[LD] Erroneous R_ARC_NONE in Relocation

[shahab-vahedi]

In this small example, you see the following output for relocations of lib.so:

$ make reloc

arc-gnu-linux-readelf -r lib.so

Relocation section '.rela.dyn' at offset 0x2d4 contains 10 entries:
 Offset     Info    Type            Sym.Value  Sym. Name + Addend
00003f0c  00000038 R_ARC_RELATIVE               9c
00003f10  00000038 R_ARC_RELATIVE               64
00004010  00000038 R_ARC_RELATIVE               0
00000000  00000000 R_ARC_NONE                   0
00000000  00000000 R_ARC_NONE                   0
00000000  00000000 R_ARC_NONE                   0
00003fec  00000536 R_ARC_GLOB_DAT    00000000   _ITM_deregisterTMClone + 0
00003ff0  00000736 R_ARC_GLOB_DAT    00000000   _ITM_registerTMCloneTa + 0
00003ff4  00000836 R_ARC_GLOB_DAT    00000000   __cxa_finalize@GLIBC_2.30 + 0
00003ff8  00000636 R_ARC_GLOB_DAT    00000000   foo + 0

Relocation section '.rela.plt' at offset 0x34c contains 2 entries:
 Offset     Info    Type            Sym.Value  Sym. Name + Addend
00004008  00000637 R_ARC_JMP_SLOT    00000000   foo + 0
0000400c  00000837 R_ARC_JMP_SLOT    00000000   __cxa_finalize@GLIBC_2.30 + 0

There are a few R_ARC_NONEs which shouldn't be.

[claziss]

The issue is due to hidden symbols which are defined in a different object, but still they are local. Because of this fact, the linker was erroneously allocating dynamic section space, while the reloc itself was solved latter on.

A simplified example is this one:

FileA.S:

	.data
	.hidden hidobj
	.global hidobj
	.type hidobj,@object
hidobj:
	.zero 4

FileB.S

	.hidden	hidobj
	.section	.text
	add	r2,pcl,@hidobj@pcl

Linker command line:
arc-elf32-ld --shared -marclinux -o libx.so fileA.o fileB.o

[vineetgarc]

FWIW, there's STAR : 9001086764- build linker to prune extraneous R_ARC_NONE dynamic relos

[claziss]

A tentative patch is here 6fb6db52d75404682fe57ffc62f339555bc0c939

[abrodkin]

@claziss what is our plan with this one?

[cupertinomiranda]

I will make a pull request with 2 commits that at least reduce number of erroneous ARC_NONE relocations.
I will need that pull request to be used to fully execute dejagnu and glibc testsuite on an HSDK before I can consider it to be merged.

This is one of those fixes that never seems to get pulled in.
BTW, dejagnu on Linux HSDK should not take more then 2h of HSDK time.

[abrodkin]

I guess that's the PR @cupertinomiranda has meant: foss-for-synopsys-dwc-arc-processors/binutils-gdb#52.