[ExpandLargeDivRem] Scalarize vector types. (#86959)

expand-large-divrem cannot handle vector types.
If overly large vector element types survive into
isel, they will likely be scalarized there, but since
isel cannot handle scalar integer types of that size,
it will assert.

Handle vector types in expand-large-divrem by
scalarizing them and then expanding the scalar type
operation. For large vectors, this results in a
*massive* code expansion, but it's better than
asserting.

Author: bevin-hansson

llvm/lib/CodeGen/ExpandLargeDivRem.cpp

@@ -54,8 +54,32 @@ static bool isSigned(unsigned int Opcode) {
   return Opcode == Instruction::SDiv || Opcode == Instruction::SRem;
 }
 
+static void scalarize(BinaryOperator *BO,
+                      SmallVectorImpl<BinaryOperator *> &Replace) {
+  VectorType *VTy = cast<FixedVectorType>(BO->getType());
+
+  IRBuilder<> Builder(BO);
+
+  unsigned NumElements = VTy->getElementCount().getFixedValue();
+  Value *Result = PoisonValue::get(VTy);
+  for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
+    Value *LHS = Builder.CreateExtractElement(BO->getOperand(0), Idx);
+    Value *RHS = Builder.CreateExtractElement(BO->getOperand(1), Idx);
+    Value *Op = Builder.CreateBinOp(BO->getOpcode(), LHS, RHS);
+    Result = Builder.CreateInsertElement(Result, Op, Idx);
+    if (auto *NewBO = dyn_cast<BinaryOperator>(Op)) {
+      NewBO->copyIRFlags(Op, true);
+      Replace.push_back(NewBO);
+    }
+  }
+  BO->replaceAllUsesWith(Result);
+  BO->dropAllReferences();
+  BO->eraseFromParent();
+}
+
 static bool runImpl(Function &F, const TargetLowering &TLI) {
   SmallVector<BinaryOperator *, 4> Replace;
+  SmallVector<BinaryOperator *, 4> ReplaceVector;
   bool Modified = false;
 
   unsigned MaxLegalDivRemBitWidth = TLI.getMaxDivRemBitWidthSupported();
@@ -71,16 +95,23 @@ static bool runImpl(Function &F, const TargetLowering &TLI) {
     case Instruction::SDiv:
     case Instruction::URem:
     case Instruction::SRem: {
-      // TODO: This doesn't handle vectors.
-      auto *IntTy = dyn_cast<IntegerType>(I.getType());
+      // TODO: This pass doesn't handle scalable vectors.
+      if (I.getOperand(0)->getType()->isScalableTy())
+        continue;
+
+      auto *IntTy = dyn_cast<IntegerType>(I.getType()->getScalarType());
       if (!IntTy || IntTy->getIntegerBitWidth() <= MaxLegalDivRemBitWidth)
         continue;
 
       // The backend has peephole optimizations for powers of two.
+      // TODO: We don't consider vectors here.
       if (isConstantPowerOfTwo(I.getOperand(1), isSigned(I.getOpcode())))
         continue;
 
-      Replace.push_back(&cast<BinaryOperator>(I));
+      if (I.getOperand(0)->getType()->isVectorTy())
+        ReplaceVector.push_back(&cast<BinaryOperator>(I));
+      else
+        Replace.push_back(&cast<BinaryOperator>(I));
       Modified = true;
       break;
     }
@@ -89,6 +120,11 @@ static bool runImpl(Function &F, const TargetLowering &TLI) {
     }
   }
 
+  while (!ReplaceVector.empty()) {
+    BinaryOperator *BO = ReplaceVector.pop_back_val();
+    scalarize(BO, Replace);
+  }
+
   if (Replace.empty())
     return false;