[RFC] Proper lowering of constant alloca operations

[Lancern]

This is related to #866 .

Since #892 , ClangIR lowers constant local variables in C/C++ to cir.alloca operations with a const flag. The presence of the const flag implies:

• The variable must be initialized by a following cir.store operation, and
• All cir.load operation that loads the cir.alloca result must produce the value stored by the cir.store operation.

An obvious optimization here is that we could eliminate all the loads and replace the loaded values with the stored initial value. LLVM already implements similar optimizations, but we need to tweak the generated LLVM IR to teach LLVM to apply those optimizations. I'm proposing several approaches here that could lead to such optimizations in LLVM, and hope we could choose one that best fits our needs.

Approach 1: use the llvm.invariant.start intrinsic

The first approach would be using the llvm.invariant.start and the llvm.invariant.end intrinsic. This pair of intrinsics tell the optimizer that a specified memory location will never change within the region bounded by the intrinsics. With this approach, the following CIR:

cir.func @test(@init: !s32i) {
  %0 = cir.alloca !s32i, !cir.ptr<!s32i>, ["var", init, const]
  cir.store @init, %0 : !s32i, !cir.ptr<!s32i>

  // example uses of %0
  %1 = cir.load %0 : !cir.ptr<!s32i>, !s32i
  cir.call @clobber(%1) : (!cir.ptr<!s32i>) -> ()
  %2 = cir.load %0 : !cir.ptr<!s32i>, !s32i

  cir.return
}

would generate the following LLVM IR:

define dso_local void @test(i32 %init) {
  %1 = alloca i32, align 4
  store i32 %init, ptr %1, align 4
  %inv = call ptr @llvm.invariant.start(ptr %1, i64 4)

  // example uses of %0
  %2 = load i32, ptr %1, align 4
  call void @clobber(ptr %1)
  %3 = load i32, ptr %1, align 4

  call void @llvm.invariant.end(ptr %inv, i64 4, ptr %1)
  ret void
}

Theoretically, the optimizer would be able to at least fold %3 into %2. Ironically, it seems that the optimizer refuses to optimize if the llvm.invariant.end intrinsic call is present, see https://godbolt.org/z/5dMv7T77e. To bypass this limitation, simply remove the call to the llvm.invariant.end intrinsic, and the optimizer works as expected.

Approach 2: use the !invariant.load metadata

A load instruction could have an !invariant.load metadata attached. The LLVM language reference says:

If a load instruction tagged with the !invariant.load metadata is executed, the memory location referenced by the load has to contain the same value at all points in the program where the memory location is dereferenceable; otherwise, the behavior is undefined.

With this approach, the CIR snippet listed earlier would emit the following LLVM IR:

define dso_local void @test(i32 %init) {
  %1 = alloca i32, align 4
  store i32 %init, ptr %1, align 4

  ; example uses of %1
  %2 = load i32, ptr %1, align 4, !invariant.load !0
  call void @clobber(ptr %1)
  %3 = load i32, ptr %1, align 4, !invariant.load !0

  ret void
}

!0 = !{}

The optimizer could then fold both load instructions to just %init, see https://godbolt.org/z/Exnh85zhx.

It's worth mentioning here that the !invariant.load metadata is already supported by the MLIR LLVMIR dialect.

Approach 3: use the !invariant.group metadata

A load instruction or a store instruction could have an !invariant.group metadata attached. Unlike !invariant.load, the !invariant.group only requires that every value loaded or stored by such instructions must be the same if those instructions load or store to the same pointer. With this approach, the CIR snippet listed earlier would emit the following LLVM IR:

define dso_local void @test(i32 %init) {
  %1 = alloca i32, align 4
  store i32 %init, ptr %1, align 4, !invariant.group !0

  ; example uses of %1
  %2 = load i32, ptr %1, align 4, !invariant.group !0
  call void @clobber(ptr %1)
  %3 = load i32, ptr %1, align 4, !invariant.group !0

  ret void
}

!0 = !{}

The optimizer could then fold both load instructions to just %init, see https://godbolt.org/z/8MsxcoqTY.

Constant local variables in inner scopes

Let's consider a slightly more complex example:

void test(std::vector<int> vec) {
  for (const int item : vec)
    do_something(item);
}

Upon each iteration, the local variable item would reuse the same memory location. But ideally we would like to still teach LLVM that item is constant during a single iteration. The second approach is infeasible since the value in the memory location changes between iterations. Thus only the first approach and the third approach is suitable for such a case.

The first approach would emit code like this:

define dso_local void @test() {
  %item = alloca i32, align 4
  ; ...
loop.body:
  store i32 %loop.ind, ptr %item, align 4
  %inv = call ptr @llvm.invariant.start(ptr %item, i64 4)
  
  ; loop body goes here, an example load instruction below
  %1 = load i32, ptr %item, align 4
  
  call void @llvm.invariant.end(ptr %inv, i64 4, ptr %1)
  br label %loop.header
}

The third approach would emit code like this:

define dso_local void @test() {
  %item = alloca i32, align 4
  ; ...
loop.body:
  %item.0 = phi ptr [ %item, %0 ], [ %item.launder, %loop.body ]
  store i32 %loop.ind, ptr %item.0, align 4, !invariant.group !0
  
  ; loop body goes here, an example load instruction below
  %1 = load i32, ptr %item.0, align 4, !invariant.group !0
  
  %item.launder = call @llvm.launder.invariant.group(ptr %item.0)
  br label %loop.body
}

!0 = !{}

The call to the llvm.launder.invariant.group intrinsic makes sure that each iteration creates a "distinct invariant group". Without this intrinsic call, the optimizer could assume that the load and store instructions would load and store the same value across all iterations of the loop.

So what do you think about these 3 lowering approaches? Or do you know any other approaches that this proposal does not mention?

[bcardosolopes]

Very nice write up, thanks for considering diferent options.

But ideally we would like to still teach LLVM that item is constant during a single iteration.

Neat!

So what do you think about these 3 lowering approaches? Or do you know any other approaches that this proposal does not mention?

The option (3) looks more appealing to me since (1) has the weird end behavior. Do you know if there's a trend in LLVM to pursue one direction more than the other? That could be one good deciding criteria, but both look good to me.

This is a good use of a CIR abstraction that allow us to give extra information to LLVM, cool to see work in this direction!

On a tagencial subject, should this be done in -O1? Or do we want to emit this even in -O0?

[Lancern]

Do you know if there's a trend in LLVM to pursue one direction more than the other? That could be one good deciding criteria, but both look good to me.

This is a good point. Actually the metadata and intrinsics involved in approach 3 is still experimental, while the llvm.invariant.start intrinsics are already there for years. Considering that the first approach still has such a strange behavior regarding llvm.invariant.end after years of development, I would prefer approach 3 too.

On a tagencial subject, should this be done in -O1? Or do we want to emit this even in -O0?

Since this is technically an optimization, I think we should emit these stuff under -O1 or higher.

[bcardosolopes]

Great, happy to see PRs in either direction. Should we add a -cc1 flag to control this? It could be enabled by default but a no version would be cool (this would be handy for later experiments to see how much perf % we get out of it).

Considering that the first approach still has such a strange behavior regarding llvm.invariant.end after years of development

One more idea: ask on LLVM discourse if folks know why this happens?

[Lancern]

Should we add a -cc1 flag to control this?

Sounds good to me!

One more idea: ask on LLVM discourse if folks know why this happens?

Great idea. Opened a thread here: https://discourse.llvm.org/t/why-llvm-invariant-end-intrinsic-would-prevent-optimizations/82984