Author: Vikash Gupta Date: 2024-07-26T15:24:06+05:30 New Revision: d65f0375910f6dfeaa94c1060e7d81fb7bdf76be
URL: https://github.com/llvm/llvm-project/commit/d65f0375910f6dfeaa94c1060e7d81fb7bdf76be DIFF: https://github.com/llvm/llvm-project/commit/d65f0375910f6dfeaa94c1060e7d81fb7bdf76be.diff LOG: [Clang] Use private address space for builtin_alloca return type for OpenCL (#95750) The __builtin_alloca was returning a flat pointer with no address space when compiled using openCL1.2 or below but worked fine with openCL2.0 and above. This accounts to the fact that later uses the concept of generic address space which supports cast to other address space(i.e to private address space which is used for stack allocation) . But, in actuality, as it returns pointer to the stack, it should be pointing to private address space irrespective of openCL version becuase builtin_alloca allocates stack memory used for current function in which it is called. Thus,it requires redefintion of the builtin function with appropraite return pointer to private address space. Added: clang/test/CodeGenOpenCL/builtins-alloca.cl Modified: clang/lib/Sema/SemaChecking.cpp Removed: ################################################################################ diff --git a/clang/lib/Sema/SemaChecking.cpp b/clang/lib/Sema/SemaChecking.cpp index 07a34fda494d8..bb30b1e289a1c 100644 --- a/clang/lib/Sema/SemaChecking.cpp +++ b/clang/lib/Sema/SemaChecking.cpp @@ -1477,6 +1477,18 @@ static bool BuiltinSEHScopeCheck(Sema &SemaRef, CallExpr *TheCall, return false; } +// In OpenCL, __builtin_alloca_* should return a pointer to address space +// that corresponds to the stack address space i.e private address space. +static void builtinAllocaAddrSpace(Sema &S, CallExpr *TheCall) { + QualType RT = TheCall->getType(); + assert((RT->isPointerType() && !(RT->getPointeeType().hasAddressSpace())) && + "__builtin_alloca has invalid address space"); + + RT = RT->getPointeeType(); + RT = S.Context.getAddrSpaceQualType(RT, LangAS::opencl_private); + TheCall->setType(S.Context.getPointerType(RT)); +} + namespace { enum PointerAuthOpKind { PAO_Strip, @@ -2214,6 +2226,9 @@ Sema::CheckBuiltinFunctionCall(FunctionDecl *FDecl, unsigned BuiltinID, case Builtin::BI__builtin_alloca_uninitialized: Diag(TheCall->getBeginLoc(), diag::warn_alloca) << TheCall->getDirectCallee(); + if (getLangOpts().OpenCL) { + builtinAllocaAddrSpace(*this, TheCall); + } break; case Builtin::BI__arithmetic_fence: if (BuiltinArithmeticFence(TheCall)) diff --git a/clang/test/CodeGenOpenCL/builtins-alloca.cl b/clang/test/CodeGenOpenCL/builtins-alloca.cl new file mode 100644 index 0000000000000..474e95e74e006 --- /dev/null +++ b/clang/test/CodeGenOpenCL/builtins-alloca.cl @@ -0,0 +1,141 @@ +// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5 +// RUN: %clang_cc1 %s -O0 -triple amdgcn-amd-amdhsa -cl-std=CL1.2 \ +// RUN: -emit-llvm -o - | FileCheck --check-prefixes=OPENCL %s +// RUN: %clang_cc1 %s -O0 -triple amdgcn-amd-amdhsa -cl-std=CL2.0 \ +// RUN: -emit-llvm -o - | FileCheck --check-prefixes=OPENCL %s +// RUN: %clang_cc1 %s -O0 -triple amdgcn-amd-amdhsa -cl-std=CL3.0 \ +// RUN: -emit-llvm -o - | FileCheck --check-prefixes=OPENCL %s +// RUN: %clang_cc1 %s -O0 -triple amdgcn-amd-amdhsa -cl-std=CL3.0 -cl-ext=+__opencl_c_generic_address_space \ +// RUN: -emit-llvm -o - | FileCheck --check-prefixes=OPENCL %s + +// OPENCL-LABEL: define dso_local void @test1_builtin_alloca( +// OPENCL-SAME: i32 noundef [[N:%.*]]) #[[ATTR0:[0-9]+]] { +// OPENCL-NEXT: [[ENTRY:.*:]] +// OPENCL-NEXT: [[N_ADDR:%.*]] = alloca i32, align 4, addrspace(5) +// OPENCL-NEXT: [[ALLOC_PTR:%.*]] = alloca ptr addrspace(5), align 4, addrspace(5) +// OPENCL-NEXT: store i32 [[N]], ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[TMP0:%.*]] = load i32, ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[CONV:%.*]] = zext i32 [[TMP0]] to i64 +// OPENCL-NEXT: [[MUL:%.*]] = mul i64 [[CONV]], 4 +// OPENCL-NEXT: [[TMP1:%.*]] = alloca i8, i64 [[MUL]], align 8, addrspace(5) +// OPENCL-NEXT: store ptr addrspace(5) [[TMP1]], ptr addrspace(5) [[ALLOC_PTR]], align 4 +// OPENCL-NEXT: ret void +// +void test1_builtin_alloca(unsigned n) { + __private float* alloc_ptr = (__private float*)__builtin_alloca(n*sizeof(int)); +} + +// OPENCL-LABEL: define dso_local void @test1_builtin_alloca_uninitialized( +// OPENCL-SAME: i32 noundef [[N:%.*]]) #[[ATTR0]] { +// OPENCL-NEXT: [[ENTRY:.*:]] +// OPENCL-NEXT: [[N_ADDR:%.*]] = alloca i32, align 4, addrspace(5) +// OPENCL-NEXT: [[ALLOC_PTR_UNINITIALIZED:%.*]] = alloca ptr addrspace(5), align 4, addrspace(5) +// OPENCL-NEXT: store i32 [[N]], ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[TMP0:%.*]] = load i32, ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[CONV:%.*]] = zext i32 [[TMP0]] to i64 +// OPENCL-NEXT: [[MUL:%.*]] = mul i64 [[CONV]], 4 +// OPENCL-NEXT: [[TMP1:%.*]] = alloca i8, i64 [[MUL]], align 8, addrspace(5) +// OPENCL-NEXT: store ptr addrspace(5) [[TMP1]], ptr addrspace(5) [[ALLOC_PTR_UNINITIALIZED]], align 4 +// OPENCL-NEXT: ret void +// +void test1_builtin_alloca_uninitialized(unsigned n) { + __private float* alloc_ptr_uninitialized = (__private float*)__builtin_alloca_uninitialized(n*sizeof(int)); +} + +// OPENCL-LABEL: define dso_local void @test1_builtin_alloca_with_align( +// OPENCL-SAME: i32 noundef [[N:%.*]]) #[[ATTR0]] { +// OPENCL-NEXT: [[ENTRY:.*:]] +// OPENCL-NEXT: [[N_ADDR:%.*]] = alloca i32, align 4, addrspace(5) +// OPENCL-NEXT: [[ALLOC_PTR_ALIGN:%.*]] = alloca ptr addrspace(5), align 4, addrspace(5) +// OPENCL-NEXT: store i32 [[N]], ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[TMP0:%.*]] = load i32, ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[CONV:%.*]] = zext i32 [[TMP0]] to i64 +// OPENCL-NEXT: [[MUL:%.*]] = mul i64 [[CONV]], 4 +// OPENCL-NEXT: [[TMP1:%.*]] = alloca i8, i64 [[MUL]], align 1, addrspace(5) +// OPENCL-NEXT: store ptr addrspace(5) [[TMP1]], ptr addrspace(5) [[ALLOC_PTR_ALIGN]], align 4 +// OPENCL-NEXT: ret void +// +void test1_builtin_alloca_with_align(unsigned n) { + __private float* alloc_ptr_align = (__private float*)__builtin_alloca_with_align((n*sizeof(int)), 8); +} + +// OPENCL-LABEL: define dso_local void @test1_builtin_alloca_with_align_uninitialized( +// OPENCL-SAME: i32 noundef [[N:%.*]]) #[[ATTR0]] { +// OPENCL-NEXT: [[ENTRY:.*:]] +// OPENCL-NEXT: [[N_ADDR:%.*]] = alloca i32, align 4, addrspace(5) +// OPENCL-NEXT: [[ALLOC_PTR_ALIGN_UNINITIALIZED:%.*]] = alloca ptr addrspace(5), align 4, addrspace(5) +// OPENCL-NEXT: store i32 [[N]], ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[TMP0:%.*]] = load i32, ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[CONV:%.*]] = zext i32 [[TMP0]] to i64 +// OPENCL-NEXT: [[MUL:%.*]] = mul i64 [[CONV]], 4 +// OPENCL-NEXT: [[TMP1:%.*]] = alloca i8, i64 [[MUL]], align 1, addrspace(5) +// OPENCL-NEXT: store ptr addrspace(5) [[TMP1]], ptr addrspace(5) [[ALLOC_PTR_ALIGN_UNINITIALIZED]], align 4 +// OPENCL-NEXT: ret void +// +void test1_builtin_alloca_with_align_uninitialized(unsigned n) { + __private float* alloc_ptr_align_uninitialized = (__private float*)__builtin_alloca_with_align_uninitialized((n*sizeof(int)), 8); +} + +// OPENCL-LABEL: define dso_local void @test2_builtin_alloca( +// OPENCL-SAME: i32 noundef [[N:%.*]]) #[[ATTR0]] { +// OPENCL-NEXT: [[ENTRY:.*:]] +// OPENCL-NEXT: [[N_ADDR:%.*]] = alloca i32, align 4, addrspace(5) +// OPENCL-NEXT: [[ALLOC_PTR:%.*]] = alloca ptr addrspace(5), align 4, addrspace(5) +// OPENCL-NEXT: store i32 [[N]], ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[TMP0:%.*]] = load i32, ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[CONV:%.*]] = zext i32 [[TMP0]] to i64 +// OPENCL-NEXT: [[TMP1:%.*]] = alloca i8, i64 [[CONV]], align 8, addrspace(5) +// OPENCL-NEXT: store ptr addrspace(5) [[TMP1]], ptr addrspace(5) [[ALLOC_PTR]], align 4 +// OPENCL-NEXT: ret void +// +void test2_builtin_alloca(unsigned n) { + __private void *alloc_ptr = __builtin_alloca(n); +} + +// OPENCL-LABEL: define dso_local void @test2_builtin_alloca_uninitialized( +// OPENCL-SAME: i32 noundef [[N:%.*]]) #[[ATTR0]] { +// OPENCL-NEXT: [[ENTRY:.*:]] +// OPENCL-NEXT: [[N_ADDR:%.*]] = alloca i32, align 4, addrspace(5) +// OPENCL-NEXT: [[ALLOC_PTR_UNINITIALIZED:%.*]] = alloca ptr addrspace(5), align 4, addrspace(5) +// OPENCL-NEXT: store i32 [[N]], ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[TMP0:%.*]] = load i32, ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[CONV:%.*]] = zext i32 [[TMP0]] to i64 +// OPENCL-NEXT: [[TMP1:%.*]] = alloca i8, i64 [[CONV]], align 8, addrspace(5) +// OPENCL-NEXT: store ptr addrspace(5) [[TMP1]], ptr addrspace(5) [[ALLOC_PTR_UNINITIALIZED]], align 4 +// OPENCL-NEXT: ret void +// +void test2_builtin_alloca_uninitialized(unsigned n) { + __private void *alloc_ptr_uninitialized = __builtin_alloca_uninitialized(n); +} + +// OPENCL-LABEL: define dso_local void @test2_builtin_alloca_with_align( +// OPENCL-SAME: i32 noundef [[N:%.*]]) #[[ATTR0]] { +// OPENCL-NEXT: [[ENTRY:.*:]] +// OPENCL-NEXT: [[N_ADDR:%.*]] = alloca i32, align 4, addrspace(5) +// OPENCL-NEXT: [[ALLOC_PTR_ALIGN:%.*]] = alloca ptr addrspace(5), align 4, addrspace(5) +// OPENCL-NEXT: store i32 [[N]], ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[TMP0:%.*]] = load i32, ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[CONV:%.*]] = zext i32 [[TMP0]] to i64 +// OPENCL-NEXT: [[TMP1:%.*]] = alloca i8, i64 [[CONV]], align 1, addrspace(5) +// OPENCL-NEXT: store ptr addrspace(5) [[TMP1]], ptr addrspace(5) [[ALLOC_PTR_ALIGN]], align 4 +// OPENCL-NEXT: ret void +// +void test2_builtin_alloca_with_align(unsigned n) { + __private void *alloc_ptr_align = __builtin_alloca_with_align(n, 8); +} + +// OPENCL-LABEL: define dso_local void @test2_builtin_alloca_with_align_uninitialized( +// OPENCL-SAME: i32 noundef [[N:%.*]]) #[[ATTR0]] { +// OPENCL-NEXT: [[ENTRY:.*:]] +// OPENCL-NEXT: [[N_ADDR:%.*]] = alloca i32, align 4, addrspace(5) +// OPENCL-NEXT: [[ALLOC_PTR_ALIGN_UNINITIALIZED:%.*]] = alloca ptr addrspace(5), align 4, addrspace(5) +// OPENCL-NEXT: store i32 [[N]], ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[TMP0:%.*]] = load i32, ptr addrspace(5) [[N_ADDR]], align 4 +// OPENCL-NEXT: [[CONV:%.*]] = zext i32 [[TMP0]] to i64 +// OPENCL-NEXT: [[TMP1:%.*]] = alloca i8, i64 [[CONV]], align 1, addrspace(5) +// OPENCL-NEXT: store ptr addrspace(5) [[TMP1]], ptr addrspace(5) [[ALLOC_PTR_ALIGN_UNINITIALIZED]], align 4 +// OPENCL-NEXT: ret void +// +void test2_builtin_alloca_with_align_uninitialized(unsigned n) { + __private void *alloc_ptr_align_uninitialized = __builtin_alloca_with_align_uninitialized(n, 8); +} _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
