[NVPTX] support switch statement with brx.idx (llvm#102400)

Add custom lowering for `BR_JT` DAG nodes to the `brx.idx` PTX
instruction ([PTX ISA 9.7.13.4. Control Flow Instructions: brx.idx]
(https://docs.nvidia.com/cuda/parallel-thread-execution/#control-flow-instructions-brx-idx)).
Depending on the heuristics in DAG selection, `switch` statements may
now be lowered using `brx.idx`

llvm/include/llvm/CodeGen/TargetLowering.h

@@ -3843,6 +3843,10 @@ class TargetLowering : public TargetLoweringBase {
   /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum.
   virtual unsigned getJumpTableEncoding() const;
 
+  virtual MVT getJumpTableRegTy(const DataLayout &DL) const {
+    return getPointerTy(DL);
+  }
+
   virtual const MCExpr *
   LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/,
                             const MachineBasicBlock * /*MBB*/, unsigned /*uid*/,

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -2977,7 +2977,7 @@ void SelectionDAGBuilder::visitJumpTable(SwitchCG::JumpTable &JT) {
   // Emit the code for the jump table
   assert(JT.SL && "Should set SDLoc for SelectionDAG!");
   assert(JT.Reg != -1U && "Should lower JT Header first!");
-  EVT PTy = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
+  EVT PTy = DAG.getTargetLoweringInfo().getJumpTableRegTy(DAG.getDataLayout());
   SDValue Index = DAG.getCopyFromReg(getControlRoot(), *JT.SL, JT.Reg, PTy);
   SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
   SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, *JT.SL, MVT::Other,
@@ -3005,12 +3005,13 @@ void SelectionDAGBuilder::visitJumpTableHeader(SwitchCG::JumpTable &JT,
   // This value may be smaller or larger than the target's pointer type, and
   // therefore require extension or truncating.
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
-  SwitchOp = DAG.getZExtOrTrunc(Sub, dl, TLI.getPointerTy(DAG.getDataLayout()));
+  SwitchOp =
+      DAG.getZExtOrTrunc(Sub, dl, TLI.getJumpTableRegTy(DAG.getDataLayout()));
 
   unsigned JumpTableReg =
-      FuncInfo.CreateReg(TLI.getPointerTy(DAG.getDataLayout()));
-  SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), dl,
-                                    JumpTableReg, SwitchOp);
+      FuncInfo.CreateReg(TLI.getJumpTableRegTy(DAG.getDataLayout()));
+  SDValue CopyTo =
+      DAG.getCopyToReg(getControlRoot(), dl, JumpTableReg, SwitchOp);
   JT.Reg = JumpTableReg;
 
   if (!JTH.FallthroughUnreachable) {

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

@@ -25,6 +25,7 @@
 #include "llvm/CodeGen/Analysis.h"
 #include "llvm/CodeGen/ISDOpcodes.h"
 #include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineJumpTableInfo.h"
 #include "llvm/CodeGen/MachineMemOperand.h"
 #include "llvm/CodeGen/SelectionDAG.h"
 #include "llvm/CodeGen/SelectionDAGNodes.h"
@@ -582,9 +583,7 @@ NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
   setOperationAction(ISD::ROTR, MVT::i8, Expand);
   setOperationAction(ISD::BSWAP, MVT::i16, Expand);
 
-  // Indirect branch is not supported.
-  // This also disables Jump Table creation.
-  setOperationAction(ISD::BR_JT, MVT::Other, Expand);
+  setOperationAction(ISD::BR_JT, MVT::Other, Custom);
   setOperationAction(ISD::BRIND, MVT::Other, Expand);
 
   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
@@ -945,6 +944,9 @@ const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
     MAKE_CASE(NVPTXISD::Dummy)
     MAKE_CASE(NVPTXISD::MUL_WIDE_SIGNED)
     MAKE_CASE(NVPTXISD::MUL_WIDE_UNSIGNED)
+    MAKE_CASE(NVPTXISD::BrxEnd)
+    MAKE_CASE(NVPTXISD::BrxItem)
+    MAKE_CASE(NVPTXISD::BrxStart)
     MAKE_CASE(NVPTXISD::Tex1DFloatS32)
     MAKE_CASE(NVPTXISD::Tex1DFloatFloat)
     MAKE_CASE(NVPTXISD::Tex1DFloatFloatLevel)
@@ -2785,6 +2787,8 @@ NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
     return LowerFP_ROUND(Op, DAG);
   case ISD::FP_EXTEND:
     return LowerFP_EXTEND(Op, DAG);
+  case ISD::BR_JT:
+    return LowerBR_JT(Op, DAG);
   case ISD::VAARG:
     return LowerVAARG(Op, DAG);
   case ISD::VASTART:
@@ -2810,6 +2814,41 @@ NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   }
 }
 
+SDValue NVPTXTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
+  SDLoc DL(Op);
+  SDValue Chain = Op.getOperand(0);
+  const auto *JT = cast<JumpTableSDNode>(Op.getOperand(1));
+  SDValue Index = Op.getOperand(2);
+
+  unsigned JId = JT->getIndex();
+  MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo();
+  ArrayRef<MachineBasicBlock *> MBBs = MJTI->getJumpTables()[JId].MBBs;
+
+  SDValue IdV = DAG.getConstant(JId, DL, MVT::i32);
+
+  // Generate BrxStart node
+  SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
+  Chain = DAG.getNode(NVPTXISD::BrxStart, DL, VTs, Chain, IdV);
+
+  // Generate BrxItem nodes
+  assert(!MBBs.empty());
+  for (MachineBasicBlock *MBB : MBBs.drop_back())
+    Chain = DAG.getNode(NVPTXISD::BrxItem, DL, VTs, Chain.getValue(0),
+                        DAG.getBasicBlock(MBB), Chain.getValue(1));
+
+  // Generate BrxEnd nodes
+  SDValue EndOps[] = {Chain.getValue(0), DAG.getBasicBlock(MBBs.back()), Index,
+                      IdV, Chain.getValue(1)};
+  SDValue BrxEnd = DAG.getNode(NVPTXISD::BrxEnd, DL, VTs, EndOps);
+
+  return BrxEnd;
+}
+
+// This will prevent AsmPrinter from trying to print the jump tables itself.
+unsigned NVPTXTargetLowering::getJumpTableEncoding() const {
+  return MachineJumpTableInfo::EK_Inline;
+}
+
 // This function is almost a copy of SelectionDAG::expandVAArg().
 // The only diff is that this one produces loads from local address space.
 SDValue NVPTXTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {

llvm/lib/Target/NVPTX/NVPTXISelLowering.h

@@ -62,6 +62,9 @@ enum NodeType : unsigned {
   BFI,
   PRMT,
   DYNAMIC_STACKALLOC,
+  BrxStart,
+  BrxItem,
+  BrxEnd,
   Dummy,
 
   LoadV2 = ISD::FIRST_TARGET_MEMORY_OPCODE,
@@ -580,6 +583,11 @@ class NVPTXTargetLowering : public TargetLowering {
     return true;
   }
 
+  // The default is the same as pointer type, but brx.idx only accepts i32
+  MVT getJumpTableRegTy(const DataLayout &) const override { return MVT::i32; }
+
+  unsigned getJumpTableEncoding() const override;
+
   bool enableAggressiveFMAFusion(EVT VT) const override { return true; }
 
   // The default is to transform llvm.ctlz(x, false) (where false indicates that
@@ -637,6 +645,8 @@ class NVPTXTargetLowering : public TargetLowering {
 
   SDValue LowerSelect(SDValue Op, SelectionDAG &DAG) const;
 
+  SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
+
   SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
 

llvm/lib/Target/NVPTX/NVPTXInstrInfo.td

@@ -3880,6 +3880,44 @@ def DYNAMIC_STACKALLOC64 :
             [(set Int64Regs:$ptr, (dyn_alloca Int64Regs:$size, (i32 timm:$align)))]>,
             Requires<[hasPTX<73>, hasSM<52>]>;
 
+
+//
+// BRX
+//
+
+def SDTBrxStartProfile : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
+def SDTBrxItemProfile : SDTypeProfile<0, 1, [SDTCisVT<0, OtherVT>]>;
+def SDTBrxEndProfile : SDTypeProfile<0, 3, [SDTCisVT<0, OtherVT>, SDTCisInt<1>, SDTCisInt<2>]>;
+
+def brx_start :
+  SDNode<"NVPTXISD::BrxStart", SDTBrxStartProfile,
+         [SDNPHasChain, SDNPOutGlue, SDNPSideEffect]>;
+def brx_item :
+  SDNode<"NVPTXISD::BrxItem", SDTBrxItemProfile,
+         [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
+def brx_end :
+  SDNode<"NVPTXISD::BrxEnd", SDTBrxEndProfile,
+         [SDNPHasChain, SDNPInGlue, SDNPSideEffect]>;
+
+let isTerminator = 1, isBranch = 1, isIndirectBranch = 1 in {
+
+  def BRX_START :
+    NVPTXInst<(outs), (ins i32imm:$id),
+              "$$L_brx_$id: .branchtargets",
+              [(brx_start (i32 imm:$id))]>;
+
+  def BRX_ITEM :
+    NVPTXInst<(outs), (ins brtarget:$target),
+              "\t$target,",
+              [(brx_item bb:$target)]>;
+
+  def BRX_END :
+    NVPTXInst<(outs), (ins brtarget:$target, Int32Regs:$val, i32imm:$id),
+              "\t$target;\n\tbrx.idx \t$val, $$L_brx_$id;",
+              [(brx_end bb:$target, (i32 Int32Regs:$val), (i32 imm:$id))]>;
+}
+
+
 include "NVPTXIntrinsics.td"
 
 //-----------------------------------

llvm/test/CodeGen/NVPTX/jump-table.ll

@@ -0,0 +1,69 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s | FileCheck %s
+; RUN: %if ptxas %{ llc < %s | %ptxas-verify %}
+
+target triple = "nvptx64-nvidia-cuda"
+
+@out = addrspace(1) global i32 0, align 4
+
+define void @foo(i32 %i) {
+; CHECK-LABEL: foo(
+; CHECK:       {
+; CHECK-NEXT:    .reg .pred %p<2>;
+; CHECK-NEXT:    .reg .b32 %r<7>;
+; CHECK-EMPTY:
+; CHECK-NEXT:  // %bb.0: // %entry
+; CHECK-NEXT:    ld.param.u32 %r2, [foo_param_0];
+; CHECK-NEXT:    setp.gt.u32 %p1, %r2, 3;
+; CHECK-NEXT:    @%p1 bra $L__BB0_6;
+; CHECK-NEXT:  // %bb.1: // %entry
+; CHECK-NEXT:    $L_brx_0: .branchtargets
+; CHECK-NEXT:     $L__BB0_2,
+; CHECK-NEXT:     $L__BB0_3,
+; CHECK-NEXT:     $L__BB0_4,
+; CHECK-NEXT:     $L__BB0_5;
+; CHECK-NEXT:    brx.idx %r2, $L_brx_0;
+; CHECK-NEXT:  $L__BB0_2: // %case0
+; CHECK-NEXT:    mov.b32 %r6, 0;
+; CHECK-NEXT:    st.global.u32 [out], %r6;
+; CHECK-NEXT:    bra.uni $L__BB0_6;
+; CHECK-NEXT:  $L__BB0_4: // %case2
+; CHECK-NEXT:    mov.b32 %r4, 2;
+; CHECK-NEXT:    st.global.u32 [out], %r4;
+; CHECK-NEXT:    bra.uni $L__BB0_6;
+; CHECK-NEXT:  $L__BB0_5: // %case3
+; CHECK-NEXT:    mov.b32 %r3, 3;
+; CHECK-NEXT:    st.global.u32 [out], %r3;
+; CHECK-NEXT:    bra.uni $L__BB0_6;
+; CHECK-NEXT:  $L__BB0_3: // %case1
+; CHECK-NEXT:    mov.b32 %r5, 1;
+; CHECK-NEXT:    st.global.u32 [out], %r5;
+; CHECK-NEXT:  $L__BB0_6: // %end
+; CHECK-NEXT:    ret;
+entry:
+  switch i32 %i, label %end [
+    i32 0, label %case0
+    i32 1, label %case1
+    i32 2, label %case2
+    i32 3, label %case3
+  ]
+
+case0:
+  store i32 0, ptr addrspace(1) @out, align 4
+  br label %end
+
+case1:
+  store i32 1, ptr addrspace(1) @out, align 4
+  br label %end
+
+case2:
+  store i32 2, ptr addrspace(1) @out, align 4
+  br label %end
+
+case3:
+  store i32 3, ptr addrspace(1) @out, align 4
+  br label %end
+
+end:
+  ret void
+}