On Tue, Dec 20, 2016 at 10:29 PM, Bernd Schmidt <bschm...@redhat.com> wrote: > The problem here is that we don't have complete coverage of lea patterns for > HImode/QImode: the combiner can't recognize a (plus (ashift reg 2) reg) > pattern it builds. > > My first idea was to canonicalize ASHIFT by constant inside PLUS to MULT. > The docs say that this is done only inside a MEM context, but that seems > misguided considering the existence of lea patterns. Maybe that's something > to revisit for gcc-8. In the meantime, the patch below is more like a > minimal fix, and it seems to produce better results at the moment anyway.
Actually, as mentioned in ix86_decompose_address, we have to handle ASHIFT anyway, since combine is quite creative when creating various combinations of PLUS, MULT and ASHIFT in the LEA context. But since ASHIFT handling is relatively minor functional addition, we just handle everything in the decomposition, with the side effect that combine sometimes doesn't canonicalize the combinations involving ASHIFT. > Bootstrapped and tested on x86_64-linux. Ok? OK. Thanks, Uros.
