C.ADDI x0, 4: illegal instruction ?

Jim Wilson jimw@sifive.com
Thu Jun 27 20:46:00 GMT 2019


On Thu, Jun 27, 2019 at 9:51 AM Alex Rocha Prado <alex.prado@nxp.com> wrote:
>   I have read all the info pointed by Jim and noticed that there is a significant difference between (doc1) “The RISC-V Instruction Set Manual Volume I: User-Level ISA Document Version 2.2” and (doc2) “The RISC-V Instruction Set Manual Volume I: Unprivileged ISA Document Version 20190621-draft”.

You are making multiple errors here.

First of all, you are assuming that the assembler should exactly
implement the ISA, but it doesn't because that isn't useful.  The
assembler for instance will accept "mv a0,a1" even though there is no
mv instruction, and no a0 or a1 register.  The assembler is more
useful when it accepts aliases like these.  The assembler still
accepts scall even though it was renamed to ecall a while ago.  Etc.

You are also pointing at current docs, claiming it says one thing,
while ignoring the fact that yesterday's docs said something
different, and last year's docs said something even more different.
You are also ignoring the fact that there are other docs, like the
RISC-V Reader, that also say other different things.  There are people
who wrote code against old versions of the docs and other docs that
were ambiguous.  It would be cruel to deliberately break their code
just because we changed some text in the ISA docs, text changes that
don't actually change any instruction encodings.

You haven't stated any clear reason why the assembler should stop
accepting c.addi x0,4.  This harms no one.  And dropping this support
clearly harms people who wrote code following old and other doc
versions.  So the reasonable choice is to continue supporting it.

As has already been stated by myself and Palmer, there are ways to fix
this, but they all require a lot of work for very little gain.  We can
add explicit support for hints which we currently don't have, and then
you could disable hint support if you don't want it.  We could add new
assembler syntax for hints instead of re-using existing instruction
syntax like c.addi and c.nop.  We could add explicit ISA version
support, so for instance if someone asks for User ISA 2.3, then we
don't allow c.addi x0,4, but we do allow it if the user asks for User
ISA 2.2.  This would be more sensible if we used the C spec version
number, except I don't think that the C spec version changed when this
problem was resolved, as there was no actual ISA change just some text
changes, and no one considered it important enough to deserve a C spec
number change.

If your actual complaint is that c.addi x0,4 should disassemble as
c.nop 4 instead, then we already have a way to fix that.  The
disasembler will accept -Mno-aliases as an option, and then we just
need to make sure that c.addi and c.nop have the aliases properly
marked.  But when I try this, I see that the instruction already
disassemble as c.nop 4 instead of c.addi x0,4, so no change seems to
be necessary and no -Mno-aliases option is necessary.

Jim



More information about the Binutils mailing list