Oleg Endo <oleg.e...@t-online.de> wrote:
> This adds the two common built-in functions __builtin_thread_pointer and
> __builtin_set_thread_pointer to the SH port.
> I've done it in a way so that hopefully it can be transitioned to target
> independent thread pointer built-ins easily, as suggested by Richard a
> while ago:
> http://gcc.gnu.org/ml/gcc-patches/2012-07/msg00946.html
>
> Originally I wanted to wait until the target independent bits are in,
> but somehow the thread mentioned above died and I got impatient.
>
> I've also added support for SH's GBR based displacement addressing
> modes. They are not used for general purpose mem loads/stores by the
> compiler, but rather when code accesses data behind the thread pointer
> (thread control block or something).
> The way GBR displacement address opportunities are discovered might not
> be the best way of doing this sort of thing, but it works. The insn
> walking could potentially slow down compile times, but it is only
> enabled for functions where the GBR is referenced, so it shouldn't be so
> bad. Alternatives and suggestions are highly appreciated. :)
>
> Tested on rev 191894 with 'make all' (c,c++) and
> 'make -k check-gcc RUNTESTFLAGS="sh.exp --target_board=sh-sim
> \{-m2/-ml,-m2/-mb,-m2a/-mb,-m4/-ml,-m4/-mb,-m4a/-ml,-m4a/-mb}"'
>
> For code that doesn't reference GBR there are no functional changes.
> TLS code that references GBR might trigger the 'sh_find_equiv_gbr_addr'
> function. Unfortunately TLS tests don't seem to work on sh-sim, so I
> could not test this part.
sh4-unknown-linux-gnu build failed during compiling libmudflap:
/exp/ldroot/dodes/xsh-gcc/./gcc/xgcc -B/exp/ldroot/dodes/xsh-gcc/./gcc/
-B/usr/local/sh4-unknown-linux-gnu/bin/ -B/usr/local/sh4-unknown-linux-gnu/lib/
-isystem /usr/local/sh4-unknown-linux-gnu/include -isystem
/usr/local/sh4-unknown-linux-gnu/sys-include -DHAVE_CONFIG_H -I.
-I../../../LOCAL/trunk/libmudflap -DLIBMUDFLAPTH -g -O2 -MT
libmudflapth_la-mf-runtime.lo -MD -MP -MF .deps/libmudflapth_la-mf-runtime.Tpo
-c ../../../LOCAL/trunk/libmudflap/mf-runtime.c -o libmudflapth_la-mf-runtime.o
../../../LOCAL/trunk/libmudflap/mf-runtime.c: In function
'begin_recursion_protect1':
../../../LOCAL/trunk/libmudflap/mf-runtime.c:152:1: internal compiler error:
Segmentation fault
}
^
0x8529c60 crash_signal
../../LOCAL/trunk/gcc/toplev.c:335
0x8771a87 sh_find_base_reg_disp
../../LOCAL/trunk/gcc/config/sh/sh.c:13344
0x8791554 sh_find_equiv_gbr_addr(rtx_def*, rtx_def*)
../../LOCAL/trunk/gcc/config/sh/sh.c:13395
0x87ce6cf gen_split_1029(rtx_def*, rtx_def**)
../../LOCAL/trunk/gcc/config/sh/sh.md:10184
0x87ea6e0 split_1
../../LOCAL/trunk/gcc/config/sh/sh.md:10183
0x87ea6e0 split_3
../../LOCAL/trunk/gcc/config/sh/sh.md:7082
0x82bf8a1 try_split(rtx_def*, rtx_def*, int)
../../LOCAL/trunk/gcc/emit-rtl.c:3503
0x849c642 split_insn
../../LOCAL/trunk/gcc/recog.c:2809
0x84a08b5 split_all_insns()
../../LOCAL/trunk/gcc/recog.c:2899
0x84a09a7 rest_of_handle_split_all_insns
../../LOCAL/trunk/gcc/recog.c:3751
Looks prev_nonnote_insn returns a barrier there:
(gdb) fr 0
#0 0x08771a87 in sh_find_base_reg_disp (insn=<optimized out>, x=0xb79ddd40,
base_reg=0x0, disp=0) at ../../LOCAL/trunk/gcc/config/sh/sh.c:13344
13344 if (p != NULL && GET_CODE (p) == SET && REG_P (XEXP (p, 0))
(gdb) p p
$1 = (rtx) 0xafafafaf
(gdb) p i
$2 = (rtx_def *) 0xb79ecc3c
(gdb) call debug_rtx(i)
(barrier 216 215 217)
> (builtin_description): Add is_enabled member.
> (shmedia_builtin, sh1_builtin): New functions.
> (signature_args): Add SH_BLTIN_VP.
> (bdesc): Use shmedia_builtin for existing built-ins. Add
> __builtin_thread_pointer and __builtin_set_thread_pointer as
> sh1_builtin.
> (sh_media_init_builtins, sh_init_builtins): Merge into single
> function sh_init_builtins. Add is_enabled checking.
> (sh_media_builtin_decl, sh_builtin_decl): Merge into single
> function sh_builtin_decl. Add is_enabled checking.
It would be better to separate this part except new thread pointer
builtins into an independent patch which should be tested also with
sh64-elf build, though now unified sh64-elf build is failing.
I'd like to commit a quick fix for sh64-elf build failure.
Regards,
kaz
