[LLVM][NFC] Use used's element type if available

[AlexVlx]

When embedding, if compiler.used exists, we should re-use it's element type instead of blindly assuming it's an unqualified pointer.

[pmatos]

I haven't touched LLVM source code for almost a year now and will defer review to others more active atm.

[jvoung]

Mind adding a test case? Perhaps something like the embed-bitcode.ll, but instead checking the round-tripping of compiler.used ?

[llvmbot]

@llvm/pr-subscribers-clang

Author: Alex Voicu (AlexVlx)

Changes

When embedding, if compiler.used exists, we should re-use it's element type instead of blindly assuming it's an unqualified pointer.

---

Full diff: https://github.com/llvm/llvm-project/pull/116804.diff

2 Files Affected:

• (added) clang/test/CodeGen/embed-bitcode-marker-with-nonzero-as.c (+8)
• (modified) llvm/lib/Bitcode/Writer/BitcodeWriter.cpp (+2-1)

diff --git a/clang/test/CodeGen/embed-bitcode-marker-with-nonzero-as.c b/clang/test/CodeGen/embed-bitcode-marker-with-nonzero-as.c
new file mode 100644
index 00000000000000..5cd0c74a03c200
--- /dev/null
+++ b/clang/test/CodeGen/embed-bitcode-marker-with-nonzero-as.c
@@ -0,0 +1,8 @@
+// RUN: %clang_cc1 -triple spirv64-amd-amdhsa -emit-llvm -fcuda-is-device -fembed-bitcode=marker -x hip %s -o - \
+// RUN:   | FileCheck %s --check-prefix=CHECK
+
+// CHECK: @llvm.embedded.module = private addrspace(1) constant [0 x i8] zeroinitializer, section ".llvmbc", align 1
+// CHECK-NEXT: @llvm.cmdline = private addrspace(1) constant [387 x i8] c"{{.*}}", section ".llvmcmd", align 1
+// CHECK-NEXT: @llvm.compiler.used = appending addrspace(1) global [5 x ptr addrspace(4)] [ptr addrspace(4) addrspacecast (ptr addrspace(1) @foo.managed to ptr addrspace(4)), ptr addrspace(4) addrspacecast (ptr addrspace(1) @foo to ptr addrspace(4)), ptr addrspace(4) addrspacecast (ptr addrspace(1) @__hip_cuid_ to ptr addrspace(4)), ptr addrspace(4) addrspacecast (ptr addrspace(1) @llvm.embedded.module to ptr addrspace(4)), ptr addrspace(4) addrspacecast (ptr addrspace(1) @llvm.cmdline to ptr addrspace(4))], section "llvm.metadata"
+
+__attribute__((managed)) int foo = 42;
diff --git a/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp b/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
index 24a4c2e8303d5a..80e12bef502ace 100644
--- a/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
+++ b/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
@@ -5501,8 +5501,9 @@ void llvm::embedBitcodeInModule(llvm::Module &M, llvm::MemoryBufferRef Buf,
   // Save llvm.compiler.used and remove it.
   SmallVector<Constant *, 2> UsedArray;
   SmallVector<GlobalValue *, 4> UsedGlobals;
-  Type *UsedElementType = PointerType::getUnqual(M.getContext());
   GlobalVariable *Used = collectUsedGlobalVariables(M, UsedGlobals, true);
+  Type *UsedElementType = Used ? Used->getValueType()->getArrayElementType()
+                               : PointerType::getUnqual(M.getContext());
   for (auto *GV : UsedGlobals) {
     if (GV->getName() != "llvm.embedded.module" &&
         GV->getName() != "llvm.cmdline")

[AlexVlx]

Mind adding a test case? Perhaps something like the embed-bitcode.ll, but instead checking the round-tripping of compiler.used ?

Done, although slightly differently; it does reflect that when there is a pre-existing compiler.used its element type is now retained.