Extend findInitExistential for cases when ApplySite argument is a global_addr

[rajbarik]

Determine InitExistential for an apply argument that is a global_addr.

Pattern 1:

%3 = global_addr @$P : $*SomeP
%4 = init_existential_addr %3 : $*SomeP, $SomeC
%5 = alloc_ref $SomeC
store %5 to %4 : $*SomeC
%10 = apply %9(%3) : $@convention(thin) (@in_guaranteed SomeP)

Pattern 2:

%3 = global_addr @$P : $*SomeP
%9 = open_existential_addr mutable_access %3 : $*SomeP to $*@opened SomeP
%15 = apply %11(%9) : $@convention(thin) (@in_guaranteed SomeP)

No test cases yet, will be merged as part of the large PR #13991

[rajbarik]

@gottesmm @slavapestov @atrick Can you review this PR? This Pr refactors the old code for findInitExistentialFromGlobalAddr to findInitExistentialFromGlobalAddrAndCopyAddr and adds a new interface for finding init_existential for an apply argument who happens to be a global_addr, findInitExistentialFromGlobalAddrAndApply . This function is currently not being used, hence no new tests. I will use this interface when #13991 lands fully.

[atrick]

This generally looks good and the comments are nice. The only comments I have are the misleading variable names above.

[atrick]

Whenever I see uses of an "apply" without the type spelled out, I like to know when I'm dealing with an ApplySite vs. ApplyInst, and which code is potentially handling try_apply. Since "AI" is an acronym for ApplyInst, this is misleading. (I suspect this is backward in so much of our code because someone did a massive replace on the type without changing variable names?) Let's not make it worse.

[atrick]

AI is misleading again here. This is not always an ApplyInst. It's just some arbitrary use of the global_addr.

[atrick]

Again, calling this AI leads me to believe that it's not potentially a try_apply.

[rajbarik]

Thanks @atrick. Should be fixed now.

[atrick]

nit: you should probably comment that Insn is assumed to be a direct user of GAI (e.g. copy_addr or apply in the pattern above) and that a valid init_existential_addr value is returned only if this routine can prove that the value it initializes is the same value in memory at the given use point. Looking at this again, I don't understand why the return type is SILValue and not InitExistentialAddrInst*.

[rajbarik]

@atrick Thanks for the suggestion -- added the assumption. The reason it returns SILValue is because for convenience in use in getAddressOfStackInit below that returns a SILValue.

[atrick]

Ok. FYI, any SingleValueInstruction such as this is castable to SILValue.

[slavapestov]

@swift-ci Please smoke test

[slavapestov]

@swift-ci Please smoke test

[slavapestov]

Please update the existing SILCombiner self concrete type peephole to use the new function and add a SIL test.