[llvm] r259675 - Revert r259662, which caused regressions on polly tests.

[Wei Mi via llvm-commits]

> 1) test/Isl/CodeGen/non-affine-phi-node-expansion.ll

> In this test case LLVM after your patch actually produces more complicated IR then before. Specifically, it does not inline the 'true'
> value any more. This is clearly a minor regression which will be cleaned up by later passes, but it might be of interest to you

For this case, the SCEV is of scConstant type which is cheaper than
normal value cached. Although in the cached value set in ExprValueMap,
there is a Constant type value (which is const true), but it is not
easy to find it out -- the cached Value set is not sorted according to
the potential cost. Existing reuse logic in SCEVExpander::expand
simply chooses the first legal element from the cached value set.

The workaround is: when the SCEV is of scConstant type, don't try the
reuse logic. simply expand it.

--- lib/Analysis/ScalarEvolutionExpander.cpp 2016-02-03 15:29:14.715431108 -0800
+++ lib/Analysis/ScalarEvolutionExpander.2.cpp 2016-02-03
15:28:57.075244422 -0800
@@ -1649,7 +1649,8 @@
   // type SCEV, it will be expanded literally, to prevent LSR's
transformed SCEV
   // from being reverted.
   if (!(LSRMode && SE.containsAddRecurrence(S))) {
-    if (Set) {
+    // If S is scConstant, it may be worse to reuse an existing Value.
+    if (S->getSCEVType() == scConstant && Set) {
       // Choose a Value from the set which dominates the insertPt.
       for (auto const &Ent : *Set) {
         if (Ent && isa<Instruction>(Ent) && S->getType() == Ent->getType() &&

Thanks,
Wei.