[PATCH] Revert r191049/r191059

[Robert Lougher]

Hi,

This patch reverts r191049/r191059 as discussed here:

http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-April/072104.html

An additional test needed to be xfailed as it relies on the
optimization.  I have CC-ed Tim Northover who added it.

OK to commit?

Thanks,
Rob.

--
Robert Lougher
SN Systems - Sony Computer Entertainment Group
-------------- next part --------------
Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp	(revision 206297)
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp	(working copy)
@@ -3555,27 +3555,6 @@
                        HasPos ? Pos : Neg).getNode();
   }
 
-  // fold (or (shl (*ext x), (*ext y)),
-  //          (srl (*ext x), (*ext (sub 32, y)))) ->
-  //   (*ext (rotl x, y)) or (*ext (rotr x, (sub 32, y)))
-  //
-  // fold (or (shl (*ext x), (*ext (sub 32, y))),
-  //          (srl (*ext x), (*ext y))) ->
-  //   (*ext (rotr x, y)) or (*ext (rotl x, (sub 32, y)))
-  if (Shifted.getOpcode() == ISD::ZERO_EXTEND ||
-      Shifted.getOpcode() == ISD::ANY_EXTEND) {
-    SDValue InnerShifted = Shifted.getOperand(0);
-    EVT InnerVT = InnerShifted.getValueType();
-    bool HasPosInner = TLI.isOperationLegalOrCustom(PosOpcode, InnerVT);
-    if (HasPosInner || TLI.isOperationLegalOrCustom(NegOpcode, InnerVT)) {
-      if (matchRotateSub(InnerPos, InnerNeg, InnerVT.getSizeInBits())) {
-        SDValue V = DAG.getNode(HasPosInner ? PosOpcode : NegOpcode, DL,
-                                InnerVT, InnerShifted, HasPosInner ? Pos : Neg);
-        return DAG.getNode(Shifted.getOpcode(), DL, VT, V).getNode();
-      }
-    }
-  }
-
   return nullptr;
 }
 
Index: test/CodeGen/ARM64/regress-interphase-shift.ll
===================================================================
--- test/CodeGen/ARM64/regress-interphase-shift.ll	(revision 206297)
+++ test/CodeGen/ARM64/regress-interphase-shift.ll	(working copy)
@@ -4,6 +4,10 @@
 ; on the phase of legalization, which led to the creation of an unexpected and
 ; unselectable "rotr" node: (i32 (rotr i32, i64)).
 
+; FIXME: This test is xfailed because it relies on an optimization that has
+; been reverted (see PR17975).
+; XFAIL: *
+
 define void @foo(i64* nocapture %d) {
 ; CHECK-LABEL: foo:
 ; CHECK: rorv
Index: test/CodeGen/X86/rotate3.ll
===================================================================
--- test/CodeGen/X86/rotate3.ll	(revision 206297)
+++ test/CodeGen/X86/rotate3.ll	(working copy)
@@ -1,76 +0,0 @@
-; Check that (or (shl x, y), (srl x, (sub 32, y))) is folded into (rotl x, y)
-; and (or (shl x, (sub 32, y)), (srl x, r)) into (rotr x, y) even if the
-; argument is zero extended. Fix for PR16726.
-
-; RUN: llc < %s -march=x86-64 -mcpu=corei7 | FileCheck %s
-
-define zeroext i8 @rolbyte(i32 %nBits_arg, i8 %x_arg) nounwind readnone {
-entry:
-  %tmp1 = zext i8 %x_arg to i32
-  %tmp3 = shl i32 %tmp1, %nBits_arg
-  %tmp8 = sub i32 8, %nBits_arg
-  %tmp10 = lshr i32 %tmp1, %tmp8
-  %tmp11 = or i32 %tmp3, %tmp10
-  %tmp12 = trunc i32 %tmp11 to i8
-  ret i8 %tmp12
-}
-; CHECK:    rolb %cl, %{{[a-z0-9]+}}
-
-
-define zeroext i8 @rorbyte(i32 %nBits_arg, i8 %x_arg) nounwind readnone {
-entry:
-  %tmp1 = zext i8 %x_arg to i32
-  %tmp3 = lshr i32 %tmp1, %nBits_arg
-  %tmp8 = sub i32 8, %nBits_arg
-  %tmp10 = shl i32 %tmp1, %tmp8
-  %tmp11 = or i32 %tmp3, %tmp10
-  %tmp12 = trunc i32 %tmp11 to i8
-  ret i8 %tmp12
-}
-; CHECK:    rorb %cl, %{{[a-z0-9]+}}
-
-define zeroext i16 @rolword(i32 %nBits_arg, i16 %x_arg) nounwind readnone {
-entry:
-  %tmp1 = zext i16 %x_arg to i32
-  %tmp3 = shl i32 %tmp1, %nBits_arg
-  %tmp8 = sub i32 16, %nBits_arg
-  %tmp10 = lshr i32 %tmp1, %tmp8
-  %tmp11 = or i32 %tmp3, %tmp10
-  %tmp12 = trunc i32 %tmp11 to i16
-  ret i16 %tmp12
-}
-; CHECK:    rolw %cl, %{{[a-z0-9]+}}
-
-define zeroext i16 @rorword(i32 %nBits_arg, i16 %x_arg) nounwind readnone {
-entry:
-  %tmp1 = zext i16 %x_arg to i32
-  %tmp3 = lshr i32 %tmp1, %nBits_arg
-  %tmp8 = sub i32 16, %nBits_arg
-  %tmp10 = shl i32 %tmp1, %tmp8
-  %tmp11 = or i32 %tmp3, %tmp10
-  %tmp12 = trunc i32 %tmp11 to i16
-  ret i16 %tmp12
-}
-; CHECK:    rorw %cl, %{{[a-z0-9]+}}
-
-define i64 @roldword(i64 %nBits_arg, i32 %x_arg) nounwind readnone {
-entry:
-  %tmp1 = zext i32 %x_arg to i64
-  %tmp3 = shl i64 %tmp1, %nBits_arg
-  %tmp8 = sub i64 32, %nBits_arg
-  %tmp10 = lshr i64 %tmp1, %tmp8
-  %tmp11 = or i64 %tmp3, %tmp10
-  ret i64 %tmp11
-}
-; CHECK:    roll %cl, %{{[a-z0-9]+}}
-
-define zeroext i64 @rordword(i64 %nBits_arg, i32 %x_arg) nounwind readnone {
-entry:
-  %tmp1 = zext i32 %x_arg to i64
-  %tmp3 = lshr i64 %tmp1, %nBits_arg
-  %tmp8 = sub i64 32, %nBits_arg
-  %tmp10 = shl i64 %tmp1, %tmp8
-  %tmp11 = or i64 %tmp3, %tmp10
-  ret i64 %tmp11
-}
-; CHECK:    rorl %cl, %{{[a-z0-9]+}}