[llvm] [InstCombine] Convert load from LUT into a select (PR #98339)

[Nikita Popov via llvm-commits]

================
@@ -1003,6 +1003,106 @@ static bool canSimplifyNullLoadOrGEP(LoadInst &LI, Value *Op) {
   return false;
 }
 
+static Value *foldLoadFromIndexedGlobal(LoadInst &LI, IRBuilderBase &Builder,
+                                        TargetLibraryInfo &TLI) {
+  if (LI.isVolatile())
+    return nullptr;
+
+  auto *GEP = dyn_cast<GetElementPtrInst>(LI.getPointerOperand());
+  if (!GEP)
+    return nullptr;
+
+  auto *GV = dyn_cast<GlobalVariable>(GEP->getPointerOperand());
+  if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
+    return nullptr;
+
+  Constant *Init = GV->getInitializer();
+  auto &DL = LI.getDataLayout();
+
+  uint64_t IndexBW = DL.getIndexTypeSizeInBits(GEP->getType());
+  APInt ConstOffset(IndexBW, 0);
+  MapVector<Value *, APInt> VariableOffsets;
+  if (!GEP->collectOffset(DL, IndexBW, VariableOffsets, ConstOffset))
+    return nullptr;
+
+  if (!ConstOffset.isZero() || VariableOffsets.size() != 1)
+    return nullptr;
+
+  auto &Step = VariableOffsets.front().second;
+  if (Step.isNonPositive())
+    return nullptr;
+  uint64_t ArraySize = DL.getTypeAllocSize(Init->getType()).getFixedValue();
+  // Don't blow up on huge arrays.
+  // This threshold is chosen based on statistics on a dataset
+  // which is collected from real-world applications.
+  constexpr uint64_t MaxArraySize = 16;
+  if (ArraySize > MaxArraySize * Step.getZExtValue())
+    return nullptr;
+
+  Value *Index = VariableOffsets.front().first;
+  if (Index->getType()->getScalarSizeInBits() != IndexBW)
+    return nullptr;
+
+  Type *LoadTy = LI.getType();
+  SmallMapVector<Constant *, uint64_t, 2> ValueMap;
----------------
nikic wrote:

`// Map from Value to single index where it occurs or MultiMapIdx if it occurs multiple times.`

https://github.com/llvm/llvm-project/pull/98339