test/SILOptimizer/redundant_load_elim_with_casts.sil

// RUN: %target-sil-opt -enable-sil-verify-all %s -module-name Swift -redundant-load-elim | %FileCheck -check-prefix=CHECK-FUTURE %s
//
// FIXME: Contains tests which are handled by old RLE, but not current one. Mostly due to casting. Eventually we probably should
// handle these cases if they turn out to be important.
//
// The problem is the current RLE uses type projection tree/path to model fields in an object. This makes it difficult for it to
// reason about casts, i.e. 1 memory with different types.
//
// Tracked by rdar://23023366

import Builtin

struct A {
  var i : Builtin.Int32
}

struct A2 {
  var i : Builtin.Int32
}

struct AA {
  var a : A
  var i : Builtin.Int32
}

class B {
  var i : Builtin.Int32
  init()
}

struct X {
  var c : B
  init()
}

enum Optional<T> {
  case none
  case some(T)
}

class E : B { }

struct C {
  var i : Builtin.Int16
}

struct D {
  var p : Builtin.RawPointer
}

sil @use : $@convention(thin) (Builtin.Int32) -> ()
sil @use_a : $@convention(thin) (@in A) -> ()
sil @escaped_a_ptr : $@convention(thin) () -> @out A
sil @escaped_a : $@convention(thin) () -> Builtin.RawPointer


struct Agg2 {
  var t : (Builtin.Int64, Builtin.Int32)
}

struct Agg1 {
  var a : Agg2
}

struct Wrapper {
  var value : Builtin.Int32
}

// CHECK-FUTURE: sil @tbaa_class_alias_nonclass
// CHECK: strong_retain [[RET:%[0-9]+]]
// CHECK: strong_retain [[RET]]
// CHECK: return
sil @tbaa_class_alias_nonclass : $@convention(thin) (@owned B, @inout Agg1) -> () {
bb0(%0 : $B, %1 : $*Agg1):
  %2 = alloc_box $<τ_0_0> { var τ_0_0 } <B>
  %2a = project_box %2 : $<τ_0_0> { var τ_0_0 } <B>, 0
  %3 = load %1 : $*Agg1
  store %3 to %1 : $*Agg1
  %5 = load %2a : $*B
  store %3 to %1 : $*Agg1
  %7 = load %2a : $*B
  strong_retain %5 : $B   //%7 and %5 should really be one load.
  strong_retain %7 : $B
  %10 = tuple()
  return %10 : $()
}

// FIXME: When RLE uses TBAA it should remove the second load.
// CHECK-FUTURE: sil @tbaa_struct
// CHECK: load
// CHECK: store
// CHECK: load
// CHECK: return
sil @tbaa_struct : $@convention(thin) (Builtin.RawPointer, A2) -> (A, A) {
bb0(%0 : $Builtin.RawPointer, %1 : $A2):
  %2 = pointer_to_address %0 : $Builtin.RawPointer to [strict] $*A
  %3 = load %2 : $*A
  %5 = pointer_to_address %0 : $Builtin.RawPointer to [strict] $*A2
  store %1 to %5 : $*A2
  %20 = load %2 : $*A
  %30 = tuple(%3 : $A, %20 : $A)
  return %30 : $(A, A)
}

// Even with TBAA, RLE should not remove the second load.
// CHECK-FUTURE: sil @tbaa_bind_memory
// CHECK: load
// CHECK: bind_memory
// CHECK: store
// CHECK: bind_memory
// CHECK: load
// CHECK: return
sil @tbaa_bind_memory : $@convention(thin) (Builtin.RawPointer, A2) -> (A, A) {
bb0(%0 : $Builtin.RawPointer, %1 : $A2):
  %2 = pointer_to_address %0 : $Builtin.RawPointer to [strict] $*A
  %3 = load %2 : $*A
  %4 = integer_literal $Builtin.Word, 1
  bind_memory %0 : $Builtin.RawPointer, %4 : $Builtin.Word to $A2
  %5 = pointer_to_address %0 : $Builtin.RawPointer to [strict] $*A2
  store %1 to %5 : $*A2
  bind_memory %0 : $Builtin.RawPointer, %4 : $Builtin.Word to $A
  %20 = load %2 : $*A
  %30 = tuple(%3 : $A, %20 : $A)
  return %30 : $(A, A)
}

// *NOTE* This does not handle raw pointer since raw pointer is only layout compatible with heap references.
// CHECK-FUTURE: sil @store_to_load_forward_unchecked_addr_cast_struct : $@convention(thin) (Optional<A>) -> () {
// CHECK: bb0([[INPUT:%[0-9]+]]
// CHECK-NEXT: [[LOCAL:%[0-9]+]] = alloc_stack
// CHECK: unchecked_trivial_bit_cast [[INPUT]] : $Optional<A> to $Builtin.Int32
// CHECK: unchecked_trivial_bit_cast [[INPUT]] : $Optional<A> to $A
// CHECK: unchecked_addr_cast [[LOCAL]] : $*Optional<A> to $*Builtin.RawPointer
// CHECK: unchecked_addr_cast [[LOCAL]] : $*Optional<A> to $*Builtin.NativeObject
// CHECK: unchecked_trivial_bit_cast [[INPUT]] : $Optional<A> to $Optional<Builtin.Int32>
// CHECK: unchecked_addr_cast [[LOCAL]] : $*Optional<A> to $*Optional<Builtin.RawPointer>
// CHECK: unchecked_addr_cast [[LOCAL]] : $*Optional<A> to $*Optional<Builtin.NativeObject>
// CHECK: return
sil @store_to_load_forward_unchecked_addr_cast_struct : $@convention(thin) (Optional<A>) -> () {
bb0(%0 : $Optional<A>):
  %1 = alloc_stack $Optional<A>
  store %0 to %1 : $*Optional<A>
  %2 = unchecked_addr_cast %1 : $*Optional<A> to $*Builtin.Int32
  %3 = load %2 : $*Builtin.Int32
  %4 = unchecked_addr_cast %1 : $*Optional<A> to $*A
  %5 = load %4 : $*A
  %6 = unchecked_addr_cast %1 : $*Optional<A> to $*Builtin.RawPointer
  %7 = load %6 : $*Builtin.RawPointer
  %8 = unchecked_addr_cast %1 : $*Optional<A> to $*Builtin.NativeObject
  %9 = load %8 : $*Builtin.NativeObject
  %10 = unchecked_addr_cast %1 : $*Optional<A> to $*Optional<Builtin.Int32>
  %11 = load %10 : $*Optional<Builtin.Int32>
  %12 = unchecked_addr_cast %1 : $*Optional<A> to $*Optional<Builtin.RawPointer>
  %13 = load %12 : $*Optional<Builtin.RawPointer>
  %14 = unchecked_addr_cast %1 : $*Optional<A> to $*Optional<Builtin.NativeObject>
  %15 = load %14 : $*Optional<Builtin.NativeObject>
  dealloc_stack %1 : $*Optional<A>
  %9999 = tuple()
  return %9999 : $()
}

struct IntPtr {
  var Val: Builtin.Int32
  var Ptr: Builtin.RawPointer
}

// unchecked_addr_cast must not be promoted unless the size of the
// source type is known to be greater or equal to the size of the
// destination type.
//
// CHECK-FUTURE: sil @store_to_load_forward_unchecked_addr_cast_nopromote : $@convention(thin) (@inout IntPtr) -> () {
// CHECK: bb0([[INPUT:%[0-9]+]] : $*IntPtr)
// CHECK-NEXT: [[LOCAL:%[0-9]+]] = alloc_stack
// CHECK: [[CAST:%[0-9]+]] = unchecked_addr_cast [[INPUT]] : $*IntPtr to $*Builtin.Int32
// CHECK: store %{{.*}} to [[CAST]]
// CHECK: unchecked_addr_cast [[CAST]] : $*Builtin.Int32 to $*IntPtr
// CHECK: return
sil @store_to_load_forward_unchecked_addr_cast_nopromote : $@convention(thin) (@inout IntPtr) -> () {
bb0(%0 : $*IntPtr):
  %1 = alloc_stack $Builtin.Int32
  %2 = unchecked_addr_cast %0 : $*IntPtr to $*Builtin.Int32
  %3 = integer_literal $Builtin.Int32, 3
  store %3 to %2 : $*Builtin.Int32
  %5 = unchecked_addr_cast %2 : $*Builtin.Int32 to $*IntPtr
  %6 = load %5 : $*IntPtr
  dealloc_stack %1 : $*Builtin.Int32
  %9999 = tuple()
  return %9999 : $()
}

// *NOTE* This does not handle raw pointer since raw pointer is layout
// compatible with heap references, but does not have reference
// semantics b/c it is a trivial type. We currently do not handle such a case.

// CHECK-FUTURE: sil @store_to_load_forward_unchecked_addr_cast_class : $@convention(thin) (Optional<B>) -> () {
// CHECK: bb0([[INPUT:%[0-9]+]]
// CHECK-NEXT: [[LOCAL:%[0-9]+]] = alloc_stack
// CHECK: unchecked_trivial_bit_cast [[INPUT]] : $Optional<B> to $Builtin.Int32
// CHECK: unchecked_ref_cast [[INPUT]] : $Optional<B> to $B
// CHECK: unchecked_trivial_bit_cast [[INPUT]] : $Optional<B> to $Builtin.RawPointer

// CHECK: unchecked_ref_cast [[INPUT]] : $Optional<B> to $Builtin.NativeObject
// CHECK: unchecked_trivial_bit_cast [[INPUT]] : $Optional<B> to $Optional<Builtin.Int32>
// CHECK: unchecked_trivial_bit_cast [[INPUT]] : $Optional<B> to $Optional<Builtin.RawPointer>
// CHECK: unchecked_ref_cast [[INPUT]] : $Optional<B> to $Optional<Builtin.NativeObject>
// CHECK: return
sil @store_to_load_forward_unchecked_addr_cast_class : $@convention(thin) (Optional<B>) -> () {
bb0(%0 : $Optional<B>):
  %1 = alloc_stack $Optional<B>
  store %0 to %1 : $*Optional<B>
  %2 = unchecked_addr_cast %1 : $*Optional<B> to $*Builtin.Int32
  %3 = load %2 : $*Builtin.Int32
  %4 = unchecked_addr_cast %1 : $*Optional<B> to $*B
  %5 = load %4 : $*B
  %6 = unchecked_addr_cast %1 : $*Optional<B> to $*Builtin.RawPointer
  %7 = load %6 : $*Builtin.RawPointer
  %8 = unchecked_addr_cast %1 : $*Optional<B> to $*Builtin.NativeObject
  %9 = load %8 : $*Builtin.NativeObject
  %10 = unchecked_addr_cast %1 : $*Optional<B> to $*Optional<Builtin.Int32>
  %11 = load %10 : $*Optional<Builtin.Int32>
  %12 = unchecked_addr_cast %1 : $*Optional<B> to $*Optional<Builtin.RawPointer>
  %13 = load %12 : $*Optional<Builtin.RawPointer>
  %14 = unchecked_addr_cast %1 : $*Optional<B> to $*Optional<Builtin.NativeObject>
  %15 = load %14 : $*Optional<Builtin.NativeObject>
  dealloc_stack %1 : $*Optional<B>
  %9999 = tuple()
  return %9999 : $()
}

// *NOTE* This does not handle raw pointer since raw pointer is only layout compatible with heap references.
// CHECK-FUTURE: sil @load_to_load_forward_unchecked_addr_cast_struct : $@convention(thin) (@inout Optional<A>) -> () {
// CHECK: bb0(

// CHECK: unchecked_trivial_bit_cast {{%[0-9]+}} : $Optional<A> to $Builtin.Int32
// CHECK: unchecked_trivial_bit_cast {{%[0-9]+}} : $Optional<A> to $A
// CHECK: unchecked_addr_cast %{{.*}} : $*Optional<A> to $*Builtin.RawPointer
// CHECK: unchecked_addr_cast %{{.*}} : $*Optional<A> to $*Builtin.NativeObject
// CHECK: unchecked_trivial_bit_cast {{%[0-9]+}} : $Optional<A> to $Optional<Builtin.Int32>
// CHECK: unchecked_addr_cast {{%[0-9]+}} : $*Optional<A> to $*Optional<Builtin.RawPointer>
// CHECK: unchecked_addr_cast
sil @load_to_load_forward_unchecked_addr_cast_struct : $@convention(thin) (@inout Optional<A>) -> () {
bb0(%0 : $*Optional<A>):
  %1 = load %0 : $*Optional<A>
  %2 = unchecked_addr_cast %0 : $*Optional<A> to $*Builtin.Int32
  %3 = load %2 : $*Builtin.Int32
  %4 = unchecked_addr_cast %0 : $*Optional<A> to $*A
  %5 = load %4 : $*A
  %6 = unchecked_addr_cast %0 : $*Optional<A> to $*Builtin.RawPointer
  %7 = load %6 : $*Builtin.RawPointer
  %8 = unchecked_addr_cast %0 : $*Optional<A> to $*Builtin.NativeObject
  %9 = load %8 : $*Builtin.NativeObject
  %10 = unchecked_addr_cast %0 : $*Optional<A> to $*Optional<Builtin.Int32>
  %11 = load %10 : $*Optional<Builtin.Int32>
  %12 = unchecked_addr_cast %0 : $*Optional<A> to $*Optional<Builtin.RawPointer>
  %13 = load %12 : $*Optional<Builtin.RawPointer>
  %14 = unchecked_addr_cast %0 : $*Optional<A> to $*Optional<Builtin.NativeObject>
  %15 = load %14 : $*Optional<Builtin.NativeObject>
  %9999 = tuple()
  return %9999 : $()
 }

// *NOTE* This does not handle raw pointer since raw pointer is layout
// compatible with heap references, but does not have reference
// semantics b/c it is a trivial type. We currently do not handle such a case.

// CHECK-FUTURE: sil @load_to_load_forward_unchecked_addr_cast_class : $@convention(thin) (@inout Optional<B>) -> () {
// CHECK: bb0({{%[0-9]+}} : $*Optional<B>):
// CHECK: unchecked_trivial_bit_cast {{%[0-9]+}} : $Optional<B> to $Builtin.Int32
// CHECK: unchecked_ref_bit_cast {{%[0-9]+}} : $Optional<B> to $B
// CHECK: unchecked_trivial_bit_cast {{%[0-9]+}} : $Optional<B> to $Builtin.RawPointer
// CHECK: unchecked_ref_bit_cast {{%[0-9]+}} : $Optional<B> to $Builtin.NativeObject
// CHECK: unchecked_trivial_bit_cast {{%[0-9]+}} : $Optional<B> to $Optional<Builtin.Int32>
// CHECK: unchecked_trivial_bit_cast {{%[0-9]+}} : $Optional<B> to $Optional<Builtin.RawPointer>
// CHECK: unchecked_ref_bit_cast {{%[0-9]+}} : $Optional<B> to $Optional<Builtin.NativeObject>
sil @load_to_load_forward_unchecked_addr_cast_class : $@convention(thin) (@inout Optional<B>) -> () {
bb0(%0 : $*Optional<B>):
  %1 = load %0 : $*Optional<B>
  %2 = unchecked_addr_cast %0 : $*Optional<B> to $*Builtin.Int32
  %3 = load %2 : $*Builtin.Int32
  %4 = unchecked_addr_cast %0 : $*Optional<B> to $*B
  %5 = load %4 : $*B
  %6 = unchecked_addr_cast %0 : $*Optional<B> to $*Builtin.RawPointer
  %7 = load %6 : $*Builtin.RawPointer
  %8 = unchecked_addr_cast %0 : $*Optional<B> to $*Builtin.NativeObject
  %9 = load %8 : $*Builtin.NativeObject
  %10 = unchecked_addr_cast %0 : $*Optional<B> to $*Optional<Builtin.Int32>
  %11 = load %10 : $*Optional<Builtin.Int32>
  %12 = unchecked_addr_cast %0 : $*Optional<B> to $*Optional<Builtin.RawPointer>
  %13 = load %12 : $*Optional<Builtin.RawPointer>
  %14 = unchecked_addr_cast %0 : $*Optional<B> to $*Optional<Builtin.NativeObject>
  %15 = load %14 : $*Optional<Builtin.NativeObject>
  %9999 = tuple()
  return %9999 : $()
}

// Don't bitcast differently sized structs.
// CHECK-FUTURE: sil @store_to_load_forward_unchecked_addr_cast_different_sized_struct
// CHECK-NOT: unchecked_trivial_bit_cast
// CHECK: return
sil @store_to_load_forward_unchecked_addr_cast_different_sized_struct : $@convention(thin) (C) -> () {
bb0(%0 : $C):
  %1 = alloc_stack $C
  store %0 to %1 : $*C
  %2 = unchecked_addr_cast %1 : $*C to $*A
  %3 = load %2 : $*A
  dealloc_stack %1 : $*C
  %9999 = tuple()
  return %9999 : $()
}


/// Make sure that we don't crash and don't optimize here.
// CHECK-FUTURE: sil @covering_store_with_unchecked_addr : $@convention(thin) (C, C) -> () {
// CHECK-NOT: unchecked_trivial_bit_cast
// CHECK: unchecked_addr_cast
// CHECK-NOT: unchecked_trivial_bit_cast
sil @covering_store_with_unchecked_addr : $@convention(thin) (C, C) -> () {
bb0(%0 : $C, %1 : $C):
  %2 = alloc_stack $C
  cond_br undef, bb1, bb2

bb1:
  store %0 to %2 : $*C
  br bb3

bb2:
  store %0 to %2 : $*C
  br bb3

bb3:
  %3 = unchecked_addr_cast %2 : $*C to $*A
  %4 = load %3 : $*A
  dealloc_stack %2 : $*C
  %9999 = tuple()
  return %9999 : $()
}

/// Make sure that we properly invalidate %3 in the following situation.
///
/// 1. We store %3 into the load map.
/// 2. We see that we are storing in %4 something we just loaded meaning that we
/// would have a dead store. We delete that store and through recursion delete %3
/// since %3's only use is %4.
/// 3. When we delete %3, we do not remove it from the load list.
/// 4. %5 can write to memory, so we try to check if it can alias %0#1. We look
/// up the load that was erased and will use it in a memory unsafe way.
//
// CHECK-FUTURE: sil @invalidate_dead_loads_with_only_store_user_correctly : $@convention(thin) () -> () {
// CHECK-NOT: load
// CHECK-NOT: {{%.*}} = store
sil @invalidate_dead_loads_with_only_store_user_correctly : $@convention(thin) () -> () {
bb0:
  %0 = alloc_stack $Optional<Builtin.Int32>
  %1 = integer_literal $Builtin.Int32, 0
  %2 = enum $Optional<Builtin.Int32>, #Optional.some!enumelt.1, %1 : $Builtin.Int32
  %3 = load %0 : $*Optional<Builtin.Int32>
  store %3 to %0 : $*Optional<Builtin.Int32>
  %5 = unchecked_take_enum_data_addr %0 : $*Optional<Builtin.Int32>, #Optional.some!enumelt.1
  dealloc_stack %0 : $*Optional<Builtin.Int32>
  %9999 = tuple()
  return %9999 : $()
}

class Empty {}
struct HoldsRef { @_hasStorage var c: Empty }

sil @mutator : $@convention(method) (@inout HoldsRef) -> ()

// Ensure we don't forward the stored value from the switch_enum
// branches past the inout appearance of the stored-to location.
// CHECK-FUTURE: sil @bad_store_forward
sil @bad_store_forward : $@convention(thin) (@guaranteed Optional<HoldsRef>) -> () {
// CHECK: bb0
bb0(%0 : $Optional<HoldsRef>):
// CHECK: [[LOC:%.*]] = alloc_stack
  %1 = alloc_stack $HoldsRef
  switch_enum %0 : $Optional<HoldsRef>, case #Optional.some!enumelt.1: bb1, case #Optional.none!enumelt: bb2

// CHECK: bb1([[ARG:%.*]] : ${{.*}}):
bb1(%3 : $HoldsRef):
// CHECK: store [[ARG]] to [[LOC]]
  store %3 to %1 : $*HoldsRef
// CHECK-NOT: br bb3(
  br bb3

// CHECK-NOT: bb2(
bb2:
  %6 = alloc_ref $Empty
// CHECK: [[STRUCT:%.*]] = struct
  %7 = struct $HoldsRef (%6 : $Empty)
// CHECK: store [[STRUCT]] to [[LOC]]
  store %7 to %1 : $*HoldsRef
// CHECK-NOT: br bb3(
  br bb3

// CHECK-NOT: bb3(
// CHECK: bb3
bb3:
// CHECK: [[FNREF:%.*]] = function_ref
  %10 = function_ref @mutator : $@convention(method) (@inout HoldsRef) -> ()
  retain_value %0 : $Optional<HoldsRef>
// CHECK: apply [[FNREF]]([[LOC]])
  %12 = apply %10(%1) : $@convention(method) (@inout HoldsRef) -> ()
// CHECK: [[ELEM:%.*]] = struct_element_addr [[LOC]]
  %13 = struct_element_addr %1 : $*HoldsRef, #HoldsRef.c
// CHECK: [[REF:%.*]] = load [[ELEM]]
  %14 = load %13 : $*Empty
// CHECK: strong_release [[REF]]
  strong_release %14 : $Empty
  %16 = tuple ()
  dealloc_stack %1 : $*HoldsRef
  return %16 : $()
}


// We internally use a map vector to represent stores. This means that when we iterate over
// the stores it should be in insertion order. Use this to test whether or not we only set
// the no-dependency bit if we do not forward a load.
// CHECK-FUTURE: sil @test_unchecked_addr_cast_3
// CHECK: bb0([[ARG1:%.*]] : $D, [[ARG2:%.*]] : $D):
// CHECK-NEXT: [[BOX1:%.*]] = alloc_stack $D
// CHECK-NEXT: [[BOX2:%.*]] = alloc_stack $D
// CHECK-NEXT: store [[ARG2]] to [[BOX2]] : $*D
// CHECK-NEXT: [[RESULT:%.*]] = unchecked_trivial_bit_cast [[ARG2]]
// CHECK-NEXT: store [[ARG2]] to [[BOX1]] : $*D
// CHECK-NEXT: dealloc_stack
// CHECK-NEXT: dealloc_stack
// CHECK-NEXT: return [[RESULT]]
sil @test_unchecked_addr_cast_3 : $@convention(thin) (D, D) -> Builtin.RawPointer {
bb0(%x : $D, %y : $D):
  %1 = alloc_stack $D
  %2 = alloc_stack $D
  store %x to %1 : $*D
  store %y to %2 : $*D
  %3 = unchecked_addr_cast %2 : $*D to $*Builtin.RawPointer
  %l1 = load %3 : $*Builtin.RawPointer
  store %y to %1 : $*D
  dealloc_stack %2 : $*D
  dealloc_stack %1 : $*D
  return %l1 : $Builtin.RawPointer
}


typealias I32 = Builtin.Int32

// Promote unchecked_addr_cast of tuples.
// (A, B, C) -> (A, B) is safe
// ((A, B), C) -> (A, B) is safe
// ((A, B), C) -> (A, B, C) is NOT safe
//
// CHECK-FUTURE: sil @unchecked_addr_cast_tuple_promote
// CHECK: bb0(%0 : $*(Builtin.Int32, Builtin.Int32, Builtin.Int32), %1 : $*((Builtin.Int32, Builtin.Int32), Builtin.Int32)):
// CHECK: load %0 : $*(Builtin.Int32, Builtin.Int32, Builtin.Int32)
// CHECK: alloc_stack $(Builtin.Int32, Builtin.Int32, Builtin.Int32)
// CHECK: store %{{.*}} to %{{.*}}#1 : $*(Builtin.Int32, Builtin.Int32, Builtin.Int32)
// CHECK: unchecked_trivial_bit_cast %{{.*}} : $(Builtin.Int32, Builtin.Int32, Builtin.Int32) to $(Builtin.Int32, Builtin.Int32)
// CHECK: tuple_extract %{{.*}} : $(Builtin.Int32, Builtin.Int32), 0
// CHECK: unchecked_trivial_bit_cast %{{.*}} : $(Builtin.Int32, Builtin.Int32, Builtin.Int32) to $(Builtin.Int32, Builtin.Int32)
// CHECK: tuple_extract %{{.*}} : $(Builtin.Int32, Builtin.Int32), 1
// CHECK: dealloc_stack %{{.*}}#0 : $*(Builtin.Int32, Builtin.Int32, Builtin.Int32)
// CHECK: load %1 : $*((Builtin.Int32, Builtin.Int32), Builtin.Int32)
// CHECK: alloc_stack $((Builtin.Int32, Builtin.Int32), Builtin.Int32)
// CHECK: store %{{.*}} to %{{.*}}#1 : $*((Builtin.Int32, Builtin.Int32), Builtin.Int32)
// CHECK: unchecked_trivial_bit_cast %{{.*}} : $((Builtin.Int32, Builtin.Int32), Builtin.Int32) to $(Builtin.Int32, Builtin.Int32)
// CHECK: tuple_extract %{{.*}} : $(Builtin.Int32, Builtin.Int32), 0
// CHECK: unchecked_trivial_bit_cast %{{.*}} : $((Builtin.Int32, Builtin.Int32), Builtin.Int32) to $(Builtin.Int32, Builtin.Int32)
// CHECK: tuple_extract %{{.*}} : $(Builtin.Int32, Builtin.Int32), 1
// CHECK: dealloc_stack %{{.*}}#0 : $*((Builtin.Int32, Builtin.Int32), Builtin.Int32)
// CHECK: alloc_stack $((Builtin.Int32, Builtin.Int32), Builtin.Int32)
// CHECK: store %{{.*}} to %{{.*}}#1 : $*((Builtin.Int32, Builtin.Int32), Builtin.Int32)
// CHECK: unchecked_addr_cast %{{.*}}#1 : $*((Builtin.Int32, Builtin.Int32), Builtin.Int32) to $*(Builtin.Int32, Builtin.Int32, Builtin.Int32)
// CHECK: tuple_element_addr %{{.*}} : $*(Builtin.Int32, Builtin.Int32, Builtin.Int32), 0
// CHECK: tuple_element_addr %{{.*}} : $*(Builtin.Int32, Builtin.Int32, Builtin.Int32), 1
// CHECK: tuple_element_addr %{{.*}} : $*(Builtin.Int32, Builtin.Int32, Builtin.Int32), 2
// CHECK: load %{{.*}} : $*Builtin.Int32
// CHECK: load %{{.*}} : $*Builtin.Int32
// CHECK: load %{{.*}} : $*Builtin.Int32
// CHECK: dealloc_stack %{{.*}}#0 : $*((Builtin.Int32, Builtin.Int32), Builtin.Int32)
// CHECK: return %{{.*}} : $((Builtin.Int32, Builtin.Int32), (Builtin.Int32, Builtin.Int32), (Builtin.Int32, Builtin.Int32, Builtin.Int32))
sil @unchecked_addr_cast_tuple_promote : $@convention(thin) (@inout (I32, I32, I32), @inout ((I32, I32), I32)) -> ((I32, I32), (I32, I32), (I32, I32, I32)) {
bb0(%0 : $*(I32, I32, I32), %1 : $*((I32, I32), I32)):
  %2 = load %0 : $*(I32, I32, I32)
  %3 = alloc_stack $(I32, I32, I32)
  store %2 to %3 : $*(I32, I32, I32)
  %5 = unchecked_addr_cast %3 : $*(I32, I32, I32) to $*(I32, I32)
  %6 = tuple_element_addr %5 : $*(I32, I32), 0
  %7 = tuple_element_addr %5 : $*(I32, I32), 1
  %8 = load %6 : $*I32
  %9 = load %7 : $*I32
  dealloc_stack %3 : $*(I32, I32, I32)
  %11 = load %1 : $*((I32, I32), I32)
  %12 = alloc_stack $((I32, I32), I32)
  store %11 to %12 : $*((I32, I32), I32)
  %14 = unchecked_addr_cast %12 : $*((I32, I32), I32) to $*(I32, I32)
  %15 = tuple_element_addr %14 : $*(I32, I32), 0
  %16 = tuple_element_addr %14 : $*(I32, I32), 1
  %17 = load %15 : $*I32
  %18 = load %16 : $*I32
  dealloc_stack %12 : $*((I32, I32), I32)
  %20 = alloc_stack $((I32, I32), I32)
  store %11 to %20 : $*((I32, I32), I32)
  %22 = unchecked_addr_cast %20 : $*((I32, I32), I32) to $*(I32, I32, I32)
  %23 = tuple_element_addr %22 : $*(I32, I32, I32), 0
  %24 = tuple_element_addr %22 : $*(I32, I32, I32), 1
  %25 = tuple_element_addr %22 : $*(I32, I32, I32), 2
  %26 = load %23 : $*I32
  %27 = load %24 : $*I32
  %28 = load %25 : $*I32
  dealloc_stack %20 : $*((I32, I32), I32)
  %30 = tuple (%8 : $I32, %9 : $I32)
  %31 = tuple (%17 : $I32, %18 : $I32)
  %32 = tuple (%26 : $I32, %27 : $I32, %28 : $I32)
  %33 = tuple (%30 : $(I32, I32), %31 : $(I32, I32), %32 : $(I32, I32, I32))
  return %33 : $((I32, I32), (I32, I32), (I32, I32, I32))
}

// Promote unchecked_addr_cast of tuple elements.
// (A, B, C) -> (A, B) is safe
// ((A, B), C) -> (A, B) is safe
// ((A, B), C) -> (A, B, C) is NOT safe
//
// CHECK-FUTURE: sil @forward_tuple_elements
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: unchecked_addr_cast
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: unchecked_addr_cast
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: unchecked_addr_cast
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: tuple_element_addr
// CHECK: return
//
// FIXME: LoadStore optimization should be able to forward these
// stores but cannot see through projections on the store side.  (I
// decided not to brute force fix this, because it may be better to
// change the canonical form of tuple load/store in this case).
sil @forward_tuple_elements : $@convention(thin) (@inout (I32, I32, I32), @inout ((I32, I32), I32)) -> ((I32, I32), (I32, I32), (I32, I32, I32)) {
bb0(%0 : $*(I32, I32, I32), %1 : $*((I32, I32), I32)):
  %4 = tuple_element_addr %0 : $*(I32, I32, I32), 0
  %5 = load %4 : $*I32
  %6 = tuple_element_addr %0 : $*(I32, I32, I32), 1
  %7 = load %6 : $*I32
  %8 = tuple_element_addr %0 : $*(I32, I32, I32), 2
  %9 = load %8 : $*I32
  %10 = alloc_stack $(I32, I32, I32)
  %11 = tuple_element_addr %10 : $*(I32, I32, I32), 0
  %12 = tuple_element_addr %10 : $*(I32, I32, I32), 1
  %13 = tuple_element_addr %10 : $*(I32, I32, I32), 2
  store %5 to %11 : $*I32
  store %7 to %12 : $*I32
  store %9 to %13 : $*I32
  %17 = unchecked_addr_cast %10 : $*(I32, I32, I32) to $*(I32, I32)
  %18 = tuple_element_addr %17 : $*(I32, I32), 0
  %19 = tuple_element_addr %17 : $*(I32, I32), 1
  %20 = load %18 : $*I32
  %21 = load %19 : $*I32
  dealloc_stack %10 : $*(I32, I32, I32)
  %23 = tuple_element_addr %1 : $*((I32, I32), I32), 0
  %24 = tuple_element_addr %23 : $*(I32, I32), 0
  %25 = load %24 : $*I32
  %26 = tuple_element_addr %23 : $*(I32, I32), 1
  %27 = load %26 : $*I32
  %28 = tuple_element_addr %1 : $*((I32, I32), I32), 1
  %29 = load %28 : $*I32
  %30 = alloc_stack $((I32, I32), I32)
  %31 = tuple_element_addr %30 : $*((I32, I32), I32), 0
  %32 = tuple_element_addr %30 : $*((I32, I32), I32), 1
  %33 = tuple_element_addr %31 : $*(I32, I32), 0
  %34 = tuple_element_addr %31 : $*(I32, I32), 1
  store %25 to %33 : $*I32
  store %27 to %34 : $*I32
  store %29 to %32 : $*I32
  %38 = unchecked_addr_cast %30 : $*((I32, I32), I32) to $*(I32, I32)
  %39 = tuple_element_addr %38 : $*(I32, I32), 0
  %40 = tuple_element_addr %38 : $*(I32, I32), 1
  %41 = load %39 : $*I32
  %42 = load %40 : $*I32
  dealloc_stack %30 : $*((I32, I32), I32)
  %44 = alloc_stack $((I32, I32), I32)
  %45 = tuple_element_addr %44 : $*((I32, I32), I32), 0
  %46 = tuple_element_addr %44 : $*((I32, I32), I32), 1
  %47 = tuple_element_addr %45 : $*(I32, I32), 0
  %48 = tuple_element_addr %45 : $*(I32, I32), 1
  store %25 to %47 : $*I32
  store %27 to %48 : $*I32
  store %29 to %46 : $*I32
  %52 = unchecked_addr_cast %44 : $*((I32, I32), I32) to $*(I32, I32, I32)
  %53 = tuple_element_addr %52 : $*(I32, I32, I32), 0
  %54 = tuple_element_addr %52 : $*(I32, I32, I32), 1
  %55 = tuple_element_addr %52 : $*(I32, I32, I32), 2
  %56 = load %53 : $*I32
  %57 = load %54 : $*I32
  %58 = load %55 : $*I32
  dealloc_stack %44 : $*((I32, I32), I32)
  %60 = tuple (%20 : $I32, %21 : $I32)
  %61 = tuple (%41 : $I32, %42 : $I32)
  %62 = tuple (%56 : $I32, %57 : $I32, %58 : $I32)
  %63 = tuple (%60 : $(I32, I32), %61 : $(I32, I32), %62 : $(I32, I32, I32))
  return %63 : $((I32, I32), (I32, I32), (I32, I32, I32))
}