/// specifies a shuffle of elements that is suitable for input to UNPCKL.
static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool HasAVX2, bool V2IsSplat = false) {
- int NumElts = VT.getVectorNumElements();
+ unsigned NumElts = VT.getVectorNumElements();
assert((VT.is128BitVector() || VT.is256BitVector()) &&
"Unsupported vector type for unpckh");
unsigned NumLanes = VT.getSizeInBits()/128;
unsigned NumLaneElts = NumElts/NumLanes;
- unsigned Start = 0;
- unsigned End = NumLaneElts;
- for (unsigned s = 0; s < NumLanes; ++s) {
- for (unsigned i = Start, j = s * NumLaneElts;
- i != End;
+ for (unsigned l = 0; l != NumLanes; ++l) {
+ for (unsigned i = l*NumLaneElts, j = l*NumLaneElts;
+ i != (l+1)*NumLaneElts;
i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
return false;
}
}
- // Process the next 128 bits.
- Start += NumLaneElts;
- End += NumLaneElts;
}
return true;
/// specifies a shuffle of elements that is suitable for input to UNPCKH.
static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool HasAVX2, bool V2IsSplat = false) {
- int NumElts = VT.getVectorNumElements();
+ unsigned NumElts = VT.getVectorNumElements();
assert((VT.is128BitVector() || VT.is256BitVector()) &&
"Unsupported vector type for unpckh");
unsigned NumLanes = VT.getSizeInBits()/128;
unsigned NumLaneElts = NumElts/NumLanes;
- unsigned Start = 0;
- unsigned End = NumLaneElts;
for (unsigned l = 0; l != NumLanes; ++l) {
- for (unsigned i = Start, j = (l*NumLaneElts)+NumLaneElts/2;
- i != End; i += 2, ++j) {
+ for (unsigned i = l*NumLaneElts, j = (l*NumLaneElts)+NumLaneElts/2;
+ i != (l+1)*NumLaneElts; i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
if (!isUndefOrEqual(BitI, j))
return false;
}
}
- // Process the next 128 bits.
- Start += NumLaneElts;
- End += NumLaneElts;
}
return true;
}
/// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
/// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
/// <0, 0, 1, 1>
-static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
- int NumElems = VT.getVectorNumElements();
- if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
+static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT,
+ bool HasAVX2) {
+ unsigned NumElts = VT.getVectorNumElements();
+
+ assert((VT.is128BitVector() || VT.is256BitVector()) &&
+ "Unsupported vector type for unpckh");
+
+ if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8 &&
+ (!HasAVX2 || (NumElts != 16 && NumElts != 32)))
return false;
// For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
// FIXME: Need a better way to get rid of this, there's no latency difference
// between UNPCKLPD and MOVDDUP, the later should always be checked first and
// the former later. We should also remove the "_undef" special mask.
- if (NumElems == 4 && VT.getSizeInBits() == 256)
+ if (NumElts == 4 && VT.getSizeInBits() == 256)
return false;
// Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
// independently on 128-bit lanes.
- unsigned NumLanes = VT.getSizeInBits() / 128;
- unsigned NumLaneElts = NumElems / NumLanes;
+ unsigned NumLanes = VT.getSizeInBits()/128;
+ unsigned NumLaneElts = NumElts/NumLanes;
- for (unsigned s = 0; s < NumLanes; ++s) {
- for (unsigned i = s * NumLaneElts, j = s * NumLaneElts;
- i != NumLaneElts * (s + 1);
+ for (unsigned l = 0; l != NumLanes; ++l) {
+ for (unsigned i = l*NumLaneElts, j = l*NumLaneElts;
+ i != (l+1)*NumLaneElts;
i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
return true;
}
-bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
+bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N, bool HasAVX2) {
SmallVector<int, 8> M;
N->getMask(M);
- return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
+ return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0), HasAVX2);
}
/// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
/// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
/// <2, 2, 3, 3>
-static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
- int NumElems = VT.getVectorNumElements();
- if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
+static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT,
+ bool HasAVX2) {
+ unsigned NumElts = VT.getVectorNumElements();
+
+ assert((VT.is128BitVector() || VT.is256BitVector()) &&
+ "Unsupported vector type for unpckh");
+
+ if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8 &&
+ (!HasAVX2 || (NumElts != 16 && NumElts != 32)))
return false;
- for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
- int BitI = Mask[i];
- int BitI1 = Mask[i+1];
- if (!isUndefOrEqual(BitI, j))
- return false;
- if (!isUndefOrEqual(BitI1, j))
- return false;
+ // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
+ // independently on 128-bit lanes.
+ unsigned NumLanes = VT.getSizeInBits()/128;
+ unsigned NumLaneElts = NumElts/NumLanes;
+
+ for (unsigned l = 0; l != NumLanes; ++l) {
+ for (unsigned i = l*NumLaneElts, j = (l*NumLaneElts)+NumLaneElts/2;
+ i != (l+1)*NumLaneElts; i += 2, ++j) {
+ int BitI = Mask[i];
+ int BitI1 = Mask[i+1];
+ if (!isUndefOrEqual(BitI, j))
+ return false;
+ if (!isUndefOrEqual(BitI1, j))
+ return false;
+ }
}
return true;
}
-bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
+bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N, bool HasAVX2) {
SmallVector<int, 8> M;
N->getMask(M);
- return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
+ return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0), HasAVX2);
}
/// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
DAG);
} else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
- assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
- EVT MiddleVT = MVT::v4i32;
+ unsigned NumBits = VT.getSizeInBits();
+ assert((NumBits == 128 || NumBits == 256) &&
+ "Expected an SSE or AVX value type!");
+ EVT MiddleVT = NumBits == 128 ? MVT::v4i32 : MVT::v8i32;
Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
Item = getShuffleVectorZeroOrUndef(Item, 0, true,
Subtarget->hasXMMInt(), DAG);
// NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
// unpckh_undef). Only use pshufd if speed is more important than size.
- if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
+ if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp, HasAVX2))
return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
- if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
+ if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp, HasAVX2))
return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
if (X86::isMOVDDUPMask(SVOp) && Subtarget->hasSSE3orAVX() &&
return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V2,
X86::getShuffleSHUFImmediate(SVOp), DAG);
- if (isUNPCKL_v_undef_Mask(M, VT))
+ if (isUNPCKL_v_undef_Mask(M, VT, HasAVX2))
return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
- if (isUNPCKH_v_undef_Mask(M, VT))
+ if (isUNPCKH_v_undef_Mask(M, VT, HasAVX2))
return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
//===--------------------------------------------------------------------===//
return DAG.getNode(ISD::MUL, dl, VT, Op, R);
}
if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
+ assert((Subtarget->hasSSE2() || Subtarget->hasAVX()) &&
+ "Need SSE2 for pslli/pcmpeq.");
+
// a = a << 5;
Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
Op.getOperand(1), DAG.getConstant(5, MVT::i32));
- ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
- ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
+ // Turn 'a' into a mask suitable for VSELECT
+ SDValue VSelM = DAG.getConstant(0x80, VT);
+ SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
+ OpVSel = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ DAG.getConstant(Intrinsic::x86_sse2_pcmpeq_b, MVT::i32),
+ OpVSel, VSelM);
- std::vector<Constant*> CVM1(16, CM1);
- std::vector<Constant*> CVM2(16, CM2);
- Constant *C = ConstantVector::get(CVM1);
- SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
- SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
- MachinePointerInfo::getConstantPool(),
- false, false, false, 16);
+ SDValue CM1 = DAG.getConstant(0x0f, VT);
+ SDValue CM2 = DAG.getConstant(0x3f, VT);
- // r = pblendv(r, psllw(r & (char16)15, 4), a);
- M = DAG.getNode(ISD::AND, dl, VT, R, M);
+ // r = VSELECT(r, psllw(r & (char16)15, 4), a);
+ SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(4, MVT::i32));
- R = DAG.getNode(ISD::VSELECT, dl, VT, Op, R, M);
+ R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
+
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
+ OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
+ OpVSel = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ DAG.getConstant(Intrinsic::x86_sse2_pcmpeq_b, MVT::i32),
+ OpVSel, VSelM);
- C = ConstantVector::get(CVM2);
- CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
- M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
- MachinePointerInfo::getConstantPool(),
- false, false, false, 16);
-
- // r = pblendv(r, psllw(r & (char16)63, 2), a);
- M = DAG.getNode(ISD::AND, dl, VT, R, M);
+ // r = VSELECT(r, psllw(r & (char16)63, 2), a);
+ M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(2, MVT::i32));
- R = DAG.getNode(ISD::VSELECT, dl, VT, Op, R, M);
+ R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
+
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
-
- // return pblendv(r, r+r, a);
- R = DAG.getNode(ISD::VSELECT, dl, VT, Op,
- R, DAG.getNode(ISD::ADD, dl, VT, R, R));
+ OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
+ OpVSel = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ DAG.getConstant(Intrinsic::x86_sse2_pcmpeq_b, MVT::i32),
+ OpVSel, VSelM);
+
+ // return VSELECT(r, r+r, a);
+ R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
+ DAG.getNode(ISD::ADD, dl, VT, R, R), R);
return R;
}
case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
- case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
isPALIGNRMask(M, VT, Subtarget->hasSSSE3orAVX()) ||
isUNPCKLMask(M, VT, Subtarget->hasAVX2()) ||
isUNPCKHMask(M, VT, Subtarget->hasAVX2()) ||
- isUNPCKL_v_undef_Mask(M, VT) ||
- isUNPCKH_v_undef_Mask(M, VT));
+ isUNPCKL_v_undef_Mask(M, VT, Subtarget->hasAVX2()) ||
+ isUNPCKH_v_undef_Mask(M, VT, Subtarget->hasAVX2()));
}
bool
// X86 Inline Assembly Support
//===----------------------------------------------------------------------===//
+// Helper to match a string separated by whitespace.
+static bool END_WITH_NULL matchAsm(StringRef s, ...) {
+ va_list ap;
+ va_start(ap, s);
+ s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
+
+ while (const char *p = va_arg(ap, const char *)) {
+ StringRef piece(p);
+ if (!s.startswith(piece)) { // Check if the piece matches.
+ va_end(ap);
+ return false;
+ }
+
+ s = s.substr(piece.size());
+ StringRef::size_type i = s.find_first_not_of(" \t");
+ if (i == 0) { // We matched a prefix.
+ va_end(ap);
+ return false;
+ }
+ s = s.substr(i);
+ }
+
+ va_end(ap);
+ return s.empty();
+}
+
bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
std::string AsmStr = IA->getAsmString();
+ IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
+ if (!Ty || Ty->getBitWidth() % 16 != 0)
+ return false;
+
// TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
SmallVector<StringRef, 4> AsmPieces;
SplitString(AsmStr, AsmPieces, ";\n");
switch (AsmPieces.size()) {
default: return false;
case 1:
- AsmStr = AsmPieces[0];
- AsmPieces.clear();
- SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
-
// FIXME: this should verify that we are targeting a 486 or better. If not,
- // we will turn this bswap into something that will be lowered to logical ops
- // instead of emitting the bswap asm. For now, we don't support 486 or lower
- // so don't worry about this.
+ // we will turn this bswap into something that will be lowered to logical
+ // ops instead of emitting the bswap asm. For now, we don't support 486 or
+ // lower so don't worry about this.
// bswap $0
- if (AsmPieces.size() == 2 &&
- (AsmPieces[0] == "bswap" ||
- AsmPieces[0] == "bswapq" ||
- AsmPieces[0] == "bswapl") &&
- (AsmPieces[1] == "$0" ||
- AsmPieces[1] == "${0:q}")) {
+ if (matchAsm(AsmPieces[0], "bswap", "$0", NULL) ||
+ matchAsm(AsmPieces[0], "bswapl", "$0", NULL) ||
+ matchAsm(AsmPieces[0], "bswapq", "$0", NULL) ||
+ matchAsm(AsmPieces[0], "bswap", "${0:q}", NULL) ||
+ matchAsm(AsmPieces[0], "bswapl", "${0:q}", NULL) ||
+ matchAsm(AsmPieces[0], "bswapq", "${0:q}", NULL)) {
// No need to check constraints, nothing other than the equivalent of
// "=r,0" would be valid here.
- IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
- if (!Ty || Ty->getBitWidth() % 16 != 0)
- return false;
return IntrinsicLowering::LowerToByteSwap(CI);
}
+
// rorw $$8, ${0:w} --> llvm.bswap.i16
if (CI->getType()->isIntegerTy(16) &&
- AsmPieces.size() == 3 &&
- (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
- AsmPieces[1] == "$$8," &&
- AsmPieces[2] == "${0:w}" &&
- IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
+ IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
+ (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}", NULL) ||
+ matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}", NULL))) {
AsmPieces.clear();
const std::string &ConstraintsStr = IA->getConstraintString();
SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
AsmPieces[0] == "~{cc}" &&
AsmPieces[1] == "~{dirflag}" &&
AsmPieces[2] == "~{flags}" &&
- AsmPieces[3] == "~{fpsr}") {
- IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
- if (!Ty || Ty->getBitWidth() % 16 != 0)
- return false;
- return IntrinsicLowering::LowerToByteSwap(CI);
- }
+ AsmPieces[3] == "~{fpsr}")
+ return IntrinsicLowering::LowerToByteSwap(CI);
}
break;
case 3:
if (CI->getType()->isIntegerTy(32) &&
- IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
- SmallVector<StringRef, 4> Words;
- SplitString(AsmPieces[0], Words, " \t,");
- if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
- Words[2] == "${0:w}") {
- Words.clear();
- SplitString(AsmPieces[1], Words, " \t,");
- if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
- Words[2] == "$0") {
- Words.clear();
- SplitString(AsmPieces[2], Words, " \t,");
- if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
- Words[2] == "${0:w}") {
- AsmPieces.clear();
- const std::string &ConstraintsStr = IA->getConstraintString();
- SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
- std::sort(AsmPieces.begin(), AsmPieces.end());
- if (AsmPieces.size() == 4 &&
- AsmPieces[0] == "~{cc}" &&
- AsmPieces[1] == "~{dirflag}" &&
- AsmPieces[2] == "~{flags}" &&
- AsmPieces[3] == "~{fpsr}") {
- IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
- if (!Ty || Ty->getBitWidth() % 16 != 0)
- return false;
- return IntrinsicLowering::LowerToByteSwap(CI);
- }
- }
- }
- }
+ IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
+ matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}", NULL) &&
+ matchAsm(AsmPieces[1], "rorl", "$$16,", "$0", NULL) &&
+ matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}", NULL)) {
+ AsmPieces.clear();
+ const std::string &ConstraintsStr = IA->getConstraintString();
+ SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
+ std::sort(AsmPieces.begin(), AsmPieces.end());
+ if (AsmPieces.size() == 4 &&
+ AsmPieces[0] == "~{cc}" &&
+ AsmPieces[1] == "~{dirflag}" &&
+ AsmPieces[2] == "~{flags}" &&
+ AsmPieces[3] == "~{fpsr}")
+ return IntrinsicLowering::LowerToByteSwap(CI);
}
if (CI->getType()->isIntegerTy(64)) {
Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
// bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
- SmallVector<StringRef, 4> Words;
- SplitString(AsmPieces[0], Words, " \t");
- if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
- Words.clear();
- SplitString(AsmPieces[1], Words, " \t");
- if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
- Words.clear();
- SplitString(AsmPieces[2], Words, " \t,");
- if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
- Words[2] == "%edx") {
- IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
- if (!Ty || Ty->getBitWidth() % 16 != 0)
- return false;
- return IntrinsicLowering::LowerToByteSwap(CI);
- }
- }
- }
+ if (matchAsm(AsmPieces[0], "bswap", "%eax", NULL) &&
+ matchAsm(AsmPieces[1], "bswap", "%edx", NULL) &&
+ matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx", NULL))
+ return IntrinsicLowering::LowerToByteSwap(CI);
}
}
break;