fergalmoran/ladybird
1
0
Fork 0
mirror of https://github.com/fergalmoran/ladybird.git synced 2025-12-22 09:19:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

AK/SIMDExtras: Fix masking logic in shuffle_or_0

Browse Source
This commit is contained in:
Diego Frias
2024-07-26 15:19:36 -07:00
committed by Ali Mohammad Pur
parent 48f1861ce9
commit a168bec7ef
1 changed files with 4 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
AK/SIMDExtras.h
View File

@@ -218,11 +218,10 @@ ALWAYS_INLINE static T shuffle_or_0_impl(T a, Control control, IndexSequence<Idx
using E = ElementOf<T>; using E = ElementOf<T>;
if constexpr (__has_builtin(__builtin_shuffle)) { if constexpr (__has_builtin(__builtin_shuffle)) {
// GCC does a very bad job at optimizing the masking, while not recognizing the shuffle idiom auto vector = __builtin_shuffle(a, control);
// So we jinx its __builtin_shuffle to work with out of bounds indices for (size_t i = 0; i < N; ++i)
// TODO: verify that this masking logic is correct (for machines with __builtin_shuffle) vector[i] = control[i] < 0 || control[i] >= N ? 0 : vector[i];
auto mask = (control >= 0) | (control < N); return vector;
return __builtin_shuffle(a, control & mask) & ~mask;
} }
// 1. Set all out of bounds values to ~0 // 1. Set all out of bounds values to ~0
// Note: This is done so that the optimization mentioned down below works // Note: This is done so that the optimization mentioned down below works