HEAD segfaults on Windows when building large packages (e.g., lsp-types)

When using nightly Windows HEAD bindists, I have noticed that it will segfault when building large packages. Unfortunately, I don't know of a minimal way to reproduce the bug other than to build a large Hackage library with head.hackage. The lsp-types-1.6.0.0 library provides a good example of this. To reproduce, do the following:

1. Download a nightly x86_64-windows-validate bindist. At the time of writing, I am using the one from here, at commit 3c0e3793.

2. Using head.hackage, check out the lsp-types-1.6.0.0 library:

   ./head.hackage/scripts/patch-tool unpack lsp-types-1.6.0.0

3. Build lsp-types-1.6.0.0 using the nightly Windows bindist. It should eventually produce a segfault that looks roughly like this one:

   $ cabal build lsp-types
   
   <elided>
   
   [45 of 52] Compiling Language.LSP.Types.Registration ( src\Language\LSP\Types\Registration.hs, C:\\Users\winferno\Documents\Hacki
   ng\Haskell\staging\dist-newstyle\build\x86_64-windows\ghc-9.5.20221105\lsp-types-1.6.0.0\build\Language\LSP\Types\Registration.o
   )
   
   src\Language\LSP\Types\Registration.hs:128:1: error: [GHC-55017]
       Illegal type variable name: `'
       When splicing a TH declaration
   Access violation in generated code when executing data at 0x7ff6992263d9
   
    Attempting to reconstruct a stack trace...
   
      Frame        Code address
    * 0xe29c6fd9f0 0x7ff6992263d9 C:\Users\winferno\Software\ghc-9.5.20221105\bin\ghc-9.5.20221105.exe+0x50363d9
   
   Error: cabal-3.8.1.0.exe: Failed to build lsp-types-1.6.0.0. The build process
   terminated with exit code 11

   I've also seen the segfault happen without the Illegal type variable name part:

   [45 of 52] Compiling Language.LSP.Types.Registration ( src\Language\LSP\Types\Registration.hs, C:\\Users\winferno\Documents\Hacki
   ng\Haskell\staging\dist-newstyle\build\x86_64-windows\ghc-9.5.20221105\lsp-types-1.6.0.0\build\Language\LSP\Types\Registration.o
   )
   
   Access violation in generated code when reading 0xffffffffffffffff
   
    Attempting to reconstruct a stack trace...
   
      Frame        Code address
    * 0xc27adfd970 0x7ff695a5c188 C:\Users\winferno\Software\ghc-9.5.20221105\bin\ghc-9.5.20221105.exe+0x186c188
   
   Error: cabal-3.8.1.0.exe: Failed to build lsp-types-1.6.0.0. The build process
   terminated with exit code 11

4. Curiously, if you re-run cabal build lsp-types, it will be able to compile the module that segfaulted above (Language.LSP.Types.Registration), but it will segfault on a later module:

   $ cabal build lsp-types
   
   <elided>
   
   [47 of 52] Compiling Language.LSP.Types.Initialize ( src\Language\LSP\Types\Initialize.hs, C:\\Users\winferno\Documents\Hacking\H
   askell\staging\dist-newstyle\build\x86_64-windows\ghc-9.5.20221105\lsp-types-1.6.0.0\build\Language\LSP\Types\Initialize.o )
   
   Access violation in generated code when executing data at 0x7ef4cc37b0ba
   
    Attempting to reconstruct a stack trace...
   
      Frame        Code address
    * 0xeb45cfda80 0x7ef4cc37b0ba
   
   Error: cabal-3.8.1.0.exe: Failed to build lsp-types-1.6.0.0. The build process
   terminated with exit code 11

   If I then re-run cabal build lsp-types a third time, then compilation will succeed.

   On the other hand, if I run cabal clean and then run cabal build lsp-tyes afresh, then I will reliably reproduce the first segfault at Language.LSP.Types.Registration observed above.

As far as I can tell, this only affects HEAD, and not any GHC version from 9.4 or earlier.