ABataev created this revision.
ABataev added a reviewer: jdoerfert.
Herald added subscribers: guansong, yaxunl, jholewinski.
Herald added a project: clang.
If the original variable is marked for allocation in the different
address space using #pragma omp allocate, need to cast the allocated
variable to its original type with the original address space.
Otherwise, the compiler may crash trying to bitcast the type of the new
allocated variable to the original type in some cases, like passing this
variable as an argument in function calls.
Repository:
rG LLVM Github Monorepo
https://reviews.llvm.org/D83696
Files:
clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
clang/test/OpenMP/nvptx_allocate_codegen.cpp
Index: clang/test/OpenMP/nvptx_allocate_codegen.cpp
===================================================================
--- clang/test/OpenMP/nvptx_allocate_codegen.cpp
+++ clang/test/OpenMP/nvptx_allocate_codegen.cpp
@@ -101,7 +101,7 @@
// CHECK: alloca float,
// CHECK-NOT: alloca double,
// CHECK: load float, float* %
-// CHECK: store double {{.+}}, double addrspace(3)* @bar_b,
+// CHECK: store double {{.+}}, double* addrspacecast (double addrspace(3)*
@bar_b to double*),
}
#pragma omp end declare target
Index: clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
===================================================================
--- clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -4770,6 +4770,7 @@
const VarDecl *VD) {
if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) {
const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
+ auto AS = LangAS::Default;
switch (A->getAllocatorType()) {
// Use the default allocator here as by default local vars are
// threadlocal.
@@ -4783,42 +4784,30 @@
case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
// TODO: implement aupport for user-defined allocators.
return Address::invalid();
- case OMPAllocateDeclAttr::OMPConstMemAlloc: {
- llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
- auto *GV = new llvm::GlobalVariable(
- CGM.getModule(), VarTy, /*isConstant=*/false,
- llvm::GlobalValue::InternalLinkage,
- llvm::Constant::getNullValue(VarTy), VD->getName(),
- /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
- CGM.getContext().getTargetAddressSpace(LangAS::cuda_constant));
- CharUnits Align = CGM.getContext().getDeclAlign(VD);
- GV->setAlignment(Align.getAsAlign());
- return Address(GV, Align);
- }
- case OMPAllocateDeclAttr::OMPPTeamMemAlloc: {
- llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
- auto *GV = new llvm::GlobalVariable(
- CGM.getModule(), VarTy, /*isConstant=*/false,
- llvm::GlobalValue::InternalLinkage,
- llvm::Constant::getNullValue(VarTy), VD->getName(),
- /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
- CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared));
- CharUnits Align = CGM.getContext().getDeclAlign(VD);
- GV->setAlignment(Align.getAsAlign());
- return Address(GV, Align);
- }
+ case OMPAllocateDeclAttr::OMPConstMemAlloc:
+ AS = LangAS::cuda_constant;
+ break;
+ case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
+ AS = LangAS::cuda_shared;
+ break;
case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
- case OMPAllocateDeclAttr::OMPCGroupMemAlloc: {
- llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
- auto *GV = new llvm::GlobalVariable(
- CGM.getModule(), VarTy, /*isConstant=*/false,
- llvm::GlobalValue::InternalLinkage,
- llvm::Constant::getNullValue(VarTy), VD->getName());
- CharUnits Align = CGM.getContext().getDeclAlign(VD);
- GV->setAlignment(Align.getAsAlign());
- return Address(GV, Align);
- }
+ case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
+ break;
}
+ llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
+ auto *GV = new llvm::GlobalVariable(
+ CGM.getModule(), VarTy, /*isConstant=*/false,
+ llvm::GlobalValue::InternalLinkage,
llvm::Constant::getNullValue(VarTy),
+ VD->getName(),
+ /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
+ CGM.getContext().getTargetAddressSpace(AS));
+ CharUnits Align = CGM.getContext().getDeclAlign(VD);
+ GV->setAlignment(Align.getAsAlign());
+ return Address(
+ CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
+ GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace(
+ VD->getType().getAddressSpace()))),
+ Align);
}
if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic)
Index: clang/test/OpenMP/nvptx_allocate_codegen.cpp
===================================================================
--- clang/test/OpenMP/nvptx_allocate_codegen.cpp
+++ clang/test/OpenMP/nvptx_allocate_codegen.cpp
@@ -101,7 +101,7 @@
// CHECK: alloca float,
// CHECK-NOT: alloca double,
// CHECK: load float, float* %
-// CHECK: store double {{.+}}, double addrspace(3)* @bar_b,
+// CHECK: store double {{.+}}, double* addrspacecast (double addrspace(3)* @bar_b to double*),
}
#pragma omp end declare target
Index: clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
===================================================================
--- clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -4770,6 +4770,7 @@
const VarDecl *VD) {
if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) {
const auto *A = VD->getAttr<OMPAllocateDeclAttr>();
+ auto AS = LangAS::Default;
switch (A->getAllocatorType()) {
// Use the default allocator here as by default local vars are
// threadlocal.
@@ -4783,42 +4784,30 @@
case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc:
// TODO: implement aupport for user-defined allocators.
return Address::invalid();
- case OMPAllocateDeclAttr::OMPConstMemAlloc: {
- llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
- auto *GV = new llvm::GlobalVariable(
- CGM.getModule(), VarTy, /*isConstant=*/false,
- llvm::GlobalValue::InternalLinkage,
- llvm::Constant::getNullValue(VarTy), VD->getName(),
- /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
- CGM.getContext().getTargetAddressSpace(LangAS::cuda_constant));
- CharUnits Align = CGM.getContext().getDeclAlign(VD);
- GV->setAlignment(Align.getAsAlign());
- return Address(GV, Align);
- }
- case OMPAllocateDeclAttr::OMPPTeamMemAlloc: {
- llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
- auto *GV = new llvm::GlobalVariable(
- CGM.getModule(), VarTy, /*isConstant=*/false,
- llvm::GlobalValue::InternalLinkage,
- llvm::Constant::getNullValue(VarTy), VD->getName(),
- /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
- CGM.getContext().getTargetAddressSpace(LangAS::cuda_shared));
- CharUnits Align = CGM.getContext().getDeclAlign(VD);
- GV->setAlignment(Align.getAsAlign());
- return Address(GV, Align);
- }
+ case OMPAllocateDeclAttr::OMPConstMemAlloc:
+ AS = LangAS::cuda_constant;
+ break;
+ case OMPAllocateDeclAttr::OMPPTeamMemAlloc:
+ AS = LangAS::cuda_shared;
+ break;
case OMPAllocateDeclAttr::OMPLargeCapMemAlloc:
- case OMPAllocateDeclAttr::OMPCGroupMemAlloc: {
- llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
- auto *GV = new llvm::GlobalVariable(
- CGM.getModule(), VarTy, /*isConstant=*/false,
- llvm::GlobalValue::InternalLinkage,
- llvm::Constant::getNullValue(VarTy), VD->getName());
- CharUnits Align = CGM.getContext().getDeclAlign(VD);
- GV->setAlignment(Align.getAsAlign());
- return Address(GV, Align);
- }
+ case OMPAllocateDeclAttr::OMPCGroupMemAlloc:
+ break;
}
+ llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType());
+ auto *GV = new llvm::GlobalVariable(
+ CGM.getModule(), VarTy, /*isConstant=*/false,
+ llvm::GlobalValue::InternalLinkage, llvm::Constant::getNullValue(VarTy),
+ VD->getName(),
+ /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal,
+ CGM.getContext().getTargetAddressSpace(AS));
+ CharUnits Align = CGM.getContext().getDeclAlign(VD);
+ GV->setAlignment(Align.getAsAlign());
+ return Address(
+ CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
+ GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace(
+ VD->getType().getAddressSpace()))),
+ Align);
}
if (getDataSharingMode(CGM) != CGOpenMPRuntimeNVPTX::Generic)
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
