[clang] [CUDA] make kernel stub ICF-proof (PR #90155)
Yaxun Liu via cfe-commits
cfe-commits at lists.llvm.org
Mon Apr 29 20:54:17 PDT 2024
https://github.com/yxsamliu updated https://github.com/llvm/llvm-project/pull/90155
>From 7a741109a6769e2e91d77a82d9a44c7f9c212c48 Mon Sep 17 00:00:00 2001
From: "Yaxun (Sam) Liu" <yaxun.liu at amd.com>
Date: Thu, 25 Apr 2024 22:23:26 -0400
Subject: [PATCH] [CUDA] make kernel stub ICF-proof
MSVC linker may merge functions which have identical
set of instructions. CUDA uses kernel stub function as key to
look up kernels in device executables. If kernel stub function
for different kernels are merged by ICF, incorrect kernels
will be launched.
To prevent ICF from merging kernel stub functions, an unique
global variable is created for each kernel stub function
and a volatile store is added to the kernel stub function.
This makes the set of instructions in each kernel function
unique.
Fixes: https://github.com/llvm/llvm-project/issues/88883
---
clang/lib/CodeGen/CGCUDANV.cpp | 27 ++++++
clang/test/CodeGenCUDA/kernel-stub-name.cu | 101 +++++++++++++--------
2 files changed, 88 insertions(+), 40 deletions(-)
diff --git a/clang/lib/CodeGen/CGCUDANV.cpp b/clang/lib/CodeGen/CGCUDANV.cpp
index 370642cb3d5364..670bc4bf72cecb 100644
--- a/clang/lib/CodeGen/CGCUDANV.cpp
+++ b/clang/lib/CodeGen/CGCUDANV.cpp
@@ -424,6 +424,33 @@ void CGNVCUDARuntime::emitDeviceStubBodyNew(CodeGenFunction &CGF,
CGM.CreateRuntimeFunction(FTy, LaunchKernelName);
CGF.EmitCall(FI, CGCallee::forDirect(cudaLaunchKernelFn), ReturnValueSlot(),
LaunchKernelArgs);
+
+ // To prevent CUDA device stub functions from being merged by ICF in MSVC
+ // environment, create an unique global variable for each kernel and write to
+ // the variable in the device stub.
+ if (CGM.getContext().getTargetInfo().getCXXABI().isMicrosoft() &&
+ !CGF.getLangOpts().HIP) {
+ llvm::Function *KernelFunction = llvm::cast<llvm::Function>(Kernel);
+ std::string GlobalVarName = (KernelFunction->getName() + ".id").str();
+
+ llvm::GlobalVariable *HandleVar =
+ CGM.getModule().getNamedGlobal(GlobalVarName);
+ if (!HandleVar) {
+ HandleVar = new llvm::GlobalVariable(
+ CGM.getModule(), CGM.Int8Ty,
+ /*Constant=*/false, KernelFunction->getLinkage(),
+ llvm::ConstantInt::get(CGM.Int8Ty, 0), GlobalVarName);
+ HandleVar->setDSOLocal(KernelFunction->isDSOLocal());
+ HandleVar->setVisibility(KernelFunction->getVisibility());
+ if (KernelFunction->hasComdat())
+ HandleVar->setComdat(CGM.getModule().getOrInsertComdat(GlobalVarName));
+ }
+
+ CGF.Builder.CreateAlignedStore(llvm::ConstantInt::get(CGM.Int8Ty, 1),
+ HandleVar, CharUnits::One(),
+ /*IsVolatile=*/true);
+ }
+
CGF.EmitBranch(EndBlock);
CGF.EmitBlock(EndBlock);
diff --git a/clang/test/CodeGenCUDA/kernel-stub-name.cu b/clang/test/CodeGenCUDA/kernel-stub-name.cu
index 23df7f5d721b56..0faea75cbbe536 100644
--- a/clang/test/CodeGenCUDA/kernel-stub-name.cu
+++ b/clang/test/CodeGenCUDA/kernel-stub-name.cu
@@ -2,7 +2,7 @@
// RUN: %clang_cc1 -triple x86_64-linux-gnu -emit-llvm %s \
// RUN: -fcuda-include-gpubinary %t -o - -x hip\
-// RUN: | FileCheck -check-prefixes=CHECK,GNU %s
+// RUN: | FileCheck -check-prefixes=CHECK,GNU,GNU-HIP,HIP %s
// RUN: %clang_cc1 -triple x86_64-linux-gnu -emit-llvm %s \
// RUN: -fcuda-include-gpubinary %t -o - -x hip\
@@ -11,7 +11,12 @@
// RUN: %clang_cc1 -triple x86_64-pc-windows-msvc -emit-llvm %s \
// RUN: -aux-triple amdgcn-amd-amdhsa -fcuda-include-gpubinary \
// RUN: %t -o - -x hip\
-// RUN: | FileCheck -check-prefixes=CHECK,MSVC %s
+// RUN: | FileCheck -check-prefixes=CHECK,MSVC,MSVC-HIP,HIP %s
+
+// RUN: %clang_cc1 -triple x86_64-pc-windows-msvc -emit-llvm %s \
+// RUN: -aux-triple nvptx64 -fcuda-include-gpubinary \
+// RUN: %t -target-sdk-version=9.2 -o - \
+// RUN: | FileCheck -check-prefixes=CHECK,MSVC,CUDA %s
// RUN: %clang_cc1 -triple x86_64-pc-windows-msvc -emit-llvm %s \
// RUN: -aux-triple amdgcn-amd-amdhsa -fcuda-include-gpubinary \
@@ -22,19 +27,23 @@
// Check kernel handles are emitted for non-MSVC target but not for MSVC target.
-// GNU: @[[HCKERN:ckernel]] = constant ptr @[[CSTUB:__device_stub__ckernel]], align 8
-// GNU: @[[HNSKERN:_ZN2ns8nskernelEv]] = constant ptr @[[NSSTUB:_ZN2ns23__device_stub__nskernelEv]], align 8
-// GNU: @[[HTKERN:_Z10kernelfuncIiEvv]] = linkonce_odr constant ptr @[[TSTUB:_Z25__device_stub__kernelfuncIiEvv]], comdat, align 8
-// GNU: @[[HDKERN:_Z11kernel_declv]] = external constant ptr, align 8
-// GNU: @[[HTDKERN:_Z20template_kernel_declIiEvT_]] = external constant ptr, align 8
-
-// MSVC: @[[HCKERN:ckernel]] = dso_local constant ptr @[[CSTUB:__device_stub__ckernel]], align 8
-// MSVC: @[[HNSKERN:"\?nskernel at ns@@YAXXZ.*"]] = dso_local constant ptr @[[NSSTUB:"\?__device_stub__nskernel at ns@@YAXXZ"]], align 8
-// MSVC: @[[HTKERN:"\?\?\$kernelfunc at H@@YAXXZ.*"]] = linkonce_odr dso_local constant ptr @[[TSTUB:"\?\?\$__device_stub__kernelfunc at H@@YAXXZ.*"]], comdat, align 8
-// MSVC: @[[HDKERN:"\?kernel_decl@@YAXXZ.*"]] = external dso_local constant ptr, align 8
-// MSVC: @[[HTDKERN:"\?\?\$template_kernel_decl at H@@YAXH.*"]] = external dso_local constant ptr, align 8
+// GNU-HIP: @[[HCKERN:ckernel]] = constant ptr @[[CSTUB:__device_stub__ckernel]], align 8
+// GNU-HIP: @[[HNSKERN:_ZN2ns8nskernelEv]] = constant ptr @[[NSSTUB:_ZN2ns23__device_stub__nskernelEv]], align 8
+// GNU-HIP: @[[HTKERN:_Z10kernelfuncIiEvv]] = linkonce_odr constant ptr @[[TSTUB:_Z25__device_stub__kernelfuncIiEvv]], comdat, align 8
+// GNU-HIP: @[[HDKERN:_Z11kernel_declv]] = external constant ptr, align 8
+// GNU-HIP: @[[HTDKERN:_Z20template_kernel_declIiEvT_]] = external constant ptr, align 8
+
+// MSVC-HIP: @[[HCKERN:ckernel]] = dso_local constant ptr @[[CSTUB:__device_stub__ckernel]], align 8
+// MSVC-HIP: @[[HNSKERN:"\?nskernel at ns@@YAXXZ.*"]] = dso_local constant ptr @[[NSSTUB:"\?__device_stub__nskernel at ns@@YAXXZ"]], align 8
+// MSVC-HIP: @[[HTKERN:"\?\?\$kernelfunc at H@@YAXXZ.*"]] = linkonce_odr dso_local constant ptr @[[TSTUB:"\?\?\$__device_stub__kernelfunc at H@@YAXXZ.*"]], comdat, align 8
+// MSVC-HIP: @[[HDKERN:"\?kernel_decl@@YAXXZ.*"]] = external dso_local constant ptr, align 8
+// MSVC-HIP: @[[HTDKERN:"\?\?\$template_kernel_decl at H@@YAXH.*"]] = external dso_local constant ptr, align 8
extern "C" __global__ void ckernel() {}
+// CUDA: @[[HCKERN:__device_stub__ckernel\.id]] = dso_local global i8 0
+// CUDA: @[[HNSKERN:"\?__device_stub__nskernel at ns@@YAXXZ\.id"]] = dso_local global i8 0
+// CUDA: @[[HTKERN:"\?\?\$__device_stub__kernelfunc at H@@YAXXZ\.id"]] = linkonce_odr dso_local global i8 0, comdat
+
namespace ns {
__global__ void nskernel() {}
} // namespace ns
@@ -60,18 +69,27 @@ extern "C" void launch(void *kern);
// Non-template kernel stub functions
-// CHECK: define{{.*}}@[[CSTUB]]
-// CHECK: call{{.*}}@hipLaunchByPtr{{.*}}@[[HCKERN]]
+// HIP: define{{.*}}@[[CSTUB]]
+// CUDA: define{{.*}}@[[CSTUB:__device_stub__ckernel]]
+// HIP: call{{.*}}@hipLaunchByPtr{{.*}}@[[HCKERN]]
+// CUDA: call{{.*}}@cudaLaunch{{.*}}@[[CSTUB]]
+// CUDA: store volatile i8 1, ptr @[[HCKERN]], align 1
+// CHECK: ret void
-// CHECK: define{{.*}}@[[NSSTUB]]
-// CHECK: call{{.*}}@hipLaunchByPtr{{.*}}@[[HNSKERN]]
+// HIP: define{{.*}}@[[NSSTUB]]
+// CUDA: define{{.*}}@[[NSSTUB:"\?__device_stub__nskernel at ns@@YAXXZ"]]
+// HIP: call{{.*}}@hipLaunchByPtr{{.*}}@[[HNSKERN]]
+// CUDA: call{{.*}}@cudaLaunch{{.*}}@[[NSSTUB]]
+// CUDA: store volatile i8 1, ptr @[[HNSKERN]], align 1
+// CHECK: ret void
// Check kernel stub is called for triple chevron.
// CHECK-LABEL: define{{.*}}@fun1()
// CHECK: call void @[[CSTUB]]()
// CHECK: call void @[[NSSTUB]]()
-// CHECK: call void @[[TSTUB]]()
+// HIP: call void @[[TSTUB]]()
+// CUDA: call void @[[TSTUB:"\?\?\$__device_stub__kernelfunc at H@@YAXXZ.*"]]()
// GNU: call void @[[DSTUB:_Z26__device_stub__kernel_declv]]()
// GNU: call void @[[TDSTUB:_Z35__device_stub__template_kernel_declIiEvT_]](
// MSVC: call void @[[DSTUB:"\?__device_stub__kernel_decl@@YAXXZ"]]()
@@ -88,7 +106,10 @@ extern "C" void fun1(void) {
// Template kernel stub functions
// CHECK: define{{.*}}@[[TSTUB]]
-// CHECK: call{{.*}}@hipLaunchByPtr{{.*}}@[[HTKERN]]
+// HIP: call{{.*}}@hipLaunchByPtr{{.*}}@[[HTKERN]]
+// CUDA: call{{.*}}@cudaLaunch{{.*}}@[[TSTUB]]
+// CUDA: store volatile i8 1, ptr @[[HTKERN]], align 1
+// CHECK: ret void
// Check declaration of stub function for external kernel.
@@ -98,11 +119,11 @@ extern "C" void fun1(void) {
// Check kernel handle is used for passing the kernel as a function pointer.
// CHECK-LABEL: define{{.*}}@fun2()
-// CHECK: call void @launch({{.*}}[[HCKERN]]
-// CHECK: call void @launch({{.*}}[[HNSKERN]]
-// CHECK: call void @launch({{.*}}[[HTKERN]]
-// CHECK: call void @launch({{.*}}[[HDKERN]]
-// CHECK: call void @launch({{.*}}[[HTDKERN]]
+// HIP: call void @launch({{.*}}[[HCKERN]]
+// HIP: call void @launch({{.*}}[[HNSKERN]]
+// HIP: call void @launch({{.*}}[[HTKERN]]
+// HIP: call void @launch({{.*}}[[HDKERN]]
+// HIP: call void @launch({{.*}}[[HTDKERN]]
extern "C" void fun2() {
launch((void *)ckernel);
launch((void *)ns::nskernel);
@@ -114,10 +135,10 @@ extern "C" void fun2() {
// Check kernel handle is used for assigning a kernel to a function pointer.
// CHECK-LABEL: define{{.*}}@fun3()
-// CHECK: store ptr @[[HCKERN]], ptr @kernel_ptr, align 8
-// CHECK: store ptr @[[HCKERN]], ptr @kernel_ptr, align 8
-// CHECK: store ptr @[[HCKERN]], ptr @void_ptr, align 8
-// CHECK: store ptr @[[HCKERN]], ptr @void_ptr, align 8
+// HIP: store ptr @[[HCKERN]], ptr @kernel_ptr, align 8
+// HIP: store ptr @[[HCKERN]], ptr @kernel_ptr, align 8
+// HIP: store ptr @[[HCKERN]], ptr @void_ptr, align 8
+// HIP: store ptr @[[HCKERN]], ptr @void_ptr, align 8
extern "C" void fun3() {
kernel_ptr = ckernel;
kernel_ptr = &ckernel;
@@ -129,11 +150,11 @@ extern "C" void fun3() {
// used with triple chevron.
// CHECK-LABEL: define{{.*}}@fun4()
-// CHECK: store ptr @[[HCKERN]], ptr @kernel_ptr
-// CHECK: call noundef i32 @{{.*hipConfigureCall}}
-// CHECK: %[[HANDLE:.*]] = load ptr, ptr @kernel_ptr, align 8
-// CHECK: %[[STUB:.*]] = load ptr, ptr %[[HANDLE]], align 8
-// CHECK: call void %[[STUB]]()
+// HIP: store ptr @[[HCKERN]], ptr @kernel_ptr
+// HIP: call noundef i32 @{{.*hipConfigureCall}}
+// HIP: %[[HANDLE:.*]] = load ptr, ptr @kernel_ptr, align 8
+// HIP: %[[STUB:.*]] = load ptr, ptr %[[HANDLE]], align 8
+// HIP: call void %[[STUB]]()
extern "C" void fun4() {
kernel_ptr = ckernel;
kernel_ptr<<<1,1>>>();
@@ -142,9 +163,9 @@ extern "C" void fun4() {
// Check kernel handle is passed to a function.
// CHECK-LABEL: define{{.*}}@fun5()
-// CHECK: store ptr @[[HCKERN]], ptr @kernel_ptr
-// CHECK: %[[HANDLE:.*]] = load ptr, ptr @kernel_ptr, align 8
-// CHECK: call void @launch(ptr noundef %[[HANDLE]])
+// HIP: store ptr @[[HCKERN]], ptr @kernel_ptr
+// HIP: %[[HANDLE:.*]] = load ptr, ptr @kernel_ptr, align 8
+// HIP: call void @launch(ptr noundef %[[HANDLE]])
extern "C" void fun5() {
kernel_ptr = ckernel;
launch((void *)kernel_ptr);
@@ -152,10 +173,10 @@ extern "C" void fun5() {
// Check kernel handle is registered.
-// CHECK-LABEL: define{{.*}}@__hip_register_globals
-// CHECK: call{{.*}}@__hipRegisterFunction{{.*}}@[[HCKERN]]{{.*}}@[[CKERN]]
-// CHECK: call{{.*}}@__hipRegisterFunction{{.*}}@[[HNSKERN]]{{.*}}@[[NSKERN]]
-// CHECK: call{{.*}}@__hipRegisterFunction{{.*}}@[[HTKERN]]{{.*}}@[[TKERN]]
+// HIP-LABEL: define{{.*}}@__hip_register_globals
+// HIP: call{{.*}}@__hipRegisterFunction{{.*}}@[[HCKERN]]{{.*}}@[[CKERN]]
+// HIP: call{{.*}}@__hipRegisterFunction{{.*}}@[[HNSKERN]]{{.*}}@[[NSKERN]]
+// HIP: call{{.*}}@__hipRegisterFunction{{.*}}@[[HTKERN]]{{.*}}@[[TKERN]]
// NEG-NOT: call{{.*}}@__hipRegisterFunction{{.*}}__device_stub
// NEG-NOT: call{{.*}}@__hipRegisterFunction{{.*}}kernel_decl
// NEG-NOT: call{{.*}}@__hipRegisterFunction{{.*}}template_kernel_decl
More information about the cfe-commits
mailing list