Optimize Specializations of Variadic Generics by Eliminating Parameter Packs

[aidan-hall]

rdar://161182876

There are two key barriers that prevent variadic generic specialisations from being optimised in most cases:

• Indirect packs in the calling convention: These force the caller to make a large number of stack allocations, since the pack is passed indirectly, and pack elements are currently always pointers, so every element must be stack-allocated separately from the pack.
• Pack loops and open_pack_element: These perform one iteration per pack element, so the loop body must be generic, with witness_method instructions that cannot be eliminated, since the type they are called on is not known.

Targeted modifications to eliminate these barriers will allow other passes to fully optimise variadic generic specialisations.

This PR only addresses the first barrier, indirect packs as function arguments.

The simplest way to eliminate indirect packs from the calling convention is effectively to move the stack allocations for the packs and pack elements inside the callee, and use the addresses of these local allocations in the main body, which is otherwise unchanged. These allocations can be eliminated by other passes once dynamic pack loops are optimised.

This transformation is performed by a function pass (currently called pack-specialization). For example, it would transform the following specialisation like so:

sil shared @$s11addTogether8doubleUp2xsxxQp_xxQptxxQp_tRvzlFSi_SdSSSJQP_Tg5 : $@convention(thin) (@pack_guaranteed Pack{Int, Double, String, Character}) -> (@pack_out Pack{Int, Double, String, Character}, @pack_out Pack{Int, Double, String, Character}) {
bb0(%0 : $*Pack{Int, Double, String, Character}, %1 : $*Pack{Int, Double, String, Character}, %2 : $*Pack{Int, Double, String, Character}):
  debug_value %2, let, name "xs", argno 1, expr op_deref
  …

sil shared @$s11addTogether8doubleUp2xsxxQp_xxQptxxQp_tRvzlFSi_SdSSSJQP_Tg5Tf8xxx_n : $@convention(thin) (Int, Double, @guaranteed String, @guaranteed Character) -> (Int, Double, @owned String, @owned Character, Int, Double, @owned String, @owned Character) {
bb0(%0 : $Int, %1 : $Double, %2 : $String, %3 : $Character):
  %4 = alloc_pack $Pack{Int, Double, String, Character}
  %5 = scalar_pack_index 0 of $Pack{Int, Double, String, Character}
  %6 = alloc_stack $Int
  store %0 to %6
  …
  debug_value %4, let, name "xs", argno 1, expr op_deref
  …

Note that the return values are all now direct, so bb0 only needs arguments corresponding to the xs parameter pack.

To optimise a pack loop, it must be unrolled. The pack element type is then statically known in each unrolled instance of the body. The existing loop unrolling pass can consistently do this in variadic generic specialisations, since the (known constant) pack length is used as the number of iterations. This still leaves constructs like the following:

  %5 = integer_literal $Builtin.Word, 0
  %6 = dynamic_pack_index %5 of $Pack{Int, Double}
  %7 = open_pack_element %6 of <each A where repeat each A : Numeric> at <Pack{Int, Double}>, shape $each A, uuid "063BDCF4-98A3-11F0-B2E8-929C59BC796C"
  %20 = witness_method $@pack_element("063BDCF4-98A3-11F0-B2E8-929C59BC796C") each A, #Numeric."*" …
  %21 = apply %20<@pack_element("063BDCF4-98A3-11F0-B2E8-929C59BC796C") each A>(%10, %14, %18, %11) … // type-defs: %7

In this context, we could easily replace all instances of the @pack_element type with Int. This would allow the appropriate witness method to be inlined. We will perform this in an instruction simplification for open_pack_element.

[aidan-hall]

@swift-ci smoke test macOS

[aidan-hall]

@swift-ci smoke test macOS

[eeckstein]

I think it makes sense to move (most of) these APIs to TypeProperties so that they are also available for AST.Type and AST.CanonicalType. The SILType implementation forwards them to the AST type, anyway

[eeckstein]

looks great!

[aidan-hall]

@swift-ci test

[aidan-hall]

@swift-ci smoke test linux

[aidan-hall]

@swift-ci smoke test macos

[aidan-hall]

@swift-ci smoke test

[aidan-hall]

@swift-ci smoke test macos

[aidan-hall]

@swift-ci smoke test linux

Many changes have been made since this approval.

[aidan-hall]

*a few

[aidan-hall]

@swift-ci test windows

[aidan-hall]

@swift-ci please smoke benchmark

[aidan-hall]

@swift-ci smoke test macos

[aidan-hall]

@swift-ci smoke test macos

[aidan-hall]

@swift-ci smoke test linux

[aidan-hall]

@swift-ci test windows

[aidan-hall]

@swift-ci test macos

[aidan-hall]

@swift-ci test windows

[aidan-hall]

@swift-ci smoke test

[aidan-hall]

@swift-ci smoke test macos

[aidan-hall]

@swift-ci Please Test Source Compatibility Release

[aidan-hall]

@swift-ci test macos

[aidan-hall]

@swift-ci test macos