Support addrspacecast initializers with isNoopAddrSpaceCast

Moves isNoopAddrSpaceCast to the TargetMachine. It logically belongs
with the DataLayout.

include/llvm/CodeGen/BasicTTIImpl.h

@@ -222,7 +222,7 @@ public:
   }
 
   bool isNoopAddrSpaceCast(unsigned FromAS, unsigned ToAS) const {
-    return getTLI()->isNoopAddrSpaceCast(FromAS, ToAS);
+    return getTLI()->getTargetMachine().isNoopAddrSpaceCast(FromAS, ToAS);
   }
 
   Value *rewriteIntrinsicWithAddressSpace(IntrinsicInst *II, Value *OldV,

include/llvm/CodeGen/TargetLowering.h

@@ -1759,17 +1759,10 @@ public:
     return "";
   }
 
-  /// Returns true if a cast between SrcAS and DestAS is a noop.
-  virtual bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const {
-    return false;
-  }
-
   /// Returns true if a cast from SrcAS to DestAS is "cheap", such that e.g. we
   /// are happy to sink it into basic blocks. A cast may be free, but not
   /// necessarily a no-op. e.g. a free truncate from a 64-bit to 32-bit pointer.
-  virtual bool isFreeAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const {
-    return isNoopAddrSpaceCast(SrcAS, DestAS);
-  }
+  virtual bool isFreeAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const;
 
   /// Return true if the pointer arguments to CI should be aligned by aligning
   /// the object whose address is being passed. If so then MinSize is set to the

include/llvm/Target/TargetMachine.h

@@ -271,6 +271,11 @@ public:
     return Options.BBSectionsFuncListBuf.get();
   }
 
+  /// Returns true if a cast between SrcAS and DestAS is a noop.
+  virtual bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const {
+    return false;
+  }
+
   /// Get a \c TargetIRAnalysis appropriate for the target.
   ///
   /// This is used to construct the new pass manager's target IR analysis pass,

lib/CodeGen/AsmPrinter/AsmPrinter.cpp

@@ -2295,6 +2295,16 @@ const MCExpr *AsmPrinter::lowerConstant(const Constant *CV) {
   }
 
   switch (CE->getOpcode()) {
+  case Instruction::AddrSpaceCast: {
+    const Constant *Op = CE->getOperand(0);
+    unsigned DstAS = CE->getType()->getPointerAddressSpace();
+    unsigned SrcAS = Op->getType()->getPointerAddressSpace();
+    if (TM.isNoopAddrSpaceCast(SrcAS, DstAS))
+      return lowerConstant(Op);
+
+    // Fallthrough to error.
+    LLVM_FALLTHROUGH;
+  }
   default: {
     // If the code isn't optimized, there may be outstanding folding
     // opportunities. Attempt to fold the expression using DataLayout as a

lib/CodeGen/CodeGenPrepare.cpp

@@ -4322,7 +4322,7 @@ bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
     unsigned SrcAS
       = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
     unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
-    if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
+    if (TLI.getTargetMachine().isNoopAddrSpaceCast(SrcAS, DestAS))
       return matchAddr(AddrInst->getOperand(0), Depth);
     return false;
   }

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

@@ -6394,7 +6394,7 @@ static void checkAddrSpaceIsValidForLibcall(const TargetLowering *TLI,
                                             unsigned AS) {
   // Lowering memcpy / memset / memmove intrinsics to calls is only valid if all
   // pointer operands can be losslessly bitcasted to pointers of address space 0
-  if (AS != 0 && !TLI->isNoopAddrSpaceCast(AS, 0)) {
+  if (AS != 0 && !TLI->getTargetMachine().isNoopAddrSpaceCast(AS, 0)) {
     report_fatal_error("cannot lower memory intrinsic in address space " +
                        Twine(AS));
   }

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -3425,7 +3425,7 @@ void SelectionDAGBuilder::visitAddrSpaceCast(const User &I) {
   unsigned SrcAS = SV->getType()->getPointerAddressSpace();
   unsigned DestAS = I.getType()->getPointerAddressSpace();
 
-  if (!TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
+  if (!TM.isNoopAddrSpaceCast(SrcAS, DestAS))
     N = DAG.getAddrSpaceCast(getCurSDLoc(), DestVT, N, SrcAS, DestAS);
 
   setValue(&I, N);

lib/CodeGen/TargetLoweringBase.cpp

@@ -801,6 +801,11 @@ bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const {
   }
 }
 
+bool TargetLoweringBase::isFreeAddrSpaceCast(unsigned SrcAS,
+                                             unsigned DestAS) const {
+  return TM.isNoopAddrSpaceCast(SrcAS, DestAS);
+}
+
 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) {
   // If the command-line option was specified, ignore this request.
   if (!JumpIsExpensiveOverride.getNumOccurrences())

lib/Target/AArch64/AArch64ISelLowering.h

@@ -462,12 +462,6 @@ public:
 
   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
 
-  /// Returns true if a cast between SrcAS and DestAS is a noop.
-  bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
-    // Addrspacecasts are always noops.
-    return true;
-  }
-
   /// This method returns a target specific FastISel object, or null if the
   /// target does not support "fast" ISel.
   FastISel *createFastISel(FunctionLoweringInfo &funcInfo,

lib/Target/AArch64/AArch64TargetMachine.h

@@ -57,6 +57,12 @@ public:
                                 SMDiagnostic &Error,
                                 SMRange &SourceRange) const override;
 
+  /// Returns true if a cast between SrcAS and DestAS is a noop.
+  bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
+    // Addrspacecasts are always noops.
+    return true;
+  }
+
 private:
   bool isLittle;
 };

lib/Target/AMDGPU/AMDGPU.h

@@ -281,8 +281,6 @@ enum TargetIndex {
 };
 }
 
-} // End namespace llvm
-
 /// OpenCL uses address spaces to differentiate between
 /// various memory regions on the hardware. On the CPU
 /// all of the address spaces point to the same memory,
@@ -339,4 +337,17 @@ namespace AMDGPUAS {
   };
 }
 
+namespace AMDGPU {
+
+// FIXME: Missing constant_32bit
+inline bool isFlatGlobalAddrSpace(unsigned AS) {
+  return AS == AMDGPUAS::GLOBAL_ADDRESS ||
+         AS == AMDGPUAS::FLAT_ADDRESS ||
+         AS == AMDGPUAS::CONSTANT_ADDRESS ||
+         AS > AMDGPUAS::MAX_AMDGPU_ADDRESS;
+}
+}
+
+} // End namespace llvm
+
 #endif

lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -1677,8 +1677,7 @@ bool AMDGPULegalizerInfo::legalizeAddrSpaceCast(
   const AMDGPUTargetMachine &TM
     = static_cast<const AMDGPUTargetMachine &>(MF.getTarget());
 
-  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
-  if (ST.getTargetLowering()->isNoopAddrSpaceCast(SrcAS, DestAS)) {
+  if (TM.isNoopAddrSpaceCast(SrcAS, DestAS)) {
     MI.setDesc(B.getTII().get(TargetOpcode::G_BITCAST));
     return true;
   }
@@ -2251,8 +2250,7 @@ bool AMDGPULegalizerInfo::legalizeAtomicCmpXChg(
   Register CmpVal = MI.getOperand(2).getReg();
   Register NewVal = MI.getOperand(3).getReg();
 
-  assert(SITargetLowering::isFlatGlobalAddrSpace(
-           MRI.getType(PtrReg).getAddressSpace()) &&
+  assert(AMDGPU::isFlatGlobalAddrSpace(MRI.getType(PtrReg).getAddressSpace()) &&
          "this should not have been custom lowered");
 
   LLT ValTy = MRI.getType(CmpVal);

lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

@@ -3232,7 +3232,7 @@ AMDGPURegisterBankInfo::getValueMappingForPtr(const MachineRegisterInfo &MRI,
   LLT PtrTy = MRI.getType(PtrReg);
   unsigned Size = PtrTy.getSizeInBits();
   if (Subtarget.useFlatForGlobal() ||
-      !SITargetLowering::isFlatGlobalAddrSpace(PtrTy.getAddressSpace()))
+      !AMDGPU::isFlatGlobalAddrSpace(PtrTy.getAddressSpace()))
     return AMDGPU::getValueMapping(AMDGPU::VGPRRegBankID, Size);
 
   // If we're using MUBUF instructions for global memory, an SGPR base register
@@ -3258,8 +3258,7 @@ AMDGPURegisterBankInfo::getInstrMappingForLoad(const MachineInstr &MI) const {
 
   const RegisterBank *PtrBank = getRegBank(PtrReg, MRI, *TRI);
 
-  if (PtrBank == &AMDGPU::SGPRRegBank &&
-      SITargetLowering::isFlatGlobalAddrSpace(AS)) {
+  if (PtrBank == &AMDGPU::SGPRRegBank && AMDGPU::isFlatGlobalAddrSpace(AS)) {
     if (isScalarLoadLegal(MI)) {
       // We have a uniform instruction so we want to use an SMRD load
       ValMapping = AMDGPU::getValueMapping(AMDGPU::SGPRRegBankID, Size);

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -526,6 +526,12 @@ const R600Subtarget *R600TargetMachine::getSubtargetImpl(
   return I.get();
 }
 
+bool AMDGPUTargetMachine::isNoopAddrSpaceCast(unsigned SrcAS,
+                                              unsigned DestAS) const {
+  return AMDGPU::isFlatGlobalAddrSpace(SrcAS) &&
+         AMDGPU::isFlatGlobalAddrSpace(DestAS);
+}
+
 TargetTransformInfo
 R600TargetMachine::getTargetTransformInfo(const Function &F) {
   return TargetTransformInfo(R600TTIImpl(this, F));

lib/Target/AMDGPU/AMDGPUTargetMachine.h

@@ -62,6 +62,8 @@ public:
             AddrSpace == AMDGPUAS::PRIVATE_ADDRESS ||
             AddrSpace == AMDGPUAS::REGION_ADDRESS) ? -1 : 0;
   }
+
+  bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
 };
 
 //===----------------------------------------------------------------------===//

lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

@@ -934,7 +934,10 @@ Value *GCNTTIImpl::rewriteIntrinsicWithAddressSpace(IntrinsicInst *II,
     Type *MaskTy = MaskOp->getType();
 
     bool DoTruncate = false;
-    if (!getTLI()->isNoopAddrSpaceCast(OldAS, NewAS)) {
+
+    const GCNTargetMachine &TM =
+        static_cast<const GCNTargetMachine &>(getTLI()->getTargetMachine());
+    if (!TM.isNoopAddrSpaceCast(OldAS, NewAS)) {
       // All valid 64-bit to 32-bit casts work by chopping off the high
       // bits. Any masking only clearing the low bits will also apply in the new
       // address space.

lib/Target/AMDGPU/SIISelLowering.cpp

@@ -1478,11 +1478,6 @@ EVT SITargetLowering::getOptimalMemOpType(
   return MVT::Other;
 }
 
-bool SITargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
-                                           unsigned DestAS) const {
-  return isFlatGlobalAddrSpace(SrcAS) && isFlatGlobalAddrSpace(DestAS);
-}
-
 bool SITargetLowering::isMemOpHasNoClobberedMemOperand(const SDNode *N) const {
   const MemSDNode *MemNode = cast<MemSDNode>(N);
   const Value *Ptr = MemNode->getMemOperand()->getValue();
@@ -1497,7 +1492,9 @@ bool SITargetLowering::isFreeAddrSpaceCast(unsigned SrcAS,
   if (SrcAS == AMDGPUAS::FLAT_ADDRESS)
     return true;
 
-  return isNoopAddrSpaceCast(SrcAS, DestAS);
+  const GCNTargetMachine &TM =
+      static_cast<const GCNTargetMachine &>(getTargetMachine());
+  return TM.isNoopAddrSpaceCast(SrcAS, DestAS);
 }
 
 bool SITargetLowering::isMemOpUniform(const SDNode *N) const {
@@ -2285,7 +2282,9 @@ SDValue SITargetLowering::LowerFormalArguments(
       if (Arg.Flags.isByRef()) {
         SDValue Ptr = lowerKernArgParameterPtr(DAG, DL, Chain, Offset);
 
-        if (!isNoopAddrSpaceCast(AMDGPUAS::CONSTANT_ADDRESS,
+        const GCNTargetMachine &TM =
+            static_cast<const GCNTargetMachine &>(getTargetMachine());
+        if (!TM.isNoopAddrSpaceCast(AMDGPUAS::CONSTANT_ADDRESS,
                                     Arg.Flags.getPointerAddrSpace())) {
           Ptr = DAG.getAddrSpaceCast(DL, VT, Ptr, AMDGPUAS::CONSTANT_ADDRESS,
                                      Arg.Flags.getPointerAddrSpace());
@@ -8506,7 +8505,7 @@ SDValue SITargetLowering::LowerATOMIC_CMP_SWAP(SDValue Op, SelectionDAG &DAG) co
   unsigned AS = AtomicNode->getAddressSpace();
 
   // No custom lowering required for local address space
-  if (!isFlatGlobalAddrSpace(AS))
+  if (!AMDGPU::isFlatGlobalAddrSpace(AS))
     return Op;
 
   // Non-local address space requires custom lowering for atomic compare

lib/Target/AMDGPU/SIISelLowering.h

@@ -275,15 +275,6 @@ public:
            AS == AMDGPUAS::PRIVATE_ADDRESS;
   }
 
-  // FIXME: Missing constant_32bit
-  static bool isFlatGlobalAddrSpace(unsigned AS) {
-    return AS == AMDGPUAS::GLOBAL_ADDRESS ||
-           AS == AMDGPUAS::FLAT_ADDRESS ||
-           AS == AMDGPUAS::CONSTANT_ADDRESS ||
-           AS > AMDGPUAS::MAX_AMDGPU_ADDRESS;
-  }
-
-  bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
   bool isFreeAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
 
   TargetLoweringBase::LegalizeTypeAction

lib/Target/ARM/ARMISelLowering.h

@@ -528,12 +528,6 @@ class VectorType;
     const TargetRegisterClass *
     getRegClassFor(MVT VT, bool isDivergent = false) const override;
 
-    /// Returns true if a cast between SrcAS and DestAS is a noop.
-    bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
-      // Addrspacecasts are always noops.
-      return true;
-    }
-
     bool shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
                                 unsigned &PrefAlign) const override;
 

lib/Target/ARM/ARMTargetMachine.h

@@ -72,6 +72,12 @@ public:
   }
 
   bool targetSchedulesPostRAScheduling() const override { return true; };
+
+  /// Returns true if a cast between SrcAS and DestAS is a noop.
+  bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
+    // Addrspacecasts are always noops.
+    return true;
+  }
 };
 
 /// ARM/Thumb little endian target machine.

lib/Target/Mips/MipsISelLowering.h

@@ -365,14 +365,6 @@ class TargetRegisterClass;
       return ABI.IsN64() ? Mips::A1_64 : Mips::A1;
     }
 
-    /// Returns true if a cast between SrcAS and DestAS is a noop.
-    bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
-      // Mips doesn't have any special address spaces so we just reserve
-      // the first 256 for software use (e.g. OpenCL) and treat casts
-      // between them as noops.
-      return SrcAS < 256 && DestAS < 256;
-    }
-
     bool isJumpTableRelative() const override {
       return getTargetMachine().isPositionIndependent();
     }

lib/Target/Mips/MipsTargetMachine.h

@@ -63,6 +63,14 @@ public:
     return TLOF.get();
   }
 
+  /// Returns true if a cast between SrcAS and DestAS is a noop.
+  bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
+    // Mips doesn't have any special address spaces so we just reserve
+    // the first 256 for software use (e.g. OpenCL) and treat casts
+    // between them as noops.
+    return SrcAS < 256 && DestAS < 256;
+  }
+
   bool isLittleEndian() const { return isLittle; }
   const MipsABIInfo &getABI() const { return ABI; }
 };

lib/Target/PowerPC/PPCISelLowering.h

@@ -1022,11 +1022,6 @@ namespace llvm {
       }
     };
 
-    bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
-      // Addrspacecasts are always noops.
-      return true;
-    }
-
     bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
                              SelectionDAG &DAG,
                              ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;

lib/Target/PowerPC/PPCTargetMachine.h

@@ -58,6 +58,11 @@ public:
     const Triple &TT = getTargetTriple();
     return (TT.getArch() == Triple::ppc64 || TT.getArch() == Triple::ppc64le);
   };
+
+  bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
+    // Addrspacecasts are always noops.
+    return true;
+  }
 };
 } // end namespace llvm
 

lib/Target/X86/X86ISelLowering.cpp

@@ -2537,17 +2537,6 @@ Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
   return TargetLowering::getSafeStackPointerLocation(IRB);
 }
 
-bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
-                                            unsigned DestAS) const {
-  assert(SrcAS != DestAS && "Expected different address spaces!");
-
-  const TargetMachine &TM = getTargetMachine();
-  if (TM.getPointerSize(SrcAS) != TM.getPointerSize(DestAS))
-    return false;
-
-  return SrcAS < 256 && DestAS < 256;
-}
-
 //===----------------------------------------------------------------------===//
 //               Return Value Calling Convention Implementation
 //===----------------------------------------------------------------------===//

lib/Target/X86/X86ISelLowering.h

@@ -1349,8 +1349,6 @@ namespace llvm {
                                           Align Alignment,
                                           SelectionDAG &DAG) const;
 
-    bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
-
     /// Customize the preferred legalization strategy for certain types.
     LegalizeTypeAction getPreferredVectorAction(MVT VT) const override;
 

lib/Target/X86/X86TargetMachine.cpp

@@ -311,6 +311,14 @@ X86TargetMachine::getSubtargetImpl(const Function &F) const {
   return I.get();
 }
 
+bool X86TargetMachine::isNoopAddrSpaceCast(unsigned SrcAS,
+                                           unsigned DestAS) const {
+  assert(SrcAS != DestAS && "Expected different address spaces!");
+  if (getPointerSize(SrcAS) != getPointerSize(DestAS))
+    return false;
+  return SrcAS < 256 && DestAS < 256;
+}
+
 //===----------------------------------------------------------------------===//
 // X86 TTI query.
 //===----------------------------------------------------------------------===//

lib/Target/X86/X86TargetMachine.h

@@ -54,6 +54,8 @@ public:
   }
 
   bool isJIT() const { return IsJIT; }
+
+  bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override;
 };
 
 } // end namespace llvm

test/CodeGen/AMDGPU/addrspacecast-initializer-unsupported.ll

@@ -0,0 +1,7 @@
+; RUN: not --crash llc -march=amdgcn -verify-machineinstrs < %s 2>&1 | FileCheck -check-prefix=ERROR %s
+
+; ERROR: LLVM ERROR: Unsupported expression in static initializer: addrspacecast ([256 x i32] addrspace(3)* @lds.arr to [256 x i32] addrspace(4)*)
+
+@lds.arr = unnamed_addr addrspace(3) global [256 x i32] undef, align 4
+
+@gv_flatptr_from_lds = unnamed_addr addrspace(2) global i32 addrspace(4)* getelementptr ([256 x i32], [256 x i32] addrspace(4)* addrspacecast ([256 x i32] addrspace(3)* @lds.arr to [256 x i32] addrspace(4)*), i64 0, i64 8), align 4

test/CodeGen/AMDGPU/addrspacecast-initializer.ll

@@ -0,0 +1,27 @@
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck %s
+
+; CHECK: global.arr:
+; CHECK: .zero 1024
+; CHECK: .size global.arr, 1024
+
+; CHECK: gv_flatptr_from_global:
+; CHECK: .quad global.arr+32
+; CHECK: .size gv_flatptr_from_global, 8
+
+; CHECK: gv_global_ptr:
+; CHECK: .quad global.arr+32
+; CHECK: .size gv_global_ptr, 8
+
+; CHECK: gv_flatptr_from_constant:
+; CHECK: .quad constant.arr+32
+; CHECK: .size	gv_flatptr_from_constant, 8
+
+@global.arr = unnamed_addr addrspace(1) global [256 x i32] undef, align 4
+@constant.arr = external unnamed_addr addrspace(4) global [256 x i32], align 4
+
+@gv_flatptr_from_global = unnamed_addr addrspace(4) global i32 addrspace(0)* getelementptr ([256 x i32], [256 x i32] addrspace(0)* addrspacecast ([256 x i32] addrspace(1)* @global.arr to [256 x i32] addrspace(0)*), i64 0, i64 8), align 4
+
+
+@gv_global_ptr = unnamed_addr addrspace(4) global i32 addrspace(1)* getelementptr ([256 x i32], [256 x i32] addrspace(1)* @global.arr, i64 0, i64 8), align 4
+
+@gv_flatptr_from_constant = unnamed_addr addrspace(4) global i32 addrspace(0)* getelementptr ([256 x i32], [256 x i32] addrspace(0)* addrspacecast ([256 x i32] addrspace(4)* @constant.arr to [256 x i32] addrspace(0)*), i64 0, i64 8), align 4