This is the mail archive of the
mailing list for the binutils project.
Re: [Patch] Gas support for MIPS Compact EH
- From: Richard Sandiford <rdsandiford at googlemail dot com>
- To: Bernd Schmidt <bernds at codesourcery dot com>
- Cc: "Moore\, Catherine" <Catherine_Moore at mentor dot com>, "binutils\ at sourceware dot org" <binutils at sourceware dot org>
- Date: Sun, 09 Feb 2014 11:11:23 +0000
- Subject: Re: [Patch] Gas support for MIPS Compact EH
- Authentication-results: sourceware.org; auth=none
- References: <FD3DCEAC5B03E9408544A1E416F11242F8FC5972 at NA-MBX-01 dot mgc dot mentorg dot com> <87k3me9jia dot fsf at talisman dot default> <FD3DCEAC5B03E9408544A1E416F11242012EAC2AE0 at NA-MBX-01 dot mgc dot mentorg dot com> <8738jt5zt1 dot fsf at talisman dot default> <52F67C03 dot 2050609 at codesourcery dot com>
Bernd Schmidt <email@example.com> writes:
> On 02/08/2014 05:34 PM, Richard Sandiford wrote:
>>>> The tc_cfi_fix_eh_ref and tc_cfi_emit_expr hooks don't seem very
>>>> consistent; the former relies on the caller to clear the bytes, whereas the
>>>> latter is supposed to do it itself.
>>> All fixed, now using a hook to return a reloc and eliminated the use of
>>> R_MIPS_EH from the assembler.
>> Hmm, but how does it work under the new scheme? It looks like gas now
>> always emits the .eh_frame_entry sections using R_MIPS_PC32, is that right?
>> But the linker chooses the .eh_frame_hdr encoding based on --pcrel-eh-reloc,
>> which also controls how R_MIPS_EH is handled. So if the:
>> DW_EH_PE_sdata4 | DW_EH_PE_datarel | DW_EH_PE_indirect
>> encoding is chosen for the .eh_frame_entry sections at link time, what
>> converts the input sections to use that encoding instead of the original
>> R_MIPS_PC32? I'd assumed R_MIPS_EH was defined the way it was to avoid
>> that kind of thing.
> What's changed is that the linker is no longer really involved in these
> decisions - the code you see in the linker-specific parts of the patch
> are there merely to deal with R_MIPS_EH relocs in object files generated
> by previous toolchains. We've always kind of already decided at compile
> time which encoding to use and passed the -pcrel-eh-reloc option to the
> linker to ensure it made the choice we wanted. What's new in this
> version of the patches is based on the realization that gcc can produce
> datarel|indirect encoding without linker help (using the new forcegpword
Ah, OK, so it's now up to the assembler writer to do the indirection
by hand? I hadn't realised that. (To be fair, the patch added the
.forcegpword directive but didn't have any examples or tests to show how
it was used, or any documentation explaining it. I deliberately didn't
complain about the latter though because most ops are undocumented.)
Using GP-relative is probably a bad idea for MIPS because the GP base
can vary in the case of multigot. When resolving the relocations in the
individual input sections, the GP used will be for that input bfd's GOT,
which isn't necessarily going to be the primary GOT. So if we're doing
the indirection by hand, why not use pcrel|indirect instead? I suppose
that amounts to using .ehword (under the new PC-relative definition)
rather than .forcegpword. I think it'd be better to drop .forcegpword
FWIW, pcrel|indirect is what the linker tries to use for .eh_frame on MIPS,
if the original input object used absolute indirect.
> On Linux targets you'll see this generated by the compiler, on
> bare-metal you'll get R_MIPS_PC32. The R_MIPS_EH reloc is longer
> produced. So it's all a lot more straightforward, directly producing an
> encoding appropriate for the target at compile time.
>> I thought R_MIPS_EH would be used for the .eh_frame_entry entries only,
>> since in that case the actual encoding of the address isn't known
>> until link time.
> I think previous versions of the code were just slightly confused - the
> idea was that datarel|indirect required things to be put into the got,
> which has to be done by the linker. It turns out that this isn't
> necessary, so there is no longer a need to use R_MIPS_EH.
> Does this clarify things?
I don't think it really answers my first question. It's still the
linker that decides what encoding goes in the .eh_frame_hdr.
Then all the following .eh_frame_entry sections must use that
encoding for the text addresses.
The input objects use relocations to mark those .eh_frame_entry text
addresses. In the original scheme those relocations were R_MIPS_EH,
giving the linker control of both the .eh_frame_hdr encoding and the
.eh_frame_entry fields that that encoding controls. That part seemed
consistent and safe. (The unsafe part was that other parts of the
assembler seemed to assume that R_MIPS_EH always meant datarel-indirect.)
In the new scheme the assembler picks an encoding-specific relocation
for those .eh_frame_entry text addresses (always PC-relative in the posted
patch), but the linker is still the one that chooses the .eh_frame_hdr
encoding. And in the posted patch the encoding used in .eh_frame_hdr is
decided by the --pcrel-eh-reloc option. The default linker behaviour is
to use datarel-indirect, which is the opposite of what the assembler now
uses. So if I do:
as foo.s -o foo.o
ld foo.o -o foo
it looked like foo.o would use a PC-relative encoding for any
.eh_frame_entry sections, but foo's .eh_frame_hdr would say that
they are datarel-indirect rather than PC-relative.
In other words, I think using datarel-indirect for the .eh_frame_hdr is
only safe if all input objects are using the old scheme. So that seems
like a bad default. And I'm not sure --pcrel-eh-reloc is safe for old
objects because of the assumption in some cases that R_MIPS_EH would
Maybe for the FSF version we should just drop the R_MIPS_EH stuff
altogether, since at this point it sounds like it would be safer
to reject the old objects than to try to guess what's happening.
The .eh_frame_hdr could then be hard-coded to use a PC-relative
encoding, like the assembler is now. It seems a bit of a shame
in a way though -- and like I say, using PC-relative wouldn't
apply well to targets like VxWorks where the gap between the text
and data segments isn't fixed at link time.