On 13-08-30 9:09 AM, Yvan Roux wrote:
Hi,
here is a request for comments on the 2 attached patches which enable
the build of GCC on ARM with LRA. The patches introduce a new
undocumented option -mlra to use LRA instead of reload, as what was
done on previous LRA support, which is here to ease the test and
comparison with reload and not supposed to be kept in 4.9.
On AArch32:
* The patch includes the previous changes of this thread, add the
handling of ASHIFTRT, LSHIFTRT, ROTATE and ROTATERT in get_index_scale
and let LRA handle the constraint in Thumb.
* The status is that the build complete in ARM mode with a couple of
regressions in the testsuite, that I'll investigate:
- gcc.c-torture/execute/20060102-1.c
- gcc.c-torture/execute/961026-1.c
- gcc.target/arm/atomic-comp-swap-release-acquire.c
and some build failures in libstdc++ at the pass manager level (there
is an invalid read in gt_ggc_mx)
* The build of libraries in Thumb mode still doesn't complete, as
Vladimir said the secondary_reload modification solves LRA cycling
but we still have some issues.
On AArch64 :
* I modified must_be_index_p to avoid the call of set_address_base
with patterns which contains a MULT.
* The build complete, but there is a couple of regression in the
testsuite I'm looking at on
- gcc.c-torture/execute/ieee/fp-cmp-4l.c
- c-c++-common/torture/complex-sign-mul-minus-one.c
for instance.
Any comments ?
Yvan, sorry for the delay with the response (I was on a long vacation).
It looks ok to me. But I have no rights to approve rtlanal.c changes.
So I am CCing this email to Richard Henderson.
Richard, sorry for bothering you, could you give an approval to the
rtanal.c changes. Neither the author of the original code (Richard
Sandiford), nor me has a right to approve rtlanal.c changes. These
changes actually were around and discussed for last 2-3 months.
Thanks.
