Fix OwnershipLiveRange for @owned values getting transformed to none values via switch_enum

[meg-gupta]

rdar://95039799

We are currently not computing OwnershipLiveRange correctly in the presence of switch_enum transforming the owned value to a none value.

We can end up with illegal SIL with LoadCopyToLoadBorrowOptimization, due to this.

Example:

sil [ossa] @switch_enum_test : $@convention(thin) (@inout StructWithEnum) -> () {
 bb0(%0 : $*StructWithEnum):
   %1 = struct_element_addr %0 : $*StructWithEnum, #StructWithEnum.val
   %2 = struct_element_addr %0 : $*StructWithEnum, #StructWithEnum.val
   %3 = load [copy] %2 : $*EnumA
   destroy_addr %1 : $*EnumA
   switch_enum %3 : $EnumA, case #EnumA.sometrivial!enumelt: bb1, case #EnumA.somenontrivial!enumelt:bb2, case #EnumA.none!enumelt: bb3

 bb1(%6 : $TrivialStruct):
   %f = function_ref @get_enum : $@convention(thin) () -> @owned EnumA
   %r = apply %f() : $@convention(thin) () -> @owned EnumA
   %ele = struct_element_addr %0 : $*StructWithEnum, #StructWithEnum.val
   store %r to [init] %ele :$*EnumA
   %9999 = tuple()
   return %9999 : $()

 bb2(%7 : @owned $NonTrivialStruct):
  unreachable

 bb3:
  unreachable
 }

can get transformed to:

sil [ossa] @switch_enum_test : $@convention(thin) (@inout StructWithEnum) -> () {

bb0(%0 : $*StructWithEnum):
  %1 = struct_element_addr %0 : $*StructWithEnum, #StructWithEnum.val 
  %2 = struct_element_addr %0 : $*StructWithEnum, #StructWithEnum.val 
  %3 = load_borrow %2 : $*EnumA                   
  destroy_addr %1 : $*EnumA                       
  switch_enum %3 : $EnumA, case #EnumA.sometrivial!enumelt: bb1, case #EnumA.somenontrivial!enumelt: bb2, case #EnumA.none!enumelt: bb3 

bb1(%6 : $TrivialStruct):                         
  end_borrow %3 : $EnumA                          
  
  %8 = function_ref @get_enum : $@convention(thin) () -> @owned EnumA 
  %9 = apply %8() : $@convention(thin) () -> @owned EnumA 
  %10 = struct_element_addr %0 : $*StructWithEnum, #StructWithEnum.val 
  store %9 to [init] %10 : $*EnumA                
  %12 = tuple ()                                  
  return %12 : $()                                

bb2(%14 : @guaranteed $NonTrivialStruct):         
  unreachable                                     

bb3:                                              
  unreachable                                     
} 

which is incorrect:
Write:   destroy_addr %1 : $*EnumA                       
SIL verification failed: Found load borrow that is invalidated by a local write?!: loadBorrowImmutabilityAnalysis.isImmutable(LBI)
Verifying instruction:
     %2 = struct_element_addr %0 : $*StructWithEnum, #StructWithEnum.val 
->   %3 = load_borrow %2 : $*EnumA                
     switch_enum %3 : $EnumA, case #EnumA.sometrivial!enumelt: bb1, case #EnumA.somenontrivial!enumelt: bb2, case #EnumA.none!enumelt: bb3 
     end_borrow %3 : $EnumA                       

[meg-gupta]

@swift-ci test

[gottesmm]

I am fine with this for 5.6 to fix this bug. That being said, lets get on main the real fix for this as we discussed: getting @atrick's patch that stops OSSA from using dead end blocks committed. That will solve this problem while allowing us to turn back on this optimization. It will also eliminate a bunch of subtle bugs that we have seen due to dead end blocks in OSSA.

[meg-gupta]

Fixed a code comment.

[meg-gupta]

@swift-ci test

[atrick]

@swift-ci benchmark

[atrick]

Thanks!

[meg-gupta]

Pasting benchmark results from the bot :

------- Performance (x86_64): -O -------

REGRESSION OLD NEW DELTA RATIO
FlattenListFlatMap 3006 3768 +25.3% 0.80x (?)

------- Code size: -O -------
------- Performance (x86_64): -Osize -------

IMPROVEMENT OLD NEW DELTA RATIO
FlattenListFlatMap 4206 2640 -37.2% 1.59x (?)
FlattenListLoop 1400 1029 -26.5% 1.36x (?)
PrefixAnySeqCntRange 15 13 -13.3% 1.15x (?)
StringWalk 1360 1240 -8.8% 1.10x (?)
StrComplexWalk 3020 2810 -7.0% 1.07x (?)

------- Code size: -Osize -------
------- Performance (x86_64): -Onone -------

REGRESSION OLD NEW DELTA RATIO
SIMDReduce.Int8 6303 7186 +14.0% 0.88x (?)

IMPROVEMENT OLD NEW DELTA RATIO
ObjectiveCBridgeFromNSDictionaryAnyObjectForced 6850 6400 -6.6% 1.07x (?)

------- Code size: -swiftlibs -------