Remove non-necessary canonicalization from the DAG

[mehdi_amini]

Summary

r236031 introduced a transformation to reduce the dependency chain
for a sequence of binop instructions.

(fadd N0: (fadd N00: (fadd z, w), N01: y), N1: x) ->
(fadd N00: (fadd z, w), (fadd N1: x, N01: y))

However as part of the implementation, a canonicalization was
introduced that commute chain of fadd to be on the left side:

(fadd (otherop, fadd)) -> (fadd (fadd, otherop))

This canonicalization does not seem required to perform the original
targetted transformation, swapping the operand locally is enough to
achieve this goal as far as I understand. Moreover the DAG is already
slow and the way it was done required a round trip through the
combiner queue and the creation of a temporary node.

[spatel]

LGTM. Thanks!

[spatel]

Closing; I used the swap suggestion in the new proposal of this patch in D9780.

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

static SDValue ReassociateBinops(SDNode *N, SelectionDAG &DAG) {
   assert(N->getNumOperands() == 2 && "Invalid node for binop reassociation");
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
   EVT VT = N->getValueType(0);
   SDLoc DL(N);
   unsigned Opcode = N->getOpcode();
 
   // Canonicalize chains of this operation to LHS to allow the following fold.
-  if (N0.getOpcode() != Opcode && N1.getOpcode() == Opcode)
-    return DAG.getNode(Opcode, DL, VT, N1, N0);
-  
+  if (N1.getOpcode() == Opcode && N1.hasOneUse() && N0.getOpcode() != Opcode)
+    std::swap(N1, N0);
   // Convert a chain of 3 dependent operations into 2 independent operations
   // and 1 dependent operation:
   //  (op N0: (op N00: (op z, w), N01: y), N1: x) ->
   //  (op N00: (op z, w), (op N1: x, N01: y))
   if (N0.getOpcode() == Opcode && N0.hasOneUse() && N1.getOpcode() != Opcode) {
     SDValue N00 = N0.getOperand(0);
     if (N00.getOpcode() == Opcode) {
       SDValue N01 = N0.getOperand(1);