Add basic support to code-gen the ARM/Thumb2 bit-field insert (BFI) instruction

and a combine pattern to use it for setting a bit-field to a constant
value. More to come for non-constant stores.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@108570 91177308-0d34-0410-b5e6-96231b3b80d8

8 files changed, 138 insertions, 36 deletions

diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 444e30e..34b662a 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -530,6 +530,9 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
   setTargetDAGCombine(ISD::SUB);
   setTargetDAGCombine(ISD::MUL);
 
+  if (Subtarget->hasV6T2Ops())
+    setTargetDAGCombine(ISD::OR);
+
   setStackPointerRegisterToSaveRestore(ARM::SP);
 
   if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
@@ -4232,6 +4235,53 @@ static SDValue PerformMULCombine(SDNode *N,
   return SDValue();
 }
 
+/// PerformORCombine - Target-specific dag combine xforms for ISD::OR
+static SDValue PerformORCombine(SDNode *N,
+                                TargetLowering::DAGCombinerInfo &DCI,
+                                const ARMSubtarget *Subtarget) {
+  // BFI is only available on V6T2+
+  if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
+  // or (and A, mask), val => ARMbfi A, val, mask
+  //   iff (val & mask) == val
+  if (N0->getOpcode() != ISD::AND)
+    return SDValue();
+
+  EVT VT = N->getValueType(0);
+  if (VT != MVT::i32)
+    return SDValue();
+
+  // The value and the mask need to be constants so we can verify this is
+  // actually a bitfield set. If the mask is 0xffff, we can do better
+  // via a movt instruction, so don't use BFI in that case.
+  ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+  if (!C)
+    return SDValue();
+  unsigned Mask = C->getZExtValue();
+  if (Mask == 0xffff)
+    return SDValue();
+  C = dyn_cast<ConstantSDNode>(N1);
+  if (!C)
+    return SDValue();
+  unsigned Val = C->getZExtValue();
+  if (ARM::isBitFieldInvertedMask(Mask) && (Val & ~Mask) != Val)
+    return SDValue();
+  Val >>= CountTrailingZeros_32(~Mask);
+
+  DebugLoc DL = N->getDebugLoc();
+  SDValue Res = DAG.getNode(ARMISD::BFI, DL, VT, N0.getOperand(0),
+                            DAG.getConstant(Val, MVT::i32),
+                            DAG.getConstant(Mask, MVT::i32));
+
+  // Do not add new nodes to DAG combiner worklist.
+  DCI.CombineTo(N, Res, false);
+
+  return SDValue();
+}
+
 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
 /// ARMISD::VMOVRRD.
 static SDValue PerformVMOVRRDCombine(SDNode *N,
@@ -4649,6 +4699,7 @@ SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::ADD:        return PerformADDCombine(N, DCI);
   case ISD::SUB:        return PerformSUBCombine(N, DCI);
   case ISD::MUL:        return PerformMULCombine(N, DCI, Subtarget);
+  case ISD::OR:         return PerformORCombine(N, DCI, Subtarget);
   case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
   case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
   case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
@@ -5380,6 +5431,21 @@ int ARM::getVFPf64Imm(const APFloat &FPImm) {
   return ((int)Sign << 7) | (Exp << 4) | Mantissa;
 }
 
+bool ARM::isBitFieldInvertedMask(unsigned v) {
+  if (v == 0xffffffff)
+    return 0;
+  // there can be 1's on either or both "outsides", all the "inside"
+  // bits must be 0's
+  unsigned int lsb = 0, msb = 31;
+  while (v & (1 << msb)) --msb;
+  while (v & (1 << lsb)) ++lsb;
+  for (unsigned int i = lsb; i <= msb; ++i) {
+    if (v & (1 << i))
+      return 0;
+  }
+  return 1;
+}
+
 /// isFPImmLegal - Returns true if the target can instruction select the
 /// specified FP immediate natively. If false, the legalizer will
 /// materialize the FP immediate as a load from a constant pool.
diff --git a/lib/Target/ARM/ARMISelLowering.h b/lib/Target/ARM/ARMISelLowering.h
index 128b72e..1c178b3 100644
--- a/lib/Target/ARM/ARMISelLowering.h
+++ b/lib/Target/ARM/ARMISelLowering.h
@@ -150,7 +150,10 @@ namespace llvm {
 
       // Floating-point max and min:
       FMAX,
-      FMIN
+      FMIN,
+
+      // Bit-field insert
+      BFI
     };
   }
 
@@ -162,6 +165,7 @@ namespace llvm {
     /// returns -1.
     int getVFPf32Imm(const APFloat &FPImm);
     int getVFPf64Imm(const APFloat &FPImm);
+    bool isBitFieldInvertedMask(unsigned v);
   }
 
   //===--------------------------------------------------------------------===//
diff --git a/lib/Target/ARM/ARMInstrInfo.td b/lib/Target/ARM/ARMInstrInfo.td
index 51fc152..7e00b4b 100644
--- a/lib/Target/ARM/ARMInstrInfo.td
+++ b/lib/Target/ARM/ARMInstrInfo.td
@@ -61,6 +61,9 @@ def SDT_ARMSYNCBARRIERV6 : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
 
 def SDT_ARMTCRET : SDTypeProfile<0, 1, [SDTCisPtrTy<0>]>;
 
+def SDT_ARMBFI : SDTypeProfile<1, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>,
+                                      SDTCisVT<2, i32>, SDTCisVT<3, i32>]>;
+
 // Node definitions.
 def ARMWrapper       : SDNode<"ARMISD::Wrapper",     SDTIntUnaryOp>;
 def ARMWrapperJT     : SDNode<"ARMISD::WrapperJT",   SDTIntBinOp>;
@@ -131,6 +134,9 @@ def ARMrbit          : SDNode<"ARMISD::RBIT", SDTIntUnaryOp>;
 def ARMtcret         : SDNode<"ARMISD::TC_RETURN", SDT_ARMTCRET, 
                         [SDNPHasChain,  SDNPOptInFlag, SDNPVariadic]>;
 
+
+def ARMbfi           : SDNode<"ARMISD::BFI", SDT_ARMBFI>;
+
 //===----------------------------------------------------------------------===//
 // ARM Instruction Predicate Definitions.
 //
@@ -221,19 +227,7 @@ def sext_16_node : PatLeaf<(i32 GPR:$a), [{
 /// e.g., 0xf000ffff
 def bf_inv_mask_imm : Operand<i32>,
                       PatLeaf<(imm), [{
-  uint32_t v = (uint32_t)N->getZExtValue();
-  if (v == 0xffffffff)
-    return 0;
-  // there can be 1's on either or both "outsides", all the "inside"
-  // bits must be 0's
-  unsigned int lsb = 0, msb = 31;
-  while (v & (1 << msb)) --msb;
-  while (v & (1 << lsb)) ++lsb;
-  for (unsigned int i = lsb; i <= msb; ++i) {
-    if (v & (1 << i))
-      return 0;
-  }
-  return 1;
+  return ARM::isBitFieldInvertedMask(N->getZExtValue());
 }] > {
   let PrintMethod = "printBitfieldInvMaskImmOperand";
 }
@@ -1858,11 +1852,11 @@ def BFC    : I<(outs GPR:$dst), (ins GPR:$src, bf_inv_mask_imm:$imm),
 }
 
 // A8.6.18  BFI - Bitfield insert (Encoding A1)
-// Added for disassembler with the pattern field purposely left blank.
-def BFI    : I<(outs GPR:$dst), (ins GPR:$src, bf_inv_mask_imm:$imm),
+def BFI    : I<(outs GPR:$dst), (ins GPR:$src, GPR:$val, bf_inv_mask_imm:$imm),
                AddrMode1, Size4Bytes, IndexModeNone, DPFrm, IIC_iUNAsi,
-               "bfi", "\t$dst, $src, $imm", "",
-               [/* For disassembly only; pattern left blank */]>,
+               "bfi", "\t$dst, $val, $imm", "$src = $dst",
+               [(set GPR:$dst, (ARMbfi GPR:$src, GPR:$val,
+                                bf_inv_mask_imm:$imm))]>,
                Requires<[IsARM, HasV6T2]> {
   let Inst{27-21} = 0b0111110;
   let Inst{6-4}   = 0b001; // Rn: Inst{3-0} != 15
diff --git a/lib/Target/ARM/ARMInstrThumb2.td b/lib/Target/ARM/ARMInstrThumb2.td
index bbe675e..d082a6b 100644
--- a/lib/Target/ARM/ARMInstrThumb2.td
+++ b/lib/Target/ARM/ARMInstrThumb2.td
@@ -1666,10 +1666,12 @@ def t2UBFX : T2I<(outs GPR:$dst), (ins GPR:$src, imm0_31:$lsb, imm0_31:$width),
 }
 
 // A8.6.18  BFI - Bitfield insert (Encoding T1)
-// Added for disassembler with the pattern field purposely left blank.
-// FIXME: Utilize this instruction in codgen.
-def t2BFI : T2I<(outs GPR:$dst), (ins GPR:$src, imm0_31:$lsb, imm0_31:$width),
-                IIC_iALUi, "bfi", "\t$dst, $src, $lsb, $width", []> {
+let Constraints = "$src = $dst" in
+def t2BFI : T2I<(outs GPR:$dst),
+                (ins GPR:$src, GPR:$val, bf_inv_mask_imm:$imm),
+                IIC_iALUi, "bfi", "\t$dst, $val, $imm",
+                [(set GPR:$dst, (ARMbfi GPR:$src, GPR:$val,
+                                 bf_inv_mask_imm:$imm))]> {
   let Inst{31-27} = 0b11110;
   let Inst{25} = 1;
   let Inst{24-20} = 0b10110;
diff --git a/lib/Target/ARM/Disassembler/ARMDisassemblerCore.cpp b/lib/Target/ARM/Disassembler/ARMDisassemblerCore.cpp
index a07ff28..57972d1 100644
--- a/lib/Target/ARM/Disassembler/ARMDisassemblerCore.cpp
+++ b/lib/Target/ARM/Disassembler/ARMDisassemblerCore.cpp
@@ -989,10 +989,12 @@ static bool DisassembleDPFrm(MCInst &MI, unsigned Opcode, uint32_t insn,
 
   // Special-case handling of BFC/BFI/SBFX/UBFX.
   if (Opcode == ARM::BFC || Opcode == ARM::BFI) {
-    // TIED_TO operand skipped for BFC and Inst{3-0} (Reg) for BFI.
-    MI.addOperand(MCOperand::CreateReg(Opcode == ARM::BFC ? 0
-                                       : getRegisterEnum(B, ARM::GPRRegClassID,
+    MI.addOperand(MCOperand::CreateReg(0));
+    if (Opcode == ARM::BFI) {
+      MI.addOperand(MCOperand::CreateReg(getRegisterEnum(B, ARM::GPRRegClassID,
                                                          decodeRm(insn))));
+      ++OpIdx;
+    }
     uint32_t mask = 0;
     if (!getBFCInvMask(insn, mask))
       return false;
diff --git a/lib/Target/ARM/Disassembler/ThumbDisassemblerCore.h b/lib/Target/ARM/Disassembler/ThumbDisassemblerCore.h
index 4b7a0bf..3e7d4d6 100644
--- a/lib/Target/ARM/Disassembler/ThumbDisassemblerCore.h
+++ b/lib/Target/ARM/Disassembler/ThumbDisassemblerCore.h
@@ -1555,6 +1555,13 @@ static bool DisassembleThumb2DPBinImm(MCInst &MI, unsigned Opcode,
     ++OpIdx;
   }
 
+  if (Opcode == ARM::t2BFI) {
+    // Add val reg operand.
+    MI.addOperand(MCOperand::CreateReg(getRegisterEnum(B, ARM::GPRRegClassID,
+                                                       decodeRn(insn))));
+    ++OpIdx;
+  }
+
   assert(OpInfo[OpIdx].RegClass < 0 && !OpInfo[OpIdx].isPredicate()
          && !OpInfo[OpIdx].isOptionalDef()
          && "Pure imm operand expected");
@@ -1567,7 +1574,7 @@ static bool DisassembleThumb2DPBinImm(MCInst &MI, unsigned Opcode,
     MI.addOperand(MCOperand::CreateImm(getIImm3Imm8(insn)));
   else if (Opcode == ARM::t2MOVi16 || Opcode == ARM::t2MOVTi16)
     MI.addOperand(MCOperand::CreateImm(getImm16(insn)));
-  else if (Opcode == ARM::t2BFC) {
+  else if (Opcode == ARM::t2BFC || Opcode == ARM::t2BFI) {
     uint32_t mask = 0;
     if (getBitfieldInvMask(insn, mask))
       MI.addOperand(MCOperand::CreateImm(mask));
@@ -1575,17 +1582,10 @@ static bool DisassembleThumb2DPBinImm(MCInst &MI, unsigned Opcode,
       return false;
   } else {
     // Handle the case of: lsb width
-    assert((Opcode == ARM::t2SBFX || Opcode == ARM::t2UBFX ||
-            Opcode == ARM::t2BFI) && "Unexpected opcode");
+    assert((Opcode == ARM::t2SBFX || Opcode == ARM::t2UBFX)
+            && "Unexpected opcode");
     MI.addOperand(MCOperand::CreateImm(getLsb(insn)));
-    if (Opcode == ARM::t2BFI) {
-      if (getMsb(insn) < getLsb(insn)) {
-        DEBUG(errs() << "Encoding error: msb < lsb\n");
-        return false;
-      }
-      MI.addOperand(MCOperand::CreateImm(getMsb(insn) - getLsb(insn) + 1));
-    } else
-      MI.addOperand(MCOperand::CreateImm(getWidthMinus1(insn) + 1));
+    MI.addOperand(MCOperand::CreateImm(getWidthMinus1(insn) + 1));
 
     ++OpIdx;
   }
diff --git a/test/CodeGen/ARM/bfi.ll b/test/CodeGen/ARM/bfi.ll
new file mode 100644
index 0000000..48ef437
--- /dev/null
+++ b/test/CodeGen/ARM/bfi.ll
@@ -0,0 +1,17 @@
+; RUN: llc -march=arm -mattr=+v6t2 < %s | FileCheck %s
+
+%struct.F = type { [3 x i8], i8 }
+
+@X = common global %struct.F zeroinitializer, align 4 ; <%struct.F*> [#uses=1]
+
+define void @f1([1 x i32] %f.coerce0) nounwind {
+entry:
+; CHECK: f1
+; CHECK: mov r2, #10
+; CHECK: bfi r1, r2, #22, #4
+  %0 = load i32* bitcast (%struct.F* @X to i32*), align 4 ; <i32> [#uses=1]
+  %1 = and i32 %0, -62914561                      ; <i32> [#uses=1]
+  %2 = or i32 %1, 41943040                        ; <i32> [#uses=1]
+  store i32 %2, i32* bitcast (%struct.F* @X to i32*), align 4
+  ret void
+}
diff --git a/test/CodeGen/Thumb2/bfi.ll b/test/CodeGen/Thumb2/bfi.ll
new file mode 100644
index 0000000..a256d67
--- /dev/null
+++ b/test/CodeGen/Thumb2/bfi.ll
@@ -0,0 +1,17 @@
+; RUN: llc -march=thumb -mattr=+v6t2 < %s | FileCheck %s
+
+%struct.F = type { [3 x i8], i8 }
+
+@X = common global %struct.F zeroinitializer, align 4 ; <%struct.F*> [#uses=1]
+
+define void @f1([1 x i32] %f.coerce0) nounwind {
+entry:
+; CHECK: f1
+; CHECK: movs r2, #10
+; CHECK: bfi r1, r2, #22, #4
+  %0 = load i32* bitcast (%struct.F* @X to i32*), align 4 ; <i32> [#uses=1]
+  %1 = and i32 %0, -62914561                      ; <i32> [#uses=1]
+  %2 = or i32 %1, 41943040                        ; <i32> [#uses=1]
+  store i32 %2, i32* bitcast (%struct.F* @X to i32*), align 4
+  ret void
+}