[llvm] improve performance of Module Analysis stage in the part of processing "other instructions" (PR #76047)

[Noel Grandin via llvm-commits]

================
@@ -319,15 +291,43 @@ void SPIRVModuleAnalysis::collectFuncNames(MachineInstr &MI,
   }
 }
 
+using InstrSignature = SmallVector<size_t>;
+using InstrTraces = std::set<InstrSignature>;
+
+// Returns a representation of an instruction as a vector of MachineOperand
+// hash values, see llvm::hash_value(const MachineOperand &MO) for details.
+// This creates a signature of the instruction with the same content
+// that MachineOperand::isIdenticalTo uses for comparison.
+static InstrSignature instrToSignature(MachineInstr &MI,
+                                       SPIRV::ModuleAnalysisInfo &MAI) {
+  InstrSignature Signature;
+  for (unsigned i = 0; i < MI.getNumOperands(); ++i) {
+    const MachineOperand &MO = MI.getOperand(i);
+    size_t h;
+    if (MO.isReg()) {
+      Register RegAlias = MAI.getRegisterAlias(MI.getMF(), MO.getReg());
+      // mimic llvm::hash_value(const MachineOperand &MO)
+      h = hash_combine(MO.getType(), (unsigned)RegAlias, MO.getSubReg(),
+                       MO.isDef());
+    } else {
+      h = hash_value(MO);
+    }
+    Signature.push_back(h);
+  }
+  return Signature;
+}
+
 // Collect the given instruction in the specified MS. We assume global register
 // numbering has already occurred by this point. We can directly compare reg
 // arguments when detecting duplicates.
 static void collectOtherInstr(MachineInstr &MI, SPIRV::ModuleAnalysisInfo &MAI,
-                              SPIRV::ModuleSectionType MSType,
+                              SPIRV::ModuleSectionType MSType, InstrTraces &IS,
                               bool Append = true) {
   MAI.setSkipEmission(&MI);
-  if (findSameInstrInMS(MI, MSType, MAI))
+  InstrSignature MISign = instrToSignature(MI, MAI);
+  if (auto FoundMI = IS.find(MISign); FoundMI != IS.end())
     return; // Found a duplicate, so don't add it.
+  IS.insert(MISign);
----------------
grandinj wrote:

faster to go:

    auto FoundMI = IS.insert(MISign);
    if (!FoundMI.second)
        return; // insert failed, so we found a duplicate

and avoid doing set lookup twice


https://github.com/llvm/llvm-project/pull/76047