return DAG.getNode(X86ISD::VTRUNC, DL, VT, Src);
}
+/// Check whether a compaction lowering can be done by dropping even
+/// elements and compute how many times even elements must be dropped.
+///
+/// This handles shuffles which take every Nth element where N is a power of
+/// two. Example shuffle masks:
+///
+/// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
+/// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
+/// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
+/// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
+/// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
+/// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
+///
+/// Any of these lanes can of course be undef.
+///
+/// This routine only supports N <= 3.
+/// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
+/// for larger N.
+///
+/// \returns N above, or the number of times even elements must be dropped if
+/// there is such a number. Otherwise returns zero.
+static int canLowerByDroppingEvenElements(ArrayRef<int> Mask,
+ bool IsSingleInput) {
+ // The modulus for the shuffle vector entries is based on whether this is
+ // a single input or not.
+ int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
+ assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
+ "We should only be called with masks with a power-of-2 size!");
+
+ uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
+
+ // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
+ // and 2^3 simultaneously. This is because we may have ambiguity with
+ // partially undef inputs.
+ bool ViableForN[3] = {true, true, true};
+
+ for (int i = 0, e = Mask.size(); i < e; ++i) {
+ // Ignore undef lanes, we'll optimistically collapse them to the pattern we
+ // want.
+ if (Mask[i] < 0)
+ continue;
+
+ bool IsAnyViable = false;
+ for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
+ if (ViableForN[j]) {
+ uint64_t N = j + 1;
+
+ // The shuffle mask must be equal to (i * 2^N) % M.
+ if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
+ IsAnyViable = true;
+ else
+ ViableForN[j] = false;
+ }
+ // Early exit if we exhaust the possible powers of two.
+ if (!IsAnyViable)
+ break;
+ }
+
+ for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
+ if (ViableForN[j])
+ return j + 1;
+
+ // Return 0 as there is no viable power of two.
+ return 0;
+}
+
// X86 has dedicated pack instructions that can handle specific truncation
// operations: PACKSS and PACKUS.
static bool matchShuffleWithPACK(MVT VT, MVT &SrcVT, SDValue &V1, SDValue &V2,
Mask, Subtarget, DAG);
}
-/// Check whether a compaction lowering can be done by dropping even
-/// elements and compute how many times even elements must be dropped.
-///
-/// This handles shuffles which take every Nth element where N is a power of
-/// two. Example shuffle masks:
-///
-/// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
-/// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
-/// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
-/// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
-/// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
-/// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
-///
-/// Any of these lanes can of course be undef.
-///
-/// This routine only supports N <= 3.
-/// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
-/// for larger N.
-///
-/// \returns N above, or the number of times even elements must be dropped if
-/// there is such a number. Otherwise returns zero.
-static int canLowerByDroppingEvenElements(ArrayRef<int> Mask,
- bool IsSingleInput) {
- // The modulus for the shuffle vector entries is based on whether this is
- // a single input or not.
- int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
- assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
- "We should only be called with masks with a power-of-2 size!");
-
- uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
-
- // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
- // and 2^3 simultaneously. This is because we may have ambiguity with
- // partially undef inputs.
- bool ViableForN[3] = {true, true, true};
-
- for (int i = 0, e = Mask.size(); i < e; ++i) {
- // Ignore undef lanes, we'll optimistically collapse them to the pattern we
- // want.
- if (Mask[i] < 0)
- continue;
-
- bool IsAnyViable = false;
- for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
- if (ViableForN[j]) {
- uint64_t N = j + 1;
-
- // The shuffle mask must be equal to (i * 2^N) % M.
- if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
- IsAnyViable = true;
- else
- ViableForN[j] = false;
- }
- // Early exit if we exhaust the possible powers of two.
- if (!IsAnyViable)
- break;
- }
-
- for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
- if (ViableForN[j])
- return j + 1;
-
- // Return 0 as there is no viable power of two.
- return 0;
-}
-
static SDValue lowerShuffleWithPERMV(const SDLoc &DL, MVT VT,
ArrayRef<int> Mask, SDValue V1,
SDValue V2, SelectionDAG &DAG) {