// Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
// source words for the shuffle, to aid later transformations.
bool AllWordsInNewV = true;
+ bool InOrder[2] = { true, true };
for (unsigned i = 0; i != 8; ++i) {
int idx = MaskVals[i];
+ if (idx != (int)i)
+ InOrder[i/4] = false;
if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
continue;
AllWordsInNewV = false;
// If we've eliminated the use of V2, and the new mask is a pshuflw or
// pshufhw, that's as cheap as it gets. Return the new shuffle.
- if (pshufhw || pshuflw) {
+ if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
MaskV.clear();
for (unsigned i = 0; i != 8; ++i)
MaskV.push_back((MaskVals[i] < 0) ? DAG.getUNDEF(MVT::i16)
--- /dev/null
+; RUN: llvm-as < %s | llc -march=x86 -mattr=sse41 -o %t -f
+; RUN: grep pshufb %t | count 1
+
+
+define <8 x i16> @shuf6(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
+entry:
+ %tmp9 = shufflevector <8 x i16> %T0, <8 x i16> %T1, <8 x i32> < i32 3, i32 2, i32 0, i32 2, i32 1, i32 5, i32 6 , i32 undef >
+ ret <8 x i16> %tmp9
+}