[clang] [Driver] Teach Barmetal toolchain about GCC installation (PR #121829)

[Garvit Gupta via cfe-commits]

================
@@ -110,20 +110,81 @@ static std::string computeBaseSysRoot(const Driver &D, bool IncludeTriple) {
   return std::string(SysRootDir);
 }
 
+static bool hasGCCToolChainAlongSideClang(const Driver &D) {
+  SmallString<128> GCCDir;
+  llvm::sys::path::append(GCCDir, D.Dir, "..", D.getTargetTriple(),
+                          "lib/crt0.o");
+  return llvm::sys::fs::exists(GCCDir);
+}
+
+std::string BareMetal::computeSysRoot() const {
+  if (!SysRoot.empty())
+    return SysRoot;
+
+  const Driver &D = getDriver();
+  if (!D.SysRoot.empty())
+    return D.SysRoot;
+
+  // Verify the GCC installation from -gcc-install-dir, --gcc-toolchain, or
+  // alongside clang. If valid, form the sysroot. Otherwise, check
+  // lib/clang-runtimes above the driver.
+  SmallString<128> SysRootDir;
+  if (GCCInstallation.isValid()) {
+    StringRef LibDir = GCCInstallation.getParentLibPath();
+    StringRef TripleStr = GCCInstallation.getTriple().str();
+    llvm::sys::path::append(SysRootDir, LibDir, "..", TripleStr);
+  } else if (hasGCCToolChainAlongSideClang(D)) {
+    // Use the triple as provided to the driver. Unlike the parsed triple
+    // this has not been normalized to always contain every field.
+    llvm::sys::path::append(SysRootDir, D.Dir, "..", D.getTargetTriple());
+  }
+
+  if (llvm::sys::fs::exists(SysRootDir))
+    return std::string(SysRootDir);
+  return computeBaseSysRoot(D, /*IncludeTriple*/ true);
+}
+
+static void addMultilibsFilePaths(const Driver &D, const MultilibSet &Multilibs,
+                                  const Multilib &Multilib,
+                                  StringRef InstallPath,
+                                  ToolChain::path_list &Paths) {
+  if (const auto &PathsCallback = Multilibs.filePathsCallback())
+    for (const auto &Path : PathsCallback(Multilib))
+      addPathIfExists(D, InstallPath + Path, Paths);
+}
+
 BareMetal::BareMetal(const Driver &D, const llvm::Triple &Triple,
                      const ArgList &Args)
-    : ToolChain(D, Triple, Args),
-      SysRoot(computeBaseSysRoot(D, /*IncludeTriple=*/true)) {
-  getProgramPaths().push_back(getDriver().Dir);
-
-  findMultilibs(D, Triple, Args);
-  SmallString<128> SysRoot(computeSysRoot());
-  if (!SysRoot.empty()) {
-    for (const Multilib &M : getOrderedMultilibs()) {
-      SmallString<128> Dir(SysRoot);
-      llvm::sys::path::append(Dir, M.osSuffix(), "lib");
-      getFilePaths().push_back(std::string(Dir));
-      getLibraryPaths().push_back(std::string(Dir));
+    : Generic_ELF(D, Triple, Args) {
+  GCCInstallation.init(Triple, Args);
+  SysRoot = computeSysRoot();
+  if (GCCInstallation.isValid()) {
+    Multilibs = GCCInstallation.getMultilibs();
+    SelectedMultilibs.assign({GCCInstallation.getMultilib()});
+    path_list &Paths = getFilePaths();
+    // Add toolchain/multilib specific file paths.
+    addMultilibsFilePaths(D, Multilibs, SelectedMultilibs.back(),
+                          GCCInstallation.getInstallPath(), Paths);
+    getFilePaths().push_back(GCCInstallation.getInstallPath().str());
+    ToolChain::path_list &PPaths = getProgramPaths();
+    // Multilib cross-compiler GCC installations put ld in a triple-prefixed
+    // directory off of the parent of the GCC installation.
+    PPaths.push_back(Twine(GCCInstallation.getParentLibPath() + "/../" +
+                           GCCInstallation.getTriple().str() + "/bin")
+                         .str());
+    PPaths.push_back((GCCInstallation.getParentLibPath() + "/../bin").str());
+    getFilePaths().push_back(SysRoot + "/lib");
+  } else {
----------------
quic-garvgupt wrote:

> Stepping back a bit, there's a bit of the logic/use-case in computeSysRoot that I don't understand. When we find a GCC ToolChain alongside clang, I don't see any logic to validate it.
 
There is no code to validate it in the constructor. However, it's use was needed in the linker job (see `Linker::ConstructJob`). This was done to preserve the behavior of RISCVToolChain Object. I am not sure why it was decided there that the mere presence of `lib/crt0.o` file alongside clang in target triple prefixed directory will indicate a GCC toolchain, however in RISCVToolchain object, toolchain found along clang  isn't considered a valid GCCInstallation and the same has been implemented here to preserve that behavior. 
One needs to look at `hasGCCToolchain`, `computeSysRoot` and constructor in RISCVToolChain.cpp for the same
 
>Is the intention here to find a valid GCC installation equivalent to the -gcc-install-dir or --gcc-toolchain, or something else?
 
Yes it is trying to find a  toolchain that is equivalent to `--gcc-toolchain`. The same is also mentioned in the comments for the function `hasGCCToolChain` in RISCVToolChain.cpp. You can also look at tests "clang/test/Driver/riscv64-toolchain-extra.c` and "clang/test/Driver/riscv32-toolchain-extra.c" to see how it works.
 
> The logic in this block is the same as if clang-runtimes with multilib.yaml were used. Is this intentional?
 
I am not sure if I understood your question completely. But yes, the logic is a bit similar but the directory for sysroot will be different. In the case when GCCInstallation isn't valid but there is a GCC toolchain adjacent, then the sysroot would be `bin/../<target-triple>/lib/crt0.o` whereas with clang-runtimes with multilib.yaml, it will be `bin/../lib/clang-runtimes/` (include triple is set to false, when multilib.yaml is found). 

Apologies for the late reply. I was out for sometime in the last month and then I my commit access got revoked. Thanks for reviewing this patch and pls let me know if you have further questions/comments.

https://github.com/llvm/llvm-project/pull/121829