[Bug c/64012] New: GCC-4.9.2 option -fcaller-saves in -O2

["airbak.li at huawei dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64012

            Bug ID: 64012
           Summary: GCC-4.9.2 option -fcaller-saves in -O2
           Product: gcc
           Version: 4.9.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: airbak.li at huawei dot com

A cross toolchain build by GCC-4.9.2, when I use this toolchain to compile
kernel-3.4.35 with option -O2, it always crash when system power up.
(CPU : armv-7 cortex-a9)

-----------Crash log------------------------------
Unable to handle kernel paging request at virtual address ffffffff
pgd = c0004000
[ffffffff] *pgd=85ffe821, *pte=00000000, *ppte=00000000
Internal error: Oops: 17 [#1] SMP ARM
Modules linked in:
CPU: 1    Not tainted  (3.4.35_hi3535 #27)
PC is at kmem_cache_alloc+0x54/0x130
LR is at con_insert_unipair+0xc0/0x10c
pc : [<c00a5764>]    lr : [<c028447c>]    psr: a0000013
sp : c5831e60  ip : c05b27d8  fp : c5831e94
r10: 00000263  r9 : 00000003  r8 : 0000005f
r7 : 0000005f  r6 : 000000d0  r5 : c5802200  r4 : ffffffff
r3 : 00000000  r2 : c5830000  r1 : c05a69e0  r0 : 0011e000
Flags: NzCv  IRQs on  FIQs on  Mode SVC_32  ISA ARM  Segment kernel
Control: 10c53c7d  Table: 8000404a  DAC: 00000015
-------------Assembler code-------------------------
c00a5710 <kmem_cache_alloc>:
c00a5710:    e1a0c00d     mov    ip, sp
c00a5714:    e92ddff0     push    {r4, r5, r6, r7, r8, r9, sl, fp, ip, lr, pc}
c00a5718:    e24cb004     sub    fp, ip, #4
c00a571c:    e24dd00c     sub    sp, sp, #12
c00a5720:    e1a0300d     mov    r3, sp
c00a5724:    e3c32d7f     bic    r2, r3, #8128    ; 0x1fc0
c00a5728:    e302c7d8     movw    ip, #10200    ; 0x27d8
c00a572c:    e3c2203f     bic    r2, r2, #63    ; 0x3f
c00a5730:    e34cc05b     movt    ip, #49243    ; 0xc05b
c00a5734:    e1a05000     mov    r5, r0
c00a5738:    e1a06001     mov    r6, r1
c00a573c:    e50be030     str    lr, [fp, #-48]    ; 0x30
c00a5740:    e5923014     ldr    r3, [r2, #20]
c00a5744:    e5951000     ldr    r1, [r5]
c00a5748:    e79c0103     ldr    r0, [ip, r3, lsl #2]
c00a574c:    e0813000     add    r3, r1, r0
c00a5750:    e593a004     ldr    sl, [r3, #4]
c00a5754:    e7914000     ldr    r4, [r1, r0]
c00a5758:    e3540000     cmp    r4, #0
c00a575c:    0a000026     beq    c00a57fc <kmem_cache_alloc+0xec>
c00a5760:    e5953014     ldr    r3, [r5, #20]
c00a5764:    e7949003     ldr    r9, [r4, r3]    --->r4 is ffffffff
c00a5768:    e10f8000     mrs    r8, CPSR
----------------------------------------------

When I compile kernel with "-O1" or "-O2 -fno-caller-saves", it is ok, no
crash.
(GCC-4.8 not have this problem)

Question:
1. Please tell me about option "-fcaller-saves" in GCC-4.9.2. Is it different
from GCC-4.8?

2. The error happen in code
"con_insert_unipair()->kmem_cache_alloc()->slab_alloc()", file path is
"linux-3.4.35\mm\slub.c".
What codes will "-fcaller-saves" optimize? Please give me some samples.

3."-fcaller-saves" may be not the root cause.
Generally, we use -O2 include "-fcaller-saves", we want to keep
"-fcaller-saves", please help me to solve it.