Jakub Jelinek <ja...@redhat.com> writes: > On Fri, Nov 08, 2024 at 05:44:48PM +0000, Richard Sandiford wrote: >> It's for https://gcc.gnu.org/pipermail/gcc-patches/2024-November/667499.html >> , >> which needs to switch to the simd clone's chosen target (SVE) in order >> to construct the correct types. Currently the patch uses: >> >> + cl_target_option_save (&cur_target, &global_options, >> &global_options_set); >> + tree new_target = DECL_FUNCTION_SPECIFIC_TARGET (node->decl); >> + cl_target_option_restore (&global_options, &global_options_set, >> + TREE_TARGET_OPTION (new_target)); >> + aarch64_override_options_internal (&global_options); >> + memcpy (m_old_have_regs_of_mode, have_regs_of_mode, >> + sizeof (have_regs_of_mode)); >> + for (int i = 0; i < NUM_MACHINE_MODES; ++i) >> + if (aarch64_sve_mode_p ((machine_mode) i)) >> + have_regs_of_mode[i] = true; >> >> to switch in and: >> >> + /* Restore current options. */ >> + cl_target_option_restore (&global_options, &global_options_set, >> &cur_target); >> + aarch64_override_options_internal (&global_options); >> + memcpy (have_regs_of_mode, m_old_have_regs_of_mode, >> + sizeof (have_regs_of_mode)); >> >> to switch back, but the idea is to replace that with: >> >> push_function_decl (node->decl); >> >> ... >> >> pop_function_decl (); > > Why do you need that? Can't you just use opt_for_fn?
The function doesn't need to query the optimisations or target directly, since it directly controls the things that it cares about. The problem instead is that we need the SVE vector modes to be enabled when doing the call to simd_clone_adjust_sve_vector_type. (There might be other reasons as well. That's just the one I'm aware of.) Richard
