llvm dev - Oct 2021 - [llvm-lit] Old echo causes test failure on windows

Hi llvm-dev,

I wanted to bring discussion from D110986 to a wider audience.

After D110986 added a unicode character to the test,
llvm\test\tools\llvm-cxxfilt\delimiters.test
started failing in our (Sony's) downstream Windows CI build (and locally for
me) because the
unicode character was being converted to '?' at some stage in the test.
The upstream Windows bots
seemed to have no such trouble. I added a workaround in D111072 that we landed
because a Chromium
bot was running into the same issue.

Digging a little deeper it turns out that the echo that lit was picking up for
me (`echo (GNU
coreutils) 5.3.0` from GnuWin32) converts unicode characters to '?'.

llvm-lit has its own built in echo which handles unicode without an issue. This
is used in the workaround
D111072; lit's echo is used iff the echo command isn't used in a
pipeline.

N.B. `echo (GNU coreutils) 8.32` shipped with git for Windows also appears to
handle unicode without
an issue.

I have a few questions:
1) Is there a reason we cannot always use lit's echo, as Hans suggests in
D111072?
2) If we can't always use lit's echo, maybe we could add an
error/warning if a "bad" echo is
detected at build-config or test-running time?

Thanks,
Orlando
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20211005/0fbda658/attachment.html>

At a minimum, I think we should stop using (and recommending) the bit-rotting
GnuWin32 tools, and instead use the tools that come with git-for-Windows.  Lit
currently checks for 5 tools (cmp, diff, echo, grep, sed) and all of these are
in the git-Windows set.  This will reduce the set of packages we depend on, even
if it doesn't reduce the set of actual tools we depend on.
--paulr

From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of via
llvm-dev
Sent: Tuesday, October 5, 2021 10:30 AM
To: llvm-dev at lists.llvm.org
Subject: [llvm-dev] [llvm-lit] Old echo causes test failure on windows

Hi llvm-dev,

I wanted to bring discussion from D110986 to a wider audience.

After D110986 added a unicode character to the test,
llvm\test\tools\llvm-cxxfilt\delimiters.test
started failing in our (Sony's) downstream Windows CI build (and locally for
me) because the
unicode character was being converted to '?' at some stage in the test.
The upstream Windows bots
seemed to have no such trouble. I added a workaround in D111072 that we landed
because a Chromium
bot was running into the same issue.

Digging a little deeper it turns out that the echo that lit was picking up for
me (`echo (GNU
coreutils) 5.3.0` from GnuWin32) converts unicode characters to '?'.

llvm-lit has its own built in echo which handles unicode without an issue. This
is used in the workaround
D111072; lit's echo is used iff the echo command isn't used in a
pipeline.

N.B. `echo (GNU coreutils) 8.32` shipped with git for Windows also appears to
handle unicode without
an issue.

I have a few questions:
1) Is there a reason we cannot always use lit's echo, as Hans suggests in
D111072?
2) If we can't always use lit's echo, maybe we could add an
error/warning if a "bad" echo is
detected at build-config or test-running time?

Thanks,
Orlando
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20211005/97aabcc6/attachment.html>

I assumed that when using the internal shell (non-optional under
Windows), lit would only use its implementation of shell-builtins [1].
If this is not the case yet, I'd vote to do so since I would
understand the option to replace the environment's shell. Shell's
generally prioritize their builtins over executables found in $PATH, I
think lit should do the same.

Michael

[1]
https://github.com/llvm/llvm-project/blob/2ac199993764e068494a69a85af098c0ae1ff37e/llvm/utils/lit/lit/TestRunner.py#L624