[clang] [MinGW] Don't look for a GCC in path if the install base has a proper mingw sysroot (#76949)

This fixes uses of the MSYS2 clang64 environment compilers, if another
set of GCC based compilers are available further back in PATH (which may
be explicitly added, or inherited unintentionally from other software
installed).

(The issue in the clang64 environment can be worked around somewhat by
installing *-gcc-compat packages which present aliases named
<triple>-gcc within the clang64 environment as well.)

This fixes msys2/MINGW-packages#11495 and
msys2/MINGW-packages#19279.

Author: mstorsjo

clang/lib/Driver/ToolChains/MinGW.cpp

@@ -471,12 +471,23 @@ findClangRelativeSysroot(const Driver &D, const llvm::Triple &LiteralTriple,
   return make_error_code(std::errc::no_such_file_or_directory);
 }
 
+static bool looksLikeMinGWSysroot(const std::string &Directory) {
+  StringRef Sep = llvm::sys::path::get_separator();
+  if (!llvm::sys::fs::exists(Directory + Sep + "include" + Sep + "_mingw.h"))
+    return false;
+  if (!llvm::sys::fs::exists(Directory + Sep + "lib" + Sep + "libkernel32.a"))
+    return false;
+  return true;
+}
+
 toolchains::MinGW::MinGW(const Driver &D, const llvm::Triple &Triple,
                          const ArgList &Args)
     : ToolChain(D, Triple, Args), CudaInstallation(D, Triple, Args),
       RocmInstallation(D, Triple, Args) {
   getProgramPaths().push_back(getDriver().getInstalledDir());
 
+  std::string InstallBase =
+      std::string(llvm::sys::path::parent_path(getDriver().getInstalledDir()));
   // The sequence for detecting a sysroot here should be kept in sync with
   // the testTriple function below.
   llvm::Triple LiteralTriple = getLiteralTriple(D, getTriple());
@@ -487,13 +498,17 @@ toolchains::MinGW::MinGW(const Driver &D, const llvm::Triple &Triple,
   else if (llvm::ErrorOr<std::string> TargetSubdir = findClangRelativeSysroot(
                getDriver(), LiteralTriple, getTriple(), SubdirName))
     Base = std::string(llvm::sys::path::parent_path(TargetSubdir.get()));
+  // If the install base of Clang seems to have mingw sysroot files directly
+  // in the toplevel include and lib directories, use this as base instead of
+  // looking for a triple prefixed GCC in the path.
+  else if (looksLikeMinGWSysroot(InstallBase))
+    Base = InstallBase;
   else if (llvm::ErrorOr<std::string> GPPName =
                findGcc(LiteralTriple, getTriple()))
     Base = std::string(llvm::sys::path::parent_path(
         llvm::sys::path::parent_path(GPPName.get())));
   else
-    Base = std::string(
-        llvm::sys::path::parent_path(getDriver().getInstalledDir()));
+    Base = InstallBase;
 
   Base += llvm::sys::path::get_separator();
   findGccLibDir(LiteralTriple);
@@ -778,9 +793,15 @@ static bool testTriple(const Driver &D, const llvm::Triple &Triple,
   if (D.SysRoot.size())
     return true;
   llvm::Triple LiteralTriple = getLiteralTriple(D, Triple);
+  std::string InstallBase =
+      std::string(llvm::sys::path::parent_path(D.getInstalledDir()));
   if (llvm::ErrorOr<std::string> TargetSubdir =
           findClangRelativeSysroot(D, LiteralTriple, Triple, SubdirName))
     return true;
+  // If the install base itself looks like a mingw sysroot, we'll use that
+  // - don't use any potentially unrelated gcc to influence what triple to use.
+  if (looksLikeMinGWSysroot(InstallBase))
+    return false;
   if (llvm::ErrorOr<std::string> GPPName = findGcc(LiteralTriple, Triple))
     return true;
   // If we neither found a colocated sysroot or a matching gcc executable,

clang/test/Driver/mingw-sysroot.cpp

@@ -14,6 +14,12 @@
 // RUN: ln -s %S/Inputs/mingw_ubuntu_posix_tree/usr/x86_64-w64-mingw32 %T/testroot-clang/x86_64-w64-mingw32
 // RUN: ln -s %S/Inputs/mingw_arch_tree/usr/i686-w64-mingw32 %T/testroot-clang/i686-w64-mingw32
 
+// RUN: rm -rf %T/testroot-clang-native
+// RUN: mkdir -p %T/testroot-clang-native/bin
+// RUN: ln -s %clang %T/testroot-clang-native/bin/clang
+// RUN: mkdir -p %T/testroot-clang-native/include/_mingw.h
+// RUN: mkdir -p %T/testroot-clang-native/lib/libkernel32.a
+
 // RUN: rm -rf %T/testroot-custom-triple
 // RUN: mkdir -p %T/testroot-custom-triple/bin
 // RUN: ln -s %clang %T/testroot-custom-triple/bin/clang
@@ -58,6 +64,28 @@
 // RUN: env "PATH=%T/testroot-gcc/bin:%PATH%" %T/testroot-gcc/bin/x86_64-w64-mingw32-clang -target x86_64-w64-mingw32 -rtlib=platform -stdlib=libstdc++ --sysroot="" -c -### %s 2>&1 | FileCheck -check-prefix=CHECK_TESTROOT_GCC %s
 
 
+// If we're executing clang from a directory with what looks like a mingw sysroot,
+// with headers in <base>/include and libs in <base>/lib, use that rather than looking
+// for another GCC in the path.
+//
+// Note, this test has a surprising quirk: We're testing with an install directory,
+// testroot-clang-native, which lacks the "x86_64-w64-mingw32" subdirectory, it only
+// has the include and lib subdirectories without any triple prefix.
+//
+// Since commit fd15cb935d7aae25ad62bfe06fe9f17cea585978, we avoid using the
+// <base>/include and <base>/lib directories when cross compiling. So technically, this
+// case testcase only works exactly as expected when running on x86_64 Windows, when
+// this target isn't considered cross compiling.
+//
+// However we do still pass the include directory <base>/x86_64-w64-mingw32/include to
+// the -cc1 interface, even if it is missing. Thus, this test looks for this path name,
+// that indicates that we did choose the right base, even if this particular directory
+// actually doesn't exist here.
+
+// RUN: env "PATH=%T/testroot-gcc/bin:%PATH%" %T/testroot-clang-native/bin/clang -target x86_64-w64-mingw32 -rtlib=compiler-rt -stdlib=libstdc++ --sysroot="" -c -### %s 2>&1 | FileCheck -check-prefix=CHECK_TESTROOT_CLANG_NATIVE %s
+// CHECK_TESTROOT_CLANG_NATIVE: "{{[^"]+}}/testroot-clang-native{{/|\\\\}}x86_64-w64-mingw32{{/|\\\\}}include"
+
+
 // If the user requests a different arch via the -m32 option, which changes
 // x86_64 into i386, check that the driver notices that it can't find a
 // sysroot for i386 but there is one for i686, and uses that one.