AMDGPU/GlobalISel: Implement LDS G_GLOBAL_VALUE

[arsenm]

Summary

Handle the simple case that lowers to a constant.

[arsenm]

Test fixes

[arsenm]

ping

[arsenm]

ping

[ronlieb]

LGTM

[arsenm]

r371424

lib/Target/AMDGPU/AMDGPUISelLowering.h

private:
   /// legalized from a smaller type VT. Need to match pre-legalized type because
   /// the generic legalization inserts the add/sub between the select and
   /// compare.
   SDValue getFFBX_U32(SelectionDAG &DAG, SDValue Op, const SDLoc &DL, unsigned Opc) const;
 
 public:
   static unsigned numBitsUnsigned(SDValue Op, SelectionDAG &DAG);
   static unsigned numBitsSigned(SDValue Op, SelectionDAG &DAG);
+  static bool hasDefinedInitializer(const GlobalValue *GV);
 
 protected:
   SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const;
   /// Split a vector store into multiple scalar stores.
   /// \returns The resulting chain.
 
   SDValue LowerFREM(SDValue Op, SelectionDAG &DAG) const;

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

case ISD::SIGN_EXTEND_INREG:
     // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
     // nothing here and let the illegal result integer be handled normally.
     return;
   default:
     return;
   }
 }
 
-static bool hasDefinedInitializer(const GlobalValue *GV) {
+bool AMDGPUTargetLowering::hasDefinedInitializer(const GlobalValue *GV) {
   const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
   if (!GVar || !GVar->hasInitializer())
     return false;
 
   return !isa<UndefValue>(GVar->getInitializer());
 }
 
 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,

lib/Target/AMDGPU/AMDGPULegalizerInfo.h

bool legalizeITOFP(MachineInstr &MI, MachineRegisterInfo &MRI,
                      MachineIRBuilder &MIRBuilder, bool Signed) const;
   bool legalizeMinNumMaxNum(MachineInstr &MI, MachineRegisterInfo &MRI,
                             MachineIRBuilder &MIRBuilder) const;
   bool legalizeExtractVectorElt(MachineInstr &MI, MachineRegisterInfo &MRI,
                                 MachineIRBuilder &MIRBuilder) const;
   bool legalizeInsertVectorElt(MachineInstr &MI, MachineRegisterInfo &MRI,
                                MachineIRBuilder &MIRBuilder) const;
 
+  bool legalizeGlobalValue(MachineInstr &MI, MachineRegisterInfo &MRI,
+                           MachineIRBuilder &MIRBuilder) const;
+
   Register getLiveInRegister(MachineRegisterInfo &MRI,
                              Register Reg, LLT Ty) const;
 
   bool loadInputValue(Register DstReg, MachineIRBuilder &B,
                       const ArgDescriptor *Arg) const;
   bool legalizePreloadedArgIntrin(
     MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B,
     AMDGPUFunctionArgInfo::PreloadedValue ArgType) const;

lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

 #include "AMDGPULegalizerInfo.h"
 #include "AMDGPUTargetMachine.h"
 #include "SIMachineFunctionInfo.h"
 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
 #include "llvm/CodeGen/TargetOpcodes.h"
 #include "llvm/CodeGen/ValueTypes.h"
 #include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/DiagnosticInfo.h"
 #include "llvm/IR/Type.h"
 #include "llvm/Support/Debug.h"
 
 #define DEBUG_TYPE "amdgpu-legalinfo"
 
 using namespace llvm;
 using namespace LegalizeActions;
 using namespace LegalizeMutations;
AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
   getActionDefinitionsBuilder(G_CONSTANT)
     .legalFor({S1, S32, S64, GlobalPtr,
                LocalPtr, ConstantPtr, PrivatePtr, FlatPtr })
     .clampScalar(0, S32, S64)
     .widenScalarToNextPow2(0)
     .legalIf(isPointer(0));
 
   setAction({G_FRAME_INDEX, PrivatePtr}, Legal);
+  getActionDefinitionsBuilder(G_GLOBAL_VALUE).customFor({LocalPtr});
+
 
   auto &FPOpActions = getActionDefinitionsBuilder(
     { G_FADD, G_FMUL, G_FNEG, G_FABS, G_FMA, G_FCANONICALIZE})
     .legalFor({S32, S64});
 
   if (ST.has16BitInsts()) {
     if (ST.hasVOP3PInsts())
       FPOpActions.legalFor({S16, V2S16});
bool AMDGPULegalizerInfo::legalizeCustom(MachineInstr &MI,
   case TargetOpcode::G_FMAXNUM:
   case TargetOpcode::G_FMINNUM_IEEE:
   case TargetOpcode::G_FMAXNUM_IEEE:
     return legalizeMinNumMaxNum(MI, MRI, MIRBuilder);
   case TargetOpcode::G_EXTRACT_VECTOR_ELT:
     return legalizeExtractVectorElt(MI, MRI, MIRBuilder);
   case TargetOpcode::G_INSERT_VECTOR_ELT:
     return legalizeInsertVectorElt(MI, MRI, MIRBuilder);
+  case TargetOpcode::G_GLOBAL_VALUE:
+    return legalizeGlobalValue(MI, MRI, MIRBuilder);
   default:
     return false;
   }
 
   llvm_unreachable("expected switch to return");
 }
 
 Register AMDGPULegalizerInfo::getSegmentAperture(
if (IdxVal.getValue() < VecTy.getNumElements())
     B.buildInsert(Dst, Vec, Ins, IdxVal.getValue() * EltTy.getSizeInBits());
   else
     B.buildUndef(Dst);
 
   MI.eraseFromParent();
   return true;
 }
 
+bool AMDGPULegalizerInfo::legalizeGlobalValue(
+  MachineInstr &MI, MachineRegisterInfo &MRI,
+  MachineIRBuilder &B) const {
+  Register DstReg = MI.getOperand(0).getReg();
+  LLT Ty = MRI.getType(DstReg);
+  unsigned AS = Ty.getAddressSpace();
+
+  const GlobalValue *GV = MI.getOperand(1).getGlobal();
+  MachineFunction &MF = B.getMF();
+  SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+
+  if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS) {
+    B.setInstr(MI);
+
+    if (!MFI->isEntryFunction()) {
+      const Function &Fn = MF.getFunction();
+      DiagnosticInfoUnsupported BadLDSDecl(
+        Fn, "local memory global used by non-kernel function", MI.getDebugLoc());
+      Fn.getContext().diagnose(BadLDSDecl);
+    }
+
+    // TODO: We could emit code to handle the initialization somewhere.
+    if (!AMDGPUTargetLowering::hasDefinedInitializer(GV)) {
+      B.buildConstant(DstReg, MFI->allocateLDSGlobal(B.getDataLayout(), *GV));
+      MI.eraseFromParent();
+      return true;
+    }
+  } else
+    return false;
+
+  const Function &Fn = MF.getFunction();
+  DiagnosticInfoUnsupported BadInit(
+    Fn, "unsupported initializer for address space", MI.getDebugLoc());
+  Fn.getContext().diagnose(BadInit);
+  return true;
+}
+
 // Return the use branch instruction, otherwise null if the usage is invalid.
 static MachineInstr *verifyCFIntrinsic(MachineInstr &MI,
                                        MachineRegisterInfo &MRI) {
   Register CondDef = MI.getOperand(0).getReg();
   if (!MRI.hasOneNonDBGUse(CondDef))
     return nullptr;
 
   MachineInstr &UseMI = *MRI.use_instr_nodbg_begin(CondDef);

lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
   case AMDGPU::G_IMPLICIT_DEF: {
     unsigned Size = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
     OpdsMapping[0] = AMDGPU::getValueMapping(AMDGPU::SGPRRegBankID, Size);
     break;
   }
   case AMDGPU::G_FCONSTANT:
   case AMDGPU::G_CONSTANT:
   case AMDGPU::G_FRAME_INDEX:
+  case AMDGPU::G_GLOBAL_VALUE:
   case AMDGPU::G_BLOCK_ADDR: {
     unsigned Size = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
     OpdsMapping[0] = AMDGPU::getValueMapping(AMDGPU::SGPRRegBankID, Size);
     break;
   }
   case AMDGPU::G_INSERT: {
     unsigned BankID = isSALUMapping(MI) ? AMDGPU::SGPRRegBankID :
                                           AMDGPU::VGPRRegBankID;

test/CodeGen/AMDGPU/GlobalISel/lds-global-non-entry-func.ll

+; Runs original SDAG test with -global-isel
+
+; RUN: not llc -global-isel -mtriple=amdgcn-amd-amdhsa -o /dev/null < %S/../lds-global-non-entry-func.ll 2>&1 | FileCheck %s
+
+@lds = internal addrspace(3) global float undef, align 4
+
+; FIXME: Selection should succeed
+; CHECK: error: <unknown>:0:0: in function func_use_lds_global void (): local memory global used by non-kernel function
+; CHECK: LLVM ERROR: cannot select: %2:sreg_32_xm0(p3) = G_GLOBAL_VALUE @lds (in function: func_use_lds_global)
+
+define void @func_use_lds_global() {
+  store float 0.0, float addrspace(3)* @lds, align 4
+  ret void
+}

test/CodeGen/AMDGPU/GlobalISel/lds-global-value.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=bonaire -verify-machineinstrs < %s | FileCheck %s
+; TODO: Replace with existing DAG tests
+
+@lds_512_4 = internal unnamed_addr addrspace(3) global [128 x i32] undef, align 4
+@lds_4_8 = addrspace(3) global i32 undef, align 8
+
+define amdgpu_kernel void @use_lds_globals(i32 addrspace(1)* %out, i32 addrspace(3)* %in) #0 {
+; CHECK-LABEL: use_lds_globals:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; CHECK-NEXT:    s_add_u32 s2, 4, 4
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    s_mov_b32 m0, -1
+; CHECK-NEXT:    ds_read_b32 v2, v2
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    s_add_u32 s0, s0, 4
+; CHECK-NEXT:    s_addc_u32 s1, s1, 0
+; CHECK-NEXT:    v_mov_b32_e32 v0, s0
+; CHECK-NEXT:    v_mov_b32_e32 v1, s1
+; CHECK-NEXT:    flat_store_dword v[0:1], v2
+; CHECK-NEXT:    v_mov_b32_e32 v0, 9
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    ds_write_b32 v1, v0
+; CHECK-NEXT:    s_endpgm
+entry:
+  %tmp0 = getelementptr [128 x i32], [128 x i32] addrspace(3)* @lds_512_4, i32 0, i32 1
+  %tmp1 = load i32, i32 addrspace(3)* %tmp0
+  %tmp2 = getelementptr i32, i32 addrspace(1)* %out, i32 1
+  store i32 %tmp1, i32 addrspace(1)* %tmp2
+  store i32 9, i32 addrspace(3)* @lds_4_8
+  ret void
+}
+
+attributes #0 = { nounwind }

test/CodeGen/AMDGPU/GlobalISel/lds-size.ll

+; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa < %S/../lds-size.ll | FileCheck -check-prefix=ALL -check-prefix=HSA %S/../lds-size.ll

test/CodeGen/AMDGPU/GlobalISel/lds-zero-initializer.ll

+; RUN: not llc -global-isel -march=amdgcn -mcpu=tonga < %S/../lds-zero-initializer.ll 2>&1 | FileCheck %s
+
+; FIXME: Select should succeed
+; CHECK: error: <unknown>:0:0: in function load_zeroinit_lds_global void (i32 addrspace(1)*, i1): unsupported initializer for address space
+; CHECK: LLVM ERROR: cannot select: %16:sreg_64(p3) = G_GLOBAL_VALUE @lds (in function: load_zeroinit_lds_global)