[Thumb] Teach ISel how to lower compares of AND bitmasks efficiently

This is essentially a recommit of r285893, but with a correctness fix. The
problem of the original commit was that this:

bic r5, r7, #31
cbz r5, .LBB2_10

got rewritten into:

lsrs  r5, r7, #5
beq .LBB2_10

The result in destination register r5 is not the same and this is incorrect
when r5 is not dead. So this fix includes checking the uses of the AND
destination register. And also, compared to the original commit, some regression
tests didn't need changing anymore because of this extra check.

For completeness, this was the original commit message:

For the common pattern (CMPZ (AND x, #bitmask), #0), we can do some more
efficient instruction selection if the bitmask is one consecutive sequence of
set bits (32 - clz(bm) - ctz(bm) == popcount(bm)).

1) If the bitmask touches the LSB, then we can remove all the upper bits and
set the flags by doing one LSLS.
2) If the bitmask touches the MSB, then we can remove all the lower bits and
set the flags with one LSRS.
3) If the bitmask has popcount == 1 (only one set bit), we can shift that bit
into the sign bit with one LSLS and change the condition query from NE/EQ to
MI/PL (we could also implement this by shifting into the carry bit and
branching on BCC/BCS).
4) Otherwise, we can emit a sequence of LSLS+LSRS to remove the upper and lower
zero bits of the mask.

1-3 require only one 16-bit instruction and can elide the CMP. 4 requires two
16-bit instructions but can elide the CMP and doesn't require materializing a
complex immediate, so is also a win.

Differential Revision: https://reviews.llvm.org/D27761

llvm-svn: 289794

2 files changed, 141 insertions, 4 deletions

diff --git a/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp b/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 9c8229c275d..70a3246e34f 100644
--- a/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -2544,7 +2544,11 @@ bool ARMBaseInstrInfo::optimizeCompareInstr(
   case ARM::EORrr:
   case ARM::EORri:
   case ARM::t2EORrr:
-  case ARM::t2EORri: {
+  case ARM::t2EORri:
+  case ARM::t2LSRri:
+  case ARM::t2LSRrr:
+  case ARM::t2LSLri:
+  case ARM::t2LSLrr: {
     // Scan forward for the use of CPSR
     // When checking against MI: if it's a conditional code that requires
     // checking of the V bit or C bit, then this is not safe to do.
diff --git a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
index e257f2faf83..c3e9591d5c7 100644
--- a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -245,6 +245,7 @@ private:
   bool tryInlineAsm(SDNode *N);
 
   void SelectConcatVector(SDNode *N);
+  void SelectCMPZ(SDNode *N, bool &SwitchEQNEToPLMI);
 
   bool trySMLAWSMULW(SDNode *N);
 
@@ -2730,6 +2731,87 @@ void ARMDAGToDAGISel::SelectConcatVector(SDNode *N) {
   ReplaceNode(N, createDRegPairNode(VT, N->getOperand(0), N->getOperand(1)));
 }
 
+static Optional<std::pair<unsigned, unsigned>>
+getContiguousRangeOfSetBits(const APInt &A) {
+  unsigned FirstOne = A.getBitWidth() - A.countLeadingZeros() - 1;
+  unsigned LastOne = A.countTrailingZeros();
+  if (A.countPopulation() != (FirstOne - LastOne + 1))
+    return Optional<std::pair<unsigned,unsigned>>();
+  return std::make_pair(FirstOne, LastOne);
+}
+
+void ARMDAGToDAGISel::SelectCMPZ(SDNode *N, bool &SwitchEQNEToPLMI) {
+  assert(N->getOpcode() == ARMISD::CMPZ);
+  SwitchEQNEToPLMI = false;
+
+  if (!Subtarget->isThumb())
+    // FIXME: Work out whether it is profitable to do this in A32 mode - LSL and
+    // LSR don't exist as standalone instructions - they need the barrel shifter.
+    return;
+
+  // select (cmpz (and X, C), #0) -> (LSLS X) or (LSRS X) or (LSRS (LSLS X))
+  SDValue And = N->getOperand(0);
+  if (!And->hasOneUse())
+    return;
+
+  SDValue Zero = N->getOperand(1);
+  if (!isa<ConstantSDNode>(Zero) || !cast<ConstantSDNode>(Zero)->isNullValue() ||
+      And->getOpcode() != ISD::AND)
+    return;
+  SDValue X = And.getOperand(0);
+  auto C = dyn_cast<ConstantSDNode>(And.getOperand(1));
+
+  if (!C || !X->hasOneUse())
+    return;
+  auto Range = getContiguousRangeOfSetBits(C->getAPIntValue());
+  if (!Range)
+    return;
+
+  // There are several ways to lower this:
+  SDNode *NewN;
+  SDLoc dl(N);
+
+  auto EmitShift = [&](unsigned Opc, SDValue Src, unsigned Imm) -> SDNode* {
+    if (Subtarget->isThumb2()) {
+      Opc = (Opc == ARM::tLSLri) ? ARM::t2LSLri : ARM::t2LSRri;
+      SDValue Ops[] = { Src, CurDAG->getTargetConstant(Imm, dl, MVT::i32),
+                        getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32),
+                        CurDAG->getRegister(0, MVT::i32) };
+      return CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
+    } else {
+      SDValue Ops[] = {CurDAG->getRegister(ARM::CPSR, MVT::i32), Src,
+                       CurDAG->getTargetConstant(Imm, dl, MVT::i32),
+                       getAL(CurDAG, dl), CurDAG->getRegister(0, MVT::i32)};
+      return CurDAG->getMachineNode(Opc, dl, MVT::i32, Ops);
+    }
+  };
+  
+  if (Range->second == 0) {
+    //  1. Mask includes the LSB -> Simply shift the top N bits off
+    NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
+    ReplaceNode(And.getNode(), NewN);
+  } else if (Range->first == 31) {
+    //  2. Mask includes the MSB -> Simply shift the bottom N bits off
+    NewN = EmitShift(ARM::tLSRri, X, Range->second);
+    ReplaceNode(And.getNode(), NewN);
+  } else if (Range->first == Range->second) {
+    //  3. Only one bit is set. We can shift this into the sign bit and use a
+    //     PL/MI comparison.
+    NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
+    ReplaceNode(And.getNode(), NewN);
+
+    SwitchEQNEToPLMI = true;
+  } else if (!Subtarget->hasV6T2Ops()) {
+    //  4. Do a double shift to clear bottom and top bits, but only in
+    //     thumb-1 mode as in thumb-2 we can use UBFX.
+    NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first);
+    NewN = EmitShift(ARM::tLSRri, SDValue(NewN, 0),
+                     Range->second + (31 - Range->first));
+    ReplaceNode(And.getNode(), NewN);
+  }
+
+}
+
 void ARMDAGToDAGISel::Select(SDNode *N) {
   SDLoc dl(N);
 
@@ -2957,6 +3039,7 @@ void ARMDAGToDAGISel::Select(SDNode *N) {
         return;
       }
     }
+
     break;
   }
   case ARMISD::VMOVRRD:
@@ -3148,9 +3231,27 @@ void ARMDAGToDAGISel::Select(SDNode *N) {
     assert(N2.getOpcode() == ISD::Constant);
     assert(N3.getOpcode() == ISD::Register);
 
-    SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
-                               cast<ConstantSDNode>(N2)->getZExtValue()), dl,
-                               MVT::i32);
+    unsigned CC = (unsigned) cast<ConstantSDNode>(N2)->getZExtValue();
+    
+    if (InFlag.getOpcode() == ARMISD::CMPZ) {
+      bool SwitchEQNEToPLMI;
+      SelectCMPZ(InFlag.getNode(), SwitchEQNEToPLMI);
+      InFlag = N->getOperand(4);
+
+      if (SwitchEQNEToPLMI) {
+        switch ((ARMCC::CondCodes)CC) {
+        default: llvm_unreachable("CMPZ must be either NE or EQ!");
+        case ARMCC::NE:
+          CC = (unsigned)ARMCC::MI;
+          break;
+        case ARMCC::EQ:
+          CC = (unsigned)ARMCC::PL;
+          break;
+        }
+      }
+    }
+
+    SDValue Tmp2 = CurDAG->getTargetConstant(CC, dl, MVT::i32);
     SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
     SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
                                              MVT::Glue, Ops);
@@ -3205,6 +3306,38 @@ void ARMDAGToDAGISel::Select(SDNode *N) {
     // Other cases are autogenerated.
     break;
   }
+
+  case ARMISD::CMOV: {
+    SDValue InFlag = N->getOperand(4);
+
+    if (InFlag.getOpcode() == ARMISD::CMPZ) {
+      bool SwitchEQNEToPLMI;
+      SelectCMPZ(InFlag.getNode(), SwitchEQNEToPLMI);
+
+      if (SwitchEQNEToPLMI) {
+        SDValue ARMcc = N->getOperand(2);
+        ARMCC::CondCodes CC =
+          (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
+
+        switch (CC) {
+        default: llvm_unreachable("CMPZ must be either NE or EQ!");
+        case ARMCC::NE:
+          CC = ARMCC::MI;
+          break;
+        case ARMCC::EQ:
+          CC = ARMCC::PL;
+          break;
+        }
+        SDValue NewARMcc = CurDAG->getConstant((unsigned)CC, dl, MVT::i32);
+        SDValue Ops[] = {N->getOperand(0), N->getOperand(1), NewARMcc,
+                         N->getOperand(3), N->getOperand(4)};
+        CurDAG->MorphNodeTo(N, ARMISD::CMOV, N->getVTList(), Ops);
+      }
+
+    }
+    // Other cases are autogenerated.
+    break;
+  }
     
   case ARMISD::VZIP: {
     unsigned Opc = 0;