R600: Swap the legality of rotl and rotr

The hardware supports rotr and not rotl.

llvm-svn: 182285

lib/Target/R600/AMDGPUISelLowering.cpp

@@ -46,6 +46,9 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
   setOperationAction(ISD::FRINT,  MVT::f32, Legal);
 
+  // The hardware supports ROTR, but not ROTL
+  setOperationAction(ISD::ROTL, MVT::i32, Expand);
+
   // Lower floating point store/load to integer store/load to reduce the number
   // of patterns in tablegen.
   setOperationAction(ISD::STORE, MVT::f32, Promote);

lib/Target/R600/AMDGPUISelLowering.h

@@ -115,7 +115,6 @@ enum {
   RET_FLAG,
   BRANCH_COND,
   // End AMDIL ISD Opcodes
-  BITALIGN,
   BUFFER_STORE,
   DWORDADDR,
   FRACT,

lib/Target/R600/AMDGPUInstrInfo.td

@@ -23,12 +23,6 @@ def AMDGPUDTIntTernaryOp : SDTypeProfile<1, 3, [
 // AMDGPU DAG Nodes
 //
 
-// out = ((a << 32) | b) >> c)
-//
-// Can be used to optimize rtol:
-// rotl(a, b) = bitalign(a, a, 32 - b)
-def AMDGPUbitalign : SDNode<"AMDGPUISD::BITALIGN", AMDGPUDTIntTernaryOp>;
-
 // This argument to this node is a dword address.
 def AMDGPUdwordaddr : SDNode<"AMDGPUISD::DWORDADDR", SDTIntUnaryOp>;
 

lib/Target/R600/AMDGPUInstructions.td

@@ -295,6 +295,12 @@ class BFEPattern <Instruction BFE> : Pat <
   (BFE $x, $y, $z)
 >;
 
+// rotr pattern
+class ROTRPattern <Instruction BIT_ALIGN> : Pat <
+  (rotr i32:$src0, i32:$src1),
+  (BIT_ALIGN $src0, $src0, $src1)
+>;
+
 include "R600Instructions.td"
 
 include "SIInstrInfo.td"

lib/Target/R600/AMDILISelLowering.cpp

@@ -138,8 +138,6 @@ void AMDGPUTargetLowering::InitAMDILLowering() {
     setOperationAction(ISD::SMUL_LOHI, VT, Expand);
     setOperationAction(ISD::UMUL_LOHI, VT, Expand);
 
-    // GPU doesn't have a rotl, rotr, or byteswap instruction
-    setOperationAction(ISD::ROTR, VT, Expand);
     setOperationAction(ISD::BSWAP, VT, Expand);
 
     // GPU doesn't have any counting operators

lib/Target/R600/R600ISelLowering.cpp

@@ -72,8 +72,6 @@ R600TargetLowering::R600TargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i1, Custom);
 
-  setOperationAction(ISD::ROTL, MVT::i32, Custom);
-
   setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
   setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
 
@@ -480,7 +478,6 @@ using namespace llvm::AMDGPUIntrinsic;
 SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   switch (Op.getOpcode()) {
   default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
-  case ISD::ROTL: return LowerROTL(Op, DAG);
   case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
   case ISD::SELECT: return LowerSELECT(Op, DAG);
   case ISD::STORE: return LowerSTORE(Op, DAG);
@@ -765,18 +762,6 @@ SDValue R600TargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const
   return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF), MVT::i32);
 }
 
-SDValue R600TargetLowering::LowerROTL(SDValue Op, SelectionDAG &DAG) const {
-  DebugLoc DL = Op.getDebugLoc();
-  EVT VT = Op.getValueType();
-
-  return DAG.getNode(AMDGPUISD::BITALIGN, DL, VT,
-                     Op.getOperand(0),
-                     Op.getOperand(0),
-                     DAG.getNode(ISD::SUB, DL, VT,
-                                 DAG.getConstant(32, MVT::i32),
-                                 Op.getOperand(1)));
-}
-
 bool R600TargetLowering::isZero(SDValue Op) const {
   if(ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) {
     return Cst->isNullValue();

lib/Target/R600/R600Instructions.td

@@ -1635,10 +1635,8 @@ let Predicates = [isEGorCayman] in {
   def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>;
   defm : BFIPatterns <BFI_INT_eg>;
 
-  def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
-    [(set i32:$dst, (AMDGPUbitalign i32:$src0, i32:$src1, i32:$src2))],
-    VecALU
-  >;
+  def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
+  def : ROTRPattern <BIT_ALIGN_INT_eg>;
 
   def MULADD_eg : MULADD_Common<0x14>;
   def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;

test/CodeGen/R600/rotr.ll

@@ -0,0 +1,29 @@
+; RUN: llc < %s -debug-only=isel -march=r600 -mcpu=redwood -o - 2>&1 | FileCheck %s
+
+; CHECK: rotr
+; CHECK: @rotr
+; CHECK: BIT_ALIGN_INT
+define void @rotr(i32 addrspace(1)* %in, i32 %x, i32 %y) {
+entry:
+  %0 = sub i32 32, %y
+  %1 = shl i32 %x, %0
+  %2 = lshr i32 %x, %y
+  %3 = or i32 %1, %2
+  store i32 %3, i32 addrspace(1)* %in
+  ret void
+}
+
+; CHECK: rotr
+; CHECK: @rotl
+; CHECK: SUB_INT {{\** T[0-9]+\.[XYZW]}}, literal.x
+; CHECK-NEXT: 32
+; CHECK: BIT_ALIGN_INT {{\** T[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW], PV.[xyzw]}}
+define void @rotl(i32 addrspace(1)* %in, i32 %x, i32 %y) {
+entry:
+  %0 = shl i32 %x, %y
+  %1 = sub i32 32, %y
+  %2 = lshr i32 %x, %1
+  %3 = or i32 %0, %2
+  store i32 %3, i32 addrspace(1)* %in
+  ret void
+}