Re: RISC-V: Relax RISCV_PCREL_* to RISCV_GPREL_*

[Palmer Dabbelt <palmer at dabbelt dot com>]

On Tue, 17 Oct 2017 10:51:36 PDT (-0700), Palmer Dabbelt wrote:
> On Tue, 10 Oct 2017 15:23:33 PDT (-0700), H.J. Lu wrote:
>> On 10/10/17, Palmer Dabbelt <palmer@dabbelt.com> wrote:
>>> I've had these two patches sitting in my tree for a few months.  I've been
>>> slowly going though and testing all our out-of-tree patches and getting
>>> them
>>> upstream, with these two as the last major ones we have.  I was hoping to
>>> go
>>> convert all our relaxations over to using R_RISCV_DELETE, with the idea
>>> being
>>> that we could avoid the O(n^2) copying we do when relaxing, but it looks
>>> like I
>>> won't get time to do that for a while.
>>>
>>> Despite the patches not being quite done, they do seem to work correctly
>>> (they're been on the branch we at SiFive make releases from for a while),
>>> enable a new optimization, and generally don't seem to hurt -- essentially
>>> they
>>> just add some code that should be used to clean up a few more places.
>>>
>>> I'll commit these to master unless anyone has any objections or notices
>>> anything I screwed up :).
>>>
>>> [PATCH 1/2] RISC-V: Add R_RISCV_DELETE, which marks bytes for
>>> [PATCH 2/2] RISC-V: Relax RISCV_PCREL_* to RISCV_GPREL_*
>>>
>>
>> For  riscv32-linux target, I see:
>>
>> FAIL: difference of two undefined symbols
>> FAIL: align
>> FAIL: .sleb128 tests (2)
>> FAIL: .sleb128 tests (4)
>> FAIL: .sleb128 tests (5)
>> FAIL: .sleb128 tests (7)
>> FAIL: relax .uleb128
>> FAIL: Disabling section padding
>> FAIL: DWARF2 10
>> FAIL: lns-duplicate
>> FAIL: lns-common-1
>> FAIL: ld-elf/compressed1d
>> FAIL: ld-elf/stab
>> FAIL: ld-scripts/fill
>> FAIL: ld-scripts/empty-address-2a
>> FAIL: ld-scripts/empty-address-2b
>> FAIL: PHDRS2
>> FAIL: ld-scripts/size-1
>>
>> Please fix them.
>>
>> Thanks.
>
> Oh, sorry about that -- it turns out we actually haven't been running the
> binutils test suite (we just run the GCC test suite).  It's something we really
> should be doing, it's just hard to find the time to actually get through these
> things with all the other issues we keep finding :).  I've managed to get the
> GAS test suite clean with these two tests, though the readelf one might be
> no good.  Are those two OK for master?
>
> I've added the binutils test suites to our toolchain regression test suite so
> we shouldn't be adding any new test failures on any commits from now on.  I'll
> keep pushing on the other test failures, but it might take a bit to get through
> all of them.
>
> Given that this doesn't add any new test failures, is it OK for master?
>
> Thanks for pointing this out!
>
> [PATCH 1/2] RISC-V: Fix readelf's R_RISCV_{ADD,SUB}{8,16,32,64}
> [PATCH 2/2] RISC-V: Mark unsupported gas testcases

I commited the original patch set, I'm still working through the test suite
failures.