[libcxx-commits] [libcxx] 0ea40bf - Revert "[libc++] Explicitly convert to masks in SIMD code (#107983)"
Thurston Dang via libcxx-commits
libcxx-commits at lists.llvm.org
Tue Sep 17 14:53:37 PDT 2024
Author: Thurston Dang
Date: 2024-09-17T21:52:33Z
New Revision: 0ea40bf02138c02e7680ce6fa8169502f2a8bd42
URL: https://github.com/llvm/llvm-project/commit/0ea40bf02138c02e7680ce6fa8169502f2a8bd42
DIFF: https://github.com/llvm/llvm-project/commit/0ea40bf02138c02e7680ce6fa8169502f2a8bd42.diff
LOG: Revert "[libc++] Explicitly convert to masks in SIMD code (#107983)"
This reverts commit 1603f99a37c5b179a21dbb8000c39a471a950927.
Reason: buildbot breakage e.g., https://lab.llvm.org/buildbot/#/builders/55/builds/2061
llvm-libc++-shared.cfg.in :: std/algorithms/alg.nonmodifying/alg.starts_with/ranges.starts_with.pass.cpp
llvm-libc++-shared.cfg.in :: std/algorithms/alg.nonmodifying/mismatch/mismatch.pass.cpp
llvm-libc++-shared.cfg.in :: std/algorithms/alg.nonmodifying/mismatch/ranges_mismatch.pass.cpp
...
(Buildbot re-run passed with the previous revision, 1fc288bf481726393c73133eef9aa73c0f78312e)
Added:
Modified:
libcxx/include/__algorithm/mismatch.h
libcxx/include/__algorithm/simd_utils.h
Removed:
################################################################################
diff --git a/libcxx/include/__algorithm/mismatch.h b/libcxx/include/__algorithm/mismatch.h
index 006ddbd80c306b..cb83347584b1a5 100644
--- a/libcxx/include/__algorithm/mismatch.h
+++ b/libcxx/include/__algorithm/mismatch.h
@@ -79,7 +79,7 @@ __mismatch_vectorized(_Iter __first1, _Iter __last1, _Iter __first2) {
}
for (size_t __i = 0; __i != __unroll_count; ++__i) {
- if (auto __cmp_res = std::__as_mask(__lhs[__i] == __rhs[__i]); !std::__all_of(__cmp_res)) {
+ if (auto __cmp_res = __lhs[__i] == __rhs[__i]; !std::__all_of(__cmp_res)) {
auto __offset = __i * __vec_size + std::__find_first_not_set(__cmp_res);
return {__first1 + __offset, __first2 + __offset};
}
@@ -91,7 +91,7 @@ __mismatch_vectorized(_Iter __first1, _Iter __last1, _Iter __first2) {
// check the remaining 0-3 vectors
while (static_cast<size_t>(__last1 - __first1) >= __vec_size) {
- if (auto __cmp_res = std::__as_mask(std::__load_vector<__vec>(__first1) == std::__load_vector<__vec>(__first2));
+ if (auto __cmp_res = std::__load_vector<__vec>(__first1) == std::__load_vector<__vec>(__first2);
!std::__all_of(__cmp_res)) {
auto __offset = std::__find_first_not_set(__cmp_res);
return {__first1 + __offset, __first2 + __offset};
@@ -108,8 +108,8 @@ __mismatch_vectorized(_Iter __first1, _Iter __last1, _Iter __first2) {
if (static_cast<size_t>(__first1 - __orig_first1) >= __vec_size) {
__first1 = __last1 - __vec_size;
__first2 = __last2 - __vec_size;
- auto __offset = std::__find_first_not_set(
- std::__as_mask(std::__load_vector<__vec>(__first1) == std::__load_vector<__vec>(__first2)));
+ auto __offset =
+ std::__find_first_not_set(std::__load_vector<__vec>(__first1) == std::__load_vector<__vec>(__first2));
return {__first1 + __offset, __first2 + __offset};
} // else loop over the elements individually
}
diff --git a/libcxx/include/__algorithm/simd_utils.h b/libcxx/include/__algorithm/simd_utils.h
index ec9840f60d87c0..56518dafa3193b 100644
--- a/libcxx/include/__algorithm/simd_utils.h
+++ b/libcxx/include/__algorithm/simd_utils.h
@@ -116,65 +116,42 @@ template <class _VecT, class _Iter>
}(make_index_sequence<__simd_vector_size_v<_VecT>>{});
}
-template <size_t _Np>
-[[__nodiscard__]] _LIBCPP_HIDE_FROM_ABI bool __all_of(__simd_vector<bool, _Np> __vec) noexcept {
- return __builtin_reduce_and(__vec);
-}
-
template <class _Tp, size_t _Np>
-[[__nodiscard__]] _LIBCPP_HIDE_FROM_ABI auto __as_mask(__simd_vector<_Tp, _Np> __vec) noexcept {
- static_assert(!is_same<_Tp, bool>::value, "vector type should not be a bool!");
- return __builtin_convertvector(__vec, __simd_vector<bool, _Np>);
-}
-
-// This uses __builtin_convertvector around the __builtin_shufflevector to work around #107981.
-template <size_t _Np>
-[[__nodiscard__]] _LIBCPP_HIDE_FROM_ABI __simd_vector<bool, 8>
-__extend_vector(__simd_vector<bool, _Np> __vec) noexcept {
- using _VecT = __simd_vector<bool, _Np>;
- if constexpr (_Np == 4) {
- return __builtin_convertvector(
- __builtin_shufflevector(__vec, _VecT{}, 0, 1, 2, 3, 4, 5, 6, 7), __simd_vector<bool, 8>);
- } else if constexpr (_Np == 2) {
- return std::__extend_vector(
- __builtin_convertvector(__builtin_shufflevector(__vec, _VecT{}, 0, 1, 2, 3), __simd_vector<bool, 4>));
- } else if constexpr (_Np == 1) {
- return std::__extend_vector(
- __builtin_convertvector(__builtin_shufflevector(__vec, _VecT{}, 0, 1), __simd_vector<bool, 2>));
- } else {
- static_assert(sizeof(_VecT) == 0, "Unexpected vector size");
- }
+[[__nodiscard__]] _LIBCPP_HIDE_FROM_ABI bool __all_of(__simd_vector<_Tp, _Np> __vec) noexcept {
+ return __builtin_reduce_and(__builtin_convertvector(__vec, __simd_vector<bool, _Np>));
}
-template <size_t _Np>
-[[__nodiscard__]] _LIBCPP_HIDE_FROM_ABI auto __to_int_mask(__simd_vector<bool, _Np> __vec) {
- if constexpr (_Np < 8) {
- return std::__bit_cast<uint8_t>(std::__extend_vector(__vec));
- } else if constexpr (_Np == 8) {
- return std::__bit_cast<uint8_t>(__vec);
- } else if constexpr (_Np == 16) {
- return std::__bit_cast<uint16_t>(__vec);
- } else if constexpr (_Np == 32) {
- return std::__bit_cast<uint32_t>(__vec);
- } else if constexpr (_Np == 64) {
- return std::__bit_cast<uint64_t>(__vec);
- } else {
- static_assert(sizeof(__simd_vector<bool, _Np>) == 0, "Unexpected vector size");
- return 0;
- }
-}
+template <class _Tp, size_t _Np>
+[[__nodiscard__]] _LIBCPP_HIDE_FROM_ABI size_t __find_first_set(__simd_vector<_Tp, _Np> __vec) noexcept {
+ using __mask_vec = __simd_vector<bool, _Np>;
-template <size_t _Np>
-[[__nodiscard__]] _LIBCPP_HIDE_FROM_ABI size_t __find_first_set(__simd_vector<bool, _Np> __vec) noexcept {
+ // This has MSan disabled du to https://github.com/llvm/llvm-project/issues/85876
+ auto __impl = [&]<class _MaskT>(_MaskT) _LIBCPP_NO_SANITIZE("memory") noexcept {
# if defined(_LIBCPP_BIG_ENDIAN)
- return std::min<size_t>(_Np, std::__countl_zero(std::__to_int_mask(__vec)));
+ return std::min<size_t>(
+ _Np, std::__countl_zero(__builtin_bit_cast(_MaskT, __builtin_convertvector(__vec, __mask_vec))));
# else
- return std::min<size_t>(_Np, std::__countr_zero(std::__to_int_mask(__vec)));
+ return std::min<size_t>(
+ _Np, std::__countr_zero(__builtin_bit_cast(_MaskT, __builtin_convertvector(__vec, __mask_vec))));
# endif
+ };
+
+ if constexpr (sizeof(__mask_vec) == sizeof(uint8_t)) {
+ return __impl(uint8_t{});
+ } else if constexpr (sizeof(__mask_vec) == sizeof(uint16_t)) {
+ return __impl(uint16_t{});
+ } else if constexpr (sizeof(__mask_vec) == sizeof(uint32_t)) {
+ return __impl(uint32_t{});
+ } else if constexpr (sizeof(__mask_vec) == sizeof(uint64_t)) {
+ return __impl(uint64_t{});
+ } else {
+ static_assert(sizeof(__mask_vec) == 0, "unexpected required size for mask integer type");
+ return 0;
+ }
}
-template <size_t _Np>
-[[__nodiscard__]] _LIBCPP_HIDE_FROM_ABI size_t __find_first_not_set(__simd_vector<bool, _Np> __vec) noexcept {
+template <class _Tp, size_t _Np>
+[[__nodiscard__]] _LIBCPP_HIDE_FROM_ABI size_t __find_first_not_set(__simd_vector<_Tp, _Np> __vec) noexcept {
return std::__find_first_set(~__vec);
}
More information about the libcxx-commits
mailing list