[llvm] AMDGPU/PromoteAlloca: Refactor into analysis / commit phases (PR #170512)

[via llvm-commits]

Nicolai =?utf-8?q?Hähnle?= <nicolai.haehnle at amd.com>,
Nicolai =?utf-8?q?Hähnle?= <nicolai.haehnle at amd.com>,
Nicolai =?utf-8?q?Hähnle?= <nicolai.haehnle at amd.com>,
Nicolai =?utf-8?q?Hähnle?= <nicolai.haehnle at amd.com>
Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/170512 at github.com>


================
@@ -237,53 +272,77 @@ FunctionPass *llvm::createAMDGPUPromoteAlloca() {
   return new AMDGPUPromoteAlloca();
 }
 
-static void collectAllocaUses(AllocaInst &Alloca,
-                              SmallVectorImpl<Use *> &Uses) {
-  SmallVector<Instruction *, 4> WorkList({&Alloca});
+bool AMDGPUPromoteAllocaImpl::collectAllocaUses(AllocaAnalysis &AA) const {
+  const auto RejectUser = [&](Instruction *Inst, Twine Msg) {
+    LLVM_DEBUG(dbgs() << "  Cannot promote alloca: " << Msg << "\n"
+                      << "    " << *Inst << "\n");
+    return false;
+  };
+
+  SmallVector<Instruction *, 4> WorkList({AA.Alloca});
   while (!WorkList.empty()) {
     auto *Cur = WorkList.pop_back_val();
+    if (find(AA.Pointers, Cur) != AA.Pointers.end())
----------------
ruiling wrote:

If the `Pointers` is only used for query, can we make it `DenseSet`. Seems like it is only used locally, do we need to put it in the `AA`?

https://github.com/llvm/llvm-project/pull/170512