[PATCH] Fix optimisations of SELECT_CC which assumed result is boolean

[Tim Northover]

Hi Oliver,

Thanks for updating the patch. Nearly there, I think, but there's one improvement we can make:

================
Comment at: lib/CodeGen/SelectionDAG/DAGCombiner.cpp:3833
@@ -3828,4 +3832,3 @@
   // fold !(x cc y) -> (x !cc y)
-  if (N1C && N1C->getAPIntValue().isAllOnesValue() &&
-      isSetCCEquivalent(N0, LHS, RHS, CC)) {
+  if (N1C && TLI.isConstTrueVal(N1C) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
     bool isInt = LHS.getValueType().isInteger();
----------------
isConstTrueVal is actually a bit more general than this (or it would be if it didn't have a bug itself), handling BUILD_VECTOR as well as ConstantSDNode. So you can probably drop the N1C check and use N1 directly.

E.g.
    define <2 x i64> @foo(<2 x i64> %lhs, <2 x i64> %rhs) {
      %tst = icmp eq <2 x i64> %lhs, %rhs
      %ntst = xor <2 x i1> %tst, <i1 1 , i1 undef>
      %btst = sext <2 x i1> %ntst to <2 x i64>
      ret <2 x i64> %btst
    }

(The undef should be a 1, I think, but that bug in isConstTrueVal (& isConstFalseVal) means the optimisation triggers with undef instead).

http://reviews.llvm.org/D6249