I hope this is the right fix for the issue.
I've regtested for arm-none-eabi on top of r15-11320-g7909a9f85f30b6 and it
applies cleanly on trunk.
The new test is derived from
libstdc++-v3/testsuite/experimental/simd/pr115454_find_last_set.cc.
Ok for trunk, releases/gcc-16 and releases/gcc-15?
--
When testing r15-10798-gae573c9d0e7f1c, I noticed that the
experimental/simd/pr115454_find_last_set.cc was failing due to that the
GNU vector type was used instead of the NEON intrinsic type.
PR libstdc++/122981
libstdc++-v3/ChangeLog:
* include/experimental/bits/simd_neon.h: Convert AArch32 NEON
vpadd_* operands to intrinsic types.
* testsuite/experimental/simd/pr122981_find_last_set_neon.cc: New test.
Signed-off-by: Torbjörn SVENSSON <[email protected]>
---
.../include/experimental/bits/simd_neon.h | 31 +++++++-----
.../simd/pr122981_find_last_set_neon.cc | 50 +++++++++++++++++++
2 files changed, 69 insertions(+), 12 deletions(-)
create mode 100644
libstdc++-v3/testsuite/experimental/simd/pr122981_find_last_set_neon.cc
diff --git a/libstdc++-v3/include/experimental/bits/simd_neon.h
b/libstdc++-v3/include/experimental/bits/simd_neon.h
index d3785db2bfd..23c205f3dc9 100644
--- a/libstdc++-v3/include/experimental/bits/simd_neon.h
+++ b/libstdc++-v3/include/experimental/bits/simd_neon.h
@@ -297,9 +297,10 @@ struct _MaskImplNeonMixin
__zero))[0];
#else
return __vector_bitcast<_UShort>(
- vpadd_s8(vpadd_s8(vpadd_s8(__lo64(__asint), __hi64(__asint)),
- __zero),
- __zero))[0];
+ vpadd_s8(vpadd_s8(vpadd_s8(__to_intrin(__lo64(__asint)),
+ __to_intrin(__hi64(__asint))),
+ __to_intrin(__zero)),
+ __to_intrin(__zero)))[0];
#endif
}
else if constexpr (sizeof(_Tp) == 2)
@@ -313,9 +314,10 @@ struct _MaskImplNeonMixin
#ifdef __aarch64__
return vaddvq_s16(__asint);
#else
- return vpadd_s16(
- vpadd_s16(vpadd_s16(__lo64(__asint), __hi64(__asint)), __zero),
- __zero)[0];
+ return vpadd_s16(vpadd_s16(vpadd_s16(__to_intrin(__lo64(__asint)),
+
__to_intrin(__hi64(__asint))),
+ __to_intrin(__zero)),
+ __to_intrin(__zero))[0];
#endif
}
else if constexpr (sizeof(_Tp) == 4)
@@ -329,8 +331,9 @@ struct _MaskImplNeonMixin
#ifdef __aarch64__
return vaddvq_s32(__asint);
#else
- return vpadd_s32(vpadd_s32(__lo64(__asint), __hi64(__asint)),
- __zero)[0];
+ return vpadd_s32(vpadd_s32(__to_intrin(__lo64(__asint)),
+ __to_intrin(__hi64(__asint))),
+ __to_intrin(__zero))[0];
#endif
}
else if constexpr (sizeof(_Tp) == 8)
@@ -353,8 +356,10 @@ struct _MaskImplNeonMixin
#ifdef __aarch64__
return vaddv_s8(__asint);
#else
- return vpadd_s8(vpadd_s8(vpadd_s8(__asint, __zero), __zero),
- __zero)[0];
+ return vpadd_s8(vpadd_s8(vpadd_s8(__to_intrin(__asint),
+ __to_intrin(__zero)),
+ __to_intrin(__zero)),
+ __to_intrin(__zero))[0];
#endif
}
else if constexpr (sizeof(_Tp) == 2)
@@ -368,7 +373,9 @@ struct _MaskImplNeonMixin
#ifdef __aarch64__
return vaddv_s16(__asint);
#else
- return vpadd_s16(vpadd_s16(__asint, __zero), __zero)[0];
+ return vpadd_s16(vpadd_s16(__to_intrin(__asint),
+ __to_intrin(__zero)),
+ __to_intrin(__zero))[0];
#endif
}
else if constexpr (sizeof(_Tp) == 4)
@@ -377,7 +384,7 @@ struct _MaskImplNeonMixin
#ifdef __aarch64__
return vaddv_s32(__asint);
#else
- return vpadd_s32(__asint, __zero)[0];
+ return vpadd_s32(__to_intrin(__asint), __to_intrin(__zero))[0];
#endif
}
else
diff --git
a/libstdc++-v3/testsuite/experimental/simd/pr122981_find_last_set_neon.cc
b/libstdc++-v3/testsuite/experimental/simd/pr122981_find_last_set_neon.cc
new file mode 100644
index 00000000000..f7f79166ae6
--- /dev/null
+++ b/libstdc++-v3/testsuite/experimental/simd/pr122981_find_last_set_neon.cc
@@ -0,0 +1,50 @@
+// { dg-options "-std=gnu++17" }
+// { dg-do compile { target c++17 } }
+// { dg-require-effective-target arm_neon_ok }
+// { dg-add-options arm_neon }
+// { dg-require-cmath "" }
+
+#include <experimental/simd>
+
+namespace stdx = std::experimental;
+
+template <typename U, int N>
+using V = stdx::simd<U, stdx::simd_abi::deduce_t<U, N>>;
+
+template <typename U, int N>
+[[gnu::noinline, gnu::noipa]]
+int reduce(typename V<U, N>::mask_type x)
+{
+ using M = typename V<U, N>::mask_type;
+ static_assert(stdx::find_last_set(M(true)) == N - 1);
+ return stdx::find_last_set(x);
+}
+
+int main()
+{
+ // 16 bytes: 16 int8_t lanes, 8 int16_t lanes, 4 int32_t lanes.
+ const int r_int8_16 = reduce<std::int8_t, 16>(typename V<std::int8_t,
16>::mask_type(true));
+ if (r_int8_16 != 15)
+ __builtin_abort();
+
+ const int r_int16_8 = reduce<std::int16_t, 8>(typename V<std::int16_t,
8>::mask_type(true));
+ if (r_int16_8 != 7)
+ __builtin_abort();
+
+ const int r_int32_4 = reduce<std::int32_t, 4>(typename V<std::int32_t,
4>::mask_type(true));
+ if (r_int32_4 != 3)
+ __builtin_abort();
+
+ // 8 bytes: 8 int8_t lanes, 4 int16_t lanes, 2 int32_t lanes.
+ const int r_int8_8 = reduce<std::int8_t, 8>(typename V<std::int8_t,
8>::mask_type(true));
+ if (r_int8_8 != 7)
+ __builtin_abort();
+
+ const int r_int16_4 = reduce<std::int16_t, 4>(typename V<std::int16_t,
4>::mask_type(true));
+ if (r_int16_4 != 3)
+ __builtin_abort();
+
+ const int r_int32_2 = reduce<std::int32_t, 2>(typename V<std::int32_t,
2>::mask_type(true));
+ if (r_int32_2 != 1)
+ __builtin_abort();
+}
--
2.43.0
