[NVPTX] Don't use stack memory when bitcasting to/from v2i8 (#113928)

`v2i8` is an unsupported type, so we hit the default legalization rules
which perform the bitcast in stack memory and is very inefficient on
GPU.

This adds a custom lowering where we pack `v2i8` into `i16` and from
there use another bitcast node to reach the final desired type. And also
the inverse unpacking `i16` into `v2i8`.

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

@@ -409,6 +409,13 @@ VectorizePTXValueVTs(const SmallVectorImpl<EVT> &ValueVTs,
   return VectorInfo;
 }
 
+static SDValue MaybeBitcast(SelectionDAG &DAG, SDLoc DL, EVT VT,
+                            SDValue Value) {
+  if (Value->getValueType(0) == VT)
+    return Value;
+  return DAG.getNode(ISD::BITCAST, DL, VT, Value);
+}
+
 // NVPTXTargetLowering Constructor.
 NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
                                          const NVPTXSubtarget &STI)
@@ -551,6 +558,10 @@ NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
   setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i8, Custom);
   setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i8, Custom);
   setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i8, Custom);
+
+  // Custom conversions to/from v2i8.
+  setOperationAction(ISD::BITCAST, MVT::v2i8, Custom);
+
   // Only logical ops can be done on v4i8 directly, others must be done
   // elementwise.
   setOperationAction(
@@ -2309,6 +2320,30 @@ NVPTXTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
   return DAG.getBuildVector(Node->getValueType(0), dl, Ops);
 }
 
+SDValue NVPTXTargetLowering::LowerBITCAST(SDValue Op, SelectionDAG &DAG) const {
+  // Handle bitcasting from v2i8 without hitting the default promotion
+  // strategy which goes through stack memory.
+  EVT FromVT = Op->getOperand(0)->getValueType(0);
+  if (FromVT != MVT::v2i8) {
+    return Op;
+  }
+
+  // Pack vector elements into i16 and bitcast to final type
+  SDLoc DL(Op);
+  SDValue Vec0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i8,
+                             Op->getOperand(0), DAG.getIntPtrConstant(0, DL));
+  SDValue Vec1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i8,
+                             Op->getOperand(0), DAG.getIntPtrConstant(1, DL));
+  SDValue Extend0 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i16, Vec0);
+  SDValue Extend1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i16, Vec1);
+  SDValue Const8 = DAG.getConstant(8, DL, MVT::i16);
+  SDValue AsInt = DAG.getNode(
+      ISD::OR, DL, MVT::i16,
+      {Extend0, DAG.getNode(ISD::SHL, DL, MVT::i16, {Extend1, Const8})});
+  EVT ToVT = Op->getValueType(0);
+  return MaybeBitcast(DAG, DL, ToVT, AsInt);
+}
+
 // We can init constant f16x2/v2i16/v4i8 with a single .b32 move.  Normally it
 // would get lowered as two constant loads and vector-packing move.
 // Instead we want just a constant move:
@@ -2817,6 +2852,8 @@ NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
     return Op;
   case ISD::BUILD_VECTOR:
     return LowerBUILD_VECTOR(Op, DAG);
+  case ISD::BITCAST:
+    return LowerBITCAST(Op, DAG);
   case ISD::EXTRACT_SUBVECTOR:
     return Op;
   case ISD::EXTRACT_VECTOR_ELT:
@@ -6127,6 +6164,28 @@ SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
   return SDValue();
 }
 
+static void ReplaceBITCAST(SDNode *Node, SelectionDAG &DAG,
+                           SmallVectorImpl<SDValue> &Results) {
+  // Handle bitcasting to v2i8 without hitting the default promotion
+  // strategy which goes through stack memory.
+  SDValue Op(Node, 0);
+  EVT ToVT = Op->getValueType(0);
+  if (ToVT != MVT::v2i8) {
+    return;
+  }
+
+  // Bitcast to i16 and unpack elements into a vector
+  SDLoc DL(Node);
+  SDValue AsInt = MaybeBitcast(DAG, DL, MVT::i16, Op->getOperand(0));
+  SDValue Vec0 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, AsInt);
+  SDValue Const8 = DAG.getConstant(8, DL, MVT::i16);
+  SDValue Vec1 =
+      DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
+                  DAG.getNode(ISD::SRL, DL, MVT::i16, {AsInt, Const8}));
+  Results.push_back(
+      DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2i8, {Vec0, Vec1}));
+}
+
 /// ReplaceVectorLoad - Convert vector loads into multi-output scalar loads.
 static void ReplaceLoadVector(SDNode *N, SelectionDAG &DAG,
                               SmallVectorImpl<SDValue> &Results) {
@@ -6412,6 +6471,9 @@ void NVPTXTargetLowering::ReplaceNodeResults(
   switch (N->getOpcode()) {
   default:
     report_fatal_error("Unhandled custom legalization");
+  case ISD::BITCAST:
+    ReplaceBITCAST(N, DAG, Results);
+    return;
   case ISD::LOAD:
     ReplaceLoadVector(N, DAG, Results);
     return;

llvm/lib/Target/NVPTX/NVPTXISelLowering.h

@@ -616,6 +616,8 @@ class NVPTXTargetLowering : public TargetLowering {
   const NVPTXSubtarget &STI; // cache the subtarget here
   SDValue getParamSymbol(SelectionDAG &DAG, int idx, EVT) const;
 
+  SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) const;
+
   SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;

llvm/test/CodeGen/NVPTX/i8x2-instructions.ll

@@ -0,0 +1,33 @@
+; RUN: llc < %s -mtriple=nvptx64-nvidia-cuda -mcpu=sm_90 -mattr=+ptx80 -asm-verbose=false \
+; RUN:          -O0 -disable-post-ra -frame-pointer=all -verify-machineinstrs \
+; RUN: | FileCheck  %s
+; RUN: %if ptxas %{                                                           \
+; RUN:   llc < %s -mtriple=nvptx64-nvidia-cuda -mcpu=sm_90 -asm-verbose=false \
+; RUN:          -O0 -disable-post-ra -frame-pointer=all -verify-machineinstrs \
+; RUN:   | %ptxas-verify -arch=sm_90                                          \
+; RUN: %}
+
+target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+
+; CHECK-LABEL: test_bitcast_2xi8_i16(
+; CHECK: ld.param.u32 	%r1, [test_bitcast_2xi8_i16_param_0];
+; CHECK: mov.b32 	{%rs1, %rs2}, %r1;
+; CHECK: shl.b16 	%rs3, %rs2, 8;
+; CHECK: and.b16  	%rs4, %rs1, 255;
+; CHECK: or.b16  	%rs5, %rs4, %rs3;
+; CHECK: cvt.u32.u16 	%r2, %rs5;
+; CHECK: st.param.b32 	[func_retval0], %r2;
+define i16 @test_bitcast_2xi8_i16(<2 x i8> %a) {
+  %res = bitcast <2 x i8> %a to i16
+  ret i16 %res
+}
+
+; CHECK-LABEL: test_bitcast_i16_2xi8(
+; CHECK: ld.param.u16 	%rs1, [test_bitcast_i16_2xi8_param_0];
+; CHECK: shr.u16 	%rs2, %rs1, 8;
+; CHECK: mov.b32 	%r1, {%rs1, %rs2};
+; CHECK: st.param.b32 	[func_retval0], %r1;
+define <2 x i8> @test_bitcast_i16_2xi8(i16 %a) {
+  %res = bitcast i16 %a to <2 x i8>
+  ret <2 x i8> %res
+}