This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [Committed] S/390: Avoid true-complementary bypass problem for reg-reg compares
Quoting Joern Rennecke <amylaar@spamcop.net>:
ar %r2,%r3
cr %r2,%r3
Sorry, this got sent prematurely due to an accident with one of my keyboards
while I was checking up on something. Although this sequence has an
add with a second operand which is identical to the complemented operand of
a following compare, they obviously can't be in the same group.
To get a really interesting example, a two-input instruction should
be scheduled immediately after a compare, or after a compare-and-branch.
I'm not sure what it would take to get the former, the scheduler seems
rather stubborn not doing this for simple testcases.
The latter would seem easy to trigger, except I can't seem to get the
cmp_and_br_signed_<mode> pattern generated. It is not generated by
the cmp / b<code> combinations, and there is no cbranch<mode> pattern.
Combine refuses to generate this pattern because it strips away the
hard register clobber of CC_REGNUM.