"H.J. Lu" <hjl.tools@gmail.com> writes:
> On Tue, Sep 15, 2020 at 7:44 AM Richard Sandiford
> <richard.sandiford@arm.com> wrote:
>>
>> Thanks for looking at this.
>>
>> "H.J. Lu" <hjl.tools@gmail.com> writes:
>> > commit 1bcb4c4faa4bd6b1c917c75b100d618faf9e628c
>> > Author: Richard Sandiford <richard.sandiford@arm.com>
>> > Date:   Wed Oct 2 07:37:10 2019 +0000
>> >
>> >     [LRA] Don't make eliminable registers live (PR91957)
>> >
>> > didn't make eliminable registers live which breaks
>> >
>> > register void *cur_pro asm("reg");
>> >
>> > where "reg" is an eliminable register.  Make fixed eliminable registers
>> > live to fix it.
>>
>> I don't think fixedness itself is the issue here: it's usual for at
>> least some registers involved in eliminations to be fixed registers.
>>
>> I think what makes this case different is instead that cur_pro/ebp
>> is a global register.  But IMO things have already gone wrong if we
>> think that a global register is eliminable.
>>
>> So I wonder if instead we should check global_regs at the beginning of:
>>
>>       for (i = 0; i < fp_reg_count; i++)
>>         if (!TEST_HARD_REG_BIT (crtl->asm_clobbers,
>>                                 HARD_FRAME_POINTER_REGNUM + i))
>>           {
>>             SET_HARD_REG_BIT (eliminable_regset,
>>                               HARD_FRAME_POINTER_REGNUM + i);
>>             if (frame_pointer_needed)
>>               SET_HARD_REG_BIT (ira_no_alloc_regs,
>>                                 HARD_FRAME_POINTER_REGNUM + i);
>>           }
>>         else if (frame_pointer_needed)
>>           error ("%s cannot be used in %<asm%> here",
>>                  reg_names[HARD_FRAME_POINTER_REGNUM + i]);
>>         else
>>           df_set_regs_ever_live (HARD_FRAME_POINTER_REGNUM + i, true);
>>
>> (ira_setup_eliminable_regset), and handle the global_regs[] case in
>> the same way as the else case, i.e. short-circuiting both of the ifs.
>>
>
> Like this?

Sorry for the delay.  I was testing this in parallel.

Bootstrapped & regression-tested on x86_64-linux-gnu.

Thanks,
Richard