[llvm] [AMDGPU] Fix a potential use-after-erase in `AMDGPUPromoteAlloca` pass (PR #174529)

[Shilei Tian via llvm-commits]

================
@@ -0,0 +1,34 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes=amdgpu-promote-alloca %s | FileCheck %s
+
+define amdgpu_kernel void @kernel(ptr readonly captures(none) %p) local_unnamed_addr {
+; CHECK-LABEL: define amdgpu_kernel void @kernel(
+; CHECK-SAME: ptr readonly captures(none) [[P:%.*]]) local_unnamed_addr {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    [[ASCAST:%.*]] = addrspacecast ptr [[P]] to ptr addrspace(1)
+; CHECK-NEXT:    [[ALLOCA:%.*]] = freeze <14 x i32> poison
+; CHECK-NEXT:    [[LOAD:%.*]] = load i32, ptr addrspace(1) [[ASCAST]], align 4
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[LOAD]], 0
+; CHECK-NEXT:    br i1 [[CMP]], label %[[BB_1:.*]], label %[[BB_2:.*]]
+; CHECK:       [[BB_2]]:
+; CHECK-NEXT:    br label %[[BB_1]]
+; CHECK:       [[BB_1]]:
+; CHECK-NEXT:    [[PROMOTEALLOCA:%.*]] = phi <14 x i32> [ [[ALLOCA]], %[[BB_2]] ], [ [[ALLOCA]], %[[ENTRY]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <14 x i32> [[PROMOTEALLOCA]], i32 0, i32 0
+; CHECK-NEXT:    ret void
+;
+entry:
+  %ascast = addrspacecast ptr %p to ptr addrspace(1)
+  %alloca = alloca [14 x i32], addrspace(5)
+  %load = load i32, ptr addrspace(1) %ascast
+  %cmp = icmp eq i32 %load, 0
+  br i1 %cmp, label %bb.1, label %bb.2
+
+bb.2:
+  store i32 poison, ptr addrspace(5) %alloca
----------------
shiltian wrote:

Aha, that's a great deal!

https://github.com/llvm/llvm-project/pull/174529