Handle (shl (anyext (shr ...))) in SimpilfyDemandedBits

This is really an extension of the current (shl (shr ...)) -> shl optimization.
The main difference is that certain upper bits must also not be demanded.

The motivating examples are the first two in the testcase, which occur
in llvmpipe output.

llvm-svn: 192783

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -635,6 +635,31 @@ bool TargetLowering::SimplifyDemandedBits(SDValue Op,
                           TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
                                           NarrowShl));
         }
+        // Repeat the SHL optimization above in cases where an extension
+        // intervenes: (shl (anyext (shr x, c1)), c2) to
+        // (shl (anyext x), c2-c1).  This requires that the bottom c1 bits
+        // aren't demanded (as above) and that the shifted upper c1 bits of
+        // x aren't demanded.
+        if (InOp.hasOneUse() &&
+            InnerOp.getOpcode() == ISD::SRL &&
+            InnerOp.hasOneUse() &&
+            isa<ConstantSDNode>(InnerOp.getOperand(1))) {
+          uint64_t InnerShAmt = cast<ConstantSDNode>(InnerOp.getOperand(1))
+            ->getZExtValue();
+          if (InnerShAmt < ShAmt &&
+              InnerShAmt < InnerBits &&
+              NewMask.lshr(InnerBits - InnerShAmt + ShAmt) == 0 &&
+              NewMask.trunc(ShAmt) == 0) {
+            SDValue NewSA =
+              TLO.DAG.getConstant(ShAmt - InnerShAmt,
+                                  Op.getOperand(1).getValueType());
+            EVT VT = Op.getValueType();
+            SDValue NewExt = TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT,
+                                             InnerOp.getOperand(0));
+            return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT,
+                                                     NewExt, NewSA));
+          }
+        }
       }
 
       KnownZero <<= SA->getZExtValue();

llvm/test/CodeGen/SystemZ/shift-10.ll

@@ -0,0 +1,67 @@
+; Test compound shifts.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
+
+; Test a shift right followed by a sign extension.  This can use two shifts.
+define i64 @f1(i32 %a) {
+; CHECK-LABEL: f1:
+; CHECK: sllg [[REG:%r[0-5]]], %r2, 62
+; CHECK: srag %r2, [[REG]], 63
+; CHECK: br %r14
+  %shr = lshr i32 %a, 1
+  %trunc = trunc i32 %shr to i1
+  %ext = sext i1 %trunc to i64
+  ret i64 %ext
+}
+
+; ...and again with the highest shift count.
+define i64 @f2(i32 %a) {
+; CHECK-LABEL: f2:
+; CHECK: sllg [[REG:%r[0-5]]], %r2, 32
+; CHECK: srag %r2, [[REG]], 63
+; CHECK: br %r14
+  %shr = lshr i32 %a, 31
+  %trunc = trunc i32 %shr to i1
+  %ext = sext i1 %trunc to i64
+  ret i64 %ext
+}
+
+; Test a left shift that of an extended right shift in a case where folding
+; is possible.
+define i64 @f3(i32 %a) {
+; CHECK-LABEL: f3:
+; CHECK: risbg %r2, %r2, 27, 181, 9
+; CHECK: br %r14
+  %shr = lshr i32 %a, 1
+  %ext = zext i32 %shr to i64
+  %shl = shl i64 %ext, 10
+  %and = and i64 %shl, 137438952960
+  ret i64 %and
+}
+
+; ...and again with a larger right shift.
+define i64 @f4(i32 %a) {
+; CHECK-LABEL: f4:
+; CHECK: risbg %r2, %r2, 30, 158, 3
+; CHECK: br %r14
+  %shr = lshr i32 %a, 30
+  %ext = sext i32 %shr to i64
+  %shl = shl i64 %ext, 33
+  %and = and i64 %shl, 8589934592
+  ret i64 %and
+}
+
+; Repeat the previous test in a case where all bits outside the
+; bottom 3 matter.  FIXME: can still use RISBG here.
+define i64 @f5(i32 %a) {
+; CHECK-LABEL: f5:
+; CHECK: srl %r2, 30
+; CHECK: sllg %r2, %r2, 33
+; CHECK: lhi %r2, 7
+; CHECK: br %r14
+  %shr = lshr i32 %a, 30
+  %ext = sext i32 %shr to i64
+  %shl = shl i64 %ext, 33
+  %or = or i64 %shl, 7
+  ret i64 %or
+}