return true;
}
+/// \brief Test whether there are elements crossing 128-bit lanes in this
+/// shuffle mask.
+///
+/// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
+/// and we routinely test for these.
+static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
+ int LaneSize = 128 / VT.getScalarSizeInBits();
+ int Size = Mask.size();
+ for (int i = 0; i < Size; ++i)
+ if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
+ return true;
+ return false;
+}
+
+/// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
+///
+/// This checks a shuffle mask to see if it is performing the same
+/// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
+/// that it is also not lane-crossing. It may however involve a blend from the
+/// same lane of a second vector.
+///
+/// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
+/// non-trivial to compute in the face of undef lanes. The representation is
+/// *not* suitable for use with existing 128-bit shuffles as it will contain
+/// entries from both V1 and V2 inputs to the wider mask.
+static bool
+is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
+ SmallVectorImpl<int> &RepeatedMask) {
+ int LaneSize = 128 / VT.getScalarSizeInBits();
+ RepeatedMask.resize(LaneSize, -1);
+ int Size = Mask.size();
+ for (int i = 0; i < Size; ++i) {
+ if (Mask[i] < 0)
+ continue;
+ if ((Mask[i] % Size) / LaneSize != i / LaneSize)
+ // This entry crosses lanes, so there is no way to model this shuffle.
+ return false;
+
+ // Ok, handle the in-lane shuffles by detecting if and when they repeat.
+ if (RepeatedMask[i % LaneSize] == -1)
+ // This is the first non-undef entry in this slot of a 128-bit lane.
+ RepeatedMask[i % LaneSize] =
+ Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
+ else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
+ // Found a mismatch with the repeated mask.
+ return false;
+ }
+ return true;
+}
+
// Hide this symbol with an anonymous namespace instead of 'static' so that MSVC
// 2013 will allow us to use it as a non-type template parameter.
namespace {
DAG.getConstant(BlendMask, MVT::i8)));
}
+ case MVT::v16i16: {
+ assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
+ SmallVector<int, 8> RepeatedMask;
+ if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
+ // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
+ assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
+ BlendMask = 0;
+ for (int i = 0; i < 8; ++i)
+ if (RepeatedMask[i] >= 16)
+ BlendMask |= 1u << i;
+ return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
+ DAG.getConstant(BlendMask, MVT::i8));
+ }
+
+ // Fall back to a fully general variable byte blend.
+ SDValue PBLENDVMask[32];
+ // Scale the blend by the number of bytes per element.
+ int Scale = VT.getScalarSizeInBits() / 8;
+ assert(Mask.size() * Scale == 32 && "Not a 256-bit vector!");
+ for (int i = 0, Size = Mask.size(); i < Size; ++i)
+ for (int j = 0; j < Scale; ++j)
+ PBLENDVMask[Scale * i + j] =
+ Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
+ : DAG.getConstant(Mask[i] < Size ? 0 : 0x80, MVT::i8);
+
+ V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1);
+ V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V2);
+ return DAG.getNode(ISD::BITCAST, DL, VT, DAG.getNode(
+ X86ISD::BLENDV, DL, MVT::v32i8, V1, V2,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PBLENDVMask)));
+ }
+
default:
llvm_unreachable("Not a supported integer vector type!");
}
}
}
-/// \brief Test whether there are elements crossing 128-bit lanes in this
-/// shuffle mask.
-///
-/// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
-/// and we routinely test for these.
-static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
- int LaneSize = 128 / VT.getScalarSizeInBits();
- int Size = Mask.size();
- for (int i = 0; i < Size; ++i)
- if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
- return true;
- return false;
-}
-
-/// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
-///
-/// This checks a shuffle mask to see if it is performing the same
-/// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
-/// that it is also not lane-crossing. It may however involve a blend from the
-/// same lane of a second vector.
-///
-/// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
-/// non-trivial to compute in the face of undef lanes. The representation is
-/// *not* suitable for use with existing 128-bit shuffles as it will contain
-/// entries from both V1 and V2 inputs to the wider mask.
-static bool
-is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
- SmallVectorImpl<int> &RepeatedMask) {
- int LaneSize = 128 / VT.getScalarSizeInBits();
- RepeatedMask.resize(LaneSize, -1);
- int Size = Mask.size();
- for (int i = 0; i < Size; ++i) {
- if (Mask[i] < 0)
- continue;
- if ((Mask[i] % Size) / LaneSize != i / LaneSize)
- // This entry crosses lanes, so there is no way to model this shuffle.
- return false;
-
- // Ok, handle the in-lane shuffles by detecting if and when they repeat.
- if (RepeatedMask[i % LaneSize] == -1)
- // This is the first non-undef entry in this slot of a 128-bit lane.
- RepeatedMask[i % LaneSize] =
- Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
- else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
- // Found a mismatch with the repeated mask.
- return false;
- }
- return true;
-}
-
/// \brief Generic routine to split a 256-bit vector shuffle into 128-bit
/// shuffles.
///
assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
- // FIXME: Actually implement this using AVX2!!!
- (void)Mask;
- return splitAndLower256BitVectorShuffle(Op, V1, V2, Subtarget, DAG);
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
+
+ // If the shuffle mask is repeated in each 128-bit lane we can use more
+ // efficient instructions that mirror the shuffles across the two 128-bit
+ // lanes.
+ SmallVector<int, 4> RepeatedMask;
+ if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
+ assert(RepeatedMask.size() == 8 && "Unexpected repeated mask size!");
+ // FIXME: It might be worth it to call into the (terribly complex) v8i16
+ // lowering here.
+
+ // Use dedicated unpack instructions for masks that match their pattern.
+ //
+ if (isShuffleEquivalent(Mask,
+ // First 128-bit lane:
+ 0, 16, 1, 17, 2, 18, 3, 19,
+ // Second 128-bit lane:
+ 8, 24, 9, 25, 10, 26, 11, 27))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
+ if (isShuffleEquivalent(Mask,
+ // First 128-bit lane:
+ 4, 20, 5, 21, 6, 22, 7, 23,
+ // Second 128-bit lane:
+ 12, 28, 13, 29, 14, 30, 15, 31))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
+ }
+
+ // There are no generalized cross-lane shuffle operations available on i16
+ // element types.
+ // FIXME: We should teach the "split and lower" path to do something more
+ // clever, or do it ourselves here. The optimal lowering of cross-lane
+ // shuffles I am aware of is to swap the lanes into a copy, shuffle both the
+ // original and the copy, and then blend to pick up the cross-lane elements.
+ // This is four instructions with a tree height of three which is better than
+ // the worst case for a gather-cross-scatter approach such as used in SSE2
+ // v8i16 lowering (where we don't have blends). While for cross-lane blends it
+ // results in a blend tree, blends are very cheap in AVX2 and newer chips. We
+ // might also want to special case situations where we can always do a single
+ // VPERMD to produce a non-lane-crossing shuffle.
+ if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
+ return splitAndLower256BitVectorShuffle(Op, V1, V2, Subtarget, DAG);
+
+ if (isSingleInputShuffleMask(Mask)) {
+ SDValue PSHUFBMask[32];
+ for (int i = 0; i < 16; ++i) {
+ if (Mask[i] == -1) {
+ PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
+ continue;
+ }
+
+ int M = i < 8 ? Mask[i] : Mask[i] - 8;
+ assert(M >= 0 && M < 8 && "Invalid single-input mask!");
+ PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
+ PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
+ }
+ return DAG.getNode(
+ ISD::BITCAST, DL, MVT::v16i16,
+ DAG.getNode(
+ X86ISD::PSHUFB, DL, MVT::v32i8,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
+ }
+
+ // Otherwise fall back on generic blend lowering.
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i16, V1, V2,
+ Mask, DAG);
}
/// \brief Handle lowering of 32-lane 8-bit integer shuffles.
projects / oota-llvm.git / blobdiff