On Fri, Oct 10, 2014 at 12:43 AM, Evgeny Stupachenko <[email protected]> wrote:
> i386 specific part of the patch:
>
> 2014-10-08 Ilya Enkovich <[email protected]>
> Vladimir Makarov <[email protected]>
> * gcc/config/i386/i386.c (ix86_use_pseudo_pic_reg): New.
> (ix86_init_pic_reg): New.
> (ix86_select_alt_pic_regnum): Add check on pseudo register.
> (ix86_save_reg): Likewise.
> (ix86_expand_prologue): Remove irrelevant code.
> (ix86_output_function_epilogue): Add check on pseudo register.
> (set_pic_reg_ever_alive): New.
> (legitimize_pic_address): Replace df_set_regs_ever_live with new
> set_pic_reg_ever_alive.
> (legitimize_tls_address): Likewise.
> (ix86_pic_register_p): New check.
> (ix86_delegitimize_address): Add check on pseudo register.
> (ix86_expand_call): Insert move from pseudo PIC register to ABI
> defined REAL_PIC_OFFSET_TABLE_REGNUM.
> (TARGET_INIT_PIC_REG): New.
> (TARGET_USE_PSEUDO_PIC_REG): New.
> (PIC_OFFSET_TABLE_REGNUM): New check.
This patch breaks glibc's ld.so on i686.
glibc has a check to make sure the PIC register is setup correctly:
/* Consistency check for position-independent code. */
#ifdef __PIC__
# define check_consistency() \
({ int __res; \
__asm__ __volatile__ \
(LOAD_PIC_REG_STR (cx) ";" \
"subl %%ebx, %%ecx;" \
"je 1f;" \
"ud2;" \
"1:\n" \
: "=c" (__res)); \
__res; })
#endif
This depends on ebx being the PIC register. Now we don't have this so
we get ud2 in some cases.
Thanks,
Andrew Pinski
