On Tue, Apr 12, 2016 at 8:10 AM, Jakub Jelinek <ja...@redhat.com> wrote: > Hi! > > Most of simplify-rtx.c attempts hard not to generate any extra instructions, > just simplify the RTL to something equivalent and let the user try to > recognize it, or whatever else it wishes. > This is especially important post-reload, when we can't create new pseudos, > or e.g. when processing (simplifying) DEBUG_INSNs, where we certainly don't > want any code around it (not to mention where exactly would it be emitted > anyway?). E.g. it uses the gen_lowpart_no_emit rtl hook instead of > gen_lowpart, etc. > As the following testcase shows, there are two exceptions, for > POINTER_EXTEND* targets under some very limited conditions it actually > calls convert_memory_address, which sometimes does what simplify-rtx.c > normally does, just return equivalent RTL expression, but in another case > can create a new pseudo (no no after reload, or for debug insns any time) > and emit some move insns. > > This patch arranges for a new argument to > convert_memory_address_addr_space_1 and calls it with that new argument set > to true, to make sure it never emits instructions or creates pseudos. > > Compared to the previous version of the patch, I've added ATTRIBUTE_UNUSED > to the new argument, so that bootstrap doesn't fail on targets that don't > define POINTERS_EXTEND*. > > Bootstrapped/regtested on x86_64-linux and i686-linux, bootstrapped on > aarch64-linux (regtest pending). > I don't have access to aarch64 ilp32 setup though, Andrew or anybody else, > could you please test it there? Also CCing H.J. just in case it affects > x86 -mx32. >
I bootstrapped and tested it on x32. There is no regression. -- H.J.
