Reviewers: davidxl, Message: I have uploaded a new patch. I did not split the patch into two because I felt the ifunc dispatch patch does nothing as a standalone. The framework now will do aggressive cloning and it will mark the clones specialized for core2 with "-mtune=core2". I have already submitted another patch that will optimize vectorization for core2.
I also tested Teresa's RAT stalls avoidance in my MV framework and it works great. I removed that from the current patch as I thought that should go in as a separate patch. Pending: Test case. On 2011/12/09 07:49:57, davidxl wrote:
It might be better to split this into two patches:
1) ifunc dispatch patch 2) auto cloning patch
The auto cloning framework is good in general, the main problem
potential
disconnection of early cloning decisions and later transformations.
the logic of deciding if loops are vectorizable is not the same as
that used in
the actual vectorization pass which can lead to either false positives (unnecessary clones) and false negatives (missing clones of
vectorizable loops).
The false negatives are entirely possible because later
transformations such as
if-cvt can expose more opportunities. To alleviate the problem, it
would be
more powerful if the framework allows more aggressive (speculative)
cloning, and
undo some of the cloning later when they turn out to be not needed.
This is
possible with some book keeping.
David
http://codereview.appspot.com/5477048/diff/1/config/i386/i386.c File config/i386/i386.c (right):
http://codereview.appspot.com/5477048/diff/1/config/i386/i386.c#newcode26331
config/i386/i386.c:26331: + THe --> The.
http://codereview.appspot.com/5477048/diff/1/config/i386/i386.c#newcode26365
config/i386/i386.c:26365: + tree-vect-stmts.cc. The loop is not
vectorizable
even if a single stmt is not even if?
http://codereview.appspot.com/5477048/diff/1/config/i386/i386.c#newcode26367
config/i386/i386.c:26367: + You need a good explanation here why a common utility function can not
be used,
but 'duplicate' the check here.
http://codereview.appspot.com/5477048/diff/1/config/i386/i386.c#newcode26524
config/i386/i386.c:26524: + if (ix86_mv_arch_string Using hardcoded target name here is not ideal -- better encode this
attribute in
the target/processor model array. For now better add a comment here.
http://codereview.appspot.com/5477048/diff/1/config/i386/i386.c#newcode26544
config/i386/i386.c:26544: + if (*cond_func_decl == NULL) Add a small example in the comment here. Also better use a helper
function.
http://codereview.appspot.com/5477048/diff/1/mversn-dispatch.c File mversn-dispatch.c (right):
http://codereview.appspot.com/5477048/diff/1/mversn-dispatch.c#newcode2330
mversn-dispatch.c:2330: static unsigned int do_auto_clone (void) Start a new line
http://codereview.appspot.com/5477048/diff/1/mversn-dispatch.c#newcode2371
mversn-dispatch.c:2371: return 0; Fix indentation
http://codereview.appspot.com/5477048/diff/1/mversn-dispatch.c#newcode2404
mversn-dispatch.c:2404: NULL, fn_ver_addr_chain); indentation
http://codereview.appspot.com/5477048/diff/1/mversn-dispatch.c#newcode2412
mversn-dispatch.c:2412: /* The current function is replaced by a ifunc
call to
the right version. s/a/an/
http://codereview.appspot.com/5477048/diff/1/mversn-dispatch.c#newcode2424
mversn-dispatch.c:2424: cond_func_addr, fn_ver_addr_chain, Indentation
http://codereview.appspot.com/5477048/diff/1/mversn-dispatch.c#newcode2431
mversn-dispatch.c:2431: gimple_set_bb (gsi_stmt (gsi), empty_bb); Fix indentation.
Please review this at http://codereview.appspot.com/5477048/ Affected files: M config/i386/i386.c M config/i386/i386.opt M doc/tm.texi M doc/tm.texi.in M mversn-dispatch.c M params.def M passes.c M target.def M tree-pass.h
