I've spent some time over the past few weeks working on problems stack users have run into on Windows, and I'd like to share the outcome. To summarize, here are the major problems I've seen encountered:
- When linking a project with a large number of libraries, GHC hits the 32k command length limit of Windows, causing linking to fail with a mysterious "gcc: command not found."
- On Windows, paths (at least by default) are limited to 260 characters. This can cause problems quickly when using either stack or cabal sandboxes, which have dist directory structures including GHC versions, Cabal versions, and sometimes a bit more metadata.
- Most users do not have a Unicode codepage (e.g., 65001 UTF-8) by default, so some characters cannot be produced by GHC. This affects both error/warning output on stdout/stderr, and dump files (e.g.,
-ddump-to-file -ddump-hi, which stack uses for detecting unlisted modules and Template Haskell files. Currently, GHC simply crashes when this occurs. This can affect non-Windows systems as well.
The result of this so far has been four GHC patches, and one recommended workaround - hopefully we can do better on that too.
Thanks to all those who have helped me get these patches in place, especially Ben Gamari, Reid Barton, Tamar Christina and Austin Seipp. If you're eager and want to test out the changes already, you can try out my GHC 7.10 branch.
Always produce UTF8-encoded dump files
This patch has already been merged and backported to GHC 7.10. The idea is simple: GHC expects input files to always be UTF-8 encoded, so generated UTF-8 encoded dump files too. Upshot: environment variables and codepage settings can no longer affect the format of these dump files, making it more reliable for tooling to parse and use these files.
Transliterate unknown characters
This patch is similarly both merged and backported. Currently, if GHC tries to print a warning that includes non-Latin characters, and the LANG variable/Windows codepage doesn't support it, you end up with a crash about the commitBuffer. This change is pretty simple: take the character encoding used by stdout and stderr, and switch on transliteration, which replaces unknown characters with a question mark (?).
GHC_CHARENC environment variable
The motivation here is that, when capturing the output of GHC, tooling like stack (and presumably cabal as well) would like to receive it in a consistent format. GHC currently has no means of setting the character encoding reliably across OSes: Windows uses the codepage, which is a quasi-global setting, whereas non-Windows uses the LANG environment variable. And even changing LANG may not be what we want; for example, setting that to
C.UTF-8 would enable smart quotes, which we don't necessary want to do.
This new variable can be used to force GHC to use a specific character encoding, regardless of other settings. I chose to do this as an environment variable instead of a command line option, so that it would be easier to have this setting trickle through multiple layers of tools (e.g., stack calling the Cabal library calling GHC).
Note: This patch has not yet been merged, and is probably due for some discussion around naming.
Use a response file for command line arguments
Response files allow us to pass compiler and linker arguments via an external file instead of the command line, avoiding the 32k limit on Windows. The response file patch does just this. This patch is still being reviewed, but I'm hopeful that it will make it in for GHC 7.10.3, to help alleviate the pain points a number of Windows users are having. I'd also like to ask people reading this who are affected by this issue to test out the patches I've made; instructions are available on the stack issue tracker.
Workaround: shorter paths
For the issue of long path names, I don't have a patch available yet, nor am I certain that I can make one. Windows in principle supports tacking
\\?\ to the beginning of an absolute path to unlock much larger path limits. However, I can't get this to be respected by GHC yet (I still need some investigation).
A workaround is to move your project directory to the root of the filesystem, and to set your
STACK_ROOT environment variable similarly to your root (e.g.,
set STACK_ROOT=c:\stack_root). This should keep you under the limit for most cases.