ThinLTO + CFI may use 8-bit relocations (R_X86_64_8 in ELF) to absolute symbols whose values are in range 0-255. Absolute symbols exist in COFF, but there is no direct analog for the relocation. That is what I suspect the cause of the assertion failure to be.

The solution I had in mind was to use IMAGE_REL_AMD64_SECREL7, and have LLD extend its meaning to allow for 8-bit values and absolute symbols (this may not work with /msvclto though).

Disabled the tests for now in r301235

I got test failures at r301243 still: https://codereview.chromium.org/2843513003

And it looks like the bot from #​0 is still red.

Ouch. r301257. I love CMake.

It looks like the MSVC linker rejects SECRELs pointing to absolute symbols, so this probably won't work with /msvclto.

C:\src\secrel7>type 1.s
.text
.data
.globl p # @​p
.p2align 3
.zero 128
p:
.byte 0x12
.byte 0x34
.byte 0x56
.byte 0x78
.long i@secrel32

C:\src\secrel7>type 2.s
.globl i
i = 42

C:\src\secrel7>\src\llvm-project\ra\bin\llvm-mc -filetype=obj -o 1.obj 1.s

C:\src\secrel7>\src\llvm-project\ra\bin\llvm-mc -filetype=obj -o 2.obj 2.s

C:\src\secrel7>link 1.obj 2.obj /entry:ExitProcess kernel32.lib /subsystem:console
Microsoft (R) Incremental Linker Version 14.00.24215.1
Copyright (C) Microsoft Corporation. All rights reserved.

1.obj : error LNK2016: absolute symbol 'i' used as target of SECREL relocation in section 2
LINK : fatal error LNK1165: link failed because of fixup errors

I ended up fixing this another way at least for now, by avoiding the use of absolute symbols on everything except x86 + ELF. https://reviews.llvm.org/D37883 should re-enable the tests.

The change to re-enable the tests landed in r313379.

mentioned in issue llvm/llvm-bugzilla-archive#32776