Windows: runtime linker regression (isatty) with GHC 9.4.5+

(Originally spun off from #23309 (comment 497937) )

Given the following two files, distilled from the text-ansi library:

-- A.hs
module A where

import Foreign.C.Types
import System.IO.Unsafe

isatty :: Bool
isatty =
  unsafePerformIO (c_isatty 1) == 1
{-# NOINLINE isatty #-}

foreign import ccall unsafe "isatty"
  c_isatty :: CInt -> IO CInt

-- B.hs
{-# LANGUAGE TemplateHaskell #-}
module B where

import Foreign.C.String
import Language.Haskell.TH

import A

$(do runIO $ print isatty
     return [])

If you compile this with GHC 9.4.4 or earlier, it will run as expected:

$ ghc-9.2.7 -fforce-recomp B.hs
[1 of 2] Compiling A                ( A.hs, A.o )
[2 of 2] Compiling B                ( B.hs, B.o )
False

If you compile this with GHC 9.4.5 or later, however, the runtime linker will crash on the isatty symbol:

$ ghc-9.4.5 -fforce-recomp B.hs
[1 of 2] Compiling A                ( A.hs, A.o )
[2 of 2] Compiling B                ( B.hs, B.o )
ghc-9.4.5.exe:  | C:\Users\winferno\Documents\Hacking\Haskell\A.o: unknown symbol `isatty'
ghc-9.4.5.exe: Could not load Object Code C:\Users\winferno\Documents\Hacking\Haskell\A.o.


<no location info>: error:

(Also, the error message appears before the "error:" part, which is odd. That isn't the main point of this issue, however.)

This is almost certainly an unintended side effect of !9475 (closed). A similar issue was diagnosed in #23309, but unlike that issue, the issue here is not caused by a missing mingwex dependency. If you compile the program with -lmingwex, the same runtime linker crash still occurs. The patch in !10360 (closed) does not fix the issue here either.