This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] x86: Replace hard frame pointer with stack pointer - UNITS_PER_WORD
On Mon, Sep 3, 2018 at 9:44 AM, Michael Matz <matz@suse.de> wrote:
> Hi,
>
> On Mon, 3 Sep 2018, H.J. Lu wrote:
>
>> > So, what's the testcase testing then? Before the patch it doesn't ICE,
>> > after the patch it doesn't ICE. What should I look out for so I can see
>> > that what the testcase is producing without the patch is wrong?
>>
>> Before the patch, debug info is wrong since it uses hard frame pointer
>> which isn't set up for the function. You can do "readelf -w" on .o file to
>> verify the debug info.
>
> Yeah, that's what I thought as well, but it's correct:
>
> % ./gcc/cc1plus -quiet -O2 -g -fno-omit-frame-pointer -fvar-tracking x.cc
> % gcc -c x.s
> % readelf -wfi x.o
> ...
> <1><8a>: Abbrev Number: 9 (DW_TAG_subprogram)
> <8b> DW_AT_specification: <0x3a>
> <8f> DW_AT_decl_line : 6
> <90> DW_AT_decl_column : 5
> <91> DW_AT_object_pointer: <0xa7>
> <95> DW_AT_low_pc : 0x0
> <9d> DW_AT_high_pc : 0x3
> <a5> DW_AT_frame_base : 1 byte block: 9c (DW_OP_call_frame_cfa)
> <a7> DW_AT_GNU_all_call_sites: 1
> ...
> <2><fe>: Abbrev Number: 11 (DW_TAG_formal_parameter)
> <ff> DW_AT_name : d
> <101> DW_AT_decl_file : 1
> <102> DW_AT_decl_line : 6
> <103> DW_AT_decl_column : 63
> <104> DW_AT_type : <0x78>
> <108> DW_AT_location : 2 byte block: 91 8 (DW_OP_fbreg: 8)
> ...
> DW_CFA_def_cfa: r7 (rsp) ofs 8
> DW_CFA_offset: r16 (rip) at cfa-8
> DW_CFA_nop
> DW_CFA_nop
> ...
>
> So, argument 'd' is supposed to be at DW_AT_frame_base + 8, which is
> %rsp+8+8, aka %rsp+16, which is correct given that it's the eigth argument
> (including the implicit this parameter).
Can we use DW_AT_frame_base when the frame pointer isn't available?
If yes,
gcc_assert ((SUPPORTS_STACK_ALIGNMENT
&& (elim == hard_frame_pointer_rtx
|| elim == stack_pointer_rtx))
|| elim == (frame_pointer_needed
? hard_frame_pointer_rtx
: stack_pointer_rtx));
should be changed to
gcc_assert (elim == hard_frame_pointer_rtx
|| elim == stack_pointer_rtx);
This will also fix:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86593
> So, can you actually show here what's broken before patch?
>
>> > You talking about this, right:
>> >
>> > /* We only use "frame base" when we're sure we're talking about the
>> > post-prologue local stack frame. We do this by *not* running
>> > register elimination until this point, and recognizing the special
>> > argument pointer and soft frame pointer rtx's. */
>> > if (reg == arg_pointer_rtx || reg == frame_pointer_rtx)
>> > {
>> > rtx elim = (ira_use_lra_p
>> > ? lra_eliminate_regs (reg, VOIDmode, NULL_RTX)
>> > : eliminate_regs (reg, VOIDmode, NULL_RTX));
>> >
>> > if (elim != reg)
>> > {
>> > ...
>> >
>> > So, why would eliminate_regs return hard_frame_pointer_rtx if no frame
>> > pointer is desired?
>>
>> Frame pointer was skipped at the last minute in
>> x86_finalize_stack_frame_flags. But eliminate_regs uses the info which
>> was computed when frame pointer was available.
>
> Let's assume something needs fixing (though I can't reproduce what right
> now) then I think changing frame_pointer_needed somehow needs to affect
> calls to {lra_,}eliminate_regs that come afterwards (by e.g. recalculating
> its info). Everything else is just asking for hacks upon hacks.
>
The only reference to hard frame pointer is in debug info. Is recomputing
eliminate info really necessary?
--
H.J.