[SPIR-V] Overhaul module analysis to improve translation speed and simplify the underlying logics (llvm#120415)

This PR is to address legacy issues with module analysis that currently
uses a complicated and not so efficient approach to trace dependencies
between SPIR-V id's via a duplicate tracker data structures and an
explicitly built dependency graph. Even a quick performance check
without any specialized benchmarks points to this part of the
implementation as a biggest bottleneck.

This PR specifically:
* eliminates a need to build a dependency graph as a data structure,
* updates the test suite (mainly, by fixing incorrect CHECK's referring
to a hardcoded order of definitions, contradicting the spec requirement
to allow certain definitions to go "in any order", see
https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_logical_layout_of_a_module),
* improves function pointers implementation so that it now passes
EXPENSIVE_CHECKS (thus removing 3 XFAIL's in the test suite).

As a quick sanity check of whether goals of the PR are achieved, we can
measure time of translation for any big LLVM IR. While testing the PR in
the local development environment, improvements of the x5 order have
been observed.

For example, the SYCL test case "group barrier" that is a ~1Mb binary IR
input shows the following values of the naive performance metric that we
can nevertheless apply here to roughly estimate effects of the PR.

before the PR:
```
$ time llc -O0 -mtriple=spirv64v1.6-unknown-unknown _group_barrier_phi.bc -o 1 --filetype=obj

real    3m33.241s
user    3m14.688s
sys     0m18.530s
```

after the PR

```
$ time llc -O0 -mtriple=spirv64v1.6-unknown-unknown _group_barrier_phi.bc -o 1 --filetype=obj

real    0m42.031s
user    0m38.834s
sys     0m3.193s
```

Next work should probably address Duplicate Tracker further, as it needs
analysis now from the perspective of what parts of it are not necessary
now, after changing the approach to implementation of the module
analysis step.

Author: VyacheslavLevytskyy

llvm/lib/Target/SPIRV/CMakeLists.txt

@@ -20,7 +20,6 @@ add_llvm_target(SPIRVCodeGen
   SPIRVCallLowering.cpp
   SPIRVInlineAsmLowering.cpp
   SPIRVCommandLine.cpp
-  SPIRVDuplicatesTracker.cpp
   SPIRVEmitIntrinsics.cpp
   SPIRVGlobalRegistry.cpp
   SPIRVInstrInfo.cpp

llvm/lib/Target/SPIRV/SPIRVAsmPrinter.cpp

@@ -274,7 +274,7 @@ void SPIRVAsmPrinter::emitInstruction(const MachineInstr *MI) {
 }
 
 void SPIRVAsmPrinter::outputModuleSection(SPIRV::ModuleSectionType MSType) {
-  for (MachineInstr *MI : MAI->getMSInstrs(MSType))
+  for (const MachineInstr *MI : MAI->getMSInstrs(MSType))
     outputInstruction(MI);
 }
 
@@ -326,7 +326,7 @@ void SPIRVAsmPrinter::outputOpMemoryModel() {
 void SPIRVAsmPrinter::outputEntryPoints() {
   // Find all OpVariable IDs with required StorageClass.
   DenseSet<Register> InterfaceIDs;
-  for (MachineInstr *MI : MAI->GlobalVarList) {
+  for (const MachineInstr *MI : MAI->GlobalVarList) {
     assert(MI->getOpcode() == SPIRV::OpVariable);
     auto SC = static_cast<SPIRV::StorageClass::StorageClass>(
         MI->getOperand(2).getImm());
@@ -336,14 +336,14 @@ void SPIRVAsmPrinter::outputEntryPoints() {
     // declaring all global variables referenced by the entry point call tree.
     if (ST->isAtLeastSPIRVVer(VersionTuple(1, 4)) ||
         SC == SPIRV::StorageClass::Input || SC == SPIRV::StorageClass::Output) {
-      MachineFunction *MF = MI->getMF();
+      const MachineFunction *MF = MI->getMF();
       Register Reg = MAI->getRegisterAlias(MF, MI->getOperand(0).getReg());
       InterfaceIDs.insert(Reg);
     }
   }
 
   // Output OpEntryPoints adding interface args to all of them.
-  for (MachineInstr *MI : MAI->getMSInstrs(SPIRV::MB_EntryPoints)) {
+  for (const MachineInstr *MI : MAI->getMSInstrs(SPIRV::MB_EntryPoints)) {
     SPIRVMCInstLower MCInstLowering;
     MCInst TmpInst;
     MCInstLowering.lower(MI, TmpInst, MAI);
@@ -381,9 +381,8 @@ void SPIRVAsmPrinter::outputGlobalRequirements() {
 
 void SPIRVAsmPrinter::outputExtFuncDecls() {
   // Insert OpFunctionEnd after each declaration.
-  SmallVectorImpl<MachineInstr *>::iterator
-      I = MAI->getMSInstrs(SPIRV::MB_ExtFuncDecls).begin(),
-      E = MAI->getMSInstrs(SPIRV::MB_ExtFuncDecls).end();
+  auto I = MAI->getMSInstrs(SPIRV::MB_ExtFuncDecls).begin(),
+       E = MAI->getMSInstrs(SPIRV::MB_ExtFuncDecls).end();
   for (; I != E; ++I) {
     outputInstruction(*I);
     if ((I + 1) == E || (*(I + 1))->getOpcode() == SPIRV::OpFunction)

llvm/lib/Target/SPIRV/SPIRVCallLowering.cpp

@@ -418,6 +418,7 @@ bool SPIRVCallLowering::lowerFormalArguments(MachineIRBuilder &MIRBuilder,
                                .addImm(FuncControl)
                                .addUse(GR->getSPIRVTypeID(FuncTy));
   GR->recordFunctionDefinition(&F, &MB.getInstr()->getOperand(0));
+  GR->addGlobalObject(&F, &MIRBuilder.getMF(), FuncVReg);
 
   // Add OpFunctionParameter instructions
   int i = 0;
@@ -431,6 +432,7 @@ bool SPIRVCallLowering::lowerFormalArguments(MachineIRBuilder &MIRBuilder,
         .addUse(GR->getSPIRVTypeID(ArgTypeVRegs[i]));
     if (F.isDeclaration())
       GR->add(&Arg, &MIRBuilder.getMF(), ArgReg);
+    GR->addGlobalObject(&Arg, &MIRBuilder.getMF(), ArgReg);
     i++;
   }
   // Name the function.

llvm/lib/Target/SPIRV/SPIRVDuplicatesTracker.cpp

@@ -211,23 +211,7 @@ class SPIRVGeneralDuplicatesTracker {
   SPIRVDuplicatesTracker<MachineInstr> MT;
   SPIRVDuplicatesTracker<SPIRV::SpecialTypeDescriptor> ST;
 
-  // NOTE: using MOs instead of regs to get rid of MF dependency to be able
-  // to use flat data structure.
-  // NOTE: replacing DenseMap with MapVector doesn't affect overall correctness
-  // but makes LITs more stable, should prefer DenseMap still due to
-  // significant perf difference.
-  using SPIRVReg2EntryTy =
-      MapVector<MachineOperand *, SPIRV::DTSortableEntry *>;
-
-  template <typename T>
-  void prebuildReg2Entry(SPIRVDuplicatesTracker<T> &DT,
-                         SPIRVReg2EntryTy &Reg2Entry,
-                         const SPIRVInstrInfo *TII);
-
 public:
-  void buildDepsGraph(std::vector<SPIRV::DTSortableEntry *> &Graph,
-                      const SPIRVInstrInfo *TII, MachineModuleInfo *MMI);
-
   void add(const Type *Ty, const MachineFunction *MF, Register R) {
     TT.add(unifyPtrType(Ty), MF, R);
   }

llvm/lib/Target/SPIRV/SPIRVDuplicatesTracker.h

@@ -721,6 +721,7 @@ Register SPIRVGlobalRegistry::buildGlobalVariable(
   }
   Reg = MIB->getOperand(0).getReg();
   DT.add(GVar, &MIRBuilder.getMF(), Reg);
+  addGlobalObject(GVar, &MIRBuilder.getMF(), Reg);
 
   // Set to Reg the same type as ResVReg has.
   auto MRI = MIRBuilder.getMRI();

llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.cpp

@@ -89,6 +89,9 @@ class SPIRVGlobalRegistry {
   // Intrinsic::spv_assign_ptr_type instructions.
   DenseMap<Value *, CallInst *> AssignPtrTypeInstr;
 
+  // Maps OpVariable and OpFunction-related v-regs to its LLVM IR definition.
+  DenseMap<std::pair<const MachineFunction *, Register>, const Value *> Reg2GO;
+
   // Add a new OpTypeXXX instruction without checking for duplicates.
   SPIRVType *createSPIRVType(const Type *Type, MachineIRBuilder &MIRBuilder,
                              SPIRV::AccessQualifier::AccessQualifier AQ =
@@ -151,15 +154,17 @@ class SPIRVGlobalRegistry {
     return DT.find(F, MF);
   }
 
-  void buildDepsGraph(std::vector<SPIRV::DTSortableEntry *> &Graph,
-                      const SPIRVInstrInfo *TII,
-                      MachineModuleInfo *MMI = nullptr) {
-    DT.buildDepsGraph(Graph, TII, MMI);
-  }
-
   void setBound(unsigned V) { Bound = V; }
   unsigned getBound() { return Bound; }
 
+  void addGlobalObject(const Value *V, const MachineFunction *MF, Register R) {
+    Reg2GO[std::make_pair(MF, R)] = V;
+  }
+  const Value *getGlobalObject(const MachineFunction *MF, Register R) {
+    auto It = Reg2GO.find(std::make_pair(MF, R));
+    return It == Reg2GO.end() ? nullptr : It->second;
+  }
+
   // Add a record to the map of function return pointer types.
   void addReturnType(const Function *ArgF, TypedPointerType *DerivedTy) {
     FunResPointerTypes[ArgF] = DerivedTy;

llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.h

@@ -47,6 +47,19 @@ bool SPIRVInstrInfo::isConstantInstr(const MachineInstr &MI) const {
   }
 }
 
+bool SPIRVInstrInfo::isSpecConstantInstr(const MachineInstr &MI) const {
+  switch (MI.getOpcode()) {
+  case SPIRV::OpSpecConstantTrue:
+  case SPIRV::OpSpecConstantFalse:
+  case SPIRV::OpSpecConstant:
+  case SPIRV::OpSpecConstantComposite:
+  case SPIRV::OpSpecConstantOp:
+    return true;
+  default:
+    return false;
+  }
+}
+
 bool SPIRVInstrInfo::isInlineAsmDefInstr(const MachineInstr &MI) const {
   switch (MI.getOpcode()) {
   case SPIRV::OpAsmTargetINTEL:

llvm/lib/Target/SPIRV/SPIRVInstrInfo.cpp

@@ -30,6 +30,7 @@ class SPIRVInstrInfo : public SPIRVGenInstrInfo {
   const SPIRVRegisterInfo &getRegisterInfo() const { return RI; }
   bool isHeaderInstr(const MachineInstr &MI) const;
   bool isConstantInstr(const MachineInstr &MI) const;
+  bool isSpecConstantInstr(const MachineInstr &MI) const;
   bool isInlineAsmDefInstr(const MachineInstr &MI) const;
   bool isTypeDeclInstr(const MachineInstr &MI) const;
   bool isDecorationInstr(const MachineInstr &MI) const;

llvm/lib/Target/SPIRV/SPIRVInstrInfo.h

@@ -1117,6 +1117,7 @@ bool SPIRVInstructionSelector::selectMemOperation(Register ResVReg,
                                             Constant::getNullValue(LLVMArrTy));
     Register VarReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
     GR.add(GV, GR.CurMF, VarReg);
+    GR.addGlobalObject(GV, GR.CurMF, VarReg);
 
     Result &=
         BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpVariable))
@@ -3509,18 +3510,25 @@ bool SPIRVInstructionSelector::selectGlobalValue(
         // References to a function via function pointers generate virtual
         // registers without a definition. We will resolve it later, during
         // module analysis stage.
+        Register ResTypeReg = GR.getSPIRVTypeID(ResType);
         MachineRegisterInfo *MRI = MIRBuilder.getMRI();
-        Register FuncVReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
-        MRI->setRegClass(FuncVReg, &SPIRV::iIDRegClass);
-        MachineInstrBuilder MB =
+        Register FuncVReg =
+            MRI->createGenericVirtualRegister(GR.getRegType(ResType));
+        MRI->setRegClass(FuncVReg, &SPIRV::pIDRegClass);
+        MachineInstrBuilder MIB1 =
+            BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
+                .addDef(FuncVReg)
+                .addUse(ResTypeReg);
+        MachineInstrBuilder MIB2 =
             BuildMI(BB, I, I.getDebugLoc(),
                     TII.get(SPIRV::OpConstantFunctionPointerINTEL))
                 .addDef(NewReg)
-                .addUse(GR.getSPIRVTypeID(ResType))
+                .addUse(ResTypeReg)
                 .addUse(FuncVReg);
         // mapping the function pointer to the used Function
-        GR.recordFunctionPointer(&MB.getInstr()->getOperand(2), GVFun);
-        return MB.constrainAllUses(TII, TRI, RBI);
+        GR.recordFunctionPointer(&MIB2.getInstr()->getOperand(2), GVFun);
+        return MIB1.constrainAllUses(TII, TRI, RBI) &&
+               MIB2.constrainAllUses(TII, TRI, RBI);
       }
       return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
           .addDef(NewReg)