefs; + unsigned LdiOp; + EVT ResVecTy = BVN->getValueType(0); + EVT ViaVecTy; + + if (!Subtarget.hasMSA() || !BVN->getValueType(0).is128BitVector()) + return std::make_pair(false, (SDNode*)NULL); + + if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, + HasAnyUndefs, 8, + !Subtarget.isLittle())) + return std::make_pair(false, (SDNode*)NULL); + + switch (SplatBitSize) { + default: + return std::make_pair(false, (SDNode*)NULL); + case 8: + LdiOp = Mips::LDI_B; + ViaVecTy = MVT::v16i8; + break; + case 16: + LdiOp = Mips::LDI_H; + ViaVecTy = MVT::v8i16; + break; + case 32: + LdiOp = Mips::LDI_W; + ViaVecTy = MVT::v4i32; + break; + case 64: + LdiOp = Mips::LDI_D; + ViaVecTy = MVT::v2i64; + break; + } + + if (!SplatValue.isSignedIntN(10)) + return std::make_pair(false, (SDNode*)NULL); + + SDValue Imm = CurDAG->getTargetConstant(SplatValue, + ViaVecTy.getVectorElementType()); + + SDNode *Res = CurDAG->getMachineNode(LdiOp, SDLoc(Node), ViaVecTy, Imm); + + if (ResVecTy != ViaVecTy) { + // If LdiOp is writing to a different register class to ResVecTy, then + // fix it up here. This COPY_TO_REGCLASS should never cause a move.v + // since the source and destination register sets contain the same + // registers. + const TargetLowering *TLI = getTargetLowering(); + MVT ResVecTySimple = ResVecTy.getSimpleVT(); + const TargetRegisterClass *RC = TLI->getRegClassFor(ResVecTySimple); + Res = CurDAG->getMachineNode(Mips::COPY_TO_REGCLASS, SDLoc(Node), + ResVecTy, SDValue(Res, 0), + CurDAG->getTargetConstant(RC->getID(), + MVT::i32)); + } + + return std::make_pair(true, Res); + } + } return std::make_pair(false, (SDNode*)NULL); |