Look for compiler-rt from subdir given by --target

[wzssyqa]

Currently, clang looks for compiler-rt only from the normalized triple subdir. While if we are configured with a non-normalized triple with -DLLVM_DEFAULT_TARGET_TRIPLE, such as triples without vendor section, clang will fail to find compiler_rt.

Let's look for compiler_rt from the subdir with name from --target option, too.

To archive this, we add a new member called Origin to class Triple.

Fixes: #87150.

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-clang-driver

Author: YunQiang Su (wzssyqa)

Changes

Currently, clang looks for compiler-rt only from the normalized triple subdir. While if we are configured with a non-normalized triple with -DLLVM_DEFAULT_TARGET_TRIPLE, such as triples without vendor section, clang will fail to find compiler_rt.

Let's look for compiler_rt from the subdir with name from --target option, too.

To archive this, we add a new member called Origin to class Triple.

Fixes: #87150.

---

Full diff: https://github.com/llvm/llvm-project/pull/88334.diff

4 Files Affected:

• (modified) clang/lib/Driver/Driver.cpp (+1)
• (modified) clang/lib/Driver/ToolChain.cpp (+6)
• (modified) llvm/include/llvm/TargetParser/Triple.h (+6)
• (modified) llvm/lib/TargetParser/Triple.cpp (+3-3)

diff --git a/clang/lib/Driver/Driver.cpp b/clang/lib/Driver/Driver.cpp
index e7335a61b10c53..4a0c939039eb31 100644
--- a/clang/lib/Driver/Driver.cpp
+++ b/clang/lib/Driver/Driver.cpp
@@ -516,6 +516,7 @@ static llvm::Triple computeTargetTriple(const Driver &D,
     TargetTriple = A->getValue();
 
   llvm::Triple Target(llvm::Triple::normalize(TargetTriple));
+  Target.setOrigin(TargetTriple);
 
   // GNU/Hurd's triples should have been -hurd-gnu*, but were historically made
   // -gnu* only, and we can not change this, so we have to detect that case as
diff --git a/clang/lib/Driver/ToolChain.cpp b/clang/lib/Driver/ToolChain.cpp
index 237092ed07e5dc..57f27a61c4060b 100644
--- a/clang/lib/Driver/ToolChain.cpp
+++ b/clang/lib/Driver/ToolChain.cpp
@@ -755,6 +755,12 @@ std::optional<std::string>
 ToolChain::getTargetSubDirPath(StringRef BaseDir) const {
   auto getPathForTriple =
       [&](const llvm::Triple &Triple) -> std::optional<std::string> {
+    if (!Triple.getOrigin().empty()) {
+      SmallString<128> Po(BaseDir);
+      llvm::sys::path::append(Po, Triple.getOrigin());
+      if (getVFS().exists(Po))
+        return std::string(Po);
+    }
     SmallString<128> P(BaseDir);
     llvm::sys::path::append(P, Triple.str());
     if (getVFS().exists(P))
diff --git a/llvm/include/llvm/TargetParser/Triple.h b/llvm/include/llvm/TargetParser/Triple.h
index f256e2b205a889..a2fc28ada0ca31 100644
--- a/llvm/include/llvm/TargetParser/Triple.h
+++ b/llvm/include/llvm/TargetParser/Triple.h
@@ -298,6 +298,8 @@ class Triple {
 private:
   std::string Data;
 
+  StringRef Origin = StringRef();
+
   /// The parsed arch type.
   ArchType Arch{};
 
@@ -425,6 +427,8 @@ class Triple {
 
   const std::string &getTriple() const { return Data; }
 
+  const StringRef getOrigin() const { return Origin; }
+
   /// Get the architecture (first) component of the triple.
   StringRef getArchName() const;
 
@@ -1058,6 +1062,8 @@ class Triple {
   /// @name Mutators
   /// @{
 
+  void setOrigin(StringRef Orig) { Origin = Orig; };
+
   /// Set the architecture (first) component of the triple to a known type.
   void setArch(ArchType Kind, SubArchType SubArch = NoSubArch);
 
diff --git a/llvm/lib/TargetParser/Triple.cpp b/llvm/lib/TargetParser/Triple.cpp
index 624679ff507a7f..ce44903d0f7d70 100644
--- a/llvm/lib/TargetParser/Triple.cpp
+++ b/llvm/lib/TargetParser/Triple.cpp
@@ -928,9 +928,9 @@ static Triple::ObjectFormatType getDefaultFormat(const Triple &T) {
 /// This stores the string representation and parses the various pieces into
 /// enum members.
 Triple::Triple(const Twine &Str)
-    : Data(Str.str()), Arch(UnknownArch), SubArch(NoSubArch),
-      Vendor(UnknownVendor), OS(UnknownOS), Environment(UnknownEnvironment),
-      ObjectFormat(UnknownObjectFormat) {
+    : Data(Str.str()), Origin(Str.getSingleStringRef()), Arch(UnknownArch),
+      SubArch(NoSubArch), Vendor(UnknownVendor), OS(UnknownOS),
+      Environment(UnknownEnvironment), ObjectFormat(UnknownObjectFormat) {
   // Do minimal parsing by hand here.
   SmallVector<StringRef, 4> Components;
   StringRef(Data).split(Components, '-', /*MaxSplit*/ 3);

[MaskRay]

Can you provide CMake configure command and the compiler-rt file paths? I am not sure we need more changes to clangDriver.

[wzssyqa]

Configure cmd

cmake ../llvm -G Ninja -DLLVM_DEFAULT_TARGET_TRIPLE=aarch64-linux-gnu -DLLVM_USE_LINKER=lld -DLLVM_ENABLE_RUNTIMES="compiler-rt;libunwind" -DLLVM_ENABLE_PROJECTS="mlir;clang;clang-tools-extra;lld" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=RelWithDebInfo

Note, in -DLLVM_DEFAULT_TARGET_TRIPLE=aarch64-linux-gnu, there is no unknown aka vendor section.

With this configure, libclang_rt.builtin.a is present in
./lib/clang/19/lib/aarch64-linux-gnu/

While clang expects it in
./lib/clang/19/lib/aarch64-unknown-linux-gnu/ #### Note "unknown"

The reason is that in computeTargetTriple, the line

llvm::Triple Target(llvm::Triple::normalize(TargetTriple));

convert the triple to normalize.

[MaskRay]

cmake -Sllvm -B/tmp/out/d-a64 -G Ninja -DLLVM_DEFAULT_TARGET_TRIPLE=aarch64-linux-gnu -DLLVM_USE_LINKER=lld -DLLVM_ENABLE_RUNTIMES="compiler-rt;libunwind" -DLLVM_ENABLE_PROJECTS="clang;lld" -DCMAKE_C_COMPILER=~/Stable/bin/clang -DCMAKE_CXX_COMPILER=~/Stable/bin/clang++ -DCMAKE_BUILD_TYPE=Debug
ninja -C /tmp/out/d-a64 builtins

gives me /tmp/out/d-a64/lib/clang/19/lib/aarch64-linux-gnu/clang_rt.*. Sanitizer targets are not available. I haven't investigated why.

Anyhow, I posted a similar but probably more complete clang driver patch: https://reviews.llvm.org/D110663 , but I concluded that detecting Debian-style x86_64-linux-gnu is not a good idea (too much complexity to the driver when both normalized and un-normalized triples are supported. Error-prone when mix-and-matched.)

It's better to do some normalization in CMake.

[wzssyqa]

It's better to do some normalization in CMake.

I will try to work in CMake, then.