On 8/8/25 21:30, Richard Sandiford wrote:
Richard Henderson <richard.hender...@linaro.org> writes:
Version 1 regressed the expansion of atomics, which means the addition
of CC clobber to all conditional branches is flawed. Version 2 goes
the other way: remove CC clobber from all conditional branches.
This requires the out-of-range TBZ->TST+B.cond expansion be removed,
falling back to the far_branch expansion TBNZ+B.
Thanks for the update. I suppose this change (patch 9) is going to be
the most controversial part. But I agree it's a necessary consequence
of the approach that the series is taking.
I have fixed target/121388, the condition inversion problem, by
requiring the immediate ranges of CB be symmetric. This omits 64
as a valid input to GE and -1 as a valid input to LE.
I have dropped the removal of cbranch<SHORT>4 expanding to CBB/CBH.
I still think that's the correct way to go, but given the change to
mov<ALLI>cc, it is no longer required to fix an ICE.
...
I've replied to some of the individual patches, but the series is ok
from my POV with those changes. Please wait till Monday to see if
there are any other comments though.
Thanks. Committed.
r~
