[MC][ELF] Work around R_MIPS_LO16 relocation handling problem

[dim]

Summary

This fixes PR49821, and avoids "ld.lld: error: test.o:(.rodata.str1.1):
offset is outside the section" errors when linking MIPS objects with
negative R_MIPS_LO16 implicit addends.

ld.lld handles R_MIPS_HI16/R_MIPS_LO16 separately, not as a whole, so it
doesn't know that an R_MIPS_HI16 with implicit addend 1 and an
R_MIPS_LO16 with implicit addend -32768 represents 32768, which is in
range of a MergeInputSection. We could introduce a new RelExpr member
(like R_RISCV_PC_INDIRECT for R_RISCV_PCREL_HI20 / R_RISCV_PCREL_LO12)
but the complexity is unnecessary given that GNU as keeps the original
symbol for this case as well.

[MaskRay]

Test?

Both R_MIPS_HI16 and R_MIPS_LO16 are considered as absolute relocations (the type is R_ABS).

R_ABS is ld.lld's annotation . It is not a standard term. You may need to mention the ld.lld qualifier.

[dim]

Add test case, and update comment to explicitly mention ld.lld.

[MaskRay]

This fixes PR49821, and avoids "ld.lld: error: test.o:(.rodata.str1.1): offset is outside the section" errors when linking certain MIPS objects

... with negative R_MIPS_LO16 implicit addends.

A root of the problem is in the R_MIPS_LO16 relocation handling. Both R_MIPS_HI16 and R_MIPS_LO16 are considered as absolute relocations (the ...

Rephrase, with something like: ld.lld handles R_MIPS_HI16/R_MIPS_LO16 separately, not as a whole, so it doesn't know that an R_MIPS_HI16 with implicit addend 1 and an R_MIPS_LO16 with implicit addend -32768 represents 32768, which is in range of a MergeInputSection. We could introduce a new RelExpr member (like R_RISCV_PC_INDIRECT for R_RISCV_PCREL_HI20 / R_RISCV_PCREL_LO12) but the complexity is unnecessary given that GNU as keeps the original symbol for this case as well.

llvm/lib/MC/ELFObjectWriter.cpp

1385

Simplify as my main comment suggests, with potentially less wording.

llvm/test/MC/Mips/mips_lo16.s

4 ↗

(On Diff #342499)

Remove hex bytes with --no-show-raw-insn

5 ↗

(On Diff #342499)

Add mips64 for RELA test

[dim]

Update comment and commit message, and add mips64 test.

[atanasyan]

LGTM

[MaskRay]

A root of the problem is in the R_MIPS_LO16 relocation handling. Both

The description needs to be updated, too.

llvm/test/MC/Mips/mips_lo16.s

1 ↗

(On Diff #342526)

Use the original local symbol for REL targets.

[dim]

Reverted due to the influence this had on some other test cases. I'll do a full check-all and go over them one by one, but it will take some time...

[MaskRay]

Please at least run check-llvm and also look at the Harbomaster failures.

[emaste]

Presumably the MIPS tests changed in 72e75ca343c6f need to be changed back

[jrtc27]

In Debian we were seeing mipsel's stage2 LLVM 12 emitting nonsense textual IR like call void @llvm.memcpy.00G8.00G8.G32(i8* align 1 %3, i8* align 1 %4, i32 0, i1 false) and %0 = tail call float @llvm.nearbyint.toH(float %a) (see https://github.com/rust-lang/rust/issues/85562 / https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=988965), and another build saw different nonsense, but each build was consistent in what it would produce, suggesting miscompilation of LLVM, not run-time memory corruption.

Cherry-picking this patch seems to have fixed that. I don't know whether this is LLVM not emitting the correct HI16/LO16 pair (both in terms of the addends and in terms of being adjacent) for llvm/lib/IR/Function.cpp's string constant pointers or whether there's a bug in GNU ld where it silently does the wrong thing, but either way this does more than just work around an LLD deficiency.

[jrtc27]

This was in libLLVM.so, too, not a statically-linked Clang, so is R_MIPS_GOT16/R_MIPS_LO16, not R_MIPS_HI16/R_MIPS_LO16, which may make a difference.

[dim]

Hm so this is definitely something that we'd want in 12.0.1? @tstellar it doesn't change any ABI, so should be perfectly safe. I'll create a PR.

[dim]

https://bugs.llvm.org/show_bug.cgi?id=50522

llvm/lib/MC/ELFObjectWriter.cpp

if (Flags & ELF::SHF_MERGE) {
       if (C != 0)
         return true;
 
       // gold<2.34 incorrectly ignored the addend for R_386_GOTOFF (9)
       // (http://sourceware.org/PR16794).
       if (TargetObjectWriter->getEMachine() == ELF::EM_386 &&
           Type == ELF::R_386_GOTOFF)
         return true;
+
+      // Work around an issue in MIPS relocation handling: both R_MIPS_HI16 and
+      // R_MIPS_LO16 are considered as absolute relocations (the type is R_ABS).
+      // When we calculate R_MIPS_HI16's addend we find a paired R_MIPS_LO16,
+      // generate a combined addend and return the correct symbol value from the
+      // InputSectionBase::getRelocTargetVA. For R_MIPS_LO16 we do not do that
+      // and use its addend as is.
+      if (TargetObjectWriter->getEMachine() == ELF::EM_MIPS &&
+          !hasRelocationAddend())
+        return true;
     }
 
     // Most TLS relocations use a got, so they need the symbol. Even those that
     // are just an offset (@tpoff), require a symbol in gold versions before
     // 5efeedf61e4fe720fd3e9a08e6c91c10abb66d42 (2014-09-26) which fixed
     // http://sourceware.org/PR16773.
     if (Flags & ELF::SHF_TLS)
       return true;