[llvm] 7ccb31a - [X86] splitVectorOp - share the same SDLoc argument instead of recreating it over and over again.
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Wed Mar 20 04:21:23 PDT 2024
Author: Simon Pilgrim
Date: 2024-03-20T11:21:09Z
New Revision: 7ccb31a5bcd60d6e691a57bf02ebd1e1ebfbe675
URL: https://github.com/llvm/llvm-project/commit/7ccb31a5bcd60d6e691a57bf02ebd1e1ebfbe675
DIFF: https://github.com/llvm/llvm-project/commit/7ccb31a5bcd60d6e691a57bf02ebd1e1ebfbe675.diff
LOG: [X86] splitVectorOp - share the same SDLoc argument instead of recreating it over and over again.
Added:
Modified:
llvm/lib/Target/X86/X86ISelLowering.cpp
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 8a6594230f0a2c..35f756ea5e1d86 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -4070,10 +4070,9 @@ static std::pair<SDValue, SDValue> splitVector(SDValue Op, SelectionDAG &DAG,
}
/// Break an operation into 2 half sized ops and then concatenate the results.
-static SDValue splitVectorOp(SDValue Op, SelectionDAG &DAG) {
+static SDValue splitVectorOp(SDValue Op, SelectionDAG &DAG, const SDLoc &dl) {
unsigned NumOps = Op.getNumOperands();
EVT VT = Op.getValueType();
- SDLoc dl(Op);
// Extract the LHS Lo/Hi vectors
SmallVector<SDValue> LoOps(NumOps, SDValue());
@@ -4096,7 +4095,8 @@ static SDValue splitVectorOp(SDValue Op, SelectionDAG &DAG) {
/// Break an unary integer operation into 2 half sized ops and then
/// concatenate the result back.
-static SDValue splitVectorIntUnary(SDValue Op, SelectionDAG &DAG) {
+static SDValue splitVectorIntUnary(SDValue Op, SelectionDAG &DAG,
+ const SDLoc &dl) {
// Make sure we only try to split 256/512-bit types to avoid creating
// narrow vectors.
EVT VT = Op.getValueType();
@@ -4107,19 +4107,20 @@ static SDValue splitVectorIntUnary(SDValue Op, SelectionDAG &DAG) {
assert(Op.getOperand(0).getValueType().getVectorNumElements() ==
VT.getVectorNumElements() &&
"Unexpected VTs!");
- return splitVectorOp(Op, DAG);
+ return splitVectorOp(Op, DAG, dl);
}
/// Break a binary integer operation into 2 half sized ops and then
/// concatenate the result back.
-static SDValue splitVectorIntBinary(SDValue Op, SelectionDAG &DAG) {
+static SDValue splitVectorIntBinary(SDValue Op, SelectionDAG &DAG,
+ const SDLoc &dl) {
// Assert that all the types match.
EVT VT = Op.getValueType();
(void)VT;
assert(Op.getOperand(0).getValueType() == VT &&
Op.getOperand(1).getValueType() == VT && "Unexpected VTs!");
assert((VT.is256BitVector() || VT.is512BitVector()) && "Unsupported VT!");
- return splitVectorOp(Op, DAG);
+ return splitVectorOp(Op, DAG, dl);
}
// Helper for splitting operands of an operation to legal target size and
@@ -20075,7 +20076,7 @@ static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
if (VT == MVT::v32i16 && !Subtarget.hasBWI()) {
assert(InVT == MVT::v32i8 && "Unexpected VT!");
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, dl);
}
if (Subtarget.hasInt256())
@@ -20656,7 +20657,7 @@ SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
if (Subtarget.hasAVX512()) {
if (InVT == MVT::v32i16 && !Subtarget.hasBWI()) {
assert(VT == MVT::v32i8 && "Unexpected VT!");
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
}
// word to byte only under BWI. Otherwise we have to promoted to v16i32
@@ -21636,7 +21637,8 @@ SDValue X86TargetLowering::LowerFP_TO_BF16(SDValue Op,
/// Depending on uarch and/or optimizing for size, we might prefer to use a
/// vector operation in place of the typical scalar operation.
-static SDValue lowerAddSubToHorizontalOp(SDValue Op, SelectionDAG &DAG,
+static SDValue lowerAddSubToHorizontalOp(SDValue Op, const SDLoc &DL,
+ SelectionDAG &DAG,
const X86Subtarget &Subtarget) {
// If both operands have other uses, this is probably not profitable.
SDValue LHS = Op.getOperand(0);
@@ -21692,7 +21694,6 @@ static SDValue lowerAddSubToHorizontalOp(SDValue Op, SelectionDAG &DAG,
// Creating a 256-bit horizontal op would be wasteful, and there is no 512-bit
// equivalent, so extract the 256/512-bit source op to 128-bit if we can.
- SDLoc DL(Op);
if (BitWidth == 256 || BitWidth == 512) {
unsigned LaneIdx = LExtIndex / NumEltsPerLane;
X = extract128BitVector(X, LaneIdx * NumEltsPerLane, DAG, DL);
@@ -21713,7 +21714,7 @@ static SDValue lowerAddSubToHorizontalOp(SDValue Op, SelectionDAG &DAG,
SDValue X86TargetLowering::lowerFaddFsub(SDValue Op, SelectionDAG &DAG) const {
assert((Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::f64) &&
"Only expecting float/double");
- return lowerAddSubToHorizontalOp(Op, DAG, Subtarget);
+ return lowerAddSubToHorizontalOp(Op, SDLoc(Op), DAG, Subtarget);
}
/// ISD::FROUND is defined to round to nearest with ties rounding away from 0.
@@ -24470,7 +24471,7 @@ static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget &Subtarget,
if (VT == MVT::v32i16 && !Subtarget.hasBWI()) {
assert(InVT == MVT::v32i8 && "Unexpected VT!");
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, dl);
}
if (Subtarget.hasInt256())
@@ -27833,7 +27834,7 @@ static SDValue LowerVectorCTLZ_AVX512CDI(SDValue Op, SelectionDAG &DAG,
// Split vector, it's Lo and Hi parts will be handled in next iteration.
if (NumElems > 16 ||
(NumElems == 16 && !Subtarget.canExtendTo512DQ()))
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, dl);
MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
@@ -27943,11 +27944,11 @@ static SDValue LowerVectorCTLZ(SDValue Op, const SDLoc &DL,
// Decompose 256-bit ops into smaller 128-bit ops.
if (VT.is256BitVector() && !Subtarget.hasInt256())
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
// Decompose 512-bit ops into smaller 256-bit ops.
if (VT.is512BitVector() && !Subtarget.hasBWI())
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
assert(Subtarget.hasSSSE3() && "Expected SSSE3 support for PSHUFB");
return LowerVectorCTLZInRegLUT(Op, DL, Subtarget, DAG);
@@ -28020,16 +28021,18 @@ static SDValue LowerCTTZ(SDValue Op, const X86Subtarget &Subtarget,
static SDValue lowerAddSub(SDValue Op, SelectionDAG &DAG,
const X86Subtarget &Subtarget) {
MVT VT = Op.getSimpleValueType();
+ SDLoc DL(Op);
+
if (VT == MVT::i16 || VT == MVT::i32)
- return lowerAddSubToHorizontalOp(Op, DAG, Subtarget);
+ return lowerAddSubToHorizontalOp(Op, DL, DAG, Subtarget);
if (VT == MVT::v32i16 || VT == MVT::v64i8)
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, DL);
assert(Op.getSimpleValueType().is256BitVector() &&
Op.getSimpleValueType().isInteger() &&
"Only handle AVX 256-bit vector integer operation");
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, DL);
}
static SDValue LowerADDSAT_SUBSAT(SDValue Op, SelectionDAG &DAG,
@@ -28043,7 +28046,7 @@ static SDValue LowerADDSAT_SUBSAT(SDValue Op, SelectionDAG &DAG,
(VT.is256BitVector() && !Subtarget.hasInt256())) {
assert(Op.getSimpleValueType().isInteger() &&
"Only handle AVX vector integer operation");
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, DL);
}
// Avoid the generic expansion with min/max if we don't have pminu*/pmaxu*.
@@ -28105,10 +28108,11 @@ static SDValue LowerADDSAT_SUBSAT(SDValue Op, SelectionDAG &DAG,
static SDValue LowerABS(SDValue Op, const X86Subtarget &Subtarget,
SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
+ SDLoc DL(Op);
+
if (VT == MVT::i16 || VT == MVT::i32 || VT == MVT::i64) {
// Since X86 does not have CMOV for 8-bit integer, we don't convert
// 8-bit integer abs to NEG and CMOV.
- SDLoc DL(Op);
SDValue N0 = Op.getOperand(0);
SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
DAG.getConstant(0, DL, VT), N0);
@@ -28119,7 +28123,6 @@ static SDValue LowerABS(SDValue Op, const X86Subtarget &Subtarget,
// ABS(vXi64 X) --> VPBLENDVPD(X, 0-X, X).
if ((VT == MVT::v2i64 || VT == MVT::v4i64) && Subtarget.hasSSE41()) {
- SDLoc DL(Op);
SDValue Src = Op.getOperand(0);
SDValue Sub =
DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Src);
@@ -28129,11 +28132,11 @@ static SDValue LowerABS(SDValue Op, const X86Subtarget &Subtarget,
if (VT.is256BitVector() && !Subtarget.hasInt256()) {
assert(VT.isInteger() &&
"Only handle AVX 256-bit vector integer operation");
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
}
if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI())
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
// Default to expand.
return SDValue();
@@ -28142,13 +28145,14 @@ static SDValue LowerABS(SDValue Op, const X86Subtarget &Subtarget,
static SDValue LowerAVG(SDValue Op, const X86Subtarget &Subtarget,
SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
+ SDLoc DL(Op);
// For AVX1 cases, split to use legal ops.
if (VT.is256BitVector() && !Subtarget.hasInt256())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, DL);
if (VT == MVT::v32i16 || VT == MVT::v64i8)
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, DL);
// Default to expand.
return SDValue();
@@ -28157,13 +28161,14 @@ static SDValue LowerAVG(SDValue Op, const X86Subtarget &Subtarget,
static SDValue LowerMINMAX(SDValue Op, const X86Subtarget &Subtarget,
SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
+ SDLoc DL(Op);
// For AVX1 cases, split to use legal ops.
if (VT.is256BitVector() && !Subtarget.hasInt256())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, DL);
if (VT == MVT::v32i16 || VT == MVT::v64i8)
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, DL);
// Default to expand.
return SDValue();
@@ -28320,15 +28325,15 @@ static SDValue LowerFMINIMUM_FMAXIMUM(SDValue Op, const X86Subtarget &Subtarget,
static SDValue LowerABD(SDValue Op, const X86Subtarget &Subtarget,
SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
+ SDLoc dl(Op);
// For AVX1 cases, split to use legal ops.
if (VT.is256BitVector() && !Subtarget.hasInt256())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, dl);
if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.useBWIRegs())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, dl);
- SDLoc dl(Op);
bool IsSigned = Op.getOpcode() == ISD::ABDS;
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
@@ -28371,10 +28376,10 @@ static SDValue LowerMUL(SDValue Op, const X86Subtarget &Subtarget,
// Decompose 256-bit ops into 128-bit ops.
if (VT.is256BitVector() && !Subtarget.hasInt256())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, dl);
if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, dl);
SDValue A = Op.getOperand(0);
SDValue B = Op.getOperand(1);
@@ -28597,10 +28602,10 @@ static SDValue LowerMULH(SDValue Op, const X86Subtarget &Subtarget,
// Decompose 256-bit ops into 128-bit ops.
if (VT.is256BitVector() && !Subtarget.hasInt256())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, dl);
if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, dl);
if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32) {
assert((VT == MVT::v4i32 && Subtarget.hasSSE2()) ||
@@ -29778,10 +29783,10 @@ static SDValue LowerShift(SDValue Op, const X86Subtarget &Subtarget,
// Decompose 256-bit shifts into 128-bit shifts.
if (VT.is256BitVector())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, dl);
if (VT == MVT::v32i16 || VT == MVT::v64i8)
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, dl);
return SDValue();
}
@@ -29858,7 +29863,7 @@ static SDValue LowerFunnelShift(SDValue Op, const X86Subtarget &Subtarget,
EltSizeInBits < 32)) {
// Pre-mask the amount modulo using the wider vector.
Op = DAG.getNode(Op.getOpcode(), DL, VT, Op0, Op1, AmtMod);
- return splitVectorOp(Op, DAG);
+ return splitVectorOp(Op, DAG, DL);
}
// Attempt to fold scalar shift as unpack(y,x) << zext(splat(z))
@@ -30020,7 +30025,7 @@ static SDValue LowerRotate(SDValue Op, const X86Subtarget &Subtarget,
// Split 256-bit integers on XOP/pre-AVX2 targets.
if (VT.is256BitVector() && (Subtarget.hasXOP() || !Subtarget.hasAVX2()))
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, DL);
// XOP has 128-bit vector variable + immediate rotates.
// +ve/-ve Amt = rotate left/right - just need to handle ISD::ROTL.
@@ -30056,7 +30061,7 @@ static SDValue LowerRotate(SDValue Op, const X86Subtarget &Subtarget,
// Split 512-bit integers on non 512-bit BWI targets.
if (VT.is512BitVector() && !Subtarget.useBWIRegs())
- return splitVectorIntBinary(Op, DAG);
+ return splitVectorIntBinary(Op, DAG, DL);
assert(
(VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8 ||
@@ -31136,11 +31141,11 @@ static SDValue LowerVectorCTPOP(SDValue Op, const SDLoc &DL,
// Decompose 256-bit ops into smaller 128-bit ops.
if (VT.is256BitVector() && !Subtarget.hasInt256())
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
// Decompose 512-bit ops into smaller 256-bit ops.
if (VT.is512BitVector() && !Subtarget.hasBWI())
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
// For element types greater than i8, do vXi8 pop counts and a bytesum.
if (VT.getScalarType() != MVT::i8) {
@@ -31264,7 +31269,7 @@ static SDValue LowerBITREVERSE_XOP(SDValue Op, SelectionDAG &DAG) {
// Decompose 256-bit ops into smaller 128-bit ops.
if (VT.is256BitVector())
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
assert(VT.is128BitVector() &&
"Only 128-bit vector bitreverse lowering supported.");
@@ -31303,11 +31308,11 @@ static SDValue LowerBITREVERSE(SDValue Op, const X86Subtarget &Subtarget,
// Split 512-bit ops without BWI so that we can still use the PSHUFB lowering.
if (VT.is512BitVector() && !Subtarget.hasBWI())
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
// Decompose 256-bit ops into smaller 128-bit ops on pre-AVX2.
if (VT.is256BitVector() && !Subtarget.hasInt256())
- return splitVectorIntUnary(Op, DAG);
+ return splitVectorIntUnary(Op, DAG, DL);
// Lower vXi16/vXi32/vXi64 as BSWAP + vXi8 BITREVERSE.
if (VT.getScalarType() != MVT::i8) {
@@ -55980,7 +55985,7 @@ static SDValue combineEXTRACT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
if (isConcatenatedNot(InVecBC.getOperand(0)) ||
isConcatenatedNot(InVecBC.getOperand(1))) {
// extract (and v4i64 X, (not (concat Y1, Y2))), n -> andnp v2i64 X(n), Y1
- SDValue Concat = splitVectorIntBinary(InVecBC, DAG);
+ SDValue Concat = splitVectorIntBinary(InVecBC, DAG, SDLoc(InVecBC));
return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT,
DAG.getBitcast(InVecVT, Concat), N->getOperand(1));
}
More information about the llvm-commits
mailing list