Lowered LLVM is preventing inlining

[bcardosolopes]

In clang/test/CIR/CodeGen/tbaa-vptr.cpp:

The LLVM IR produced by OG is as follows:

define dso_local void @_ZN1BD0Ev(ptr noundef nonnull align 8 dereferenceable(9) initializes((0, 8)) %this) unnamed_addr #0 align 2 {
entry:
  store ptr getelementptr inbounds nuw inrange(-16, 16) (i8, ptr @_ZTV1B, i64 16), ptr %this, align 8, !tbaa !5
  %0 = getelementptr inbounds nuw i8, ptr %this, i64 8
  tail call void @_ZN6MemberD1Ev(ptr noundef nonnull align 1 dereferenceable(1) %0) #3
  tail call void @_ZN1AD2Ev(ptr noundef nonnull align 8 dereferenceable(9) %this) #3
  tail call void @_ZdlPvm(ptr noundef %this, i64 noundef 16) #4
  ret void
}

In contrast, the LLVM IR generated by ClangIR is as follows:

define dso_local void @_ZN1BD0Ev(ptr %0) #1 {
  tail call void @_ZN1BD1Ev(ptr %0) #0
  tail call void @_ZdlPvm(ptr %0, i64 16)
  ret void
}

Is there something missing in the ClangIR?

Originally posted by @PikachuHyA in #1463 (comment)

[AdUhTkJm]

The behaviour seems to be correct, just that the destructor is not inlined. It might be more obvious if we demangle the names.

This is simplified output from OG with -O0:

; non-deleting destructor
define dso_local void @B::~B()(ptr %0) {
  %2 = alloca ptr, align 8
  store ptr %0, ptr %2, align 8
  %3 = load ptr, ptr %2, align 8
  store ptr getelementptr ({ [4 x ptr] }, ptr @vtable for B, i32 0, i32 0, i32 2), ptr %3
  %4 = getelementptr i8, ptr %3, i64 8
  call void @Member::~Member()(ptr %4)
  call void @A::~A()(ptr %3)
  ret void
}

; deleting destructor
define dso_local void @B::~B()(ptr %0) {
  %2 = alloca ptr, align 8
  store ptr %0, ptr %2, align 8
  %3 = load ptr, ptr %2, align 8
  call void @B::~B()(ptr %3) ; this is the non-deleting one
  call void @operator delete(void*, unsigned long)(ptr %3, i64 16)
  ret void
}

As we can see, the OG output in the example comes from inlining the non-deleting destructor into the deleting one:

define dso_local void @_ZN1BD0Ev(ptr %this) {
entry:
  store ptr getelementptr (i8, ptr @_ZTV1B, i64 16), ptr %this, align 8
  %0 = getelementptr inbounds nuw i8, ptr %this, i64 8
  tail call void @_ZN6MemberD1Ev(ptr %0) ; Member::~Member()
  tail call void @_ZN1AD2Ev(ptr %this) ; A::~A()
  tail call void @_ZdlPvm(ptr %this, i64 16) ; operator delete(void *, unsigned long)
  ret void
}

While the Clang IR one does not perform the inline:

define dso_local void @_ZN1BD0Ev(ptr %0) #1 {
  tail call void @_ZN1BD1Ev(ptr %0) ; B::~B(), which is the non-deleting one
  tail call void @_ZdlPvm(ptr %0, i64 16) ; operator delete(void *, unsigned long)
  ret void
}

(cc @PikachuHyA)

[bcardosolopes]

Thanks for checking. Seems like CIR -> LLVM lowering is probably lacking some attributes that would make the inliner kick in - the idea is that LLVM should be able to continue doing same opts after lowered to CIR.

[yimdx]

I would like to work on this to learn ClangIR!

[siya100]

hey @bcardosolopes i want to work on this issue to delve into ClangIR

[bcardosolopes]

I'm fine with whoever submits the PR, thanks!