| Message ID | 548B4C26.10709@redhat.com |
|---|---|
| State | New |
| Headers | show |
Hi,
After this commit, GCC build fails for ARM targets:
--target=arm-none-eabi --with-mode=arm --with-cpu=cortex-a9 --with-fpu=neon
/obj-arm-none-eabi/gcc1/./gcc/xgcc -B/obj-arm-none-eabi/gcc1/./gcc/
-B/tools/arm-none-eabi/bin/ -B/tools/arm-none-eabi/lib/ -isystem
/tools/arm-none-eabi/include -isystem /tools/arm-none-eabi/sys-include
-g -O2 -mfloat-abi=hard -O2 -g -O2 -DIN_GCC
-DCROSS_DIRECTORY_STRUCTURE -W -Wall -Wno-narrowing -Wwrite-strings
-Wcast-qual -Wstrict-prototypes -Wmissing-prototypes
-Wold-style-definition -isystem ./include -fno-inline -g
-DIN_LIBGCC2 -fbuilding-libgcc -fno-stack-protector -Dinhibit_libc
-fno-inline -I. -I. -I../../.././gcc -I/trunk/libgcc -I/trunk/libgcc/.
-I/trunk/libgcc/../gcc -I/trunk/libgcc/../include -DHAVE_CC_TLS -o
_mulhelperDQ.o -MT _mulhelperDQ.o -MD -MP -MF _mulhelperDQ.dep
-DL_mulhelper -DDQ_MODE -c /trunk/libgcc/fixed-bit.c
-fvisibility=hidden -DHIDE_EXPORTS
/trunk/libgcc/fixed-bit.c: In function '__gnu_mulhelperdq':
/trunk/libgcc/fixed-bit.c:371:1: error: unable to generate reloads for:
}
^
(insn 55 63 59 2 (set (reg:DI 124 [ D.7630 ])
(mult:DI (zero_extend:DI (subreg:SI (reg:DI 112 [ D.7628 ]) 4))
(zero_extend:DI (subreg:SI (reg:DI 111 [ D.7628 ]) 4))))
/trunk/libgcc/fixed-bit.c:307 54 {*umulsidi3_v6}
(nil))
/trunk/libgcc/fixed-bit.c:371:1: internal compiler error: in
curr_insn_transform, at lra-constraints.c:3383
0x9d4875 _fatal_insn(char const*, rtx_def const*, char const*, int, char const*)
/trunk/gcc/rtl-error.c:110
0x904ec4 curr_insn_transform
/trunk/gcc/lra-constraints.c:3383
0x9057ce lra_constraints(bool)
/trunk/gcc/lra-constraints.c:4324
0x8f5eb1 lra(_IO_FILE*)
/trunk/gcc/lra.c:2277
0x8b4959 do_reload
/trunk/gcc/ira.c:5391
0x8b4959 execute
/trunk/gcc/ira.c:5561
Please submit a full bug report,
Can you have a look?
Thanks,
Christophe.
On 12 December 2014 at 21:12, Vladimir Makarov <vmakarov@redhat.com> wrote:
> The following patch fixes
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64110
>
> The patch was successfully tested and bootstrapped on x86/x86-64.
>
> Committed as rev. 218688.
>
> 2014-12-12 Vladimir Makarov <vmakarov@redhat.com>
>
> PR target/64110
> * lra-constraints.c (process_alt_operands): Refuse alternative
> when reload pseudo of given class can not hold value of given
> mode.
>
> 2014-12-12 Vladimir Makarov <vmakarov@redhat.com>
>
> PR target/64110
> * gcc.target/i386/pr64110.c: New.
>
On 2014-12-15 9:14 AM, Christophe Lyon wrote: > Hi, > > After this commit, GCC build fails for ARM targets: > --target=arm-none-eabi --with-mode=arm --with-cpu=cortex-a9 --with-fpu=neon > full bug report, > > Can you have a look? > Sure. Sorry for inconvenience. LRA/reload bug fixing is a complicated thing in most cases.
Hi, This commit will cause another GCC build fail for ARM targets. The details are descripted in the following Bugzilla linker: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64323. Could you help me to have a look? Thanks, Hale. > -----Original Message----- > From: gcc-patches-owner@gcc.gnu.org [mailto:gcc-patches- > owner@gcc.gnu.org] On Behalf Of Vladimir Makarov > Sent: Saturday, December 13, 2014 4:12 AM > To: GCC Patches > Subject: patch to fix PR64110 > > The following patch fixes > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64110 > > The patch was successfully tested and bootstrapped on x86/x86-64. > > Committed as rev. 218688. > > 2014-12-12 Vladimir Makarov <vmakarov@redhat.com> > > PR target/64110 > * lra-constraints.c (process_alt_operands): Refuse alternative > when reload pseudo of given class can not hold value of given > mode. > > 2014-12-12 Vladimir Makarov <vmakarov@redhat.com> > > PR target/64110 > * gcc.target/i386/pr64110.c: New.
This issue is already fixed by your commit r218760. Thanks. Hale. > -----Original Message----- > From: gcc-patches-owner@gcc.gnu.org [mailto:gcc-patches- > owner@gcc.gnu.org] On Behalf Of Hale Wang > Sent: Tuesday, December 16, 2014 10:17 AM > To: 'Vladimir Makarov'; GCC Patches > Subject: RE: patch to fix PR64110 > > Hi, > > This commit will cause another GCC build fail for ARM targets. The details are > descripted in the following Bugzilla linker: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64323. > > Could you help me to have a look? > Thanks, > > Hale. > > > -----Original Message----- > > From: gcc-patches-owner@gcc.gnu.org [mailto:gcc-patches- > > owner@gcc.gnu.org] On Behalf Of Vladimir Makarov > > Sent: Saturday, December 13, 2014 4:12 AM > > To: GCC Patches > > Subject: patch to fix PR64110 > > > > The following patch fixes > > > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64110 > > > > The patch was successfully tested and bootstrapped on x86/x86-64. > > > > Committed as rev. 218688. > > > > 2014-12-12 Vladimir Makarov <vmakarov@redhat.com> > > > > PR target/64110 > > * lra-constraints.c (process_alt_operands): Refuse alternative > > when reload pseudo of given class can not hold value of given > > mode. > > > > 2014-12-12 Vladimir Makarov <vmakarov@redhat.com> > > > > PR target/64110 > > * gcc.target/i386/pr64110.c: New. > > >
Index: lra-constraints.c =================================================================== --- lra-constraints.c (revision 218685) +++ lra-constraints.c (working copy) @@ -2267,6 +2267,29 @@ process_alt_operands (int only_alternati goto fail; } + /* Alternative loses if it required class pseudo can not + hold value of required mode. Such insns can be + described by insn definitions with mode iterators. + Don't use ira_prohibited_class_mode_regs here as it + is common practice for constraints to use a class + which does not have actually enough regs to hold the + value (e.g. x86 AREG for mode requiring more one + general reg). */ + if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode + && ! hard_reg_set_empty_p (this_alternative_set) + && ! HARD_REGNO_MODE_OK (ira_class_hard_regs + [this_alternative][0], + GET_MODE (*curr_id->operand_loc[nop]))) + { + if (lra_dump_file != NULL) + fprintf + (lra_dump_file, + " alt=%d: reload pseudo for op %d " + " can not hold the mode value -- refuse\n", + nalt, nop); + goto fail; + } + /* Check strong discouragement of reload of non-constant into class THIS_ALTERNATIVE. */ if (! CONSTANT_P (op) && ! no_regs_p Index: testsuite/gcc.target/i386/pr64110.c =================================================================== --- testsuite/gcc.target/i386/pr64110.c (revision 0) +++ testsuite/gcc.target/i386/pr64110.c (working copy) @@ -0,0 +1,17 @@ +/* { dg-do compile } */ +/* { dg-options "-O3 -march=core-avx2" } */ + +int foo (void); +int a; +short *b; + +void +bar (short x) +{ + while (a--) + { + int i, j = foo (); + for (i = 0; i < j; ++i) + *b++ = x; + } +}
The following patch fixes https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64110 The patch was successfully tested and bootstrapped on x86/x86-64. Committed as rev. 218688. 2014-12-12 Vladimir Makarov <vmakarov@redhat.com> PR target/64110 * lra-constraints.c (process_alt_operands): Refuse alternative when reload pseudo of given class can not hold value of given mode. 2014-12-12 Vladimir Makarov <vmakarov@redhat.com> PR target/64110 * gcc.target/i386/pr64110.c: New.