On Thu, Oct 12, 2017 at 8:39 AM, Uros Bizjak <ubiz...@gmail.com> wrote: > On Thu, Oct 12, 2017 at 8:32 AM, Uros Bizjak <ubiz...@gmail.com> wrote: >> On Wed, Oct 11, 2017 at 10:59 PM, Jakub Jelinek <ja...@redhat.com> wrote: >>> Hi! >>> >>> As can be seen on the testcase below, the *<rotate_insn><mode>3_mask >>> insn/splitter is able to optimize only the case when the and is >>> performed in SImode and then the result subreged into QImode, >>> while if the computation is already in QImode, we don't handle it. >>> >>> Fixed by adding another pattern, bootstrapped/regtested on x86_64-linux and >>> i686-linux, ok for trunk? >> >> We probably want to add this variant to *all* *_mask splitters (there >> are a few of them in i386.md, please grep for "Avoid useless >> masking"). Which finally begs a question - should we implement this >> simplification in a generic, target-independent way? OTOH, we already >> have SHIFT_COUNT_TRUNCATED and shift_truncation_mask hooks, but last >> time I try the former, there were some problems in the testsuite on >> x86. I guess there are several targets that would benefit from >> removing useless masking of count operands. > > Oh, and there is a strange x86 exception in the comment for > SHIFT_COUNT_TRUNCATED. I'm not sure what "(real or pretended) > bit-field operation" means, but variable-count BT instruction with > non-memory operand (we never generate variable-count BTx with memory > operand) masks its count operand as well.
I forgot that SSE shifts don't truncate their count operand. This is the reason that the removal of mask is implemented in the *.md file, but it would really be nice if the same functionality can be achieved in a more generic way, without pattern explosion. Uros.
