Add Support for Creating and Deleting Unicode Files and Directories in Lit

[jmittert]

Summary

This enables lit to work with unicode file names via mkdir, rm, and redirection. Lit still uses utf-8 internally, but converts to utf-16 on Windows, or just utf-8 bytes on everything else.

Taking my best guess at a reviewer.

[smeenai]

Could you upload this with context? The easiest way is to use arcanist (https://llvm.org/docs/Phabricator.html). Otherwise use -U9999 when generating the diff to upload to Phabricator.

[jmittert]

Recreated with context/arc diff

[serge-sans-paille]

utils/lit/lit/TestRunner.py

1109

I'm not 100% sure here, but this looks very redundant and non-maintainable. Why not always open the file in binary mode?

[jmittert]

Updated to open the script file in binary mode regardless of platform.

[serge-sans-paille]

utils/lit/lit/TestRunner.py

1087

That's nce, but now, you need to use the b prefix for all strings, and for decoding o f the joined command, right?

[jmittert]

Ugh, I had broken something somewhere else which caused the tests look like
they were always passing...

I've now properly fixed the binary mode.

[serge-sans-paille]

utils/lit/lit/util.py

444

Bad news: I've been discussing with a Python Core Dev and he acknowledge this function to be ultra dangerous. We should find a way to have everything right without this tricky call.

My take would be to first have the whole script work in unicode mode (i.e. in Python3) and then do the minimal work to have it work on Python2. Does that makes sense to you?

[jmittert]

utils/lit/lit/util.py

444

Let me check that I understand correctly. Right now, lit uses strings on python3 which are unicode aware, and strings/bytes on python2 which are not unicode aware which causes issues when utf8 characters are passed to os.path on Windows.

What about alternatively adding a to_unicode a la the to_string and to_bytes in util.py which would return str on python3, and unicode on python2?

Then adding something like

s = some_string
if windows:
    s = to_unicode(s)
os.path.some_func(s)

before each call to os.path.some_func?

[serge-sans-paille]

utils/lit/lit/util.py

444

Your first statement is correct. However I don't understand why there's anything windows specific in there. Let think about it this way: in Python2, strings where used to represent two different kind of objects : raw bytes, and unicode string. They are different types in Python3. So my advice would be:

1. have the code work in Python3, inserting bytes to string conversion whenever needed. These conversion should not be platform specific.

2. try to run the same code under Python2 and see where it fails.

Does that look like a reasonable path to you?

[jmittert]

utils/lit/lit/util.py

444

Hmm, I was under the impression that that was exactly what I've done here.

Lit works as of now with python3 on both Windows and Unix. The "convertToLocalEncoding" is essentially the fix up for when I ran it in python2 and saw where it failed. In order for the os.path functions to handle unicode properly on python 2, they need to take a unicode string on Windows. Otherwise, the os.path functions create/open a file with the literal utf-8 bytes which causes a broken filename for every other application that expects utf-16 on Windows. On Unix, utf-8 means things work fine in python2 as is. Passing a unicode type to os.path on unix/python2 gives an error about not being able to decode to ascii, but just passing bytes works.

Right now, with python 2 lit passes utf-8 bytes around (the python2 string/byte type) which works for Unix, but needs converting on Windows. What I have right now shouldn't result in any functional changes except for python2 with Windows.

[jmittert]

@serge-sans-paille ping

[serge-sans-paille]

@jmittert I tried the following (simpler) patch on Linux and it seems to work nice for both Python2 and Python3

Index: lit/TestRunner.py
===================================================================
--- lit/TestRunner.py	(revision 353501)
+++ lit/TestRunner.py	(working copy)
@@ -345,7 +345,7 @@
     exitCode = 0
     for dir in args:
         if not os.path.isabs(dir):
-            dir = os.path.realpath(os.path.join(cmd_shenv.cwd, dir))
+            dir = os.path.realpath(to_bytes(os.path.join(cmd_shenv.cwd, dir)))
         if parent:
             lit.util.mkdir_p(dir)
         else:
@@ -599,7 +599,7 @@
     exitCode = 0
     for path in args:
         if not os.path.isabs(path):
-            path = os.path.realpath(os.path.join(cmd_shenv.cwd, path))
+            path = os.path.realpath(to_bytes(os.path.join(cmd_shenv.cwd, path)))
         if force and not os.path.exists(path):
             continue
         try:
@@ -695,7 +695,7 @@
         else:
             # Make sure relative paths are relative to the cwd.
             redir_filename = os.path.join(cmd_shenv.cwd, name)
-            fd = open(redir_filename, mode)
+            fd = open(to_bytes(redir_filename), mode)
         # Workaround a Win32 and/or subprocess bug when appending.
         #
         # FIXME: Actually, this is probably an instance of PR6753.

What do you think of this path?

[jmittert]

What do you think of this path?

This doesn't work on Windows with Python 2 because it to_bytes doesn't convert the bytes to UTF16. It will work on Python 3 with Windows because py3 strings are already unicode aware. Running with python 2 creates the garbled 中文 directory on Windows because it tries to interpret the UTF8 as UTF16. Running with python 3 properly produces the 中文 directory.

For example, adding a quick test with

# RUN: mkdir -p  c:/Users/jmittertreiner/Output/中文

Produces

S:\build\Ninja-DebugAssert\llbuild-windows-amd64> dir C:\Users\jmittertreiner\Output\                                                                                     
                                                                                                                                                                          
                                                                                                                                                                          
    Directory: C:\Users\jmittertreiner\Output                                                                                                                             
                                                                                                                                                                          
                                                                                                                                                                          
Mode                LastWriteTime         Length Name                                                                                                                     
----                -------------         ------ ----                                                                                                                     
d-----         2/8/2019   9:48 AM                中文             <-- Running with Python 2                                                                                                      
d-----         2/8/2019   9:49 AM                中文              <-- Running with Python 3

[compnerd]

utils/lit/lit/util.py

444

@serge-sans-paille this is a nifty workaround for Windows. When on the python side, a unicode object is passed, the C side will use the W versions of the file APIs rather than A versions which is required to access any non-ASCII file path. Additionally, the use of the W variant of the APIs permits the use of NT style paths (\\?\ prefixed paths) which bypass the Win32 layer and go right to the kernel to avoid the 261 character limit.

[jmittert]

utils/lit/lit/util.py

444

@serge-sans-paille

@compnerd mentioned to me that you wanted to have this function changed to be called something along the lines of "convertToUnicode". I should mention though that this only converts to Unicode on Windows, On non Windows platforms it converts to bytes, not the unicode type. Think that convertToPlatformEncoding is maybe more descriptive instead?

[jmittert]

utils/lit/lit/util.py

444

@serge-sans-paille ping again

[efriedma]

utils/lit/lit/util.py

440

The inner "if" is redundant; on Python2, bytes is an alias for str, so both sides have the same meaning.

This function seems dangerous in the sense that it isn't clear what type of data it's expecting as input. Probably you should have separate functions for each of the possible cases: input is bytes, input is str, or input is Python2 unicode/Python3 str. (Not sure which of those you actually need, but those are the reasonable possibilities, I think.)

[jmittert]

utils/lit/lit/util.py

440

I'm fairly certain the if isn't redundant: on python 2, if we get bytes(/str) we want to decode that to a unicode. In python 2, .decode returns a unicode, not a str like in python 3.

Part of the reason it has confusing inputs is because what input it gets depends on they python version. It gets bytes/str on python 2 and str on python 3. I figured it was safer to handle all text cases (python 2 bytes/str/unicode, python 3 str/bytes). At least that way it's idempotent. I do agree it's dangerous in that the type it outputs is different based on the platform and version, but I've tried to limit use of it to only right at the edge before os calls.

[efriedma]

utils/lit/lit/util.py

440

on python 2, if we get bytes(/str) we want to decode that to a unicode

I didn't mean it's a no-op, just that isinstance(text, str) and isinstance(text, bytes) always return the same result on Python2, so you don't need the explicit version_info check.

I figured it was safer to handle all text cases

Making the function handle any string-like input is "safer" in the sense that you're less likely to get a runtime error in this function, but it makes the code harder to understand, and it's more likely to lead to confusing results if some other part of the code isn't handling strings consistently.

[jmittert]

Rather than using the ambiguous (and not particularly safe)
convertToLocalEncoding, define a to_unicode and use that as well as the
existing to_bytes depending on the platform.

[jmittert]

utils/lit/lit/util.py

440

Ah, I see what you mean, yeah, that makes sense to do.

I've removed the convertToLocalEncoding and instead added to_unicode, which I think is significantly less confusing.

[serge-sans-paille]

@jmittert sorry for the long delay, but I'm finally fine with this patch now. I like how it explicitly emphasizes on the Windows/Linux difference. However the patch needs to be rebased against master, can you update it?

[jmittert]

Rebased!

utils/lit/lit/TestRunner.py

for o, a in opts:
             assert False, "unhandled option"
 
     if len(args) == 0:
         raise InternalShellError(cmd, "Error: 'mkdir' is missing an operand")
 
     stderr = StringIO()
     exitCode = 0
     for dir in args:
+        dir = lit.util.convertToLocalEncoding(dir)
+        cwd = lit.util.convertToLocalEncoding(cmd_shenv.cwd)
         if not os.path.isabs(dir):
-            dir = os.path.realpath(os.path.join(cmd_shenv.cwd, dir))
+            dir = os.path.realpath(os.path.join(cwd, dir))
         if parent:
             lit.util.mkdir_p(dir)
         else:
             try:
                 os.mkdir(dir)
             except OSError as err:
                 stderr.write("Error: 'mkdir' command failed, %s\n" % str(err))
                 exitCode = 1
def on_rm_error(func, path, exc_info):
         # path contains the path of the file that couldn't be removed
         # let's just assume that it's read-only and remove it.
         os.chmod(path, stat.S_IMODE( os.stat(path).st_mode) | stat.S_IWRITE)
         os.remove(path)
 
     stderr = StringIO()
     exitCode = 0
     for path in args:
+        path = lit.util.convertToLocalEncoding(path)
+        cwd = lit.util.convertToLocalEncoding(cmd_shenv.cwd)
         if not os.path.isabs(path):
-            path = os.path.realpath(os.path.join(cmd_shenv.cwd, path))
+            path = os.path.realpath(os.path.join(cwd, path))
         if force and not os.path.exists(path):
             continue
         try:
             if os.path.isdir(path):
                 if not recursive:
                     stderr.write("Error: %s is a directory\n" % path)
                     exitCode = 1
                 shutil.rmtree(path, onerror = on_rm_error if force else None)
for (index, r) in enumerate(redirects):
             fd = tempfile.TemporaryFile(mode=mode)
         elif kIsWindows and name == '/dev/tty':
             # Simulate /dev/tty on Windows.
             # "CON" is a special filename for the console.
             fd = open("CON", mode)
         else:
             # Make sure relative paths are relative to the cwd.
             redir_filename = os.path.join(cmd_shenv.cwd, name)
-            fd = open(redir_filename, mode)
+            fd = open(lit.util.convertToLocalEncoding(redir_filename), mode)
         # Workaround a Win32 and/or subprocess bug when appending.
         #
         # FIXME: Actually, this is probably an instance of PR6753.
         if mode == 'a':
             fd.seek(0, 2)
         # Mutate the underlying redirect list so that we can redirect stdout
         # and stderr to the same place without opening the file twice.
         r[2] = fd

 def executeScript(test, litConfig, tmpBase, commands, cwd):
     bashPath = litConfig.getBashPath()
     isWin32CMDEXE = (litConfig.isWindows and not bashPath)
     script = tmpBase + '.script'
     if isWin32CMDEXE:
         script += '.bat'
 
     # Write script file
-    mode = 'w'
-    if litConfig.isWindows and not isWin32CMDEXE:
-      mode += 'b'  # Avoid CRLFs when writing bash scripts.
-    f = open(script, mode)
+    f = open(script, 'wb')
     if isWin32CMDEXE:
         for i, ln in enumerate(commands):
             commands[i] = re.sub(kPdbgRegex, "echo '\\1' > nul && ", ln)
         if litConfig.echo_all_commands:
             f.write('@echo on\n')
         else:
             f.write('@echo off\n')
         f.write('\n@if %ERRORLEVEL% NEQ 0 EXIT\n'.join(commands))
     else:
         for i, ln in enumerate(commands):
             commands[i] = re.sub(kPdbgRegex, ": '\\1'; ", ln)
         if test.config.pipefail:
-            f.write('set -o pipefail;')
+            f.write('set -o pipefail;\n')
         if litConfig.echo_all_commands:
-            f.write('set -x;')
-        f.write('{ ' + '; } &&\n{ '.join(commands) + '; }')
+            f.write('set -x;\n')
+        f.write('{ ' + '; } &&\n{ '.join(commands) + '; }\n')
     f.write('\n')
     f.close()
 
     if isWin32CMDEXE:
         command = ['cmd','/c', script]
     else:
         if bashPath:
             command = [bashPath, script]
         else:
             command = ['/bin/sh', script]
         if litConfig.useValgrind:
             # FIXME: Running valgrind on sh is overkill. We probably could just
             # run on clang with no real loss.
             command = litConfig.valgrindArgs + command

utils/lit/lit/util.py

try:
         for child in children_iterator:
             try:
                 child.kill()
             except psutil.NoSuchProcess:
                 pass
         psutilProc.kill()
     except psutil.NoSuchProcess:
         pass
+
+def convertToLocalEncoding(text):
+    """This function converts utf-8 text into a format which the local system prefers
+    to work with.
+    """
+    if platform.system() == 'Windows':
+        if sys.version_info < (3,0):
+            # On Windows and Python2, we want to use 'unicode' so it gets converted
+            # to UTF-16
+            return text.decode('utf-8') if isinstance(text, str) else test
+        else:
+            # On Windows and Python3, we want to use a unicode string so it gets
+            # converted to UTF-16
+            return text.decode('utf-8') if isinstance(text, bytes) else test
+
+    # On non Windows, we just want bytes so python don't try to
+    # convert it to ascii
+    return text if isinstance(text, bytes) else test.encode('utf-8')

utils/lit/tests/Inputs/shtest-shell/rm-unicode-0.txt

+# Check creating and removing unicode files
+#
+# RUN: mkdir -p  Output/中文
+# RUN: echo "" > Output/中文/你好.txt
+# RUN: rm Output/中文/你好.txt
+# RUN: echo "" > Output/中文/你好.txt
+# RUN: rm -r Output/中文

utils/lit/tests/shtest-shell.py

 # CHECK: error: command failed with exit status: 1
 # CHECK: ***
 
 # CHECK: FAIL: shtest-shell :: rm-error-3.txt
 # CHECK: *** TEST 'shtest-shell :: rm-error-3.txt' FAILED ***
 # CHECK: Exit Code: 1
 # CHECK: ***
 
+# CHECK: PASS: shtest-shell :: rm-unicode-0.txt
 # CHECK: PASS: shtest-shell :: sequencing-0.txt
 # CHECK: XFAIL: shtest-shell :: sequencing-1.txt
 # CHECK: PASS: shtest-shell :: valid-shell.txt
 # CHECK: Failing Tests (27)