https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109069
--- Comment #5 from John Platts <john_platts at hotmail dot com> --- Here is another test program that shows the same code generation bug when a splat followed by a vec_sld is incorrectly optimized by gcc 12.2.0 on powerpc64-linux-gnu and powerpc64le-linux-gnu with the -mcpu=power10 -O2 options: #pragma push_macro("vector") #pragma push_macro("pixel") #pragma push_macro("bool") #undef vector #undef pixel #undef bool #include <altivec.h> #pragma pop_macro("vector") #pragma pop_macro("pixel") #pragma pop_macro("bool") #include <stdint.h> #include <type_traits> template<class T> struct MakeSimdVectorType { typedef T type __attribute__((__vector_size__(16))); }; template<class T> using SimdVectorType = typename MakeSimdVectorType<T>::type; template<class T, std::enable_if_t<(sizeof(T) == 1 && std::is_integral_v<std::decay_t<T>>)>* = nullptr> static inline SimdVectorType<T> Splat(T val) { return reinterpret_cast<SimdVectorType<T>>( vec_splats(static_cast<unsigned char>(val))); } template<class T, std::enable_if_t<(sizeof(T) == 2 && std::is_integral_v<std::decay_t<T>>)>* = nullptr> static inline SimdVectorType<T> Splat(T val) { return reinterpret_cast<SimdVectorType<T>>( vec_splats(static_cast<unsigned short>(val))); } template<class T, std::enable_if_t<(sizeof(T) == 4 && std::is_integral_v<std::decay_t<T>>)>* = nullptr> static inline SimdVectorType<T> Splat(T val) { return reinterpret_cast<SimdVectorType<T>>( vec_splats(static_cast<unsigned int>(val))); } template<class T, std::enable_if_t<(sizeof(T) == 8 && std::is_integral_v<std::decay_t<T>>)>* = nullptr> static inline SimdVectorType<T> Splat(T val) { return reinterpret_cast<SimdVectorType<T>>( vec_splats(static_cast<unsigned long long>(val))); } static inline __vector float Splat(float val) { return vec_splats(val); } static inline __vector double Splat(double val) { return vec_splats(val); } using AltivecUCharVectType = __vector unsigned char; template<int kShiftAmount, class T> AltivecUCharVectType SplatAndShift(T val) { const auto splatResult = Splat(val); return vec_sld(reinterpret_cast<AltivecUCharVectType>(splatResult), reinterpret_cast<AltivecUCharVectType>(splatResult), kShiftAmount); } template<int kShiftAmount, class T> AltivecUCharVectType SplatAndShift_2(T val) { auto splatResult = Splat(val); __asm__("" : "+wa" (splatResult)); return vec_sld(reinterpret_cast<AltivecUCharVectType>(splatResult), reinterpret_cast<AltivecUCharVectType>(splatResult), kShiftAmount); } auto SplatAndShift_I16_1() { return SplatAndShift<5>(int16_t{-32346}); } auto SplatAndShift_I16_2() { return SplatAndShift_2<5>(int16_t{-32346}); } auto SplatAndShift_I32_1() { return SplatAndShift<3>(int32_t{-1394373889}); } auto SplatAndShift_I32_2() { return SplatAndShift_2<3>(int32_t{-1394373889}); } Here is the assembly code that is generated for the above code on powerpc64le-linux-gnu with the -O2 -mcpu=power10 options: _Z19SplatAndShift_I16_1v: xxspltiw 34,2175173030 blr _Z19SplatAndShift_I16_2v: xxspltiw 34,2175173030 vsldoi 2,2,2,5 blr _Z19SplatAndShift_I32_1v: xxspltiw 34,2900593407 blr _Z19SplatAndShift_I32_2v: xxspltiw 34,2900593407 vsldoi 2,2,2,3 blr Here is the assembly code that is generated for the above code on powerpc64-linux-gnu with the -O2 -mcpu=power10 options: _Z19SplatAndShift_I16_1v: .quad .L._Z19SplatAndShift_I16_1v,.TOC.@tocbase,0 .L._Z19SplatAndShift_I16_1v: xxspltiw 34,2175173030 blr _Z19SplatAndShift_I16_2v: .quad .L._Z19SplatAndShift_I16_2v,.TOC.@tocbase,0 .L._Z19SplatAndShift_I16_2v: xxspltiw 34,2175173030 vsldoi 2,2,2,5 blr _Z19SplatAndShift_I32_1v: .quad .L._Z19SplatAndShift_I32_1v,.TOC.@tocbase,0 .L._Z19SplatAndShift_I32_1v: xxspltiw 34,2900593407 blr _Z19SplatAndShift_I32_2v: .quad .L._Z19SplatAndShift_I32_2v,.TOC.@tocbase,0 .L._Z19SplatAndShift_I32_2v: xxspltiw 34,2900593407 vsldoi 2,2,2,3 blr