[llvm] [InstCombine] Canonicalize `gep T, (gep i8, base, C1), (Index +nsw C2)` into `gep T, (gep i8, base, C1 + C2 * sizeof(T)), Index` (PR #76177)

[Nikita Popov via llvm-commits]

================
@@ -2038,6 +2038,54 @@ static Instruction *foldSelectGEP(GetElementPtrInst &GEP,
   return SelectInst::Create(Cond, NewTrueC, NewFalseC, "", nullptr, Sel);
 }
 
+// Canonicalization:
+// gep T, (gep i8, base, C1), (Index +nsw C2) into
+// gep T, (gep i8, base, C1 + C2 * sizeof(T)), Index
+static Instruction *canonicalizeGEPOfConstGEPI8(GetElementPtrInst &GEP,
+                                                GEPOperator *Src,
+                                                InstCombinerImpl &IC) {
+  if (GEP.getNumIndices() != 1)
+    return nullptr;
+  auto &DL = IC.getDataLayout();
+  if (!Src->getSourceElementType()->isIntegerTy(8) ||
+      !Src->hasAllConstantIndices())
+    return nullptr;
+  Value *VarIndex;
+  const APInt *C2;
+  Type *PtrTy = Src->getType()->getScalarType();
+  unsigned IndexSizeInBits = DL.getIndexTypeSizeInBits(PtrTy);
+  if (!(GEP.getOperand(1)->getType()->getScalarSizeInBits() >=
+            IndexSizeInBits &&
+        match(GEP.getOperand(1), m_Add(m_Value(VarIndex), m_APInt(C2)))) &&
+      !match(GEP.getOperand(1),
+             m_SExtOrSelf(
+                 m_CombineOr(m_NSWAdd(m_Value(VarIndex), m_APInt(C2)),
+                             m_DisjointOr(m_Value(VarIndex), m_APInt(C2))))))
+    return nullptr;
+  Type *BaseType = GEP.getSourceElementType();
+  APInt C1(IndexSizeInBits, 0);
+  // Add the offset for Src (which is fully constant).
+  if (!Src->accumulateConstantOffset(DL, C1))
+    return nullptr;
+  APInt TypeSize(IndexSizeInBits, DL.getTypeAllocSize(BaseType));
+  bool Overflow = false;
+  APInt C3 = TypeSize.smul_ov(C2->sext(TypeSize.getBitWidth()), Overflow);
----------------
nikic wrote:

I don't think the overflow checks are needed (for correctness)?

https://github.com/llvm/llvm-project/pull/76177