[NFC][LLVM] Use @llvm.compiler.used/@llvm.used helpers in Clang and BitcodeWriter #162660

jmmartinez · 2025-10-09T14:04:39Z

The address space for the pointers stored in the llvm.compiler.used/llvm.used should be 0, but this was not always respected.

After #164432 Clang always emits addrspace 0 for the elements of llvm.compiler.used/llvm.used.

To avoid inconsistencies in the future, move the helpers to GlobalValue and use them in the 2 other places where these variables are manually updated: BitcodeWriter and CodeGenModule.

llvmbot · 2025-10-09T14:05:22Z

@llvm/pr-subscribers-llvm-ir
@llvm/pr-subscribers-backend-amdgpu

@llvm/pr-subscribers-llvm-transforms

Author: Juan Manuel Martinez Caamaño (jmmartinez)

Changes

At the moment, the address space for the pointer stored in the llvm.compiler.used/llvm.used is not well defined:

Clang uses the language default address space,
LLVM helpers from ModuleUtils use the 0 address space,
and BitcodeWriter (for fembed-bitcode) tries to preserve the element type if it already exists, 0 otherwise.

This PR doesn't solve this issue. It only adds one test that documents one problem with it, and makes the BitcodeWriter use the llvm helpers to reduce the places where these variables change.

Without this patch, the test added in the PR fails with an assertion:

UtilsTests: /home/juamarti/llvm/_llvm/llvm/lib/IR/Constants.cpp:1327:
    static Constant *llvm::ConstantArray::getImpl(ArrayType *, ArrayRef&lt;Constant *&gt;):
    Assertion `C-&gt;getType() == Ty-&gt;getElementType() &amp;&amp; "Wrong type in array element initializer"' failed.

Full diff: https://github.com/llvm/llvm-project/pull/162660.diff

4 Files Affected:

(modified) llvm/lib/Bitcode/Writer/BitcodeWriter.cpp (+30-40)
(modified) llvm/lib/Bitcode/Writer/CMakeLists.txt (+1)
(modified) llvm/lib/Transforms/Utils/ModuleUtils.cpp (+2-1)
(modified) llvm/unittests/Transforms/Utils/ModuleUtilsTest.cpp (+17)

diff --git a/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp b/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
index 7ed140d392fca..f7afa393f3e00 100644
--- a/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
+++ b/llvm/lib/Bitcode/Writer/BitcodeWriter.cpp
@@ -75,6 +75,7 @@
 #include "llvm/Support/SHA1.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/TargetParser/Triple.h"
+#include "llvm/Transforms/Utils/ModuleUtils.h"
 #include <algorithm>
 #include <cassert>
 #include <cstddef>
@@ -5850,25 +5851,25 @@ static const char *getSectionNameForCommandline(const Triple &T) {
 void llvm::embedBitcodeInModule(llvm::Module &M, llvm::MemoryBufferRef Buf,
                                 bool EmbedBitcode, bool EmbedCmdline,
                                 const std::vector<uint8_t> &CmdArgs) {
-  // Save llvm.compiler.used and remove it.
-  SmallVector<Constant *, 2> UsedArray;
-  SmallVector<GlobalValue *, 4> UsedGlobals;
-  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")
-      UsedArray.push_back(
-          ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, UsedElementType));
-  }
-  if (Used)
-    Used->eraseFromParent();
 
   // Embed the bitcode for the llvm module.
   std::string Data;
   ArrayRef<uint8_t> ModuleData;
   Triple T(M.getTargetTriple());
+  SmallVector<GlobalValue *, 2> NewGlobals;
+
+  auto IsCmdOrBitcode = [&](Constant *C) {
+    GlobalVariable *GV = dyn_cast<GlobalVariable>(C);
+    StringRef Name = GV ? GV->getName() : "";
+    if (EmbedBitcode && Name == "llvm.embedded.module")
+      return true;
+    if (EmbedCmdline && Name == "llvm.cmdline")
+      return true;
+    return false;
+  };
+
+  if (EmbedBitcode || EmbedCmdline)
+    removeFromUsedLists(M, IsCmdOrBitcode);
 
   if (EmbedBitcode) {
     if (Buf.getBufferSize() == 0 ||
@@ -5887,23 +5888,22 @@ void llvm::embedBitcodeInModule(llvm::Module &M, llvm::MemoryBufferRef Buf,
   }
   llvm::Constant *ModuleConstant =
       llvm::ConstantDataArray::get(M.getContext(), ModuleData);
-  llvm::GlobalVariable *GV = new llvm::GlobalVariable(
+  llvm::GlobalVariable *EmbeddedModule = new llvm::GlobalVariable(
       M, ModuleConstant->getType(), true, llvm::GlobalValue::PrivateLinkage,
       ModuleConstant);
-  GV->setSection(getSectionNameForBitcode(T));
+  EmbeddedModule->setSection(getSectionNameForBitcode(T));
   // Set alignment to 1 to prevent padding between two contributions from input
   // sections after linking.
-  GV->setAlignment(Align(1));
-  UsedArray.push_back(
-      ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, UsedElementType));
+  EmbeddedModule->setAlignment(Align(1));
+  NewGlobals.push_back(EmbeddedModule);
   if (llvm::GlobalVariable *Old =
           M.getGlobalVariable("llvm.embedded.module", true)) {
     assert(Old->hasZeroLiveUses() &&
            "llvm.embedded.module can only be used once in llvm.compiler.used");
-    GV->takeName(Old);
+    EmbeddedModule->takeName(Old);
     Old->eraseFromParent();
   } else {
-    GV->setName("llvm.embedded.module");
+    EmbeddedModule->setName("llvm.embedded.module");
   }
 
   // Skip if only bitcode needs to be embedded.
@@ -5913,30 +5913,20 @@ void llvm::embedBitcodeInModule(llvm::Module &M, llvm::MemoryBufferRef Buf,
                               CmdArgs.size());
     llvm::Constant *CmdConstant =
         llvm::ConstantDataArray::get(M.getContext(), CmdData);
-    GV = new llvm::GlobalVariable(M, CmdConstant->getType(), true,
-                                  llvm::GlobalValue::PrivateLinkage,
-                                  CmdConstant);
-    GV->setSection(getSectionNameForCommandline(T));
-    GV->setAlignment(Align(1));
-    UsedArray.push_back(
-        ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, UsedElementType));
+    GlobalVariable *CmdLine = new llvm::GlobalVariable(
+        M, CmdConstant->getType(), true, llvm::GlobalValue::PrivateLinkage,
+        CmdConstant);
+    CmdLine->setSection(getSectionNameForCommandline(T));
+    CmdLine->setAlignment(Align(1));
     if (llvm::GlobalVariable *Old = M.getGlobalVariable("llvm.cmdline", true)) {
       assert(Old->hasZeroLiveUses() &&
              "llvm.cmdline can only be used once in llvm.compiler.used");
-      GV->takeName(Old);
+      CmdLine->takeName(Old);
       Old->eraseFromParent();
     } else {
-      GV->setName("llvm.cmdline");
+      CmdLine->setName("llvm.cmdline");
     }
+    NewGlobals.push_back(CmdLine);
+    appendToCompilerUsed(M, NewGlobals);
   }
-
-  if (UsedArray.empty())
-    return;
-
-  // Recreate llvm.compiler.used.
-  ArrayType *ATy = ArrayType::get(UsedElementType, UsedArray.size());
-  auto *NewUsed = new GlobalVariable(
-      M, ATy, false, llvm::GlobalValue::AppendingLinkage,
-      llvm::ConstantArray::get(ATy, UsedArray), "llvm.compiler.used");
-  NewUsed->setSection("llvm.metadata");
 }
diff --git a/llvm/lib/Bitcode/Writer/CMakeLists.txt b/llvm/lib/Bitcode/Writer/CMakeLists.txt
index 2c508ca9fae95..5bbb872a90341 100644
--- a/llvm/lib/Bitcode/Writer/CMakeLists.txt
+++ b/llvm/lib/Bitcode/Writer/CMakeLists.txt
@@ -15,4 +15,5 @@ add_llvm_component_library(LLVMBitWriter
   ProfileData
   Support
   TargetParser
+  TransformUtils
   )
diff --git a/llvm/lib/Transforms/Utils/ModuleUtils.cpp b/llvm/lib/Transforms/Utils/ModuleUtils.cpp
index 596849ecab742..d1acb0ff1ad6b 100644
--- a/llvm/lib/Transforms/Utils/ModuleUtils.cpp
+++ b/llvm/lib/Transforms/Utils/ModuleUtils.cpp
@@ -138,10 +138,11 @@ static void appendToUsedList(Module &M, StringRef Name, ArrayRef<GlobalValue *>
 
   SmallSetVector<Constant *, 16> Init;
   collectUsedGlobals(GV, Init);
+  Type *ArrayEltTy = GV ? GV->getValueType()->getArrayElementType()
+                        : PointerType::getUnqual(M.getContext());
   if (GV)
     GV->eraseFromParent();
 
-  Type *ArrayEltTy = llvm::PointerType::getUnqual(M.getContext());
   for (auto *V : Values)
     Init.insert(ConstantExpr::getPointerBitCastOrAddrSpaceCast(V, ArrayEltTy));
 
diff --git a/llvm/unittests/Transforms/Utils/ModuleUtilsTest.cpp b/llvm/unittests/Transforms/Utils/ModuleUtilsTest.cpp
index d4094c5307060..0cc408af43bc5 100644
--- a/llvm/unittests/Transforms/Utils/ModuleUtilsTest.cpp
+++ b/llvm/unittests/Transforms/Utils/ModuleUtilsTest.cpp
@@ -69,6 +69,23 @@ TEST(ModuleUtils, AppendToUsedList2) {
   EXPECT_EQ(1, getListSize(*M, "llvm.used"));
 }
 
+TEST(ModuleUtils, AppendToUsedList3) {
+  LLVMContext C;
+
+  std::unique_ptr<Module> M = parseIR(C, R"(
+          @x = addrspace(1) global [2 x i32] zeroinitializer, align 4
+          @y = addrspace(2) global [2 x i32] zeroinitializer, align 4
+          @llvm.compiler.used = appending global [1 x ptr addrspace (3)] [ptr addrspace(3) addrspacecast (ptr addrspace (1) @x to ptr addrspace(3))]
+      )");
+  GlobalVariable *X = M->getNamedGlobal("x");
+  GlobalVariable *Y = M->getNamedGlobal("y");
+  EXPECT_EQ(1, getListSize(*M, "llvm.compiler.used"));
+  appendToCompilerUsed(*M, X);
+  EXPECT_EQ(1, getListSize(*M, "llvm.compiler.used"));
+  appendToCompilerUsed(*M, Y);
+  EXPECT_EQ(2, getListSize(*M, "llvm.compiler.used"));
+}
+
 using AppendFnType = decltype(&appendToGlobalCtors);
 using TransformFnType = decltype(&transformGlobalCtors);
 using ParamType = std::tuple<StringRef, AppendFnType, TransformFnType>;

arsenm · 2025-10-10T02:31:02Z

At the moment, the address space for the pointer stored in the llvm.compiler.used/llvm.used is not well defined:

It should be defined to 0

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

llvm/unittests/Transforms/Utils/ModuleUtilsTest.cpp

jmmartinez · 2025-10-10T08:10:59Z

At the moment, the address space for the pointer stored in the llvm.compiler.used/llvm.used is not well defined:

It should be defined to 0

In that case, should the LLVM-IR verifier (or at least the linter) reject bitcode where these variables have the wrong type?

I still have to check what breaks exactly when clang emits an addrspace(0) ptr in llvm.used/llvm.compiler.used. AFAIK it is the case for SPIRV, which should be solved by #162678 .

github-actions · 2025-10-10T08:17:31Z

✅ With the latest revision this PR passed the C/C++ code formatter.

arsenm · 2025-10-10T08:17:38Z

In that case, should the LLVM-IR verifier (or at least the linter) reject bitcode where these variables have the wrong type?

Yes. verifier failure.

I still have to check what breaks exactly when clang emits an addrspace(0) ptr in llvm.used/llvm.compiler.used. AFAIK it is the case for SPIRV, which should be solved by #162678 .

SPIR-V probably shouldn't just drop these on the floor. It probably will need special case code to stripPointerCasts and re-package the references into whatever SPIRV type it prefers

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

jmmartinez · 2025-10-21T08:25:06Z

Ping !

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

llvm/lib/IR/Globals.cpp

…ace(0) (#164432) By convention the AS of the elements of `llvm.compiler.used` & `llvm.used` is 0. However, the AS of `CGM.Int8PtrTy` is not always 0. This leaves some LLVM helpers (`appendToUsed/appendToCompilerUsed/removeFromUsedLists`) unusable. This patch makes the AS of the elements of these variables to be 0. This PR is related to #162660

…@llvm.used if it already exists This new test fails with: /home/juamarti/llvm/_llvm/llvm/lib/IR/Constants.cpp:1327: static Constant *llvm::ConstantArray::getImpl(ArrayType *, ArrayRef<Constant *>): Assertion `C->getType() == Ty->getElementType() && "Wrong type in array element initializer"' failed.

…already exists At the moment, the pointer type stored in the llvm.compiler.used/llvm.used is not well fixed. The frontend uses a pointer to the default address space (which may not be 0; for example, it is 4 for SPIRV). This patch makes `appendToUsed/appendToCompilerUsed` match the behaviour in BitcodeWriter.cpp: if the variable already exists, preserve its element type, otherwise use `ptr addrspace (0)`. This fixes the following error in the newly added test: UtilsTests: /home/juamarti/llvm/_llvm/llvm/lib/IR/Constants.cpp:1327: static Constant *llvm::ConstantArray::getImpl(ArrayType *, ArrayRef<Constant *>): Assertion `C->getType() == Ty->getElementType() && "Wrong type in array element initializer"' failed.

…mentation

GlobalValue; and remove dependency between BitcodeWriter & TransformUtils

…ents addrspace(0) (#164432) By convention the AS of the elements of `llvm.compiler.used` & `llvm.used` is 0. However, the AS of `CGM.Int8PtrTy` is not always 0. This leaves some LLVM helpers (`appendToUsed/appendToCompilerUsed/removeFromUsedLists`) unusable. This patch makes the AS of the elements of these variables to be 0. This PR is related to llvm/llvm-project#162660

… can now use unqual only

jmmartinez · 2025-11-03T15:49:17Z

I've updated the PR after #164432 was merged.

The PR now only moves around the helpers and updates BitcodeWriter/CodeGenModule to use them.

…ace(0) (llvm#164432) By convention the AS of the elements of `llvm.compiler.used` & `llvm.used` is 0. However, the AS of `CGM.Int8PtrTy` is not always 0. This leaves some LLVM helpers (`appendToUsed/appendToCompilerUsed/removeFromUsedLists`) unusable. This patch makes the AS of the elements of these variables to be 0. This PR is related to llvm#162660

arsenm · 2025-11-05T03:56:27Z

llvm/unittests/IR/UsedGlobalTest.cpp

+}
+
+static int getListSize(Module &M, StringRef Name) {
+  auto *List = M.getGlobalVariable(Name);


Suggested change

auto *List = M.getGlobalVariable(Name);

const GlobalVariable *List = M.getGlobalVariable(Name);

arsenm · 2025-11-05T03:56:38Z

llvm/unittests/IR/UsedGlobalTest.cpp

+  return Mod;
+}
+
+static int getListSize(Module &M, StringRef Name) {


Suggested change

static int getListSize(Module &M, StringRef Name) {

static size_t getListSize(Module &M, StringRef Name) {

jmmartinez requested review from AlexVlx, arsenm, eugenis and nikic October 9, 2025 14:04

jmmartinez self-assigned this Oct 9, 2025

llvmbot added the llvm:transforms label Oct 9, 2025

jmmartinez requested a review from shiltian October 9, 2025 14:06

arsenm reviewed Oct 10, 2025

View reviewed changes

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp Outdated Show resolved Hide resolved

llvm/unittests/Transforms/Utils/ModuleUtilsTest.cpp Outdated Show resolved Hide resolved

jmmartinez force-pushed the fix/workaround_for_llvm_compiler_used branch from 3d08e3a to 7b445e0 Compare October 10, 2025 08:14

arsenm reviewed Oct 10, 2025

View reviewed changes

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp Outdated Show resolved Hide resolved

jmmartinez mentioned this pull request Oct 10, 2025

[amd-spirv-backend][Driver] Develop opt-in pipeline for AMDGPU ROCm/llvm-project#305

Open

jmmartinez force-pushed the fix/workaround_for_llvm_compiler_used branch from e9a8d50 to e686175 Compare October 16, 2025 08:09

llvmbot added backend:AMDGPU llvm:codegen PGO Profile Guided Optimizations backend:NVPTX llvm:ir labels Oct 16, 2025

jmmartinez requested a review from arsenm October 20, 2025 07:14

nikic reviewed Oct 21, 2025

View reviewed changes

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp Outdated Show resolved Hide resolved

llvm/lib/IR/Globals.cpp Outdated Show resolved Hide resolved

jmmartinez mentioned this pull request Oct 21, 2025

[Clang] Make the AS of llvm.compiler.used & llvm.used elements addrspace(0) #164432

Merged

jmmartinez force-pushed the fix/workaround_for_llvm_compiler_used branch from e9a6253 to b076e53 Compare October 21, 2025 14:40

jmmartinez requested a review from nikic October 22, 2025 10:04

jmmartinez added 8 commits November 3, 2025 15:32

[NFC][BitcodeWriter] Use appendToCompilerUsed instead of custom imple…

bad3df5

…mentation

[Review] Remove lambda from embedBitcodeInModule

5279d3b

[Review] 'addrspace (n)' -> 'addrspace(n)'

7691a6f

[Review] Forgot again to run clang-format

68d30f5

[Review] Move llvm.used/llvm.compiler.used helpers from ModuleUtils to

57690b1

GlobalValue; and remove dependency between BitcodeWriter & TransformUtils

[Review] forgot to move the unittests

89d2468

[Review] Inline lambda with reference to string, to a trivial version

1b7c060

jmmartinez added 3 commits November 3, 2025 16:23

[Update] Use helpers in Clang's side

6005afe

[Review] header is not used anymore

00d3bfb

[Review] Since Clang doesn't emit elements in the wrong addrspace, we…

0195220

… can now use unqual only

jmmartinez force-pushed the fix/workaround_for_llvm_compiler_used branch from b076e53 to 0195220 Compare November 3, 2025 15:39

jmmartinez added 3 commits November 3, 2025 16:40

[Review] Remove unused header

443b060

[Update] Update test

2956d76

[Review] remove more unused headers

4609dd0

jmmartinez changed the title ~~[LLVM] Maintain element type of @llvm.compiler.used/@llvm.used if it already exists~~ [LLVM] Use @llvm.compiler.used/@llvm.used helpers in Clang and BitcodeWriter Nov 3, 2025

jmmartinez changed the title ~~[LLVM] Use @llvm.compiler.used/@llvm.used helpers in Clang and BitcodeWriter~~ [NFC][LLVM] Use @llvm.compiler.used/@llvm.used helpers in Clang and BitcodeWriter Nov 3, 2025

jmmartinez mentioned this pull request Nov 3, 2025

[Clang] Make the AS of llvm.compiler.used & llvm.used elements addrspace(0) (#164432) ROCm/llvm-project#477

Closed

arsenm approved these changes Nov 5, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[NFC][LLVM] Use @llvm.compiler.used/@llvm.used helpers in Clang and BitcodeWriter #162660

[NFC][LLVM] Use @llvm.compiler.used/@llvm.used helpers in Clang and BitcodeWriter #162660

Uh oh!

jmmartinez commented Oct 9, 2025 •

edited

Loading

Uh oh!

llvmbot commented Oct 9, 2025 •

edited

Loading

Uh oh!

arsenm commented Oct 10, 2025

Uh oh!

Uh oh!

Uh oh!

jmmartinez commented Oct 10, 2025

Uh oh!

github-actions bot commented Oct 10, 2025 •

edited

Loading

Uh oh!

arsenm commented Oct 10, 2025

Uh oh!

Uh oh!

jmmartinez commented Oct 21, 2025

Uh oh!

Uh oh!

Uh oh!

jmmartinez commented Nov 3, 2025

Uh oh!

arsenm Nov 5, 2025

Uh oh!

arsenm Nov 5, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

	auto *List = M.getGlobalVariable(Name);
	const GlobalVariable *List = M.getGlobalVariable(Name);

	static int getListSize(Module &M, StringRef Name) {
	static size_t getListSize(Module &M, StringRef Name) {

[NFC][LLVM] Use @llvm.compiler.used/@llvm.used helpers in Clang and BitcodeWriter #162660

Are you sure you want to change the base?

[NFC][LLVM] Use @llvm.compiler.used/@llvm.used helpers in Clang and BitcodeWriter #162660

Uh oh!

Conversation

jmmartinez commented Oct 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Oct 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

arsenm commented Oct 10, 2025

Uh oh!

Uh oh!

Uh oh!

jmmartinez commented Oct 10, 2025

Uh oh!

github-actions bot commented Oct 10, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

arsenm commented Oct 10, 2025

Uh oh!

Uh oh!

jmmartinez commented Oct 21, 2025

Uh oh!

Uh oh!

Uh oh!

jmmartinez commented Nov 3, 2025

Uh oh!

arsenm Nov 5, 2025

Choose a reason for hiding this comment

Uh oh!

arsenm Nov 5, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

jmmartinez commented Oct 9, 2025 •

edited

Loading

llvmbot commented Oct 9, 2025 •

edited

Loading

github-actions bot commented Oct 10, 2025 •

edited

Loading