nop and addi x0, x0, 0 should be identical, are not

[sorear]

According to the "Assembly Programmer's Handbook" in the ISA manual, nop is a pseudo-instruction and is equivalent to addi x0, x0, 0. However the current assembler treats them differently and is unable to output a compressed encoding for the latter:

$ cat asdf.s
floor:
        nop
        addi x0, x0, 0
$ as -march=rv32ic asdf.s
$ objdump -d a.out

a.out:     file format elf32-littleriscv


Disassembly of section .text:

00000000 <floor>:
   0:   0001                nop
   2:   00000013            nop

If this behavior is deliberate it should be documented somewhere.

Context: http://lkml.iu.edu/hypermail/linux/kernel/1801.1/06322.html

[jim-wilson]

It looks like an oversight. In the base instruction set, nop is an alias for addi, but in the compressed intsruction set, c.nop and a.addi are different instructions with similar encodings. c.addi with a zero immediate is a hint. c.nop with a non-zero immediate is a hint (in v2.3 draft). So we can't compress addi x0,x0,0 to c.addi, but we could compress it to c.nop. A rule for that is missing in the opcode table.

[sorear]

My perspective is that there are four C instructions corresponding to base addi, c.addi, c.addi4spn, c.addi16sp, and c.nop, and they should all be handled; the current behavior additionally breaks the conceptual model where pseudoinstructions are expanded before compression.

[jim-wilson]

c.addi, c.addi4spn, and c.addi16sp are already handled. It is only the c.nop support that is missing.

[jim-wilson]

I added a patch upstream. This just missed the binutils 2.30 branch. I don't think it is important enough to backport to that release branch, which would require additional testing for little or no gain.