On Dec 30, 2011, at 3:20 PM, Richard Guenther wrote: > On Sat, Dec 24, 2011 at 8:00 PM, Steven Bosscher <stevenb....@gmail.com> > wrote: >> On Tue, Dec 20, 2011 at 9:46 AM, Tristan Gingold <ging...@adacore.com> wrote: >>> Hi, >>> >>> currently alpha/vms backend emits a trampoline entry point for all nested >>> functions. This is a waste of code space, as although nested functions are >>> very common in Ada, address of nested functions are only seldom taken. >>> >>> The fact that the address of a function is taken seems only be available in >>> cgraph. Is it OK to use cgraph in alpha.c ? >> >> Since no-one has answered yet, I'll just toss in my $0.02. >> (Hold on to them a bit, they may be worth a million Euro soon :-) >> >> I think that in general you cannot rely on cgraph in the backends, >> this has to be analyzed case-by-case. In your case I'm not sure, but I >> think it should be OK. >> >> Your patch uses cgraph in alpha_start_function, which is apparently >> only used for the target hook ASM_DECLARE_FUNCTION_NAME. This hook is >> called from varasm.c:assemble_start_function(), and this is in turn >> only called from final.c:rest_of_handle_final() to generate assembly >> from RTL, and from cgraphunit.c:assemble_thunk() to output assembly >> for MI thunks. AFAICT cgraph should be correct and complete at the >> stage when those two functions are called. Therefore your patch should >> be OK. >> >> Perhaps Honza can throw in his 0.02h? > > Err - are not _all_ backends using trampolines to represent address-taken > nested functions?
No. > At least I remeber to see them for x86 and plain C > nested functions as well. So - is this really a target issue? Yes. > At the time > the address of a nested fn is taken the C frontend arranges to (dynamically?) > create the trampoline via some built-in. Some backends (AIX, ia64, powerpc64-elf, VMS) don't need trampolines at all because they use function descriptors. Tristan.
