yaxunl updated this revision to Diff 331363.
yaxunl added a comment.
bug fix. only remove unused casts.
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D98783/new/
https://reviews.llvm.org/D98783
Files:
clang/lib/CodeGen/CodeGenModule.cpp
clang/lib/CodeGen/CodeGenModule.h
clang/test/CodeGenCUDA/unused-global-var.cu
Index: clang/test/CodeGenCUDA/unused-global-var.cu
===================================================================
--- /dev/null
+++ clang/test/CodeGenCUDA/unused-global-var.cu
@@ -0,0 +1,50 @@
+// REQUIRES: amdgpu-registered-target
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -fcuda-is-device -x hip %s \
+// RUN: -std=c++11 -O3 -mllvm -amdgpu-internalize-symbols -emit-llvm -o - \
+// RUN: | FileCheck %s
+
+#include "Inputs/cuda.h"
+
+// AMDGPU internalize unused global variables for whole-program compilation
+// (-fno-gpu-rdc for each TU, or -fgpu-rdc for LTO), which are then
+// eliminated by global DCE. If there are invisible unused address space casts
+// for global variables, the internalization and elimination of unused global
+// variales will be hindered. This test makes sure no such address space
+// casts.
+
+// Check unused device/constant variables are eliminated.
+
+// CHECK-NOT: @v1
+__device__ int v1;
+
+// CHECK-NOT: @v2
+__constant__ int v2;
+
+// CHECK-NOT: @_ZL2v3
+constexpr int v3 = 1;
+
+// Check managed variables are always kept.
+
+// CHECK: @v4
+__managed__ int v4;
+
+// Check used device/constant variables are not eliminated.
+// CHECK: @u1
+__device__ int u1;
+
+// CHECK: @u2
+__constant__ int u2;
+
+// Check u3 is kept because its address is taken.
+// CHECK: @_ZL2u3
+constexpr int u3 = 2;
+
+// Check u4 is not kept because it is not ODR-use.
+// CHECK-NOT: @_ZL2u4
+constexpr int u4 = 3;
+
+__device__ int fun1(const int& x);
+
+__global__ void kern1(int *x) {
+ *x = u1 + u2 + fun1(u3) + u4;
+}
Index: clang/lib/CodeGen/CodeGenModule.h
===================================================================
--- clang/lib/CodeGen/CodeGenModule.h
+++ clang/lib/CodeGen/CodeGenModule.h
@@ -371,6 +371,14 @@
llvm::SmallVector<std::pair<llvm::GlobalValue *, llvm::Constant *>, 8>
GlobalValReplacements;
+ /// Potentially unused address space casts of global variables to be cleaned
+ /// up. In CUDA/HIP, global variables are emitted as global variables in
+ /// device or constant address space which are then casted to default address
+ /// space. If the global variables are not used, the address space casts
+ /// become invisible LLVM constants, causing spurious use of the global
+ /// variables which prevents them from being erased.
+ llvm::DenseSet<llvm::Constant *> GlobalVarCasts;
+
/// Variables for which we've emitted globals containing their constant
/// values along with the corresponding globals, for opportunistic reuse.
llvm::DenseMap<const VarDecl*, llvm::GlobalVariable*> InitializerConstants;
Index: clang/lib/CodeGen/CodeGenModule.cpp
===================================================================
--- clang/lib/CodeGen/CodeGenModule.cpp
+++ clang/lib/CodeGen/CodeGenModule.cpp
@@ -768,6 +768,11 @@
// that might affect the DLL storage class or the visibility, and
// before anything that might act on these.
setVisibilityFromDLLStorageClass(LangOpts, getModule());
+
+ // Remove unused address space casts of global variables.
+ for (auto *Cast : GlobalVarCasts)
+ if (Cast->use_empty())
+ Cast->destroyConstant();
}
void CodeGenModule::EmitOpenCLMetadata() {
@@ -3938,9 +3943,14 @@
: (LangOpts.OpenCL ? LangAS::opencl_global : LangAS::Default);
assert(getContext().getTargetAddressSpace(ExpectedAS) ==
Ty->getPointerAddressSpace());
- if (AddrSpace != ExpectedAS)
- return getTargetCodeGenInfo().performAddrSpaceCast(*this, GV, AddrSpace,
- ExpectedAS, Ty);
+ if (AddrSpace != ExpectedAS) {
+ auto *Cast = getTargetCodeGenInfo().performAddrSpaceCast(
+ *this, GV, AddrSpace, ExpectedAS, Ty);
+ // Record address space casts of global variables for cleaning up if unused.
+ if (Cast != GV)
+ GlobalVarCasts.insert(Cast);
+ return Cast;
+ }
return GV;
}
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
