Re: [PATCH][auto-inc-dec.c] Account for cost of move operation in FORM_PRE_ADD and FORM_POST_ADD cases

[Kyrill Tkachov]

On 28/10/15 17:23, Christophe Lyon wrote:

Hi Kyrill,

Hi Christophe,

On 26 October 2015 at 12:52, Kyrill Tkachov <kyrylo.tkac...@arm.com> wrote:

On 26/10/15 11:28, Bernd Schmidt wrote:

On 10/26/2015 12:12 PM, Bernd Schmidt wrote:

But isn't that balanced by the fact that it doesn't seem to take into
account the gain of removing the inc_insn either? So I think this can't
be right.

Argh, misread the code. The patch is OK with the change I suggested.

Thanks!
Here's what I committed with r229344.

Since this commit, I've noticed:

FAIL: gcc.target/arm/lp1243022.c scan-rtl-dump subreg2 "REG_INC"
with --target arm-none-eabi --with-thumb --with-cpu=cortex-a9

I think this is a testcase issue, I'll look into updating it separately.

as well as ICEs:
   gcc.target/aarch64/advsimd-intrinsics/vldX.c   -O2  (internal compiler error)
   gcc.target/aarch64/advsimd-intrinsics/vldX.c   -O2 -flto
-fno-use-linker-plugin -flto-partition=none  (internal compiler error)

with --target arm-none-linux-gnueabihf --with-thumb
--with-cpu=cortex-a15 --with-fpu=neon-vfpv4
and --target arm-none-linux-gnueabihf --with-thumb
--with-cpu=cortex-a57 --with-fpu=crypto-neon-fp-armv8

See for a more synthetic view:
http://people.linaro.org/~christophe.lyon/cross-validation/gcc/trunk/229344/report-build-info.html

Can you have a look?

The ICE backtrace is:
0x7a4494 change_address_1
        $SRC/gcc/emit-rtl.c:2132
0x7a7453 adjust_address_1(rtx_def*, machine_mode, long, int, int, int, long)
        $SRC/gcc/emit-rtl.c:2270
0xab792d gen_lowpart_general(machine_mode, rtx_def*)
        $SRC/gcc/rtlhooks.c:90
0xfd721d gen_split_47(rtx_insn*, rtx_def**)
        $SRC/gcc/config/arm/arm.md:4336
0x12473f2 split_11
        $SRC/gcc/config/arm/arm.md:4331
0x12473f2 split_insns(rtx_def*, rtx_insn*)
        $SRC/gcc/config/arm/sync.md:361
0x7af30e try_split(rtx_def*, rtx_insn*, int)
        $SRC/gcc/emit-rtl.c:3664
0xa53dc5 split_insn
        $SRC/gcc/recog.c:2874
0xa5ccf5 split_all_insns()
        $SRC/gcc/recog.c:2964
0xa5cde3 rest_of_handle_split_after_reload
        $SRC/gcc/recog.c:3900
0xa5cde3 execute
        $SRC/gcc/recog.c:3929


Looks like a latent issue exposed by my patch.
Could you please file a BZ entry if get the chance?

Thanks,
Kyrill

Thanks,

Christophe.

Kyrill

2015-10-26  Kyrylo Tkachov  <kyrylo.tkac...@arm.com>

     * auto-inc-dec.c (insert_move_insn_before): Delete.
     (attempt_change): Remember to cost the simple move in the
     FORM_PRE_ADD and FORM_POST_ADD cases.

Bernd