PENG

@tstellar Can this commit be backported to 13.0.1 ?

ping?

But this approach is fine be me, but the check should be in isLeafProc().

LGTM. I can land this for you if you have no commit access, just give me the name/email you want to be credited as.

The fix LGTM, there's only a minor nit wrt the test placement.

LGTM.
My only suggestion would be to drop the popcount test as the bug affects uaddlv4h_from_v8i8, but that's only a minor nit.

I'm not sure this transform is sound, if you consider a simple case such as uaddlv4h_from_v8i8 and perform the expansion by hand you'll see wildly different results.
Given an input vector of u8{1,1,1,1,1,1,1,1} one would expect uaddlp to turn it into u16{2,2,2,2} and uaddlv to fold it into u16{0x8}. On the other hand, replacing everything with a single uaddlv operating on a v0.16b yields u16{0x404} as it skips the pairwise sum step.

ping?

Please upload the diff with full context, check out the contribution guide for more info.

Ping?

In D102985#2830779, @jrtc27 wrote:

Downstream we turned it into a std::pair<uint64_t, uint64_t> (https://github.com/CTSRD-CHERI/llvm-project/blob/de58a2a25c32471b4b3c2aec77fee624e24120cd/lld/ELF/Relocations.cpp#L136-L179), though something more general where it's parameterised on the number of words might be nicer rather than hand-rolling a pair version.

Here's the updated patch, I'm now able to link and run several small programs linked against glibc (in both PIC/PIE mode), shared libraries and some basic examples w/ ifuncs.

In D102985#2829310, @jrtc27 wrote:

In D102985#2829308, @LemonBoy wrote:

In D102985#2829305, @jrtc27 wrote:

Ok, so you need R_RELAX_GOT_OFF to match R_RELAX_GOT_PC?

A relocation that behaves as R_GOT_OFF or, when it's possible to relax it, as a R_GOTREL.

Yes, the R_GOT_OFF version of what R_RELAX_GOT_PC is to R_GOT_PC (e.g. movq foo@GOTPCREL(%rip), %rax -> leaq foo(%rip), %rax, where the latter means leaq foo@PCREL(%rip) but that's implied by the %rip base).

In D102985#2829305, @jrtc27 wrote:

Ok, so you need R_RELAX_GOT_OFF to match R_RELAX_GOT_PC?

In D102985#2827315, @MaskRay wrote:

Is dynamic linking fully functional?

Ping

Add back the RELA bit, no explicit test is needed as the existing ones are already covering that bit.

Alright, folding this back into D102986 (with the getDynRel fix?)

LGTM

In D104185#2815375, @jrtc27 wrote:

The diff looks fine but:

(a) the commit title only mentions truncstore whilst this also fixes both general truncation and extension
(b) the commit body doesn't mention extension

Sorry for the noise, the test got stale since I first uploaded it. I'll push again once this is sorted out.

Address review comments.

In D103232#2787468, @efriedma wrote:

The transform makes sense on targets that don't have atomic operations on floating-point registers, but that isn't all targets. In particular, the GPU targets have floating-point atomic operations, and bitcasting like this might get in the way of the natural lowering there.

If you think it makes sense to use a dedicated lowering hook, rather than extending AtomicExpansionKind, that's fine, I guess.

In D103232#2785963, @efriedma wrote:

I don't really like unconditionally messing with the type of the atomic operation; not sure what kind of impact that will have.

Maybe it would make sense to add a new AtomicExpansionKind?

In D102985#2776038, @MaskRay wrote:

I think it is useful to set up a minimal goal. When you fix/implement GOT/PLT, make sure some basic programs with GOT/PLT are really working.

If we do "fix GOT a" "fot GOT b" "fix PLT c", it is very likely that a subsequent commit may update an earlier commit. This is less elegant.

So what's the state of this port? When I implemented ppc32 support and added riscv support, I made sure glibc/musl ppc32/riscv dynamically linking programs work.

Some (candidate) milestones: (1) support -fno-pic -no-pie -nostdlib link (2) support dynamic linking (many stuff are here: RELATIVE/COPY/GOT/PLT/...) (3) support tls (4) support extra things that glibc static links require

In D102985#2776028, @MaskRay wrote:

I dislike that we are fixing these stuff increasingly with so many patches.

See the history of Arch/PPC.cpp - most stuff are implemented with very few commits.

In D102986#2776010, @jrtc27 wrote:

Can we not have a test that uses textrel; what's wrong with:

.data
x: .xword y

or

.data
x: .xword x

Depends on D102986.

Make this patch depend on D102985

I apparently misunderstood the description about reserved GOT entries.

Rebased on top of D102908.
Correct computation of relaxed tp-relative offset.
Turn add into or/xor during the relaxation.

Reserve 3 slots in the GOT, this detail was well-hidden in the documentation.
Address some stylistic nits.

Fix test formatting.
Force use of rela over rel, forgot to commit this change with the previous patch.

Reinstate the alloca initialization check.

Add a test w/ partially-initialized alloca.

In D102862#2772849, @MaskRay wrote:

Is IE supported now? I think the correct order is LE, IE, GD/LD.

Address review comments.

In D102153#2764464, @rnk wrote:

I don't see why this optimization should be declared illegal. Why isn't it the caller's responsibility to store these narrow integers canonically? C doesn't have narrow integers, but GCC doesn't do this. If a variable is passed in memory, it just uses the memory that was passed into it.

Cleaner and nicer logic.

Ping?

LGTM

LGTM and matches GAS' behaviour.

Address review comments.

Is "tail reloc" a standard term for this?

Distinguish %gdop and TLS-relocations applied to ld and ldx at parser level.

Ping?

Rebased. I've moved the test to icmp-null.ll and made it a bit nicer.

In D101975#2744037, @peter.smith wrote:

LGTM too.

Could I trouble you to add a test case for something like:

        .text
        .arm
        .thumb_func
        nop
sym:
        bx lr

This just makes sure we get the nop before sym in Thumb state even though the symbol doesn't start immediately after the thumb_func.

Ping?

Fix LLD test.

Weekly ping.