AMDGPU/GlobalISel: Implement select() for @llvm.amdgcn.cvt.pkrtz

Reviewers: arsenm, nhaehnle

Reviewed By: arsenm

Subscribers: kzhuravl, wdng, yaxunl, rovka, kristof.beyls, dstuttard, tpr, t-tye, llvm-commits

Differential Revision: https://reviews.llvm.org/D45907

llvm-svn: 334757

llvm/lib/Target/AMDGPU/AMDGPUGISel.td

@@ -46,5 +46,38 @@ class GISelVop2Pat <
   (inst src0_vt:$src0, src1_vt:$src1)
 >;
 
+class GISelVop2CommutePat <
+  SDPatternOperator node,
+  Instruction inst,
+  ValueType dst_vt,
+  ValueType src0_vt = dst_vt, ValueType src1_vt = src0_vt>   : GCNPat <
+
+  (dst_vt (node (src1_vt VGPR_32:$src1), (src0_vt (sd_vsrc0 src0_vt:$src0)))),
+  (inst src0_vt:$src0, src1_vt:$src1)
+>;
+
+multiclass GISelVop2IntrPat <
+  SDPatternOperator node, Instruction inst,
+  ValueType dst_vt, ValueType src_vt = dst_vt> {
+
+  def : GISelVop2Pat <node, inst, dst_vt, src_vt>;
+
+  // FIXME: Intrinsics aren't marked as commutable, so we need to add an explcit
+  // pattern to handle commuting.  This is another reason why legalizing to a
+  // generic machine instruction may be better that matching the intrinsic
+  // directly.
+  def : GISelVop2CommutePat <node, inst, dst_vt, src_vt>;
+}
+
 def : GISelSop2Pat <or, S_OR_B32, i32>;
 def : GISelVop2Pat <or, V_OR_B32_e32, i32>;
+
+// FIXME: Select directly to _e32 so we don't need to deal with modifiers.
+// FIXME: We can't re-use SelectionDAG patterns here because they match
+// against a custom SDNode and we would need to create a generic machine
+// instruction that is equivalent to the custom SDNode.  This would also require
+// us to custom legalize the intrinsic to the new generic machine instruction,
+// but I can't get custom legalizing of intrinsic to work and I'm not sure if
+// this is even supported yet.
+defm : GISelVop2IntrPat <
+  int_amdgcn_cvt_pkrtz, V_CVT_PKRTZ_F16_F32_e32, v2f16, f32>;

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

@@ -159,6 +159,19 @@ bool AMDGPUInstructionSelector::selectG_GEP(MachineInstr &I) const {
   return selectG_ADD(I);
 }
 
+bool AMDGPUInstructionSelector::selectG_INTRINSIC(MachineInstr &I,
+                                          CodeGenCoverage &CoverageInfo) const {
+  unsigned IntrinsicID =  I.getOperand(1).getIntrinsicID();
+
+  switch (IntrinsicID) {
+  default:
+    break;
+  case Intrinsic::amdgcn_cvt_pkrtz:
+    return selectImpl(I, CoverageInfo);
+  }
+  return false;
+}
+
 bool AMDGPUInstructionSelector::selectG_STORE(MachineInstr &I) const {
   MachineBasicBlock *BB = I.getParent();
   MachineFunction *MF = BB->getParent();
@@ -522,6 +535,8 @@ bool AMDGPUInstructionSelector::select(MachineInstr &I,
     return selectG_CONSTANT(I);
   case TargetOpcode::G_GEP:
     return selectG_GEP(I);
+  case TargetOpcode::G_INTRINSIC:
+    return selectG_INTRINSIC(I, CoverageInfo);
   case TargetOpcode::G_LOAD:
     return selectG_LOAD(I);
   case TargetOpcode::G_STORE:

llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h

@@ -63,6 +63,7 @@ private:
   bool selectG_CONSTANT(MachineInstr &I) const;
   bool selectG_ADD(MachineInstr &I) const;
   bool selectG_GEP(MachineInstr &I) const;
+  bool selectG_INTRINSIC(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;
   bool hasVgprParts(ArrayRef<GEPInfo> AddrInfo) const;
   void getAddrModeInfo(const MachineInstr &Load, const MachineRegisterInfo &MRI,
                        SmallVectorImpl<GEPInfo> &AddrInfo) const;

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-amdgcn.cvt.pkrtz.mir

@@ -0,0 +1,44 @@
+# RUN: llc -march=amdgcn -run-pass=instruction-select -verify-machineinstrs -global-isel %s -o - | FileCheck %s -check-prefixes=GCN
+
+--- |
+  define void @cvt_pkrtz(i32 addrspace(1)* %global0) { ret void }
+...
+---
+
+name:            cvt_pkrtz
+legalized:       true
+regBankSelected: true
+
+# GCN-LABEL: name: cvt_pkrtz
+body: |
+  bb.0:
+    liveins: $sgpr0, $vgpr0, $vgpr1, $vgpr3_vgpr4
+
+    ; GCN: [[SGPR0:%[0-9]+]]:sreg_32_xm0 = COPY $sgpr0
+    %0:sgpr(s32) = COPY $sgpr0
+    ; GCN: [[VGPR0:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    %1:vgpr(s32) = COPY $vgpr0
+    ; GCN: [[VGPR1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+    %2:vgpr(s32) = COPY $vgpr1
+    %3:vgpr(s64) = COPY $vgpr3_vgpr4
+
+    ; cvt_pkrtz vs
+    ; GCN: V_CVT_PKRTZ_F16_F32_e32 [[SGPR0]], [[VGPR0]]
+    %4:vgpr(<2 x s16>) = G_INTRINSIC intrinsic(@llvm.amdgcn.cvt.pkrtz), %1, %0
+
+    ; cvt_pkrtz sv
+    ; GCN: V_CVT_PKRTZ_F16_F32_e32 [[SGPR0]], [[VGPR0]]
+    %5:vgpr(<2 x s16>) = G_INTRINSIC intrinsic(@llvm.amdgcn.cvt.pkrtz), %0, %1
+
+    ; cvt_pkrtz vv
+    ; GCN: V_CVT_PKRTZ_F16_F32_e32 [[VGPR0]], [[VGPR1]]
+    %6:vgpr(<2 x s16>) = G_INTRINSIC intrinsic(@llvm.amdgcn.cvt.pkrtz), %1, %2
+
+    %7:vgpr(s32) = G_BITCAST %4
+    %8:vgpr(s32) = G_BITCAST %5
+    %9:vgpr(s32) = G_BITCAST %6
+    G_STORE %7, %3 :: (store 4 into %ir.global0)
+    G_STORE %8, %3 :: (store 4 into %ir.global0)
+    G_STORE %9, %3 :: (store 4 into %ir.global0)
+...
+---