[PATCH] D45228: AMDGPU/SI: Handle BitCast of GEP in promoting alloca to vector

[Changpeng Fang via Phabricator via llvm-commits]

cfang added inline comments.


================
Comment at: lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp:340-341
+  }
+  case Instruction::BitCast: {
+    if (isa<GetElementPtrInst>(Used)) {
+      for (User *BCUser : Inst->users()) {
----------------
arsenm wrote:
> This shouldn't need to special case GEP users. You really want something like stripPointerCasts, but not one that looks through addrspacecast.
> 
> Could use a test with 0 index GEPs which should show the same problem.
I could not understand this comment.

Here we are handling the case of load(bitcast(gep))!
We are checking the bitcase and collect the load. We will transfer to vector load later, which depends on the gep for the index.
So we have to have the gep, and collect the load/store.


https://reviews.llvm.org/D45228