[llvm-branch-commits] [llvm] [AMDGPU] Allocate and reserve registers for address space 13 objects (PR #214248)

Matt Arsenault via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Mon Nov 2 20:53:04 PST 2026


================
@@ -386,6 +420,20 @@ bool AMDGPUPromoteAllocaImpl::run(Function &F, bool PromoteToLDS) {
       LLVM_DEBUG(dbgs() << "Analyzing: " << *AI << '\n');
 
       AllocaAnalysis AA{AI};
+
+      // An alloca already in the VGPR address space is not promoted, only
+      // given a place there. Required for functionality, so it runs whether
+      // or not optimizations are enabled.
+      if (AI->getAddressSpace() == AMDGPUAS::VGPR) {
+        if (IsLatePass)
+          Allocas.push_back(std::move(AA));
+        continue;
+      }
----------------
arsenm wrote:

There are more bugs here but I think it's critical this feature not touch promote alloca, other than to inform promote alloca that it should hold off on using more VGPRs for the occupancy target 

https://github.com/llvm/llvm-project/pull/214248


More information about the llvm-branch-commits mailing list