[llvm][Mips] Assembler does not assemble `j $ra` correctly for r6

[alexrp]

❯ cat test.s
.global __start
__start:
    j $ra
❯ clang -target mipsisa64r6-linux-gnu -nostdlib test.s
❯ objdump --disassemble=__start a.out

a.out:     file format elf64-tradbigmips


Disassembly of section .text:

00000000000003f0 <__start>:
 3f0:   03e00008        .word   0x3e00008
 3f4:   00000000        nop
❯ mips64-linux-gnuabi64-gcc -march=mips64r6 -nostdlib test.s
❯ objdump --disassemble=__start a.out

a.out:     file format elf64-tradbigmips


Disassembly of section .text:

0000000120000150 <__start>:
   120000150:   03e00009        jr      ra
   120000154:   00000000        nop
        ...

It works fine for versions prior to r6.

This occurs in musl: https://git.musl-libc.org/cgit/musl/tree/crt/mips64/crtn.s

[alexrp]

jr $ra works as expected; it's just the j $ra handling that's wrong.

[llvmbot]

@llvm/issue-subscribers-backend-mips

Author: Alex Rønne Petersen (alexrp)

``` ❯ cat test.s .global __start __start: j $ra ❯ clang -target mipsisa64r6-linux-gnu -nostdlib test.s ❯ objdump --disassemble=__start a.out

a.out: file format elf64-tradbigmips

Disassembly of section .text:

00000000000003f0 <__start>:
3f0: 03e00008 .word 0x3e00008
3f4: 00000000 nop
❯ mips64-linux-gnuabi64-gcc -march=mips64r6 -nostdlib test.s
❯ objdump --disassemble=__start a.out

a.out: file format elf64-tradbigmips

Disassembly of section .text:

0000000120000150 <__start>:
120000150: 03e00009 jr ra
120000154: 00000000 nop
...

It works fine for versions prior to r6.

This occurs in musl: https://git.musl-libc.org/cgit/musl/tree/crt/mips64/crtn.s
</details>

[yingopq]

jr $ra works as expected; it's just the j $ra handling that's wrong.

The MIPS64 instruction manual states "The Jump instruction has been deprecated in Release 6. Use BC instead."

[alexrp]

The MIPS64 instruction manual states "The Jump instruction has been deprecated in Release 6. Use BC instead."

That's for a different instruction ("Jump", i.e. to an offset). j <reg> is supposed to assemble to jr <reg> ("Jump Register", i.e. register indirect). It does so, but as noted in the issue description, it uses the pre-r6 encoding even for r6. GNU as assembles it correctly.

It's a bit confusing because assemblers have historically accepted the j <reg> "alias" form which doesn't make a lot of sense relative to the ISA manual. Unfortunately, this form is quite common in the wild, used by both glibc and musl.

[yingopq]

The MIPS64 instruction manual states "The Jump instruction has been deprecated in Release 6. Use BC instead."

That's for a different instruction ("Jump", i.e. to an offset). j <reg> is supposed to assemble to jr <reg> ("Jump Register", i.e. register indirect). It does so, but as noted in the issue description, it uses the pre-r6 encoding even for r6. GNU as assembles it correctly.

It's a bit confusing because assemblers have historically accepted the j <reg> form which doesn't make a lot of sense relative to the ISA manual. Unfortunately, this form is quite common in the wild, used by both glibc and musl.

OK, I would do investigation.