On Mon, Nov 21, 2016 at 12:49 AM, Uros Bizjak <ubiz...@gmail.com> wrote: > On Sun, Nov 20, 2016 at 10:00 PM, Uros Bizjak <ubiz...@gmail.com> wrote: >>>> This also eliminate quite a few zero-extensions in the compile.exp >>>> testsuite >>>> at -O2 on the SPARC. Tested on x86-64/Linux and SPARC/Solaris. >>>> >>>> >>>> 2016-11-07 Eric Botcazou <ebotca...@adacore.com> >>>> >>>> PR rtl-optimization/59461 >>>> * doc/rtl.texi (paradoxical subregs): Add missing word. >>>> * combine.c (reg_nonzero_bits_for_combine): Do not discard results >>>> in modes with precision larger than that of last_set_mode. >>>> * rtlanal.c (nonzero_bits1) <SUBREG>: If WORD_REGISTER_OPERATIONS is >>>> set and LOAD_EXTEND_OP is appropriate, propagate results from inner >>>> REGs to paradoxical SUBREGs. >>>> (num_sign_bit_copies1) <SUBREG>: Likewise. Check that the mode is not >>>> larger than a word before invoking LOAD_EXTEND_OP on it. >>> >>> I have installed it after testing on ARM/EABI and IA-64/Linux. >> >> Just a heads-up for now, the above patch introduced following >> testsuite failures on Alpha [1]: >> >> FAIL: gcc.dg/atomic/stdatomic-compare-exchange-1.c -O2 execution test >> FAIL: gcc.dg/atomic/stdatomic-compare-exchange-1.c -O3 -g execution test >> FAIL: gcc.dg/atomic/stdatomic-compare-exchange-1.c -O2 -flto >> -fno-use-linker-plugin -flto-partition=none execution test >> FAIL: gcc.dg/atomic/stdatomic-compare-exchange-1.c -O2 -flto >> -fuse-linker-plugin -fno-fat-lto-objects execution test >> FAIL: gcc.dg/atomic/stdatomic-compare-exchange-2.c -O2 execution test >> FAIL: gcc.dg/atomic/stdatomic-compare-exchange-2.c -O3 -g execution test >> FAIL: gcc.dg/atomic/stdatomic-compare-exchange-2.c -O2 -flto >> -fno-use-linker-plugin -flto-partition=none execution test >> FAIL: gcc.dg/atomic/stdatomic-compare-exchange-2.c -O2 -flto >> -fuse-linker-plugin -fno-fat-lto-objects execution test >> >> On a first look, it looks that something is wrong with sign-extensions. >> >> I will try to look into this a bit more. > > Your patch exposes a problem in the ree pass where: > > 50: $22:QI=$8:QI > REG_EQUAL [`expected'] > 43: [$2:DI+low(`expected')]=$8:QI > 51: $3:DI=sign_extend($22:QI) > > gets converted to: > > 50: $3:DI=sign_extend($8:QI) > 141: $22:DI=$3:DI > 43: [$2:DI+low(`expected')]=$8:QI > > $22:DI is used later in a fallthrough block, and the above sequences > are certainly not equal wrt. to $22. The first one does have all high > bits of $22 zero, the second one doesn't. > > -fno-ree fixes the failures. > > I will fill a PR tomorrow.
PR 78437 [1]. [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78437 Uros.
