[SYCL] Switch to use plain array in sycl::vec in more cases (intel#17656)

AlexeySachkov · web-flow · commit 1cd73d96798e · 2025-03-27T19:50:16.000+01:00
The problem with using `std::array` in `sycl::vec` is that we cannot compile it in some environments (namely, Windows) because the former may use something that is illegal in SYCL device code. intel#17025 fixed that, but only did so under preview breaking changes mode, which does not satisfy some of our customers immediately. This PR introduces two main changes: - it allows to opt-in for new behavior through passing `-D__SYCL_USE_NEW_VEC_IMPL=1` macro without using `-fpreview-breaking-changes` flag. That allows for a more gradual opt-in from customers who are interested in this fix - it switches the imlpementation to use the new approach with C-style arrays if their size and alignment is the same as for the corresponding `std::array` - in that case their memory layout is expected to be absolutely the same and therefore it should be safe to use the new approach without fear of some ABI incompatibilities. This allows for customers to benefit from the fix without specifying any extra macro (which should be the case for the most common platforms out there)
diff --git a/sycl/include/sycl/detail/type_traits/vec_marray_traits.hpp b/sycl/include/sycl/detail/type_traits/vec_marray_traits.hpp
@@ -8,14 +8,6 @@
 
 #pragma once
 
-#ifndef __SYCL_USE_NEW_VEC_IMPL
-#if defined(__INTEL_PREVIEW_BREAKING_CHANGES)
-#define __SYCL_USE_NEW_VEC_IMPL 1
-#else
-#define __SYCL_USE_NEW_VEC_IMPL 0
-#endif
-#endif
-
 #include <cstddef>
 #include <type_traits>
 
diff --git a/sycl/include/sycl/detail/vector_convert.hpp b/sycl/include/sycl/detail/vector_convert.hpp
@@ -57,9 +57,7 @@
 #include <sycl/detail/generic_type_traits.hpp> // for is_sigeninteger, is_s...
 #include <sycl/exception.hpp>                  // for errc
 
-#ifdef __INTEL_PREVIEW_BREAKING_CHANGES
 #include <sycl/detail/memcpy.hpp>
-#endif
 #include <sycl/ext/oneapi/bfloat16.hpp> // bfloat16
 #include <sycl/vector.hpp>
 
diff --git a/sycl/include/sycl/vector.hpp b/sycl/include/sycl/vector.hpp
@@ -22,6 +22,15 @@
 #endif
 #endif // __clang__
 
+// See vec::DataType definitions for more details
+#ifndef __SYCL_USE_PLAIN_ARRAY_AS_VEC_STORAGE
+#if defined(__INTEL_PREVIEW_BREAKING_CHANGES)
+#define __SYCL_USE_PLAIN_ARRAY_AS_VEC_STORAGE 1
+#else
+#define __SYCL_USE_PLAIN_ARRAY_AS_VEC_STORAGE 0
+#endif
+#endif
+
 #if !defined(__HAS_EXT_VECTOR_TYPE__) && defined(__SYCL_DEVICE_ONLY__)
 #error "SYCL device compiler is built without ext_vector_type support"
 #endif
@@ -86,6 +95,9 @@ struct elem {
 };
 
 namespace detail {
+// To be defined in tests, trick to access vec's private methods
+template <typename T1, int T2> class vec_base_test;
+
 template <typename VecT, typename OperationLeftT, typename OperationRightT,
           template <typename> class OperationCurrentT, int... Indexes>
 class SwizzleOp;
@@ -144,16 +156,34 @@ template <typename DataT, int NumElements> class vec_base {
   static constexpr size_t AdjustedNum = (NumElements == 3) ? 4 : NumElements;
   // This represent type of underlying value. There should be only one field
   // in the class, so vec<float, 16> should be equal to float16 in memory.
-#if defined(__INTEL_PREVIEW_BREAKING_CHANGES) &&                               \
-    defined(__SYCL_USE_NEW_VEC_IMPL)
-  using DataType = DataT[AdjustedNum];
+  //
+  // In intel/llvm#14130 we incorrectly used std::array as an underlying storage
+  // for vec data. The problem with std::array is that it comes from the C++
+  // STL headers which we do not control and they may use something that is
+  // illegal in SYCL device code. One of specific examples is use of debug
+  // assertions in MSVC's STL implementation.
+  //
+  // The better approach is to use plain C++ array, but the problem here is that
+  // C++ specification does not provide any guarantees about the memory layout
+  // of std::array and therefore directly switching to it would technically be
+  // an ABI-break, even though the practical chances of encountering the issue
+  // are low.
+  //
+  // To play it safe, we only switch to use plain array if both its size and
+  // alignment match those of std::array, or unless the new behavior is forced
+  // via __SYCL_USE_PLAIN_ARRAY_AS_VEC_STORAGE or preview breaking changes mode.
+  using DataType = std::conditional_t<
+#if __SYCL_USE_PLAIN_ARRAY_AS_VEC_STORAGE
+      true,
 #else
-  using DataType = std::array<DataT, AdjustedNum>;
-  // Assuming that std::array has the same size as the underlying array.
-  // C++ standard does not guarantee that, but it is true for most popular
-  // implementations.
-  static_assert(sizeof(DataType) == sizeof(DataT[AdjustedNum]));
+      sizeof(std::array<DataT, AdjustedNum>) == sizeof(DataT[AdjustedNum]) &&
+          alignof(std::array<DataT, AdjustedNum>) ==
+              alignof(DataT[AdjustedNum]),
 #endif
+      DataT[AdjustedNum], std::array<DataT, AdjustedNum>>;
+
+  // To allow testing of private methods
+  template <typename T1, int T2> friend class detail::vec_base_test;
 
 protected:
   // fields
diff --git a/sycl/test/abi/layout_vec.cpp b/sycl/test/abi/layout_vec.cpp
@@ -12,8 +12,7 @@ SYCL_EXTERNAL void foo(sycl::vec<int, 4>) {}
 
 // CHECK:      0 | class sycl::vec<int, 4>
 // ignore empty base classes
-// CHECK:      0 |   struct std::array<int, 4> m_Data
-// CHECK-NEXT: 0 |     typename {{.+}}::_Type _M_elems
+// CHECK:      0 |   DataType m_Data
 // CHECK-NEXT:   | [sizeof=16, dsize=16, align=16,
 // CHECK-NEXT:   |  nvsize=16, nvalign=16]
 
@@ -23,7 +22,6 @@ SYCL_EXTERNAL void foo(sycl::vec<bool, 16>) {}
 
 // CHECK:      0 | class sycl::vec<_Bool, 16>
 // ignore empty base classes
-// CHECK:      0 |   struct std::array<_Bool, 16> m_Data
-// CHECK-NEXT: 0 |     typename {{.+}}::_Type _M_elems
+// CHECK:      0 |   DataType m_Data
 // CHECK-NEXT:   | [sizeof=16, dsize=16, align=16,
 // CHECK-NEXT:   |  nvsize=16, nvalign=16]
diff --git a/sycl/test/basic_tests/vectors/storage.cpp b/sycl/test/basic_tests/vectors/storage.cpp
@@ -0,0 +1,43 @@
+// RUN: %clangxx -fsycl -Xclang -verify %s -fsyntax-only
+// RUN: %clangxx -fsycl -Xclang -verify %s -fsyntax-only -fpreview-breaking-changes
+// RUN: %clangxx -fsycl -Xclang -verify %s -fsyntax-only -D__SYCL_USE_PLAIN_ARRAY_AS_VEC_STORAGE=1
+// expected-no-diagnostics
+
+#include <sycl/vector.hpp>
+
+#include <type_traits>
+
+namespace sycl {
+namespace detail {
+template <typename T, int N> class vec_base_test {
+public:
+  static void do_check() {
+    constexpr bool uses_std_array =
+        std::is_same_v<typename sycl::vec<T, N>::DataType, std::array<T, N>>;
+    constexpr bool uses_plain_array =
+        std::is_same_v<typename sycl::vec<T, N>::DataType, T[N]>;
+
+    constexpr bool std_array_and_plain_array_have_the_same_layout =
+        sizeof(std::array<T, N>) == sizeof(T[N]) &&
+        alignof(std::array<T, N>) == alignof(T[N]);
+
+#if defined(__INTEL_PREVIEW_BREAKING_CHANGES) ||                               \
+    __SYCL_USE_PLAIN_ARRAY_AS_VEC_STORAGE
+    static_assert(uses_plain_array,
+                  "We must use plain array regardless of "
+                  "layout, because user is opted-in for a potential ABI-break");
+#else
+    static_assert(std_array_and_plain_array_have_the_same_layout ==
+                      uses_plain_array,
+                  "If layouts are the same, we must use safer plain array "
+                  "instead of std::array, or vice versa");
+    static_assert(
+        !std_array_and_plain_array_have_the_same_layout == uses_std_array,
+        "If layouts are not the same, we must use std::array to preserve ABI");
+#endif
+  }
+};
+} // namespace detail
+} // namespace sycl
+
+int main() { sycl::detail::vec_base_test<int, 4>::do_check(); }
diff --git a/sycl/test/check_device_code/vector/bf16_builtins.cpp b/sycl/test/check_device_code/vector/bf16_builtins.cpp
@@ -69,7 +69,7 @@ SYCL_EXTERNAL auto TestFMin(vec<bfloat16, 2> a, vec<bfloat16, 2> b) {
 }
 
 // CHECK-LABEL: define dso_local spir_func void @_Z8TestFMaxN4sycl3_V13vecINS0_3ext6oneapi8bfloat16ELi3EEES5_(
-// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.71") align 8 captures(none) initializes((0, 8)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.71") align 8 captures(none) [[A:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.71") align 8 captures(none) [[B:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META24:![0-9]+]] !sycl_fixed_targets [[META7]] {
+// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.70") align 8 captures(none) initializes((0, 8)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.70") align 8 captures(none) [[A:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.70") align 8 captures(none) [[B:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META24:![0-9]+]] !sycl_fixed_targets [[META7]] {
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[VEC_ADDR_I_I_I_I12_I:%.*]] = alloca <3 x float>, align 16
 // CHECK-NEXT:    [[DST_I_I_I_I13_I:%.*]] = alloca [4 x %"class.sycl::_V1::ext::oneapi::bfloat16"], align 2
@@ -123,7 +123,7 @@ SYCL_EXTERNAL auto TestFMax(vec<bfloat16, 3> a, vec<bfloat16, 3> b) {
 }
 
 // CHECK-LABEL: define dso_local spir_func void @_Z9TestIsNanN4sycl3_V13vecINS0_3ext6oneapi8bfloat16ELi4EEE(
-// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.149") align 8 captures(none) initializes((0, 8)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.188") align 8 captures(none) [[A:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META38:![0-9]+]] !sycl_fixed_targets [[META7]] {
+// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.146") align 8 captures(none) initializes((0, 8)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.184") align 8 captures(none) [[A:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META38:![0-9]+]] !sycl_fixed_targets [[META7]] {
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[VEC_ADDR_I_I_I_I_I:%.*]] = alloca <4 x i16>, align 8
 // CHECK-NEXT:    [[DST_I_I_I_I_I:%.*]] = alloca [4 x float], align 4
@@ -149,7 +149,7 @@ SYCL_EXTERNAL auto TestIsNan(vec<bfloat16, 4> a) {
 }
 
 // CHECK-LABEL: define dso_local spir_func void @_Z8TestFabsN4sycl3_V13vecINS0_3ext6oneapi8bfloat16ELi8EEE(
-// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.342") align 16 captures(none) initializes((0, 16)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.342") align 16 captures(none) [[A:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META48:![0-9]+]] !sycl_fixed_targets [[META7]] {
+// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.336") align 16 captures(none) initializes((0, 16)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.336") align 16 captures(none) [[A:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META48:![0-9]+]] !sycl_fixed_targets [[META7]] {
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[VEC_ADDR_I_I_I_I2_I:%.*]] = alloca <8 x float>, align 32
 // CHECK-NEXT:    [[DST_I_I_I_I3_I:%.*]] = alloca [8 x %"class.sycl::_V1::ext::oneapi::bfloat16"], align 2
@@ -185,7 +185,7 @@ SYCL_EXTERNAL auto TestFabs(vec<bfloat16, 8> a) {
 }
 
 // CHECK-LABEL: define dso_local spir_func void @_Z8TestCeilN4sycl3_V13vecINS0_3ext6oneapi8bfloat16ELi8EEE(
-// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.342") align 16 captures(none) initializes((0, 16)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.342") align 16 captures(none) [[A:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META59:![0-9]+]] !sycl_fixed_targets [[META7]] {
+// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.336") align 16 captures(none) initializes((0, 16)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.336") align 16 captures(none) [[A:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META59:![0-9]+]] !sycl_fixed_targets [[META7]] {
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[VEC_ADDR_I_I_I_I2_I:%.*]] = alloca <8 x float>, align 32
 // CHECK-NEXT:    [[DST_I_I_I_I3_I:%.*]] = alloca [8 x %"class.sycl::_V1::ext::oneapi::bfloat16"], align 2
@@ -221,7 +221,7 @@ SYCL_EXTERNAL auto TestCeil(vec<bfloat16, 8> a) {
 }
 
 // CHECK-LABEL: define dso_local spir_func void @_Z7TestFMAN4sycl3_V13vecINS0_3ext6oneapi8bfloat16ELi16EEES5_S5_(
-// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.420") align 32 captures(none) initializes((0, 32)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.420") align 32 captures(none) [[A:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.420") align 32 captures(none) [[B:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.420") align 32 captures(none) [[C:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META70:![0-9]+]] !sycl_fixed_targets [[META7]] {
+// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.412") align 32 captures(none) initializes((0, 32)) [[AGG_RESULT:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.412") align 32 captures(none) [[A:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.412") align 32 captures(none) [[B:%.*]], ptr noundef readonly byval(%"class.sycl::_V1::vec.412") align 32 captures(none) [[C:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META70:![0-9]+]] !sycl_fixed_targets [[META7]] {
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    [[VEC_ADDR_I_I_I_I14_I:%.*]] = alloca <16 x float>, align 64
 // CHECK-NEXT:    [[DST_I_I_I_I15_I:%.*]] = alloca [16 x %"class.sycl::_V1::ext::oneapi::bfloat16"], align 2
diff --git a/sycl/test/check_device_code/vector/convert_bfloat.cpp b/sycl/test/check_device_code/vector/convert_bfloat.cpp
@@ -63,7 +63,7 @@ SYCL_EXTERNAL auto TestBFtoFDeviceRZ(vec<bfloat16, 3> &inp) {
 }
 
 // CHECK-LABEL: define dso_local spir_func void @_Z19TestBFtointDeviceRZRN4sycl3_V13vecINS0_3ext6oneapi8bfloat16ELi3EEE(
-// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.71") align 16 captures(none) [[AGG_RESULT:%.*]], ptr addrspace(4) noundef readonly align 8 captures(none) dereferenceable(8) [[INP:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META18:![0-9]+]] !sycl_fixed_targets [[META7]] {
+// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.70") align 16 captures(none) [[AGG_RESULT:%.*]], ptr addrspace(4) noundef readonly align 8 captures(none) dereferenceable(8) [[INP:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META18:![0-9]+]] !sycl_fixed_targets [[META7]] {
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    tail call void @llvm.experimental.noalias.scope.decl(metadata [[META19:![0-9]+]])
 // CHECK-NEXT:    [[LOADVECN_I_I:%.*]] = load <4 x i16>, ptr addrspace(4) [[INP]], align 8, !noalias [[META19]]
@@ -90,7 +90,7 @@ SYCL_EXTERNAL auto TestBFtointDeviceRZ(vec<bfloat16, 3> &inp) {
 }
 
 // CHECK-LABEL: define dso_local spir_func void @_Z20TestBFtointDeviceRNERN4sycl3_V13vecINS0_3ext6oneapi8bfloat16ELi1EEE(
-// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.110") align 4 captures(none) initializes((0, 4)) [[AGG_RESULT:%.*]], ptr addrspace(4) noundef readonly align 2 captures(none) dereferenceable(2) [[INP:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META24:![0-9]+]] !sycl_fixed_targets [[META7]] {
+// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.108") align 4 captures(none) initializes((0, 4)) [[AGG_RESULT:%.*]], ptr addrspace(4) noundef readonly align 2 captures(none) dereferenceable(2) [[INP:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META24:![0-9]+]] !sycl_fixed_targets [[META7]] {
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    tail call void @llvm.experimental.noalias.scope.decl(metadata [[META25:![0-9]+]])
 // CHECK-NEXT:    [[TMP0:%.*]] = load i16, ptr addrspace(4) [[INP]], align 2, !tbaa [[TBAA11]], !noalias [[META25]]
@@ -182,7 +182,7 @@ SYCL_EXTERNAL auto TestInttoBFDeviceRZ(vec<int, 3> &inp) {
 }
 
 // CHECK-LABEL: define dso_local spir_func void @_Z19TestLLtoBFDeviceRTPRN4sycl3_V13vecIxLi1EEE(
-// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.149") align 2 captures(none) initializes((0, 2)) [[AGG_RESULT:%.*]], ptr addrspace(4) noundef readonly align 8 captures(none) dereferenceable(8) [[INP:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META42:![0-9]+]] !sycl_fixed_targets [[META7]] {
+// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.146") align 2 captures(none) initializes((0, 2)) [[AGG_RESULT:%.*]], ptr addrspace(4) noundef readonly align 8 captures(none) dereferenceable(8) [[INP:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META42:![0-9]+]] !sycl_fixed_targets [[META7]] {
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    tail call void @llvm.experimental.noalias.scope.decl(metadata [[META43:![0-9]+]])
 // CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr addrspace(4) [[INP]], align 8, !tbaa [[TBAA46:![0-9]+]], !noalias [[META43]]
@@ -195,7 +195,7 @@ SYCL_EXTERNAL auto TestLLtoBFDeviceRTP(vec<long long, 1> &inp) {
 }
 
 // CHECK-LABEL: define dso_local spir_func void @_Z22TestShorttoBFDeviceRTNRN4sycl3_V13vecIsLi2EEE(
-// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.229") align 4 captures(none) [[AGG_RESULT:%.*]], ptr addrspace(4) noundef readonly align 4 captures(none) dereferenceable(4) [[INP:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META48:![0-9]+]] !sycl_fixed_targets [[META7]] {
+// CHECK-SAME: ptr addrspace(4) dead_on_unwind noalias writable writeonly sret(%"class.sycl::_V1::vec.224") align 4 captures(none) [[AGG_RESULT:%.*]], ptr addrspace(4) noundef readonly align 4 captures(none) dereferenceable(4) [[INP:%.*]]) local_unnamed_addr #[[ATTR0]] !srcloc [[META48:![0-9]+]] !sycl_fixed_targets [[META7]] {
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    tail call void @llvm.experimental.noalias.scope.decl(metadata [[META49:![0-9]+]])
 // CHECK-NEXT:    [[TMP0:%.*]] = load <2 x i16>, ptr addrspace(4) [[INP]], align 4, !tbaa [[TBAA11]], !noalias [[META49]]
diff --git a/sycl/test/check_device_code/vector/math_ops.cpp b/sycl/test/check_device_code/vector/math_ops.cpp
diff --git a/sycl/test/regression/vec_array_windows.cpp b/sycl/test/regression/vec_array_windows.cpp
