[clang] [CIR] Implement CIRBasicAliasAnalysis::getUnderlyingObject (PR #215683)

Andy Kaylor via cfe-commits cfe-commits at lists.llvm.org
Thu Aug 13 09:54:50 PDT 2026


================
@@ -20,17 +21,117 @@ using namespace cir;
 // Helpers
 //===----------------------------------------------------------------------===//
 
+static constexpr unsigned MaxLookupDepth = 6;
+
 mlir::Value CIRBasicAliasAnalysis::getUnderlyingObject(mlir::Value val) {
   LDBG() << "Getting underlying object for: " << val;
 
-  // TODO: Walk through cir.ptr_stride, cir.cast, cir.get_member, etc.
-  // to find the root allocation (cir.alloca, cir.global_addr, function args).
-  LDBG() << "Not yet implemented";
+  for (unsigned depth = 0; depth < MaxLookupDepth; ++depth) {
+    mlir::Operation *defOp = val.getDefiningOp();
+    if (!defOp) {
+      LDBG() << "No defining operation, stopping";
+      break; // Block argument (e.g. function parameter) — stop here.
+    }
+
+    // Bitcast and address-space casts don't change the underlying object.
+    // array_to_ptrdecay produces an element pointer to the same storage as
+    // the array pointer, so strip through it too.
+    if (auto castOp = mlir::dyn_cast<cir::CastOp>(defOp)) {
+      if (castOp.isAllocaPreservingCast() ||
+          castOp.getKind() == cir::CastKind::array_to_ptrdecay) {
+        LDBG() << "Walking past cast operation";
+        val = castOp.getSrc();
+        continue;
+      }
+      LDBG() << "Opaque cast operation, stopping";
+      break;
+    }
+
+    // Pointer stride: only strip through when we can prove the access stays
----------------
andykaylor wrote:

I'm going to add offset handling and bounds checking in my next PR in this series. This comment is an artifact of my having already explored that. It should have been removed. I found a strange thing with LLVM alias analysis here. If the pointer stride goes beyond the end of the alloca object, BasicAA says it doesn't alias with anything.

https://godbolt.org/z/hYqno8ea8

Apparently, this is intentional, saying the out-of-bounds offset is UB.

I think in the current state it doesn't make sense to have a shared helper to detect the zero offset because the helper would still need op-specific code to get the offset. When I introduce the more complete offset handling this will naturally become `offset == 0` for everything.

https://github.com/llvm/llvm-project/pull/215683


More information about the cfe-commits mailing list