http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56965
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Attachment #32105|0 |1
is obsolete| |
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 32106
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=32106&action=edit
/suse/rguenther/p
Somewhat micro-optimized patch as otherwise it slows down the common cases too
much it seems (testing on insn-recog). It's still slower than the non-sorting
version (but if we make it stronger by collecting more relevant FIELD_DECLs
we'll hit the trade-off limit faster ...).
Some statistics on insn-recog, fieldsx vs. fieldsy length:
184 rtl pre "ncr-1-1 == 1" 6192389
185 hoist "ncr-1-1 == 1" 252507
201 dse1 "ncr-1-1 == 1" 6743
227 dse2 "ncr-1-1 == 1" 6743
239 sched2 "ncr-1-1 == 1" 6743
which is interesting ;) The most common case is just one. Very easy to
very micro-optimize but still shows non-noise regression of about 1%
for the pathological case above.
We can also optimize the length 1 vs. arbitrary length case and avoid
sorting the 2nd vector (not done yet).
