Attachment: Download

Issue Description:

The issue

I have this pretty simple demo program

import Darwin

@inline(never)
func FOObarBUZquxQUUX(_ fn: () -> Int32) -> Int32 {
    #if WITH_WEL
        return withExtendedLifetime(fn) { fn in
            return fn()
        }
    #elseif NO_WEL
        return fn()
    #else
        fatalError("Compile with -DWITH_WEL or -DNO_WEL")
    #endif
}

for i in 0..<100_000_000 {
    let a = FOObarBUZquxQUUX {
        return getpid() + Int32(extendingOrTruncating: i)
    }
    if a != getpid() + Int32(extendingOrTruncating: i) {
        fatalError()
    }
}

which takes 20s to run if compiled with -D WITH_WEL to enable the withExtendedLifetime but only 0.5s to run with -D NO_WEL to disable withExtendedLifetime. I really can't understand why.

Demo run

$ for f in WITH NO; do echo $f && rm -f test && swiftc -O -D ${f}_WEL test.swift && swiftc -O -D ${f}_WEL test.swift -emit-sil > /tmp/${f}_WEL.sil && time ./test; done
WITH

real    0m23.835s
user    0m23.387s
sys 0m0.100s
NO

real    0m0.552s
user    0m0.524s
sys 0m0.007s

beginning of an analysis

• Attached, please find a diff of the generated SILs of the two programs, on the left hand side is the one with the withExtendedLifetime and on the right the one without.

• the difference seems to be that in the case where I use withExtendedLifetime it does a {{ %5 = partial_apply %4(%2) : $@convention(thin) (@in (), @owned @callee_owned () -> Int32) -> @out Int32 // users: %20, %19}} which doesn't seem to happen without it. I wonder what needs to be partially applied? The only thing I can think of is that it's building a "useless" thunk just so it has something it can retain/release for withExtendedLifetime but I hope that's not quite true.

• the version with withExtendedLifetime causes a malloc and a free every time it's run, that seems to be for the thunk that's created above.