When generating llvm.used, we may need an addrspacecast instead of a bitcast.

[jholewinski]

Summary

This is especially important for targets that use multiple address spaces,
and commonly place global variables in address spaces other than zero.

Fixes PR22383

[arsenm]

LGTM

[jingyue]

Only two minor comments. Thanks!

lib/CodeGen/CodeGenModule.cpp

939 ↗

(On Diff #19152)

The indentation can be improved :)

test/CodeGenCUDA/llvm-used.cu

8 ↗

(On Diff #19152)

Is attribute((used)) necessary for this test?

[jholewinski]

Comments inline.

lib/CodeGen/CodeGenModule.cpp

939 ↗

(On Diff #19152)

What would you suggest? I lowered the indentation to keep it in 80 columns.

test/CodeGenCUDA/llvm-used.cu

8 ↗

(On Diff #19152)

That is the test case attached to the bug report. The attribute may not be strictly required in the input .cu file, but without it the test won't add the variable to the llvm.used list which is the point of the bug.

[jingyue]

lib/CodeGen/CodeGenModule.cpp

939 ↗

(On Diff #19152)

e.g., four spaces from the last line. I'm following clang-format.

for (unsigned i = 0, e = List.size(); i != e; ++i) {
  UsedArray[i] =
      llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(
          cast<llvm::Constant>(&*List[i]),

test/CodeGenCUDA/llvm-used.cu

8 ↗

(On Diff #19152)

I see. I was confused.

cfe/trunk/lib/CodeGen/CodeGenModule.cpp

static void emitUsed(CodeGenModule &CGM, StringRef Name,
   if (List.empty())
     return;
 
   // Convert List to what ConstantArray needs.
   SmallVector<llvm::Constant*, 8> UsedArray;
   UsedArray.resize(List.size());
   for (unsigned i = 0, e = List.size(); i != e; ++i) {
     UsedArray[i] =
-     llvm::ConstantExpr::getBitCast(cast<llvm::Constant>(&*List[i]),
-                                    CGM.Int8PtrTy);
+        llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(
+            cast<llvm::Constant>(&*List[i]), CGM.Int8PtrTy);
   }
 
   if (UsedArray.empty())
     return;
   llvm::ArrayType *ATy = llvm::ArrayType::get(CGM.Int8PtrTy, UsedArray.size());
 
   auto *GV = new llvm::GlobalVariable(
       CGM.getModule(), ATy, false, llvm::GlobalValue::AppendingLinkage,

cfe/trunk/test/CodeGenCUDA/llvm-used.cu

+// RUN: %clang_cc1 -emit-llvm %s -o - -fcuda-is-device -triple nvptx64-unknown-unknown | FileCheck %s
+
+
+// Make sure we emit the proper addrspacecast for llvm.used.  PR22383 exposed an
+// issue where we were generating a bitcast instead of an addrspacecast.
+
+// CHECK: @llvm.used = appending global [1 x i8*] [i8* addrspacecast (i8 addrspace(1)* bitcast ([0 x i32] addrspace(1)* @a to i8 addrspace(1)*) to i8*)], section "llvm.metadata"
+__attribute__((device)) __attribute__((__used__)) int a[] = {};