[clang] [Clang] Add fake use emission to Clang with -fextend-lifetimes (PR #110102)

[Stephen Tozer via cfe-commits]

================
@@ -1664,6 +1710,17 @@ CodeGenFunction::EmitAutoVarAlloca(const VarDecl &D) {
                                          emission.getOriginalAllocatedAddress(),
                                          emission.getSizeForLifetimeMarkers());
 
+  // Analogous to lifetime markers, we use a 'cleanup' to emit fake.use
+  // calls for local variables. We are exempting volatile variables and
+  // non-scalars larger than 4 times the size of an unsigned int (32 bytes).
+  // Larger non-scalars are often allocated in memory and may create unnecessary
+  // overhead.
+  if (CGM.getCodeGenOpts().ExtendLifetimes) {
+    if (extendLifetime(getContext(), CurCodeDecl, D, CXXABIThisDecl))
+      EHStack.pushCleanup<FakeUse>(NormalFakeUse,
+                                   emission.getAllocatedAddress());
----------------
SLTozer wrote:

Looks like the difference is about allocas in a non-default address space - essentially we'll get either:
```
  %i = alloca i32, align 4, addrspace(5)
  %i.ascast = addrspacecast ptr addrspace(5) %i to ptr
  store i32 %someval, ptr %i.ascast, align 4
; getAllocatedAddress
  %fake.use = load i32, ptr %i.ascast, align 4
  call void (...) @llvm.fake.use(i32 %fake.use)
; getOriginalAllocatedAddress
  %fake.use = load i32, ptr addrspace(5) %i, align 4
  call void (...) @llvm.fake.use(i32 %fake.use)
```
In practice, both of them result in extended variable lifetimes (the fake use will continue to track `%someval` in both cases), so I _think_ this is unimportant. On the one hand, the pointer that we use to store and load is `%i.ascast`, so it seems consistent for the fake use to use that; on the other hand, if a `#dbg_declare` is created, it will refer to `ptr addrspace(5) %i`, so that could be more appropriate for the fake use. In summary: I don't know, and I don't think it matters.

https://github.com/llvm/llvm-project/pull/110102