This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.
Index Nav: | [Date Index] [Subject Index] [Author Index] [Thread Index] | |
---|---|---|
Message Nav: | [Date Prev] [Date Next] | [Thread Prev] [Thread Next] |
Other format: | [Raw text] |
On Fri, Dec 2, 2016 at 5:11 PM, Uros Bizjak <ubizjak@gmail.com> wrote: > On Fri, Dec 2, 2016 at 3:21 PM, Jakub Jelinek <jakub@redhat.com> wrote: >> Hi! >> >> While the https://gcc.gnu.org/ml/gcc-patches/2016-01/msg01087.html >> added *andndi3_doubleword, I don't understand how it can actually work. >> The problem is that this pattern is for combine, and needs combining >> of DImode NOT setter with DImode AND user. But there is no DImode >> pattern for one_cmpldi2, so it is lowered early into two one_cmplsi2 >> patterns and I can't see how combine would be able to figure stuff out from >> that. >> >> This patch: >> 1) adds one_cmpldi2 pattern for stv purposes (which splits into two >> one_cmplsi2 after reload) >> 2) teaches the 32-bit stv pass to handle NOT (as xor all-ones) >> 3) renames the old *andndi3_doubleword to *andndi3_doubleword_bmi, as it >> is for -mbmi only, and adds another *andndi3_doubleword pattern that is >> meant to live just from combine till the stv pass, or worse case till >> following split1 pass when it is split back into not followed by and; >> this change makes it possible to use pandn in stv pass, even without >> -mbmi > > Please use attached (lightly tested) patch to implement point 3) > above. The patch splits insn after reload, as is the case with all STV > patterns. Now attached for real. Uros.
Attachment:
p.diff.txt
Description: Text document
Index Nav: | [Date Index] [Subject Index] [Author Index] [Thread Index] | |
---|---|---|
Message Nav: | [Date Prev] [Date Next] | [Thread Prev] [Thread Next] |