Driver tests: set `--sysroot=""` to support clang with `DEFAULT_SYSROOT`

[broadwaylamb]

Summary

When testing clang that has been compiled with -DDEFAULT_SYSROOT set to some path,
some tests would fail. Override sysroot to be empty string for the tests to succeed when clang is configured with DEFAULT_SYSROOT.

[broadwaylamb]

Add --sysroot="" to some driver regression tests

These tests otherwise fail if clang is configured with DEFAULT_SYSROOT.

[broadwaylamb]

Remove unrelated change

[compnerd]

I think that this is pretty easy to forget. Fortunately, last argument wins. Why not sink this into the %clang substitution in lit? That ensures that we run with an empty sysroot and then when the test needs to adjust the sysroot, it can do so explicitly.

[broadwaylamb]

In D66834#1652756, @compnerd wrote:

I think that this is pretty easy to forget. Fortunately, last argument wins. Why not sink this into the %clang substitution in lit? That ensures that we run with an empty sysroot and then when the test needs to adjust the sysroot, it can do so explicitly.

This totally makes sense.

I've just tried to do it, but unfortunately some tests are failing, for example, Driver/cc1-response-files.c. The problem is that %clang is expanded to /path/to/clang --sysroot=, but the succeeding flags (such as -cc1) may be incompatible with --sysroot.

Or am I missing something?

[sepavloff]

In D66834#1653334, @broadwaylamb wrote:

In D66834#1652756, @compnerd wrote:

I think that this is pretty easy to forget. Fortunately, last argument wins. Why not sink this into the %clang substitution in lit? That ensures that we run with an empty sysroot and then when the test needs to adjust the sysroot, it can do so explicitly.

I've just tried to do it, but unfortunately some tests are failing, for example, Driver/cc1-response-files.c. The problem is that %clang is expanded to /path/to/clang --sysroot=, but the succeeding flags (such as -cc1) may be incompatible with --sysroot.

Does the issue manifests itself with -cc1 only? We usually use %clang_cc1 in such cases, so probably those tests require update.

[broadwaylamb]

In D66834#1685260, @sepavloff wrote:

In D66834#1653334, @broadwaylamb wrote:

In D66834#1652756, @compnerd wrote:

I think that this is pretty easy to forget. Fortunately, last argument wins. Why not sink this into the %clang substitution in lit? That ensures that we run with an empty sysroot and then when the test needs to adjust the sysroot, it can do so explicitly.

I've just tried to do it, but unfortunately some tests are failing, for example, Driver/cc1-response-files.c. The problem is that %clang is expanded to /path/to/clang --sysroot=, but the succeeding flags (such as -cc1) may be incompatible with --sysroot.

Does the issue manifests itself with -cc1 only? We usually use %clang_cc1 in such cases, so probably those tests require update.

Some Driver tests take a list of arguments from a file (for example clang/test/Driver/cc1-response-files.c), so we probably cannot use %clang_cc1 there

[sepavloff]

In D66834#1685921, @broadwaylamb wrote:

In D66834#1685260, @sepavloff wrote:

In D66834#1653334, @broadwaylamb wrote:

In D66834#1652756, @compnerd wrote:

I think that this is pretty easy to forget. Fortunately, last argument wins. Why not sink this into the %clang substitution in lit? That ensures that we run with an empty sysroot and then when the test needs to adjust the sysroot, it can do so explicitly.

I've just tried to do it, but unfortunately some tests are failing, for example, Driver/cc1-response-files.c. The problem is that %clang is expanded to /path/to/clang --sysroot=, but the succeeding flags (such as -cc1) may be incompatible with --sysroot.

Does the issue manifests itself with -cc1 only? We usually use %clang_cc1 in such cases, so probably those tests require update.

Some Driver tests take a list of arguments from a file (for example clang/test/Driver/cc1-response-files.c), so we probably cannot use %clang_cc1 there

The obvious solution in this case is to implement support of an environment variable, say SYSROOT or CLANG_SYSROOT, which would override hard-coded path (but not that specified by --sysroot option). Unfortunately some developers don't like use of environment variables, as it creates 'secret' control channel, and such patch would have high chance to be rejected. So probably this is all we can do here.

This patch looks good for me.

[broadwaylamb]

Can someone please commit this?

clang/unittests/Driver/ToolChainTest.cpp

const char *EmptyFiles[] = {
       "/lib/arm-linux-gnueabi/.keep",
       "/lib/arm-linux-gnueabihf/.keep"};
 
   for (const char *Path : EmptyFiles)
     InMemoryFileSystem->addFile(Path, 0,
                                 llvm::MemoryBuffer::getMemBuffer("\n"));
 
   std::unique_ptr<Compilation> C(TheDriver.BuildCompilation(
-      {"-fsyntax-only", "--gcc-toolchain=", "foo.cpp"}));
+      {"-fsyntax-only", "--gcc-toolchain=", "--sysroot=", "foo.cpp"}));
   EXPECT_TRUE(C);
 
   std::string S;
   {
     llvm::raw_string_ostream OS(S);
     C->getDefaultToolChain().printVerboseInfo(OS);
   }
 #if _WIN32