> - int32_t positions[num_keys]; > + int32_t *positions = alloca(sizeof(int32_t) * num_keys);
A general comment, using this patch as an example... I wonder if adding const has any effect on the compilers' optimizers: int32_t * const positions = alloca(sizeof(int32_t) * num_keys); Perhaps not in this specific case, but maybe elsewhere. Especially if passing the array by reference. We're not good at adding const and restrict to function parameters, so declaring the alloca()'d arrays const might compensate for that.
