[PATCH][TUPLES] Tuplification of tree-ssa-phiopt.c
Doug Kwan (關振德)
dougkwan@google.com
Fri Apr 18 02:18:00 GMT 2008
The boolean tree has at most two operators, one is from the original
cond_stmt, and one is a TRUTH_NOT_EXPR and all operands are already
valid gimple operands. I think it is always gimplifiable. So far it
passes the testsuite and bootstrap without any problems. The
conversion from boolean to phi result type is done sepearately. I
found that folding conversion with everything sometimes produces
un-gimplifiable expression tree and that causes
force_gimple_operand_gsi to assert.
So there is currently no check in conditional_replacement for valid gimple.
-Doug
2008/4/17 Andrew Pinski <pinskia@gmail.com>:
> On Thu, Apr 17, 2008 at 3:55 PM, Doug Kwan (關振德) <dougkwan@google.com> wrote:
>
> > Construct a boolean tree and let fold and force_gimple_operand_gsi
> > to do optimization and code genaration.
>
> I always wanted to fix up PHI-opt this way but never got around to it.
> This is a good improvement and in some cases, you should be able to
> remove some of the checks where we would have produced invalid gimple.
>
> Thanks,
> Andrew Pinski
>
More information about the Gcc-patches
mailing list