[X86][SSE] Prefer trunc(movd(x)) to pextrb(x,0)

[RKSimon]

Summary

If we're extracting the 0'th index of a v16i8 vector we're better off using MOVD than PEXTRB, unless we're storing the value or we require the implicit zero extension of PEXTRB.

The biggest perf diff is on SLM targets where MOVD (uops=1, lat=3 tp=1) is notably faster than PEXTRB (uops=2, lat=5, tp=4).

This matches what we already do for PEXTRW.

[lebedev.ri]

Test changes seem to be omitted from the diff?

[RKSimon]

In D76138#1921775, @lebedev.ri wrote:

Test changes seem to be omitted from the diff?

I loathe git at times :-)

[RKSimon]

rebase, with tests this time

[craig.topper]

LGTM

llvm/lib/Target/X86/X86ISelLowering.cpp

static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
   SDValue Idx = Op.getOperand(1);
   assert(isa<ConstantSDNode>(Idx) && "Constant index expected");
   SDLoc dl(Op);
 
   if (!Vec.getSimpleValueType().is128BitVector())
     return SDValue();
 
   if (VT.getSizeInBits() == 8) {
+    // If IdxVal is 0, it's cheaper to do a move instead of a pextrb, unless
+    // we're going to zero extend the register or fold the store.
+    if (llvm::isNullConstant(Idx) && !MayFoldIntoZeroExtend(Op) &&
+        !MayFoldIntoStore(Op))
+      return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
+                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
+                                     DAG.getBitcast(MVT::v4i32, Vec), Idx));
+
     SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32, Vec, Idx);
     return DAG.getNode(ISD::TRUNCATE, dl, VT, Extract);
   }
 
   if (VT == MVT::f32) {
     // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
     // the result back to FR32 register. It's only worth matching if the
     // result has a single use which is a store or a bitcast to i32.  And in